refactor(cli): isolate local agent runtime

- Fail outbound sends while the app is rebuilding instead of returning quietly, so ChannelManager retries and logs the failure instead of counting the message as delivered
- Close the bot's HTTPX pools during teardown: Application.shutdown() skips them when initialize() never finished, leaking one pool per startup retry
- Propagate terminal startup errors (rejected token, bad proxy, bound webhook port) and clear the running flag instead of retrying forever while the channel still reports itself healthy
- Restrict startup retries to NetworkError/TimedOut, the exceptions HTTPXRequest wraps every httpx failure into
- Scrub the rejected-token failure so PTB's token-bearing message never reaches the log

.github/workflows/ci.yml

@@ -96,8 +96,9 @@ jobs:
             os: windows-latest
             python-version: "3.14"
             coverage: false
-            # Keep each test file in one worker while using both hosted-runner cores.
-            pytest_args: "-n 2 --dist loadfile"
+            # Real PowerShell/process-tree tests run serially below. Keep
+            # them out of xdist so workers never share a Windows console.
+            pytest_args: "-n 2 --dist loadfile --ignore=tests/tools/test_exec_platform.py"
 
     steps:
       - uses: actions/checkout@v4
@@ -147,6 +148,13 @@ jobs:
           ${{ matrix.pytest_args }}
           --durations=25 --durations-min=1.0
 
+      - name: Run Windows process compatibility tests
+        if: runner.os == 'Windows'
+        run: >-
+          uv run --no-sync python -m pytest
+          tests/tools/test_exec_platform.py
+          --durations=25 --durations-min=1.0
+
   webui:
     runs-on: ubuntu-latest
     timeout-minutes: 15
@@ -179,6 +187,77 @@ jobs:
         working-directory: webui
         run: bun run build
 
+  tui:
+    name: ${{ matrix.name }}
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: Terminal UI
+            os: ubuntu-latest
+          - name: Terminal UI (Windows)
+            os: windows-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: 1.3.13
+
+      - name: Install TUI dependencies
+        working-directory: tui
+        run: bun install --frozen-lockfile
+
+      - name: Check TUI
+        working-directory: tui
+        run: bun run check
+
+      - name: Test TUI
+        working-directory: tui
+        run: bun run test
+
+      - name: Test TUI in a real pseudo-terminal
+        if: runner.os != 'Windows'
+        working-directory: tui
+        run: python3 scripts/pty_smoke.py
+
+      - name: Set up Python for ConPTY smoke test
+        if: runner.os == 'Windows'
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.13'
+
+      - name: Test TUI in a real ConPTY terminal
+        if: runner.os == 'Windows'
+        working-directory: tui
+        shell: pwsh
+        run: |
+          python -m pip install --disable-pip-version-check pywinpty==3.0.5
+          python scripts/conpty_smoke.py
+
+      - name: Build TUI
+        working-directory: tui
+        run: bun run build
+
+      - name: Verify licensed Linux release archive
+        if: runner.os == 'Linux'
+        working-directory: tui
+        run: |
+          bun scripts/release-notices.ts linux-x64
+          python3 scripts/package-release.py linux-x64
+
+      - name: Verify licensed Windows release archive
+        if: runner.os == 'Windows'
+        working-directory: tui
+        shell: pwsh
+        run: |
+          bun scripts/release-notices.ts win32-x64
+          python scripts/package-release.py win32-x64
+
   docker:
     runs-on: ubuntu-latest
     timeout-minutes: 20
@@ -189,6 +268,44 @@ jobs:
       - name: Build image with default channel dependencies
         run: docker build -t nanobot:test .
 
+      - name: Verify Docker Compose startup and privilege boundary
+        env:
+          HOME: ${{ runner.temp }}
+        run: |
+          docker compose run --rm --no-deps --build -T nanobot-cli status
+          docker compose run --rm --no-deps -T --entrypoint sh nanobot-cli -s <<'OUTER'
+          set -eu
+          field() {
+            awk -v key="$1:" '$1 == key { print $2 }' /proc/self/status
+          }
+          test "$(id -u)" = "0"
+          test "$(field NoNewPrivs)" = "1"
+          setpriv --reuid=nanobot --regid=nanobot --init-groups sh -s <<'INNER'
+          set -eu
+          field() {
+            awk -v key="$1:" '$1 == key { print $2 }' /proc/self/status
+          }
+          test "$(id -u)" = "1000"
+          test "$(field NoNewPrivs)" = "1"
+          for capability_set in CapInh CapPrm CapEff CapAmb; do
+            test "$(field "$capability_set")" = "0000000000000000"
+          done
+          INNER
+          OUTER
+          docker compose -f docker-compose.yml -f docker-compose.bwrap.yml --profile cli \
+            config --format json > "${RUNNER_TEMP}/bwrap-compose.json"
+          python - <<'PY'
+          import json
+          import os
+          from pathlib import Path
+
+          config = json.loads(Path(os.environ["RUNNER_TEMP"], "bwrap-compose.json").read_text())
+          for service_name in ("nanobot-gateway", "nanobot-api", "nanobot-cli"):
+              service = config["services"][service_name]
+              assert {"CHOWN", "SETGID", "SETUID", "SYS_ADMIN"} <= set(service["cap_add"])
+              assert "no-new-privileges:true" in service["security_opt"]
+          PY
+
       - name: Verify default WhatsApp dependencies
         run: docker run --rm --entrypoint python nanobot:test -c "import neonize, segno"
 

.github/workflows/tui-release.yml

@@ -0,0 +1,96 @@
+name: Publish Terminal UI
+
+on:
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: Existing release tag (for example, v0.3.1)
+        required: true
+        type: string
+      compliance_reviewed:
+        description: Confirm notices, source offer, source archive, and relinking were reviewed
+        required: true
+        type: boolean
+        default: false
+
+permissions:
+  contents: write
+
+jobs:
+  build:
+    if: ${{ inputs.compliance_reviewed }}
+    name: ${{ matrix.target }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    strategy:
+      fail-fast: false
+      matrix:
+        target:
+          - darwin-arm64
+          - darwin-x64
+          - linux-arm64
+          - linux-x64
+          - win32-x64
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.tag }}
+          fetch-depth: 0
+
+      - name: Verify release tag
+        env:
+          GH_TOKEN: ${{ github.token }}
+          TAG: ${{ inputs.tag }}
+        shell: bash
+        run: |
+          [[ "$TAG" =~ ^v[0-9]+\.[0-9]+\.[0-9]+([.-][0-9A-Za-z.-]+)?$ ]]
+          gh release view "$TAG" >/dev/null
+          test "$(git rev-parse HEAD)" = "$(git rev-list -n 1 "refs/tags/$TAG")"
+
+      - name: Set up Bun
+        uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: 1.3.13
+
+      - name: Install dependencies
+        working-directory: tui
+        run: bun install --frozen-lockfile
+
+      - name: Install ${{ matrix.target }} native dependencies
+        working-directory: tui
+        run: bun scripts/prepare-target.ts ${{ matrix.target }}
+
+      - name: Build ${{ matrix.target }}
+        working-directory: tui
+        run: bun run build -- ${{ matrix.target }}
+
+      - name: Ad-hoc sign macOS executable
+        if: startsWith(matrix.target, 'darwin-')
+        uses: indygreg/apple-code-sign-action@44d0985b7f4363198e80b6fea63ac3e9dd3e9957 # v1
+        with:
+          input_path: tui/dist/nanobot-tui-${{ matrix.target }}
+          rcodesign_version: 0.29.0
+
+      - name: Build notices and release archive
+        working-directory: tui
+        env:
+          TARGET: ${{ matrix.target }}
+        run: |
+          bun scripts/release-notices.ts "$TARGET"
+          python3 scripts/package-release.py "$TARGET"
+
+      - name: Upload release assets
+        env:
+          GH_TOKEN: ${{ github.token }}
+          TAG: ${{ inputs.tag }}
+          TARGET: ${{ matrix.target }}
+        shell: bash
+        run: |
+          gh release view "$TAG" >/dev/null
+          asset="nanobot-tui-${TARGET}"
+          if [[ "$TARGET" == win32-* ]]; then asset="${asset}.exe"; fi
+          gh release upload "$TAG" \
+            "tui/dist/${asset}.zip" \
+            "tui/dist/${asset}.zip.sha256" \
+            --clobber

.gitignore

@@ -16,6 +16,8 @@ webui/node_modules/
 webui/dist/
 webui/coverage/
 webui/.vite/
+tui/node_modules/
+tui/dist/
 *.tsbuildinfo
 
 # Python bytecode & caches

CONTRIBUTING.md

@@ -136,6 +136,29 @@ GitHub Actions' free tier:
 If your change genuinely needs to step outside this, please call it out
 explicitly in the PR description so it can be discussed before merge.
 
+## Release Packaging Contract
+
+A stable install must never combine Python from one version with a TUI from another. Publish in
+this order:
+
+1. Set the package version and publish the matching GitHub release tag (`vX.Y.Z`).
+2. Review the pinned Bun/OpenTUI licenses, source offer, and relinking materials for that tag.
+3. Manually run **Publish Terminal UI** for the exact tag and confirm the compliance review input.
+4. Wait for every platform archive and checksum to appear on the release, then publish the same
+   `X.Y.Z` package to PyPI.
+
+The wheel contains the built WebUI. The native TUI stays a platform-specific release sidecar so
+users download only the archive for their machine. Each archive must contain the executable,
+target-specific third-party notices, project and runtime licenses, corresponding application
+source, a written source offer, relinking instructions, and a checksum manifest. Never upload a
+naked TUI executable. Source checkouts use an editable Python install, run `tui/` with Bun, and
+rebuild stale `webui/` assets locally.
+
+The confirmation is an operational commitment, not a cosmetic checkbox. Before accepting it,
+verify that the exact Bun/WebKit revisions remain retrievable and that the project can honor the
+archive's corresponding-source offer for its full stated period. Preserve published archives and
+their source materials.
+
 ## Questions?
 
 If you have questions, ideas, or half-formed insights, you are warmly welcome here.

README.md

@@ -77,7 +77,12 @@ nanobot is a self-hosted personal AI agent runtime. It can:
 
 Pick **one** install method:
 
-Prerequisites: Python 3.11 or newer. Git is only needed for a source install. Published packages already include the WebUI; a current-source install needs `bun` or `npm` to build it.
+| Track | Install with | Update with | What runs |
+|---|---|---|---|
+| Stable | installer, `uv`, or pip | the same package tool | one released Python/WebUI/TUI version |
+| Current source | editable Git checkout | `git pull --ff-only` + editable dependency sync | Python, WebUI, and TUI from that checkout |
+
+Prerequisites: Python 3.11 or newer. Git and [Bun](https://bun.sh/) are only needed for a source install. Published packages include the WebUI and fetch a checksummed, version-matched TUI archive—with its licenses, notices, corresponding application source, source offer, and relinking instructions—on first use.
 
 If terminals, API keys, or config files are new to you, use the guided zero-background walkthrough in [Start Without Technical Background](./docs/start-without-technical-background.md) instead of this compact README path.
 
@@ -97,7 +102,7 @@ irm https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.ps1 | i
 
 The default command installs or upgrades `nanobot-ai` from PyPI. On a fresh local desktop, it then starts `nanobot webui` so you can configure the first provider and model in **Settings → Models**. SSH, headless, existing-config, and older-release paths keep the terminal setup wizard. The installer avoids system-wide pip installs by using an active virtual environment, `uv`, `pipx`, or a managed venv under `~/.nanobot/venv`. It also prints the exact command it used to run nanobot; reuse that full command below if `nanobot` is not on `PATH`.
 
-To preview the plan without changing your environment, pass `--dry-run`; combine it with `--dev` when you want to preview the main-branch install.
+To preview the plan without changing your environment, pass `--dry-run`.
 
 ```bash
 curl -fsSL https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.sh | sh -s -- --dry-run
@@ -107,16 +112,6 @@ curl -fsSL https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.
 & ([scriptblock]::Create((irm https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.ps1))) --dry-run
 ```
 
-To install the current `main` branch instead, pass `--dev`:
-
-```bash
-curl -fsSL https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.sh | sh -s -- --dev
-```
-
-```powershell
-& ([scriptblock]::Create((irm https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.ps1))) --dev
-```
-
 If you prefer to inspect the script first, open [`scripts/install.sh`](./scripts/install.sh) or [`scripts/install.ps1`](./scripts/install.ps1).
 
 **Install with `uv`**
@@ -135,15 +130,27 @@ If pip reports `externally-managed-environment` on macOS or Linux, use the one-c
 
 **Install from source**
 
-`bun` or `npm` must be available. From an activated virtual environment:
+Clone the repository and install it in editable mode. Bun is required because the source
+checkout runs the matching TUI directly instead of downloading an older release binary.
 
 ```bash
 git clone https://github.com/HKUDS/nanobot.git
 cd nanobot
-python -m pip install .
+python -m venv .venv
 ```
 
-On Windows, if pip reports that it cannot launch `npm`, run `cd webui`, `npm.cmd install --package-lock=false`, `npm.cmd run build`, and `cd ..` in order, then retry the install. Contributors who need an editable checkout should follow [`CONTRIBUTING.md`](./CONTRIBUTING.md) and [`webui/README.md`](./webui/README.md).
+Activate it with `source .venv/bin/activate` on macOS/Linux or
+`.venv\Scripts\Activate.ps1` in Windows PowerShell, then run:
+
+```bash
+python -m pip install -e .
+```
+
+After that, the normal commands are identical to a stable install. `nanobot agent` runs the TUI
+from this checkout, and `nanobot webui` rebuilds stale frontend assets automatically. A later
+`git pull --ff-only` updates the Python, TUI, and WebUI source together; rerun
+`python -m pip install -e .` when Python dependencies change. Contributors should also read
+[`CONTRIBUTING.md`](./CONTRIBUTING.md).
 
 Verify the install:
 
@@ -161,7 +168,7 @@ If `nanobot` is not on `PATH`, invoke it through the method that installed it: r
 nanobot webui
 ```
 
-This is the recommended first run. The launcher creates the config and workspace when needed, safely enables the local WebSocket channel after confirmation, starts the gateway, and opens [`http://127.0.0.1:8765`](http://127.0.0.1:8765). A fresh install can open before a model is configured, so setup continues in the browser instead of beginning in a JSON file. The first-run WebUI binds to localhost by default and is not exposed to your LAN.
+This is the recommended first run. The launcher creates the config and workspace when needed, safely enables the local WebSocket channel after confirmation, starts or joins the shared local gateway, and opens [`http://127.0.0.1:8765`](http://127.0.0.1:8765). A fresh install can open before a model is configured, so setup continues in the browser instead of beginning in a JSON file. The first-run WebUI binds to localhost by default and is not exposed to your LAN.
 
 **Your first three steps**
 
@@ -174,10 +181,10 @@ Any normal reply means the provider, model, workspace, and browser gateway are w
 **Keep nanobot running after you close the terminal**
 
 ```bash
-nanobot webui --background
+nanobot gateway --background
 ```
 
-This starts the same full gateway as `nanobot webui`, opens the browser, and leaves channels and automations running after the launcher exits. Complete first-time model setup with foreground `nanobot webui` before switching to background mode.
+This is the only command that promotes the shared gateway to persistent background mode. It leaves channels and automations running after every local TUI and WebUI launcher exits. Complete first-time model setup with `nanobot webui` before switching to background mode; open the same localhost WebUI again afterward.
 
 ```bash
 nanobot gateway status
@@ -202,7 +209,7 @@ Use `nanobot gateway --background` for the same direct entry point without keepi
 nanobot agent
 ```
 
-This opens an interactive terminal chat with the same configured model, workspace, and tools while keeping its own CLI session history. It does not open a browser or keep chat channels and automations running after you exit. Type `exit` or press `Ctrl+C` when you are done.
+This opens the native terminal client with the same configured model, workspace, tools, streaming protocol, and session engine as the WebUI. Use `/sessions` to switch saved conversations, `/new-chat` to preserve this conversation and start another one, `/branch` to fork from a completed reply, `/context` to inspect the compacted summary and raw message suffix available to the agent, or `/diff` to review the latest turn's file changes. Type `@` to mention an installed app, configured MCP server, or saved session. While nanobot is working, `Enter` steers the current turn, `Tab` queues a visible follow-up for the next turn, and `Option+Up` on macOS (`Alt+Up` on Windows/Linux) returns the latest queued message for editing. Press `Shift+Enter` to add a newline; `Ctrl+J` is the universal fallback for terminals that cannot distinguish modified Enter keys. Use `PageUp` at the top to load earlier transcript pages. The next launch returns to your last session unless `--session` selects another WebSocket session; use `--classic` to resume a session from another channel. The existing nanobot `/new` command keeps its original behavior: it resets the current chat. `nanobot agent` and `nanobot webui` share one on-demand local gateway: either command can start it, each launcher releases only its own client, and the last interactive launcher to exit stops it. Use `nanobot gateway --background` when the gateway must stay alive with no local clients. Type `exit` or press `Ctrl+C` when you are done. Use `nanobot agent --classic` only when you need the compatibility Python prompt.
 
 For one request and an immediate exit, use:
 
@@ -241,7 +248,7 @@ Prefer your own infrastructure? Follow the [deployment guide](./docs/deployment.
 
 ## 🌐 WebUI
 
-The WebUI ships **inside the published wheel** with no separate frontend build. It is the browser workbench for persistent topics, visible agent activity, workspace controls, Apps, Skills, Automations, and settings.
+The WebUI ships **inside the published wheel** with no separate frontend build. It is the browser workbench for persistent topics, temporary chats, visible agent activity, workspace controls, Apps, Skills, Automations, and settings.
 
 <p align="center">
   <img src="images/nanobot_webui.png" alt="nanobot webui preview" width="900">
@@ -250,9 +257,10 @@ The WebUI ships **inside the published wheel** with no separate frontend build.
 Use it to:
 
 - keep separate topics for different tasks and projects;
+- use temporary chats when a conversation should not be saved to history or memory;
 - inspect reasoning, tool calls, file edits, diffs, command output, and generated artifacts;
 - switch models and workspaces without leaving the conversation;
-- configure providers, chat channels, Apps, Skills, and Automations from one place.
+- configure providers and chat channels, connect Apps, discover Skills, and manage Automations from one place.
 
 See the [WebUI guide](./docs/webui.md) for LAN access, background operation, workspace controls, and the full feature tour. Working on the frontend itself? Use [`webui/README.md`](./webui/README.md).
 

THIRD_PARTY_NOTICES.md

@@ -1,7 +1,11 @@
 # Third-Party Notices
 
 The following third-party components are redistributed as part of the packaged
-nanobot Python distribution (`pip install nanobot-ai`).
+nanobot Python distribution (`pip install nanobot-ai`). Native TUI executables are distributed
+separately in per-platform release archives. Each TUI archive carries its generated
+`THIRD_PARTY_NOTICES.txt`, project and runtime licenses, corresponding application source,
+written source offer, and relinking instructions; those target-specific notices are generated by
+`tui/scripts/release-notices.ts` and are not duplicated below.
 
 ---
 

conftest.py

@@ -6,6 +6,7 @@ import os
 import ssl
 import sys
 from collections.abc import Iterator
+from pathlib import Path
 
 import certifi
 import pytest
@@ -22,6 +23,32 @@ def _isolate_nanobot_log_activation() -> Iterator[None]:
         logger.enable("nanobot")
 
 
+@pytest.fixture(autouse=True)
+def _isolate_sessions_root(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Iterator[None]:
+    """Redirect session storage away from the real active config data directory.
+
+    Session storage lives under the active runtime data root (outside the workspace,
+    per ADR-0001), so without redirection tests would write into the real home.
+    """
+    runtime_root = tmp_path.parent / f"{tmp_path.name}-runtime-root"
+    legacy_root = tmp_path.parent / f"{tmp_path.name}-legacy-sessions-root"
+
+    def runtime_subdir(name: str) -> Path:
+        path = runtime_root / name
+        path.mkdir(parents=True, exist_ok=True)
+        return path
+
+    monkeypatch.setattr(
+        "nanobot.session.manager.get_runtime_subdir",
+        runtime_subdir,
+    )
+    monkeypatch.setattr(
+        "nanobot.session.manager.get_legacy_sessions_dir",
+        lambda: legacy_root,
+    )
+    yield
+
+
 @pytest.fixture(scope="session", autouse=True)
 def _use_windows_system_ca_for_default_http_clients() -> Iterator[None]:
     """Avoid reparsing certifi's CA bundle for every offline HTTP client.

docker-compose.yml

@@ -8,6 +8,15 @@ x-common-config: &common-config
     - ~/.nanobot:/home/nanobot/.nanobot
   cap_drop:
     - ALL
+  # Entrypoint uses these to fix bind-mount ownership and drop to the nanobot user.
+  cap_add:
+    - CHOWN
+    - SETGID
+    - SETUID
+  # Prevent the non-root process from regaining capabilities through setuid
+  # binaries or file capabilities left inside the container image.
+  security_opt:
+    - no-new-privileges:true
 
 services:
   nanobot-gateway:

docs/README.md

@@ -19,7 +19,7 @@ The recommended first-run path is:
 3. Configure a provider and model in **Settings → Models**.
 4. Send `Hello!` before configuring anything else.
 
-Most people do not need to edit JSON for the first run. The WebUI handles the initial provider, model, and local browser settings. SSH, headless, existing-config, and older-release installs retain `nanobot onboard --wizard` as a terminal fallback. After the WebUI opens, use **Settings** for models and built-in capabilities, **Settings → Channels** for chat apps, and **Apps** for CLI App or MCP integrations.
+Most people do not need to edit JSON for the first run. The WebUI handles the initial provider, model, and local browser settings. SSH, headless, existing-config, and older-release installs retain `nanobot onboard --wizard` as a terminal fallback. After the WebUI opens, use **Settings** for models and built-in capabilities, **Settings → Channels** for chat apps, and **Apps** for Agent Plugins, CLI Apps, and MCP integrations.
 
 ## Add One Capability
 
@@ -32,6 +32,7 @@ Pick the row that matches what you want to accomplish next:
 | Choose a hosted, OAuth, company, or local model | [Provider Cookbook](./provider-cookbook.md) |
 | Add model fallbacks | [Configure Model Fallback](./guides/configure-model-fallback.md) |
 | Enable web search | [Configure Web Search](./guides/configure-web-search.md) |
+| Manage Agent Plugins, CLI Apps, or MCP integrations | [WebUI Apps](./webui.md#apps) |
 | Add an MCP tool server | [Configure MCP Tools](./guides/configure-mcp-tools.md) |
 | Generate images | [Image Generation](./image-generation.md) |
 | Schedule work or create a local trigger | [Automations](./automations.md) |

docs/architecture.md

@@ -51,6 +51,13 @@ Main files:
 - feeds tool results back into the model;
 - stops when a final answer is produced or runtime limits are hit.
 
+MCP connections are application-owned infrastructure. Composition roots create
+an `MCPProvider`, share its `ToolRegistry` with `AgentLoop`, await `connect()`
+before use, and guarantee `aclose()` during shutdown; the loop does not manage
+that lifecycle. `AgentLoop.from_config()` therefore requires a caller-owned
+`ToolRegistry`; callers using MCP share it with their application-owned
+`MCPProvider`.
+
 Keep this split in mind when debugging. If a problem is about channel routing, session keys, workspace selection, or outbound delivery, start in `agent/loop.py`. If it is about provider calls, tool calls, streaming, or iteration limits, start in `agent/runner.py`.
 
 ## Providers
@@ -142,7 +149,7 @@ Defaults:
 |---|---|
 | Config | `~/.nanobot/config.json` |
 | Workspace | `~/.nanobot/workspace/` |
-| Sessions | `<workspace>/sessions/*.jsonl` |
+| Sessions | `<config-dir>/sessions/<workspace-id>/*.jsonl` (default: `~/.nanobot/sessions/...`) |
 | Memory | `<workspace>/memory/` |
 | Cron store | `<workspace>/cron/jobs.json` |
 | WebUI/media/log runtime data | config directory subdirectories such as `webui/`, `media/`, and `logs/` |
@@ -157,7 +164,7 @@ a WebUI chat may select a separate project:
 
 | Concern | Path owner |
 |---|---|
-| Sessions, `SOUL.md`, `USER.md`, memory, and custom skills | Configured agent workspace |
+| Session namespace, `SOUL.md`, `USER.md`, memory, and custom skills | Configured agent workspace |
 | Project `AGENTS.md`, relative tool paths, and shell working directory | Effective project workspace |
 | Workspace access mode and project metadata | Session workspace scope |
 
@@ -173,7 +180,7 @@ Session history is the near-term conversation replay. Memory is the longer-term
 
 | Store | File area |
 |---|---|
-| Session JSONL files | `<workspace>/sessions/` |
+| Session JSONL files | `<config-dir>/sessions/<workspace-id>/` |
 | Long-term memory | `<workspace>/memory/MEMORY.md` |
 | Consolidation source history | `<workspace>/memory/history.jsonl` |
 | Bootstrap identity files | `<workspace>/SOUL.md`, `<workspace>/USER.md`, templates under `nanobot/templates/` |
@@ -201,8 +208,10 @@ When changing tools, channels, file access, WebUI workspace behavior, or network
 | Provider | Add `ProviderSpec` in `providers/registry.py`, add schema field in `config/schema.py`, implement provider only if the generic backend is not enough |
 | Channel | Export a `ChannelPlugin` descriptor, keep its runtime and optional setup surfaces in one package, and follow [`channel-package-guide.md`](./channel-package-guide.md) |
 | Tool | Implement a tool under `agent/tools/` or expose a plugin entry point |
-| MCP | Add `tools.mcpServers` config |
-| Skill | Add workspace skill files under `<workspace>/skills/` or built-in skills under `nanobot/skills/` |
+| Agent Plugin | Add a v1 package under `<workspace>/plugins/` and enable it from Apps |
+| MCP | Add `tools.mcpServers` config or bundle the server in an Agent Plugin |
+| Skill | Add workspace skills under `<workspace>/skills/`, bundle them in an Agent Plugin, or add built-in skills under `nanobot/skills/` |
+| CLI App | Add it to the CLI Apps catalog; the installer owns its executable lifecycle and writes a skills-only Agent Plugin |
 
 Prefer existing registry/discovery patterns over ad hoc wiring.
 

docs/chat-commands.md

@@ -47,7 +47,7 @@ Use `/model` to inspect the current runtime model:
 /model
 ```
 
-The response shows the current session's model and preset, plus the available preset names. Named presets come from the top-level `modelPresets` config and are the recommended way to configure model choices. `default` is always available and represents the model settings from direct `agents.defaults.*` fields.
+The response shows the current session's model and preset, plus the available preset names. Each key under the top-level `modelPresets` config is the preset's canonical name everywhere nanobot displays or references it. `default` is always available and represents the model settings from direct `agents.defaults.*` fields.
 
 To switch presets for future turns:
 

docs/cli-reference.md

@@ -88,13 +88,48 @@ follow the printed WebUI **Settings → Models** or `nanobot onboard --wizard` r
 |---|---|
 | `nanobot agent -m "Hello!"` | Send one message and exit |
 | `nanobot agent` | Start interactive terminal chat |
-| `nanobot agent --session <id>` | Use a specific session key |
+| `nanobot agent --session <id>` | Use a WebSocket session key; add `--classic` for another channel |
 | `nanobot agent --workspace <path>` | Override workspace |
 | `nanobot agent --config <path>` | Use a specific config file |
-| `nanobot agent --no-markdown` | Print plain text instead of Rich-rendered Markdown |
-| `nanobot agent --logs` | Show runtime logs while chatting |
+| `nanobot agent --classic` | Use the compatibility Python prompt instead of the native terminal UI |
+| `nanobot agent --theme auto\|dark\|light` | Auto-detect the terminal appearance or force a TUI palette |
+| `nanobot agent --no-markdown` | Use the classic prompt and print plain text instead of Markdown |
+| `nanobot agent --logs` | Use the classic prompt and show runtime logs while chatting |
 
-In interactive mode, `Enter` sends the current message. Press `Alt+Enter` to add a newline before sending.
+Inside the native TUI, `/sessions` switches saved conversations, `/new-chat` starts another saved
+conversation, and `/context` explains the compacted summary and raw session suffix available to
+the next agent turn. `/branch` forks a saved conversation from a completed reply, and `/diff`
+opens the latest turn's file changes as a full-screen unified diff.
+`PageUp` loads older transcript pages when you reach the top. The default
+launch returns to the last attached TUI session; `--session` selects a specific session instead.
+
+## Session Storage and Rollback
+
+Session JSONL files live under `<config-dir>/sessions/<workspace-id>/`, outside the
+agent-readable workspace. On the first upgraded start, nanobot safely migrates existing
+`<workspace>/sessions/*.jsonl` files after verifying an atomic copy. Stop every old nanobot
+process that uses the workspace before upgrading; old and new binaries must not write the
+same session concurrently.
+
+To prepare a downgrade, stop nanobot and copy the current sessions back to the path understood
+by older releases:
+
+```bash
+nanobot sessions restore-workspace --config ./bot-a/config.json --workspace ./bot-a/workspace
+```
+
+The command never deletes the external store and refuses to overwrite a different existing
+workspace file. Back up both the config directory and workspace before changing versions.
+
+Interactive mode uses nanobot's native TypeScript terminal UI. It talks to the same local gateway as the WebUI, so streaming, tool progress, and WebSocket sessions share one protocol instead of maintaining a second agent loop. If no gateway is running, either client starts it on demand. Exiting one TUI or WebUI launcher releases only that client; the last interactive launcher stops the on-demand gateway. A small gateway watchdog also reclaims an on-demand process if its last client crashes. Only an explicit `nanobot gateway --background` promotes it to persistent mode. `nanobot gateway restart` restarts a detached gateway without changing that lifetime; restart an attached foreground gateway in its owning terminal. `nanobot gateway stop` ends either mode.
+
+The default `--theme auto` mode probes the terminal's real foreground and background colors before first paint and follows supported live appearance changes. Use `--theme light` or `--theme dark` when a terminal or multiplexer does not report its colors reliably. The model preset and workspace access labels above the composer can be clicked to open their selectors; arrow keys, `Enter`, and `Esc` provide the same controls without a mouse. Access changes still pass through the gateway's local-trust and active-turn policy checks.
+
+`Enter` sends the current message. While a turn is active, `Enter` steers it immediately, `Tab` queues a visible follow-up for the next turn, and `Option+Up` on macOS (`Alt+Up` on Windows/Linux) returns the latest queued message to the composer. Press `Shift+Enter` to add a newline; `Ctrl+J` is the universal fallback when a terminal cannot distinguish modified Enter keys. `Alt+Enter` and `Ctrl+Enter` are also accepted when distinguishable. Use `Up`/`Down` at the composer edge to recall prompts from the current saved session. Large pastes appear as a compact placeholder in the composer but are sent unchanged. Type `/` to discover nanobot commands and terminal navigation in one palette, or type `@` to complete installed apps, configured MCP servers, and saved sessions. Use the arrow keys to choose an item and `Tab` to complete it. `/sessions` opens a searchable conversation picker, `/new-chat` preserves the current conversation and starts another one, and `/branch` forks from a completed reply. `/diff` opens a read-only unified diff for the newest turn; use `Left`/`Right` to switch edits and `Esc` to close it. The core `/new` command retains its cross-channel behavior and resets the current chat. `Ctrl+C` copies a selection, stops a running turn, clears a non-empty composer, or exits when idle. Use `PageUp`/`PageDown` to scroll, `Ctrl+Home`/`Ctrl+End` to jump to the transcript edges, and `Ctrl+O` to expand or collapse long tool traces. When you leave the bottom, the TUI shows a scrollbar and a `Ctrl+End` hint until you return. The footer reports provider token/cache usage when available. Selections copy through OSC 52 when the terminal supports it. The transcript reflows when the terminal is resized, and exiting restores the previous screen.
+
+Packaged releases fetch a version-matched, checksummed terminal archive for macOS (Apple Silicon and Intel), Linux (x64 and ARM64), or Windows x64 on first use. The cache keeps the executable together with its licenses, third-party notices, source offer, relinking instructions, and corresponding TUI source. Windows ARM64 must currently use `--classic` because the Bun runtime disables the FFI required by OpenTUI on that platform. Set `NANOBOT_TUI_NO_DOWNLOAD=1` or pass `--classic` to keep the Python-only path. A local source install requires Bun and runs its own `tui/` source while the original checkout remains available; it never silently falls back to a release binary.
+
+Non-interactive input/output, `--logs`, and `--no-markdown` automatically retain the classic prompt so existing scripts and diagnostic workflows do not acquire terminal control sequences or silently ignore their options.
 
 Interactive mode exits with `exit`, `quit`, `/exit`, `/quit`, `:q`, or `Ctrl+D`.
 
@@ -103,7 +138,7 @@ Interactive mode exits with `exit`, `quit`, `/exit`, `/quit`, `:q`, or `Ctrl+D`.
 | Command | Description |
 |---|---|
 | `nanobot webui` | Create config/workspace if needed, enable the local WebUI channel after confirmation, start the gateway, and open `http://127.0.0.1:8765` |
-| `nanobot webui --background` | Start or reuse a background gateway, then open the WebUI |
+| `nanobot webui --background` | Deprecated; prints the equivalent explicit `nanobot gateway --background` command and exits |
 | `nanobot webui --dev` | Start the gateway and Vite together at `http://127.0.0.1:5173`, with live frontend updates |
 | `nanobot webui --no-open` | Prepare and start the WebUI without opening a browser |
 | `nanobot webui --port <port>` | Set the WebUI/WebSocket port |
@@ -112,9 +147,12 @@ Interactive mode exits with `exit`, `quit`, `/exit`, `/quit`, `:q`, or `Ctrl+D`.
 
 First-run WebUI setup binds to `127.0.0.1` by default. Use manual configuration and a WebUI password before exposing the WebSocket channel beyond localhost.
 
-`--dev` is a foreground source-checkout workflow and cannot be combined with `--background`.
+`--dev` is a foreground source-checkout workflow. Persistent gateway lifecycle is deliberately
+owned only by `nanobot gateway --background`; `nanobot webui --background` prints migration
+guidance instead of silently changing process ownership.
 It installs frontend dependencies when `webui/node_modules` is missing, proxies to the configured
-WebSocket channel port, and stops Vite together with the foreground gateway.
+WebSocket channel port, and stops Vite when the launcher exits. The shared on-demand gateway stops
+only when no other interactive client still holds it.
 
 ## Gateway
 
@@ -128,7 +166,7 @@ WebSocket channel port, and stops Vite together with the foreground gateway.
 | `nanobot gateway --workspace <path>` | Override workspace |
 | `nanobot gateway --config <path>` | Use a specific config file |
 | `nanobot gateway --background` | Start the gateway as a background process |
-| `nanobot gateway status` | Show the recorded background gateway PID, state file, and log file |
+| `nanobot gateway status` | Show PID, foreground/background launch mode, explicit/on-demand lifetime, live client count, state, and logs |
 | `nanobot gateway logs --no-follow` | Print recent background gateway logs and exit |
 | `nanobot gateway logs` | Follow background gateway logs |
 | `nanobot gateway restart` | Restart the recorded background gateway with the current config |

docs/concepts.md

@@ -26,7 +26,8 @@ The default instance lives under `~/.nanobot/`:
 | Path | Meaning |
 |---|---|
 | `~/.nanobot/config.json` | Instance configuration: providers, model defaults, channels, tools, gateway, API, and runtime options |
-| `~/.nanobot/workspace/` | Agent workspace: memory, sessions, heartbeat tasks, cron jobs, skills, and generated artifacts |
+| `~/.nanobot/workspace/` | Agent workspace: memory, heartbeat tasks, cron jobs, skills, and generated artifacts |
+| `~/.nanobot/sessions/<workspace-id>/` | Session history stored outside the agent-accessible workspace; the opaque ID follows workspace moves |
 
 You can override both with command flags:
 
@@ -125,13 +126,39 @@ nanobot uses two related stores:
 
 | Store | Location | Purpose |
 |---|---|---|
-| Sessions | `<workspace>/sessions/*.jsonl` | Recent conversation turns replayed into context |
+| Sessions | `<config-dir>/sessions/<workspace-id>/*.jsonl` | Recent conversation turns replayed into context |
 | Memory | `<workspace>/memory/MEMORY.md` and `<workspace>/memory/history.jsonl` | Long-term facts and consolidated history |
 
 Dream is a periodic consolidation job. It reads accumulated history and updates workspace memory so useful context can survive beyond short session replay.
 
+The configured workspace contains a `.nanobot/workspace-id` file. It contains only an
+opaque random identifier—never conversation content or credentials. Keep it with workspace
+backups: it lets nanobot find the same external session namespace after the workspace is
+renamed, moved, or restored. A live copy opened alongside the original receives a new ID so
+the two workspaces do not share conversations accidentally.
+
 See [`memory.md`](./memory.md) for the detailed design.
 
+## Apps and Agent Plugins
+
+Agent Plugins are nanobot's common package and activation boundary for
+installable capabilities. They organize existing extension types instead of
+replacing them:
+
+| Part | Role |
+|---|---|
+| Agent Plugin | Installable package that can bundle skills, MCP servers, or both |
+| Skill | Workflow guidance loaded progressively or invoked with `$skill-name` |
+| MCP server | Runtime tools exposed to the agent |
+| CLI App | Locally managed executable whose adapter is packaged and activated like a plugin |
+| Apps | WebUI surface for reviewing and managing these capabilities |
+
+Native providers, channels, built-in tools, standalone workspace skills, and
+directly configured MCP servers keep their existing extension paths. See
+[`webui.md#apps`](./webui.md#apps) for the user-facing flow and
+[`configuration.md#agent-plugins-v1`](./configuration.md#agent-plugins-v1) for
+the package contract.
+
 ## Tools and Safety
 
 Tools are discovered automatically from built-in modules and plugin entry points. Common tool groups include:

docs/configuration.md

@@ -330,7 +330,11 @@ By default, OpenAI uses `apiType: "auto"`: nanobot calls Chat Completions normal
 
 Valid `apiType` values are exactly `auto`, `chat_completions`, and `responses`.
 
-`extraBody` follows the selected OpenAI API surface. With Chat Completions, nanobot passes it through as the SDK `extra_body` value. With Responses, configure it in Responses API body shape; nanobot merges ordinary top-level fields into the Responses request body, appends `extraBody.tools` after generated function tools, and merges `extraBody.include` without duplicates:
+`extraBody` follows the selected OpenAI API surface. With Chat Completions, nanobot passes
+ordinary fields through as the SDK `extra_body` value; list-valued `extraBody.tools` is handled
+specially and appended after generated function tools. With Responses, configure it in Responses
+API body shape; nanobot merges ordinary top-level fields into the Responses request body, appends
+`extraBody.tools` after generated function tools, and merges `extraBody.include` without duplicates:
 
 ```json
 {
@@ -356,7 +360,7 @@ request, while other tools such as `web_fetch` remain available.
 <details>
 <summary><b>DeepSeek native web search</b></summary>
 
-DeepSeek V4 Flash uses DeepSeek's native Responses API. Its provider-hosted web search is
+DeepSeek V4 Flash and Pro use DeepSeek's native Responses API. Their provider-hosted web search is
 enabled by default because it does not require a separate paid add-on. Turn it off from the
 WebUI provider settings, or with:
 
@@ -373,9 +377,9 @@ WebUI provider settings, or with:
 }
 ```
 
-The switch applies to `deepseek-v4-flash`; DeepSeek models that remain on Chat Completions
-cannot use this Responses tool. Native search calls appear in the WebUI activity stream, and
-their opaque output items are preserved for multi-turn Responses state replay.
+The switch applies to `deepseek-v4-flash` and `deepseek-v4-pro`; DeepSeek models that remain on
+Chat Completions cannot use this Responses tool. Native search calls appear in the WebUI activity
+stream, and their opaque output items are preserved for multi-turn Responses state replay.
 
 </details>
 
@@ -387,7 +391,7 @@ Providers that use the Responses API can keep reasoning context across a
 conversation, which helps with multi-step tasks. Supported providers can also
 compact long conversations automatically.
 
-nanobot preserves Responses conversation state automatically for OpenAI Responses, OpenAI Codex, Azure OpenAI, DeepSeek V4 Flash, and compatible GitHub Copilot models.
+nanobot preserves Responses conversation state automatically for OpenAI Responses, OpenAI Codex, Azure OpenAI, DeepSeek V4, and compatible GitHub Copilot models.
 Native compaction is also automatic when the provider supports it. The
 threshold is derived from the active model's context window and reserved output
 headroom; no provider configuration is required.
@@ -1400,21 +1404,6 @@ Existing configs do not need to change. Direct `agents.defaults.model`, `provide
 {
   "modelPresets": {
     "fast": {
-      "provider": "openrouter",
-      "model": "anthropic/claude-sonnet-4.5",
-      "maxTokens": 4096,
-      "contextWindowTokens": 65536
-    }
-  },
-  "agents": {
-    "defaults": {
-      "modelPreset": "fast",
-      "fallbackModels": ["deep", "localSmall"]
-    }
-  },
-  "modelPresets": {
-    "fast": {
-      "label": "Fast",
       "model": "gpt-4.1-mini",
       "provider": "openai",
       "maxTokens": 4096,
@@ -1423,7 +1412,6 @@ Existing configs do not need to change. Direct `agents.defaults.model`, `provide
       "reasoningEffort": "low"
     },
     "deep": {
-      "label": "Deep",
       "model": "claude-opus-4-5",
       "provider": "anthropic",
       "maxTokens": 8192,
@@ -1431,22 +1419,28 @@ Existing configs do not need to change. Direct `agents.defaults.model`, `provide
       "reasoningEffort": "high"
     },
     "localSmall": {
-      "label": "Local Small",
       "model": "llama3.2",
       "provider": "ollama",
       "maxTokens": 4096,
       "contextWindowTokens": 32768,
       "temperature": 0.2
     }
+  },
+  "agents": {
+    "defaults": {
+      "modelPreset": "fast",
+      "fallbackModels": ["deep", "localSmall"]
+    }
   }
 }
 ```
 
-`modelPresets` is a top-level object. The keys under it (`fast`, `deep`, `coding`, etc.) are user-defined preset names. Each preset supports:
+`modelPresets` is a top-level object. Each key (`fast`, `deep`, `coding`, etc.) is the preset's one canonical name: it is shown in the interface, passed to `/model <name>`, and referenced by defaults, fallbacks, sessions, and Dream. New and renamed presets must be unique ignoring case. Existing keys accepted by earlier releases remain loadable so upgrades do not break startup. Each preset supports:
+
+Older configs may still contain a `label` inside a preset. It is accepted when loading for compatibility but ignored; the object key remains the canonical name.
 
 | Field | Description |
 |-------|-------------|
-| `label` | Optional display name shown in model lists. |
 | `model` | Model name to use for this preset. |
 | `provider` | Provider name, or `"auto"` to use provider auto-detection. |
 | `maxTokens` | Maximum completion/output tokens. |
@@ -1917,6 +1911,14 @@ Create a key at [serper.dev](https://serper.dev). You can also set `SERPER_API_K
 
 nanobot by default uses [Jina Reader](https://jina.ai/reader/), a third-party API, to convert arbitrary pages into Markdown format for easy digestion by the LLM, with a local fallback based on [readability-lxml](https://github.com/buriy/python-readability) if the former fails.
 
+> [!NOTE]
+> Using the remote reader means the fetched URL itself is disclosed to the
+> third-party service. URLs that visibly carry credentials (userinfo, signed-URL
+> or token-style query parameters) are detected and fetched locally instead, but
+> secrets embedded in a URL's *path* (for example bot-token or webhook-style
+> URLs) cannot be reliably detected. Set `useJinaReader: false` if fetched URLs
+> must never leave the machine.
+
 If you want to always use the local conversion, you can force it using:
 
 ```json
@@ -1971,15 +1973,52 @@ Add MCP servers to your `config.json`:
 }
 ```
 
-Two transport modes are supported:
+MCP servers can run locally over stdio or connect remotely over HTTP:
 
-| Mode | Config | Example |
+| Connection | Config | Example |
 |------|--------|---------|
 | **Stdio** | `command` + `args` | Local process via `npx` / `uvx` |
-| **HTTP** | `url` + `headers` (optional) | Remote endpoint (`https://mcp.example.com/sse`) |
+| **Streamable HTTP / SSE** | `url` + `headers` (optional) | Remote endpoint (`https://mcp.example.com/mcp`) |
+
+Remote HTTP servers may use browser OAuth instead of static headers. In the
+WebUI, open **Apps → MCP → Add MCP server**, choose **Custom**, select HTTP or
+SSE, and choose **OAuth** under **Authentication**. Save the server, then choose
+**Connect**. For manual configuration, add `auth: "oauth"` and open
+**Apps → MCP** to connect. Known presets such as Xmind, Notion, and Linear add
+the config automatically on first click.
+
+```json
+{
+  "tools": {
+    "mcpServers": {
+      "notion": {
+        "type": "streamableHttp",
+        "url": "https://mcp.notion.com/mcp",
+        "auth": "oauth"
+      }
+    }
+  }
+}
+```
+
+nanobot opens the server's authorization page and handles the callback through
+the gateway. The tools become available immediately when hot reload succeeds;
+otherwise the WebUI asks for a restart. OAuth tokens and dynamic client
+registration data are stored in the nanobot data directory under
+`auth/mcp.json`; they are not written to `config.json`. Removing the MCP server
+from Apps also removes its saved OAuth credentials. Normal gateway startup never
+opens a browser or registers a new OAuth client when credentials are
+missing—interactive authorization starts only after a user clicks **Connect**.
+
+For a remotely accessed WebUI, HTTPS is recommended. Configure
+`channels.websocket.publicWsUrl` with the browser-facing `wss://` endpoint so
+nanobot can register the matching HTTPS callback and finish automatically. A
+loopback WebUI may use HTTP. When a remote WebUI is served over plain HTTP,
+nanobot instead registers a localhost callback and asks you to paste the complete
+callback URL from the browser address bar after authorization.
 
 > [!IMPORTANT]
-> HTTP/SSE MCP URLs are validated before probing or connecting, and every outgoing MCP HTTP request is validated again before redirects are followed. `localhost`, `127.0.0.1`, RFC1918/private IPs, CGNAT/Tailscale ranges, link-local addresses, and cloud metadata endpoints are blocked by default. This can break previously working local or private HTTP MCP configs until the endpoint is explicitly allowed with `tools.ssrfWhitelist`, preferably with a single-host CIDR such as `127.0.0.1/32`, `::1/128`, or `192.168.1.50/32`. Stdio MCP servers are not affected.
+> HTTP/SSE MCP URLs are validated before probing or connecting, and every outgoing MCP HTTP request—including OAuth metadata, client registration, token exchange, and redirects—is validated again. `localhost`, `127.0.0.1`, RFC1918/private IPs, CGNAT/Tailscale ranges, link-local addresses, and cloud metadata endpoints are blocked by default. This can break previously working local or private HTTP MCP configs until the endpoint is explicitly allowed with `tools.ssrfWhitelist`, preferably with a single-host CIDR such as `127.0.0.1/32`, `::1/128`, or `192.168.1.50/32`. Stdio MCP servers are not affected.
 
 Use `toolTimeout` to override the default 30s per-call timeout for slow servers:
 
@@ -2054,7 +2093,7 @@ For API keys, tokens, and other secrets, see [Environment Variables for Secrets]
 | `tools.ssrfWhitelist` | `[]` | CIDR ranges exempted from the shared SSRF guard used by web fetches and HTTP/SSE MCP connections. Prefer exact host CIDRs such as `192.168.1.50/32`; broad ranges increase SSRF exposure. |
 | `channels.*.allowFrom` | omitted | Access control per channel. Omit to use pairing-only mode; set `["*"]` to allow everyone; or list specific user IDs. See [Pairing](#pairing) for details. |
 
-**Docker security**: The official Docker image runs as a non-root user (`nanobot`, UID 1000) with bubblewrap pre-installed. The default `docker-compose.yml` drops all Linux capabilities and keeps Docker's default AppArmor/seccomp profiles enabled. If you enable `"tools.exec.sandbox": "bwrap"` inside Docker, start Compose with `docker-compose.bwrap.yml` as an additional override so bubblewrap can create nested namespaces.
+**Docker security**: The official Docker image runs as a non-root user (`nanobot`, UID 1000) with bubblewrap pre-installed. The default `docker-compose.yml` drops all Linux capabilities except the `CHOWN`, `SETGID`, and `SETUID` capabilities required by the root entrypoint to initialize bind-mount ownership and become UID 1000. It enables `no-new-privileges` so the final non-root process cannot regain those bootstrap capabilities, and keeps Docker's default AppArmor/seccomp profiles enabled. If you enable `"tools.exec.sandbox": "bwrap"` inside Docker, start Compose with `docker-compose.bwrap.yml` as an additional override so bubblewrap can create nested namespaces. The host must also allow unprivileged user namespaces; the override cannot bypass a host-level namespace restriction.
 
 
 ## Pairing
@@ -2306,6 +2345,20 @@ Disabled skills are excluded from the main agent's skill summary, from always-on
 |--------|---------|-------------|
 | `agents.defaults.disabledSkills` | `[]` | List of skill directory names to exclude from loading. Applies to both built-in skills and workspace skills. |
 
+### Agent Plugins v1
+
+nanobot discovers [Agent Plugins](https://agent-plugins.org/) under `<workspace>/plugins/`; a v1 package has `plugin.json` and may add `mcp.json`, `skills/<name>/SKILL.md`, or both. Agent Plugins are the common package and activation boundary for installable capabilities; they do not replace native providers, channels, tools, standalone workspace skills, or directly configured MCP servers.
+
+Directory presence means installed; activation is explicit in **Apps**. Skills use progressive loading and `$skill-name` invocation, with workspace > plugin > built-in precedence.
+Enabled `stdio` servers receive contained `PLUGIN_ROOT` and isolated `PLUGIN_DATA` paths; explicit
+`tools.mcpServers` entries win collisions. Invalid or escaping components are ignored.
+An enabled package is treated as immutable: changing any packaged file disables it until the user
+reviews and enables it again. Runtime state belongs under `PLUGIN_DATA`, not the package root.
+
+Enabled plugins run as the nanobot user; permissions are descriptive, not an OS sandbox. The optional `extensions.dev.nanobot.logo` accepts a contained PNG, JPEG, or WebP up to 256 KiB.
+
+CLI Apps use the same skills-only package layout while their installer manages executables, updates, and removal. Future catalogs can place packages before using this activation path.
+
 ## Tool Hint Max Length
 
 Tool hints are the short progress messages shown when the agent calls tools (e.g. `$ cd …/project && npm test`). By default, these are truncated at 40 characters, which can make long commands hard to read.

docs/deployment.md

@@ -11,7 +11,7 @@ Check these once before Render, Docker, systemd, or LaunchAgent:
 | `nanobot status` shows the expected config and workspace | Confirms the process will read the instance you meant to run |
 | `nanobot agent -m "Hello!"` works | Proves install, config, provider, model, and workspace writes before adding a service layer |
 | Secrets are in environment variables or protected config files | API keys, bot tokens, OAuth state, and chat credentials should not be world-readable |
-| `~/.nanobot/` or your custom config/workspace path is persistent | Sessions, memory, channel login state, generated artifacts, and cron jobs live there |
+| The active config directory (including `sessions/`) and workspace are persistent | Sessions follow `--config`; memory, generated artifacts, and the workspace identity marker follow the workspace |
 | Channel access control is intentional | Use `allowFrom`, pairing, WebSocket `token`/`tokenIssueSecret`, or private test channels before exposing the bot |
 | Ports are planned | Gateway health defaults to local-only `127.0.0.1:18790`; WebUI/WebSocket defaults to `8765`; `nanobot serve` defaults to `8900` |
 | Logs are easy to reach | Use `docker compose logs`, `journalctl`, LaunchAgent log files, or `nanobot gateway --verbose` while diagnosing startup |
@@ -160,8 +160,11 @@ docker compose logs -f nanobot-gateway                   # view logs
 docker compose down                                      # stop
 ```
 
-The default Compose file drops all Linux capabilities and keeps Docker's default
-AppArmor/seccomp profiles enabled. If you explicitly set
+The default Compose file drops all Linux capabilities except `CHOWN`, `SETUID`, and
+`SETGID`, which the root entrypoint needs to fix bind-mount ownership and become UID
+1000. It also enables `no-new-privileges`, so the non-root process cannot regain those
+bootstrap capabilities through setuid binaries or file capabilities. Docker's default
+AppArmor/seccomp profiles remain enabled. If you explicitly set
 `"tools.exec.sandbox": "bwrap"` in `~/.nanobot/config.json`, add the bwrap
 override file when starting containers:
 
@@ -170,8 +173,10 @@ docker compose -f docker-compose.yml -f docker-compose.bwrap.yml up -d nanobot-g
 docker compose -f docker-compose.yml -f docker-compose.bwrap.yml run --rm nanobot-cli agent -m "Hello!"
 ```
 
-The override grants `CAP_SYS_ADMIN` and disables AppArmor/seccomp confinement for
-the container so bubblewrap can create its nested namespaces. Use it only when the
+The override adds `CAP_SYS_ADMIN` and disables AppArmor/seccomp confinement for the
+container so bubblewrap can create its nested namespaces. It preserves
+`no-new-privileges`. The host must also allow unprivileged user namespaces; the
+override cannot bypass a host-level namespace restriction. Use it only when the
 bwrap sandbox is enabled.
 
 ### Docker
@@ -197,6 +202,8 @@ vim ~/.nanobot/config.json
 # health endpoint on 18790.
 docker run \
   --cap-drop ALL \
+  --cap-add CHOWN --cap-add SETGID --cap-add SETUID \
+  --security-opt no-new-privileges:true \
   -v ~/.nanobot:/home/nanobot/.nanobot \
   -p 18790:18790 -p 8765:8765 \
   nanobot gateway
@@ -205,7 +212,9 @@ docker run \
 # bubblewrap needs for nested namespaces. Without them, `bwrap` may exit with
 # `clone3: Operation not permitted`.
 docker run \
-  --cap-drop ALL --cap-add SYS_ADMIN \
+  --cap-drop ALL \
+  --cap-add CHOWN --cap-add SETGID --cap-add SETUID --cap-add SYS_ADMIN \
+  --security-opt no-new-privileges:true \
   --security-opt apparmor=unconfined \
   --security-opt seccomp=unconfined \
   -v ~/.nanobot:/home/nanobot/.nanobot \

docs/guides/ai-agent-webui.md

@@ -45,7 +45,8 @@ outside the workspace when the gateway allows it.
 
 ## Production notes
 
-- Use `nanobot webui --background` when you do not want to keep a terminal open.
+- Use `nanobot gateway --background` when you do not want to keep a terminal open, then open the
+  configured WebUI URL in a browser.
 - Use `nanobot gateway status`, `logs`, `restart`, and `stop` to manage a
   background gateway.
 - If you expose the WebUI beyond localhost, set a token issue secret and review

docs/guides/configure-mcp-tools.md

@@ -30,10 +30,15 @@ remote HTTP endpoint.
 For local interactive setup:
 
 1. Run `nanobot webui` and open **Apps**.
-2. Choose a known integration preset, or add a custom stdio, HTTP, or SSE server.
+2. Choose a known MCP server preset, or add a custom stdio, HTTP, or SSE server.
+   For a custom OAuth server, choose **OAuth** under **Authentication**, save it,
+   and click **Connect**. Presets such as Xmind, Notion, and Linear go straight to
+   **Connect**. Approve access in the browser window. HTTPS and localhost WebUIs
+   return automatically. From a remote plain-HTTP WebUI, copy the complete
+   localhost callback URL from the browser address bar and paste it into nanobot.
 3. Limit the enabled tools when the server exposes more than the task needs.
 4. Save and restart when prompted.
-5. Mention the integration with `@` in the next message and ask for a small test action.
+5. Mention the connected MCP server with `@` in the next message and ask for a small test action.
 
 For manual or deployment-managed config, add this to `~/.nanobot/config.json`:
 
@@ -58,12 +63,16 @@ Restart nanobot and ask a question that requires the MCP tool.
 - Prefer `enabledTools` over exposing every tool by default.
 - Use `toolTimeout` for slow MCP operations.
 - Use HTTP MCP only for endpoints you trust.
+- For deployment-managed OAuth servers, set `auth` to `oauth` and complete the
+  browser connection from **Apps → MCP**.
 - Keep MCP server commands stable and versioned in deployment docs or scripts.
 
 ## Security notes
 
 - Stdio MCP starts a local process; review the command before enabling it.
-- HTTP/SSE MCP uses nanobot's SSRF guard.
+- HTTP/SSE MCP uses nanobot's SSRF guard, including OAuth discovery, registration,
+  token exchange, and redirects.
+- OAuth credentials live in the nanobot data directory, not in `config.json`.
 - Allow private HTTP MCP hosts only with narrow `tools.ssrfWhitelist` CIDRs.
 - Do not place secrets in command arguments when environment variables or
   headers can be used.

docs/guides/configure-model-fallback.md

@@ -32,16 +32,14 @@ with ones you control:
 ```json
 {
   "modelPresets": {
-    "fast": {
-      "label": "Fast",
+    "Fast": {
       "provider": "primary-provider",
       "model": "primary-model-id",
       "maxTokens": 4096,
       "contextWindowTokens": 65536,
       "temperature": 0.1
     },
-    "deep": {
-      "label": "Deep",
+    "Deep": {
       "provider": "fallback-provider",
       "model": "fallback-model-id",
       "maxTokens": 4096,
@@ -51,8 +49,8 @@ with ones you control:
   },
   "agents": {
     "defaults": {
-      "modelPreset": "fast",
-      "fallbackModels": ["deep"]
+      "modelPreset": "Fast",
+      "fallbackModels": ["Deep"]
     }
   }
 }
@@ -69,7 +67,7 @@ for common providers.
   how much context can fit.
 - Put cheaper or faster fallbacks before expensive ones when acceptable.
 - Use `/model <preset>` for runtime switching without editing config.
-- Keep labels human-readable for WebUI model lists.
+- Keep preset names human-readable; the same name appears in the WebUI and `/model`.
 
 ## Security notes
 

docs/guides/configure-ollama-prompt-cache.md

@@ -179,8 +179,7 @@ Merge this preset into `~/.nanobot/config.json` and select it:
     }
   },
   "modelPresets": {
-    "ollamaPrefixStable": {
-      "label": "Ollama Llama 3.1 prefix-stable",
+    "Ollama Llama 3.1 prefix-stable": {
       "provider": "ollama",
       "model": "llama3.1:8b-prefix-stable-v1",
       "maxTokens": 2048,
@@ -190,7 +189,7 @@ Merge this preset into `~/.nanobot/config.json` and select it:
   },
   "agents": {
     "defaults": {
-      "modelPreset": "ollamaPrefixStable"
+      "modelPreset": "Ollama Llama 3.1 prefix-stable"
     }
   }
 }

docs/guides/configure-openai-compatible-provider.md

@@ -41,8 +41,7 @@ Merge this into `~/.nanobot/config.json`:
     }
   },
   "modelPresets": {
-    "primary": {
-      "label": "Custom",
+    "Custom": {
       "provider": "custom",
       "model": "provider-model-name",
       "maxTokens": 4096,
@@ -52,7 +51,7 @@ Merge this into `~/.nanobot/config.json`:
   },
   "agents": {
     "defaults": {
-      "modelPreset": "primary"
+      "modelPreset": "Custom"
     }
   }
 }

docs/guides/configure-web-search.md

@@ -81,6 +81,10 @@ in the WebUI or logs.
 - Web fetch and HTTP MCP share an SSRF guard.
 - Private, loopback, link-local, and cloud metadata addresses are blocked by
   default.
+- With `useJinaReader` enabled (the default), fetched URLs are disclosed to the
+  remote reader service. Credential-bearing URLs (userinfo or token/signature
+  query parameters) are fetched locally instead; path-embedded secrets cannot
+  be detected, so disable the remote reader when URLs must stay local.
 - Add `tools.ssrfWhitelist` only for narrow trusted CIDRs.
 - Do not give public chat users unrestricted web and shell access without
   review.

docs/guides/deploy-nanobot-gateway.md

@@ -37,18 +37,20 @@ nanobot gateway
 For WebUI background usage:
 
 ```bash
-nanobot webui --background
+nanobot gateway --background
 nanobot gateway status
 nanobot gateway logs
 ```
 
+Open the configured WebUI URL in a browser, or run `nanobot webui` as a foreground client.
+
 ## Production notes
 
 - Docker Compose is the most repeatable Linux container path.
 - systemd user services are useful for Linux user-level gateway deployments.
 - macOS LaunchAgent keeps the gateway alive after login.
-- Persist config, workspace, sessions, memory files, channel login state, and
-  generated artifacts.
+- Persist the active config directory's `sessions/` folder together with the workspace
+  (including `.nanobot/workspace-id`), memory files, channel login state, and generated artifacts.
 - Restart the gateway after editing `config.json`.
 
 ## Security notes

docs/multiple-instances.md

@@ -52,12 +52,13 @@ nanobot webui -c ~/.nanobot-telegram/config.json
 nanobot agent -c ~/.nanobot-telegram/config.json -w /tmp/nanobot-telegram-test
 ```
 
-> `nanobot agent` starts a local CLI agent using the selected workspace/config. It does not attach to or proxy through an already running `nanobot gateway` process.
+> Interactive `nanobot agent` and `nanobot webui` commands with the same `--config` and explicit `--workspace` selectors share one gateway instance. Different selectors produce isolated runtime state and processes. The one-shot and `--classic` agent paths remain direct local executions.
 
 | Component | Resolved From | Example |
 |-----------|---------------|---------|
 | **Config** | `--config` path | `~/.nanobot-A/config.json` |
 | **Workspace** | `--workspace` or config | `~/.nanobot-A/workspace/` |
+| **Sessions** | config directory + workspace ID | `~/.nanobot-A/sessions/<workspace-id>/` |
 | **Cron Jobs** | workspace directory | `~/.nanobot-A/workspace/cron/` |
 | **Media / runtime state** | config directory | `~/.nanobot-A/media/` |
 
@@ -126,6 +127,6 @@ nanobot gateway --config ~/.nanobot-telegram/config.json --workspace /tmp/nanobo
 ## Notes
 
 - Each instance must use a different port if they run at the same time
-- Use a different workspace per instance if you want isolated memory, sessions, and skills
+- Session data follows the active config directory; use a different workspace per instance to isolate memory, skills, and the stable session namespace ID
 - `--workspace` overrides the workspace defined in the config file
 - Cron jobs are stored in the active workspace; runtime media/state is derived from the config directory

docs/provider-cookbook.md

@@ -71,7 +71,6 @@ This recipe applies when one API key routes many hosted model families.
   },
   "modelPresets": {
     "primary": {
-      "label": "Primary",
       "provider": "openrouter",
       "model": "anthropic/claude-sonnet-4.5",
       "maxTokens": 4096,
@@ -113,7 +112,6 @@ OpenCode Zen:
   },
   "modelPresets": {
     "primary": {
-      "label": "OpenCode Zen",
       "provider": "opencode_zen",
       "model": "opencode/deepseek-v4-pro",
       "maxTokens": 4096,
@@ -140,7 +138,6 @@ OpenCode Go:
   },
   "modelPresets": {
     "primary": {
-      "label": "OpenCode Go",
       "provider": "opencode_go",
       "model": "opencode-go/deepseek-v4-flash",
       "maxTokens": 4096,
@@ -182,7 +179,6 @@ This recipe applies when you have an OpenAI API key and want to call OpenAI dire
   },
   "modelPresets": {
     "primary": {
-      "label": "OpenAI",
       "provider": "openai",
       "model": "gpt-5",
       "maxTokens": 4096,
@@ -219,7 +215,6 @@ This recipe applies when your key comes from Anthropic and your model name is an
   },
   "modelPresets": {
     "primary": {
-      "label": "Anthropic",
       "provider": "anthropic",
       "model": "claude-sonnet-4-5",
       "maxTokens": 4096,
@@ -255,7 +250,6 @@ If you use an Anthropic-compatible proxy, keep the preset provider as `anthropic
   },
   "modelPresets": {
     "primary": {
-      "label": "Anthropic proxy",
       "provider": "anthropic",
       "model": "claude-sonnet-4-5",
       "maxTokens": 4096,
@@ -286,7 +280,6 @@ This recipe applies when your key comes from Kimi's Coding Plan endpoint. Nanobo
   },
   "modelPresets": {
     "kimiCoding": {
-      "label": "Kimi Coding",
       "provider": "kimi_coding",
       "model": "kimi-for-coding",
       "maxTokens": 4096,
@@ -324,7 +317,6 @@ This recipe applies to an OpenAI-compatible service that is not a named nanobot
   },
   "modelPresets": {
     "primary": {
-      "label": "Custom",
       "provider": "custom",
       "model": "provider-model-name",
       "maxTokens": 4096,
@@ -364,7 +356,6 @@ For multiple custom endpoints, do not overload the single `custom` block. Name e
   },
   "modelPresets": {
     "work": {
-      "label": "Work proxy",
       "provider": "workProxy",
       "model": "gpt-4o-mini",
       "maxTokens": 4096,
@@ -372,7 +363,6 @@ For multiple custom endpoints, do not overload the single `custom` block. Name e
       "temperature": 0.1
     },
     "lab": {
-      "label": "Lab local",
       "provider": "lab-local",
       "model": "served-model-name",
       "maxTokens": 4096,
@@ -408,7 +398,6 @@ ollama pull llama3.2
   },
   "modelPresets": {
     "local": {
-      "label": "Local",
       "provider": "ollama",
       "model": "llama3.2",
       "maxTokens": 2048,
@@ -453,7 +442,6 @@ This recipe applies when a local server exposes an OpenAI-compatible `/v1` API.
   },
   "modelPresets": {
     "local": {
-      "label": "Local",
       "provider": "vllm",
       "model": "served-model-name",
       "maxTokens": 4096,
@@ -480,7 +468,6 @@ For LM Studio, use its local base URL and provider name:
   },
   "modelPresets": {
     "local": {
-      "label": "LM Studio",
       "provider": "lm_studio",
       "model": "local-model",
       "maxTokens": 2048,
@@ -505,7 +492,6 @@ This recipe applies when one provider sometimes rate-limits, one model is expens
 {
   "modelPresets": {
     "fast": {
-      "label": "Fast",
       "provider": "openrouter",
       "model": "anthropic/claude-sonnet-4.5",
       "maxTokens": 4096,
@@ -513,7 +499,6 @@ This recipe applies when one provider sometimes rate-limits, one model is expens
       "temperature": 0.1
     },
     "deep": {
-      "label": "Deep",
       "provider": "anthropic",
       "model": "claude-sonnet-4-5",
       "maxTokens": 4096,
@@ -521,7 +506,6 @@ This recipe applies when one provider sometimes rate-limits, one model is expens
       "temperature": 0.1
     },
     "local": {
-      "label": "Local",
       "provider": "ollama",
       "model": "llama3.2",
       "maxTokens": 2048,
@@ -580,14 +564,12 @@ Use this after you have more than one preset and are chatting through a supporte
 {
   "modelPresets": {
     "fast": {
-      "label": "Fast",
       "provider": "openrouter",
       "model": "anthropic/claude-sonnet-4.5",
       "maxTokens": 4096,
       "contextWindowTokens": 65536
     },
     "local": {
-      "label": "Local",
       "provider": "ollama",
       "model": "llama3.2",
       "maxTokens": 2048,

docs/providers.md

@@ -100,6 +100,63 @@ Gateway-style setup for model IDs served through OpenRouter.
 
 Use the model ID exactly as OpenRouter lists it.
 
+To opt into OpenRouter server-managed search and fetch, add:
+
+```json
+{
+  "providers": {
+    "openrouter": {
+      "extraBody": {
+        "tools": [
+          { "type": "openrouter:web_search" },
+          { "type": "openrouter:web_fetch" }
+        ]
+      }
+    }
+  }
+}
+```
+
+Chat Completions-compatible OpenRouter
+[server tools](https://openrouter.ai/docs/guides/features/server-tools), such as those above, are
+appended to nanobot's generated functions. This keeps unrelated local tools such as `write_file`
+available in the same request. Responses-only server tools require an API surface that the
+OpenRouter provider does not currently enable.
+
+### OrcaRouter Gateway
+
+[OrcaRouter](https://www.orcarouter.ai) is an OpenAI-compatible model routing gateway. Configure
+the built-in `orcarouter` provider and use a model ID from OrcaRouter's catalog:
+
+```json
+{
+  "providers": {
+    "orcarouter": {
+      "apiKey": "${ORCAROUTER_API_KEY}"
+    }
+  },
+  "modelPresets": {
+    "primary": {
+      "provider": "orcarouter",
+      "model": "orcarouter/auto",
+      "maxTokens": 8192,
+      "contextWindowTokens": 65536
+    }
+  },
+  "agents": {
+    "defaults": {
+      "modelPreset": "primary"
+    }
+  }
+}
+```
+
+Use the model ID exactly as OrcaRouter lists it. `orcarouter/auto` routes to a
+suitable upstream automatically; explicit IDs such as
+`anthropic/claude-sonnet-4.6` or `openai/gpt-5` are also accepted. OrcaRouter API keys start with
+`sk-orca-`. The WebUI can load the account's model catalog after the API key is saved under
+**Settings → Models**.
+
 ### Eden AI Gateway
 
 Eden AI exposes an OpenAI-compatible chat-completions endpoint at
@@ -264,7 +321,7 @@ Arbitrary custom provider names are OpenAI-compatible only; they do not use the
 
 `providers.openai.apiType` may be set when you need to force a specific OpenAI API surface. Other providers reject `apiType`; leave it unset outside `providers.openai`. Replace the model with a model ID available to your OpenAI account. Direct OpenAI Responses, OpenAI Codex, Azure OpenAI Responses, and eligible GitHub Copilot models share [opaque Responses state retention](./configuration.md#responses-state-and-compaction); native compaction is enabled only where the backend supports it. The WebUI exposes provider-native switches for OpenAI web search, Codex Fast mode, DeepSeek web search, and Grok X Search. These switches write the corresponding raw provider request fields under `extraBody`.
 
-DeepSeek is the model-level exception in the OpenAI-compatible provider: `deepseek-v4-flash` automatically uses DeepSeek's native Responses API, while `deepseek-v4-pro` remains on Chat Completions. Its native `web_search` tool is enabled by default and shows its lifecycle in WebUI chat activity; set `providers.deepseek.extraBody.tools` to `[]` to disable it.
+DeepSeek is the model-level exception in the OpenAI-compatible provider: `deepseek-v4-flash` and `deepseek-v4-pro` automatically use DeepSeek's native Responses API. Its native `web_search` tool is enabled by default and shows its lifecycle in WebUI chat activity; set `providers.deepseek.extraBody.tools` to `[]` to disable it.
 
 ### Custom OpenAI-Compatible Endpoint
 
@@ -576,7 +633,6 @@ Model presets are the recommended model configuration surface. Use them when you
 {
   "modelPresets": {
     "fast": {
-      "label": "Fast",
       "provider": "openrouter",
       "model": "anthropic/claude-sonnet-4.5",
       "maxTokens": 4096,
@@ -584,7 +640,6 @@ Model presets are the recommended model configuration surface. Use them when you
       "temperature": 0.1
     },
     "deep": {
-      "label": "Deep",
       "provider": "anthropic",
       "model": "claude-opus-4-5",
       "maxTokens": 8192,
@@ -610,7 +665,6 @@ Fallbacks are useful for transient provider failures, rate limits, or model avai
 {
   "modelPresets": {
     "fast": {
-      "label": "Fast",
       "provider": "openrouter",
       "model": "anthropic/claude-sonnet-4.5",
       "maxTokens": 4096,
@@ -618,7 +672,6 @@ Fallbacks are useful for transient provider failures, rate limits, or model avai
       "temperature": 0.1
     },
     "deep": {
-      "label": "Deep",
       "provider": "anthropic",
       "model": "claude-opus-4-5",
       "maxTokens": 8192,
@@ -626,7 +679,6 @@ Fallbacks are useful for transient provider failures, rate limits, or model avai
       "temperature": 0.1
     },
     "localSmall": {
-      "label": "Local Small",
       "provider": "ollama",
       "model": "llama3.2",
       "maxTokens": 4096,

docs/quick-start.md

@@ -12,7 +12,7 @@ These repository docs follow current `main`. The recommended installer uses the
 - Access to one supported AI provider, company endpoint, or local model server.
 - The credential, endpoint URL, and model ID required by that service. Local providers such as Ollama may not require a key.
 
-Git is only needed for a source install. The published package already contains the WebUI. A current-source install needs `bun` or `npm` so its WebUI bundle can be built.
+Git and [Bun](https://bun.sh/) are only needed for a source install. The published package already contains the WebUI and fetches a checksummed, version-matched TUI archive with its licenses, notices, corresponding application source, source offer, and relinking instructions on first use.
 
 ## 1. Install nanobot
 
@@ -30,7 +30,7 @@ curl -fsSL https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.
 irm https://raw.githubusercontent.com/HKUDS/nanobot/main/scripts/install.ps1 | iex
 ```
 
-The installer chooses an active virtual environment, `uv`, `pipx`, or a managed environment under `~/.nanobot/venv`. It installs the stable PyPI release unless you explicitly pass `--dev`. At the end it prints the exact command it used to run nanobot; if `nanobot` is not on `PATH`, reuse that full command in the examples below.
+The installer chooses an active virtual environment, `uv`, `pipx`, or a managed environment under `~/.nanobot/venv`. It installs the stable PyPI release. At the end it prints the exact command it used to run nanobot; if `nanobot` is not on `PATH`, reuse that full command in the examples below.
 
 If you prefer to inspect the scripts first, open [`install.sh`](../scripts/install.sh) or [`install.ps1`](../scripts/install.ps1).
 
@@ -48,7 +48,8 @@ The WebUI launcher creates or updates:
 | Path | Purpose |
 |---|---|
 | `~/.nanobot/config.json` | Provider, model, WebUI, channel, tool, and runtime settings |
-| `~/.nanobot/workspace/` | Sessions, memory, skills, automations, and generated files |
+| `~/.nanobot/workspace/` | Memory, skills, automations, and generated files |
+| `~/.nanobot/sessions/<workspace-id>/` | Recent session history stored outside the workspace; the ID remains stable across workspace moves |
 
 If the installer did not open the browser, run:
 
@@ -78,7 +79,7 @@ Most other providers can say `not set`. This command validates local setup but d
 
 ## 4. Get the First Reply
 
-If the installer-started WebUI is no longer running, run `nanobot webui` again. Leave that terminal open; the first-run WebUI is bound to localhost, so other devices on your network cannot reach it.
+If the installer-started WebUI is no longer running, run `nanobot webui` again. Leave that launcher open; the first-run WebUI is bound to localhost, so other devices on your network cannot reach it.
 
 Send:
 
@@ -88,7 +89,7 @@ Hello!
 
 Any normal assistant answer is success. It proves that nanobot can load the config, reach the selected model, use the workspace, and serve the browser UI.
 
-Leave the terminal open while using the WebUI. If you prefer a managed background process, stop the foreground process with `Ctrl+C`, then run:
+Interactive WebUI and TUI launchers share one on-demand gateway. Closing one launcher leaves it running for the others; closing the last launcher stops it. If you prefer a persistent background process, press `Ctrl+C`, then run:
 
 ```bash
 nanobot gateway --background
@@ -111,7 +112,11 @@ Then start an interactive terminal chat with:
 nanobot agent
 ```
 
-In interactive mode, `Enter` sends and `Alt+Enter` inserts a newline. Exit with `exit`, `/exit`, `:q`, or `Ctrl+D`.
+In interactive mode, `Enter` sends and `Shift+Enter` inserts a newline (`Ctrl+J` is the
+universal fallback). While a turn is running,
+`Enter` steers it, `Tab` queues a follow-up, and `Option+Up` on macOS (`Alt+Up` on
+Windows/Linux) edits the latest queued message. Exit
+with `exit`, `/exit`, `:q`, or `Ctrl+D`.
 
 ## Choose One Next Step
 
@@ -150,18 +155,28 @@ If pip reports `externally-managed-environment`, use the recommended installer,
 
 **Current source**
 
-`bun` or `npm` must be available. Activate a virtual environment first, then run:
+Clone the repository and install it in editable mode. Bun is required so the checkout can run
+its matching native TUI instead of mixing current Python with an older release binary.
 
 ```bash
 git clone https://github.com/HKUDS/nanobot.git
 cd nanobot
-python -m pip install .
+python -m venv .venv
+```
+
+Activate it with `source .venv/bin/activate` on macOS/Linux or
+`.venv\Scripts\Activate.ps1` in Windows PowerShell, then run:
+
+```bash
+python -m pip install -e .
 nanobot webui
 ```
 
-On Windows, if `python -m pip install .` reports that it cannot launch `npm`, run `cd webui`, `npm.cmd install --package-lock=false`, `npm.cmd run build`, and `cd ..` in order, then retry the install.
-
-The source path follows current `main` and can be newer than the published package. A non-editable install triggers the build hook that bundles the current WebUI. For editable Python or frontend development, follow [`../CONTRIBUTING.md`](../CONTRIBUTING.md) and [`../webui/README.md`](../webui/README.md).
+The source path follows current `main` and can be newer than the published package. The editable
+install keeps Python pointed at the checkout; `nanobot agent` runs `tui/` with Bun, and
+`nanobot webui` automatically rebuilds `webui/` when its bundled assets are stale. All normal
+commands remain the same as a stable install. For development details, follow
+[`../CONTRIBUTING.md`](../CONTRIBUTING.md).
 
 If the package is installed but the shell cannot find `nanobot`, use the runner that owns the installation. The recommended installer prints the exact command to reuse. Common forms are:
 
@@ -220,11 +235,15 @@ python -m pip install -U nanobot-ai
 For a source checkout:
 
 ```bash
-git pull
-python -m pip install .
+git pull --ff-only
+python -m pip install -e .
 ```
 
-Then check `nanobot --version`. Run `nanobot onboard --refresh` when you want to add newly introduced default fields while preserving existing settings.
+Because the install is editable, normal source changes are visible immediately. Re-running the
+install synchronizes any changed Python dependencies; the TUI and WebUI refresh their own
+dependencies/assets when launched. Then check `nanobot --version`. Run
+`nanobot onboard --refresh` when you want to add newly introduced default fields while preserving
+existing settings.
 
 ## If the First Reply Fails
 

docs/start-without-technical-background.md

@@ -160,4 +160,4 @@ Run:
 nanobot webui
 ```
 
-Leave that terminal open while you use nanobot. To stop it, return to the terminal and press `Ctrl+C`. Use `nanobot webui --background` only after the normal foreground start and model setup work; then manage it with `nanobot gateway status`, `logs`, `restart`, and `stop`.
+Leave that launcher open while you use nanobot. Pressing `Ctrl+C` disconnects it; the shared gateway stops when it was the last local WebUI or TUI client. After the normal foreground start and model setup work, use `nanobot gateway --background` when you want the gateway to stay online with no clients; then manage it with `nanobot gateway status`, `logs`, `restart`, and `stop`.

docs/troubleshooting.md

@@ -270,6 +270,12 @@ http://127.0.0.1:8765
 
 If accessing from another device, bind the WebSocket channel to `0.0.0.0` and set `token` or `tokenIssueSecret`. The WebSocket channel refuses public binds without a token or token issue secret.
 
+| Symptom | Check |
+|---|---|
+| A temporary chat disappeared after a reload or reconnect | This is expected. Temporary chats exist only for the current WebUI connection and are not saved to history or memory. Use a regular topic for anything you need to retain. |
+| A skills.sh install says that `npx` is required | Install Node.js with `npx` on the gateway machine, or choose a SkillHub skill that does not require `npx`. |
+| A remote browser says skill installation is disabled | Install from a same-machine WebUI. For a private deployment where every authenticated user is trusted to install third-party skill instructions or scripts, explicitly enable `tools.webuiAllowRemotePackageInstall`. |
+
 See [`webui.md#lan-access`](./webui.md#lan-access) for LAN setup and [`../webui/README.md`](../webui/README.md) for frontend development.
 
 ## Chat App Problems
@@ -313,7 +319,8 @@ See [`chat-apps.md`](./chat-apps.md) for channel-specific setup.
 |---|---|
 | Conversation context seems wrong | Confirm the active workspace and session. WebUI chats and chat app threads may use different sessions. |
 | Memory does not update immediately | Dream consolidation is periodic; recent turns still live in session history. |
-| Old sessions appear after moving config | Session files are stored under `<workspace>/sessions/`; verify the workspace path. |
+| Sessions disappear after changing `--config` | Sessions follow the config directory at `<config-dir>/sessions/<workspace-id>/`; use the original config path or copy that `sessions/` directory into the new config directory while nanobot is stopped. |
+| Sessions disappear after moving a workspace | Keep the workspace's `.nanobot/workspace-id` file with the move or backup. If it was lost, restore that marker from backup before starting nanobot. |
 | You want one shared session across devices | Set `agents.defaults.unifiedSession` intentionally; otherwise keep separate sessions. |
 
 ## Collect Useful Evidence

docs/webui.md

@@ -1,10 +1,10 @@
 # Nanobot WebUI: Browser Workbench for Self-Hosted AI Agents
 
-<!-- Meta description: Run nanobot from a browser WebUI with persistent topics, visible tool activity, workspace controls, Apps, MCP presets, Skills, settings, and Automations. -->
+<!-- Meta description: Run nanobot from a browser WebUI with persistent and temporary chats, visible tool activity, workspace controls, Apps, skill discovery, settings, and Automations. -->
 
-The WebUI is nanobot's browser workbench for persistent topics, visible
-agent activity, workspace controls, Apps, Skills, settings, and Automations in
-one place.
+The WebUI is nanobot's browser workbench for persistent topics, temporary
+chats, visible agent activity, workspace controls, Apps, skill discovery,
+settings, and Automations in one place.
 
 The published `nanobot-ai` wheel already includes the WebUI bundle. You only need
 the `webui/` source directory when you are changing the frontend itself.
@@ -19,21 +19,24 @@ nanobot webui
 
 `nanobot webui` creates the config/workspace when needed, enables the local
 WebSocket channel after confirmation, generates a WebUI bootstrap secret when
-one is missing, starts the gateway, and opens the browser. With a fresh config,
+one is missing, starts or joins the same on-demand gateway used by the native
+TUI, and opens the browser. With a fresh config,
 it can open before a model is configured so you can finish setup in **Settings
 → Models**. The first-run path binds the WebUI to `127.0.0.1` by default, so
 it is not available from other devices on your LAN.
 
-Run it in the background when you do not want to keep a terminal open:
+After model setup, explicitly promote the shared gateway when you do not want to keep a client open:
 
 ```bash
-nanobot webui --background
+nanobot gateway --background
 ```
 
-Complete first-time model setup in a foreground `nanobot webui` session before using
-`--background`.
+`nanobot webui --background` is retained only to print migration guidance. This keeps one
+unambiguous owner for persistent process lifecycle.
 
-Manage the background gateway with `nanobot gateway status`, `nanobot gateway
+Each foreground WebUI or TUI launcher releases only its own client. The last
+interactive launcher stops an on-demand gateway. `nanobot gateway --background` makes the
+gateway persistent; manage it with `nanobot gateway status`, `nanobot gateway
 logs`, `nanobot gateway restart`, and `nanobot gateway stop`.
 
 Manual config still works. Same-machine localhost WebUI access can run without
@@ -72,14 +75,14 @@ This path avoids hand-editing `config.json` for normal setup. Use the reference
 
 | Area | Use it for |
 |---|---|
-| Topics | Start, switch, search, fork, and delete browser topics |
+| Topics | Start persistent topics or temporary chats; switch, search, reorder, fork, or delete persistent topics |
 | Agent activity | See thinking, tool calls, file edits with diffs, command output, and generated artifacts in context |
 | Workspace | Pick the project workspace before asking for file or shell work |
 | Access | Choose the access mode for local capabilities allowed by your gateway configuration |
 | Composer | Send text, images, voice input, slash commands, and `@` mentions for topics, Apps, or MCP presets |
 | Channels | Connect and validate chat platforms, install their optional support, and manage saved channel setup |
 | Apps | Install, test, update, and use local CLI App adapters and MCP presets |
-| Skills | Inspect available built-in and workspace skills before relying on them |
+| Skills | Inspect and manage installed skills, or discover skills from supported marketplaces |
 | Automations | Review, search, run, pause, edit, and delete scheduled and local-trigger agent turns |
 | Settings | Adjust models, providers, image generation, voice, web tools, runtime, and safety options |
 
@@ -90,6 +93,10 @@ workspace selection, and linked automations. Use a new topic when you want a
 separate context; use fork when you want to continue from an existing point
 without changing the original thread.
 
+Drag a topic within its current sidebar group to keep frequently used work in
+your preferred order. Drag a topic from the sidebar into the composer when you
+want to reference it in the next message instead of switching to it.
+
 The message timeline shows both user-visible replies and agent activity. Long
 tool or reasoning sections can be expanded when you need the details.
 
@@ -103,11 +110,34 @@ File previews follow the active session access mode. Restricted workspace access
 previews only files under the selected workspace. Full Access can preview files
 outside the workspace when that access mode is allowed by the gateway.
 
+## Temporary Chats
+
+Use a temporary chat for a conversation that should not be added to nanobot's
+topic history or long-term memory:
+
+1. Select **New topic**.
+2. Select the **Temporary chat** control in the page header.
+3. Send the first message.
+
+You can keep more than one temporary chat open and switch between them under
+**Temporary chats** in the sidebar while the current WebUI connection remains
+open. Reloading or closing the page, restarting the gateway, or losing the
+WebSocket connection ends all of them. They cannot be recovered afterward.
+
+Temporary does not mean consequence-free. Requests still go to the configured
+model provider, and tools can still change files, run commands, or affect
+external services. Temporary chats always use the default workspace in
+Restricted mode; the project picker and Full Access are unavailable. Commands
+and tools that create durable goals, automations, or subagent work are also
+unavailable. Use a regular topic when you need reusable context, scheduled work,
+or a result you must retain.
+
 ## Workspace and Access
 
 Use the workspace picker before starting project-specific work. This gives the
 agent the right project context for file paths, shell commands, and session
-metadata.
+metadata. A locally hosted WebUI opens the operating system's folder chooser
+when one is available; remote deployments keep the manual absolute path entry.
 
 Selecting a project does not replace the configured agent workspace. The two
 paths have different responsibilities:
@@ -145,7 +175,8 @@ clients.
 The composer supports plain messages, image attachments, voice input when
 transcription is configured, slash commands, and `@` mentions for installed Apps
 or MCP presets. Select another topic from the `@` menu to attach a stable
-reference; plain text that happens to start with `@` does not attach history.
+reference, or drag that topic from the sidebar into the composer. Plain text
+that happens to start with `@` does not attach history.
 Restricted chats offer topics from the same project, while Full Access chats can
 reference any WebUI topic. Nanobot reads a referenced topic only when its history
 is relevant and can link it in the response. The model badge shows the current
@@ -171,14 +202,23 @@ Test a new channel with a private DM. When a supported channel sends a pairing c
 
 ## Apps
 
-Open Apps from the sidebar to manage tools that nanobot can attach to a chat
-turn. The default **Ready** view shows only tools that can be used immediately:
+Open Apps from the sidebar to review and manage installable capabilities. The
+default **Ready** view shows only capabilities that can be used immediately:
 
-- **Apps** are local command-line adapters that nanobot runs on your machine.
-  Installing an adapter does not modify the native desktop or web app it
-  connects to.
-- **Integrations** are MCP servers. Presets provide known configurations, and
-  the custom integration panel accepts stdio, HTTP, and SSE servers.
+- **Agent Plugins** are local packages that can bundle skills, MCP servers, or
+  both. A package under `<workspace>/plugins/` is installed but remains inactive
+  until you enable it in Apps.
+- **CLI Apps** are local command-line adapters that nanobot runs on your
+  machine. Their installer manages the executable and exposes its adapter
+  through the same plugin activation model. Installing an adapter does not
+  modify the native desktop or web app it connects to.
+- **MCP** lists Model Context Protocol servers. Presets provide known
+  configurations, and the **Add MCP server** panel accepts stdio, HTTP, and SSE
+  servers. Custom HTTP/SSE servers can use no authentication, OAuth, or request
+  headers. After saving an OAuth server, choose **Connect** to open its sign-in
+  page. Presets such as Xmind, Notion, and Linear already use OAuth. HTTPS and
+  localhost WebUIs return automatically; a remote plain-HTTP WebUI shows one
+  field for pasting the complete localhost callback URL.
 
 Apps intentionally does not list nanobot runtime support packages such as
 `api` or `bedrock`. Those packages enable providers, servers, or channels; they
@@ -187,6 +227,7 @@ are not tools that can be attached to a turn with `@`. Manage them from
 included in nanobot and activate automatically when a file is attached. The
 equivalent CLI for optional integrations remains `nanobot plugins`. See
 [`cli-reference.md`](./cli-reference.md#optional-features).
+That command manages nanobot runtime extras, not Agent Plugin packages.
 
 Some MCP presets connect to hosted keyless endpoints. For example, the Firecrawl
 preset uses Firecrawl's hosted MCP endpoint for search, scrape, crawl, and
@@ -199,15 +240,26 @@ endpoint and exposes `web_search` and `web_fetch` without requiring an API key.
 It is an optional integration and does not replace nanobot's built-in web search
 provider; mention `@parallel-search` when a turn should use it.
 
-After an App or integration is available, mention it from the composer with
-`@` to attach that tool to the next message.
+After a CLI App or MCP server is available, mention it from the composer with
+`@` to attach that tool to the next message. Plugin-provided skills participate
+in normal skill discovery and can be invoked with `$skill-name`.
 
 ## Skills
 
-The Skills view shows the skill instructions available to the agent, including
-built-in skills and workspace-provided skills. Check this view when you want to
-know whether nanobot already has a focused workflow for a task before you ask it
-to perform that task.
+Open **Skills → Installed** to review built-in and workspace-provided skills.
+You can search and filter them, inspect their instructions and setup
+requirements, enable or disable them, and delete workspace skills you no longer
+want.
+
+Open **Skills → Discover** to browse or search skills from skills.sh and
+SkillHub. A marketplace skill is copied into the active agent workspace after
+you confirm the installation. skills.sh installation requires Node.js with
+`npx`; SkillHub installation does not.
+
+Marketplace skills are third-party instructions and may include executable
+scripts. Review the source and instructions before installing one, and enable
+only skills you trust with the same files, tools, and credentials available to
+your agent.
 
 ## Automations
 
@@ -295,10 +347,10 @@ trusts. Configure [`sslCertfile` and `sslKeyfile`](./websocket.md#tlsssl) on the
 WebSocket channel and open `https://<your-host>:8765`, or terminate HTTPS at a
 reverse proxy and use that proxy's HTTPS URL.
 
-Remote WebUI clients with a valid token can view and use Apps. Actions that
-install missing nanobot support packages, such as adding a channel dependency,
-are blocked by default. To let trusted remote administrators change the Python
-environment through the WebUI, opt in explicitly:
+Remote WebUI clients with a valid token can view and use Apps and installed
+skills. Actions that install missing nanobot support packages or third-party
+marketplace skills are blocked by default. To let trusted remote administrators
+perform those installations through the WebUI, opt in explicitly:
 
 ```json
 {
@@ -309,12 +361,13 @@ environment through the WebUI, opt in explicitly:
 ```
 
 Use this only for a private deployment where every authenticated WebUI user is
-trusted to change the Python environment that nanobot runs in. If you publish
-the WebUI through Nginx, Caddy, Cloudflare Tunnel, or a similar service, treat it
-as remote access and leave package installs disabled unless that is intentional.
+trusted to change nanobot's Python environment and install workspace skill
+instructions or scripts. If you publish the WebUI through Nginx, Caddy,
+Cloudflare Tunnel, or a similar service, treat it as remote access and leave
+package and skill installs disabled unless that is intentional.
 
 Optional feature installs use pip's configured package index, including
-`PIP_INDEX_URL`.
+`PIP_INDEX_URL`. skills.sh marketplace installs use `npx` instead.
 
 Leave remote package installs disabled when the WebUI is exposed beyond a
 private, trusted network.

nanobot/agent/context.py

@@ -42,25 +42,11 @@ def session_extra(metadata: Mapping[str, Any] | None) -> dict[str, Any]:
     )
 
 
-async def connect_mcp(state: Any, tools: ToolRegistry) -> None:
-    await mcp_tools.connect_missing_servers(state, tools)
-
-
-async def close_mcp(state: Any) -> None:
-    await mcp_tools.close_mcp_servers(state)
-
-
 async def handle_runtime_control(state: Any, msg: InboundMessage, tools: ToolRegistry) -> bool:
     if msg.metadata.get(INBOUND_META_RUNTIME_CONTROL) == RUNTIME_CONTROL_SESSION_DISCARD:
         await state.discard_session(msg.session_key)
         return True
-    for handler in (
-        image_generation_tools.handle_runtime_control,
-        mcp_tools.handle_runtime_control,
-    ):
-        if await handler(state, msg, tools):
-            return True
-    return False
+    return await image_generation_tools.handle_runtime_control(state, msg, tools)
 
 
 class ContextBuilder:

nanobot/agent/loop.py

@@ -36,6 +36,7 @@ from nanobot.agent.tools.exec_session import ExecSessionManager
 from nanobot.agent.tools.file_state import FileStateStore, bind_file_states, reset_file_states
 from nanobot.agent.tools.message import MessageTool
 from nanobot.agent.tools.registry import ToolRegistry
+from nanobot.agent.tools.runtime_control import AgentRuntimeControl
 from nanobot.agent.tools.self import MyTool
 from nanobot.agent.turn_delivery import (
     TurnDelivery,
@@ -43,7 +44,7 @@ from nanobot.agent.turn_delivery import (
 )
 from nanobot.agent.turn_delivery import TurnRoute as TurnRoute
 from nanobot.agent.turn_hooks import AgentTurnHookSpec, build_agent_turn_hook
-from nanobot.bus.events import InboundMessage, OutboundMessage
+from nanobot.bus.events import INBOUND_META_USER_SHELL, InboundMessage, OutboundMessage
 from nanobot.bus.outbound_events import StreamedResponseEvent
 from nanobot.bus.queue import MessageBus
 from nanobot.bus.runtime_events import RuntimeEventBus
@@ -75,6 +76,7 @@ from nanobot.session.goal_state import (
 from nanobot.session.history_visibility import HIDDEN_HISTORY_META
 from nanobot.session.keys import UNIFIED_SESSION_KEY, remember_last_channel
 from nanobot.session.manager import (
+    SESSION_CACHE_MAX_SIZE,
     Session,
     SessionManager,
     replay_max_messages_for_context,
@@ -94,11 +96,9 @@ from nanobot.utils.runtime import (
 )
 
 if TYPE_CHECKING:
-    from nanobot.agent.tools.mcp import MCPConnection
     from nanobot.config.schema import (
         ChannelsConfig,
         Config,
-        MCPServerConfig,
         ProviderConfig,
         ToolsConfig,
     )
@@ -163,6 +163,7 @@ class TurnContext:
     turn_wall_started_at: float = field(default_factory=time.time)
     visible_run_started_at: float | None = None
     turn_latency_ms: int | None = None
+    usage: dict[str, int] = field(default_factory=dict)
 
     def require_runtime(self) -> LLMRuntime:
         """Return the runtime established by the BUILD stage."""
@@ -197,6 +198,11 @@ class AgentLoop:
     def tool_names(self) -> list[str]:
         return self.tools.tool_names
 
+    @property
+    def last_usage(self) -> Mapping[str, int]:
+        """Latest aggregate usage exposed through the runtime-control snapshot."""
+        return self._last_usage
+
     @property
     def provider(self) -> LLMProvider:
         """Provider selected for future turn admissions."""
@@ -265,7 +271,7 @@ class AgentLoop:
         cron_service: CronService | None = None,
         restrict_to_workspace: bool = False,
         session_manager: SessionManager | None = None,
-        mcp_servers: dict[str, MCPServerConfig] | None = None,
+        tool_registry: ToolRegistry | None = None,
         channels_config: ChannelsConfig | None = None,
         timezone: str | None = None,
         session_ttl_minutes: int = 0,
@@ -372,11 +378,15 @@ class AgentLoop:
 
         self.context = ContextBuilder(workspace, timezone=timezone, disabled_skills=disabled_skills)
         self.sessions = session_manager or SessionManager(workspace)
-        self.sessions.set_file_cap_archiver(self.context.memory.raw_archive)
-        self.tools = ToolRegistry()
         # One file-read/write tracker per logical session. The tool registry is
         # shared by this loop, so tools resolve the active state via contextvars.
-        self._file_state_store = FileStateStore()
+        self._file_state_store = FileStateStore(max_sessions=SESSION_CACHE_MAX_SIZE)
+        # SessionManager owns every durable deletion entrypoint, including the
+        # WebUI and fork rollback paths.  Observe that boundary once instead of
+        # duplicating cleanup in each consumer.
+        self.sessions.set_delete_observer(self._file_state_store.discard)
+        self.sessions.set_file_cap_archiver(self.context.memory.raw_archive)
+        self.tools = tool_registry if tool_registry is not None else ToolRegistry()
         self._exec_session_manager = ExecSessionManager()
         self.runner = AgentRunner()
         self.subagents = SubagentManager(
@@ -393,14 +403,11 @@ class AgentLoop:
         )
         self._unified_session = unified_session
         self._running = False
-        self._mcp_servers = mcp_servers or {}
-        self._mcp_stacks: dict[str, MCPConnection] = {}
-        self._mcp_connecting = False
         self._runtime_context_providers: list[RuntimeContextProvider] = []
         self._active_tasks: dict[str, set[asyncio.Task[Any]]] = {}
         self._discarding_sessions: set[str] = set()
         self._background_tasks: set[asyncio.Task[Any]] = set()
-        self._close_mcp_lock = asyncio.Lock()
+        self._close_lock = asyncio.Lock()
         self._session_locks: weakref.WeakValueDictionary[str, asyncio.Lock] = (
             weakref.WeakValueDictionary()
         )
@@ -448,7 +455,6 @@ class AgentLoop:
         if model_preset:
             self.set_model_preset(model_preset, publish_update=False)
         self._register_default_tools(provider_snapshot_loader=provider_snapshot_loader)
-        self._runtime_vars: dict[str, Any] = {}
         self._current_iteration: int = 0
         self.commands = CommandRouter()
         register_builtin_commands(self.commands)
@@ -458,10 +464,15 @@ class AgentLoop:
         cls,
         config: Config,
         bus: MessageBus | None = None,
+        *,
+        tool_registry: ToolRegistry,
         **extra: Any,
     ) -> AgentLoop:
         """Create an AgentLoop from config with the common parameter set.
 
+        The tool registry is caller-owned so application composition can share
+        it with infrastructure such as an ``MCPProvider``.
+
         Extra keyword arguments are forwarded to ``AgentLoop.__init__``,
         allowing callers to override or extend the standard config-derived
         parameters (e.g. ``cron_service``, ``session_manager``).
@@ -471,6 +482,12 @@ class AgentLoop:
         if bus is None:
             bus = MessageBus()
         defaults = config.agents.defaults
+        if "session_manager" not in extra:
+            data_dir = config.runtime_data_dir
+            extra["session_manager"] = SessionManager(
+                config.workspace_path,
+                sessions_root=data_dir / "sessions" if data_dir is not None else None,
+            )
         provider = extra.pop("provider", None) or make_provider(config)
         resolved = config.resolve_preset()
         model = extra.pop("model", None) or resolved.model
@@ -494,7 +511,6 @@ class AgentLoop:
             provider_retry_mode=defaults.provider_retry_mode,
             tool_hint_max_length=defaults.tool_hint_max_length,
             restrict_to_workspace=config.tools.restrict_to_workspace,
-            mcp_servers=config.tools.mcp_servers,
             channels_config=config.channels,
             timezone=defaults.timezone,
             unified_session=defaults.unified_session,
@@ -509,6 +525,7 @@ class AgentLoop:
             restart_mode=config.gateway.restart_mode,
             provider_snapshot_loader=provider_snapshot_loader,
             preset_snapshot_loader=preset_snapshot_loader,
+            tool_registry=tool_registry,
             **extra,
         )
 
@@ -517,9 +534,15 @@ class AgentLoop:
         self.subagents.max_iterations = self.max_iterations
 
     def invalidate_runtime_config(self) -> None:
-        """Invalidate runtime config and notify clients to refresh its catalog."""
+        """Invalidate runtime config for lazy refresh at the next admission."""
         self.runtime_resolver.invalidate()
-        self._publish_runtime_selection(self.runtime_resolver.runtime)
+
+    def refresh_runtime_config(self) -> LLMRuntime:
+        """Refresh runtime config now and publish the canonical selection."""
+        self.runtime_resolver.invalidate()
+        runtime = self.runtime_resolver.admit()
+        self._publish_runtime_selection(runtime)
+        return runtime
 
     def runtime_for_session(
         self,
@@ -623,19 +646,18 @@ class AgentLoop:
         loader = ToolLoader()
         registered = loader.load(ctx, self.tools)
 
-        # MyTool needs runtime state reference — manual registration
+        # MyTool receives only the explicit runtime-control capability.
         if self.tools_config.my.enable:
             self.tools.register(
-                MyTool(runtime_state=self, modify_allowed=self.tools_config.my.allow_set)
+                MyTool(
+                    runtime_control=AgentRuntimeControl(self),
+                    modify_allowed=self.tools_config.my.allow_set,
+                )
             )
             registered.append("my")
 
         logger.info("Registered {} tools: {}", len(registered), registered)
 
-    async def _connect_mcp(self) -> None:
-        """Connect configured MCP servers."""
-        await agent_context.connect_mcp(self, self.tools)
-
     def register_runtime_context_provider(
         self,
         provider: RuntimeContextProvider,
@@ -780,6 +802,7 @@ class AgentLoop:
         dispatch_fn: Callable[[CommandContext], Awaitable[OutboundMessage | None]],
     ) -> None:
         """Dispatch a command directly from the run() loop and publish the result."""
+        async def dispatch_and_publish() -> None:
             ctx = CommandContext(msg=msg, session=None, key=key, raw=raw, loop=self)
             result = await dispatch_fn(ctx)
             if result:
@@ -787,6 +810,59 @@ class AgentLoop:
             else:
                 logger.warning("Command '{}' matched but dispatch returned None", raw)
 
+        # A shell command may run for up to the configured exec timeout. Keep
+        # the inbound consumer responsive when it runs beside an active turn.
+        if (msg.metadata or {}).get(INBOUND_META_USER_SHELL) is True:
+            self.schedule_background(dispatch_and_publish())
+            return
+        await dispatch_and_publish()
+
+    async def execute_user_shell_command(self, ctx: CommandContext) -> OutboundMessage:
+        """Execute one trusted user command with the active workspace policy."""
+        metadata = dict(ctx.msg.metadata or {})
+        tool = self.tools.get("exec")
+        if tool is None:
+            content = "Shell execution is disabled in this nanobot configuration."
+        else:
+            session = ctx.session or self.sessions.get_or_create(ctx.key)
+            scope = self.workspace_scopes.for_turn(
+                channel=ctx.msg.channel,
+                message_metadata=metadata,
+                session_metadata=session.metadata,
+            )
+            request_token = bind_request_context(RequestContext(
+                channel=ctx.msg.channel,
+                chat_id=ctx.msg.chat_id,
+                message_id=metadata.get("message_id"),
+                session_key=ctx.key,
+                original_user_text=f"!{ctx.args.strip()}",
+                runtime=ctx.runtime,
+                metadata=metadata,
+                sender_id=ctx.msg.sender_id,
+                turn_id=metadata.get("webui_turn_id"),
+                workspace=scope.project_path,
+            ))
+            workspace_token = bind_workspace_scope(scope)
+            turn_scope_stack = ExitStack()
+            try:
+                for turn_scope in ctx.turn_scopes:
+                    turn_scope_stack.enter_context(turn_scope)
+                result = await tool.execute(
+                    command=ctx.args.strip(),
+                    working_dir=str(scope.project_path),
+                )
+                content = str(result)
+            finally:
+                turn_scope_stack.close()
+                reset_workspace_scope(workspace_token)
+                reset_request_context(request_token)
+        return OutboundMessage(
+            channel=ctx.msg.channel,
+            chat_id=ctx.msg.chat_id,
+            content=content,
+            metadata={**metadata, "render_as": "text"},
+        )
+
     async def _cancel_active_tasks(self, key: str) -> int:
         """Cancel and await all active work for *key*.
 
@@ -808,8 +884,13 @@ class AgentLoop:
             self.sessions.invalidate(key)
             await self._cancel_active_tasks(key)
         finally:
+            self.discard_session_file_state(key)
             self._discarding_sessions.discard(key)
 
+    def discard_session_file_state(self, key: str) -> None:
+        """Forget ephemeral file-read state for a reset or removed session."""
+        self._file_state_store.discard(key)
+
     def _effective_session_key(self, msg: InboundMessage) -> str:
         """Return the session key used for task routing and mid-turn injections."""
         if self._unified_session and not msg.session_key_override:
@@ -1147,7 +1228,6 @@ class AgentLoop:
         """Run the agent loop, dispatching messages as tasks to stay responsive to /stop."""
         self._running = True
         try:
-            await self._connect_mcp()
             logger.info("Agent loop started")
 
             while self._running:
@@ -1238,8 +1318,7 @@ class AgentLoop:
                 active_tasks.add(task)
                 task.add_done_callback(active_tasks.discard)
         finally:
-            # MCP stdio transports use AnyIO cancel scopes; close them from the task that opened them.
-            await self.close_mcp()
+            await self.aclose()
 
     async def _dispatch(self, msg: InboundMessage) -> None:
         """Process a message: per-session serial, cross-session concurrent."""
@@ -1357,24 +1436,24 @@ class AgentLoop:
                 await delivery.idle()
                 await self._publish_next_deferred_automation_turn(session_key)
 
-    async def close_mcp(self) -> None:
-        """Stop active work, then close exec, subagent, and MCP resources.
+    async def aclose(self) -> None:
+        """Stop active work, then close resources owned by the agent loop.
 
         Resource teardown must still run if cancellation interrupts task draining.
         Gateway shutdown deliberately bounds this coroutine, so keeping the cleanup
         phase in ``finally`` prevents a timed-out background task from leaving
         subprocess transports alive after the event loop closes.
         """
-        # The agent loop closes itself from ``run()`` while gateway shutdown also
+        # The loop closes itself from ``run()`` while application shutdown also
         # performs a guaranteed final close. Serialize those owners so they cannot
-        # tear down the same subprocess transports concurrently.
-        close_lock = getattr(self, "_close_mcp_lock", None)
+        # tear down the same resources concurrently.
+        close_lock = getattr(self, "_close_lock", None)
         if close_lock is None:
-            close_lock = self._close_mcp_lock = asyncio.Lock()
+            close_lock = self._close_lock = asyncio.Lock()
         async with close_lock:
-            await self._close_mcp_unlocked()
+            await self._aclose_unlocked()
 
-    async def _close_mcp_unlocked(self) -> None:
+    async def _aclose_unlocked(self) -> None:
         errors: list[BaseException] = []
         active_task_groups = getattr(self, "_active_tasks", {})
         active_tasks = tuple({task for tasks in active_task_groups.values() for task in tasks})
@@ -1397,7 +1476,6 @@ class AgentLoop:
         cleanup_steps = (
             self.subagents.close,
             self._exec_session_manager.close_all,
-            lambda: agent_context.close_mcp(self),
         )
         for cleanup in cleanup_steps:
             try:
@@ -1770,7 +1848,7 @@ class AgentLoop:
                 session.provider_state = None
                 self.sessions.save(session)
             ctx.input_persisted_early = True
-        ctx.delivery.record_runtime(runtime)
+        await ctx.delivery.runtime_admitted(runtime)
 
         ctx.request_context = self._request_context_for_turn(ctx)
         if ctx.kind is TurnKind.USER:
@@ -1874,6 +1952,8 @@ class AgentLoop:
         ctx.all_messages = all_msgs
         ctx.stop_reason = stop_reason
         ctx.had_injections = had_injections
+        ctx.usage = dict(self._last_usage)
+        ctx.delivery.record_usage(ctx.usage)
         if ctx.kind is TurnKind.USER:
             await turn_continuation.maybe_continue_turn(ctx)
 
@@ -1899,6 +1979,8 @@ class AgentLoop:
             else ctx.turn_wall_started_at
         )
         ctx.turn_latency_ms = max(0, int((time.time() - latency_started_at) * 1000))
+        if ctx.usage and not ctx.ephemeral:
+            session.metadata["_last_usage"] = dict(ctx.usage)
         self._save_turn(
             session, ctx.all_messages, ctx.save_skip,
             turn_latency_ms=ctx.turn_latency_ms,
@@ -2286,7 +2368,6 @@ class AgentLoop:
         """Process an external message directly and return the outbound payload."""
         if channel == "system":
             raise ValueError("channel 'system' is reserved for internal messages")
-        await self._connect_mcp()
         metadata: dict[str, Any] = {}
         if not persist_user_message:
             metadata[turn_continuation.SKIP_USER_PERSIST_META] = True

nanobot/agent/memory.py

@@ -769,27 +769,24 @@ class MemoryStore:
         return f"{prefix}\n\n{diff_body}"
 
     @staticmethod
-    def prune_dream_sessions(sessions_dir: Path, *, keep: int = 10) -> None:
+    def prune_dream_sessions(sessions: SessionManager, *, keep: int = 10) -> None:
         """Remove the oldest Dream session files, keeping only the N most recent.
 
         Only current base64url-encoded Dream session keys are considered.
         Non-dream session files are never touched.
         """
-        dream_files: list[Path] = []
+        with sessions.locked_session_files() as sessions_dir:
+            dream_files: list[tuple[Path, str]] = []
             for path in sessions_dir.glob("*.jsonl"):
                 decoded_key = SessionManager.decode_storage_key(path.stem)
                 if decoded_key is not None and decoded_key.startswith("dream:"):
-                dream_files.append(path)
-        dream_files.sort(key=lambda p: p.stat().st_mtime)
-        if len(dream_files) <= keep:
-            return
+                    dream_files.append((path, decoded_key))
+            dream_files.sort(key=lambda item: item[0].stat().st_mtime)
 
-        to_remove = dream_files[: len(dream_files) - keep]
-        for path in to_remove:
-            try:
-                path.unlink()
+            for path, key in dream_files[: max(0, len(dream_files) - keep)]:
+                if sessions.delete_session(key):
                     logger.debug("Pruned old dream session: {}", path.stem)
-            except OSError:
+                else:
                     logger.warning("Failed to prune dream session {}", path)
 
 

nanobot/agent/model_presets.py

@@ -79,7 +79,9 @@ def normalize_preset_name(name: str | None, presets: dict[str, ModelPresetConfig
     if not isinstance(name, str) or not name.strip():
         raise ValueError("model_preset must be a non-empty string")
     name = name.strip()
-    if name not in presets:
-        raise KeyError(f"model_preset {name!r} not found. Available: {', '.join(presets) or '(none)'}")
+    if name in presets:
         return name
-
+    matches = [candidate for candidate in presets if candidate.casefold() == name.casefold()]
+    if len(matches) == 1:
+        return matches[0]
+    raise KeyError(f"model_preset {name!r} not found. Available: {', '.join(presets) or '(none)'}")

nanobot/agent/plugins.py

@@ -0,0 +1,512 @@
+"""Load and activate locally installed Agent Plugin packages."""
+
+from __future__ import annotations
+
+import base64
+import json
+import re
+from dataclasses import dataclass, replace
+from hashlib import sha256
+from pathlib import Path
+from typing import cast
+
+from loguru import logger
+from pydantic import ValidationError
+
+from nanobot.agent.skills import parse_skill_metadata, valid_skill_metadata
+from nanobot.config.loader import get_config_path
+from nanobot.config.schema import MCPServerConfig
+
+AGENT_PLUGIN_SCHEMA = "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json"
+AGENT_PLUGIN_MCP_SCHEMA = "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json"
+
+_PLUGIN_NAME = re.compile(r"^(?!.*(?:--|\.\.))[a-z0-9](?:[a-z0-9.-]*[a-z0-9])?$")
+_MCP_SERVER_FIELDS = {"type", "command", "args", "env", "cwd"}
+_MAX_LOGO_BYTES = 256 * 1024
+
+
+@dataclass(frozen=True, slots=True)
+class _PackageSnapshot:
+    root: Path
+    fingerprint: str
+    skill_dirs: tuple[Path, ...]
+
+
+@dataclass(frozen=True, slots=True)
+class _SkillCacheEntry:
+    skills: tuple[tuple[str, Path], ...]
+    packages: tuple[_PackageSnapshot, ...]
+
+
+_SKILL_CACHE: dict[tuple[Path, Path], _SkillCacheEntry] = {}
+
+
+@dataclass(frozen=True)
+class AgentPlugin:
+    """A validated, locally installed Agent Plugins v1 package."""
+
+    name: str
+    root: Path
+    description: str
+    repository: str
+    display_name: str
+    category: str
+    accent_color: str | None
+    logo: str | None
+    permissions: tuple[str, ...]
+    mcp_servers: tuple[str, ...] = ()
+    enabled: bool = False
+
+
+def _installed_plugins(workspace: Path) -> list[AgentPlugin]:
+    """Return installed packages found under ``<workspace>/plugins/*``."""
+    workspace = workspace.expanduser().resolve()
+    root = _contained(workspace / "plugins", workspace, directory=True)
+    if root is None:
+        return []
+    plugins: dict[str, AgentPlugin | None] = {}
+    for candidate in _children(root, "Agent Plugins directory"):
+        plugin_root = _contained(candidate, root, directory=True)
+        if plugin_root is None:
+            continue
+        plugin = _load_manifest(plugin_root)
+        if plugin is not None:
+            if plugin.name in plugins:
+                logger.warning("Ignoring duplicate Agent Plugin identity '{}'", plugin.name)
+                plugins[plugin.name] = None
+            else:
+                plugins[plugin.name] = plugin
+    return [plugin for plugin in plugins.values() if plugin is not None]
+
+
+def enabled_agent_plugin_skills(workspace: Path) -> list[tuple[str, Path]]:
+    """Verify and return skills from plugins the user has explicitly enabled."""
+    skills: list[tuple[str, Path]] = []
+    packages: list[_PackageSnapshot] = []
+    for plugin in _installed_plugins(workspace):
+        plugin_skills = _discover_plugin_skills(plugin.name, plugin.root)
+        fingerprint = _enabled_package_fingerprint(workspace, plugin)
+        if fingerprint is None:
+            continue
+        skills.extend(plugin_skills)
+        if plugin_skills:
+            packages.append(
+                _PackageSnapshot(
+                    root=plugin.root,
+                    fingerprint=fingerprint,
+                    skill_dirs=tuple(path.parent for _name, path in plugin_skills),
+                )
+            )
+
+    key = _skill_cache_key(workspace)
+    _SKILL_CACHE[key] = _SkillCacheEntry(tuple(skills), tuple(packages))
+    return skills
+
+
+def enabled_agent_plugin_skill_dirs(
+    workspace: Path,
+    *,
+    requested_path: str | Path | None = None,
+) -> tuple[Path, ...]:
+    """Return skill roots authorized for one read, revalidating their package."""
+    key = _skill_cache_key(workspace)
+    cached = _SKILL_CACHE.get(key)
+    if cached is None:
+        enabled_agent_plugin_skills(workspace)
+        cached = _SKILL_CACHE.get(key)
+    if cached is None:
+        return ()
+
+    target = (
+        Path(requested_path).expanduser().resolve(strict=False)
+        if requested_path is not None
+        else None
+    )
+    packages = tuple(
+        package
+        for package in cached.packages
+        if target is None
+        or any(target == root or target.is_relative_to(root) for root in package.skill_dirs)
+    )
+    if any(_package_fingerprint(package.root) != package.fingerprint for package in packages):
+        # Re-run the full activation check so a changed package loses its
+        # marker and cannot become readable again through this cache.
+        _invalidate_skill_cache(workspace)
+        enabled_agent_plugin_skills(workspace)
+        return ()
+
+    if target is None:
+        return tuple(root for package in packages for root in package.skill_dirs)
+    return tuple(
+        root
+        for package in packages
+        for root in package.skill_dirs
+        if target == root or target.is_relative_to(root)
+    )
+
+
+def _skill_cache_key(workspace: Path) -> tuple[Path, Path]:
+    return (
+        workspace.expanduser().resolve(),
+        get_config_path().expanduser().resolve(),
+    )
+
+
+def _invalidate_skill_cache(workspace: Path) -> None:
+    _SKILL_CACHE.pop(_skill_cache_key(workspace), None)
+
+
+def _package_fingerprint(root: Path) -> str | None:
+    """Hash package paths, link targets, and file contents."""
+    digest = sha256()
+    try:
+        for candidate in sorted(root.rglob("*")):
+            relative = candidate.relative_to(root).as_posix()
+            digest.update(relative.encode())
+            if candidate.is_symlink():
+                digest.update(b"\0link\0")
+                digest.update(candidate.readlink().as_posix().encode())
+            elif candidate.is_file():
+                digest.update(b"\0file\0")
+                digest.update(candidate.read_bytes())
+            elif candidate.is_dir():
+                digest.update(b"\0dir\0")
+            else:
+                return None
+            digest.update(b"\0")
+    except OSError:
+        return None
+    return digest.hexdigest()
+
+
+def _load_manifest(plugin_root: Path) -> AgentPlugin | None:
+    payload = _read_object(plugin_root / "plugin.json", plugin_root)
+    if payload is None:
+        return None
+    if payload.get("$schema") != AGENT_PLUGIN_SCHEMA:
+        return None
+    name = payload.get("name")
+    if (
+        not isinstance(name, str)
+        or len(name) > 64
+        or _PLUGIN_NAME.fullmatch(name) is None
+    ):
+        logger.warning("Ignoring Agent Plugin manifest in '{}': invalid name", plugin_root)
+        return None
+    extension = payload.get("extensions")
+    extension_payload = cast(dict[str, object], extension) if isinstance(extension, dict) else {}
+    nanobot_value = extension_payload.get("dev.nanobot")
+    nanobot = cast(dict[str, object], nanobot_value) if isinstance(nanobot_value, dict) else {}
+    return AgentPlugin(
+        name=name,
+        root=plugin_root,
+        description=_string(payload.get("description")),
+        repository=_string(payload.get("repository")),
+        display_name=_string(nanobot.get("displayName")) or name,
+        category=_string(nanobot.get("category")) or "Plugin",
+        accent_color=_accent_color(nanobot.get("accentColor")),
+        logo=_plugin_logo(nanobot.get("logo"), plugin_root),
+        permissions=_string_tuple(nanobot.get("permissions")),
+    )
+
+
+def agent_plugin_mcp_servers(
+    workspace: Path,
+    configured: dict[str, MCPServerConfig] | None = None,
+) -> dict[str, MCPServerConfig]:
+    """Merge explicitly enabled plugin MCP servers with user configuration.
+
+    User configuration wins on the unlikely event of a namespaced collision.
+    """
+    servers: dict[str, MCPServerConfig] = {}
+    for plugin in _installed_plugins(workspace):
+        if not _enabled(workspace, plugin):
+            continue
+        plugin_servers = _plugin_mcp_servers(workspace, plugin)
+        for name, server in plugin_servers.items():
+            # ``--`` cannot occur in a valid plugin identity, so multi-server
+            # namespaces cannot collide with a single-server plugin name.
+            host_name = plugin.name if len(plugin_servers) == 1 else f"{plugin.name}--{name}"
+            servers[host_name] = server
+    configured = configured or {}
+    if collisions := servers.keys() & configured.keys():
+        logger.warning("Configured MCP servers override Agent Plugins: {}", ", ".join(sorted(collisions)))
+    return servers | configured
+
+
+def discover_agent_plugins(workspace: Path) -> list[AgentPlugin]:
+    """Return component and lifecycle state for discovered plugins."""
+    return [
+        replace(
+            plugin,
+            mcp_servers=tuple(sorted(_plugin_mcp_servers(workspace, plugin))),
+            enabled=_enabled(workspace, plugin),
+        )
+        for plugin in _installed_plugins(workspace)
+    ]
+
+
+def set_agent_plugin_enabled(workspace: Path, name: str, enabled: bool) -> None:
+    """Enable or disable one installed plugin."""
+    plugin = next((item for item in _installed_plugins(workspace) if item.name == name), None)
+    if plugin is None:
+        raise ValueError(f"unknown Agent Plugin '{name}'")
+    data = _plugin_data_dir(workspace, plugin.name, create=True)
+    marker = data / "enabled"
+    if enabled:
+        activation = _activation_marker(plugin)
+        if activation is None:
+            raise RuntimeError(f"Agent Plugin '{name}' changed while it was being enabled")
+        marker.write_text(activation, encoding="utf-8")
+        marker.chmod(0o600)
+    else:
+        marker.unlink(missing_ok=True)
+    _invalidate_skill_cache(workspace)
+
+
+def _string(value: object) -> str:
+    return value.strip() if isinstance(value, str) else ""
+
+
+def _string_tuple(value: object) -> tuple[str, ...]:
+    items = cast(list[object], value) if isinstance(value, list) else []
+    return tuple(item.strip() for item in items if isinstance(item, str) and item.strip())
+
+
+def _accent_color(value: object) -> str | None:
+    return value if isinstance(value, str) and re.fullmatch(r"#[0-9a-fA-F]{6}", value) else None
+
+
+def _plugin_logo(value: object, plugin_root: Path) -> str | None:
+    """Resolve nanobot's optional packaged logo extension."""
+    if value is None:
+        return None
+    if not isinstance(value, str) or not value.startswith("./"):
+        logger.warning("Ignoring invalid Agent Plugin logo in '{}'", plugin_root)
+        return None
+    logo = _contained(plugin_root / value[2:], plugin_root)
+    try:
+        data = logo.read_bytes() if logo is not None else b""
+        suffix = logo.suffix.lower() if logo is not None else ""
+        if len(data) <= _MAX_LOGO_BYTES and (
+            suffix == ".png" and data.startswith(b"\x89PNG\r\n\x1a\n")
+            or suffix in {".jpg", ".jpeg"} and data.startswith(b"\xff\xd8\xff")
+            or suffix == ".webp" and data.startswith(b"RIFF") and data[8:12] == b"WEBP"
+        ):
+            mime = "jpeg" if suffix in {".jpg", ".jpeg"} else suffix[1:]
+            return f"data:image/{mime};base64,{base64.b64encode(data).decode('ascii')}"
+    except OSError:
+        pass
+    logger.warning("Ignoring invalid Agent Plugin logo in '{}'", plugin_root)
+    return None
+
+
+def _plugin_mcp_servers(workspace: Path, plugin: AgentPlugin) -> dict[str, MCPServerConfig]:
+    payload = _read_object(plugin.root / "mcp.json", plugin.root)
+    if payload is None:
+        return {}
+    raw_servers = payload.get("mcpServers")
+    if (
+        payload.keys() != {"$schema", "mcpServers"}
+        or payload.get("$schema") != AGENT_PLUGIN_MCP_SCHEMA
+        or not isinstance(raw_servers, dict)
+    ):
+        logger.warning("Ignoring invalid MCP component for Agent Plugin '{}'", plugin.name)
+        return {}
+
+    data = _plugin_data_dir(workspace, plugin.name, create=True)
+    servers: dict[str, MCPServerConfig] = {}
+    for name, raw in cast(dict[str, object], raw_servers).items():
+        if not name or len(name) > 128 or any(ord(char) < 32 for char in name):
+            logger.warning("Ignoring invalid MCP server name in Agent Plugin '{}'", plugin.name)
+            continue
+        server = _plugin_mcp_server(raw, plugin.root, data)
+        if server is None:
+            logger.warning("Ignoring invalid MCP server '{}' in Agent Plugin '{}'", name, plugin.name)
+            continue
+        servers[name] = server
+    return servers
+
+
+def _plugin_mcp_server(raw: object, root: Path, data: Path) -> MCPServerConfig | None:
+    if not isinstance(raw, dict):
+        return None
+    payload = cast(dict[str, object], raw)
+    if payload.keys() - _MCP_SERVER_FIELDS:
+        return None
+    try:
+        server = MCPServerConfig.model_validate(payload)
+    except ValidationError:
+        return None
+    command = _stdio_command(server.command, root)
+    cwd = _stdio_cwd(payload.get("cwd"), root, data)
+    if server.type != "stdio" or command is None or cwd is None:
+        return None
+    if {"PLUGIN_ROOT", "PLUGIN_DATA"} & server.env.keys():
+        return None
+    return server.model_copy(
+        update={
+            "command": command,
+            "args": [_expand(item, root, data) for item in server.args],
+            "env": {
+                **{key: _expand(value, root, data) for key, value in server.env.items()},
+                "PYTHONDONTWRITEBYTECODE": "1",
+                "PLUGIN_ROOT": str(root),
+                "PLUGIN_DATA": str(data),
+            },
+            "cwd": str(cwd),
+        }
+    )
+
+
+def _stdio_command(value: object, root: Path) -> str | None:
+    if not isinstance(value, str) or not value:
+        return None
+    if value.startswith("./"):
+        executable = _contained(root / value[2:], root)
+        return str(executable) if executable is not None else None
+    if any(char.isspace() for char in value) or "/" in value or "\\" in value:
+        return None
+    return value
+
+
+def _stdio_cwd(value: object, root: Path, data: Path) -> Path | None:
+    if value is None:
+        return root
+    if not isinstance(value, str):
+        return None
+    if value.startswith("./"):
+        return _contained(root / value[2:], root, directory=True)
+    for placeholder, base in (("${PLUGIN_ROOT}", root), ("${PLUGIN_DATA}", data)):
+        if value == placeholder or value.startswith(f"{placeholder}/"):
+            relative = value[len(placeholder):].lstrip("/")
+            candidate = (base / relative).resolve()
+            if not candidate.is_relative_to(base):
+                return None
+            if base == data:
+                candidate.mkdir(parents=True, exist_ok=True)
+                candidate.chmod(0o700)
+            return candidate if candidate.is_dir() else None
+    return None
+
+
+def _expand(value: str, root: Path, data: Path) -> str:
+    return value.replace("${PLUGIN_ROOT}", str(root)).replace("${PLUGIN_DATA}", str(data))
+
+
+def _plugin_data_dir(workspace: Path, name: str, *, create: bool) -> Path:
+    workspace_id = sha256(str(workspace.expanduser().resolve()).encode()).hexdigest()[:12]
+    current = get_config_path().expanduser().resolve().parent
+    for segment in ("plugin-data", workspace_id, name):
+        path = current / segment
+        if create:
+            path.mkdir(parents=True, exist_ok=True)
+        try:
+            resolved = path.resolve(strict=create)
+        except OSError as exc:
+            raise RuntimeError("Agent Plugin data directory is unavailable") from exc
+        if not resolved.is_relative_to(current):
+            raise RuntimeError("Agent Plugin data directory escapes its parent")
+        if create:
+            resolved.chmod(0o700)
+        current = resolved
+    return current
+
+
+def _enabled_package_fingerprint(workspace: Path, plugin: AgentPlugin) -> str | None:
+    """Return the content fingerprint when this exact package is enabled."""
+    marker = _plugin_data_dir(workspace, plugin.name, create=False) / "enabled"
+    try:
+        if not marker.is_file():
+            return None
+        current = marker.read_text(encoding="utf-8")
+        activation = _activation_marker(plugin)
+        if activation is None:
+            marker.unlink(missing_ok=True)
+            _invalidate_skill_cache(workspace)
+            return None
+        payload = cast(dict[str, object], json.loads(activation))
+        fingerprint = payload.get("fingerprint")
+        if not isinstance(fingerprint, str):
+            return None
+        if current == activation:
+            return fingerprint
+        if current == str(plugin.root):
+            marker.write_text(activation, encoding="utf-8")
+            marker.chmod(0o600)
+            return fingerprint
+        marker.unlink(missing_ok=True)
+        _invalidate_skill_cache(workspace)
+        return None
+    except (OSError, json.JSONDecodeError):
+        _invalidate_skill_cache(workspace)
+        return None
+
+
+def _enabled(workspace: Path, plugin: AgentPlugin) -> bool:
+    return _enabled_package_fingerprint(workspace, plugin) is not None
+
+
+def _activation_marker(plugin: AgentPlugin) -> str | None:
+    """Bind activation to one immutable package snapshot."""
+    fingerprint = _package_fingerprint(plugin.root)
+    if fingerprint is None:
+        return None
+    return json.dumps(
+        {"fingerprint": fingerprint, "root": str(plugin.root)},
+        separators=(",", ":"),
+        sort_keys=True,
+    )
+
+
+def _discover_plugin_skills(plugin_name: str, plugin_root: Path) -> list[tuple[str, Path]]:
+    skills_root = _contained(plugin_root / "skills", plugin_root, directory=True)
+    if skills_root is None:
+        return []
+
+    skills: list[tuple[str, Path]] = []
+    for candidate in _children(skills_root, f"Agent Plugin '{plugin_name}' skills"):
+        skill_root = _contained(candidate, skills_root, directory=True)
+        if skill_root is None:
+            continue
+        skill_file = _contained(skill_root / "SKILL.md", plugin_root)
+        if skill_file is None:
+            continue
+        try:
+            metadata = parse_skill_metadata(skill_file.read_text(encoding="utf-8"))
+        except (OSError, UnicodeError):
+            metadata = None
+        if metadata is None or not valid_skill_metadata(metadata, candidate.name):
+            logger.warning("Ignoring Agent Plugin '{}' skill '{}': invalid metadata", plugin_name, candidate.name)
+            continue
+        skills.append((candidate.name, skill_file))
+    return skills
+
+
+def _children(root: Path, label: str) -> list[Path]:
+    try:
+        return sorted(root.iterdir(), key=lambda path: path.name)
+    except OSError as exc:
+        logger.warning("Could not inspect {}: {}", label, exc)
+        return []
+
+
+def _contained(path: Path, root: Path, *, directory: bool = False) -> Path | None:
+    try:
+        resolved = path.resolve(strict=True)
+    except OSError:
+        return None
+    expected_kind = resolved.is_dir() if directory else resolved.is_file()
+    return resolved if expected_kind and resolved.is_relative_to(root) else None
+
+
+def _read_object(path: Path, root: Path) -> dict[str, object] | None:
+    contained = _contained(path, root)
+    if contained is None:
+        return None
+    try:
+        value = cast(object, json.loads(contained.read_text(encoding="utf-8")))
+    except (OSError, UnicodeError, json.JSONDecodeError) as exc:
+        logger.warning("Ignoring invalid Agent Plugin component '{}': {}", contained, exc)
+        return None
+    return cast(dict[str, object], value) if isinstance(value, dict) else None

nanobot/agent/runner.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import inspect
 import os
+import time
 from collections.abc import Awaitable, Callable, Iterable
 from copy import deepcopy
 from dataclasses import dataclass, field
@@ -932,6 +933,27 @@ class AgentRunner:
 
         progress_state: dict[str, bool] | None = None
         active_hosted_tools: dict[str, dict[str, Any]] = {}
+        request_started_at = 0.0
+        first_output_at: float | None = None
+        generation_started_at: float | None = None
+        generation_elapsed_s = 0.0
+
+        def _generation_delta(delta: str) -> None:
+            nonlocal first_output_at, generation_started_at
+            if not delta:
+                return
+            now = time.perf_counter()
+            if first_output_at is None:
+                first_output_at = now
+            if generation_started_at is None:
+                generation_started_at = now
+
+        def _pause_generation() -> None:
+            nonlocal generation_elapsed_s, generation_started_at
+            if generation_started_at is None:
+                return
+            generation_elapsed_s += max(0.0, time.perf_counter() - generation_started_at)
+            generation_started_at = None
 
         async def _provider_tool_event(event: dict[str, Any]) -> None:
             if event.get("kind") != "hosted_tool":
@@ -950,6 +972,7 @@ class AgentRunner:
             thinking_buf = ""
 
             async def _stream(delta: str) -> None:
+                _generation_delta(delta)
                 if delta:
                     context.streamed_content = True
                 await hook.on_stream(context, delta)
@@ -958,6 +981,7 @@ class AgentRunner:
                 nonlocal thinking_buf
                 if not delta:
                     return
+                _generation_delta(delta)
                 prev_clean = strip_reasoning_tags(thinking_buf)
                 thinking_buf += delta
                 new_clean = strip_reasoning_tags(thinking_buf)
@@ -967,6 +991,7 @@ class AgentRunner:
                     await hook.emit_reasoning(incremental)
 
             async def _stream_recover() -> None:
+                _pause_generation()
                 await hook.on_stream_end(context, resuming=True)
 
             coro = spec.runtime.provider.chat_stream_with_retry(
@@ -986,6 +1011,7 @@ class AgentRunner:
                 nonlocal stream_buf
                 if not delta:
                     return
+                _generation_delta(delta)
                 prev_clean = strip_think(stream_buf)
                 stream_buf += delta
                 new_clean = strip_think(stream_buf)
@@ -1027,6 +1053,7 @@ class AgentRunner:
             if is_streaming_request and timeout_s is not None
             else timeout_s
         )
+        request_started_at = time.perf_counter()
         try:
             response = (
                 await coro if outer_timeout_s is None
@@ -1045,6 +1072,11 @@ class AgentRunner:
                     finish_reason="error",
                     error_kind="timeout",
                 )
+        _pause_generation()
+        if first_output_at is not None:
+            response.ttft_ms = max(0, round((first_output_at - request_started_at) * 1000))
+        if generation_elapsed_s > 0:
+            response.generation_ms = max(1, round(generation_elapsed_s * 1000))
         # chat_stream_with_retry may recover internally, so only fail unfinished
         # hosted calls after the provider returns its final error response.
         if response.finish_reason == "error":
@@ -1288,10 +1320,18 @@ class AgentRunner:
         if total > 0:
             usage["total_tokens"] = total
             usage.setdefault("provider_tokens", total)
-            return usage
-        if response.finish_reason == "error":
+        elif response.finish_reason == "error":
             return {}
-        return self._estimate_response_usage(spec, messages, response)
+        else:
+            usage = self._estimate_response_usage(spec, messages, response)
+        completion = usage.get("completion_tokens", 0)
+        if response.generation_ms is not None and completion > 0:
+            usage["generation_ms"] = response.generation_ms
+            usage["measured_completion_tokens"] = completion
+        if response.ttft_ms is not None:
+            usage["ttft_ms"] = response.ttft_ms
+            usage["timed_requests"] = 1
+        return usage
 
     def _estimate_response_usage(
         self,

nanobot/agent/skills.py

@@ -17,9 +17,35 @@ _STRIP_SKILL_FRONTMATTER = re.compile(
     r"^---\s*\r?\n(.*?)\r?\n---\s*\r?\n?",
     re.DOTALL,
 )
+_SKILL_NAME = re.compile(r"^(?!.*--)[a-z0-9](?:[a-z0-9-]*[a-z0-9])?$")
 _SKILL_REFERENCE = re.compile(r"(?<![\w$])\$([A-Za-z0-9_-]+)")
 
 
+def parse_skill_metadata(content: str) -> dict[str, object] | None:
+    """Parse a skill document's YAML frontmatter."""
+    if not (match := _STRIP_SKILL_FRONTMATTER.match(content)):
+        return None
+    try:
+        parsed = yaml.safe_load(match.group(1))
+    except yaml.YAMLError:
+        return None
+    if not isinstance(parsed, dict):
+        return None
+    return {str(key): value for key, value in cast(dict[object, object], parsed).items()}
+
+
+def valid_skill_metadata(metadata: dict[str, object], name: str) -> bool:
+    """Return whether metadata satisfies the Agent Skills identity contract."""
+    description = metadata.get("description")
+    return (
+        metadata.get("name") == name
+        and len(name) <= 64
+        and _SKILL_NAME.fullmatch(name) is not None
+        and isinstance(description, str)
+        and 1 <= len(description.strip()) <= 1024
+    )
+
+
 class SkillsLoader:
     """
     Loader for agent skills.
@@ -34,6 +60,15 @@ class SkillsLoader:
         self.builtin_skills = builtin_skills_dir or BUILTIN_SKILLS_DIR
         self.disabled_skills = disabled_skills or set()
 
+    def _skill_aliases(self) -> dict[str, str]:
+        """Return compatibility aliases owned by installed CLI Apps."""
+        from nanobot.apps.cli import CliAppManager
+
+        try:
+            return CliAppManager(workspace=self.workspace).installed_skill_aliases()
+        except OSError:
+            return {}
+
     def _skill_entries_from_dir(self, base: Path, source: str, *, skip_names: set[str] | None = None) -> list[dict[str, str]]:
         if not base.exists():
             return []
@@ -60,15 +95,33 @@ class SkillsLoader:
         Returns:
             List of skill info dicts with 'name', 'path', 'source'.
         """
+        from nanobot.agent.plugins import enabled_agent_plugin_skills
+
+        plugin_skills = enabled_agent_plugin_skills(self.workspace)
         skills = self._skill_entries_from_dir(self.workspace_skills, "workspace")
-        workspace_names = {entry["name"] for entry in skills}
+        seen_names = {entry["name"] for entry in skills}
+        for name, path in plugin_skills:
+            if name in seen_names:
+                continue
+            skills.append(
+                {
+                    "name": name,
+                    "path": str(path),
+                    "source": "plugin",
+                }
+            )
+            seen_names.add(name)
         if self.builtin_skills and self.builtin_skills.exists():
             skills.extend(
-                self._skill_entries_from_dir(self.builtin_skills, "builtin", skip_names=workspace_names)
+                self._skill_entries_from_dir(self.builtin_skills, "builtin", skip_names=seen_names)
             )
 
         if self.disabled_skills:
-            skills = [s for s in skills if s["name"] not in self.disabled_skills]
+            disabled = set(self.disabled_skills)
+            for legacy, canonical in self._skill_aliases().items():
+                if legacy in disabled or canonical in disabled:
+                    disabled.update((legacy, canonical))
+            skills = [s for s in skills if s["name"] not in disabled]
 
         if filter_unavailable:
             return [skill for skill in skills if self._check_requirements(self._get_skill_meta(skill["name"]))]
@@ -84,14 +137,11 @@ class SkillsLoader:
         Returns:
             Skill content or None if not found.
         """
-        roots = [self.workspace_skills]
-        if self.builtin_skills:
-            roots.append(self.builtin_skills)
-        for root in roots:
-            path = root / name / "SKILL.md"
-            if path.exists():
-                return path.read_text(encoding="utf-8")
-        return None
+        skills = self.list_skills(filter_unavailable=False)
+        available = {skill["name"] for skill in skills}
+        resolved = name if name in available else self._skill_aliases().get(name, name)
+        entry = next((skill for skill in skills if skill["name"] == resolved), None)
+        return Path(entry["path"]).read_text(encoding="utf-8") if entry else None
 
     def load_skills_for_context(self, skill_names: list[str]) -> str:
         """
@@ -118,9 +168,11 @@ class SkillsLoader:
             entry["name"]
             for entry in self.list_skills(filter_unavailable=True)
         }
+        aliases = self._skill_aliases()
         invoked: list[str] = []
         for match in _SKILL_REFERENCE.finditer(text):
-            name = match.group(1)
+            requested = match.group(1)
+            name = requested if requested in available else aliases.get(requested, requested)
             if name in available and name not in invoked:
                 invoked.append(name)
         return invoked
@@ -145,6 +197,7 @@ class SkillsLoader:
         sections: list[str] = []
         groups = (
             ("Workspace skills", "workspace", self.workspace_skills),
+            ("Agent Plugin skills", "plugin", self.workspace / "plugins"),
             ("Built-in skills", "builtin", self.builtin_skills),
         )
         for label, source, root in groups:
@@ -278,21 +331,4 @@ class SkillsLoader:
         Returns:
             Metadata dict or None.
         """
-        content = self.load_skill(name)
-        if not content or not content.startswith("---"):
-            return None
-        match = _STRIP_SKILL_FRONTMATTER.match(content)
-        if not match:
-            return None
-        try:
-            parsed = yaml.safe_load(match.group(1))
-        except yaml.YAMLError:
-            return None
-        if not isinstance(parsed, dict):
-            return None
-        # yaml.safe_load returns native types (int, bool, list, etc.);
-        # keep values as-is so downstream consumers get correct types.
-        metadata: dict[str, object] = {}
-        for key, value in cast(dict[object, object], parsed).items():
-            metadata[str(key)] = value
-        return metadata
+        return parse_skill_metadata(self.load_skill(name) or "")

nanobot/agent/subagent.py

@@ -5,6 +5,7 @@ import json
 import time
 import uuid
 import warnings
+from collections.abc import Mapping
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Callable, TypedDict
@@ -157,6 +158,10 @@ class SubagentManager:
         self._task_statuses: dict[str, SubagentStatus] = {}
         self._session_tasks: dict[str, set[str]] = {}  # session_key -> {task_id, ...}
 
+    def runtime_statuses(self) -> Mapping[str, SubagentStatus]:
+        """Return the observable task statuses used by runtime-control snapshots."""
+        return self._task_statuses
+
     def set_provider(self, provider: LLMProvider, model: str) -> None:
         """Update the deprecated runtime source used by legacy ``spawn`` calls."""
         warnings.warn(

nanobot/agent/tools/_windows_job.py

@@ -0,0 +1,218 @@
+"""Windows Job Object ownership for subprocess trees."""
+
+from __future__ import annotations
+
+import ctypes
+from ctypes import wintypes
+
+_CREATE_SUSPENDED = 0x00000004
+_PROCESS_SET_QUOTA = 0x0100
+_PROCESS_TERMINATE = 0x0001
+_TH32CS_SNAPTHREAD = 0x00000004
+_THREAD_SUSPEND_RESUME = 0x0002
+_JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE = 0x00002000
+_JOB_OBJECT_EXTENDED_LIMIT_INFORMATION = 9
+_INVALID_HANDLE_VALUE = ctypes.c_void_p(-1).value
+
+
+class _IoCounters(ctypes.Structure):
+    _fields_ = [
+        ("ReadOperationCount", ctypes.c_ulonglong),
+        ("WriteOperationCount", ctypes.c_ulonglong),
+        ("OtherOperationCount", ctypes.c_ulonglong),
+        ("ReadTransferCount", ctypes.c_ulonglong),
+        ("WriteTransferCount", ctypes.c_ulonglong),
+        ("OtherTransferCount", ctypes.c_ulonglong),
+    ]
+
+
+class _BasicLimitInformation(ctypes.Structure):
+    _fields_ = [
+        ("PerProcessUserTimeLimit", ctypes.c_longlong),
+        ("PerJobUserTimeLimit", ctypes.c_longlong),
+        ("LimitFlags", wintypes.DWORD),
+        ("MinimumWorkingSetSize", ctypes.c_size_t),
+        ("MaximumWorkingSetSize", ctypes.c_size_t),
+        ("ActiveProcessLimit", wintypes.DWORD),
+        ("Affinity", ctypes.c_size_t),
+        ("PriorityClass", wintypes.DWORD),
+        ("SchedulingClass", wintypes.DWORD),
+    ]
+
+
+class _ExtendedLimitInformation(ctypes.Structure):
+    _fields_ = [
+        ("BasicLimitInformation", _BasicLimitInformation),
+        ("IoInfo", _IoCounters),
+        ("ProcessMemoryLimit", ctypes.c_size_t),
+        ("JobMemoryLimit", ctypes.c_size_t),
+        ("PeakProcessMemoryUsed", ctypes.c_size_t),
+        ("PeakJobMemoryUsed", ctypes.c_size_t),
+    ]
+
+
+class _ThreadEntry32(ctypes.Structure):
+    _fields_ = [
+        ("dwSize", wintypes.DWORD),
+        ("cntUsage", wintypes.DWORD),
+        ("th32ThreadID", wintypes.DWORD),
+        ("th32OwnerProcessID", wintypes.DWORD),
+        ("tpBasePri", wintypes.LONG),
+        ("tpDeltaPri", wintypes.LONG),
+        ("dwFlags", wintypes.DWORD),
+    ]
+
+
+_kernel32 = ctypes.WinDLL("kernel32", use_last_error=True)
+_kernel32.CreateJobObjectW.argtypes = [ctypes.c_void_p, wintypes.LPCWSTR]
+_kernel32.CreateJobObjectW.restype = wintypes.HANDLE
+_kernel32.SetInformationJobObject.argtypes = [
+    wintypes.HANDLE,
+    ctypes.c_int,
+    ctypes.c_void_p,
+    wintypes.DWORD,
+]
+_kernel32.SetInformationJobObject.restype = wintypes.BOOL
+_kernel32.OpenProcess.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
+_kernel32.OpenProcess.restype = wintypes.HANDLE
+_kernel32.AssignProcessToJobObject.argtypes = [wintypes.HANDLE, wintypes.HANDLE]
+_kernel32.AssignProcessToJobObject.restype = wintypes.BOOL
+_kernel32.TerminateProcess.argtypes = [wintypes.HANDLE, wintypes.UINT]
+_kernel32.TerminateProcess.restype = wintypes.BOOL
+_kernel32.TerminateJobObject.argtypes = [wintypes.HANDLE, wintypes.UINT]
+_kernel32.TerminateJobObject.restype = wintypes.BOOL
+_kernel32.CreateToolhelp32Snapshot.argtypes = [wintypes.DWORD, wintypes.DWORD]
+_kernel32.CreateToolhelp32Snapshot.restype = wintypes.HANDLE
+_kernel32.Thread32First.argtypes = [wintypes.HANDLE, ctypes.POINTER(_ThreadEntry32)]
+_kernel32.Thread32First.restype = wintypes.BOOL
+_kernel32.Thread32Next.argtypes = [wintypes.HANDLE, ctypes.POINTER(_ThreadEntry32)]
+_kernel32.Thread32Next.restype = wintypes.BOOL
+_kernel32.OpenThread.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
+_kernel32.OpenThread.restype = wintypes.HANDLE
+_kernel32.ResumeThread.argtypes = [wintypes.HANDLE]
+_kernel32.ResumeThread.restype = wintypes.DWORD
+_kernel32.CloseHandle.argtypes = [wintypes.HANDLE]
+_kernel32.CloseHandle.restype = wintypes.BOOL
+
+
+def _win_error(operation: str) -> OSError:
+    code = ctypes.get_last_error()
+    return OSError(code, f"{operation} failed (Windows error {code})")
+
+
+def _close_handle(handle: int | None) -> None:
+    if handle:
+        _kernel32.CloseHandle(handle)
+
+
+def _set_kill_on_close(handle: int, enabled: bool) -> None:
+    info = _ExtendedLimitInformation()
+    if enabled:
+        info.BasicLimitInformation.LimitFlags = _JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE
+    if not _kernel32.SetInformationJobObject(
+        handle,
+        _JOB_OBJECT_EXTENDED_LIMIT_INFORMATION,
+        ctypes.byref(info),
+        ctypes.sizeof(info),
+    ):
+        raise _win_error("SetInformationJobObject")
+
+
+def _resume_primary_thread(pid: int) -> None:
+    snapshot = _kernel32.CreateToolhelp32Snapshot(_TH32CS_SNAPTHREAD, 0)
+    if snapshot == _INVALID_HANDLE_VALUE:
+        raise _win_error("CreateToolhelp32Snapshot")
+    try:
+        entry = _ThreadEntry32()
+        entry.dwSize = ctypes.sizeof(entry)
+        found = _kernel32.Thread32First(snapshot, ctypes.byref(entry))
+        while found:
+            if entry.th32OwnerProcessID == pid:
+                thread = _kernel32.OpenThread(
+                    _THREAD_SUSPEND_RESUME,
+                    False,
+                    entry.th32ThreadID,
+                )
+                if not thread:
+                    raise _win_error("OpenThread")
+                try:
+                    if _kernel32.ResumeThread(thread) == 0xFFFFFFFF:
+                        raise _win_error("ResumeThread")
+                    return
+                finally:
+                    _close_handle(thread)
+            found = _kernel32.Thread32Next(snapshot, ctypes.byref(entry))
+        raise RuntimeError(f"suspended process {pid} has no resumable thread")
+    finally:
+        _close_handle(snapshot)
+
+
+class WindowsJob:
+    """Own a process tree even after its root process exits."""
+
+    creation_flags = _CREATE_SUSPENDED
+
+    def __init__(self, handle: int) -> None:
+        self._handle: int | None = handle
+
+    @classmethod
+    def create(cls) -> WindowsJob:
+        handle = _kernel32.CreateJobObjectW(None, None)
+        if not handle:
+            raise _win_error("CreateJobObjectW")
+        try:
+            _set_kill_on_close(handle, True)
+        except Exception:
+            _close_handle(handle)
+            raise
+        return cls(handle)
+
+    def assign_and_resume(self, pid: int) -> None:
+        """Atomically establish tree ownership before the root can spawn."""
+        if self._handle is None:
+            raise RuntimeError("Windows job is already closed")
+        process = _kernel32.OpenProcess(
+            _PROCESS_SET_QUOTA | _PROCESS_TERMINATE,
+            False,
+            pid,
+        )
+        if not process:
+            error = _win_error("OpenProcess")
+            self.close()
+            raise error
+
+        if not _kernel32.AssignProcessToJobObject(self._handle, process):
+            error = _win_error("AssignProcessToJobObject")
+            _kernel32.TerminateProcess(process, 1)
+            _close_handle(process)
+            self.close()
+            raise error
+
+        try:
+            _resume_primary_thread(pid)
+        except Exception:
+            self.terminate()
+            raise
+        finally:
+            _close_handle(process)
+
+    def release(self) -> None:
+        """Release ownership after successful output collection."""
+        if self._handle is None:
+            return
+        _set_kill_on_close(self._handle, False)
+        self.close()
+
+    def terminate(self) -> None:
+        """Terminate every process in the job and close its handle."""
+        if self._handle is None:
+            return
+        try:
+            _kernel32.TerminateJobObject(self._handle, 1)
+        finally:
+            self.close()
+
+    def close(self) -> None:
+        handle = self._handle
+        self._handle = None
+        _close_handle(handle)

nanobot/agent/tools/base.py

@@ -1,6 +1,7 @@
 """Base class for agent tools."""
 from __future__ import annotations
 
+import math
 import typing
 from abc import ABC, abstractmethod
 from collections.abc import Callable
@@ -67,6 +68,8 @@ class Schema(ABC):
             return [f"{label} should be number"]
         if t in _JSON_TYPE_MAP and t not in ("integer", "number") and not isinstance(val, _JSON_TYPE_MAP[t]):
             return [f"{label} should be {t}"]
+        if t == "number" and isinstance(val, float) and not math.isfinite(val):
+            return [f"{label} must be finite"]
 
         errors: list[str] = []
         if "enum" in schema and val not in schema["enum"]:

nanobot/agent/tools/exec_session.py

@@ -209,7 +209,11 @@ class _ExecSession:
                     timeout=2.0,
                 )
             # Safety-net reap after normal exit.
-            from nanobot.agent.tools.shell import _reap_pid  # pyright: ignore[reportPrivateUsage]
+            from nanobot.agent.tools.shell import (  # pyright: ignore[reportPrivateUsage]
+                ExecTool,
+                _reap_pid,  # pyright: ignore[reportPrivateUsage]
+            )
+            ExecTool._release_process_tree(self.process)  # pyright: ignore[reportPrivateUsage]
             _reap_pid(self.process.pid)  # pyright: ignore[reportPrivateUsage]
         elif yield_time_ms > 0:
             await self._wait_for_buffered_output()

nanobot/agent/tools/file_state.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import hashlib
 import os
+from collections import OrderedDict
 from contextvars import ContextVar, Token
 from dataclasses import dataclass
 from pathlib import Path
@@ -135,21 +136,30 @@ class FileStates:
 
 
 class FileStateStore:
-    """Lookup table for per-session file read/write state."""
+    """Bounded lookup table for per-session file read/write state."""
 
-    __slots__ = ("_states_by_key",)
+    __slots__ = ("_max_sessions", "_states_by_key")
 
-    def __init__(self) -> None:
-        self._states_by_key: dict[str, FileStates] = {}
+    def __init__(self, *, max_sessions: int = 128) -> None:
+        if max_sessions <= 0:
+            raise ValueError("max_sessions must be positive")
+        self._max_sessions = max_sessions
+        self._states_by_key: OrderedDict[str, FileStates] = OrderedDict()
 
     def for_session(self, session_key: str | None) -> FileStates:
         key = session_key or "__default__"
-        states = self._states_by_key.get(key)
+        states = self._states_by_key.pop(key, None)
         if states is None:
             states = FileStates()
         self._states_by_key[key] = states
+        while len(self._states_by_key) > self._max_sessions:
+            self._states_by_key.popitem(last=False)
         return states
 
+    def discard(self, session_key: str | None) -> None:
+        """Forget file state when a session is reset or removed."""
+        self._states_by_key.pop(session_key or "__default__", None)
+
     def clear(self) -> None:
         self._states_by_key.clear()
 

nanobot/agent/tools/filesystem.py

@@ -148,9 +148,31 @@ class _FsTool(Tool):
         )
 
     def _resolve_read(self, path: str) -> Path:
+        plugin_skill_dirs: list[Path] = []
+        if self._workspace is not None:
+            from nanobot.agent.plugins import enabled_agent_plugin_skill_dirs
+
+            try:
+                access = current_tool_workspace(
+                    self._workspace,
+                    restrict_to_workspace=self._restrict_to_workspace,
+                    sandbox_restricts_workspace=self._sandbox_restricts_workspace,
+                )
+                if self._effective_allowed_root(access.allowed_root) is not None:
+                    candidate = Path(path).expanduser()
+                    if not candidate.is_absolute() and access.project_path is not None:
+                        candidate = access.project_path / candidate
+                    plugin_skill_dirs = list(
+                        enabled_agent_plugin_skill_dirs(
+                            Path(self._workspace),
+                            requested_path=candidate.resolve(strict=False),
+                        )
+                    )
+            except (OSError, RuntimeError):
+                pass
         return self._resolve_with_extra(
             path,
-            self._extra_read_allowed_dirs,
+            [*self._extra_read_allowed_dirs, *plugin_skill_dirs],
             self._extra_read_allowed_files,
             include_media_dir=True,
             extra_files_require_allowed_root=True,
@@ -827,7 +849,8 @@ class EditFileTool(_FsTool):
     def description(self) -> str:
         return (
             "Perform a small, exact replacement in one file by replacing "
-            "old_text with new_text. Use this for narrow text substitutions "
+            "old_text with new_text. When replacing text in an existing file, "
+            "old_text and new_text must be different. Use this for narrow text substitutions "
             "with old_text copied from read_file. For multi-file, structural, "
             "or generated code edits, prefer apply_patch. If old_text matches "
             "multiple times, provide more context or set occurrence, line_hint, "
@@ -862,9 +885,12 @@ class EditFileTool(_FsTool):
                 return ToolResult.error("Error: expected_replacements must be >= 1.")
 
             fp = self._resolve_write(path)
+            file_exists = fp.exists()
+            if file_exists and old_text == new_text:
+                return ToolResult.error("Error: new_text must be different from old_text.")
 
             # Create-file semantics: old_text='' + file doesn't exist → create
-            if not fp.exists():
+            if not file_exists:
                 if old_text == "":
                     fp.parent.mkdir(parents=True, exist_ok=True)
                     fp.write_text(new_text, encoding="utf-8")

nanobot/agent/tools/loader.py

@@ -19,7 +19,7 @@ if TYPE_CHECKING:
 
 _SKIP_MODULES = frozenset({
     "base", "schema", "registry", "context", "loader", "config",
-    "file_state", "sandbox", "mcp", "__init__", "runtime_state",
+    "file_state", "sandbox", "mcp", "__init__", "runtime_control",
 })
 
 

nanobot/agent/tools/mcp.py

@@ -1,4 +1,6 @@
-"""MCP client: connects to MCP servers and wraps their tools as native nanobot tools."""
+"""MCP client and dynamic tool-provider lifecycle."""
+
+from __future__ import annotations
 
 import asyncio
 import hashlib
@@ -7,23 +9,15 @@ import os
 import re
 import shutil
 import urllib.parse
-from collections.abc import AsyncIterator, Awaitable, Callable
+from collections.abc import AsyncIterator, Awaitable, Callable, Iterable, Mapping
 from contextlib import AsyncExitStack, suppress
-from typing import TYPE_CHECKING, Any, Mapping, Protocol, cast
-from weakref import WeakKeyDictionary
+from typing import TYPE_CHECKING, Any, Literal, Protocol, cast
 
 import httpx
 from loguru import logger
 
 from nanobot.agent.tools.base import Tool, ToolResult
 from nanobot.agent.tools.registry import ToolRegistry
-from nanobot.bus.events import (
-    INBOUND_META_RUNTIME_CONTROL,
-    RUNTIME_CONTROL_ACK,
-    RUNTIME_CONTROL_MCP_RELOAD,
-    InboundMessage,
-)
-from nanobot.bus.queue import MessageBus
 from nanobot.security.network import (
     PinnedDNSAsyncTransport,
     env_proxy_applies_to_url,
@@ -38,7 +32,8 @@ if TYPE_CHECKING:
     from mcp.types import Prompt, Resource
     from mcp.types import Tool as MCPToolDefinition
 
-    from nanobot.config.schema import MCPServerConfig
+    from nanobot.agent.tools.mcp_oauth import MCPOAuthHandlers
+    from nanobot.config.schema import Config, MCPServerConfig
 
 # Transient connection errors that warrant a single retry.
 # These typically happen when an MCP server restarts or a network
@@ -59,14 +54,37 @@ _WINDOWS_SHELL_LAUNCHERS: frozenset[str] = frozenset(("npx", "npm", "pnpm", "yar
 # Characters allowed in tool names by model providers (Anthropic, OpenAI, etc.).
 # Replace anything outside [a-zA-Z0-9_-] with underscore and collapse runs.
 _SANITIZE_RE = re.compile(r"_+")
-_RELOAD_LOCKS: WeakKeyDictionary[Any, asyncio.Lock] = WeakKeyDictionary()
 _ReconnectCallback = Callable[[str, str, Tool], Awaitable[Tool | None]]
+MCPServerLoader = Callable[[], Mapping[str, "MCPServerConfig"]]
+MCPRuntimeStatus = Literal["connecting", "connected", "failed"]
 
 
 class MCPConnection(Protocol):
     async def aclose(self) -> None: ...
 
 
+async def _close_mcp_connection(name: str, connection: MCPConnection) -> None:
+    try:
+        await connection.aclose()
+    except asyncio.CancelledError:
+        if task_is_cancelling():
+            raise
+        logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
+    except (RuntimeError, BaseExceptionGroup):
+        logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
+
+
+async def _close_mcp_connections(connections: Mapping[str, MCPConnection]) -> None:
+    cancellation: asyncio.CancelledError | None = None
+    for name, connection in connections.items():
+        try:
+            await _close_mcp_connection(name, connection)
+        except asyncio.CancelledError as exc:
+            cancellation = cancellation or exc
+    if cancellation is not None:
+        raise cancellation
+
+
 class _OwnedMCPConnection:
     """Close an MCP transport from the task that originally opened it."""
 
@@ -184,6 +202,25 @@ def _is_transient(exc: BaseException) -> bool:
     return type(exc).__name__ in _TRANSIENT_EXC_NAMES
 
 
+def _is_transient_connection_failure(exc: BaseException) -> bool:
+    if isinstance(exc, BaseExceptionGroup):
+        group = cast(BaseExceptionGroup[BaseException], exc)
+        return bool(group.exceptions) and all(
+            _is_transient_connection_failure(nested) for nested in group.exceptions
+        )
+    return isinstance(exc, (httpx.ConnectError, httpx.ConnectTimeout)) or _is_transient(exc)
+
+
+def _log_mcp_connection_failure(name: str, exc: BaseException, hint: str = "") -> None:
+    if _is_transient_connection_failure(exc):
+        logger.warning("MCP server '{}': transient connection failure", name)
+        logger.opt(exception=exc).debug(
+            "MCP server '{}' transient connection failure details", name
+        )
+        return
+    logger.opt(exception=exc).error("MCP server '{}': failed to connect: {}", name, hint)
+
+
 def _is_session_terminated(exc: BaseException) -> bool:
     """Return True when the MCP SDK reports a dead client session."""
     if _is_transient(exc):
@@ -468,11 +505,11 @@ class _MCPWrapperBase(Tool):
     """Common reconnect handling for wrappers bound to one MCP server session."""
 
     _plugin_discoverable = False
-    _session: "ClientSession"
+    _session: ClientSession
     _server_name: str
     _name: str
 
-    def _set_mcp_connection(self, session: "ClientSession", server_name: str) -> None:
+    def _set_mcp_connection(self, session: ClientSession, server_name: str) -> None:
         self._session = session
         self._server_name = server_name
         self._reconnect: _ReconnectCallback | None = None
@@ -562,9 +599,9 @@ class MCPToolWrapper(_MCPWrapperBase):
 
     def __init__(
         self,
-        session: "ClientSession",
+        session: ClientSession,
         server_name: str,
-        tool_def: "MCPToolDefinition",
+        tool_def: MCPToolDefinition,
         tool_timeout: int = 30,
     ):
         self._set_mcp_connection(session, server_name)
@@ -724,9 +761,9 @@ class MCPResourceWrapper(_MCPWrapperBase):
 
     def __init__(
         self,
-        session: "ClientSession",
+        session: ClientSession,
         server_name: str,
-        resource_def: "Resource",
+        resource_def: Resource,
         resource_timeout: int = 30,
     ):
         self._set_mcp_connection(session, server_name)
@@ -828,9 +865,9 @@ class MCPPromptWrapper(_MCPWrapperBase):
 
     def __init__(
         self,
-        session: "ClientSession",
+        session: ClientSession,
         server_name: str,
-        prompt_def: "Prompt",
+        prompt_def: Prompt,
         prompt_timeout: int = 30,
     ):
         self._set_mcp_connection(session, server_name)
@@ -961,7 +998,10 @@ class MCPPromptWrapper(_MCPWrapperBase):
 
 
 async def connect_mcp_servers(
-    mcp_servers: "dict[str, MCPServerConfig]", registry: ToolRegistry
+    mcp_servers: dict[str, MCPServerConfig],
+    registry: ToolRegistry,
+    *,
+    oauth_handlers: Mapping[str, MCPOAuthHandlers] | None = None,
 ) -> dict[str, MCPConnection]:
     """Connect to configured MCP servers and register their tools, resources, prompts.
 
@@ -975,11 +1015,8 @@ async def connect_mcp_servers(
     from mcp.client.streamable_http import streamable_http_client
 
     async def open_single_server(
-        name: str, cfg: "MCPServerConfig"
-    ) -> tuple[str, AsyncExitStack | None]:
-        server_stack = AsyncExitStack()
-        await server_stack.__aenter__()
-
+        name: str, cfg: MCPServerConfig, server_stack: AsyncExitStack
+    ) -> bool:
         try:
             transport_type = cfg.type
             if not transport_type:
@@ -991,8 +1028,7 @@ async def connect_mcp_servers(
                     )
                 else:
                     logger.warning("MCP server '{}': no command or url configured, skipping", name)
-                    await server_stack.aclose()
-                    return name, None
+                    return False
 
             if transport_type in {"sse", "streamableHttp"}:
                 ok, error = validate_url_target(cfg.url)
@@ -1003,8 +1039,30 @@ async def connect_mcp_servers(
                         _redact_url(cfg.url),
                         error,
                     )
-                    await server_stack.aclose()
-                    return name, None
+                    return False
+
+            oauth_auth: httpx.Auth | None = None
+            if cfg.auth == "oauth":
+                if transport_type not in {"sse", "streamableHttp"}:
+                    logger.warning(
+                        "MCP server '{}': OAuth requires an SSE or Streamable HTTP transport",
+                        name,
+                    )
+                    return False
+                from nanobot.agent.tools.mcp_oauth import (
+                    MCPAuthorizationRequiredError,
+                    create_mcp_oauth_auth,
+                )
+
+                try:
+                    oauth_auth = await create_mcp_oauth_auth(
+                        name,
+                        cfg.url,
+                        (oauth_handlers or {}).get(name),
+                    )
+                except MCPAuthorizationRequiredError:
+                    logger.info("MCP server '{}': waiting for browser authorization", name)
+                    return False
 
             if transport_type == "stdio":
                 command, args, env = _normalize_windows_stdio_command(
@@ -1022,8 +1080,7 @@ async def connect_mcp_servers(
             elif transport_type == "sse":
                 if not await _probe_http_url(cfg.url):
                     logger.warning("MCP server '{}': {} unreachable, skipping", name, _redact_url(cfg.url))
-                    await server_stack.aclose()
-                    return name, None
+                    return False
 
                 def httpx_client_factory(
                     headers: dict[str, str] | None = None,
@@ -1044,31 +1101,37 @@ async def connect_mcp_servers(
                         **_pinned_transport_kwargs(),
                     )
 
+                sse_kwargs: dict[str, Any] = {
+                    "httpx_client_factory": httpx_client_factory,
+                }
+                if oauth_auth is not None:
+                    sse_kwargs["auth"] = oauth_auth
                 read, write = await server_stack.enter_async_context(
-                    sse_client(cfg.url, httpx_client_factory=httpx_client_factory)
+                    sse_client(cfg.url, **sse_kwargs)
                 )
             elif transport_type == "streamableHttp":
                 if not await _probe_http_url(cfg.url):
                     logger.warning("MCP server '{}': {} unreachable, skipping", name, _redact_url(cfg.url))
-                    await server_stack.aclose()
-                    return name, None
+                    return False
 
-                http_client = await server_stack.enter_async_context(
-                    httpx.AsyncClient(
-                        headers=cfg.headers or None,
-                        event_hooks={"request": [_validate_mcp_request_url]},
-                        follow_redirects=True,
-                        timeout=httpx.Timeout(30.0, connect=10.0),
+                http_client_kwargs: dict[str, Any] = {
+                    "headers": cfg.headers or None,
+                    "event_hooks": {"request": [_validate_mcp_request_url]},
+                    "follow_redirects": True,
+                    "timeout": httpx.Timeout(30.0, connect=10.0),
                     **_pinned_transport_kwargs(),
-                    )
+                }
+                if oauth_auth is not None:
+                    http_client_kwargs["auth"] = oauth_auth
+                http_client = await server_stack.enter_async_context(
+                    httpx.AsyncClient(**http_client_kwargs)
                 )
                 read, write, _ = await server_stack.enter_async_context(
                     streamable_http_client(cfg.url, http_client=http_client)
                 )
             else:
                 logger.warning("MCP server '{}': unknown transport type '{}'", name, transport_type)
-                await server_stack.aclose()
-                return name, None
+                return False
 
             read = _filter_malformed_mcp_progress_notifications(read, name)
             session = await server_stack.enter_async_context(ClientSession(read, write))
@@ -1171,7 +1234,7 @@ async def connect_mcp_servers(
             logger.info(
                 "MCP server '{}': connected, {} capabilities registered", name, registered_count
             )
-            return name, server_stack
+            return True
 
         except Exception as e:
             hint = ""
@@ -1190,43 +1253,41 @@ async def connect_mcp_servers(
                     " Hint: this looks like stdio protocol pollution. Make sure the MCP server writes "
                     "only JSON-RPC to stdout and sends logs/debug output to stderr instead."
                 )
-            logger.exception("MCP server '{}': failed to connect: {}", name, hint)
-            with suppress(Exception):
-                await server_stack.aclose()
-            return name, None
+            _log_mcp_connection_failure(name, e, hint)
+            return False
 
     async def connect_single_server(
-        name: str, cfg: "MCPServerConfig"
+        name: str, cfg: MCPServerConfig
     ) -> tuple[str, MCPConnection | None]:
         loop = asyncio.get_running_loop()
         ready: asyncio.Future[bool] = loop.create_future()
         close_requested = asyncio.Event()
 
         async def own_connection() -> None:
-            stack: AsyncExitStack | None = None
             try:
-                _, stack = await open_single_server(name, cfg)
+                async with AsyncExitStack() as stack:
+                    connected = await open_single_server(name, cfg, stack)
                     if not ready.done():
-                    ready.set_result(stack is not None)
-                if stack is not None:
+                        ready.set_result(connected)
+                    if connected:
                         await close_requested.wait()
             except BaseException as exc:
                 if not ready.done():
                     ready.set_exception(exc)
                 raise
-            finally:
-                if stack is not None:
-                    await stack.aclose()
 
         owner = asyncio.create_task(own_connection(), name=f"mcp:{name}")
         connection = _OwnedMCPConnection(owner, close_requested)
         try:
             connected = await ready
-        except BaseException:
+        except BaseException as exc:
             close_requested.set()
             owner.cancel()
             with suppress(BaseException):
                 await asyncio.shield(owner)
+            if isinstance(exc, asyncio.CancelledError) and not task_is_cancelling():
+                logger.warning("MCP server '{}': connection cancelled by server/SDK", name)
+                return name, None
             raise
         if not connected:
             await connection.aclose()
@@ -1234,15 +1295,29 @@ async def connect_mcp_servers(
         return name, connection
 
     server_stacks: dict[str, MCPConnection] = {}
+    attempted_names: list[str] = []
 
+    try:
         for name, cfg in mcp_servers.items():
+            attempted_names.append(name)
             try:
                 result = await connect_single_server(name, cfg)
             except Exception as e:
-            logger.exception("MCP server '{}' connection failed: {}", name, e)
+                _log_mcp_connection_failure(name, e)
                 continue
             if result[1] is not None:
                 server_stacks[result[0]] = result[1]
+    except BaseException:
+        # Callers can bound readiness/reload with a timeout. If cancellation
+        # interrupts a later server, ownership of earlier connections has not
+        # transferred yet, so roll the whole batch back before propagating it.
+        for name in attempted_names:
+            _unregister_server_tools(registry, name)
+        try:
+            await _close_mcp_connections(server_stacks)
+        except BaseException as cleanup_exc:
+            logger.debug("MCP batch rollback cleanup error (can be ignored): {}", cleanup_exc)
+        raise
 
     return server_stacks
 
@@ -1253,53 +1328,158 @@ def session_extra(metadata: Mapping[str, Any] | None) -> dict[str, Any]:
     return {"mcp_presets": mcp_presets} if isinstance(mcp_presets, list) and mcp_presets else {}
 
 
-async def connect_missing_servers(state: Any, registry: ToolRegistry) -> None:
-    """Connect configured MCP servers that are not currently live."""
-    async with _reload_lock(state):
-        if getattr(state, "_mcp_closing", False):
-            return
-        missing_servers = {
-            name: cfg for name, cfg in state._mcp_servers.items() if name not in state._mcp_stacks
+def _configured_servers(config: Config) -> dict[str, MCPServerConfig]:
+    from nanobot.agent.plugins import agent_plugin_mcp_servers
+
+    return agent_plugin_mcp_servers(
+        config.workspace_path,
+        config.tools.mcp_servers,
+    )
+
+
+def _load_current_servers() -> dict[str, MCPServerConfig]:
+    from nanobot.config.loader import load_config, resolve_config_env_vars
+
+    return _configured_servers(resolve_config_env_vars(load_config()))
+
+
+class MCPProvider:
+    """Own configured MCP connections and their dynamic tool registrations."""
+
+    def __init__(
+        self,
+        servers: Mapping[str, MCPServerConfig],
+        registry: ToolRegistry,
+        *,
+        server_loader: MCPServerLoader | None = None,
+    ) -> None:
+        self._servers = dict(servers)
+        self._registry = registry
+        self._server_loader = server_loader or _load_current_servers
+        self._connections: dict[str, MCPConnection] = {}
+        self._runtime_statuses: dict[str, MCPRuntimeStatus] = {}
+        self._lock = asyncio.Lock()
+        self._closing = False
+
+    @classmethod
+    def from_config(
+        cls,
+        config: Config,
+        registry: ToolRegistry,
+        *,
+        server_loader: MCPServerLoader | None = None,
+    ) -> MCPProvider:
+        return cls(
+            _configured_servers(config),
+            registry,
+            server_loader=server_loader,
+        )
+
+    @property
+    def configured_server_names(self) -> set[str]:
+        return set(self._servers)
+
+    @property
+    def connected_server_names(self) -> set[str]:
+        return set(self._connections)
+
+    def runtime_status(self) -> dict[str, MCPRuntimeStatus]:
+        """Return the latest connection-attempt result for configured servers."""
+        return {
+            name: status
+            for name, status in self._runtime_statuses.items()
+            if name in self._servers
         }
-        if state._mcp_connecting or not missing_servers:
+
+    def _set_runtime_status(
+        self,
+        server_names: Iterable[str],
+        status: MCPRuntimeStatus,
+    ) -> None:
+        for name in server_names:
+            self._runtime_statuses[name] = status
+
+    def _record_connection_result(
+        self,
+        attempted: Iterable[str],
+        connected: Iterable[str],
+    ) -> None:
+        attempted_names = set(attempted)
+        connected_names = set(connected)
+        self._set_runtime_status(connected_names, "connected")
+        self._set_runtime_status(attempted_names - connected_names, "failed")
+
+    async def connect(self) -> None:
+        """Connect configured servers that are not currently live."""
+        async with self._lock:
+            if self._closing:
                 return
-        state._mcp_connecting = True
+            configured_missing = {
+                name: cfg
+                for name, cfg in self._servers.items()
+                if name not in self._connections
+            }
+            oauth_servers = {
+                name: cfg
+                for name, cfg in configured_missing.items()
+                if cfg.auth == "oauth"
+            }
+            authorization_pending: set[str] = set()
+            if oauth_servers:
+                from nanobot.agent.tools.mcp_oauth import mcp_oauth_has_credentials
+
+                authorization_pending = {
+                    name
+                    for name, cfg in oauth_servers.items()
+                    if not mcp_oauth_has_credentials(name, cfg.url)
+                }
+            for name in authorization_pending:
+                self._runtime_statuses.pop(name, None)
+            missing_servers = {
+                name: cfg
+                for name, cfg in configured_missing.items()
+                if name not in authorization_pending
+            }
+            if not missing_servers:
+                return
+            self._set_runtime_status(missing_servers, "connecting")
             try:
-            connected = await connect_mcp_servers(missing_servers, registry)
-            if getattr(state, "_mcp_closing", False):
-                for connection in connected.values():
-                    await connection.aclose()
+                connected = await connect_mcp_servers(missing_servers, self._registry)
+                if self._closing:
+                    await _close_mcp_connections(connected)
                     return
-            state._mcp_stacks.update(connected)
-            _attach_reconnect_handlers(state, registry, connected)
+                self._connections.update(connected)
+                self._record_connection_result(missing_servers, connected)
+                self._attach_reconnect_handlers(connected)
                 if connected:
                     logger.info("MCP connected servers: {}", sorted(connected))
                 else:
-                logger.warning("No MCP servers connected successfully (will retry next message)")
+                    logger.warning(
+                        "No MCP servers connected successfully "
+                        "(will retry on the next readiness check)"
+                    )
             except asyncio.CancelledError:
+                self._set_runtime_status(missing_servers, "failed")
                 if task_is_cancelling():
                     raise
-            logger.warning("MCP connection cancelled (will retry next message)")
-        except BaseException as e:
-            logger.warning("Failed to connect MCP servers (will retry next message): {}", e)
-        finally:
-            state._mcp_connecting = False
+                logger.warning(
+                    "MCP connection cancelled (will retry on the next readiness check)"
+                )
+            except BaseException as exc:
+                self._set_runtime_status(missing_servers, "failed")
+                logger.warning(
+                    "Failed to connect MCP servers "
+                    "(will retry on the next readiness check): {}",
+                    exc,
+                )
 
-
-async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
-    """Reconcile live MCP connections with the current config file."""
-    async with _reload_lock(state):
-        if getattr(state, "_mcp_closing", False):
-            return {
-                "ok": False,
-                "message": "MCP connections are shutting down.",
-                "requires_restart": True,
-            }
+    async def reload(self) -> dict[str, Any]:
+        """Reconcile live MCP connections with the current configuration."""
+        async with self._lock:
+            if self._closing:
+                return self._closing_result()
             try:
-            from nanobot.config.loader import load_config, resolve_config_env_vars
-
-            config = resolve_config_env_vars(load_config())
-            next_servers = dict(config.tools.mcp_servers)
+                next_servers = dict(self._server_loader())
             except Exception as exc:
                 logger.warning("MCP hot reload could not read config: {}", exc)
                 return {
@@ -1309,49 +1489,69 @@ async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
                     "error": str(exc),
                 }
 
-        current_servers = dict(state._mcp_servers)
+            current_servers = dict(self._servers)
             current_names = set(current_servers)
             next_names = set(next_servers)
+            from nanobot.agent.tools.mcp_oauth import mcp_oauth_has_credentials
+
+            authorization_pending = {
+                name
+                for name, cfg in next_servers.items()
+                if cfg.auth == "oauth" and not mcp_oauth_has_credentials(name, cfg.url)
+            }
             removed = sorted(current_names - next_names)
             added = sorted(next_names - current_names)
             changed = sorted(
                 name
                 for name in current_names & next_names
-            if _server_signature(current_servers[name]) != _server_signature(next_servers[name])
+                if _server_signature(current_servers[name])
+                != _server_signature(next_servers[name])
             )
 
             tools_removed = 0
             for name in [*removed, *changed]:
-            tools_removed += _unregister_server_tools(registry, name)
-            await _close_server(state, name)
+                tools_removed += _unregister_server_tools(self._registry, name)
+                await self._close_server(name)
 
-        state._mcp_servers = next_servers
+            for name in [*removed, *authorization_pending]:
+                self._runtime_statuses.pop(name, None)
+
+            self._servers = next_servers
             retry_missing = sorted(
                 name
                 for name in next_names
-            if name not in state._mcp_stacks and name not in set(added) | set(changed)
+                if name not in self._connections
+                and name not in set(added) | set(changed)
+                and name not in authorization_pending
+            )
+            to_connect_names = sorted(
+                (set(added) | set(changed) | set(retry_missing))
+                - authorization_pending
             )
-        to_connect_names = sorted(set(added) | set(changed) | set(retry_missing))
             to_connect = {name: next_servers[name] for name in to_connect_names}
             connected: dict[str, MCPConnection] = {}
             if to_connect:
-            connected = await connect_mcp_servers(to_connect, registry)
-            if getattr(state, "_mcp_closing", False):
-                for connection in connected.values():
-                    await connection.aclose()
-                return {
-                    "ok": False,
-                    "message": "MCP connections are shutting down.",
-                    "requires_restart": True,
-                }
-            state._mcp_stacks.update(connected)
-            _attach_reconnect_handlers(state, registry, connected)
+                self._set_runtime_status(to_connect, "connecting")
+                try:
+                    connected = await connect_mcp_servers(to_connect, self._registry)
+                except BaseException:
+                    self._set_runtime_status(to_connect, "failed")
+                    raise
+                if self._closing:
+                    await _close_mcp_connections(connected)
+                    return self._closing_result()
+                self._connections.update(connected)
+                self._record_connection_result(to_connect, connected)
+                self._attach_reconnect_handlers(connected)
 
             failed = sorted(set(to_connect) - set(connected))
             unchanged = not removed and not added and not changed and not retry_missing
             ok = not failed
             if failed:
-            message = "MCP config reloaded, but some servers did not connect: " + ", ".join(failed)
+                message = (
+                    "MCP config reloaded, but some servers did not connect: "
+                    + ", ".join(failed)
+                )
             elif unchanged:
                 message = "MCP config is already live."
             elif retry_missing and not added and not changed and not removed:
@@ -1360,7 +1560,8 @@ async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
                 message = "MCP config reloaded without restarting nanobot."
 
             logger.info(
-            "MCP hot reload: added={} changed={} removed={} retried={} connected={} failed={} tools_removed={}",
+                "MCP hot reload: added={} changed={} removed={} retried={} "
+                "connected={} failed={} tools_removed={}",
                 added,
                 changed,
                 removed,
@@ -1376,114 +1577,51 @@ async def reload_servers(state: Any, registry: ToolRegistry) -> dict[str, Any]:
                 "changed": changed,
                 "removed": removed,
                 "retried": retry_missing,
-            "connected": sorted(state._mcp_stacks),
-            "configured": sorted(state._mcp_servers),
+                "connected": sorted(self._connections),
+                "configured": sorted(self._servers),
                 "failed": failed,
                 "tools_removed": tools_removed,
                 "requires_restart": False,
             }
 
-
-async def request_mcp_reload(
-    bus: MessageBus,
-    *,
-    timeout: float = 15.0,
-) -> dict[str, Any]:
-    """Ask the running agent loop to reconcile live MCP connections."""
-    loop = asyncio.get_running_loop()
-    ack: asyncio.Future[dict[str, Any]] = loop.create_future()
-    await bus.publish_inbound(
-        InboundMessage(
-            channel="system",
-            sender_id="webui-settings",
-            chat_id="runtime",
-            content=RUNTIME_CONTROL_MCP_RELOAD,
-            metadata={
-                INBOUND_META_RUNTIME_CONTROL: RUNTIME_CONTROL_MCP_RELOAD,
-                RUNTIME_CONTROL_ACK: ack,
-            },
-        )
-    )
-    try:
-        result = await asyncio.wait_for(ack, timeout=timeout)
-    except asyncio.TimeoutError:
+    @staticmethod
+    def _closing_result() -> dict[str, Any]:
         return {
             "ok": False,
-            "message": "MCP hot reload timed out. Restart nanobot to pick up changes.",
-            "requires_restart": True,
-        }
-    return result if isinstance(cast(object, result), dict) else {
-        "ok": False,
-        "message": "MCP hot reload returned an unexpected response.",
+            "message": "MCP connections are shutting down.",
             "requires_restart": True,
         }
 
-
-async def handle_runtime_control(state: Any, msg: InboundMessage, registry: ToolRegistry) -> bool:
-    metadata = msg.metadata if isinstance(cast(object, msg.metadata), dict) else {}
-    control = metadata.get(INBOUND_META_RUNTIME_CONTROL)
-    if control != RUNTIME_CONTROL_MCP_RELOAD:
-        return False
-
-    ack = metadata.get(RUNTIME_CONTROL_ACK)
-    try:
-        result = await reload_servers(state, registry)
-    except Exception as exc:
-        logger.exception("MCP hot reload failed")
-        result = {
-            "ok": False,
-            "message": "MCP hot reload failed. Restart nanobot to pick up changes.",
-            "requires_restart": True,
-            "error": str(exc),
-        }
-    if isinstance(ack, asyncio.Future) and not ack.done():
-        cast(asyncio.Future[dict[str, Any]], ack).set_result(result)
-    return True
-
-
-def _reload_lock(state: Any) -> asyncio.Lock:
-    try:
-        return _RELOAD_LOCKS[state]
-    except KeyError:
-        lock = asyncio.Lock()
-        _RELOAD_LOCKS[state] = lock
-        return lock
-
-
-def _attach_reconnect_handlers(
-    state: Any,
-    registry: ToolRegistry,
-    server_names: Mapping[str, Any] | set[str] | list[str] | tuple[str, ...],
-) -> None:
-    async def reconnect(server_name: str, tool_name: str, stale_tool: Tool) -> Tool | None:
-        return await _refresh_terminated_server(
-            state,
-            registry,
+    def _attach_reconnect_handlers(self, server_names: Iterable[str]) -> None:
+        async def reconnect(
+            server_name: str,
+            tool_name: str,
+            stale_tool: Tool,
+        ) -> Tool | None:
+            return await self._refresh_terminated_server(
                 server_name,
                 tool_name,
                 stale_tool,
             )
 
         for server_name in server_names:
-        for tool_name in list(registry.tool_names):
-            tool = registry.get(tool_name)
+            for tool_name in list(self._registry.tool_names):
+                tool = self._registry.get(tool_name)
                 if not _tool_belongs_to_server(tool, tool_name, server_name):
                     continue
                 if isinstance(tool, _MCPWrapperBase):
                     tool.set_reconnect_handler(reconnect)
 
-
-async def _refresh_terminated_server(
-    state: Any,
-    registry: ToolRegistry,
+    async def _refresh_terminated_server(
+        self,
         server_name: str,
         tool_name: str,
         stale_tool: Tool,
-) -> Tool | None:
-    async with _reload_lock(state):
-        if getattr(state, "_mcp_closing", False):
+    ) -> Tool | None:
+        async with self._lock:
+            if self._closing:
                 return None
-        cfg = state._mcp_servers.get(server_name)
+            cfg = self._servers.get(server_name)
             if cfg is None:
                 logger.warning(
                     "MCP server '{}' session terminated but is no longer configured",
@@ -1491,29 +1629,56 @@ async def _refresh_terminated_server(
                 )
                 return None
 
-        current_tool = registry.get(tool_name)
+            current_tool = self._registry.get(tool_name)
             if (
                 current_tool is not None
                 and current_tool is not stale_tool
-            and server_name in state._mcp_stacks
+                and server_name in self._connections
             ):
                 return current_tool
 
-        logger.warning("MCP server '{}' session terminated; refreshing connection", server_name)
-        _unregister_server_tools(registry, server_name)
-        await _close_server(state, server_name)
+            logger.warning(
+                "MCP server '{}' session terminated; refreshing connection",
+                server_name,
+            )
+            _unregister_server_tools(self._registry, server_name)
+            await self._close_server(server_name)
 
-        connected = await connect_mcp_servers({server_name: cfg}, registry)
-        if getattr(state, "_mcp_closing", False):
-            for connection in connected.values():
-                await connection.aclose()
+            self._set_runtime_status({server_name}, "connecting")
+            connected = await connect_mcp_servers(
+                {server_name: cfg},
+                self._registry,
+            )
+            if self._closing:
+                await _close_mcp_connections(connected)
                 return None
-        state._mcp_stacks.update(connected)
-        _attach_reconnect_handlers(state, registry, connected)
+            self._connections.update(connected)
+            self._record_connection_result({server_name}, connected)
+            self._attach_reconnect_handlers(connected)
             if server_name not in connected:
-            logger.warning("MCP server '{}' reconnect failed after session termination", server_name)
+                logger.warning(
+                    "MCP server '{}' reconnect failed after session termination",
+                    server_name,
+                )
                 return None
-        return registry.get(tool_name)
+            return self._registry.get(tool_name)
+
+    async def _close_server(self, server_name: str) -> None:
+        connection = self._connections.pop(server_name, None)
+        if connection is None:
+            return
+        await _close_mcp_connection(server_name, connection)
+
+    async def aclose(self) -> None:
+        """Close every connection while excluding reconnect and hot reload."""
+        self._closing = True
+        async with self._lock:
+            connections = dict(self._connections)
+            self._connections.clear()
+            self._runtime_statuses.clear()
+            for name in self._servers:
+                _unregister_server_tools(self._registry, name)
+            await _close_mcp_connections(connections)
 
 
 def _server_signature(cfg: Any) -> Any:
@@ -1540,34 +1705,3 @@ def _unregister_server_tools(registry: ToolRegistry, server_name: str) -> int:
             registry.unregister(tool_name)
             removed += 1
     return removed
-
-
-async def _close_server(state: Any, server_name: str) -> None:
-    stack = state._mcp_stacks.pop(server_name, None)
-    if stack is None:
-        return
-    try:
-        await stack.aclose()
-    except asyncio.CancelledError:
-        if task_is_cancelling():
-            raise
-        logger.debug("MCP server '{}' cleanup error (can be ignored)", server_name)
-    except (RuntimeError, BaseExceptionGroup):
-        logger.debug("MCP server '{}' cleanup error (can be ignored)", server_name)
-
-
-async def close_mcp_servers(state: Any) -> None:
-    """Close every MCP connection while excluding reconnect and hot reload."""
-    state._mcp_closing = True
-    async with _reload_lock(state):
-        connections = list(state._mcp_stacks.items())
-        state._mcp_stacks.clear()
-        for name, connection in connections:
-            try:
-                await connection.aclose()
-            except asyncio.CancelledError:
-                if task_is_cancelling():
-                    raise
-                logger.debug("MCP server '{}' cleanup error (can be ignored)", name)
-            except (RuntimeError, BaseExceptionGroup):
-                logger.debug("MCP server '{}' cleanup error (can be ignored)", name)

nanobot/agent/tools/mcp_oauth.py

@@ -0,0 +1,401 @@
+"""OAuth support for remote MCP servers.
+
+This module intentionally owns MCP OAuth end to end. Provider OAuth has a
+different lifecycle and storage contract, so sharing a higher-level workflow
+would couple unrelated extension boundaries.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import json
+import os
+import secrets
+from collections.abc import Awaitable, Callable
+from contextlib import suppress
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, TypedDict, cast
+
+from filelock import FileLock
+from loguru import logger
+from mcp.client.auth import OAuthClientProvider
+from mcp.shared.auth import OAuthClientInformationFull, OAuthClientMetadata, OAuthToken
+from pydantic import AnyHttpUrl, AnyUrl
+
+from nanobot.config.paths import get_data_dir
+from nanobot.utils.helpers import _write_text_atomic  # pyright: ignore[reportPrivateUsage]
+
+MCP_OAUTH_CALLBACK_PATH = "/auth/mcp/callback"
+_STORE_VERSION = 1
+_STORE_LOCK_TIMEOUT_S = 15
+_DEFAULT_REDIRECT_URI = f"http://127.0.0.1{MCP_OAUTH_CALLBACK_PATH}"
+_CLIENT_URI = AnyHttpUrl("https://github.com/HKUDS/nanobot")
+_LOGO_URI = AnyHttpUrl(
+    "https://raw.githubusercontent.com/HKUDS/nanobot/main/"
+    "webui/public/brand/nanobot_apple_touch.png"
+)
+
+
+class _StoredServer(TypedDict, total=False):
+    server_fingerprint: str
+    write_lease: str
+    tokens: dict[str, Any]
+    client_info: dict[str, Any]
+    redirect_uri: str
+
+
+class _CredentialStore(TypedDict):
+    version: int
+    servers: dict[str, _StoredServer]
+    generations: dict[str, str]
+
+
+class MCPAuthorizationRequiredError(RuntimeError):
+    """Raised when a background MCP connection needs interactive authorization."""
+
+
+@dataclass(frozen=True)
+class MCPOAuthHandlers:
+    """Browser callbacks supplied only for a user-initiated OAuth attempt."""
+
+    redirect_uri: str
+    redirect_handler: Callable[[str], Awaitable[None]]
+    callback_handler: Callable[[], Awaitable[tuple[str, str | None]]]
+    reset_credentials: bool = False
+
+
+def _store_path() -> Path:
+    return get_data_dir() / "auth" / "mcp.json"
+
+
+def _server_fingerprint(server_url: str) -> str:
+    return hashlib.sha256(server_url.strip().encode("utf-8")).hexdigest()
+
+
+def _empty_store() -> _CredentialStore:
+    return {"version": _STORE_VERSION, "servers": {}, "generations": {}}
+
+
+def _stored_server(value: object) -> _StoredServer | None:
+    if not isinstance(value, dict):
+        return None
+    raw = cast(dict[object, object], value)
+    entry: _StoredServer = {}
+    fingerprint = raw.get("server_fingerprint")
+    if isinstance(fingerprint, str):
+        entry["server_fingerprint"] = fingerprint
+    write_lease = raw.get("write_lease")
+    if isinstance(write_lease, str) and write_lease:
+        entry["write_lease"] = write_lease
+    redirect_uri = raw.get("redirect_uri")
+    if isinstance(redirect_uri, str):
+        entry["redirect_uri"] = redirect_uri
+    tokens = raw.get("tokens")
+    if isinstance(tokens, dict):
+        token_values = cast(dict[object, object], tokens)
+        if all(isinstance(key, str) for key in token_values):
+            entry["tokens"] = cast(dict[str, Any], token_values)
+    client_info = raw.get("client_info")
+    if isinstance(client_info, dict):
+        client_values = cast(dict[object, object], client_info)
+        if all(isinstance(key, str) for key in client_values):
+            entry["client_info"] = cast(dict[str, Any], client_values)
+    return entry
+
+
+def _read_store_unlocked(path: Path) -> _CredentialStore:
+    try:
+        raw = cast(object, json.loads(path.read_text(encoding="utf-8")))
+    except FileNotFoundError:
+        return _empty_store()
+    except (OSError, ValueError, TypeError) as exc:
+        logger.warning("Could not read MCP OAuth credentials: {}", type(exc).__name__)
+        return _empty_store()
+    if not isinstance(raw, dict):
+        return _empty_store()
+    payload = cast(dict[object, object], raw)
+    raw_servers = payload.get("servers")
+    if not isinstance(raw_servers, dict):
+        return _empty_store()
+    servers: dict[str, _StoredServer] = {}
+    for name, value in cast(dict[object, object], raw_servers).items():
+        entry = _stored_server(value)
+        if isinstance(name, str) and entry is not None:
+            servers[name] = entry
+    generations: dict[str, str] = {}
+    raw_generations = payload.get("generations")
+    if isinstance(raw_generations, dict):
+        for name, value in cast(dict[object, object], raw_generations).items():
+            if isinstance(name, str) and isinstance(value, str) and value:
+                generations[name] = value
+    return {
+        "version": _STORE_VERSION,
+        "servers": servers,
+        "generations": generations,
+    }
+
+
+def _with_store_lock(path: Path) -> FileLock:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    return FileLock(str(path.with_suffix(".lock")), timeout=_STORE_LOCK_TIMEOUT_S)
+
+
+def _write_store_unlocked(path: Path, payload: _CredentialStore) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with suppress(OSError):
+        os.chmod(path.parent, 0o700)
+    _write_text_atomic(path, json.dumps(payload, indent=2, ensure_ascii=False))
+    with suppress(OSError):
+        os.chmod(path, 0o600)
+
+
+class MCPOAuthStorage:
+    """Persistent MCP SDK token storage, isolated by config name and server URL."""
+
+    def __init__(self, server_name: str, server_url: str) -> None:
+        self.server_name = server_name
+        self.server_fingerprint = _server_fingerprint(server_url)
+        self._observed_generation = self._read_generation_sync()
+        self._write_lease: str | None = None
+
+    def _read_generation_sync(self) -> str | None:
+        path = _store_path()
+        if not path.exists():
+            return None
+        # Writes replace the whole file atomically, so this observes either side
+        # of a concurrent deletion without blocking the async connection path.
+        return _read_store_unlocked(path)["generations"].get(self.server_name)
+
+    def _generation_is_current(self, payload: _CredentialStore) -> bool:
+        return payload["generations"].get(self.server_name) == self._observed_generation
+
+    def _entry_unlocked(self, payload: _CredentialStore) -> _StoredServer | None:
+        servers = payload["servers"]
+        entry = servers.get(self.server_name)
+        if entry is None or entry.get("server_fingerprint") != self.server_fingerprint:
+            return None
+        return entry
+
+    def _bind_entry_unlocked(
+        self,
+        payload: _CredentialStore,
+        *,
+        create: bool,
+    ) -> tuple[_StoredServer | None, bool]:
+        if not self._generation_is_current(payload):
+            return None, False
+        entry = self._entry_unlocked(payload)
+        if self._write_lease is not None:
+            if entry is None or entry.get("write_lease") != self._write_lease:
+                return None, False
+            return entry, False
+        if entry is None:
+            if not create:
+                return None, False
+            self._write_lease = secrets.token_urlsafe(24)
+            entry = _StoredServer(
+                server_fingerprint=self.server_fingerprint,
+                write_lease=self._write_lease,
+            )
+            payload["servers"][self.server_name] = entry
+            return entry, True
+        write_lease = entry.get("write_lease")
+        changed = not isinstance(write_lease, str) or not write_lease
+        if changed:
+            write_lease = secrets.token_urlsafe(24)
+            entry["write_lease"] = write_lease
+        self._write_lease = write_lease
+        return entry, changed
+
+    def _read_entry_sync(self) -> _StoredServer | None:
+        path = _store_path()
+        with _with_store_lock(path):
+            payload = _read_store_unlocked(path)
+            entry, changed = self._bind_entry_unlocked(payload, create=False)
+            if changed:
+                _write_store_unlocked(path, payload)
+            return entry
+
+    def _update_entry_sync(
+        self,
+        update: Callable[[_StoredServer], None],
+        *,
+        create: bool = True,
+        claim: bool = False,
+    ) -> bool:
+        path = _store_path()
+        with _with_store_lock(path):
+            payload = _read_store_unlocked(path)
+            if claim:
+                # A browser flow owns subsequent SDK writes until another flow
+                # claims the entry or the configured server is removed.
+                if not self._generation_is_current(payload):
+                    logger.info(
+                        "Ignored stale MCP OAuth credential claim for '{}'",
+                        self.server_name,
+                    )
+                    return False
+                entry = self._entry_unlocked(payload)
+                if entry is None:
+                    entry = _StoredServer(server_fingerprint=self.server_fingerprint)
+                    payload["servers"][self.server_name] = entry
+                self._write_lease = secrets.token_urlsafe(24)
+                entry["write_lease"] = self._write_lease
+            else:
+                entry, _ = self._bind_entry_unlocked(payload, create=create)
+                if entry is None:
+                    if self._write_lease is not None:
+                        logger.info(
+                            "Ignored stale MCP OAuth credential update for '{}'",
+                            self.server_name,
+                        )
+                    return False
+            update(entry)
+            payload["version"] = _STORE_VERSION
+            _write_store_unlocked(path, payload)
+            return True
+
+    async def get_tokens(self) -> OAuthToken | None:
+        entry = await asyncio.to_thread(self._read_entry_sync)
+        raw = entry.get("tokens") if entry is not None else None
+        if not isinstance(raw, dict):
+            return None
+        try:
+            return OAuthToken.model_validate(raw)
+        except (ValueError, TypeError):
+            logger.warning("Ignoring invalid MCP OAuth tokens for '{}'", self.server_name)
+            return None
+
+    async def set_tokens(self, tokens: OAuthToken) -> None:
+        raw = tokens.model_dump(mode="json", exclude_none=True)
+
+        def update(entry: _StoredServer) -> None:
+            entry["tokens"] = raw
+
+        await asyncio.to_thread(self._update_entry_sync, update)
+
+    async def clear_tokens(self) -> None:
+        def update(entry: _StoredServer) -> None:
+            entry.pop("tokens", None)
+
+        await asyncio.to_thread(self._update_entry_sync, update, create=False)
+
+    async def get_client_info(self) -> OAuthClientInformationFull | None:
+        entry = await asyncio.to_thread(self._read_entry_sync)
+        raw = entry.get("client_info") if entry is not None else None
+        if not isinstance(raw, dict):
+            return None
+        try:
+            return OAuthClientInformationFull.model_validate(raw)
+        except (ValueError, TypeError):
+            logger.warning("Ignoring invalid MCP OAuth client info for '{}'", self.server_name)
+            return None
+
+    async def set_client_info(self, client_info: OAuthClientInformationFull) -> None:
+        raw = client_info.model_dump(mode="json", exclude_none=True)
+
+        def update(entry: _StoredServer) -> None:
+            entry["client_info"] = raw
+
+        await asyncio.to_thread(self._update_entry_sync, update)
+
+    async def redirect_uri(self) -> str | None:
+        entry = await asyncio.to_thread(self._read_entry_sync)
+        value = entry.get("redirect_uri") if entry is not None else None
+        return value if isinstance(value, str) and value else None
+
+    async def prepare_redirect_uri(self, redirect_uri: str, *, reset: bool = False) -> None:
+        def update(entry: _StoredServer) -> None:
+            changed = entry.get("redirect_uri") != redirect_uri
+            if reset:
+                entry.pop("tokens", None)
+                entry.pop("client_info", None)
+            elif changed:
+                # Dynamic registrations bind a client to its redirect URI.
+                entry.pop("client_info", None)
+            entry["redirect_uri"] = redirect_uri
+
+        claimed = await asyncio.to_thread(self._update_entry_sync, update, claim=True)
+        if not claimed:
+            raise MCPAuthorizationRequiredError("MCP authorization was cancelled")
+
+    def has_credentials(self) -> bool:
+        entry = self._read_entry_sync()
+        raw_tokens = entry.get("tokens") if entry is not None else None
+        if not isinstance(raw_tokens, dict):
+            return False
+        tokens = cast(dict[str, object], raw_tokens)
+        access_token = tokens.get("access_token")
+        return isinstance(access_token, str) and bool(access_token)
+
+
+async def _missing_callback() -> tuple[str, str | None]:
+    raise MCPAuthorizationRequiredError("MCP server requires browser authorization")
+
+
+async def create_mcp_oauth_auth(
+    server_name: str,
+    server_url: str,
+    handlers: MCPOAuthHandlers | None = None,
+) -> OAuthClientProvider:
+    """Build the official MCP SDK OAuth provider for one configured server."""
+    storage = MCPOAuthStorage(server_name, server_url)
+    if handlers is not None:
+        await storage.prepare_redirect_uri(
+            handlers.redirect_uri,
+            reset=handlers.reset_credentials,
+        )
+        redirect_uri = handlers.redirect_uri
+        redirect_handler = handlers.redirect_handler
+        callback_handler = handlers.callback_handler
+    else:
+        if not await asyncio.to_thread(storage.has_credentials):
+            # Do not perform discovery or dynamic registration from a background
+            # startup. Interactive OAuth begins only after an explicit user action.
+            raise MCPAuthorizationRequiredError("MCP server requires browser authorization")
+        redirect_uri = await storage.redirect_uri() or _DEFAULT_REDIRECT_URI
+
+        async def authorization_required(_authorization_url: str) -> None:
+            await storage.clear_tokens()
+            raise MCPAuthorizationRequiredError("MCP server requires browser authorization")
+
+        redirect_handler = authorization_required
+        callback_handler = _missing_callback
+
+    metadata = OAuthClientMetadata(
+        redirect_uris=[AnyUrl(redirect_uri)],
+        token_endpoint_auth_method="none",
+        client_name="nanobot",
+        client_uri=_CLIENT_URI,
+        logo_uri=_LOGO_URI,
+        software_id="https://github.com/HKUDS/nanobot",
+    )
+    return OAuthClientProvider(
+        server_url,
+        metadata,
+        storage,
+        redirect_handler=redirect_handler,
+        callback_handler=callback_handler,
+        timeout=300,
+    )
+
+
+def mcp_oauth_has_credentials(server_name: str, server_url: str) -> bool:
+    """Return whether this exact configured MCP instance has an access token."""
+    return MCPOAuthStorage(server_name, server_url).has_credentials()
+
+
+def delete_mcp_oauth_credentials(server_name: str) -> bool:
+    """Delete credentials for one config name without touching other MCP instances."""
+    path = _store_path()
+    with _with_store_lock(path):
+        payload = _read_store_unlocked(path)
+        servers = payload["servers"]
+        removed = servers.pop(server_name, None) is not None
+        # Rotate even when no entry exists so a flow created before removal cannot
+        # claim the name later and resurrect credentials.
+        payload["generations"][server_name] = secrets.token_urlsafe(24)
+        _write_store_unlocked(path, payload)
+        return removed

nanobot/agent/tools/runtime_control.py

@@ -0,0 +1,318 @@
+"""Explicit runtime state boundary used by :class:`MyTool`."""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Protocol, TypeAlias, runtime_checkable
+
+if TYPE_CHECKING:
+    from nanobot.agent.subagent import SubagentManager, SubagentStatus
+    from nanobot.agent.tools.shell import ExecToolConfig
+    from nanobot.agent.tools.web import WebToolsConfig
+    from nanobot.config.schema import ModelPresetConfig
+    from nanobot.utils.llm_runtime import LLMRuntime
+
+
+JsonScalar: TypeAlias = str | int | float | bool | None
+JsonValue: TypeAlias = JsonScalar | list["JsonValue"] | dict[str, "JsonValue"]
+
+
+RUNTIME_SNAPSHOT_KEYS = frozenset({
+    "model",
+    "model_preset",
+    "model_presets",
+    "max_iterations",
+    "context_window_tokens",
+    "workspace",
+    "provider_retry_mode",
+    "max_tool_result_chars",
+    "current_iteration",
+    "_current_iteration",
+    "tool_names",
+    "web_config",
+    "exec_config",
+    "subagents",
+    "_last_usage",
+})
+
+RUNTIME_COMMAND_KEYS = frozenset({
+    "model",
+    "model_preset",
+    "max_iterations",
+    "context_window_tokens",
+    "provider_retry_mode",
+    "max_tool_result_chars",
+    "workspace",
+})
+
+
+@dataclass(frozen=True, slots=True)
+class RuntimeSnapshot:
+    """Detached, allowlisted values available to self-inspection."""
+
+    model: str
+    model_preset: str | None
+    model_presets: dict[str, dict[str, object]]
+    max_iterations: int
+    context_window_tokens: int
+    workspace: Path | str
+    provider_retry_mode: str
+    max_tool_result_chars: int
+    current_iteration: int
+    tool_names: list[str]
+    web_config: dict[str, object]
+    exec_config: dict[str, object]
+    subagent_statuses: dict[str, dict[str, object]]
+    last_usage: dict[str, int]
+    scratchpad: dict[str, JsonValue]
+
+    def as_mapping(self) -> Mapping[str, object]:
+        """Return the fixed public names understood by ``MyTool``."""
+        values: dict[str, object] = {
+            "model": self.model,
+            "model_preset": self.model_preset,
+            "model_presets": self.model_presets,
+            "max_iterations": self.max_iterations,
+            "context_window_tokens": self.context_window_tokens,
+            "workspace": self.workspace,
+            "provider_retry_mode": self.provider_retry_mode,
+            "max_tool_result_chars": self.max_tool_result_chars,
+            "current_iteration": self.current_iteration,
+            "_current_iteration": self.current_iteration,
+            "tool_names": self.tool_names,
+            "web_config": self.web_config,
+            "exec_config": self.exec_config,
+            "subagents": {"_task_statuses": self.subagent_statuses},
+            "_last_usage": self.last_usage,
+        }
+        assert values.keys() == RUNTIME_SNAPSHOT_KEYS
+        return values
+
+
+@runtime_checkable
+class RuntimeControl(Protocol):
+    """The complete runtime capability exposed to ``MyTool``."""
+
+    def snapshot(self) -> RuntimeSnapshot: ...
+
+    def set_model(self, model: str) -> LLMRuntime: ...
+
+    def set_model_preset(
+        self,
+        name: str,
+        *,
+        session_key: str | None,
+    ) -> LLMRuntime: ...
+
+    def set_max_iterations(self, value: int) -> None: ...
+
+    def set_context_window_tokens(self, value: int) -> LLMRuntime: ...
+
+    def set_provider_retry_mode(self, value: str) -> None: ...
+
+    def set_max_tool_result_chars(self, value: int) -> None: ...
+
+    def set_workspace_display(self, value: str) -> None: ...
+
+    def set_scratchpad(self, key: str, value: JsonValue, *, max_keys: int) -> None: ...
+
+
+class _RuntimeControlTarget(Protocol):
+    """Narrow structural dependency required by ``AgentRuntimeControl``."""
+
+    max_iterations: int
+    provider_retry_mode: str
+    max_tool_result_chars: int
+    web_config: WebToolsConfig
+    exec_config: ExecToolConfig
+    subagents: SubagentManager
+
+    @property
+    def model(self) -> str: ...
+
+    @property
+    def model_preset(self) -> str | None: ...
+
+    @property
+    def model_presets(self) -> Mapping[str, ModelPresetConfig]: ...
+
+    @property
+    def context_window_tokens(self) -> int: ...
+
+    @property
+    def workspace(self) -> Path: ...
+
+    @property
+    def current_iteration(self) -> int: ...
+
+    @property
+    def tool_names(self) -> list[str]: ...
+
+    @property
+    def last_usage(self) -> Mapping[str, int]: ...
+
+    def set_runtime_model(self, model: str) -> LLMRuntime: ...
+
+    def set_runtime_context_window(self, context_window_tokens: int) -> LLMRuntime: ...
+
+    def set_model_preset(self, name: str | None) -> LLMRuntime: ...
+
+    def set_session_model_preset(self, session_key: str, name: str) -> LLMRuntime: ...
+
+
+class AgentRuntimeControl:
+    """Allowlisted adapter from agent-loop state to ``RuntimeControl``."""
+
+    def __init__(self, target: _RuntimeControlTarget) -> None:
+        self.__target = target
+        self.__scratchpad: dict[str, JsonValue] = {}
+        self.__workspace_display: str | None = None
+
+    def snapshot(self) -> RuntimeSnapshot:
+        target = self.__target
+        return RuntimeSnapshot(
+            model=target.model,
+            model_preset=target.model_preset,
+            model_presets=_snapshot_model_presets(target.model_presets),
+            max_iterations=target.max_iterations,
+            context_window_tokens=target.context_window_tokens,
+            workspace=(
+                self.__workspace_display
+                if self.__workspace_display is not None
+                else target.workspace
+            ),
+            provider_retry_mode=target.provider_retry_mode,
+            max_tool_result_chars=target.max_tool_result_chars,
+            current_iteration=target.current_iteration,
+            tool_names=list(target.tool_names),
+            web_config=_snapshot_web_config(target.web_config),
+            exec_config=_snapshot_exec_config(target.exec_config),
+            subagent_statuses=_snapshot_subagent_statuses(target.subagents),
+            last_usage=dict(target.last_usage),
+            scratchpad=_snapshot_json_mapping(self.__scratchpad),
+        )
+
+    def set_model(self, model: str) -> LLMRuntime:
+        return self.__target.set_runtime_model(model)
+
+    def set_model_preset(
+        self,
+        name: str,
+        *,
+        session_key: str | None,
+    ) -> LLMRuntime:
+        if session_key is not None:
+            return self.__target.set_session_model_preset(session_key, name)
+        return self.__target.set_model_preset(name)
+
+    def set_max_iterations(self, value: int) -> None:
+        self.__target.max_iterations = value
+        self.__target.subagents.max_iterations = value
+
+    def set_context_window_tokens(self, value: int) -> LLMRuntime:
+        return self.__target.set_runtime_context_window(value)
+
+    def set_provider_retry_mode(self, value: str) -> None:
+        self.__target.provider_retry_mode = value
+
+    def set_max_tool_result_chars(self, value: int) -> None:
+        self.__target.max_tool_result_chars = value
+
+    def set_workspace_display(self, value: str) -> None:
+        """Preserve MyTool display compatibility without changing path enforcement."""
+        self.__workspace_display = value
+
+    def set_scratchpad(self, key: str, value: JsonValue, *, max_keys: int) -> None:
+        if key not in self.__scratchpad and len(self.__scratchpad) >= max_keys:
+            raise ValueError(f"scratchpad is full (max {max_keys} keys)")
+        self.__scratchpad[key] = value
+
+
+def _snapshot_model_presets(
+    presets: Mapping[str, ModelPresetConfig],
+) -> dict[str, dict[str, object]]:
+    return {
+        name: {
+            "model": preset.model,
+            "provider": preset.provider,
+            "max_tokens": preset.max_tokens,
+            "context_window_tokens": preset.context_window_tokens,
+            "temperature": preset.temperature,
+            "reasoning_effort": preset.reasoning_effort,
+        }
+        for name, preset in presets.items()
+    }
+
+
+def _snapshot_web_config(config: WebToolsConfig) -> dict[str, object]:
+    return {
+        "enable": config.enable,
+        # Proxy URLs may embed credentials. Presence is enough for diagnosis.
+        "proxy": "<configured>" if config.proxy else config.proxy,
+        "user_agent": config.user_agent,
+        "search": {
+            "provider": config.search.provider,
+            "base_url": config.search.base_url,
+            "max_results": config.search.max_results,
+            "timeout": config.search.timeout,
+        },
+        "fetch": {
+            "use_jina_reader": config.fetch.use_jina_reader,
+        },
+    }
+
+
+def _snapshot_exec_config(config: ExecToolConfig) -> dict[str, object]:
+    return {
+        "enable": config.enable,
+        "timeout": config.timeout,
+        "path_prepend": config.path_prepend,
+        "path_append": config.path_append,
+        "sandbox": config.sandbox,
+        "sandbox_ro_binds": list(config.sandbox_ro_binds),
+        "sandbox_rw_binds": list(config.sandbox_rw_binds),
+        "allowed_env_keys": list(config.allowed_env_keys),
+        "allow_patterns": list(config.allow_patterns),
+        "deny_patterns": list(config.deny_patterns),
+    }
+
+
+def _snapshot_subagent_statuses(
+    manager: SubagentManager,
+) -> dict[str, dict[str, object]]:
+    return {
+        task_id: _snapshot_subagent_status(status)
+        for task_id, status in manager.runtime_statuses().items()
+    }
+
+
+def _snapshot_subagent_status(status: SubagentStatus) -> dict[str, object]:
+    return {
+        "task_id": status.task_id,
+        "label": status.label,
+        "task_description": status.task_description,
+        "started_at": status.started_at,
+        "phase": status.phase,
+        "iteration": status.iteration,
+        "tool_events": [dict(event) for event in status.tool_events],
+        "usage": dict(status.usage),
+        "stop_reason": status.stop_reason,
+        "error": status.error,
+    }
+
+
+def _snapshot_json_mapping(values: Mapping[str, JsonValue]) -> dict[str, JsonValue]:
+    return {key: _snapshot_json_value(value) for key, value in values.items()}
+
+
+def _snapshot_json_value(value: JsonValue) -> JsonValue:
+    if isinstance(value, list):
+        return [_snapshot_json_value(item) for item in value]
+    if isinstance(value, dict):
+        return {
+            key: _snapshot_json_value(item)
+            for key, item in value.items()
+        }
+    return value

nanobot/agent/tools/runtime_state.py

@@ -1,76 +0,0 @@
-"""RuntimeState protocol: agent loop state exposed to MyTool."""
-
-from __future__ import annotations
-
-from pathlib import Path
-from typing import TYPE_CHECKING, Any, Protocol
-
-if TYPE_CHECKING:
-    from nanobot.agent.subagent import SubagentManager
-    from nanobot.agent.tools.shell import ExecToolConfig
-    from nanobot.agent.tools.web import WebToolsConfig
-    from nanobot.utils.llm_runtime import LLMRuntime
-
-
-class RuntimeState(Protocol):
-    """Minimum contract that MyTool requires from its runtime state provider.
-
-    In practice, this is always satisfied by ``AgentLoop``.  MyTool also
-    accesses arbitrary attributes dynamically (via ``getattr`` / ``setattr``)
-    for dot-path inspection and modification; those paths are validated at
-    runtime rather than by this protocol.
-    """
-
-    @property
-    def model(self) -> str: ...
-
-    @property
-    def max_iterations(self) -> int: ...
-
-    @property
-    def current_iteration(self) -> int: ...
-
-    @property
-    def tool_names(self) -> list[str]: ...
-
-    @property
-    def workspace(self) -> Path: ...
-
-    @property
-    def provider_retry_mode(self) -> str: ...
-
-    @property
-    def max_tool_result_chars(self) -> int: ...
-
-    @property
-    def context_window_tokens(self) -> int: ...
-
-    @property
-    def web_config(self) -> WebToolsConfig: ...
-
-    @property
-    def exec_config(self) -> ExecToolConfig: ...
-
-    @property
-    def subagents(self) -> SubagentManager: ...
-
-    @property
-    def _runtime_vars(self) -> dict[str, Any]: ...
-
-    @property
-    def _last_usage(self) -> dict[str, int]: ...
-
-    def _sync_subagent_runtime_limits(self) -> None: ...
-
-    def set_runtime_model(self, model: str) -> LLMRuntime: ...
-
-    def set_runtime_context_window(self, context_window_tokens: int) -> LLMRuntime: ...
-
-    def set_session_model_preset(
-        self,
-        session_key: str,
-        name: str,
-    ) -> LLMRuntime: ...
-
-    @property
-    def model_preset(self) -> str | None: ...

nanobot/agent/tools/self.py

@@ -1,8 +1,7 @@
 """MyTool: runtime state inspection and configuration for the agent loop."""
 
-# RuntimeState intentionally exposes a narrow set of AgentLoop internals to
-# this manually registered tool. Tool.execute accepts heterogeneous schemas.
-# pyright: reportPrivateUsage=false, reportIncompatibleMethodOverride=false
+# Tool.execute accepts heterogeneous schemas.
+# pyright: reportIncompatibleMethodOverride=false
 
 from __future__ import annotations
 
@@ -14,7 +13,13 @@ from loguru import logger
 
 from nanobot.agent.tools.base import Tool, ToolResult
 from nanobot.agent.tools.context import current_request_context, current_request_session_key
-from nanobot.agent.tools.runtime_state import RuntimeState
+from nanobot.agent.tools.runtime_control import (
+    RUNTIME_COMMAND_KEYS,
+    RUNTIME_SNAPSHOT_KEYS,
+    JsonValue,
+    RuntimeControl,
+    RuntimeSnapshot,
+)
 from nanobot.config_base import Base
 
 if TYPE_CHECKING:
@@ -28,25 +33,28 @@ class MyToolConfig(Base):
     allow_set: bool = False
 
 
-def _has_real_attr(obj: Any, key: str) -> bool:
-    """Check if obj has a real (explicitly set) attribute, not auto-generated by mock."""
-    if isinstance(obj, dict):
-        return key in obj
-    d = getattr(obj, "__dict__", None)
-    if d is not None and key in d:
-        return True
-    for cls in type(obj).__mro__:
-        if key in cls.__dict__:
-            return True
-    return False
-
-
 def _is_subagent_status(value: object) -> TypeGuard[SubagentStatus]:
     from nanobot.agent.subagent import SubagentStatus
 
     return isinstance(value, SubagentStatus)
 
 
+def _is_subagent_status_snapshot(value: object) -> TypeGuard[Mapping[str, object]]:
+    if not isinstance(value, Mapping):
+        return False
+    return all(
+        field in value
+        for field in ("task_id", "label", "task_description", "started_at", "phase")
+    )
+
+
+def _is_string_mapping(value: object) -> TypeGuard[Mapping[str, object]]:
+    if not isinstance(value, Mapping):
+        return False
+    mapping = cast(Mapping[object, object], value)
+    return all(isinstance(key, str) for key in mapping)
+
+
 class MyTool(Tool):
     """Check and set the agent loop's runtime configuration."""
 
@@ -70,7 +78,7 @@ class MyTool(Tool):
         "runner", "sessions", "consolidator",
         "dream", "auto_compact", "context", "commands",
         # Sensitive runtime state (credentials, message routing, task tracking)
-        "_mcp_servers", "_mcp_stacks", "_pending_queues",
+        "_pending_queues",
         "_session_locks", "_active_tasks", "_background_tasks",
         # Security boundaries (inspect + modify both blocked)
         "restrict_to_workspace", "channels_config",
@@ -79,7 +87,10 @@ class MyTool(Tool):
 
     READ_ONLY = frozenset({
         "subagents",  # observable but replacing it would break the system
+        "tool_names",
+        "current_iteration",
         "_current_iteration",  # updated by runner only
+        "_last_usage",
         "exec_config",  # inspect allowed (e.g. check sandbox), modify blocked
         "web_config",  # inspect allowed (e.g. check enable), modify blocked
         "model_presets",  # config-derived catalog; changes require config reload
@@ -103,13 +114,6 @@ class MyTool(Tool):
         "private_key", "access_token", "refresh_token", "auth",
     })
 
-    @classmethod
-    def _is_sensitive_field_name(cls, name: str) -> bool:
-        lowered = name.lower()
-        return lowered in cls._SENSITIVE_NAMES or any(
-            part in cls._SENSITIVE_NAMES for part in lowered.split("_")
-        )
-
     RESTRICTED: dict[str, dict[str, Any]] = {
         "max_iterations":        {"type": int, "min": 1,   "max": 100},
         "context_window_tokens": {"type": int, "min": 4096, "max": 1_000_000},
@@ -123,15 +127,15 @@ class MyTool(Tool):
         "context_window_tokens",
     })
 
-    def __init__(self, runtime_state: RuntimeState, modify_allowed: bool = True) -> None:
-        self._runtime_state = runtime_state
+    def __init__(self, runtime_control: RuntimeControl, modify_allowed: bool = True) -> None:
+        self._runtime_control = runtime_control
         self._modify_allowed = modify_allowed
 
     def __deepcopy__(self, memo: dict[int, Any]) -> MyTool:
         cls = self.__class__
         result = cls.__new__(cls)
         memo[id(self)] = result
-        result._runtime_state = self._runtime_state
+        result._runtime_control = self._runtime_control
         result._modify_allowed = self._modify_allowed
         return result
 
@@ -208,9 +212,12 @@ class MyTool(Tool):
     # Path resolution
     # ------------------------------------------------------------------
 
-    def _resolve_path(self, path: str) -> tuple[Any, str | None]:
+    def _resolve_path(
+        self,
+        snapshot: RuntimeSnapshot,
+        path: str,
+    ) -> tuple[object | None, str | None]:
         parts = path.split(".")
-        obj: Any = self._runtime_state
         for part in parts:
             if part in self._DENIED_ATTRS or part.startswith("__"):
                 return None, f"'{part}' is not accessible"
@@ -218,17 +225,13 @@ class MyTool(Tool):
                 return None, f"'{part}' is not accessible"
             if part.lower() in self._SENSITIVE_NAMES:
                 return None, f"'{part}' is not accessible"
-            try:
-                if isinstance(obj, Mapping):
-                    mapping = cast(Mapping[str, Any], obj)
-                    if part in mapping:
-                        obj = mapping[part]
-                    else:
+        obj: object = snapshot.as_mapping()
+        for part in parts:
+            if not _is_string_mapping(obj):
+                return None, f"'{part}' not found"
+            if part not in obj:
                 return None, f"'{part}' not found in mapping"
-                else:
-                    obj = getattr(obj, part)
-            except (KeyError, AttributeError) as e:
-                return None, f"'{part}' not found: {e}"
+            obj = obj[part]
         return obj, None
 
     @staticmethod
@@ -242,20 +245,48 @@ class MyTool(Tool):
     # ------------------------------------------------------------------
 
     @staticmethod
-    def _format_status(st: "SubagentStatus", indent: str = "  ") -> str:
-        elapsed = time.monotonic() - st.started_at
-        tool_summary = ", ".join(
-            f"{e.get('name', '?')}({e.get('status', '?')})" for e in st.tool_events[-5:]
-        ) or "none"
+    def _format_status(
+        st: "SubagentStatus | Mapping[str, object]",
+        indent: str = "  ",
+    ) -> str:
+        if isinstance(st, Mapping):
+            started_at = st.get("started_at", time.monotonic())
+            raw_events = st.get("tool_events", [])
+            phase = st.get("phase", "unknown")
+            iteration = st.get("iteration", 0)
+            usage = st.get("usage", {})
+            error = st.get("error")
+            stop_reason = st.get("stop_reason")
+        else:
+            started_at = st.started_at
+            raw_events = st.tool_events
+            phase = st.phase
+            iteration = st.iteration
+            usage = st.usage
+            error = st.error
+            stop_reason = st.stop_reason
+        elapsed = time.monotonic() - (
+            float(started_at) if isinstance(started_at, (int, float)) else time.monotonic()
+        )
+        tool_events = cast(list[object], raw_events) if isinstance(raw_events, list) else []
+        tool_summaries: list[str] = []
+        for raw_event in tool_events[-5:]:
+            if not isinstance(raw_event, Mapping):
+                continue
+            event = cast(Mapping[str, object], raw_event)
+            tool_summaries.append(
+                f"{event.get('name', '?')}({event.get('status', '?')})"
+            )
+        tool_summary = ", ".join(tool_summaries) or "none"
         lines = [
-            f"{indent}phase: {st.phase}, iteration: {st.iteration}, elapsed: {elapsed:.1f}s",
+            f"{indent}phase: {phase}, iteration: {iteration}, elapsed: {elapsed:.1f}s",
             f"{indent}tools: {tool_summary}",
-            f"{indent}usage: {st.usage or 'n/a'}",
+            f"{indent}usage: {usage or 'n/a'}",
         ]
-        if st.error:
-            lines.append(f"{indent}error: {st.error}")
-        if st.stop_reason:
-            lines.append(f"{indent}stop_reason: {st.stop_reason}")
+        if error:
+            lines.append(f"{indent}error: {error}")
+        if stop_reason:
+            lines.append(f"{indent}stop_reason: {stop_reason}")
         return "\n".join(lines)
 
     @staticmethod
@@ -264,29 +295,38 @@ class MyTool(Tool):
             header = f"Subagent [{val.task_id}] '{val.label}'"
             detail = MyTool._format_status(val, "  ")
             return f"{header}\n  task: {val.task_description}\n{detail}"
-        # SubagentManager: delegate to its _task_statuses dict
-        task_statuses = getattr(val, "_task_statuses", None)
-        if isinstance(task_statuses, dict):
-            return MyTool._format_value(task_statuses, key)
+        if _is_subagent_status_snapshot(val):
+            header = f"Subagent [{val['task_id']}] '{val['label']}'"
+            detail = MyTool._format_status(val, "  ")
+            return f"{header}\n  task: {val['task_description']}\n{detail}"
         if isinstance(val, Mapping):
             mapping = cast(Mapping[object, object], val)
         else:
             mapping = None
+        if mapping and set(mapping) == {"_task_statuses"}:
+            task_statuses = mapping["_task_statuses"]
+            if isinstance(task_statuses, Mapping):
+                return MyTool._format_value(task_statuses, key)
         if (
             mapping
-            and _is_subagent_status(next(iter(mapping.values())))
+            and (
+                _is_subagent_status(next(iter(mapping.values())))
+                or _is_subagent_status_snapshot(next(iter(mapping.values())))
+            )
         ):
-            status_mapping: Mapping[object, SubagentStatus] = cast(Any, mapping)
             prefix = f"{key}: " if key else ""
-            lines = [f"{prefix}{len(status_mapping)} subagent(s):"]
-            for tid, st in status_mapping.items():
+            lines = [f"{prefix}{len(mapping)} subagent(s):"]
+            for tid, st in mapping.items():
+                if _is_subagent_status(st):
                     detail = MyTool._format_status(st, "    ")
-                lines.append(f"  [{tid}] '{st.label}'\n{detail}")
+                    label = st.label
+                elif _is_subagent_status_snapshot(st):
+                    detail = MyTool._format_status(st, "    ")
+                    label = st.get("label", "?")
+                else:
+                    continue
+                lines.append(f"  [{tid}] '{label}'\n{detail}")
             return "\n".join(lines)
-        dynamic_value = cast(Any, val)
-        if hasattr(dynamic_value, "tool_names"):
-            tool_names: Any = getattr(dynamic_value, "tool_names")
-            return f"tools: {len(tool_names)} registered — {tool_names}"
         # Scalar types — repr is fine
         if isinstance(val, (str, int, float, bool, type(None))):
             r = repr(val)
@@ -311,32 +351,6 @@ class MyTool(Tool):
                 return f"{key}: [{len(sequence)} items]" if key else f"[{len(sequence)} items]"
             r = repr(sequence)
             return f"{key}: {r}" if key else r
-        # Complex object — small Pydantic models: show values; others: show field names for navigation
-        value_type = type(cast(object, val))
-        cls_name = value_type.__name__
-        model_fields = cast(object, getattr(value_type, "model_fields", None))
-        if isinstance(model_fields, Mapping) and model_fields:
-            fields = list(cast(Mapping[str, object], model_fields).keys())
-            if len(fields) <= 8:
-                # Small config objects: show field=value pairs
-                pairs: list[str] = []
-                for f in fields:
-                    fv = getattr(val, f, "?")
-                    if MyTool._is_sensitive_field_name(f):
-                        continue
-                    if isinstance(fv, (str, int, float, bool, type(None))):
-                        pairs.append(f"{f}={fv!r}")
-                    else:
-                        pairs.append(f"{f}=<{type(fv).__name__}>")
-                preview = ", ".join(pairs)
-                return f"{key}: {preview}" if key else preview
-        else:
-            attributes = cast(dict[str, Any], getattr(val, "__dict__", {}))
-            fields = [name for name in attributes if not name.startswith("__")]
-        if fields:
-            preview = ", ".join(str(f) for f in fields[:20])
-            suffix = ", ..." if len(fields) > 20 else ""
-            return f"{key}: <{cls_name}> [{preview}{suffix}]" if key else f"<{cls_name}> [{preview}{suffix}]"
         r = repr(val)
         return f"{key}: {r}" if key else r
 
@@ -366,7 +380,12 @@ class MyTool(Tool):
         runtime = request_ctx.runtime if request_ctx is not None else None
         if runtime is None or key not in self._MODEL_RUNTIME_FIELDS:
             return False, None
-        return True, getattr(runtime, key)
+        values: dict[str, object] = {
+            "model": runtime.model,
+            "model_preset": runtime.model_preset,
+            "context_window_tokens": runtime.context_window_tokens,
+        }
+        return True, values[key]
 
     def _inspect(self, key: str | None) -> str:
         if not key:
@@ -375,62 +394,64 @@ class MyTool(Tool):
             request_ctx = current_request_context()
             if request_ctx is None:
                 return ToolResult.error("Error: current request context is unavailable")
+            request_values: dict[str, str | None] = {
+                "channel": request_ctx.channel,
+                "chat_id": request_ctx.chat_id,
+                "sender_id": request_ctx.sender_id,
+            }
             if key == "request":
-                return self._format_value(
-                    {field: getattr(request_ctx, field) for field in self._REQUEST_FIELDS},
-                    key,
-                )
+                return self._format_value(request_values, key)
             field = key.removeprefix("request.")
             if field not in self._REQUEST_FIELDS:
                 return ToolResult.error(f"Error: '{key}' not found")
-            return self._format_value(getattr(request_ctx, field), key)
+            return self._format_value(request_values[field], key)
         if "." not in key:
             found, value = self._current_runtime_value(key)
             if found:
                 return self._format_value(value, key)
+        snapshot = self._runtime_control.snapshot()
         top = key.split(".")[0]
         if top in self._DENIED_ATTRS or top.startswith("__"):
             return ToolResult.error(f"Error: '{top}' is not accessible")
-        obj, err = self._resolve_path(key)
+        obj, err = self._resolve_path(snapshot, key)
         if err:
-            # "scratchpad" alias for _runtime_vars
             if key == "scratchpad":
-                rv = self._runtime_state._runtime_vars
-                return self._format_value(rv, "scratchpad") if rv else "scratchpad is empty"
-            # Fallback: check _runtime_vars for simple keys stored by modify
-            if "." not in key and key in self._runtime_state._runtime_vars:
-                return self._format_value(self._runtime_state._runtime_vars[key], key)
+                return (
+                    self._format_value(snapshot.scratchpad, "scratchpad")
+                    if snapshot.scratchpad
+                    else "scratchpad is empty"
+                )
+            if "." not in key and key in snapshot.scratchpad:
+                return self._format_value(snapshot.scratchpad[key], key)
             return ToolResult.error(f"Error: {err}")
-        # Guard against mock auto-generated attributes
-        if "." not in key and not _has_real_attr(self._runtime_state, key):
-            if key in self._runtime_state._runtime_vars:
-                return self._format_value(self._runtime_state._runtime_vars[key], key)
-            return ToolResult.error(f"Error: '{key}' not found")
         return self._format_value(obj, key)
 
     def _inspect_all(self) -> str:
-        state = self._runtime_state
+        snapshot = self._runtime_control.snapshot()
+        values = snapshot.as_mapping()
         parts: list[str] = []
-        # RESTRICTED keys
         for k in self.RESTRICTED:
             found, value = self._current_runtime_value(k)
-            parts.append(self._format_value(value if found else getattr(state, k, None), k))
+            parts.append(self._format_value(value if found else values[k], k))
         found, value = self._current_runtime_value("model_preset")
         parts.append(self._format_value(
-            value if found else state.model_preset,
+            value if found else snapshot.model_preset,
             "model_preset",
         ))
-        # Other useful top-level keys shown in description
-        for k in ("workspace", "provider_retry_mode", "max_tool_result_chars", "_current_iteration", "web_config", "exec_config", "workspace_sandbox", "subagents"):
-            if _has_real_attr(state, k):
-                parts.append(self._format_value(getattr(state, k, None), k))
-        # Token usage
-        usage = state._last_usage
-        if usage:
-            parts.append(self._format_value(usage, "_last_usage"))
-        rv = state._runtime_vars
-        if rv:
-            parts.append(self._format_value(rv, "scratchpad"))
+        for k in (
+            "workspace",
+            "provider_retry_mode",
+            "max_tool_result_chars",
+            "_current_iteration",
+            "web_config",
+            "exec_config",
+            "subagents",
+        ):
+            parts.append(self._format_value(values[k], k))
+        if snapshot.last_usage:
+            parts.append(self._format_value(snapshot.last_usage, "_last_usage"))
+        if snapshot.scratchpad:
+            parts.append(self._format_value(snapshot.scratchpad, "scratchpad"))
         return "\n".join(parts)
 
     # -- modify --
@@ -454,48 +475,49 @@ class MyTool(Tool):
             if leaf.lower() in self._SENSITIVE_NAMES:
                 self._audit("modify", f"BLOCKED sensitive leaf '{leaf}'")
                 return ToolResult.error(f"Error: '{leaf}' is not accessible")
-            parent, err = self._resolve_path(parent_path)
+            snapshot = self._runtime_control.snapshot()
+            _parent, err = self._resolve_path(snapshot, parent_path)
             if err:
                 return ToolResult.error(f"Error: {err}")
-            if isinstance(parent, dict):
-                parent[leaf] = value
-            else:
-                setattr(parent, leaf, value)
-            self._audit("modify", f"{key} = {value!r}")
-            return f"Set {key} = {value!r}"
+            self._audit("modify", f"READ_ONLY {key}")
+            return ToolResult.error(f"Error: '{key}' is read-only and cannot be modified")
         if key == "model_preset":
             return self._modify_model_preset(value)
         if key in self.RESTRICTED:
             return self._modify_restricted(key, value)
-        return self._modify_free(key, value)
+        if key in RUNTIME_COMMAND_KEYS:
+            return self._modify_runtime_setting(key, value)
+        if key in RUNTIME_SNAPSHOT_KEYS:
+            self._audit("modify", f"READ_ONLY {key}")
+            return ToolResult.error(f"Error: '{key}' is read-only and cannot be modified")
+        return self._modify_scratchpad(key, value)
 
     def _modify_model_preset(self, value: Any) -> str:
         if not isinstance(value, str) or not value.strip():
             return ToolResult.error("Error: 'model_preset' must be a non-empty string")
         name = value.strip()
         session_key = current_request_session_key()
-        if session_key:
+        old = self._runtime_control.snapshot().model_preset
         try:
-                runtime = self._runtime_state.set_session_model_preset(
-                    session_key,
+            runtime = self._runtime_control.set_model_preset(
                 name,
+                session_key=session_key,
             )
         except (KeyError, ValueError) as exc:
             message = str(exc.args[0]) if exc.args else str(exc)
             punctuation = "" if message.endswith((".", "!", "?")) else "."
             return ToolResult.error(f"Error: {message}{punctuation}")
+        if session_key:
             self._audit("modify", f"model_preset = {name!r}")
             return (
                 f"Set model_preset = {name!r} for the next turn; "
                 f"model will be {runtime.model!r}; "
                 f"context_window_tokens will be {runtime.context_window_tokens!r}"
             )
-        result = self._modify_free("model_preset", name)
-        if isinstance(result, ToolResult) and result.is_error:
-            return result if result.endswith((".", "!", "?")) else ToolResult.error(f"{result}.")
+        self._audit("modify", f"model_preset: {old!r} -> {name!r}")
         return (
-            f"{result}; model is now {self._runtime_state.model!r}; "
-            f"context_window_tokens is now {self._runtime_state.context_window_tokens!r}"
+            f"Set model_preset = {name!r} (was {old!r}); model is now {runtime.model!r}; "
+            f"context_window_tokens is now {runtime.context_window_tokens!r}"
         )
 
     def _modify_restricted(self, key: str, value: Any) -> str:
@@ -508,7 +530,7 @@ class MyTool(Tool):
                 value = expected(value)
             except (ValueError, TypeError):
                 return ToolResult.error(f"Error: '{key}' must be {expected.__name__}, got {type(value).__name__}")
-        old = getattr(self._runtime_state, key)
+        old = self._runtime_control.snapshot().as_mapping()[key]
         if "min" in spec and value < spec["min"]:
             return ToolResult.error(f"Error: '{key}' must be >= {spec['min']}")
         if "max" in spec and value > spec["max"]:
@@ -521,41 +543,46 @@ class MyTool(Tool):
                 "during an active session; use a configured model_preset"
             )
         if key == "model":
-            self._runtime_state.set_runtime_model(cast(str, value))
+            self._runtime_control.set_model(cast(str, value))
         elif key == "context_window_tokens":
-            self._runtime_state.set_runtime_context_window(cast(int, value))
+            self._runtime_control.set_context_window_tokens(cast(int, value))
         else:
-            setattr(self._runtime_state, key, value)
-        if key == "max_iterations" and hasattr(
-            self._runtime_state,
-            "_sync_subagent_runtime_limits",
-        ):
-            self._runtime_state._sync_subagent_runtime_limits()
+            self._runtime_control.set_max_iterations(cast(int, value))
         self._audit("modify", f"{key}: {old!r} -> {value!r}")
         return f"Set {key} = {value!r} (was {old!r})"
 
-    def _modify_free(self, key: str, value: Any) -> str:
-        if _has_real_attr(self._runtime_state, key):
-            old = getattr(self._runtime_state, key)
-            if isinstance(old, (str, int, float, bool)):
-                old_t: type[Any] = type(old)
+    def _modify_runtime_setting(self, key: str, value: Any) -> str:
+        old = self._runtime_control.snapshot().as_mapping()[key]
+        if key == "workspace":
+            if not isinstance(value, str):
+                return ToolResult.error(
+                    f"Error: 'workspace' expects str, got {type(value).__name__}"
+                )
+            self._runtime_control.set_workspace_display(value)
+            self._audit("modify", f"workspace: {old!r} -> {value!r}")
+            return f"Set workspace = {value!r} (was {old!r})"
+        old_t = type(old)
         new_t = cast(type[Any], type(value))
         if old_t is float and new_t is int:
-                    pass  # int → float coercion allowed
+            pass
         elif old_t is not new_t:
             self._audit(
                 "modify",
                 f"REJECTED type mismatch {key}: expects {old_t.__name__}, got {new_t.__name__}",
             )
-                    return ToolResult.error(f"Error: '{key}' expects {old_t.__name__}, got {new_t.__name__}")
-            try:
-                setattr(self._runtime_state, key, value)
-            except (ValueError, KeyError) as e:
-                message = str(e.args[0] if isinstance(e, KeyError) and e.args else e).strip('"')
-                self._audit("modify", f"REJECTED {key}: {message}")
-                return ToolResult.error(f"Error: {message}")
+            return ToolResult.error(
+                f"Error: '{key}' expects {old_t.__name__}, got {new_t.__name__}"
+            )
+        if key == "provider_retry_mode":
+            self._runtime_control.set_provider_retry_mode(cast(str, value))
+        elif key == "max_tool_result_chars":
+            self._runtime_control.set_max_tool_result_chars(cast(int, value))
+        else:
+            raise AssertionError(f"Unhandled runtime command: {key}")
         self._audit("modify", f"{key}: {old!r} -> {value!r}")
         return f"Set {key} = {value!r} (was {old!r})"
+
+    def _modify_scratchpad(self, key: str, value: Any) -> str:
         if callable(value):
             self._audit("modify", f"REJECTED callable {key}")
             return ToolResult.error("Error: cannot store callable values")
@@ -563,12 +590,16 @@ class MyTool(Tool):
         if err:
             self._audit("modify", f"REJECTED {key}: {err}")
             return ToolResult.error(f"Error: {err}")
-        if key not in self._runtime_state._runtime_vars and len(self._runtime_state._runtime_vars) >= self._MAX_RUNTIME_KEYS:
+        try:
+            self._runtime_control.set_scratchpad(
+                key,
+                cast(JsonValue, value),
+                max_keys=self._MAX_RUNTIME_KEYS,
+            )
+        except ValueError as exc:
             self._audit("modify", f"REJECTED {key}: max keys ({self._MAX_RUNTIME_KEYS}) reached")
-            return ToolResult.error(f"Error: scratchpad is full (max {self._MAX_RUNTIME_KEYS} keys). Remove unused keys first.")
-        old = self._runtime_state._runtime_vars.get(key)
-        self._runtime_state._runtime_vars[key] = value
-        self._audit("modify", f"scratchpad.{key}: {old!r} -> {value!r}")
+            return ToolResult.error(f"Error: {exc}. Remove unused keys first.")
+        self._audit("modify", f"scratchpad.{key} = {value!r}")
         return f"Set scratchpad.{key} = {value!r}"
 
     @classmethod

nanobot/agent/tools/shell.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import os
 import re
+import shlex
 import shutil
 import signal
 import subprocess
@@ -12,7 +13,8 @@ import sys
 from contextlib import suppress
 from dataclasses import dataclass
 from pathlib import Path, PureWindowsPath
-from typing import Any
+from typing import Any, Protocol, cast
+from urllib.parse import unquote
 
 from loguru import logger
 from pydantic import Field
@@ -42,6 +44,17 @@ from nanobot.security.workspace_access import current_scope_allows_loopback, cur
 from nanobot.security.workspace_policy import is_path_within
 
 _IS_WINDOWS = sys.platform == "win32"
+_PROCESS_TREE_OWNER_ATTR = "_nanobot_process_tree_owner"
+
+
+class _ProcessTreeOwner(Protocol):
+    creation_flags: int
+
+    def assign_and_resume(self, pid: int) -> None: ...
+
+    def release(self) -> None: ...
+
+    def terminate(self) -> None: ...
 
 
 def _reap_pid(pid: int) -> None:
@@ -326,6 +339,7 @@ class ExecTool(Tool):
                 prepared.env,
                 prepared.shell_program,
                 prepared.login,
+                process_tree=True,
             )
 
             try:
@@ -334,10 +348,10 @@ class ExecTool(Tool):
                     timeout=prepared.timeout,
                 )
             except asyncio.TimeoutError:
-                await self._kill_process(process)
+                await self._kill_process_tree(process)
                 return ToolResult.error(f"Error: Command timed out after {prepared.timeout} seconds")
             except asyncio.CancelledError:
-                await self._kill_process(process)
+                await self._kill_process_tree(process)
                 raise
 
             # Safety-net reap: asyncio *should* have reaped the child via
@@ -368,13 +382,14 @@ class ExecTool(Tool):
                     + result[-half:]
                 )
 
+            self._release_process_tree(process)
             return result
 
         except Exception as e:
             # Kill and reap the child if it was spawned but an unexpected
             # error prevented communicate() from completing.
             if process is not None:
-                await self._kill_process(process)
+                await self._kill_process_tree(process)
             return ToolResult.error(f"Error executing command: {str(e)}")
 
     async def _execute_session(
@@ -537,22 +552,31 @@ class ExecTool(Tool):
     ) -> asyncio.subprocess.Process:
         """Launch *command* in a platform-appropriate shell."""
         if _IS_WINDOWS:
+            windows_job = None
+            process = None
+            creation_flags = 0
+            if process_tree and sys.platform == "win32":
+                windows_job = ExecTool._create_windows_job()
+                creation_flags = windows_job.creation_flags
             # Default to PowerShell so single-line and multi-line commands
             # share the same shell semantics.  cmd.exe is reachable via the
             # explicit shell="cmd" parameter (see _resolve_shell).
             default_program = shutil.which("pwsh") or shutil.which("powershell") or "powershell"
             program = shell_program or default_program
             program_name = PureWindowsPath(program).name.lower()
+            try:
                 if program_name in ("cmd", "cmd.exe"):
                     cmd_env = {**env, "COMSPEC": program}
-                return await asyncio.create_subprocess_shell(
+                    process = await asyncio.create_subprocess_shell(
                         command,
                         stdin=stdin,
                         stdout=asyncio.subprocess.PIPE,
                         stderr=asyncio.subprocess.PIPE,
                         cwd=cwd,
                         env=cmd_env,
+                        creationflags=creation_flags,
                     )
+                else:
                     command = ExecTool._normalize_powershell_command(command)
                     command = (
                         "[Console]::OutputEncoding = [System.Text.UTF8Encoding]::new($false)\n"
@@ -561,14 +585,25 @@ class ExecTool(Tool):
                         f"{command}\n"
                         "if ($LASTEXITCODE -ne $null) { exit $LASTEXITCODE }"
                     )
-            return await asyncio.create_subprocess_exec(
+                    process = await asyncio.create_subprocess_exec(
                         program, "-NoProfile", "-NonInteractive", "-Command", command,
                         stdin=stdin,
                         stdout=asyncio.subprocess.PIPE,
                         stderr=asyncio.subprocess.PIPE,
                         cwd=cwd,
                         env=env,
+                        creationflags=creation_flags,
                     )
+                if windows_job is not None:
+                    windows_job.assign_and_resume(process.pid)
+                    setattr(process, _PROCESS_TREE_OWNER_ATTR, windows_job)
+                return process
+            except BaseException:
+                if windows_job is not None:
+                    windows_job.terminate()
+                if process is not None:
+                    await ExecTool._kill_process(process)
+                raise
         shell_program = shell_program or shutil.which("bash") or "/bin/bash"
         args: list[str] = [shell_program]
         shell_name = Path(shell_program).name.lower()
@@ -687,11 +722,12 @@ class ExecTool(Tool):
     @staticmethod
     async def _kill_process_tree(process: asyncio.subprocess.Process) -> None:
         """Kill a session process and descendants, then reap the root process."""
-        if process.returncode is not None:
-            _reap_pid(process.pid)
-            return
+        owner = ExecTool._process_tree_owner(process)
         try:
-            if _IS_WINDOWS:
+            if owner is not None:
+                owner.terminate()
+            elif _IS_WINDOWS:
+                if process.returncode is None:
                     with suppress(OSError, asyncio.TimeoutError):
                         await asyncio.wait_for(
                             asyncio.to_thread(
@@ -715,8 +751,36 @@ class ExecTool(Tool):
             with suppress(asyncio.TimeoutError):
                 await asyncio.wait_for(process.wait(), timeout=5.0)
         finally:
+            if owner is not None:
+                ExecTool._drop_process_tree_owner(process)
             _reap_pid(process.pid)
 
+    @staticmethod
+    def _process_tree_owner(
+        process: asyncio.subprocess.Process,
+    ) -> _ProcessTreeOwner | None:
+        # _spawn is the only writer for this private ownership marker.
+        return cast(_ProcessTreeOwner | None, vars(process).get(_PROCESS_TREE_OWNER_ATTR))
+
+    @staticmethod
+    def _create_windows_job() -> _ProcessTreeOwner:
+        from nanobot.agent.tools._windows_job import WindowsJob
+
+        return WindowsJob.create()
+
+    @staticmethod
+    def _drop_process_tree_owner(process: asyncio.subprocess.Process) -> None:
+        with suppress(AttributeError):
+            delattr(process, _PROCESS_TREE_OWNER_ATTR)
+
+    @staticmethod
+    def _release_process_tree(process: asyncio.subprocess.Process) -> None:
+        owner = ExecTool._process_tree_owner(process)
+        if owner is None:
+            return
+        owner.release()
+        ExecTool._drop_process_tree_owner(process)
+
     def _build_env(self) -> dict[str, str]:
         """Build a minimal environment for subprocess execution.
 
@@ -826,12 +890,27 @@ class ExecTool(Tool):
             for raw in self._extract_absolute_paths(cmd):
                 try:
                     expanded = os.path.expandvars(raw.strip())
+                    # Python's expanduser() intentionally does not implement
+                    # shell directory-stack forms. ``~+`` is the active cwd,
+                    # while ``~-`` and indexed forms can resolve outside it;
+                    # normalize the former and fail closed on the latter.
+                    if expanded == "~+":
+                        p = cwd_path
+                    elif expanded.startswith("~+/"):
+                        p = (cwd_path / expanded[3:]).resolve()
+                    elif re.match(r"^~(?:-|[+-]\d+)(?:/|$)", expanded):
+                        return ToolResult.error(
+                            "Error: Command blocked by safety guard "
+                            "(path outside working dir)"
+                            + _WORKSPACE_BOUNDARY_NOTE
+                        )
+                    else:
+                        p = Path(expanded).expanduser().resolve()
                     # Match against the un-resolved path first.  On Linux,
                     # /dev/stderr is a symlink to /proc/self/fd/2 and
                     # ``Path.resolve()`` would mask the device-file intent.
                     if self._is_benign_device_path(expanded):
                         continue
-                    p = Path(expanded).expanduser().resolve()
                 except Exception:
                     continue
 
@@ -914,7 +993,9 @@ class ExecTool(Tool):
                 ):
                     current.append(ch)
                     operator_len = 1
-                elif ch in {";", "|"}:
+                # A newline separates commands just like ";" does, so a payload
+                # smuggled onto its own line must be checked on its own too.
+                elif ch in {";", "|", "\n", "\r"}:
                     operator_len = 1
 
             if operator_len:
@@ -948,9 +1029,134 @@ class ExecTool(Tool):
             r"(?<![A-Za-z])(?:[A-Za-z]:[^\s\"'|><;]*|\\\\[^\s\"'|><;]+(?:\\[^\s\"'|><;]+)*)",
             command
         )
-        posix_paths = re.findall(r"(?:^|[\s|>='\"])(/[^\s\"'>;|<]+)", command) # POSIX: /absolute only
-        home_paths = re.findall(r"(?:^|[\s>='\"])(~[/+][^\s\"'>;|<]*)", command) # POSIX/Windows home shortcut: ~/ or ~+
-        return win_paths + posix_paths + home_paths
+        try:
+            lexer = shlex.shlex(command, posix=True, punctuation_chars="();<>|&")
+            lexer.whitespace_split = True
+            lexer.commenters = ""
+            tokens = list(lexer)
+        except ValueError:
+            # Keep malformed quoting fail-closed. The shell will normally reject
+            # it too, but a conservative raw scan must not turn it into a bypass.
+            tokens = [command]
+
+        paths = [*win_paths]
+        seen = set(win_paths)
+        for index, token in enumerate(tokens):
+            for path in ExecTool._extract_posix_paths_from_token(token):
+                if path not in seen:
+                    paths.append(path)
+                    seen.add(path)
+            if index > 0 and tokens[index - 1] in {"-c", "-lc", "--command"}:
+                for path in ExecTool._extract_absolute_paths(token):
+                    if path not in seen:
+                        paths.append(path)
+                        seen.add(path)
+        return paths
+
+    @staticmethod
+    def _extract_posix_paths_from_token(token: str) -> list[str]:
+        """Extract local POSIX/home paths from one shell-decoded token.
+
+        ``shlex`` separates real grouping/redirection operators while preserving
+        parentheses and spaces that were quoted or escaped as part of a path.
+        Embedded scripts (for example ``sh -c \"cat /tmp/x\"``) still need a
+        small boundary scan. Colons are not general boundaries: treating them
+        as such misclassifies URLs, ``host:/remote`` and ``C:/Windows``. They
+        are considered only inside a syntactically valid assignment, where
+        shells expand each colon-delimited tilde component.
+        """
+        paths: list[str] = []
+        for match in re.finditer(
+            r"file://(?:[^/\s\"']+)?(/[^\s\"'<>|;&]*)",
+            token,
+            flags=re.IGNORECASE,
+        ):
+            uri_prefix = token[: match.start()]
+            raw_path = match.group(1)
+            if uri_prefix.count("(") > uri_prefix.count(")"):
+                raw_path = raw_path.split(")", 1)[0]
+            if uri_prefix.count("{") > uri_prefix.count("}"):
+                raw_path = raw_path.split(",", 1)[0].split("}", 1)[0]
+            raw_path = raw_path.split("?", 1)[0].split("#", 1)[0]
+            if raw_path:
+                paths.append(unquote(raw_path))
+        boundary_chars = frozenset(" \t\r\n=({,<>|;&\"'")
+        i = 0
+        while i < len(token):
+            is_posix = token[i] == "/"
+            home_match = re.match(
+                r"~(?:[+-](?:\d+)?|[A-Za-z0-9_.@-]+)?(?=/|:|$)",
+                token[i:],
+            )
+            is_home = home_match is not None
+            if not is_posix and not is_home:
+                i += 1
+                continue
+
+            prefix = token[:i]
+            parameter_default = (
+                i >= 2 and token[i - 2] == ":" and token[i - 1] in "-+?="
+            )
+            word_start = max(
+                (prefix.rfind(char) for char in " \t\r\n<>|;&"),
+                default=-1,
+            ) + 1
+            word_prefix = prefix[word_start:]
+            assignment_component = bool(
+                re.fullmatch(
+                    r"(?:[A-Za-z_][A-Za-z0-9_]*|--?[A-Za-z0-9_.-]+)="
+                    r"(?:[^:=\s]*:)*",
+                    word_prefix,
+                )
+            )
+            at_boundary = i == 0 or token[i - 1] in boundary_chars
+            if is_home:
+                # A shell word beginning with ``~`` is a separate shlex token.
+                # Mid-token expansion is valid only after ``=`` or a colon in
+                # an assignment. This avoids PromQL/Loki ``=~`` and ``|~``
+                # match operators while covering PATH-like values.
+                at_boundary = i == 0 or assignment_component
+            if not at_boundary and not parameter_default:
+                i += 1
+                continue
+
+            if re.search(r"[A-Za-z][A-Za-z0-9+.-]*://", word_prefix) or re.match(
+                r"(?:[^/:=\s]+@)?[^/:=\s]+:$",
+                word_prefix,
+            ):
+                # HTTP-style URL path/query fragments and scp-style remote paths
+                # are not local filesystem references. ``file://`` paths were
+                # decoded above. Windows drive paths are already captured by the
+                # platform-specific expression above.
+                i += 1
+                continue
+
+            assignment_value = assignment_component
+            if i == 0 or assignment_value:
+                end = len(token)
+                if assignment_value:
+                    separator = token.find(":", i)
+                    if separator >= 0:
+                        end = separator
+            elif token[i - 1] in {"'", '"'}:
+                quote = token[i - 1]
+                closing = token.find(quote, i)
+                end = len(token) if closing < 0 else closing
+            else:
+                end_chars = set(" \t\r\n\"'<>|;&")
+                if prefix.count("(") > prefix.count(")"):
+                    end_chars.add(")")
+                if prefix.count("{") > prefix.count("}"):
+                    end_chars.update({",", "}"})
+                end = i
+                while end < len(token) and token[end] not in end_chars:
+                    end += 1
+
+            candidate = token[i:end]
+            if candidate:
+                paths.append(candidate)
+            i = max(end, i + 1)
+        return paths
 
     @staticmethod
     def _normalize_bind_roots(paths: list[str] | None) -> list[Path]:

nanobot/agent/tools/web.py

@@ -11,7 +11,7 @@ import os
 import re
 from collections.abc import Callable
 from typing import Any, cast
-from urllib.parse import quote, urljoin, urlparse
+from urllib.parse import parse_qsl, quote, urljoin, urlparse
 
 import httpx
 from loguru import logger
@@ -148,6 +148,59 @@ def _unsafe_url_request_error(exc: BaseException) -> str | None:
     return str(exc) if isinstance(exc, UnsafeURLRequestError) else None
 
 
+# Forwarding a URL to the remote Jina reader discloses it to a third party, so
+# URLs that embed credential material (userinfo, signed-URL parameters, token
+# or key query values) must never leave the machine. Matching is by parameter
+# name: over-matching only costs the local readability fallback, while
+# under-matching leaks a secret.
+_CREDENTIAL_QUERY_PARAMS = frozenset({
+    "access_token", "api-key", "api-token", "apikey", "api_key", "api_token",
+    "auth", "authorization", "client_assertion", "client_secret", "code",
+    "credential", "credentials", "id_token", "jwt", "key", "password",
+    "passwd", "private_key", "pwd", "refresh_token", "samlresponse", "secret",
+    "session_id", "session_token", "sessionid", "sig", "signature", "sso_token",
+    "ticket", "token",
+})
+_CREDENTIAL_QUERY_PREFIXES = ("x-amz-", "x-goog-")
+
+
+def _url_carries_credentials(url: str) -> bool:
+    try:
+        parsed = urlparse(url)
+    except ValueError:
+        return True
+    if parsed.username is not None or parsed.password is not None:
+        return True
+    # Some frameworks still accept semicolons as query separators. Treating
+    # them as separators here may over-match a value, but the safe consequence
+    # is only using the local extractor instead of disclosing a credential.
+    query = parsed.query.replace(";", "&")
+    for name, _value in parse_qsl(query, keep_blank_values=True):
+        lowered = name.strip().lower()
+        if lowered in _CREDENTIAL_QUERY_PARAMS or lowered.startswith(_CREDENTIAL_QUERY_PREFIXES):
+            return True
+    return False
+
+
+def _redact_url_for_log(url: str) -> str:
+    """Return only a URL's origin, excluding userinfo, path, query, and fragment."""
+    try:
+        parsed = urlparse(url)
+        hostname = parsed.hostname
+        if not parsed.scheme or hostname is None:
+            return "<redacted URL>"
+        if ":" in hostname:
+            hostname = f"[{hostname}]"
+        try:
+            port = parsed.port
+        except ValueError:
+            port = None
+        authority = f"{hostname}:{port}" if port is not None else hostname
+        return f"{parsed.scheme}://{authority}"
+    except ValueError:
+        return "<redacted URL>"
+
+
 async def _get_with_safe_redirects(
     client: httpx.AsyncClient,
     url: str,
@@ -191,13 +244,14 @@ async def _stream_with_safe_redirects(
     client: httpx.AsyncClient,
     url: str,
     headers: dict[str, str] | None = None,
-) -> tuple[httpx.Response | None, Any | None, str | None]:
+) -> tuple[httpx.Response | None, Any | None, str | None, bool]:
     """Open a streamed response while validating every redirect target first."""
     current_url = url
+    chain_carries_credentials = _url_carries_credentials(url)
     for _ in range(MAX_REDIRECTS + 1):
         is_valid, error_msg, _ = _resolve_url_safe(current_url)
         if not is_valid:
-            return None, None, f"Redirect blocked: {error_msg}"
+            return None, None, f"Redirect blocked: {error_msg}", chain_carries_credentials
 
         stream = client.stream(
             "GET",
@@ -210,26 +264,39 @@ async def _stream_with_safe_redirects(
         except httpx.RequestError as exc:
             unsafe_error = _unsafe_url_request_error(exc)
             if unsafe_error is not None:
-                return None, None, f"Redirect blocked: {unsafe_error}"
+                return (
+                    None,
+                    None,
+                    f"Redirect blocked: {unsafe_error}",
+                    chain_carries_credentials,
+                )
             raise
         is_redirect = 300 <= response.status_code < 400
         if not is_redirect:
-            return response, stream, None
+            return response, stream, None, chain_carries_credentials
 
         location = response.headers.get("location")
         if not location:
-            return response, stream, None
+            return response, stream, None, chain_carries_credentials
 
         next_url = urljoin(str(response.url), location)
+        chain_carries_credentials = (
+            chain_carries_credentials or _url_carries_credentials(next_url)
+        )
         is_valid, error_msg = _validate_url_safe(next_url)
         if not is_valid:
             await stream.__aexit__(None, None, None)
-            return None, None, f"Redirect blocked: {error_msg}"
+            return None, None, f"Redirect blocked: {error_msg}", chain_carries_credentials
 
         await stream.__aexit__(None, None, None)
         current_url = next_url
 
-    return None, None, f"Too many redirects: exceeded limit of {MAX_REDIRECTS}"
+    return (
+        None,
+        None,
+        f"Too many redirects: exceeded limit of {MAX_REDIRECTS}",
+        chain_carries_credentials,
+    )
 
 
 def _format_results(query: str, items: list[dict[str, Any]], n: int) -> str:
@@ -1043,20 +1110,26 @@ class WebFetchTool(Tool):
         if not is_valid:
             return json.dumps({"error": f"URL validation failed: {error_msg}", "url": url}, ensure_ascii=False)
 
-        # Detect and fetch images directly to avoid Jina's textual image captioning
+        # Detect and fetch images directly to avoid Jina's textual image captioning.
+        # This local preflight also proves that no credential-bearing URL occurs
+        # in the redirect chain before the original URL may be sent to Jina.
+        jina_remote_safe = False
         try:
             async with httpx.AsyncClient(
                 **_fetch_client_kwargs(self.proxy, 15.0),
             ) as client:
-                r, stream, redirect_error = await _stream_with_safe_redirects(
+                r, stream, redirect_error, chain_carries_credentials = (
+                    await _stream_with_safe_redirects(
                         client,
                         url,
                         headers={"User-Agent": self.user_agent},
                     )
+                )
                 if redirect_error:
                     return json.dumps({"error": redirect_error, "url": url}, ensure_ascii=False)
                 if r is None:
                     return json.dumps({"error": "Fetch failed", "url": url}, ensure_ascii=False)
+                jina_remote_safe = not chain_carries_credentials
 
                 try:
                     ctype = r.headers.get("content-type", "")
@@ -1071,10 +1144,14 @@ class WebFetchTool(Tool):
             unsafe_error = _unsafe_url_request_error(e)
             if unsafe_error is not None:
                 return json.dumps({"error": f"URL validation failed: {unsafe_error}", "url": url}, ensure_ascii=False)
-            logger.debug("Pre-fetch image detection failed for {}: {}", url, e)
+            logger.debug(
+                "Pre-fetch image detection failed for {} ({})",
+                _redact_url_for_log(url),
+                type(e).__name__,
+            )
 
         result = None
-        if self.config.use_jina_reader:
+        if self.config.use_jina_reader and jina_remote_safe:
             result = await self._fetch_jina(url, max_chars)
         if result is None:
             result = await self._fetch_readability(url, extract_mode, max_chars)
@@ -1082,13 +1159,23 @@ class WebFetchTool(Tool):
 
     async def _fetch_jina(self, url: str, max_chars: int) -> str | None:
         """Try fetching via Jina Reader API. Returns None on failure."""
+        if _url_carries_credentials(url):
+            logger.debug(
+                "Skipping Jina Reader for {}: URL carries credential material",
+                _redact_url_for_log(url),
+            )
+            return None
+        # httpx already drops the fragment when building the request; strip it
+        # explicitly so client-side-only data (OAuth implicit flows put tokens
+        # there) stays out of this path even if the transport changes.
+        forwarded_url = url.split("#", 1)[0]
         try:
             headers = {"Accept": "application/json", "User-Agent": self.user_agent}
             jina_key = os.environ.get("JINA_API_KEY", "")
             if jina_key:
                 headers["Authorization"] = f"Bearer {jina_key}"
             async with httpx.AsyncClient(proxy=self.proxy, timeout=20.0) as client:
-                r = await client.get(f"https://r.jina.ai/{url}", headers=headers)
+                r = await client.get(f"https://r.jina.ai/{forwarded_url}", headers=headers)
                 if r.status_code == 429:
                     logger.debug("Jina Reader rate limited, falling back to readability")
                     return None
@@ -1113,7 +1200,11 @@ class WebFetchTool(Tool):
                 "untrusted": True, "text": text,
             }, ensure_ascii=False)
         except Exception as e:
-            logger.debug("Jina Reader failed for {}, falling back to readability: {}", url, e)
+            logger.debug(
+                "Jina Reader failed for {}, falling back to readability ({})",
+                _redact_url_for_log(url),
+                type(e).__name__,
+            )
             return None
 
     async def _fetch_readability(self, url: str, extract_mode: str, max_chars: int) -> Any:
@@ -1144,7 +1235,11 @@ class WebFetchTool(Tool):
                     text = self._extract_readable_html(r.text, extract_mode)
                     extractor = "readability"
                 except Exception as e:
-                    logger.warning("Readability failed for {}, using raw HTML fallback: {}", url, e)
+                    logger.warning(
+                        "Readability failed for {}, using raw HTML fallback ({})",
+                        _redact_url_for_log(url),
+                        type(e).__name__,
+                    )
                     text, extractor = _normalize(_strip_tags(r.text)), "html"
             else:
                 text, extractor = r.text, "raw"
@@ -1160,10 +1255,18 @@ class WebFetchTool(Tool):
                 "untrusted": True, "text": text,
             }, ensure_ascii=False)
         except httpx.ProxyError as e:
-            logger.exception("WebFetch proxy error for {}", url)
+            logger.warning(
+                "WebFetch proxy error for {} ({})",
+                _redact_url_for_log(url),
+                type(e).__name__,
+            )
             return json.dumps({"error": f"Proxy error: {e}", "url": url}, ensure_ascii=False)
         except Exception as e:
-            logger.exception("WebFetch error for {}", url)
+            logger.warning(
+                "WebFetch error for {} ({})",
+                _redact_url_for_log(url),
+                type(e).__name__,
+            )
             return json.dumps({"error": str(e), "url": url}, ensure_ascii=False)
 
     def _extract_readable_html(self, html_content: str, extract_mode: str) -> str:

nanobot/agent/turn_delivery.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 
 import dataclasses
 import time
-from collections.abc import Awaitable, Callable
+from collections.abc import Awaitable, Callable, Mapping
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any, cast
 
@@ -189,12 +189,23 @@ class TurnDelivery:
                 started_at=started_at,
             )
 
-    def record_runtime(self, runtime: LLMRuntime) -> None:
+    async def runtime_admitted(self, runtime: LLMRuntime) -> None:
+        """Record the immutable runtime and expose it at the lifecycle seam."""
+        if self.route.publish_lifecycle:
+            await self.runtime_event_publisher.turn_runtime_admitted(
+                self.delivery_message,
+                self.session_key,
+                runtime,
+            )
+            return
         self.runtime_event_publisher.record_turn_runtime(self.session_key, runtime)
 
     def record_latency(self, latency_ms: int | None) -> None:
         self.runtime_event_publisher.record_turn_latency(self.session_key, latency_ms)
 
+    def record_usage(self, usage: Mapping[str, int]) -> None:
+        self.runtime_event_publisher.record_turn_usage(self.session_key, usage)
+
     def background_response(
         self,
         content: str | None,

nanobot/api/server.py

@@ -48,6 +48,7 @@ _AGENT_LOOP_KEY = web.AppKey[Any]("agent_loop")
 _MODEL_NAME_KEY = web.AppKey[str]("model_name")
 _REQUEST_TIMEOUT_KEY = web.AppKey[float]("request_timeout")
 _SESSION_LOCKS_KEY = web.AppKey[dict[str, asyncio.Lock]]("session_locks")
+_PREPARE_AGENT_KEY = web.AppKey[Callable[[], Awaitable[None]] | None]("prepare_agent")
 _MISSING = object()
 
 
@@ -66,6 +67,17 @@ def _app_value(
         return app.get(legacy_key, default)
 
 
+async def _prepare_agent(app: Any) -> None:
+    prepare: Callable[[], Awaitable[None]] | None = _app_value(
+        app,
+        _PREPARE_AGENT_KEY,
+        "prepare_agent",
+        None,
+    )
+    if prepare is not None:
+        await prepare()
+
+
 # ---------------------------------------------------------------------------
 # Response helpers
 # ---------------------------------------------------------------------------
@@ -346,8 +358,9 @@ async def handle_chat_completions(request: web.Request) -> web.Response | web.St
             nonlocal stream_failed
             try:
                 async with session_lock:
-                    response = await asyncio.wait_for(
-                        agent_loop.process_direct(
+                    async with asyncio.timeout(timeout_s):
+                        await _prepare_agent(request.app)
+                        response = await agent_loop.process_direct(
                             content=text,
                             media=media_paths if media_paths else None,
                             session_key=session_key,
@@ -355,8 +368,6 @@ async def handle_chat_completions(request: web.Request) -> web.Response | web.St
                             chat_id=API_CHAT_ID,
                             on_stream=_on_stream,
                             on_stream_end=_on_stream_end,
-                        ),
-                        timeout=timeout_s,
                         )
                     if not emitted_content:
                         response_text = _response_text(response)
@@ -390,15 +401,14 @@ async def handle_chat_completions(request: web.Request) -> web.Response | web.St
     try:
         async with session_lock:
             try:
-                response = await asyncio.wait_for(
-                    agent_loop.process_direct(
+                async with asyncio.timeout(timeout_s):
+                    await _prepare_agent(request.app)
+                    response = await agent_loop.process_direct(
                         content=text,
                         media=media_paths if media_paths else None,
                         session_key=session_key,
                         channel="api",
                         chat_id=API_CHAT_ID,
-                    ),
-                    timeout=timeout_s,
                     )
                 response_text = _response_text(response)
                 if not response_text or not response_text.strip():
@@ -452,6 +462,7 @@ def create_app(
     model_name: str = "nanobot",
     request_timeout: float = 120.0,
     api_key: str = "",
+    prepare_agent: Callable[[], Awaitable[None]] | None = None,
 ) -> web.Application:
     """Create the aiohttp application.
 
@@ -460,12 +471,14 @@ def create_app(
         model_name: Model name reported in responses.
         request_timeout: Per-request timeout in seconds.
         api_key: Optional API key for Bearer-token authentication on API routes.
+        prepare_agent: Optional application-owned readiness callback run before each turn.
     """
     app = web.Application(client_max_size=20 * 1024 * 1024)  # 20MB for base64 images
     app[_AGENT_LOOP_KEY] = agent_loop
     app[_MODEL_NAME_KEY] = model_name
     app[_REQUEST_TIMEOUT_KEY] = request_timeout
     app[_SESSION_LOCKS_KEY] = {}  # per-user locks, keyed by session_key
+    app[_PREPARE_AGENT_KEY] = prepare_agent
 
     @web.middleware
     async def auth_middleware(

nanobot/apps/cli/service.py

@@ -20,6 +20,7 @@ from urllib.parse import urlparse
 import httpx
 from loguru import logger
 
+from nanobot.agent.skills import parse_skill_metadata, valid_skill_metadata
 from nanobot.apps.protocol import app_manifest, compact_dict
 from nanobot.config.paths import get_runtime_subdir
 from nanobot.security.workspace_policy import is_path_within
@@ -27,6 +28,7 @@ from nanobot.security.workspace_policy import is_path_within
 CLI_ANYTHING_REGISTRY_URL = "https://hkuds.github.io/CLI-Anything/registry.json"
 CLI_ANYTHING_PUBLIC_REGISTRY_URL = "https://hkuds.github.io/CLI-Anything/public_registry.json"
 CLI_ANYTHING_RAW_BASE = "https://raw.githubusercontent.com/HKUDS/CLI-Anything/main"
+AGENT_PLUGIN_SCHEMA = "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json"
 NANOBOT_EXTENSION_REGISTRY_URL = "https://raw.githubusercontent.com/Re-bin/nanobot-extension/main/registry.json"
 NANOBOT_EXTENSION_RAW_BASE = "https://raw.githubusercontent.com/Re-bin/nanobot-extension/main"
 _CATALOG_SOURCES = (
@@ -210,11 +212,27 @@ def _as_object_dict(value: object) -> dict[str, Any] | None:
     return cast(dict[str, Any], value) if isinstance(value, dict) else None
 
 
-def _safe_skill_name(name: str) -> str:
+def _skill_name(name: str, *, legacy: bool = False) -> str:
     clean = _SAFE_NAME_RE.sub("-", name.lower()).strip("-")
+    if not legacy:
+        clean = clean.replace("_", "-")
     return f"cli-app-{clean or 'app'}"
 
 
+def _plugin_skill_relative_path(name: str) -> str:
+    skill_name = _skill_name(name)
+    return f"plugins/{skill_name}/skills/{skill_name}/SKILL.md"
+
+
+def cli_app_skill_relative_path(workspace: Path, name: str) -> str:
+    """Return a CLI App's skill path, including the legacy location."""
+    canonical = _plugin_skill_relative_path(name)
+    legacy = f"skills/{_skill_name(name, legacy=True)}/SKILL.md"
+    if not (workspace / canonical).is_file() and (workspace / legacy).is_file():
+        return legacy
+    return canonical
+
+
 def _has_shell_meta(command: str) -> bool:
     return any(char in command for char in _SHELL_META_CHARS)
 
@@ -442,6 +460,16 @@ class CliAppManager:
         """Return registry names explicitly installed through CLI Apps."""
         return sorted(str(name) for name in self._load_installed())
 
+    def installed_skill_aliases(self) -> dict[str, str]:
+        """Map pre-plugin CLI App skill names to their portable identities."""
+        aliases: dict[str, str] = {}
+        for name in self.installed_names():
+            legacy = _skill_name(name, legacy=True)
+            canonical = _skill_name(name)
+            if legacy != canonical:
+                aliases[legacy] = canonical
+        return aliases
+
     def _fetch_registry(
         self,
         url: str,
@@ -613,7 +641,7 @@ class CliAppManager:
                     "name": installed_name,
                     "entry_point": entry_point,
                     "source": str(data.get("source") or ""),
-                    "skill": f"skills/{_safe_skill_name(installed_name)}/SKILL.md",
+                    "skill": cli_app_skill_relative_path(self.workspace, installed_name),
                     "tool": "run_cli_app",
                 }
             )
@@ -639,9 +667,6 @@ class CliAppManager:
         install_cmd = str(app.get("install_cmd") or "")
         return not _has_shell_meta(install_cmd)
 
-    def _skill_path(self, name: str) -> Path:
-        return self.workspace / "skills" / _safe_skill_name(name) / "SKILL.md"
-
     def _app_payload(
         self,
         app: dict[str, Any],
@@ -677,7 +702,7 @@ class CliAppManager:
             "status": status,
             "logo_url": logo_url,
             "brand_color": brand_color,
-            "skill_installed": self._skill_path(name).is_file(),
+            "skill_installed": (self.workspace / cli_app_skill_relative_path(self.workspace, name)).is_file(),
             "manifest": self._manifest_payload(app, logo_url=logo_url, brand_color=brand_color),
         }
 
@@ -713,7 +738,8 @@ class CliAppManager:
         name = str(app["name"])
         entry_point = str(app.get("entry_point") or "")
         strategy = self._strategy(app)
-        skill_path = f"skills/{_safe_skill_name(name)}/SKILL.md"
+        skill_path = _plugin_skill_relative_path(name)
+        plugin_path = f"plugins/{_skill_name(name)}"
         capabilities = [
             compact_dict({
                 "type": "cli",
@@ -726,13 +752,13 @@ class CliAppManager:
         install = compact_dict({
             "supported": install_supported,
             "strategy": strategy,
-            "managed_paths": [skill_path],
+            "managed_paths": [plugin_path],
             "verification": ["entry_point_available"] if entry_point else [],
         })
         remove = compact_dict({
             "supported": strategy != "unsupported",
             "strategy": strategy,
-            "managed_paths": [skill_path],
+            "managed_paths": [plugin_path],
             "verification": (
                 ["package_manager_ok", "entry_point_absent", "managed_paths_absent"]
                 if strategy not in {"bundled", "unsupported"}
@@ -964,6 +990,35 @@ class CliAppManager:
             return None
         raise CliAppError("this CLI app uses an unsupported install strategy")
 
+    def _subprocess_env(self) -> dict[str, str]:
+        """Minimal env for CLI app subprocesses — no API keys or secrets.
+
+        Mirrors the shell tool's allowlist so installed apps cannot read
+        provider credentials from the parent process environment.
+        """
+        if sys.platform == "win32":
+            sr = os.environ.get("SYSTEMROOT", r"C:\Windows")
+            env = {
+                "SYSTEMROOT": sr,
+                "COMSPEC": os.environ.get("COMSPEC", f"{sr}\\system32\\cmd.exe"),
+                "USERPROFILE": os.environ.get("USERPROFILE", ""),
+                "HOMEDRIVE": os.environ.get("HOMEDRIVE", "C:"),
+                "HOMEPATH": os.environ.get("HOMEPATH", "\\"),
+                "TEMP": os.environ.get("TEMP", f"{sr}\\Temp"),
+                "TMP": os.environ.get("TMP", f"{sr}\\Temp"),
+                "PATHEXT": os.environ.get("PATHEXT", ".COM;.EXE;.BAT;.CMD"),
+                "PATH": os.environ.get("PATH", f"{sr}\\system32;{sr}"),
+                "PYTHONUNBUFFERED": "1",
+            }
+            return env
+        return {
+            "HOME": os.environ.get("HOME", "/tmp"),
+            "LANG": os.environ.get("LANG", "C.UTF-8"),
+            "TERM": os.environ.get("TERM", "dumb"),
+            "PATH": os.environ.get("PATH", "/usr/bin:/bin"),
+            "PYTHONUNBUFFERED": "1",
+        }
+
     def _run_argv(self, argv: list[str], *, timeout: int) -> subprocess.CompletedProcess[str]:
         command = subprocess.list2cmdline(argv)
         logger.info("CLI Apps: running {}", command)
@@ -974,6 +1029,7 @@ class CliAppManager:
             encoding="utf-8",
             errors="replace",
             timeout=timeout,
+            env=self._subprocess_env(),
         )
         logger.info("CLI Apps: command exited with code {}: {}", result.returncode, command)
         output = (result.stderr or result.stdout or "").strip()
@@ -1032,11 +1088,10 @@ class CliAppManager:
         name = str(app.get("name") or "unknown")
         display = str(app.get("display_name") or name)
         entry = str(app.get("entry_point") or f"cli-anything-{name}")
-        description = _catalog_description(app) or f"Use {display} from nanobot."
+        description = (_catalog_description(app) or f"Use {display} from nanobot.")[:1024]
         return f"""---
-name: {_safe_skill_name(name)}
-description: >-
-  {description}
+name: {_skill_name(name)}
+description: {json.dumps(description, ensure_ascii=False)}
 ---
 
 # {display}
@@ -1056,10 +1111,17 @@ Prefer machine-readable output when the CLI supports `--json`.
 """
 
     def _with_nanobot_skill_note(self, content: str, app: dict[str, Any]) -> str:
+        name = str(app.get("name") or "unknown")
+        skill_name = _skill_name(name)
+        metadata = parse_skill_metadata(content)
+        if metadata is None or not valid_skill_metadata(metadata | {"name": skill_name}, skill_name):
+            content = self._fallback_skill(app)
+        content, replaced = re.subn(r"(?m)^name\s*:.*$", f"name: {skill_name}", content, count=1)
+        if not replaced:
+            content = content.replace("---\n", f"---\nname: {skill_name}\n", 1)
         marker = "<!-- nanobot-cli-app-note -->"
         if marker in content:
             return content
-        name = str(app.get("name") or "unknown")
         note = f"""{marker}
 ## Nanobot execution
 
@@ -1073,24 +1135,42 @@ Use the `run_cli_app` tool with `name="{name}"` for command execution. Do not in
         return note + "\n" + content
 
     def install_skill(self, app: dict[str, Any]) -> Path:
-        path = self._skill_path(str(app["name"]))
+        name = str(app["name"])
+        path = self.workspace / _plugin_skill_relative_path(name)
         path.parent.mkdir(parents=True, exist_ok=True)
         content = self._fetch_skill_content(app) or self._fallback_skill(app)
         content = self._with_nanobot_skill_note(content, app)
         path.write_text(content, encoding="utf-8")
+        plugin_root = path.parents[2]
+        manifest = compact_dict({
+            "$schema": AGENT_PLUGIN_SCHEMA,
+            "name": _skill_name(str(app["name"])),
+            "version": str(app.get("version") or ""),
+            "description": _catalog_description(app),
+        })
+        _write_json(plugin_root / "plugin.json", manifest)
+        legacy_dir = self.workspace / "skills" / _skill_name(str(app["name"]), legacy=True)
+        if legacy_dir.is_dir():
+            shutil.rmtree(legacy_dir)
         return path
 
     def remove_skill(self, name: str) -> None:
-        skill_dir = self._skill_path(name).parent
-        if skill_dir.is_dir():
-            shutil.rmtree(skill_dir)
+        plugin_root = (self.workspace / _plugin_skill_relative_path(name)).parents[2]
+        if plugin_root.is_dir():
+            shutil.rmtree(plugin_root)
+        legacy_dir = self.workspace / "skills" / _skill_name(name, legacy=True)
+        if legacy_dir.is_dir():
+            shutil.rmtree(legacy_dir)
 
     def _record_installed(self, app: dict[str, Any]) -> dict[str, Any]:
+        from nanobot.agent.plugins import set_agent_plugin_enabled
+
         installed = self._load_installed()
         entry = self._installed_entry(app)
         installed[str(app["name"])] = entry
         self._save_installed(installed)
         self.install_skill(app)
+        set_agent_plugin_enabled(self.workspace, _skill_name(str(app["name"])), True)
         return entry
 
     def install(self, name: str) -> dict[str, Any]:
@@ -1381,7 +1461,7 @@ Use the `run_cli_app` tool with `name="{name}"` for command execution. Do not in
                 encoding="utf-8",
                 errors="replace",
                 timeout=effective_timeout,
-                env=os.environ.copy(),
+                env=self._subprocess_env(),
             )
         except subprocess.TimeoutExpired:
             return f"CLI app '{name}' timed out after {effective_timeout}s"

nanobot/apps/cli/utils.py

@@ -20,6 +20,8 @@ def runtime_lines_for_request(
     """Return CLI App annotations from an immutable request snapshot."""
     structured = metadata.get("cli_apps") if isinstance(metadata, Mapping) else None
     if isinstance(structured, list):
+        from nanobot.apps.cli.service import cli_app_skill_relative_path
+
         structured_items = cast(list[Any], structured)
         mentions = [
             cast(Mapping[str, Any], item) for item in structured_items
@@ -32,7 +34,7 @@ def runtime_lines_for_request(
                 f"@{str(item['name']).strip().lower()} "
                 f"(installed; tool=run_cli_app; "
                 f"entry_point={str(item.get('entry_point') or 'unknown')}; "
-                f"skill=skills/cli-app-{str(item['name']).strip().lower()}/SKILL.md). "
+                f"skill={cli_app_skill_relative_path(workspace, str(item['name']))}). "
                 "Read the skill when useful, then run this app with `run_cli_app`; do not bypass it with shell."
                 for item in mentions
                 if str(item.get("name") or "").strip()

nanobot/bus/events.py

@@ -12,11 +12,11 @@ if TYPE_CHECKING:
 # render it and other channels may ignore unknown keys.
 OUTBOUND_META_AGENT_UI = "_agent_ui"
 
-# Internal-only inbound metadata used by in-process channels to ask the agent
-# loop to update runtime state without going through a user session.
+# Internal-only inbound metadata minted by trusted transports and runtime
+# services. Never accept these keys verbatim from an untrusted client.
 INBOUND_META_RUNTIME_CONTROL = "_runtime_control"
+INBOUND_META_USER_SHELL = "_user_shell"
 RUNTIME_CONTROL_ACK = "_ack"
-RUNTIME_CONTROL_MCP_RELOAD = "mcp_reload"
 RUNTIME_CONTROL_IMAGE_GENERATION_RELOAD = "image_generation_reload"
 RUNTIME_CONTROL_SESSION_DISCARD = "session_discard"
 

nanobot/bus/outbound_events.py

@@ -58,6 +58,8 @@ class StreamedResponseEvent(OutboundEvent):
 class TurnEndEvent(OutboundEvent):
     latency_ms: int | None = None
     goal_state: dict[str, Any] | None = None
+    usage: dict[str, int] | None = None
+    context_window_tokens: int | None = None
 
 
 @dataclass(frozen=True)
@@ -84,9 +86,11 @@ class RuntimeModelUpdatedEvent(OutboundEvent):
 
 @dataclass(frozen=True)
 class TurnModelUpdatedEvent(OutboundEvent):
-    """The fallback model currently handling one chat turn."""
+    """The canonical preset and concrete model handling one chat turn."""
 
     model: str
+    model_preset: str | None = None
+    context_window_tokens: int | None = None
 
 
 def outbound_message_for_event(
@@ -171,6 +175,12 @@ def _legacy_event_from_metadata(msg: OutboundMessage) -> OutboundEvent | None:
         return TurnEndEvent(
             latency_ms=_metadata_int(meta, "latency_ms"),
             goal_state=cast(dict[str, Any], goal_state) if isinstance(goal_state, dict) else None,
+            usage=(
+                cast(dict[str, int], meta.get("usage"))
+                if isinstance(meta.get("usage"), dict)
+                else None
+            ),
+            context_window_tokens=_metadata_int(meta, "context_window_tokens"),
         )
     if meta.get("_session_updated"):
         return SessionUpdatedEvent(scope=_metadata_str(meta, "_session_update_scope"))

nanobot/bus/runtime_events.py

@@ -10,7 +10,7 @@ from __future__ import annotations
 import asyncio
 import contextlib
 import inspect
-from collections.abc import Awaitable, Callable
+from collections.abc import Awaitable, Callable, Mapping
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
 
@@ -40,6 +40,14 @@ class SessionTurnStarted:
     context: RuntimeEventContext
 
 
+@dataclass(frozen=True)
+class TurnRuntimeAdmitted:
+    """The immutable model runtime selected for one admitted turn."""
+
+    context: RuntimeEventContext
+    runtime: LLMRuntime
+
+
 @dataclass(frozen=True)
 class TurnRunStatusChanged:
     """Visible run status changed for a turn."""
@@ -56,6 +64,7 @@ class TurnCompleted:
     context: RuntimeEventContext
     latency_ms: int | None = None
     runtime: LLMRuntime | None = None
+    usage: dict[str, int] = field(default_factory=dict)
 
 
 @dataclass(frozen=True)
@@ -85,6 +94,7 @@ class RuntimeModelChanged:
 
 RuntimeEvent = (
     SessionTurnStarted
+    | TurnRuntimeAdmitted
     | SessionTurnPersisted
     | TurnRunStatusChanged
     | TurnCompleted
@@ -93,6 +103,7 @@ RuntimeEvent = (
 )
 RuntimeEventType = (
     type[SessionTurnStarted]
+    | type[TurnRuntimeAdmitted]
     | type[SessionTurnPersisted]
     | type[TurnRunStatusChanged]
     | type[TurnCompleted]
@@ -159,6 +170,7 @@ class RuntimeEventPublisher:
         self.bus = bus or RuntimeEventBus()
         self._turn_latency_ms: dict[str, int] = {}
         self._turn_runtime: dict[str, LLMRuntime] = {}
+        self._turn_usage: dict[str, dict[str, int]] = {}
 
     @staticmethod
     def _context(
@@ -184,9 +196,17 @@ class RuntimeEventPublisher:
         if latency_ms is not None:
             self._turn_latency_ms[session_key] = int(latency_ms)
 
+    def record_turn_usage(self, session_key: str, usage: Mapping[str, int]) -> None:
+        self._turn_usage[session_key] = {
+            key: int(value)
+            for key, value in usage.items()
+            if type(value) is int and value >= 0
+        }
+
     def clear_turn(self, session_key: str) -> None:
         self._turn_latency_ms.pop(session_key, None)
         self._turn_runtime.pop(session_key, None)
+        self._turn_usage.pop(session_key, None)
 
     async def session_turn_started(
         self,
@@ -204,6 +224,26 @@ class RuntimeEventPublisher:
             )
         )
 
+    async def turn_runtime_admitted(
+        self,
+        msg: InboundMessage,
+        session_key: str,
+        runtime: LLMRuntime,
+    ) -> None:
+        """Record and publish the runtime selected for one turn."""
+        self.record_turn_runtime(session_key, runtime)
+        await self.bus.publish(
+            TurnRuntimeAdmitted(
+                context=self._context(
+                    channel=msg.channel,
+                    chat_id=msg.chat_id,
+                    session_key=session_key,
+                    metadata=msg.metadata,
+                ),
+                runtime=runtime,
+            )
+        )
+
     async def run_status_changed(
         self,
         msg: InboundMessage,
@@ -265,6 +305,7 @@ class RuntimeEventPublisher:
                 ),
                 latency_ms=self._turn_latency_ms.pop(session_key, None),
                 runtime=self._turn_runtime.pop(session_key, None),
+                usage=self._turn_usage.pop(session_key, {}),
             )
         )
 

nanobot/channels/feishu/webui/FeishuAssistantsPanel.tsx

@@ -15,6 +15,7 @@ import type {
   NanobotFeatureInfo,
   NanobotFeaturesPayload,
 } from "@/lib/types";
+import { useClient } from "@/providers/ClientProvider";
 
 import { FeishuConnectFlow } from "./FeishuConnectFlow";
 
@@ -33,7 +34,6 @@ export function FeishuAssistantsPanel({
 
   return (
     <ChannelInstancesPanel
-      token={token}
       feature={feature}
       showBrandLogos={showBrandLogos}
       chatAppsDocsUrl={chatAppsDocsUrl}
@@ -59,7 +59,7 @@ export function FeishuAssistantsPanel({
           />
         ),
         footer: (
-          <div className="mt-4 overflow-hidden rounded-[16px] border border-border/70 bg-background px-4 py-4">
+          <div className="mt-4 overflow-hidden rounded-floating border border-border/70 bg-background px-4 py-4">
             <div className="text-[13px] font-semibold text-foreground">
               {tx("custom.createAnother", "Create another assistant")}
             </div>
@@ -92,6 +92,7 @@ function FeishuInstanceAction({
   instance: NanobotChannelInstanceInfo;
   onFeaturesUpdate: (payload: NanobotFeaturesPayload) => void;
 }) {
+  const { client } = useClient();
   const { t } = useTranslation();
   const tx = channelTranslator(t, "feishu");
   const [busy, setBusy] = useState(false);
@@ -114,7 +115,7 @@ function FeishuInstanceAction({
     setError(null);
     try {
       onFeaturesUpdate(
-        await enableNanobotFeature(token, "feishu", { instanceId: instance.id }),
+        await enableNanobotFeature(client, "feishu", { instanceId: instance.id }),
       );
     } catch (err) {
       setError((err as Error).message);
@@ -143,7 +144,7 @@ function FeishuInstanceAction({
         </Button>
       </div>
       {error ? (
-        <div className="mt-3 rounded-[12px] border border-destructive/20 px-3 py-2 text-[12px] leading-5 text-destructive">
+        <div className="mt-3 rounded-control border border-destructive/20 px-3 py-2 text-[12px] leading-5 text-destructive">
           {error}
         </div>
       ) : null}

nanobot/channels/feishu/webui/index.tsx

@@ -1,7 +1,13 @@
+import { lazy } from "react";
+
 import type { ChannelUiContribution } from "@/channel-plugins/types";
 import { chatAppGuideUrl } from "@/components/settings/channels/catalog";
 
-import { FeishuAssistantsPanel } from "./FeishuAssistantsPanel";
+const FeishuAssistantsPanel = lazy(() =>
+  import("./FeishuAssistantsPanel").then(({ FeishuAssistantsPanel: component }) => ({
+    default: component,
+  })),
+);
 
 export default {
   Panel: FeishuAssistantsPanel,

nanobot/channels/manager.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 import asyncio
 import hashlib
 import inspect
-from collections.abc import Callable, Iterable
+from collections.abc import Awaitable, Callable, Iterable, Mapping
 from contextlib import suppress
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, cast
@@ -95,24 +95,36 @@ class ChannelManager:
         cron_service: CronService | None = None,
         local_trigger_store: LocalTriggerStore | None = None,
         webui_runtime_model_name: Callable[[], str | None] | None = None,
+        webui_refresh_runtime_config: Callable[[], None] | None = None,
         webui_cron_pending_job_ids: Callable[[str], set[str]] | None = None,
         webui_local_trigger_pending_ids: Callable[[str], set[str]] | None = None,
         webui_static_dist: bool = True,
         webui_runtime_surface: str = "browser",
         webui_runtime_capabilities: dict[str, Any] | None = None,
+        webui_mcp_runtime_status: Callable[[], Mapping[str, str]] | None = None,
+        webui_mcp_reload: Callable[[], Awaitable[dict[str, Any]]] | None = None,
         webui_skill_state_action: Callable[[set[str]], None] | None = None,
+        config_path: Path | None = None,
     ):
+        if config_path is None:
+            from nanobot.config.loader import get_config_path
+
+            config_path = get_config_path()
         self.config = config
+        self._config_path = config_path.expanduser().resolve(strict=False)
         self.bus = bus
         self._session_manager = session_manager
         self._cron_service = cron_service
         self._local_trigger_store = local_trigger_store
         self._webui_runtime_model_name = webui_runtime_model_name
+        self._webui_refresh_runtime_config = webui_refresh_runtime_config
         self._webui_cron_pending_job_ids = webui_cron_pending_job_ids
         self._webui_local_trigger_pending_ids = webui_local_trigger_pending_ids
         self._webui_static_dist = webui_static_dist
         self._webui_runtime_surface = webui_runtime_surface
         self._webui_runtime_capabilities = dict(webui_runtime_capabilities or {})
+        self._webui_mcp_runtime_status = webui_mcp_runtime_status
+        self._webui_mcp_reload = webui_mcp_reload
         self._webui_skill_state_action = webui_skill_state_action
         self.channels: dict[str, BaseChannel] = {}
         self._channel_owners: dict[str, str] = {}
@@ -170,8 +182,10 @@ class ChannelManager:
                 static_dist_path=static_path,
                 workspace_path=workspace,
                 default_restrict_to_workspace=self.config.tools.restrict_to_workspace,
+                config_path=self._config_path,
                 disabled_skills=set(self.config.agents.defaults.disabled_skills),
                 runtime_model_name=self._webui_runtime_model_name,
+                refresh_runtime_config=self._webui_refresh_runtime_config,
                 runtime_surface=self._webui_runtime_surface,
                 runtime_capabilities_overrides=self._webui_runtime_capabilities,
                 cron_service=self._cron_service,
@@ -180,6 +194,8 @@ class ChannelManager:
                 local_trigger_pending_ids=self._webui_local_trigger_pending_ids,
                 channel_feature_action=self.apply_channel_feature_action,
                 channel_runtime_status=self.get_status,
+                mcp_runtime_status=self._webui_mcp_runtime_status,
+                mcp_reload=self._webui_mcp_reload,
                 skill_state_action=self._webui_skill_state_action,
                 logger=logger,
             )

nanobot/channels/matrix/runtime.py

@@ -968,6 +968,11 @@ class MatrixChannel(BaseChannel):
             meta["thread_reply_to_event_id"] = reply_to
         return meta
 
+    def _thread_session_key(self, room_id: str, event: RoomMessage) -> str | None:
+        if not (root_id := self._event_thread_root_id(event)):
+            return None
+        return f"{self.name}:{room_id}:thread:{root_id}"
+
     @staticmethod
     def _build_thread_relates_to(metadata: dict[str, Any] | None) -> dict[str, Any] | None:
         if not metadata:
@@ -1171,6 +1176,7 @@ class MatrixChannel(BaseChannel):
             await self._handle_message(
                 sender_id=event.sender, chat_id=room.room_id,
                 content=event.body, metadata=self._base_metadata(room, event),
+                session_key=self._thread_session_key(room.room_id, event),
                 is_dm=self._is_direct_room(room),
             )
         except Exception:
@@ -1209,6 +1215,7 @@ class MatrixChannel(BaseChannel):
                 content="\n".join(parts),
                 media=[attachment["path"]] if attachment else [],
                 metadata=meta,
+                session_key=self._thread_session_key(room.room_id, event),
                 is_dm=self._is_direct_room(room),
             )
         except Exception:

nanobot/channels/matrix/tests/test_matrix_channel.py

@@ -971,6 +971,81 @@ async def test_on_message_sets_thread_metadata_when_threaded_event() -> None:
     assert metadata["thread_root_event_id"] == "$root1"
     assert metadata["thread_reply_to_event_id"] == "$reply1"
     assert metadata["event_id"] == "$reply1"
+    assert handled[0]["session_key"] == "matrix:!room:matrix.org:thread:$root1"
+
+
+@pytest.mark.asyncio
+async def test_on_message_keeps_matrix_thread_sessions_independent() -> None:
+    channel = MatrixChannel(_make_config(), MessageBus())
+    client = _FakeAsyncClient("", "", "", None)
+    channel.client = client
+
+    handled: list[dict[str, object]] = []
+
+    async def _fake_handle_message(**kwargs) -> None:
+        handled.append(kwargs)
+
+    channel._handle_message = _fake_handle_message  # type: ignore[method-assign]
+
+    room = SimpleNamespace(room_id="!room:matrix.org", display_name="Test room", member_count=3)
+
+    def _thread_event(body: str, event_id: str, root_id: str) -> SimpleNamespace:
+        return SimpleNamespace(
+            sender="@alice:matrix.org",
+            body=body,
+            event_id=event_id,
+            source={
+                "content": {
+                    "m.relates_to": {
+                        "rel_type": "m.thread",
+                        "event_id": root_id,
+                    }
+                }
+            },
+        )
+
+    await channel._on_message(room, _thread_event("Plan the wedding", "$reply1", "$root1"))
+    await channel._on_message(room, _thread_event("Pick a gift", "$reply2", "$root1"))
+    await channel._on_message(room, _thread_event("/new", "$reply3", "$root2"))
+
+    assert [message["chat_id"] for message in handled] == [
+        "!room:matrix.org",
+        "!room:matrix.org",
+        "!room:matrix.org",
+    ]
+    assert [message["session_key"] for message in handled] == [
+        "matrix:!room:matrix.org:thread:$root1",
+        "matrix:!room:matrix.org:thread:$root1",
+        "matrix:!room:matrix.org:thread:$root2",
+    ]
+    assert handled[2]["content"] == "/new"
+
+
+@pytest.mark.asyncio
+async def test_on_message_keeps_non_threaded_room_session() -> None:
+    channel = MatrixChannel(_make_config(), MessageBus())
+    client = _FakeAsyncClient("", "", "", None)
+    channel.client = client
+
+    handled: list[dict[str, object]] = []
+
+    async def _fake_handle_message(**kwargs) -> None:
+        handled.append(kwargs)
+
+    channel._handle_message = _fake_handle_message  # type: ignore[method-assign]
+
+    room = SimpleNamespace(room_id="!room:matrix.org", display_name="Test room", member_count=3)
+    event = SimpleNamespace(
+        sender="@alice:matrix.org",
+        body="Hello",
+        event_id="$event1",
+        source={"content": {}},
+    )
+
+    await channel._on_message(room, event)
+
+    assert len(handled) == 1
+    assert handled[0]["session_key"] is None
 
 
 @pytest.mark.asyncio
@@ -1076,6 +1151,7 @@ async def test_on_media_message_sets_thread_metadata_when_threaded_event(
     assert metadata["thread_root_event_id"] == "$root1"
     assert metadata["thread_reply_to_event_id"] == "$event1"
     assert metadata["event_id"] == "$event1"
+    assert handled[0]["session_key"] == "matrix:!room:matrix.org:thread:$root1"
 
 
 @pytest.mark.asyncio

nanobot/channels/telegram/runtime.py

@@ -25,9 +25,9 @@ from telegram import (
     Update,
     User,
 )
-from telegram.error import BadRequest, NetworkError, TimedOut
+from telegram.error import BadRequest, InvalidToken, NetworkError, TimedOut
 from telegram.ext import Application, CallbackQueryHandler, ContextTypes, MessageHandler, filters
-from telegram.request import HTTPXRequest
+from telegram.request import BaseRequest, HTTPXRequest
 
 from nanobot.bus.events import OutboundMessage
 from nanobot.bus.outbound_events import ProgressEvent
@@ -38,6 +38,7 @@ from nanobot.config.paths import get_media_dir
 from nanobot.config.schema import Base
 from nanobot.security.network import validate_url_target
 from nanobot.utils.helpers import split_message
+from nanobot.utils.logging_bridge import redirect_lib_logging
 
 TELEGRAM_MAX_MESSAGE_LEN = 4000  # Telegram message character limit
 # Telegram's actual API limit is 4096; we split raw markdown at 4000 as a
@@ -53,6 +54,42 @@ TELEGRAM_REPLY_CONTEXT_MAX_LEN = TELEGRAM_MAX_MESSAGE_LEN  # Max length for repl
 TelegramApplication: TypeAlias = Application[Any, Any, Any, Any, Any, Any]
 _T = TypeVar("_T")
 
+# A healthy getUpdates long poll completes every ~10s even with no traffic;
+# PTB retries timeouts silently, so stalls must be detected here.
+POLL_STALE_SECONDS = 120.0
+POLL_WATCH_INTERVAL = 1.0
+RESTART_BACKOFF_INITIAL_SECONDS = 5.0
+RESTART_BACKOFF_MAX_SECONDS = 300.0
+# How long a send waits out a rebuild; short because ChannelManager dispatches
+# every channel from one serial loop.
+APP_RESTART_SEND_WAIT_SECONDS = 2.0
+
+
+class _LivenessTrackedRequest(BaseRequest):
+    """Wrap the getUpdates request pool, reporting each completed round trip."""
+
+    __slots__ = ("inner", "_on_round_trip")
+
+    def __init__(self, inner: BaseRequest, on_round_trip: Callable[[], None]) -> None:
+        super().__init__()
+        self.inner = inner
+        self._on_round_trip = on_round_trip
+
+    @property
+    def read_timeout(self) -> float | None:
+        return self.inner.read_timeout
+
+    async def initialize(self) -> None:
+        await self.inner.initialize()
+
+    async def shutdown(self) -> None:
+        await self.inner.shutdown()
+
+    async def do_request(self, *args: Any, **kwargs: Any) -> tuple[int, bytes]:
+        result = await self.inner.do_request(*args, **kwargs)
+        self._on_round_trip()
+        return result
+
 
 def _split_telegram_markdown(content: str, max_len: int) -> list[str]:
     """Split raw Telegram Markdown without leaving fenced code blocks unbalanced."""
@@ -477,6 +514,9 @@ class TelegramChannel(BaseChannel):
         self._inbound_buffers: dict[str, list[_QueuedTelegramUpdate]] = {}
         self._inbound_workers: dict[str, asyncio.Task[None]] = {}
         self._rich_send_disabled: bool = False  # Latch off if Bot API < 10.1
+        self._last_poll_ok: float = 0.0  # monotonic time of last getUpdates round trip
+        self._app_ready = asyncio.Event()  # cleared while the app is being rebuilt
+        self._teardown_lock = asyncio.Lock()
 
     def _require_app(self) -> TelegramApplication:
         if self._app is None:
@@ -516,13 +556,63 @@ class TelegramChannel(BaseChannel):
         return content
 
     async def start(self) -> None:
-        """Start the Telegram bot."""
+        """Start the Telegram bot, rebuilding the app whenever polling stalls."""
         if not self.config.token:
             self.logger.error("bot token not configured")
             return
 
-        self._running = True
+        redirect_lib_logging("telegram")
+        redirect_lib_logging("httpx", level="WARNING")
 
+        self._running = True
+        backoff = RESTART_BACKOFF_INITIAL_SECONDS
+        while self._running:
+            try:
+                await self._start_app()
+            except InvalidToken:
+                # A config error, not a blip: fail the channel. The scrubbed
+                # re-raise keeps PTB's token-bearing message out of the log.
+                await self._teardown_app()
+                self._running = False
+                self.logger.error("bot token rejected by Telegram")
+                raise RuntimeError("Telegram bot token was rejected by the server") from None
+            except Exception as e:
+                await self._teardown_app()
+                if not self._running:
+                    break
+                if not self._is_transient_startup_error(e):
+                    # Never heals on its own: fail instead of retrying forever
+                    # while ChannelManager keeps reporting the channel running.
+                    self._running = False
+                    self.logger.error("startup failed: {}", self._format_telegram_error(e))
+                    raise
+                self.logger.error(
+                    "startup failed: {}; retrying in {:.0f}s",
+                    self._format_telegram_error(e),
+                    backoff,
+                )
+                await self._idle(backoff)
+                backoff = min(backoff * 2, RESTART_BACKOFF_MAX_SECONDS)
+                continue
+
+            backoff = RESTART_BACKOFF_INITIAL_SECONDS
+            if not self._running:
+                # stop() ran while _start_app() was mid-flight and tore down the
+                # previous (possibly None) app; this one would leak otherwise.
+                await self._teardown_app()
+                break
+            stalled = await self._watch_polling()
+            if not stalled or not self._running:
+                break
+            self.logger.warning(
+                "polling stalled: no getUpdates round trip for {:.0f}s; "
+                "rebuilding connection pools and restarting",
+                time.monotonic() - self._last_poll_ok,
+            )
+            await self._teardown_app()
+
+    async def _start_app(self) -> None:
+        """Build, initialize and start the Telegram application."""
         proxy = self.config.proxy or None
 
         # Separate pools so long-polling (getUpdates) never starves outbound sends.
@@ -544,7 +634,7 @@ class TelegramChannel(BaseChannel):
             Application.builder()
             .token(self.config.token)
             .request(api_request)
-            .get_updates_request(poll_request)
+            .get_updates_request(_LivenessTrackedRequest(poll_request, self._note_poll_ok))
         )
         self._app = builder.build()
         self._app.add_error_handler(self._on_error)
@@ -621,16 +711,80 @@ class TelegramChannel(BaseChannel):
                 max_connections=self.config.webhook_max_connections,
             )
         else:
-            # Start polling (this runs until stopped)
+            self._last_poll_ok = time.monotonic()
             await cast(Any, self._app.updater).start_polling(
                 allowed_updates=allowed_updates,
                 drop_pending_updates=False,  # Process pending messages on startup
                 error_callback=self._on_polling_error,
             )
 
-        # Keep running until stopped
+        self._app_ready.set()
+
+    @staticmethod
+    def _is_transient_startup_error(exc: Exception) -> bool:
+        """Report whether a startup failure is worth retrying.
+
+        HTTPXRequest wraps every httpx failure into NetworkError/TimedOut, so
+        anything else is terminal: a bad proxy raises ValueError, an already
+        bound webhook port raises OSError.
+        """
+        return isinstance(exc, NetworkError | TimedOut | asyncio.TimeoutError)
+
+    async def _wait_for_app(self) -> TelegramApplication | None:
+        """Return the live app, briefly waiting out an in-flight rebuild.
+
+        Returning quietly while ``start()`` rebuilds would let the manager count
+        the message as delivered, so raise once the wait runs out. None means the
+        channel is stopped: nothing left to deliver.
+        """
+        if self._app_ready.is_set() and self._app is not None:
+            return self._app
+        if not self._running:
+            return None
+        with suppress(asyncio.TimeoutError):
+            await asyncio.wait_for(self._app_ready.wait(), APP_RESTART_SEND_WAIT_SECONDS)
+        if not self._app_ready.is_set() or self._app is None:
+            raise RuntimeError("Telegram application is restarting; message not delivered")
+        return self._app
+
+    def _note_poll_ok(self) -> None:
+        # HTTP error statuses count too: the watchdog detects transport stalls,
+        # not logical failures.
+        self._last_poll_ok = time.monotonic()
+
+    async def _watch_polling(self) -> bool:
+        """Idle until stop(); in polling mode, return True when getUpdates goes stale."""
+        watch = self.config.mode != "webhook"
         while self._running:
-            await asyncio.sleep(1)
+            await asyncio.sleep(POLL_WATCH_INTERVAL)
+            if watch and time.monotonic() - self._last_poll_ok > POLL_STALE_SECONDS:
+                return True
+        return False
+
+    async def _idle(self, seconds: float) -> None:
+        """Sleep in short steps so stop() stays responsive."""
+        deadline = time.monotonic() + seconds
+        while self._running and time.monotonic() < deadline:
+            await asyncio.sleep(POLL_WATCH_INTERVAL)
+
+    async def _teardown_app(self) -> None:
+        """Shut down the application, tolerating partially started state."""
+        async with self._teardown_lock:
+            app, self._app = self._app, None
+            self._app_ready.clear()
+            if not app:
+                return
+            for step in (cast(Any, app.updater).stop, app.stop, app.shutdown):
+                try:
+                    await step()
+                except Exception as e:
+                    self.logger.debug("teardown step failed: {}", e)
+            # Application.shutdown() skips the HTTPX pools unless initialize()
+            # finished, so a failed startup leaks one per retry. This is idempotent.
+            try:
+                await app.bot.shutdown()
+            except Exception as e:
+                self.logger.debug("bot shutdown failed: {}", e)
 
     async def stop(self) -> None:
         """Stop the Telegram bot."""
@@ -652,10 +806,9 @@ class TelegramChannel(BaseChannel):
 
         if self._app:
             self.logger.info("Stopping bot...")
-            await cast(Any, self._app.updater).stop()
-            await self._app.stop()
-            await self._app.shutdown()
-            self._app = None
+        # Join an in-flight supervisor teardown before ChannelManager cancels
+        # start(), otherwise cancellation can strand the old HTTPX pools.
+        await self._teardown_app()
 
     @staticmethod
     def _get_media_type(path: str) -> str:
@@ -746,7 +899,8 @@ class TelegramChannel(BaseChannel):
 
     async def send(self, msg: OutboundMessage) -> None:
         """Send a message through Telegram."""
-        if not self._app:
+        app = await self._wait_for_app()
+        if app is None:
             self.logger.warning("bot not running")
             return
 
@@ -785,11 +939,11 @@ class TelegramChannel(BaseChannel):
             try:
                 media_type = self._get_media_type(media_path)
                 sender = {
-                    "photo": self._app.bot.send_photo,
-                    "video": self._app.bot.send_video,
-                    "voice": self._app.bot.send_voice,
-                    "audio": self._app.bot.send_audio,
-                }.get(media_type, self._app.bot.send_document)
+                    "photo": app.bot.send_photo,
+                    "video": app.bot.send_video,
+                    "voice": app.bot.send_voice,
+                    "audio": app.bot.send_audio,
+                }.get(media_type, app.bot.send_document)
                 param = {
                     "photo": "photo",
                     "video": "video",
@@ -829,7 +983,7 @@ class TelegramChannel(BaseChannel):
             except Exception:
                 filename = media_path.rsplit("/", 1)[-1]
                 self.logger.exception("Failed to send media {}", media_path)
-                await self._app.bot.send_message(
+                await app.bot.send_message(
                     chat_id=chat_id,
                     text=f"[Failed to send: {filename}]",
                     reply_parameters=reply_params,
@@ -957,7 +1111,8 @@ class TelegramChannel(BaseChannel):
         merge_next: bool = False,
     ) -> None:
         """Progressive message editing: send on first delta, edit on subsequent ones."""
-        if not self._app:
+        app = await self._wait_for_app()
+        if app is None:
             return
         meta = metadata or {}
         int_chat_id = int(chat_id)
@@ -996,7 +1151,7 @@ class TelegramChannel(BaseChannel):
                     # Delete the streaming preview message
                     try:
                         await self._call_with_retry(
-                            self._app.bot.delete_message,
+                            app.bot.delete_message,
                             chat_id=int_chat_id, message_id=buf.message_id,
                         )
                     except Exception:
@@ -1010,7 +1165,7 @@ class TelegramChannel(BaseChannel):
             extra_html_chunks = html_chunks[1:]
             try:
                 await self._call_with_retry(
-                    self._app.bot.edit_message_text,
+                    app.bot.edit_message_text,
                     chat_id=int_chat_id, message_id=buf.message_id,
                     text=primary_html, parse_mode="HTML",
                 )
@@ -1027,7 +1182,7 @@ class TelegramChannel(BaseChannel):
                 primary_plain = split_message(raw_text, TELEGRAM_MAX_MESSAGE_LEN)[0] if len(raw_text) > TELEGRAM_MAX_MESSAGE_LEN else raw_text
                 try:
                     await self._call_with_retry(
-                        self._app.bot.edit_message_text,
+                        app.bot.edit_message_text,
                         chat_id=int_chat_id, message_id=buf.message_id,
                         text=primary_plain,
                     )
@@ -1040,7 +1195,7 @@ class TelegramChannel(BaseChannel):
             for extra_html_chunk in extra_html_chunks:
                 try:
                     await self._call_with_retry(
-                        self._app.bot.send_message,
+                        app.bot.send_message,
                         chat_id=int_chat_id, text=extra_html_chunk,
                         parse_mode="HTML",
                         **thread_kwargs,
@@ -1070,7 +1225,7 @@ class TelegramChannel(BaseChannel):
             preview = _strip_md_block(buf.text)
             try:
                 sent = await self._call_with_retry(
-                    self._app.bot.send_message,
+                    app.bot.send_message,
                     chat_id=int_chat_id, text=preview,
                     **stream_thread_kwargs,
                 )
@@ -1087,7 +1242,7 @@ class TelegramChannel(BaseChannel):
             preview = _strip_md_block(buf.text)
             try:
                 await self._call_with_retry(
-                    self._app.bot.edit_message_text,
+                    app.bot.edit_message_text,
                     chat_id=int_chat_id, message_id=buf.message_id,
                     text=preview,
                 )

nanobot/channels/telegram/tests/test_telegram_channel.py

@@ -61,6 +61,10 @@ class _FakeBot:
         self.sent_messages: list[dict] = []
         self.sent_media: list[dict] = []
         self.get_me_calls = 0
+        self.shutdown_calls = 0
+
+    async def shutdown(self) -> None:
+        self.shutdown_calls += 1
 
     async def get_me(self):
         self.get_me_calls += 1
@@ -153,6 +157,14 @@ class _FakeBuilder:
         return self.app
 
 
+def _install_ready_app(channel: TelegramChannel) -> _FakeApp:
+    """Install the ready app state expected by ordinary send tests."""
+    app = _FakeApp(lambda: None)
+    channel._app = app
+    channel._app_ready.set()
+    return app
+
+
 def _make_telegram_update(
     *,
     chat_type: str = "group",
@@ -337,7 +349,7 @@ async def test_start_creates_separate_pools_with_proxy(monkeypatch) -> None:
     assert api_req.kwargs["connection_pool_size"] == 32
     assert poll_req.kwargs["connection_pool_size"] == 4
     assert builder.request_value is api_req
-    assert builder.get_updates_request_value is poll_req
+    assert builder.get_updates_request_value.inner is poll_req
     assert callable(app.updater.start_polling_kwargs["error_callback"])
     assert any(cmd.command == "status" for cmd in app.bot.commands)
     assert any(cmd.command == "history" for cmd in app.bot.commands)
@@ -378,6 +390,311 @@ async def test_start_respects_custom_pool_config(monkeypatch) -> None:
     assert poll_req.kwargs["pool_timeout"] == 10.0
 
 
+@pytest.mark.asyncio
+async def test_stalled_polling_triggers_pool_rebuild(monkeypatch) -> None:
+    """When no getUpdates round trip completes for too long, the app is rebuilt."""
+    _FakeHTTPXRequest.clear()
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    bus = MessageBus()
+    channel = TelegramChannel(config, bus)
+
+    apps: list[_FakeApp] = []
+
+    def on_start_polling() -> None:
+        if len(apps) >= 2:
+            channel._running = False
+
+    def make_builder():
+        app = _FakeApp(on_start_polling)
+        apps.append(app)
+        return _FakeBuilder(app)
+
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=make_builder),
+    )
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.POLL_STALE_SECONDS", -1.0)
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.POLL_WATCH_INTERVAL", 0.0)
+
+    await channel.start()
+
+    assert len(apps) == 2
+    assert apps[0].updater.start_polling_kwargs is not None
+    assert apps[1].updater.start_polling_kwargs is not None
+    # 2 fresh pools per app
+    assert len(_FakeHTTPXRequest.instances) == 4
+
+
+@pytest.mark.asyncio
+async def test_startup_failure_retries_with_backoff(monkeypatch) -> None:
+    """Transient startup failures back off and retry until the app comes up."""
+    from telegram.error import NetworkError
+
+    _FakeHTTPXRequest.clear()
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    bus = MessageBus()
+    channel = TelegramChannel(config, bus)
+
+    apps: list[_FakeApp] = []
+
+    def make_builder():
+        app = _FakeApp(lambda: setattr(channel, "_running", False))
+        if len(apps) < 2:
+            async def _fail() -> None:
+                raise NetworkError("connect failed")
+
+            app.initialize = _fail
+        apps.append(app)
+        return _FakeBuilder(app)
+
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=make_builder),
+    )
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.POLL_WATCH_INTERVAL", 0.0)
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.RESTART_BACKOFF_INITIAL_SECONDS", 0.0)
+
+    await channel.start()
+
+    assert len(apps) == 3
+    assert apps[0].updater.start_polling_kwargs is None
+    assert apps[1].updater.start_polling_kwargs is None
+    assert apps[2].updater.start_polling_kwargs is not None
+    # Pools must be closed via the bot: app.shutdown() skips them here.
+    assert apps[0].bot.shutdown_calls == 1
+    assert apps[1].bot.shutdown_calls == 1
+
+
+@pytest.mark.asyncio
+async def test_terminal_startup_error_is_not_retried(monkeypatch) -> None:
+    """Config errors (bad proxy, bound webhook port) must fail the channel."""
+    _FakeHTTPXRequest.clear()
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    bus = MessageBus()
+    channel = TelegramChannel(config, bus)
+
+    apps: list[_FakeApp] = []
+
+    def make_builder():
+        app = _FakeApp(lambda: None)
+
+        async def _fail() -> None:
+            raise ValueError("Unknown scheme for proxy URL")
+
+        app.initialize = _fail
+        apps.append(app)
+        return _FakeBuilder(app)
+
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=make_builder),
+    )
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.RESTART_BACKOFF_INITIAL_SECONDS", 0.0)
+
+    with pytest.raises(ValueError, match="proxy URL"):
+        await channel.start()
+
+    assert len(apps) == 1  # no retry loop
+    assert channel._app is None
+    assert channel.is_running is False
+
+
+@pytest.mark.asyncio
+async def test_invalid_token_stops_without_retry(monkeypatch) -> None:
+    """A rejected token is a config error: fail the channel instead of retrying."""
+    from telegram.error import InvalidToken
+
+    _FakeHTTPXRequest.clear()
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    bus = MessageBus()
+    channel = TelegramChannel(config, bus)
+
+    apps: list[_FakeApp] = []
+
+    def make_builder():
+        app = _FakeApp(lambda: None)
+
+        async def _reject() -> None:
+            raise InvalidToken("token rejected by Telegram")
+
+        app.initialize = _reject
+        apps.append(app)
+        return _FakeBuilder(app)
+
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=make_builder),
+    )
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.RESTART_BACKOFF_INITIAL_SECONDS", 0.0)
+
+    with pytest.raises(RuntimeError) as excinfo:
+        await channel.start()
+
+    assert len(apps) == 1
+    assert channel._app is None
+    assert channel.is_running is False
+    assert "123:abc" not in str(excinfo.value)  # token must not reach the log
+
+
+@pytest.mark.asyncio
+async def test_stop_during_startup_does_not_leak_app(monkeypatch) -> None:
+    """stop() landing while _start_app() is mid-flight must not leave the app running."""
+    _FakeHTTPXRequest.clear()
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    bus = MessageBus()
+    channel = TelegramChannel(config, bus)
+
+    # Simulate stop() winning the race just before start_polling returns.
+    app = _FakeApp(lambda: setattr(channel, "_running", False))
+    builder = _FakeBuilder(app)
+
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=lambda: builder),
+    )
+
+    await channel.start()
+
+    assert channel._app is None  # torn down, not leaked
+
+
+@pytest.mark.asyncio
+async def test_stop_waits_for_inflight_watchdog_teardown(monkeypatch) -> None:
+    """Manager cancellation after stop() must not interrupt an active teardown."""
+    _FakeHTTPXRequest.clear()
+    channel = TelegramChannel(
+        TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
+        MessageBus(),
+    )
+    teardown_started = asyncio.Event()
+    finish_teardown = asyncio.Event()
+    app = _FakeApp(lambda: None)
+
+    async def slow_updater_stop() -> None:
+        teardown_started.set()
+        await finish_teardown.wait()
+
+    app.updater.stop = slow_updater_stop
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.HTTPXRequest", _FakeHTTPXRequest)
+    monkeypatch.setattr(
+        "nanobot.channels.telegram.runtime.Application",
+        SimpleNamespace(builder=lambda: _FakeBuilder(app)),
+    )
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.POLL_STALE_SECONDS", -1.0)
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.POLL_WATCH_INTERVAL", 0.0)
+
+    start_task = asyncio.create_task(channel.start())
+    await teardown_started.wait()
+    assert channel._app is None
+
+    stop_task = asyncio.create_task(channel.stop())
+    await asyncio.sleep(0)
+    assert not stop_task.done()
+
+    finish_teardown.set()
+    await stop_task
+    await start_task
+
+    assert app.bot.shutdown_calls == 1
+
+
+@pytest.mark.asyncio
+async def test_send_during_rebuild_fails_instead_of_dropping(monkeypatch) -> None:
+    """A send that cannot reach Telegram must raise so the manager can retry."""
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.APP_RESTART_SEND_WAIT_SECONDS", 0.0)
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    channel = TelegramChannel(config, MessageBus())
+
+    # Mid-rebuild: still running, but no app to send through.
+    channel._running = True
+    channel._app = None
+
+    msg = OutboundMessage(channel="telegram", chat_id="123", content="hello")
+    with pytest.raises(RuntimeError, match="restarting"):
+        await channel.send(msg)
+    with pytest.raises(RuntimeError, match="restarting"):
+        await channel.send_delta("123", "hello", stream_id="s1")
+
+    # Stopped: nothing to deliver, so stay quiet.
+    channel._running = False
+    await channel.send(msg)
+    await channel.send_delta("123", "hello", stream_id="s1")
+
+
+@pytest.mark.asyncio
+async def test_send_waits_for_rebuild_to_finish(monkeypatch) -> None:
+    """A fast rebuild is waited out rather than surfaced as a delivery failure."""
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.APP_RESTART_SEND_WAIT_SECONDS", 5.0)
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    channel = TelegramChannel(config, MessageBus())
+
+    app = _FakeApp(lambda: None)
+    channel._running = True
+    channel._app = None
+
+    async def _finish_rebuild() -> None:
+        await asyncio.sleep(0)
+        channel._app = app
+        channel._app_ready.set()
+
+    rebuild = asyncio.create_task(_finish_rebuild())
+    await channel.send(OutboundMessage(channel="telegram", chat_id="123", content="hello"))
+    await rebuild
+
+    assert [m["text"] for m in app.bot.sent_messages] == ["hello"]
+
+
+@pytest.mark.asyncio
+async def test_send_waits_for_partially_initialized_app(monkeypatch) -> None:
+    """A built app is not available for sends until startup marks it ready."""
+    monkeypatch.setattr("nanobot.channels.telegram.runtime.APP_RESTART_SEND_WAIT_SECONDS", 5.0)
+    config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
+    channel = TelegramChannel(config, MessageBus())
+
+    app = _FakeApp(lambda: None)
+    channel._running = True
+    channel._app = app
+
+    send_task = asyncio.create_task(
+        channel.send(OutboundMessage(channel="telegram", chat_id="123", content="hello"))
+    )
+    await asyncio.sleep(0)
+    assert not send_task.done()
+
+    channel._app_ready.set()
+    await send_task
+
+    assert [m["text"] for m in app.bot.sent_messages] == ["hello"]
+
+
+@pytest.mark.asyncio
+async def test_liveness_tracked_request_stamps_on_round_trip() -> None:
+    from nanobot.channels.telegram.runtime import _LivenessTrackedRequest
+
+    stamps: list[int] = []
+
+    class _Inner:
+        read_timeout = 5.0
+
+        async def initialize(self) -> None:
+            pass
+
+        async def shutdown(self) -> None:
+            pass
+
+        async def do_request(self, *args, **kwargs):
+            return 200, b"{}"
+
+    wrapped = _LivenessTrackedRequest(_Inner(), lambda: stamps.append(1))
+    assert await wrapped.do_request(url="https://example.org", method="POST") == (200, b"{}")
+    assert stamps == [1]
+
+
 def test_webhook_config_requires_https_url_and_secret() -> None:
     with pytest.raises(ValueError, match="webhook_url is required"):
         TelegramConfig(enabled=True, token="123:abc", mode="webhook")
@@ -479,7 +796,7 @@ async def test_send_text_retries_on_timeout() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     call_count = 0
     original_send = channel._app.bot.send_message
@@ -514,7 +831,7 @@ async def test_send_text_gives_up_after_max_retries() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     async def always_timeout(**kwargs):
         raise TimedOut()
@@ -541,7 +858,7 @@ async def test_send_rich_capability_error_latches_and_falls_back() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], rich_messages=True),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.do_api_request = AsyncMock(side_effect=BadRequest("Method not found"))
 
     await channel.send(OutboundMessage(channel="telegram", chat_id="123", content="**hello**"))
@@ -560,7 +877,7 @@ async def test_send_rich_bad_request_does_not_latch_capability() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], rich_messages=True),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.do_api_request = AsyncMock(
         side_effect=BadRequest("Bad Request: message to reply not found")
     )
@@ -579,7 +896,7 @@ async def test_rich_messages_default_skips_send_rich_message() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.do_api_request = AsyncMock()
 
     await channel.send(OutboundMessage(channel="telegram", chat_id="123", content="**hello**"))
@@ -666,7 +983,7 @@ async def test_send_delta_stream_end_raises_and_keeps_buffer_on_failure() -> Non
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock(side_effect=RuntimeError("boom"))
     channel._stream_bufs["123"] = _StreamBuf(text="hello", message_id=7, last_edit=0.0)
 
@@ -682,7 +999,7 @@ async def test_send_delta_merge_next_preserves_buffer() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._stream_bufs["123"] = _StreamBuf(
         text="first-",
@@ -711,7 +1028,7 @@ async def test_send_delta_stream_end_treats_not_modified_as_success() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock(side_effect=BadRequest("Message is not modified"))
     channel._stream_bufs["123"] = _StreamBuf(text="hello", message_id=7, last_edit=0.0, stream_id="s:0")
 
@@ -731,7 +1048,7 @@ async def test_send_delta_stream_end_does_not_fallback_on_network_timeout(
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     monkeypatch.setattr("nanobot.channels.telegram.runtime._SEND_RETRY_BASE_DELAY", 0)
     # _call_with_retry retries TimedOut up to 3 times, so the mock will be called
     # multiple times – but all calls must be with parse_mode="HTML" (no plain fallback).
@@ -759,7 +1076,7 @@ async def test_send_delta_stream_end_does_not_fallback_on_network_error() -> Non
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock(side_effect=NetworkError("connection reset"))
     channel._stream_bufs["123"] = _StreamBuf(text="hello", message_id=7, last_edit=0.0)
 
@@ -783,7 +1100,7 @@ async def test_send_delta_stream_end_falls_back_on_bad_request() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     # First call (HTML) raises BadRequest, second call (plain) succeeds
     channel._app.bot.edit_message_text = AsyncMock(
@@ -815,7 +1132,7 @@ async def test_send_delta_stream_end_splits_oversized_reply() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=99))
 
@@ -849,7 +1166,7 @@ async def test_send_delta_stream_end_html_expansion_does_not_overflow() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=99))
 
@@ -880,7 +1197,7 @@ async def test_send_delta_stream_end_splits_long_code_block_before_html_renderin
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=99))
 
@@ -909,7 +1226,7 @@ async def test_send_delta_new_stream_id_replaces_stale_buffer() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._stream_bufs["123"] = _StreamBuf(
         text="hello",
         message_id=7,
@@ -933,7 +1250,7 @@ async def test_send_delta_incremental_edit_treats_not_modified_as_success() -> N
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._stream_bufs["123"] = _StreamBuf(text="hello", message_id=7, last_edit=0.0, stream_id="s:0")
     channel._app.bot.edit_message_text = AsyncMock(side_effect=BadRequest("Message is not modified"))
 
@@ -951,7 +1268,7 @@ async def test_send_delta_incremental_edit_splits_oversized_buffer() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=99))
 
@@ -988,7 +1305,7 @@ async def test_send_delta_incremental_html_expansion_does_not_overflow() -> None
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock()
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=99))
 
@@ -1022,7 +1339,7 @@ async def test_send_delta_incremental_html_parse_failure_falls_back_to_plain() -
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.edit_message_text = AsyncMock(
         side_effect=[BadRequest("Can't parse entities"), None]
     )
@@ -1056,7 +1373,7 @@ async def test_send_delta_initial_send_keeps_message_in_thread() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     await channel.send_delta(
         "123",
@@ -1129,7 +1446,7 @@ def test_is_allowed_rejects_invalid_legacy_telegram_sender_shapes() -> None:
 async def test_send_progress_keeps_message_in_topic() -> None:
     config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"])
     channel = TelegramChannel(config, MessageBus())
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     await channel.send(
         OutboundMessage(
@@ -1148,7 +1465,7 @@ async def test_send_progress_keeps_message_in_topic() -> None:
 async def test_send_reply_infers_topic_from_message_id_cache() -> None:
     config = TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], reply_to_message=True)
     channel = TelegramChannel(config, MessageBus())
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._message_threads[("123", 10)] = 42
 
     await channel.send(
@@ -1170,7 +1487,7 @@ async def test_send_remote_media_url_after_security_validation(monkeypatch) -> N
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     monkeypatch.setattr("nanobot.channels.telegram.runtime.validate_url_target", lambda url: (True, ""))
 
     await channel.send(
@@ -1198,7 +1515,7 @@ async def test_send_local_media_preserves_filename(tmp_path: Path) -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     attachment = tmp_path / "report.final.md"
     attachment.write_bytes(b"# Report\n")
 
@@ -1228,7 +1545,7 @@ async def test_send_blocks_unsafe_remote_media_url(monkeypatch) -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     monkeypatch.setattr(
         "nanobot.channels.telegram.runtime.validate_url_target",
         lambda url: (False, "Blocked: example.com resolves to private/internal address 127.0.0.1"),
@@ -1259,7 +1576,7 @@ async def test_group_policy_mention_ignores_unmentioned_group_message() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="mention"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     handled = []
 
@@ -1281,7 +1598,7 @@ async def test_group_policy_mention_accepts_text_mention_and_caches_bot_identity
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="mention"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     handled = []
 
@@ -1305,7 +1622,7 @@ async def test_group_policy_mention_accepts_caption_mention() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="mention"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     handled = []
 
@@ -1331,7 +1648,7 @@ async def test_group_policy_mention_accepts_reply_to_bot() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="mention"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     handled = []
 
@@ -1353,7 +1670,7 @@ async def test_group_policy_open_accepts_plain_group_message() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     handled = []
 
@@ -1381,7 +1698,7 @@ async def test_extract_reply_context_no_reply() -> None:
 async def test_extract_reply_context_with_text() -> None:
     """When reply has text, return prefixed string."""
     channel = TelegramChannel(TelegramConfig(enabled=True, token="123:abc"), MessageBus())
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     reply = SimpleNamespace(text="Hello world", caption=None, from_user=SimpleNamespace(id=2, username="testuser", first_name="Test"))
     message = SimpleNamespace(reply_to_message=reply)
     assert await channel._extract_reply_context(message) == "[Reply to @testuser: Hello world]"
@@ -1391,7 +1708,7 @@ async def test_extract_reply_context_with_text() -> None:
 async def test_extract_reply_context_with_caption_only() -> None:
     """When reply has only caption (no text), caption is used."""
     channel = TelegramChannel(TelegramConfig(enabled=True, token="123:abc"), MessageBus())
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     reply = SimpleNamespace(text=None, caption="Photo caption", from_user=SimpleNamespace(id=2, username=None, first_name="Test"))
     message = SimpleNamespace(reply_to_message=reply)
     assert await channel._extract_reply_context(message) == "[Reply to Test: Photo caption]"
@@ -1401,7 +1718,7 @@ async def test_extract_reply_context_with_caption_only() -> None:
 async def test_extract_reply_context_truncation() -> None:
     """Reply text is truncated at TELEGRAM_REPLY_CONTEXT_MAX_LEN."""
     channel = TelegramChannel(TelegramConfig(enabled=True, token="123:abc"), MessageBus())
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     long_text = "x" * (TELEGRAM_REPLY_CONTEXT_MAX_LEN + 100)
     reply = SimpleNamespace(text=long_text, caption=None, from_user=SimpleNamespace(id=2, username=None, first_name=None))
     message = SimpleNamespace(reply_to_message=reply)
@@ -1428,7 +1745,7 @@ async def test_on_message_includes_reply_context() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
     async def capture_handle(**kwargs) -> None:
         handled.append(kwargs)
@@ -1460,7 +1777,7 @@ async def test_download_message_media_returns_path_when_download_succeeds(
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.get_file = AsyncMock(
         return_value=SimpleNamespace(download_to_drive=AsyncMock(return_value=None))
     )
@@ -1587,7 +1904,7 @@ async def test_on_message_reply_to_media_fallback_when_download_fails() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.get_file = None
     handled = []
     async def capture_handle(**kwargs) -> None:
@@ -1670,7 +1987,7 @@ async def test_forward_command_does_not_inject_reply_context() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
     async def capture_handle(**kwargs) -> None:
         handled.append(kwargs)
@@ -1690,7 +2007,7 @@ async def test_forward_command_pairs_unauthorized_private_user(monkeypatch) -> N
         TelegramConfig(enabled=True, token="123:abc", allow_from=["999"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     monkeypatch.setattr(
         "nanobot.channels.base.generate_code", lambda _ch, _sid: "ABCD-EFGH"
     )
@@ -1707,7 +2024,7 @@ async def test_forward_command_preserves_dream_log_args_and_strips_bot_suffix()
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
 
     async def capture_handle(**kwargs) -> None:
@@ -1728,7 +2045,7 @@ async def test_forward_command_normalizes_telegram_safe_dream_aliases() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
 
     async def capture_handle(**kwargs) -> None:
@@ -1803,7 +2120,7 @@ async def test_on_start_sends_pairing_code_to_unauthorized_private_user(monkeypa
         TelegramConfig(enabled=True, token="123:abc", allow_from=["999"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     update = _make_telegram_update(text="/start", chat_type="private")
     update.message.reply_text = AsyncMock()
     monkeypatch.setattr(
@@ -1823,7 +2140,7 @@ async def test_on_help_sends_pairing_code_to_unauthorized_private_user(monkeypat
         TelegramConfig(enabled=True, token="123:abc", allow_from=["999"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     update = _make_telegram_update(text="/help", chat_type="private")
     update.message.reply_text = AsyncMock()
     monkeypatch.setattr(
@@ -1845,7 +2162,7 @@ async def test_on_message_pairs_unauthorized_private_user_before_side_effects(
         TelegramConfig(enabled=True, token="123:abc", allow_from=["999"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     started_typing: list[str] = []
     channel._start_typing = lambda chat_id: started_typing.append(chat_id)
     channel._add_reaction = AsyncMock(return_value=None)
@@ -1870,7 +2187,7 @@ async def test_on_message_location_content() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
     async def capture_handle(**kwargs) -> None:
         handled.append(kwargs)
@@ -1892,7 +2209,7 @@ async def test_on_message_location_with_text() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], group_policy="open"),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     handled = []
     async def capture_handle(**kwargs) -> None:
         handled.append(kwargs)
@@ -1956,7 +2273,7 @@ async def test_send_text_does_not_fallback_on_network_timeout() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     call_count = 0
 
@@ -1993,7 +2310,7 @@ async def test_send_text_does_not_fallback_on_network_error() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     call_count = 0
 
@@ -2030,7 +2347,7 @@ async def test_send_text_falls_back_on_bad_request() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     original_send = channel._app.bot.send_message
     html_call_count = 0
@@ -2068,7 +2385,7 @@ async def test_send_text_bad_request_plain_fallback_exhausted() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     call_count = 0
 
@@ -2191,7 +2508,7 @@ async def test_send_delta_mid_stream_strips_markdown() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"]),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
     channel._app.bot.send_message = AsyncMock(return_value=SimpleNamespace(message_id=42))
     channel._app.bot.edit_message_text = AsyncMock()
 
@@ -2292,7 +2609,7 @@ async def test_send_falls_back_buttons_to_inline_text_when_flag_off() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], inline_keyboards=False),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     await channel.send(
         OutboundMessage(
@@ -2320,7 +2637,7 @@ async def test_send_uses_native_keyboard_when_flag_on() -> None:
         TelegramConfig(enabled=True, token="123:abc", allow_from=["*"], inline_keyboards=True),
         MessageBus(),
     )
-    channel._app = _FakeApp(lambda: None)
+    _install_ready_app(channel)
 
     await channel.send(
         OutboundMessage(

nanobot/channels/websocket/runtime.py

@@ -3,14 +3,17 @@
 from __future__ import annotations
 
 import asyncio
+import hashlib
 import hmac
 import ipaddress
 import json
 import re
 import ssl
+import time
 import uuid
 from collections.abc import Callable
 from contextlib import suppress
+from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Self, TypeGuard, cast
 from urllib.parse import urlsplit, urlunsplit
@@ -21,6 +24,7 @@ from websockets.exceptions import ConnectionClosed
 from websockets.http11 import Request as WsRequest
 
 from nanobot.bus.events import (
+    INBOUND_META_USER_SHELL,
     OUTBOUND_META_AGENT_UI,
     OutboundMessage,
 )
@@ -36,7 +40,7 @@ from nanobot.bus.outbound_events import (
 )
 from nanobot.bus.queue import MessageBus
 from nanobot.channels.base import BaseChannel
-from nanobot.command.builtin import builtin_command_starts_agent_turn
+from nanobot.command.builtin import USER_SHELL_COMMAND, builtin_command_starts_agent_turn
 from nanobot.config.schema import Base
 from nanobot.runtime_context import (
     RUNTIME_CONTEXT_INPUT_META,
@@ -49,6 +53,7 @@ from nanobot.security.workspace_access import (
     WorkspaceScopeError,
 )
 from nanobot.session.goal_state import goal_state_ws_blob
+from nanobot.session.model_selection import model_preset_from_metadata
 from nanobot.session.webui_turns import (
     clear_websocket_turn_if_current,
     clear_websocket_turns,
@@ -58,6 +63,7 @@ from nanobot.session.webui_turns import (
     websocket_turn_transcript_persistence_failed,
     websocket_turn_wall_started_at,
 )
+from nanobot.utils.helpers import safe_filename
 from nanobot.webui.cli_apps_api import normalize_cli_app_mentions
 from nanobot.webui.forking import handle_webui_fork_chat
 from nanobot.webui.gateway_services import GatewayServices
@@ -92,6 +98,8 @@ from nanobot.webui.websocket_logging import websockets_server_logger
 
 # Plain HTTP WebUI routes also run through websockets.process_request.
 _WEBUI_HTTP_OPEN_TIMEOUT_S = 360.0
+_WEBUI_REQUEST_CACHE_TTL_S = 5 * 60.0
+_WEBUI_REQUEST_CACHE_MAX = 256
 
 
 _ROUTING_ASSERTION_HEADERS = frozenset(
@@ -348,6 +356,21 @@ def _is_websocket_upgrade(request: WsRequest) -> bool:
     return True
 
 
+@dataclass(frozen=True)
+class _WebUIRequestResult:
+    result: Any = None
+    status: int | None = None
+    message: str | None = None
+
+
+@dataclass
+class _WebUIRequestOperation:
+    action: str
+    payload_digest: bytes
+    task: asyncio.Task[_WebUIRequestResult]
+    completed_at: float | None = None
+
+
 class WebSocketChannel(BaseChannel):
     """Run a local WebSocket server; forward text/JSON messages to the message bus."""
 
@@ -373,6 +396,17 @@ class WebSocketChannel(BaseChannel):
         self._conn_default: dict[ServerConnection, str] = {}
         # Connections authenticated with a one-time token from /webui/bootstrap.
         self._webui_connections: set[ServerConnection] = set()
+        # Delivery tasks are connection-bound, while operations are keyed only
+        # by request_id so reconnect retries join or replay the original work.
+        self._webui_request_tasks: dict[
+            tuple[ServerConnection, str],
+            asyncio.Task[None],
+        ] = {}
+        self._webui_request_operations: dict[str, _WebUIRequestOperation] = {}
+        # Preserve request/response order for mutations from one
+        # UI. Without this, an earlier slow settings response can overwrite a
+        # newer settings snapshot in the client.
+        self._webui_request_locks: dict[ServerConnection, asyncio.Lock] = {}
         self._stop_event: asyncio.Event | None = None
         self._server_task: asyncio.Task[None] | None = None
 
@@ -402,6 +436,35 @@ class WebSocketChannel(BaseChannel):
         self._subs.setdefault(chat_id, set()).add(connection)
         self._conn_chats.setdefault(connection, set()).add(chat_id)
 
+    def _attached_model_fields(self, chat_id: str) -> dict[str, Any]:
+        """Expose small session runtime facts on the attach handshake."""
+        sessions = self.gateway.session_manager
+        if sessions is None:
+            return {}
+        snapshot = sessions.read_session_metadata(f"websocket:{chat_id}")
+        raw_metadata = snapshot.get("metadata") if snapshot is not None else None
+        metadata = cast(dict[str, object], raw_metadata) if isinstance(raw_metadata, dict) else None
+        fields: dict[str, Any] = {}
+        try:
+            fields["model_preset"] = model_preset_from_metadata(metadata)
+        except ValueError:
+            self.logger.warning("ignoring invalid model preset metadata for chat_id={}", chat_id)
+            fields["model_preset"] = None
+        if isinstance(metadata, dict):
+            usage = metadata.get("_last_usage")
+            if isinstance(usage, dict):
+                sanitized_usage: dict[str, int | float] = {}
+                for key, value in cast(dict[object, object], usage).items():
+                    if (
+                        isinstance(key, str)
+                        and isinstance(value, (int, float))
+                        and not isinstance(value, bool)
+                        and value >= 0
+                    ):
+                        sanitized_usage[key] = value
+                fields["usage"] = sanitized_usage
+        return fields
+
     def _detach(self, connection: ServerConnection, chat_id: str) -> None:
         chats = self._conn_chats.get(connection)
         if chats is not None:
@@ -447,7 +510,12 @@ class WebSocketChannel(BaseChannel):
         """Attach and hydrate a newly created WebUI chat fork."""
         scope = self._workspaces.scope_for_session_key(fork_key)
         self._attach(connection, fork_id)
-        await self._send_event(connection, "attached", chat_id=fork_id)
+        await self._send_event(
+            connection,
+            "attached",
+            chat_id=fork_id,
+            **self._attached_model_fields(fork_id),
+        )
         await self._send_event(
             connection,
             "session_updated",
@@ -469,9 +537,10 @@ class WebSocketChannel(BaseChannel):
             await self._discard_connection_owned_chat(connection, cid)
         self._conn_default.pop(connection, None)
         self._webui_connections.discard(connection)
+        self._discard_webui_request_lock_if_idle(connection)
 
-    async def _maybe_push_active_goal_state(self, chat_id: str) -> None:
-        """Replay an active sustained goal from session metadata after *chat_id* is subscribed.
+    async def _maybe_push_persisted_goal_state(self, chat_id: str) -> None:
+        """Replay actionable goal state after *chat_id* is subscribed.
 
         Goal metadata lives on the session JSONL and survives gateway restarts, but
         connected clients normally see it via ``goal_state`` / ``turn_end`` frames.
@@ -485,7 +554,7 @@ class WebSocketChannel(BaseChannel):
         if not isinstance(meta, dict):
             meta = {}
         blob = goal_state_ws_blob(cast(dict[str, Any], meta))
-        if not blob.get("active"):
+        if not blob.get("active") and blob.get("status") != "blocked":
             return
         await self.send_goal_state(chat_id, blob)
 
@@ -503,7 +572,7 @@ class WebSocketChannel(BaseChannel):
 
     async def _hydrate_after_subscribe(self, chat_id: str) -> None:
         """Replay persisted or actively running per-chat state after subscribe."""
-        await self._maybe_push_active_goal_state(chat_id)
+        await self._maybe_push_persisted_goal_state(chat_id)
         await self._maybe_push_turn_run_wall_clock(chat_id)
 
     async def _send_event(
@@ -523,6 +592,61 @@ class WebSocketChannel(BaseChannel):
         except Exception as e:
             self.logger.warning("failed to send {} event: {}", event, e)
 
+    async def _broadcast_webui_event(self, event: str, **fields: Any) -> None:
+        for connection in tuple(self._webui_connections):
+            await self._send_event(connection, event, **fields)
+
+    async def _broadcast_user_message(
+        self,
+        origin: ServerConnection,
+        chat_id: str,
+        text: str,
+        *,
+        turn_id: str | None,
+        starts_turn: bool,
+        media_paths: list[str],
+        media_names: list[str | None],
+        cli_apps: list[dict[str, Any]],
+        mcp_presets: list[dict[str, Any]],
+        session_mentions: list[SessionMention],
+    ) -> None:
+        """Project one accepted user message to the other clients on the chat.
+
+        The origin already has an optimistic row and receives canonical turn
+        ownership in ``message_accepted``. Peers need the ingress projection.
+        """
+        body: dict[str, Any] = {
+            "event": "user_message",
+            "chat_id": chat_id,
+            "text": text,
+            "starts_turn": starts_turn,
+        }
+        if turn_id is not None:
+            body["turn_id"] = turn_id
+        media = self._media.augment_transcript_user_media(media_paths)
+        for attachment, name in zip(media, media_names, strict=False):
+            if name:
+                attachment["name"] = name
+        if media:
+            body["media_urls"] = media
+        if cli_apps:
+            body["cli_apps"] = cli_apps
+        if mcp_presets:
+            body["mcp_presets"] = mcp_presets
+        if session_mentions:
+            body["session_mentions"] = session_mentions
+        active_turn_id = websocket_turn_id(chat_id)
+        if active_turn_id is not None:
+            body["active_turn_id"] = active_turn_id
+        started_at = websocket_turn_wall_started_at(chat_id)
+        if active_turn_id is not None and started_at is not None:
+            body["started_at"] = started_at
+        raw = json.dumps(body, ensure_ascii=False)
+        for connection in tuple(self._subs.get(chat_id, ())):
+            if connection is origin:
+                continue
+            await self._safe_send_to(connection, raw, label=" user_message ")
+
     @classmethod
     def default_config(cls) -> dict[str, Any]:
         return WebSocketConfig().model_dump(by_alias=True)
@@ -758,6 +882,9 @@ class WebSocketChannel(BaseChannel):
     ) -> None:
         """Route one typed inbound envelope (``new_chat`` / ``attach`` / ``message``)."""
         t = envelope.get("type")
+        if t == "webui_request":
+            await self._start_webui_request(connection, envelope)
+            return
         if t == "new_chat":
             new_id = str(uuid.uuid4())
             scope = await self._workspace_scope_or_error(
@@ -771,7 +898,12 @@ class WebSocketChannel(BaseChannel):
                 return
             self._workspaces.persist_scope(new_id, scope)
             self._attach(connection, new_id)
-            await self._send_event(connection, "attached", chat_id=new_id)
+            await self._send_event(
+                connection,
+                "attached",
+                chat_id=new_id,
+                **self._attached_model_fields(new_id),
+            )
             await self._send_event(
                 connection,
                 "session_updated",
@@ -822,7 +954,12 @@ class WebSocketChannel(BaseChannel):
                 await self._send_event(connection, "error", detail=exc.detail, chat_id=cid)
                 return
             self._attach(connection, cid)
-            await self._send_event(connection, "attached", chat_id=cid)
+            await self._send_event(
+                connection,
+                "attached",
+                chat_id=cid,
+                **self._attached_model_fields(cid),
+            )
             await self._hydrate_after_subscribe(cid)
             return
         if t == "set_sidebar_state":
@@ -838,7 +975,7 @@ class WebSocketChannel(BaseChannel):
                 )
                 return
             try:
-                await asyncio.to_thread(
+                saved_state = await asyncio.to_thread(
                     write_webui_sidebar_state,
                     cast(dict[str, Any], state),
                 )
@@ -849,6 +986,11 @@ class WebSocketChannel(BaseChannel):
                     detail="invalid_sidebar_state",
                 )
                 return
+            await self._broadcast_webui_event(
+                "sidebar_state_updated",
+                state=saved_state,
+            )
+            return
         if t == "set_workspace_scope":
             cid = envelope.get("chat_id")
             if not _is_valid_chat_id(cid):
@@ -872,6 +1014,10 @@ class WebSocketChannel(BaseChannel):
             if scope is None:
                 return
             self._workspaces.persist_scope(cid, scope)
+            # Other clients on the same gateway only need an invalidation; they
+            # can reload the authoritative session row without receiving a
+            # local project path that belongs to another connection.
+            await self.send_session_updated(cid, scope="metadata")
             await self._send_event(
                 connection,
                 "session_updated",
@@ -881,7 +1027,10 @@ class WebSocketChannel(BaseChannel):
             )
             return
         if t == "transcribe_audio":
-            event, payload = await webui_transcription_event(envelope)
+            event, payload = await webui_transcription_event(
+                envelope,
+                config_path=self.gateway.settings.config.path,
+            )
             await self._send_event(connection, event, **payload)
             return
         if t == "message":
@@ -944,6 +1093,7 @@ class WebSocketChannel(BaseChannel):
 
             raw_media = envelope.get("media")
             media_paths: list[str] = []
+            media_names: list[str | None] = []
             if raw_media is not None:
                 if not isinstance(raw_media, list):
                     await self._send_event(
@@ -964,6 +1114,12 @@ class WebSocketChannel(BaseChannel):
                         **rejection_fields,
                     )
                     return
+                for item in cast(list[Any], raw_media):
+                    attachment = cast(dict[str, Any], item) if isinstance(item, dict) else {}
+                    name = attachment.get("name")
+                    media_names.append(
+                        (safe_filename(name) or None) if isinstance(name, str) else None
+                    )
                 if temporary_policy is not None:
                     self._temporary_chats.register_media(connection, cid, media_paths)
 
@@ -1017,10 +1173,25 @@ class WebSocketChannel(BaseChannel):
                 metadata["webui"] = True
                 metadata.update(self._transcripts.client_turn_metadata(envelope.get("turn_id")))
             trusted_webui = metadata.get("webui") is True and connection in self._webui_connections
+            is_user_shell = (
+                trusted_webui
+                and envelope.get("user_shell") is True
+                and content.startswith("!")
+            )
+            if is_user_shell:
+                metadata[INBOUND_META_USER_SHELL] = True
+            dispatch_content = (
+                f"{USER_SHELL_COMMAND} {content[1:].lstrip()}"
+                if is_user_shell
+                else content
+            )
             cli_apps = normalize_cli_app_mentions(envelope.get("cli_apps"))
             if cli_apps:
                 metadata["cli_apps"] = cli_apps
-            mcp_presets = normalize_mcp_preset_mentions(envelope.get("mcp_presets"))
+            mcp_presets = normalize_mcp_preset_mentions(
+                envelope.get("mcp_presets"),
+                config_path=self.gateway.settings.config.path,
+            )
             if mcp_presets:
                 metadata["mcp_presets"] = mcp_presets
             session_mentions: list[SessionMention] = []
@@ -1039,7 +1210,7 @@ class WebSocketChannel(BaseChannel):
             self._workspaces.persist_scope(cid, scope)
             is_webui = metadata.get("webui") is True
             queued_owner = None
-            if is_webui and builtin_command_starts_agent_turn(content):
+            if is_webui and not is_user_shell and builtin_command_starts_agent_turn(content):
                 queued_owner = register_queued_websocket_turn_if_idle(cid, turn_id)
                 if queued_owner is not None:
                     metadata[WEBSOCKET_TURN_OWNER_METADATA_KEY] = queued_owner
@@ -1076,7 +1247,7 @@ class WebSocketChannel(BaseChannel):
                 await self._handle_message(
                     sender_id=client_id,
                     chat_id=cid,
-                    content=content,
+                    content=dispatch_content,
                     media=media_paths or None,
                     metadata=metadata,
                     is_dm=False,
@@ -1095,16 +1266,282 @@ class WebSocketChannel(BaseChannel):
             finally:
                 if not accepted and queued_owner is not None:
                     clear_websocket_turn_if_current(cid, queued_owner)
+            if is_webui:
+                await self._broadcast_user_message(
+                    connection,
+                    cid,
+                    content,
+                    turn_id=turn_id,
+                    starts_turn=queued_owner is not None,
+                    media_paths=media_paths,
+                    media_names=media_names,
+                    cli_apps=cli_apps,
+                    mcp_presets=mcp_presets,
+                    session_mentions=session_mentions,
+                )
             if is_webui and turn_id:
+                active_turn_id = websocket_turn_id(cid)
+                started_at = websocket_turn_wall_started_at(cid)
                 await self._send_event(
                     connection,
                     "message_accepted",
                     chat_id=cid,
                     turn_id=turn_id,
+                    starts_turn=queued_owner is not None,
+                    **(
+                        {"active_turn_id": active_turn_id}
+                        if active_turn_id is not None
+                        else {}
+                    ),
+                    **(
+                        {"started_at": started_at}
+                        if active_turn_id is not None and started_at is not None
+                        else {}
+                    ),
                 )
             return
         await self._send_event(connection, "error", detail=f"unknown type: {t!r}")
 
+    async def _start_webui_request(
+        self,
+        connection: ServerConnection,
+        envelope: dict[str, Any],
+    ) -> None:
+        request_id = envelope.get("request_id")
+        if not isinstance(request_id, str) or re.fullmatch(
+            r"[A-Za-z0-9._:-]{1,128}",
+            request_id,
+        ) is None:
+            await self._send_event(
+                connection,
+                "error",
+                detail="invalid webui request_id",
+            )
+            return
+        if connection not in self._webui_connections:
+            await self._send_webui_response(
+                connection,
+                request_id,
+                status=403,
+                message="access_denied",
+            )
+            return
+
+        action = envelope.get("action")
+        payload = envelope.get("payload")
+        if not isinstance(action, str) or re.fullmatch(
+            r"[a-z][a-z0-9_.]{0,127}",
+            action,
+        ) is None:
+            await self._send_webui_response(
+                connection,
+                request_id,
+                status=400,
+                message="invalid WebUI mutation action",
+            )
+            return
+        if not isinstance(payload, dict):
+            await self._send_webui_response(
+                connection,
+                request_id,
+                status=400,
+                message="WebUI mutation payload must be an object",
+            )
+            return
+
+        payload_digest = hashlib.sha256(
+            json.dumps(
+                payload,
+                ensure_ascii=False,
+                sort_keys=True,
+                separators=(",", ":"),
+            ).encode("utf-8")
+        ).digest()
+        self._prune_webui_request_operations()
+        operation = self._webui_request_operations.get(request_id)
+        is_replay = operation is not None
+        if operation is not None and (
+            operation.action != action or operation.payload_digest != payload_digest
+        ):
+            await self._send_webui_response(
+                connection,
+                request_id,
+                status=409,
+                message="request_id was already used for a different WebUI mutation",
+            )
+            return
+        if operation is None:
+            operation_task = asyncio.create_task(
+                self._execute_webui_request(
+                    connection,
+                    action,
+                    cast(dict[str, Any], payload),
+                )
+            )
+            new_operation = _WebUIRequestOperation(
+                action=action,
+                payload_digest=payload_digest,
+                task=operation_task,
+            )
+            operation = new_operation
+            self._webui_request_operations[request_id] = new_operation
+
+            def mark_complete(_task: asyncio.Task[_WebUIRequestResult]) -> None:
+                current = self._webui_request_operations.get(request_id)
+                if current is not new_operation:
+                    return
+                new_operation.completed_at = time.monotonic()
+                self._prune_webui_request_operations()
+
+            operation_task.add_done_callback(mark_complete)
+
+        key = (connection, request_id)
+        if key in self._webui_request_tasks:
+            return
+        delivery_task = asyncio.create_task(
+            self._deliver_webui_request(
+                connection,
+                request_id,
+                operation.task,
+                sequence=is_replay,
+            )
+        )
+        self._webui_request_tasks[key] = delivery_task
+
+    def _prune_webui_request_operations(self) -> None:
+        now = time.monotonic()
+        for request_id, operation in tuple(self._webui_request_operations.items()):
+            if (
+                operation.completed_at is not None
+                and now - operation.completed_at >= _WEBUI_REQUEST_CACHE_TTL_S
+            ):
+                self._webui_request_operations.pop(request_id, None)
+
+        completed = sorted(
+            (
+                (operation.completed_at, request_id)
+                for request_id, operation in self._webui_request_operations.items()
+                if operation.completed_at is not None
+            ),
+            key=lambda item: item[0],
+        )
+        for _, request_id in completed[:-_WEBUI_REQUEST_CACHE_MAX]:
+            self._webui_request_operations.pop(request_id, None)
+
+    def _discard_webui_request_lock_if_idle(self, connection: ServerConnection) -> None:
+        if connection in self._webui_connections:
+            return
+        if any(task_connection is connection for task_connection, _ in self._webui_request_tasks):
+            return
+        self._webui_request_locks.pop(connection, None)
+
+    async def _deliver_webui_request(
+        self,
+        connection: ServerConnection,
+        request_id: str,
+        operation_task: asyncio.Task[_WebUIRequestResult],
+        *,
+        sequence: bool = False,
+    ) -> None:
+        try:
+            if sequence:
+                # Make replayed work the predecessor for subsequent mutations on
+                # this connection without blocking its receive loop.
+                lock = self._webui_request_locks.setdefault(connection, asyncio.Lock())
+                async with lock:
+                    result = await asyncio.shield(operation_task)
+                    await self._send_webui_response(
+                        connection,
+                        request_id,
+                        result=result.result,
+                        status=result.status,
+                        message=result.message,
+                    )
+                return
+            result = await asyncio.shield(operation_task)
+            await self._send_webui_response(
+                connection,
+                request_id,
+                result=result.result,
+                status=result.status,
+                message=result.message,
+            )
+        finally:
+            self._webui_request_tasks.pop((connection, request_id), None)
+            self._discard_webui_request_lock_if_idle(connection)
+
+    async def _execute_webui_request(
+        self,
+        connection: ServerConnection,
+        action: str,
+        payload: dict[str, Any],
+    ) -> _WebUIRequestResult:
+        try:
+            lock = self._webui_request_locks.setdefault(connection, asyncio.Lock())
+            async with lock:
+                response = await self._http_router.dispatch_webui_mutation(
+                    connection,
+                    action,
+                    payload,
+                )
+                status = response.status_code
+                body = bytes(response.body).decode("utf-8", errors="replace").strip()
+                if 200 <= status < 300:
+                    try:
+                        result = json.loads(body)
+                    except json.JSONDecodeError:
+                        return _WebUIRequestResult(
+                            status=502,
+                            message="WebUI mutation returned an invalid response",
+                        )
+                    if action == "sidebar.update" and isinstance(result, dict):
+                        await self._broadcast_webui_event(
+                            "sidebar_state_updated",
+                            state=result,
+                        )
+                    return _WebUIRequestResult(result=result)
+                return _WebUIRequestResult(
+                    status=status,
+                    message=body or response.reason_phrase,
+                )
+        except asyncio.CancelledError:
+            raise
+        except Exception:
+            self.logger.exception("WebUI mutation '{}' failed", action)
+            return _WebUIRequestResult(
+                status=500,
+                message="WebUI mutation failed",
+            )
+
+    async def _send_webui_response(
+        self,
+        connection: ServerConnection,
+        request_id: str,
+        *,
+        result: Any = None,
+        status: int | None = None,
+        message: str | None = None,
+    ) -> None:
+        if status is None:
+            await self._send_event(
+                connection,
+                "webui_response",
+                request_id=request_id,
+                ok=True,
+                result=result,
+            )
+            return
+        await self._send_event(
+            connection,
+            "webui_response",
+            request_id=request_id,
+            ok=False,
+            error={
+                "status": status,
+                "message": message or "WebUI mutation failed",
+            },
+        )
+
     async def _workspace_scope_or_error(
         self,
         connection: ServerConnection,
@@ -1145,6 +1582,19 @@ class WebSocketChannel(BaseChannel):
             except Exception as e:
                 self.logger.warning("server task error during shutdown: {}", e)
             self._server_task = None
+        delivery_tasks = tuple(self._webui_request_tasks.values())
+        operation_tasks = tuple(
+            operation.task for operation in self._webui_request_operations.values()
+        )
+        for task in (*delivery_tasks, *operation_tasks):
+            task.cancel()
+        if delivery_tasks:
+            await asyncio.gather(*delivery_tasks, return_exceptions=True)
+        if operation_tasks:
+            await asyncio.gather(*operation_tasks, return_exceptions=True)
+        self._webui_request_tasks.clear()
+        self._webui_request_locks.clear()
+        self._webui_request_operations.clear()
         self._subs.clear()
         self._conn_chats.clear()
         self._conn_default.clear()
@@ -1231,6 +1681,8 @@ class WebSocketChannel(BaseChannel):
                 await self.send_turn_model_updated(
                     msg.chat_id,
                     model_name=event.model,
+                    model_preset=event.model_preset,
+                    context_window_tokens=event.context_window_tokens,
                 )
             return
         if isinstance(event, GoalStateSyncEvent):
@@ -1275,6 +1727,8 @@ class WebSocketChannel(BaseChannel):
                 msg.chat_id,
                 latency_ms=event.latency_ms,
                 goal_state=event.goal_state,
+                usage=event.usage,
+                context_window_tokens=event.context_window_tokens,
                 metadata=msg.metadata,
                 turn_owner=turn_owner if isinstance(turn_owner, str) else None,
             )
@@ -1301,6 +1755,9 @@ class WebSocketChannel(BaseChannel):
             "chat_id": msg.chat_id,
             "text": wire_text,
         }
+        turn_id = msg.metadata.get(WEBUI_TURN_METADATA_KEY)
+        if isinstance(turn_id, str) and turn_id:
+            payload["turn_id"] = turn_id
         if msg.media:
             payload["media"] = msg.media
             urls: list[dict[str, str]] = []
@@ -1489,16 +1946,25 @@ class WebSocketChannel(BaseChannel):
         latency_ms: int | None = None,
         *,
         goal_state: dict[str, Any] | None = None,
+        usage: dict[str, int] | None = None,
+        context_window_tokens: int | None = None,
         metadata: dict[str, Any] | None = None,
         turn_owner: str | None = None,
     ) -> None:
         """Signal that the agent has fully finished processing the current turn."""
         conns = list(self._subs.get(chat_id, ()))
         body: dict[str, Any] = {"event": "turn_end", "chat_id": chat_id}
+        turn_id = (metadata or {}).get(WEBUI_TURN_METADATA_KEY)
+        if isinstance(turn_id, str) and turn_id:
+            body["turn_id"] = turn_id
         if latency_ms is not None:
             body["latency_ms"] = int(latency_ms)
         if goal_state is not None:
             body["goal_state"] = goal_state
+        if usage:
+            body["usage"] = usage
+        if context_window_tokens is not None:
+            body["context_window_tokens"] = int(context_window_tokens)
         canonical_webui_turn = (metadata or {}).get("webui") is True
         prior_persistence_failure = (
             canonical_webui_turn
@@ -1598,6 +2064,8 @@ class WebSocketChannel(BaseChannel):
         chat_id: str,
         *,
         model_name: Any,
+        model_preset: Any = None,
+        context_window_tokens: Any = None,
     ) -> None:
         """Notify one chat's subscribers which model is handling its current request."""
         conns = list(self._subs.get(chat_id, ()))
@@ -1612,6 +2080,10 @@ class WebSocketChannel(BaseChannel):
             "chat_id": chat_id,
             "model_name": model_name.strip(),
         }
+        if isinstance(model_preset, str) and model_preset.strip():
+            body["model_preset"] = model_preset.strip()
+        if isinstance(context_window_tokens, int) and context_window_tokens > 0:
+            body["context_window_tokens"] = context_window_tokens
         raw = json.dumps(body, ensure_ascii=False)
         for connection in conns:
             await self._safe_send_to(connection, raw, label=" turn_model_updated ")

nanobot/channels/websocket/tests/conftest.py

@@ -0,0 +1,14 @@
+"""Shared isolation for WebSocket tests that persist runtime state."""
+
+from pathlib import Path
+
+import pytest
+
+
+@pytest.fixture(autouse=True)
+def isolate_websocket_runtime_data(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Keep transcripts and other runtime files out of the active user data directory."""
+    monkeypatch.setattr("nanobot.config.paths.get_data_dir", lambda: tmp_path)

nanobot/channels/websocket/tests/test_websocket_envelope_media.py

@@ -129,9 +129,47 @@ async def test_webui_message_acceptance_echoes_turn_id() -> None:
         "event": "message_accepted",
         "chat_id": "abc123",
         "turn_id": "turn-accepted",
+        "starts_turn": True,
+        "active_turn_id": "turn-accepted",
+        "started_at": wth.websocket_turn_wall_started_at("abc123"),
     }
 
 
+@pytest.mark.asyncio
+async def test_webui_message_projects_attachments_to_other_clients(tmp_path: Path) -> None:
+    channel = _make_channel()
+    origin = AsyncMock()
+    peer = AsyncMock()
+    channel._attach(origin, "abc123")
+    channel._attach(peer, "abc123")
+    channel._webui_connections.add(origin)
+    envelope = {
+        "type": "message",
+        "chat_id": "abc123",
+        "content": "please inspect @drawio",
+        "webui": True,
+        "turn_id": "turn-shared",
+        "media": [{"data_url": _tiny_png_data_url(), "name": "shot.png"}],
+        "cli_apps": [{"name": "DrawIO", "entry_point": "cli-anything-drawio"}],
+    }
+
+    with patch("nanobot.webui.media_gateway.get_media_dir", return_value=tmp_path):
+        await channel._dispatch_envelope(origin, "client-1", envelope)
+
+    event = json.loads(peer.send.await_args.args[0])
+    assert event["event"] == "user_message"
+    assert event["turn_id"] == "turn-shared"
+    assert event["text"] == "please inspect @drawio"
+    assert event["cli_apps"] == [{
+        "name": "drawio",
+        "entry_point": "cli-anything-drawio",
+    }]
+    assert event["media_urls"][0]["kind"] == "image"
+    assert event["media_urls"][0]["name"] == "shot.png"
+    assert event["media_urls"][0]["url"].startswith("/api/media/")
+    assert json.loads(origin.send.await_args.args[0])["event"] == "message_accepted"
+
+
 @pytest.mark.asyncio
 async def test_message_text_policy_is_independent_from_transport_limit() -> None:
     channel = _make_channel()

nanobot/channels/websocket/tests/test_websocket_integration.py

@@ -334,6 +334,56 @@ async def test_independent_sessions(bus: MagicMock) -> None:
         await t
 
 
+@pytest.mark.asyncio
+async def test_same_session_projects_one_turn_to_both_clients(bus: MagicMock) -> None:
+    ch = _ch(bus, 29925)
+    t = asyncio.create_task(ch.start())
+    try:
+        async with WsTestClient("ws://127.0.0.1:29925/", client_id="terminal-a") as a:
+            async with WsTestClient("ws://127.0.0.1:29925/", client_id="terminal-b") as b:
+                chat_id = (await a.recv_ready()).chat_id
+                await b.recv_ready()
+                await b.send_json({"type": "attach", "chat_id": chat_id})
+                attached = await b.recv()
+                assert attached.event == "attached"
+                assert attached.chat_id == chat_id
+
+                await a.send_json(
+                    {
+                        "type": "message",
+                        "chat_id": chat_id,
+                        "content": "hello from terminal A",
+                        "webui": True,
+                        "turn_id": "turn-a",
+                    }
+                )
+
+                accepted = await a.recv()
+                projected = await b.recv()
+                assert accepted.event == "message_accepted"
+                assert accepted.raw["turn_id"] == "turn-a"
+                assert accepted.raw["starts_turn"] is True
+                assert accepted.raw["active_turn_id"] == "turn-a"
+                assert projected.raw == {
+                    "event": "user_message",
+                    "chat_id": chat_id,
+                    "text": "hello from terminal A",
+                    "starts_turn": True,
+                    "turn_id": "turn-a",
+                    "active_turn_id": "turn-a",
+                    "started_at": projected.raw["started_at"],
+                }
+
+                await ch.send_delta(chat_id, "shared reply", stream_id="stream-a")
+                assert (await a.recv_delta()).text == "shared reply"
+                assert (await b.recv_delta()).text == "shared reply"
+
+                assert bus.publish_inbound.await_count == 1
+    finally:
+        await ch.stop()
+        await t
+
+
 @pytest.mark.asyncio
 async def test_disconnected_client_cleanup(bus: MagicMock) -> None:
     ch = _ch(bus, 29914)

nanobot/channels/weixin/connect.py

@@ -22,6 +22,7 @@ class WeixinConnectSession:
     channel: WeixinChannel
     current_poll_base_url: str
     refresh_count: int
+    force: bool
     created_wall: float
     deadline: float
     last_error: str | None = None
@@ -72,7 +73,7 @@ class WeixinConnectStore:
 
         channel.connect_open_client()
         try:
-            qrcode_id, qr_url = await channel.connect_fetch_qr_code()
+            qrcode_id, qr_url = await channel.connect_fetch_qr_code(force=force)
         except Exception as exc:
             await self._close_channel(channel)
             raise ChannelConnectError(
@@ -89,6 +90,7 @@ class WeixinConnectStore:
             channel=channel,
             current_poll_base_url=channel.connect_base_url,
             refresh_count=0,
+            force=force,
             created_wall=now_wall,
             deadline=time.monotonic() + 600,
         )
@@ -187,7 +189,7 @@ class WeixinConnectStore:
                 }
             try:
                 session.qrcode_id, session.qr_url = (
-                    await session.channel.connect_fetch_qr_code()
+                    await session.channel.connect_fetch_qr_code(force=session.force)
                 )
             except Exception as exc:
                 self._sessions.pop(session_id, None)
@@ -204,6 +206,17 @@ class WeixinConnectStore:
             )
 
         if status == "binded_redirect":
+            if session.force:
+                self._sessions.pop(session_id, None)
+                await self._close_channel(session.channel)
+                return {
+                    "session_id": session_id,
+                    "status": "failed",
+                    "message": (
+                        "Unable to complete a new WeChat login. "
+                        "Start again and scan with the account you want to connect."
+                    ),
+                }
             if not session.channel.connect_load_state():
                 self._sessions.pop(session_id, None)
                 await self._close_channel(session.channel)
@@ -234,7 +247,7 @@ class WeixinConnectStore:
                 }
             try:
                 session.qrcode_id, session.qr_url = (
-                    await session.channel.connect_fetch_qr_code()
+                    await session.channel.connect_fetch_qr_code(force=session.force)
                 )
             except Exception as exc:
                 self._sessions.pop(session_id, None)

nanobot/channels/weixin/runtime.py

@@ -486,6 +486,35 @@ class WeixinChannel(BaseChannel):
         if base_url:
             self.config.base_url = base_url
         self._save_state(force=True)
+        self._persist_connect_credentials(token=token, base_url=base_url)
+
+    def _persist_connect_credentials(self, *, token: str, base_url: str) -> None:
+        """Write the QR-login token and base_url back to config.json.
+
+        The connect flow saves account state to ``account.json`` (via
+        ``_save_state``), but the WebUI's post-connect ``enable`` step calls
+        ``set_channel_config_enabled`` which reads config.json. Without
+        persisting the token here, that step would overwrite it with the
+        default empty value, losing the freshly obtained credential.
+        """
+        from nanobot.config.loader import get_config_path, load_config, save_config
+
+        try:
+            full_config = load_config()
+            section = getattr(full_config.channels, "weixin", None)
+            if section is not None and hasattr(section, "model_dump"):
+                values = section.model_dump(mode="json", by_alias=True)
+            elif isinstance(section, dict):
+                values = dict(cast(dict[str, Any], section))
+            else:
+                values = {}
+            values["token"] = token
+            if base_url:
+                values["baseUrl"] = base_url
+            setattr(full_config.channels, "weixin", values)
+            save_config(full_config, get_config_path())
+        except Exception:
+            self.logger.exception("Failed to persist WeChat credentials to config.json")
 
     # ------------------------------------------------------------------
     # HTTP helpers  (matches api.ts buildHeaders / apiFetch)
@@ -726,9 +755,9 @@ class WeixinChannel(BaseChannel):
                 break
         return tokens
 
-    async def _fetch_qr_code(self) -> tuple[str, str]:
-        """Fetch a fresh QR code. Returns (qrcode_id, scan_url)."""
-        local_tokens = self._local_token_list()
+    async def _fetch_qr_code(self, *, force: bool = False) -> tuple[str, str]:
+        """Fetch a QR code without existing credentials when forced."""
+        local_tokens = [] if force else self._local_token_list()
         data = await self._api_post(
             "ilink/bot/get_bot_qrcode?bot_type=3",
             {"local_token_list": local_tokens},
@@ -755,11 +784,11 @@ class WeixinChannel(BaseChannel):
             raise RuntimeError(f"Failed to get QR code from WeChat API: {data}")
         return qrcode_id, (qrcode_img_content or qrcode_id)
 
-    async def _qr_login(self) -> bool:
-        """Perform QR code login flow. Returns True on success."""
+    async def _qr_login(self, *, force: bool = False) -> bool:
+        """Perform QR login; forced flows accept only newly confirmed credentials."""
         try:
             refresh_count = 0
-            qrcode_id, scan_url = await self._fetch_qr_code()
+            qrcode_id, scan_url = await self._fetch_qr_code(force=force)
             self._print_qr_code(scan_url)
             current_poll_base_url = self.config.base_url
             verify_code = ""
@@ -825,11 +854,16 @@ class WeixinChannel(BaseChannel):
                     if refresh_count > MAX_QR_REFRESH_COUNT:
                         self.logger.warning("WeChat verification failed too many times")
                         return False
-                    qrcode_id, scan_url = await self._fetch_qr_code()
+                    qrcode_id, scan_url = await self._fetch_qr_code(force=force)
                     current_poll_base_url = self.config.base_url
                     self._print_qr_code(scan_url)
                     continue
                 elif status == "binded_redirect":
+                    if force:
+                        self.logger.error(
+                            "Forced WeChat login returned an existing binding without new credentials"
+                        )
+                        return False
                     if self._token or self._load_state():
                         self.logger.info("WeChat account is already connected")
                         return True
@@ -846,7 +880,7 @@ class WeixinChannel(BaseChannel):
                             MAX_QR_REFRESH_COUNT,
                         )
                         return False
-                    qrcode_id, scan_url = await self._fetch_qr_code()
+                    qrcode_id, scan_url = await self._fetch_qr_code(force=force)
                     current_poll_base_url = self.config.base_url
                     verify_code = ""
                     self._print_qr_code(scan_url)
@@ -893,8 +927,8 @@ class WeixinChannel(BaseChannel):
         self._client = self._new_http_client(httpx.Timeout(60, connect=30))
         self._running = True
 
-    async def connect_fetch_qr_code(self) -> tuple[str, str]:
-        return await self._fetch_qr_code()
+    async def connect_fetch_qr_code(self, *, force: bool = False) -> tuple[str, str]:
+        return await self._fetch_qr_code(force=force)
 
     async def connect_poll_qr_code(
         self,
@@ -947,14 +981,14 @@ class WeixinChannel(BaseChannel):
         if force:
             self._token = ""
             self._get_updates_buf = ""
-        if self._token or self._load_state():
+        if self._token or (not force and self._load_state()):
             return True
 
         # Initialize HTTP client for the login flow
         self._client = self._new_http_client(httpx.Timeout(60, connect=30))
         self._running = True  # Enable polling loop in _qr_login()
         try:
-            return await self._qr_login()
+            return await self._qr_login(force=force)
         finally:
             self._running = False
             if self._client:

nanobot/channels/weixin/tests/test_connect.py

@@ -25,7 +25,9 @@ async def test_weixin_connect_store_saves_confirmed_qr_login(
     )
     monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
+    async def fake_fetch_qr_code(
+        self: WeixinChannel, **_kwargs: Any
+    ) -> tuple[str, str]:
         return "qr-1", "https://qr.example/1"
 
     async def fake_api_get_with_base(
@@ -64,6 +66,63 @@ async def test_weixin_connect_store_saves_confirmed_qr_login(
     assert saved["token"] == "wx-token"
     assert saved["base_url"] == "https://weixin.example"
 
+    # Token and base_url must also be persisted to config.json so the
+    # post-connect enable step does not overwrite them with empty defaults.
+    config_data = json.loads(config_path.read_text(encoding="utf-8"))
+    weixin_cfg = config_data.get("channels", {}).get("weixin", {})
+    assert weixin_cfg.get("token") == "wx-token"
+    assert weixin_cfg.get("baseUrl") == "https://weixin.example"
+    assert weixin_cfg.get("stateDir") == str(state_dir)
+
+
+@pytest.mark.asyncio
+async def test_weixin_connect_persists_credentials_without_channels_config(
+    tmp_path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """When config.json has no channels key at all, connect must still write
+    the obtained token and base_url back to config.json."""
+    config_path = tmp_path / "config.json"
+    # config.json with NO channels key — the bug scenario
+    config_path.write_text(
+        json.dumps({"agents": {"defaults": {"model": "test"}}}),
+        encoding="utf-8",
+    )
+    monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
+
+    async def fake_fetch_qr_code(
+        self: WeixinChannel, **_kwargs: Any
+    ) -> tuple[str, str]:
+        return "qr-1", "https://qr.example/1"
+
+    async def fake_api_get_with_base(
+        self: WeixinChannel,
+        *,
+        base_url: str,
+        endpoint: str,
+        params: dict[str, Any],
+        auth: bool,
+    ) -> dict[str, str]:
+        return {
+            "status": "confirmed",
+            "bot_token": "wx-token",
+            "baseurl": "https://weixin.example",
+            "ilink_user_id": "wx-user",
+        }
+
+    monkeypatch.setattr(WeixinChannel, "_fetch_qr_code", fake_fetch_qr_code)
+    monkeypatch.setattr(WeixinChannel, "_api_get_with_base", fake_api_get_with_base)
+
+    store = WeixinConnectStore()
+    started = await store.start()
+    completed = await store.poll(started["session_id"])
+    assert completed["status"] == "succeeded"
+
+    config_data = json.loads(config_path.read_text(encoding="utf-8"))
+    weixin_cfg = config_data.get("channels", {}).get("weixin", {})
+    assert weixin_cfg.get("token") == "wx-token"
+    assert weixin_cfg.get("baseUrl") == "https://weixin.example"
+
 
 @pytest.mark.asyncio
 async def test_weixin_reconnect_keeps_existing_account_until_scan_succeeds(
@@ -86,14 +145,31 @@ async def test_weixin_reconnect_keeps_existing_account_until_scan_succeeds(
     )
     monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
-        return "qr-reconnect", "https://qr.example/reconnect"
+    observed_force: list[bool] = []
+
+    async def fake_fetch_qr_code(
+        self: WeixinChannel,
+        *,
+        force: bool = False,
+    ) -> tuple[str, str]:
+        observed_force.append(force)
+        return f"qr-reconnect-{len(observed_force)}", "https://qr.example/reconnect"
+
+    async def fake_api_get_with_base(
+        self: WeixinChannel,
+        **_kwargs: Any,
+    ) -> dict[str, str]:
+        return {"status": "expired"}
 
     monkeypatch.setattr(WeixinChannel, "_fetch_qr_code", fake_fetch_qr_code)
+    monkeypatch.setattr(WeixinChannel, "_api_get_with_base", fake_api_get_with_base)
 
     store = WeixinConnectStore()
     started = await store.start(force=True)
+    refreshed = await store.poll(started["session_id"])
 
+    assert refreshed["status"] == "pending"
+    assert observed_force == [True, True]
     assert json.loads(state_file.read_text(encoding="utf-8")) == existing
     cancelled = await store.cancel(started["session_id"])
     assert cancelled["status"] == "cancelled"
@@ -116,7 +192,9 @@ async def test_weixin_cancel_wins_over_inflight_confirmation(
     poll_started = asyncio.Event()
     release_poll = asyncio.Event()
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
+    async def fake_fetch_qr_code(
+        self: WeixinChannel, **_kwargs: Any
+    ) -> tuple[str, str]:
         return "qr-cancel", "https://qr.example/cancel"
 
     async def fake_api_get_with_base(
@@ -162,7 +240,9 @@ async def test_weixin_connect_store_handles_verification_code(
     )
     monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
+    async def fake_fetch_qr_code(
+        self: WeixinChannel, **_kwargs: Any
+    ) -> tuple[str, str]:
         return "qr-verify", "https://qr.example/verify"
 
     responses = [
@@ -204,7 +284,7 @@ async def test_weixin_connect_store_handles_verification_code(
 
 
 @pytest.mark.asyncio
-async def test_weixin_connect_store_treats_existing_binding_as_success(
+async def test_weixin_connect_store_rejects_existing_binding_during_forced_login(
     tmp_path,
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
@@ -221,7 +301,12 @@ async def test_weixin_connect_store_treats_existing_binding_as_success(
     )
     monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
+    async def fake_fetch_qr_code(
+        self: WeixinChannel,
+        *,
+        force: bool = False,
+    ) -> tuple[str, str]:
+        assert force is True
         return "qr-existing", "https://qr.example/existing"
 
     async def fake_api_get_with_base(
@@ -237,8 +322,8 @@ async def test_weixin_connect_store_treats_existing_binding_as_success(
     started = await store.start(force=True)
     completed = await store.poll(started["session_id"])
 
-    assert completed["status"] == "succeeded"
-    assert "already connected" in completed["message"]
+    assert completed["status"] == "failed"
+    assert "new WeChat login" in completed["message"]
     assert json.loads((state_dir / "account.json").read_text())["token"] == "working-token"
 
 
@@ -255,7 +340,9 @@ async def test_weixin_connect_store_rejects_existing_binding_without_local_crede
     )
     monkeypatch.setattr("nanobot.config.loader._current_config_path", config_path)
 
-    async def fake_fetch_qr_code(self: WeixinChannel) -> tuple[str, str]:
+    async def fake_fetch_qr_code(
+        self: WeixinChannel, **_kwargs: Any
+    ) -> tuple[str, str]:
         return "qr-missing", "https://qr.example/missing"
 
     async def fake_api_get_with_base(
@@ -268,7 +355,7 @@ async def test_weixin_connect_store_rejects_existing_binding_without_local_crede
     monkeypatch.setattr(WeixinChannel, "_api_get_with_base", fake_api_get_with_base)
 
     store = WeixinConnectStore()
-    started = await store.start(force=True)
+    started = await store.start(force=False)
     completed = await store.poll(started["session_id"])
 
     assert completed["status"] == "failed"

nanobot/channels/weixin/tests/test_weixin_channel.py

@@ -196,6 +196,86 @@ def test_save_state_with_empty_runtime_token_preserves_persisted_account(tmp_pat
     assert json.loads((tmp_path / "account.json").read_text()) == persisted
 
 
+@pytest.mark.asyncio
+async def test_login_force_ignores_persisted_account_through_qr_flow(tmp_path) -> None:
+    persisted = {
+        "token": "persisted-token",
+        "get_updates_buf": "persisted-cursor",
+        "context_tokens": {"wx-user": "ctx-persisted"},
+        "typing_tickets": {"wx-user": {"ticket": "ticket-persisted"}},
+        "base_url": "https://persisted.example",
+    }
+    channel = WeixinChannel(
+        WeixinConfig(
+            enabled=True,
+            allow_from=["*"],
+            token="configured-token",
+            state_dir=str(tmp_path),
+        ),
+        MessageBus(),
+    )
+    (tmp_path / "account.json").write_text(
+        json.dumps(persisted),
+        encoding="utf-8",
+    )
+    channel._print_qr_code = lambda _url: None
+    channel._api_post = AsyncMock(
+        side_effect=[
+            {"qrcode": "qr-1", "qrcode_img_content": "url-1"},
+            {"qrcode": "qr-2", "qrcode_img_content": "url-2"},
+        ]
+    )
+    channel._api_get_with_base = AsyncMock(
+        side_effect=[
+            {"status": "expired"},
+            {"status": "binded_redirect"},
+        ]
+    )
+
+    ok = await channel.login(force=True)
+
+    assert ok is False
+    assert [call.args[1]["local_token_list"] for call in channel._api_post.await_args_list] == [
+        [],
+        [],
+    ]
+    assert channel._token == ""
+    assert channel._get_updates_buf == ""
+    assert channel._context_tokens == {}
+    assert channel._typing_tickets == {}
+    assert channel.config.base_url == "https://ilinkai.weixin.qq.com"
+    assert json.loads((tmp_path / "account.json").read_text()) == persisted
+
+
+@pytest.mark.asyncio
+async def test_login_without_force_reuses_persisted_account(tmp_path) -> None:
+    channel = WeixinChannel(
+        WeixinConfig(enabled=True, allow_from=["*"], state_dir=str(tmp_path)),
+        MessageBus(),
+    )
+    (tmp_path / "account.json").write_text(
+        json.dumps(
+            {
+                "token": "persisted-token",
+                "get_updates_buf": "persisted-cursor",
+                "context_tokens": {"wx-user": "ctx-persisted"},
+                "base_url": "https://persisted.example",
+            }
+        ),
+        encoding="utf-8",
+    )
+    channel._qr_login = AsyncMock(return_value=False)
+
+    ok = await channel.login(force=False)
+
+    assert ok is True
+    channel._qr_login.assert_not_awaited()
+    assert channel._token == "persisted-token"
+    assert channel._get_updates_buf == "persisted-cursor"
+    assert channel._context_tokens == {"wx-user": "ctx-persisted"}
+    assert channel.config.base_url == "https://persisted.example"
+
+
 @pytest.mark.asyncio
 async def test_process_message_deduplicates_inbound_ids() -> None:
     channel, bus = _make_channel()

nanobot/channels/weixin/webui/WeixinPanel.tsx

@@ -27,33 +27,16 @@ import type {
   NanobotFeatureInfo,
 } from "@/lib/types";
 import { cn } from "@/lib/utils";
+import { useClient } from "@/providers/ClientProvider";
 
 import {
   WEIXIN_AUTH_EXPIRED_MESSAGE,
   WeixinConnectFlow,
 } from "./WeixinConnectFlow";
-
-export const WEIXIN_PRIMARY_FIELD_KEYS = [
-  "channels.weixin.sendProgress",
-  "channels.weixin.sendToolHints",
-  "channels.weixin.streaming",
-] as const;
-
-export const WEIXIN_ADVANCED_FIELD_KEYS = [
-  "channels.weixin.allowFrom",
-  "channels.weixin.token",
-  "channels.weixin.replyProgressMessages",
-  "channels.weixin.replyProgressMaxMessages",
-  "channels.weixin.contextMessageBudget",
-  "channels.weixin.blockStreaming",
-  "channels.weixin.blockStreamingMinChars",
-  "channels.weixin.blockStreamingMaxMessages",
-  "channels.weixin.baseUrl",
-  "channels.weixin.cdnBaseUrl",
-  "channels.weixin.routeTag",
-  "channels.weixin.stateDir",
-  "channels.weixin.pollTimeout",
-] as const;
+import {
+  WEIXIN_ADVANCED_FIELD_KEYS,
+  WEIXIN_PRIMARY_FIELD_KEYS,
+} from "./presentation";
 
 export function WeixinPanel({
   token,
@@ -64,6 +47,7 @@ export function WeixinPanel({
   onAction,
   onFeaturesUpdate,
 }: ChannelPluginPanelProps) {
+  const { client } = useClient();
   const { t, i18n } = useTranslation();
   const tx = (key: string, fallback: string) => t(key, { defaultValue: fallback });
   const channelTx = channelTranslator(t, "weixin");
@@ -150,7 +134,7 @@ export function WeixinPanel({
     setSaveState("idle");
     try {
       const payload = await configureChannel(
-        context.token,
+        client,
         "weixin",
         channelValuesForSave(editableFieldsRef.current, values),
         { enable: context.enabled },
@@ -168,7 +152,7 @@ export function WeixinPanel({
     } finally {
       setSaving(false);
     }
-  }, []);
+  }, [client]);
 
   useEffect(() => {
     if (
@@ -213,7 +197,7 @@ export function WeixinPanel({
   });
 
   return (
-    <aside className="min-h-full rounded-[20px] bg-settings-surface p-5">
+    <aside className="min-h-full rounded-panel bg-settings-surface p-5">
       <div className="flex items-start justify-between gap-4">
         <div className="flex min-w-0 items-start gap-3">
           <WeixinLogo showBrandLogos={showBrandLogos} />
@@ -267,7 +251,7 @@ export function WeixinPanel({
       </div>
 
       {runtimeError ? (
-        <div className="mt-4 rounded-[12px] border border-destructive/20 bg-destructive/5 px-3 py-2 text-[12px] leading-5 text-destructive">
+        <div className="mt-4 rounded-control border border-destructive/20 bg-destructive/5 px-3 py-2 text-[12px] leading-5 text-destructive">
           {runtimeError}
         </div>
       ) : null}
@@ -320,7 +304,7 @@ export function WeixinPanel({
         {saveError ? (
           <div
             role="alert"
-            className="rounded-[12px] border border-destructive/20 bg-destructive/5 px-3 py-2 text-[12px] leading-5 text-destructive"
+            className="rounded-control border border-destructive/20 bg-destructive/5 px-3 py-2 text-[12px] leading-5 text-destructive"
           >
             {saveError}
           </div>
@@ -429,7 +413,7 @@ function WeixinLogo({ showBrandLogos }: { showBrandLogos: boolean }) {
   const { logoUrl, onLogoError, onLogoLoad } = useLogoFallback(logoUrls);
   if (showBrandLogos && logoUrl) {
     return (
-      <span className="grid h-10 w-10 shrink-0 place-items-center rounded-[12px] bg-background">
+      <span className="grid h-10 w-10 shrink-0 place-items-center rounded-control bg-background">
         <img
           src={logoUrl}
           alt=""
@@ -444,7 +428,7 @@ function WeixinLogo({ showBrandLogos }: { showBrandLogos: boolean }) {
   }
   return (
     <span
-      className="flex h-10 w-10 shrink-0 items-center justify-center rounded-[12px] bg-background text-[11px] font-bold"
+      className="flex h-10 w-10 shrink-0 items-center justify-center rounded-control bg-background text-[11px] font-bold"
       style={{ color: "#07C160" }}
       aria-hidden
     >

nanobot/channels/weixin/webui/index.tsx

@@ -1,12 +1,21 @@
+import { lazy } from "react";
+
 import type { ChannelUiContribution } from "@/channel-plugins/types";
 import { chatAppGuideUrl } from "@/components/settings/channels/catalog";
 
-import { WeixinConnectFlow } from "./WeixinConnectFlow";
 import {
   WEIXIN_ADVANCED_FIELD_KEYS,
   WEIXIN_PRIMARY_FIELD_KEYS,
-  WeixinPanel,
-} from "./WeixinPanel";
+} from "./presentation";
+
+const WeixinPanel = lazy(() =>
+  import("./WeixinPanel").then(({ WeixinPanel: component }) => ({ default: component })),
+);
+const WeixinConnectFlow = lazy(() =>
+  import("./WeixinConnectFlow").then(({ WeixinConnectFlow: component }) => ({
+    default: component,
+  })),
+);
 
 export default {
   Panel: WeixinPanel,

nanobot/channels/weixin/webui/presentation.ts

@@ -0,0 +1,21 @@
+export const WEIXIN_PRIMARY_FIELD_KEYS = [
+  "channels.weixin.sendProgress",
+  "channels.weixin.sendToolHints",
+  "channels.weixin.streaming",
+] as const;
+
+export const WEIXIN_ADVANCED_FIELD_KEYS = [
+  "channels.weixin.allowFrom",
+  "channels.weixin.token",
+  "channels.weixin.replyProgressMessages",
+  "channels.weixin.replyProgressMaxMessages",
+  "channels.weixin.contextMessageBudget",
+  "channels.weixin.blockStreaming",
+  "channels.weixin.blockStreamingMinChars",
+  "channels.weixin.blockStreamingMaxMessages",
+  "channels.weixin.baseUrl",
+  "channels.weixin.cdnBaseUrl",
+  "channels.weixin.routeTag",
+  "channels.weixin.stateDir",
+  "channels.weixin.pollTimeout",
+] as const;

nanobot/cli/agent.py

@@ -1,52 +1,18 @@
-"""Direct and interactive agent CLI command."""
+"""Agent CLI command."""
 
-import asyncio
-import signal
 import sys
-from collections.abc import Awaitable, Callable
-from types import FrameType
-from typing import Any
 
 import typer
 from rich.console import Console
 
-from nanobot import __logo__
-from nanobot.agent.hooks import create_file_edit_activity_hook
-from nanobot.agent.loop import AgentLoop
-from nanobot.bus.outbound_events import (
-    StreamDeltaEvent,
-    StreamedResponseEvent,
-    StreamEndEvent,
-    outbound_event_from_message,
-)
-from nanobot.cli import terminal as cli_terminal
-from nanobot.cli.log_control import _set_nanobot_logs
-from nanobot.cli.runtime_config import (
-    _load_runtime_config,
-    _migrate_cron_store,
-    _model_display,
-    _print_agent_start_error,
-)
-from nanobot.cli.stream import StreamRenderer, ThinkingSpinner
-from nanobot.config.paths import is_default_workspace
-from nanobot.utils.helpers import (
-    sanitize_surrogates as _sanitize_surrogates,
-)
-from nanobot.utils.helpers import (
-    sync_workspace_templates,
-)
-from nanobot.utils.restart import (
-    consume_restart_notice_from_env,
-    format_restart_completed_message,
-    should_show_cli_restart_notice,
-)
+from nanobot.cli.runtime_config import _load_runtime_config
 
 console = Console()
 
 
 def agent(
-    message: str = typer.Option(None, "--message", "-m", help="Message to send to the agent"),
-    session_id: str = typer.Option("cli:direct", "--session", "-s", help="Session ID"),
+    message: str | None = typer.Option(None, "--message", "-m", help="Message to send to the agent"),
+    session_id: str | None = typer.Option(None, "--session", "-s", help="Session ID"),
     workspace: str | None = typer.Option(None, "--workspace", "-w", help="Workspace directory"),
     config: str | None = typer.Option(None, "--config", "-c", help="Path to config file"),
     markdown: bool = typer.Option(
@@ -59,294 +25,63 @@ def agent(
         "--logs/--no-logs",
         help="Show nanobot runtime logs during chat",
     ),
-):
-    """Interact with the agent directly."""
-    from nanobot.bus.queue import MessageBus
-    from nanobot.cron.service import CronService
-    from nanobot.providers.factory import make_provider
-    from nanobot.providers.image_generation import image_gen_provider_configs
-
+    classic: bool = typer.Option(
+        False,
+        "--classic",
+        help="Use the compatibility Python prompt instead of the terminal UI",
+    ),
+    theme: str = typer.Option(
+        "auto",
+        "--theme",
+        help="Native terminal UI appearance: auto, dark, or light",
+    ),
+) -> None:
+    """Chat in the terminal or send one message non-interactively."""
     runtime_config = _load_runtime_config(config, workspace)
+    theme = theme.strip().lower()
+    if theme not in {"auto", "dark", "light"}:
+        raise typer.BadParameter("must be auto, dark, or light", param_hint="--theme")
+
+    if message is None and not classic:
+        from nanobot.cli.tui_launcher import TuiSessionError, TuiUnavailableError, launch_tui
+        from nanobot.config.loader import get_config_path
+
+        if not sys.stdin.isatty() or not sys.stdout.isatty():
+            raise typer.BadParameter(
+                "the native TUI requires an interactive terminal; use --message for "
+                "one-shot input or --classic for the compatibility prompt",
+                param_hint="terminal",
+            )
+        if not markdown:
+            raise typer.BadParameter("--no-markdown requires --classic", param_hint="--no-markdown")
+        if logs:
+            raise typer.BadParameter("--logs requires --classic", param_hint="--logs")
         try:
-        provider = make_provider(runtime_config)
-    except ValueError as exc:
-        _print_agent_start_error(exc)
-        raise typer.Exit(1) from exc
-
-    sync_workspace_templates(runtime_config.workspace_path)
-
-    bus = MessageBus()
-
-    # Preserve existing single-workspace installs, but keep custom workspaces clean.
-    if is_default_workspace(runtime_config.workspace_path):
-        _migrate_cron_store(runtime_config)
-
-    # Create cron service with workspace-scoped store
-    cron_store_path = runtime_config.workspace_path / "cron" / "jobs.json"
-    cron = CronService(cron_store_path)
-
-    _set_nanobot_logs(logs)
-
-    try:
-        agent_loop = AgentLoop.from_config(
+            exit_code = launch_tui(
                 runtime_config,
-            bus,
-            provider=provider,
-            cron_service=cron,
-            image_generation_provider_configs=image_gen_provider_configs(runtime_config),
-            hook_factories=[create_file_edit_activity_hook],
+                config_path=get_config_path().resolve(strict=False),
+                workspace_override=workspace,
+                session_id=session_id,
+                theme=theme,
             )
-    except ValueError as exc:
-        _print_agent_start_error(exc)
-        raise typer.Exit(1) from exc
-    restart_notice = consume_restart_notice_from_env()
-    if restart_notice and should_show_cli_restart_notice(restart_notice, session_id):
-        cli_terminal._print_agent_response(
-            format_restart_completed_message(restart_notice.started_at_raw),
-            render_markdown=False,
-        )
-
-    # Shared reference for progress callbacks
-    _thinking: ThinkingSpinner | None = None
-
-    def _make_progress(
-        renderer: StreamRenderer | None = None,
-    ) -> Callable[..., Awaitable[None]]:
-        reasoning_buffer = cli_terminal._ReasoningBuffer()
-
-        async def _cli_progress(
-            content: str,
-            *,
-            tool_hint: bool = False,
-            reasoning: bool = False,
-            **_kwargs: Any,
-        ) -> None:
-            ch = agent_loop.channels_config
-
-            if _kwargs.get("reasoning_end"):
-                if ch and not ch.show_reasoning:
-                    reasoning_buffer.clear()
-                else:
-                    cli_terminal._flush_cli_reasoning(reasoning_buffer, _thinking, renderer)
-                return
-
-            if reasoning:
-                if ch and not ch.show_reasoning:
-                    reasoning_buffer.clear()
-                    return
-                text = reasoning_buffer.add(content)
-                if text:
-                    cli_terminal._print_cli_reasoning(text, _thinking, renderer)
-                return
-            if ch and tool_hint and not ch.send_tool_hints:
-                return
-            if ch and not tool_hint and not ch.send_progress:
-                return
-            cli_terminal._print_cli_progress_line(content, _thinking, renderer)
-
-        return _cli_progress
-
-    if message:
-        # Single message mode — direct call, no bus needed
-        async def run_once() -> None:
-            renderer = StreamRenderer(
-                render_markdown=markdown,
-                bot_name=runtime_config.agents.defaults.bot_name,
-                bot_icon=runtime_config.agents.defaults.bot_icon,
-            )
-            response = await agent_loop.process_direct(
-                message,
-                session_id,
-                on_progress=_make_progress(renderer),
-                on_stream=renderer.on_delta,
-                on_stream_end=renderer.on_end,
-            )
-            if not renderer.streamed:
-                await renderer.close()
-                print_kwargs: dict[str, Any] = {}
-                if renderer.header_printed:
-                    print_kwargs["show_header"] = False
-                cli_terminal._print_agent_response(
-                    response.content if response else "",
-                    render_markdown=markdown,
-                    metadata=response.metadata if response else None,
-                    **print_kwargs,
-                )
-            await agent_loop.close_mcp()
-
-        asyncio.run(run_once())
-    else:
-        # Interactive mode — route through bus like other channels
-        from nanobot.bus.events import InboundMessage
-
-        cli_terminal._init_prompt_session()
-        _model, _preset_tag = _model_display(runtime_config)
-        _icon = runtime_config.agents.defaults.bot_icon or __logo__
+        except TuiSessionError as exc:
+            raise typer.BadParameter(str(exc), param_hint="--session") from exc
+        except TuiUnavailableError as exc:
+            console.print(f"[red]Native TUI unavailable: {exc}[/red]")
             console.print(
-            f"{_icon} Interactive mode [bold blue]({_model})[/bold blue]{_preset_tag} "
-            "— type [bold]exit[/bold] or [bold]Ctrl+C[/bold] to quit\n"
+                "[dim]Use `nanobot agent --classic` only if you want the compatibility prompt.[/dim]"
             )
+            raise typer.Exit(1) from exc
+        if exit_code:
+            raise typer.Exit(exit_code)
+        return
 
-        if ":" in session_id:
-            cli_channel, cli_chat_id = session_id.split(":", 1)
-        else:
-            cli_channel, cli_chat_id = "cli", session_id
+    from nanobot.cli.agent_runtime import run_local_agent
 
-        def _handle_signal(signum: int, _frame: FrameType | None) -> None:
-            sig_name = signal.Signals(signum).name
-            cli_terminal._restore_terminal()
-            console.print(f"\nReceived {sig_name}, goodbye!")
-            sys.exit(0)
-
-        signal.signal(signal.SIGINT, _handle_signal)
-        signal.signal(signal.SIGTERM, _handle_signal)
-        # SIGHUP is not available on Windows
-        if hasattr(signal, "SIGHUP"):
-            signal.signal(signal.SIGHUP, _handle_signal)
-        # Ignore SIGPIPE to prevent silent process termination when writing to closed pipes
-        # SIGPIPE is not available on Windows
-        if hasattr(signal, "SIGPIPE"):
-            signal.signal(signal.SIGPIPE, signal.SIG_IGN)
-
-        async def run_interactive() -> None:
-            bus_task = asyncio.create_task(agent_loop.run())
-            turn_done = asyncio.Event()
-            turn_done.set()
-            turn_response: list[Any] = []
-            renderer: StreamRenderer | None = None
-            reasoning_buffer = cli_terminal._ReasoningBuffer()
-
-            async def _consume_outbound() -> None:
-                while True:
-                    try:
-                        msg = await asyncio.wait_for(bus.consume_outbound(), timeout=1.0)
-                        event = outbound_event_from_message(msg)
-
-                        if isinstance(event, StreamDeltaEvent):
-                            if renderer:
-                                await renderer.on_delta(msg.content)
-                            continue
-                        if isinstance(event, StreamEndEvent):
-                            if renderer:
-                                await renderer.on_end(
-                                    resuming=event.resuming,
+    run_local_agent(
+        runtime_config,
+        message=message,
+        session_id=session_id or "cli:direct",
+        markdown=markdown,
+        logs=logs,
     )
-                            continue
-                        if isinstance(event, StreamedResponseEvent):
-                            if msg.content and renderer and not renderer.streamed:
-                                await renderer.close()
-                                print_kwargs: dict[str, Any] = {}
-                                if renderer.header_printed:
-                                    print_kwargs["show_header"] = False
-                                cli_terminal._print_agent_response(
-                                    msg.content,
-                                    render_markdown=markdown,
-                                    metadata=msg.metadata,
-                                    **print_kwargs,
-                                )
-                            turn_done.set()
-                            continue
-
-                        if await cli_terminal._maybe_print_interactive_progress(
-                            msg,
-                            None,
-                            agent_loop.channels_config,
-                            renderer,
-                            reasoning_buffer,
-                        ):
-                            continue
-
-                        if not turn_done.is_set():
-                            if msg.content:
-                                turn_response.append(msg)
-                            turn_done.set()
-                        elif msg.content:
-                            await cli_terminal._print_interactive_response(
-                                msg.content,
-                                render_markdown=markdown,
-                                metadata=msg.metadata,
-                            )
-
-                    except asyncio.TimeoutError:
-                        continue
-                    except asyncio.CancelledError:
-                        break
-
-            outbound_task = asyncio.create_task(_consume_outbound())
-
-            try:
-                while True:
-                    try:
-                        cli_terminal._flush_pending_tty_input()
-                        # Stop spinner before user input to avoid prompt_toolkit conflicts
-                        if renderer:
-                            renderer.stop_for_input()
-                        user_input = _sanitize_surrogates(
-                            await cli_terminal._read_interactive_input_async()
-                        )
-                        command = user_input.strip()
-                        if not command:
-                            continue
-
-                        if cli_terminal._is_exit_command(command):
-                            cli_terminal._restore_terminal()
-                            console.print("\nGoodbye!")
-                            break
-
-                        turn_done.clear()
-                        turn_response.clear()
-                        reasoning_buffer.clear()
-                        renderer = StreamRenderer(
-                            render_markdown=markdown,
-                            bot_name=runtime_config.agents.defaults.bot_name,
-                            bot_icon=runtime_config.agents.defaults.bot_icon,
-                        )
-
-                        await bus.publish_inbound(
-                            InboundMessage(
-                                channel=cli_channel,
-                                sender_id="user",
-                                chat_id=cli_chat_id,
-                                content=user_input,
-                                metadata={"_wants_stream": True},
-                            )
-                        )
-
-                        await turn_done.wait()
-
-                        if turn_response:
-                            response_msg = turn_response[0]
-                            content = response_msg.content
-                            meta = response_msg.metadata
-                            if content and not isinstance(
-                                response_msg.event,
-                                StreamedResponseEvent,
-                            ):
-                                if renderer:
-                                    await renderer.close()
-                                print_kwargs: dict[str, Any] = {}
-                                if renderer and renderer.header_printed:
-                                    print_kwargs["show_header"] = False
-                                cli_terminal._print_agent_response(
-                                    content,
-                                    render_markdown=markdown,
-                                    metadata=meta,
-                                    **print_kwargs,
-                                )
-                        elif renderer and not renderer.streamed:
-                            await renderer.close()
-                    except KeyboardInterrupt:
-                        cli_terminal._restore_terminal()
-                        console.print("\nGoodbye!")
-                        break
-                    except EOFError:
-                        cli_terminal._restore_terminal()
-                        console.print("\nGoodbye!")
-                        break
-            finally:
-                agent_loop.stop()
-                outbound_task.cancel()
-                await asyncio.gather(bus_task, outbound_task, return_exceptions=True)
-                await agent_loop.close_mcp()
-
-        asyncio.run(run_interactive())

nanobot/cli/agent_runtime.py

@@ -0,0 +1,308 @@
+"""Python runtime for one-shot agent calls and the compatibility prompt."""
+
+import asyncio
+import signal
+import sys
+from types import FrameType
+from typing import Any
+
+import typer
+
+from nanobot import __logo__
+from nanobot.agent.hooks import create_file_edit_activity_hook
+from nanobot.agent.loop import AgentLoop
+from nanobot.agent.tools.mcp import MCPProvider
+from nanobot.agent.tools.registry import ToolRegistry
+from nanobot.bus.events import InboundMessage, OutboundMessage
+from nanobot.bus.outbound_events import (
+    StreamDeltaEvent,
+    StreamedResponseEvent,
+    StreamEndEvent,
+    outbound_event_from_message,
+)
+from nanobot.bus.queue import MessageBus
+from nanobot.cli import terminal as cli_terminal
+from nanobot.cli.log_control import _set_nanobot_logs
+from nanobot.cli.runtime_config import (
+    _migrate_cron_store,
+    _model_display,
+    _print_agent_start_error,
+)
+from nanobot.cli.stream import StreamRenderer
+from nanobot.config.paths import is_default_workspace
+from nanobot.config.schema import Config
+from nanobot.cron.service import CronService
+from nanobot.providers.factory import make_provider
+from nanobot.providers.image_generation import image_gen_provider_configs
+from nanobot.utils.helpers import sanitize_surrogates, sync_workspace_templates
+from nanobot.utils.restart import (
+    consume_restart_notice_from_env,
+    format_restart_completed_message,
+    should_show_cli_restart_notice,
+)
+
+
+def run_local_agent(
+    config: Config,
+    *,
+    message: str | None,
+    session_id: str,
+    markdown: bool,
+    logs: bool,
+) -> None:
+    """Run without the gateway: once for a message, otherwise as the classic prompt."""
+    runtime = _LocalAgent(config, logs=logs, session_id=session_id)
+    if message is not None:
+        asyncio.run(runtime.run_once(message, session_id=session_id, markdown=markdown))
+    else:
+        runtime.run_classic(session_id=session_id, markdown=markdown)
+
+
+class _LocalAgent:
+    def __init__(self, config: Config, *, logs: bool, session_id: str) -> None:
+        self.config = config
+        try:
+            provider = make_provider(config)
+        except ValueError as exc:
+            _print_agent_start_error(exc)
+            raise typer.Exit(1) from exc
+
+        sync_workspace_templates(config.workspace_path)
+        if is_default_workspace(config.workspace_path):
+            _migrate_cron_store(config)
+
+        self.bus = MessageBus()
+        tools = ToolRegistry()
+        self.mcp = MCPProvider.from_config(config, tools)
+        _set_nanobot_logs(logs)
+        try:
+            self.loop = AgentLoop.from_config(
+                config,
+                self.bus,
+                provider=provider,
+                cron_service=CronService(config.workspace_path / "cron" / "jobs.json"),
+                image_generation_provider_configs=image_gen_provider_configs(config),
+                hook_factories=[create_file_edit_activity_hook],
+                tool_registry=tools,
+            )
+        except ValueError as exc:
+            _print_agent_start_error(exc)
+            raise typer.Exit(1) from exc
+
+        notice = consume_restart_notice_from_env()
+        if notice and should_show_cli_restart_notice(notice, session_id):
+            cli_terminal._print_agent_response(
+                format_restart_completed_message(notice.started_at_raw),
+                render_markdown=False,
+            )
+
+    async def close(self) -> None:
+        try:
+            await self.loop.aclose()
+        finally:
+            await self.mcp.aclose()
+
+    def renderer(self, markdown: bool) -> StreamRenderer:
+        return StreamRenderer(
+            render_markdown=markdown,
+            bot_name=self.config.agents.defaults.bot_name,
+            bot_icon=self.config.agents.defaults.bot_icon,
+        )
+
+    async def run_once(self, message: str, *, session_id: str, markdown: bool) -> None:
+        try:
+            await self.mcp.connect()
+            renderer = self.renderer(markdown)
+            reasoning_buffer = cli_terminal._ReasoningBuffer()
+
+            async def report(
+                content: str,
+                *,
+                tool_hint: bool = False,
+                reasoning: bool = False,
+                **kwargs: Any,
+            ) -> None:
+                channel_config = self.loop.channels_config
+                if kwargs.get("reasoning_end"):
+                    if channel_config and not channel_config.show_reasoning:
+                        reasoning_buffer.clear()
+                    else:
+                        cli_terminal._flush_cli_reasoning(reasoning_buffer, None, renderer)
+                    return
+                if reasoning:
+                    if channel_config and not channel_config.show_reasoning:
+                        reasoning_buffer.clear()
+                        return
+                    text = reasoning_buffer.add(content)
+                    if text:
+                        cli_terminal._print_cli_reasoning(text, None, renderer)
+                    return
+                if channel_config and tool_hint and not channel_config.send_tool_hints:
+                    return
+                if channel_config and not tool_hint and not channel_config.send_progress:
+                    return
+                cli_terminal._print_cli_progress_line(content, None, renderer)
+
+            response = await self.loop.process_direct(
+                message,
+                session_id,
+                on_progress=report,
+                on_stream=renderer.on_delta,
+                on_stream_end=renderer.on_end,
+            )
+            if renderer.streamed:
+                return
+            await renderer.close()
+            cli_terminal._print_agent_response(
+                response.content if response else "",
+                render_markdown=markdown,
+                metadata=response.metadata if response else None,
+                **({"show_header": False} if renderer.header_printed else {}),
+            )
+        finally:
+            await self.close()
+
+    def run_classic(self, *, session_id: str, markdown: bool) -> None:
+        cli_terminal._init_prompt_session()
+        model, preset_tag = _model_display(self.config)
+        icon = self.config.agents.defaults.bot_icon or __logo__
+        cli_terminal.console.print(
+            f"{icon} Interactive mode [bold blue]({model})[/bold blue]{preset_tag} "
+            "— type [bold]exit[/bold] or [bold]Ctrl+C[/bold] to quit\n"
+        )
+        channel, chat_id = (
+            session_id.split(":", 1) if ":" in session_id else ("cli", session_id)
+        )
+        self._install_signal_handlers()
+        asyncio.run(self._run_classic_loop(channel, chat_id, markdown=markdown))
+
+    @staticmethod
+    def _install_signal_handlers() -> None:
+        def exit_on_signal(signum: int, _frame: FrameType | None) -> None:
+            cli_terminal._restore_terminal()
+            cli_terminal.console.print(f"\nReceived {signal.Signals(signum).name}, goodbye!")
+            sys.exit(0)
+
+        signal.signal(signal.SIGINT, exit_on_signal)
+        signal.signal(signal.SIGTERM, exit_on_signal)
+        if hasattr(signal, "SIGHUP"):
+            signal.signal(signal.SIGHUP, exit_on_signal)
+        if hasattr(signal, "SIGPIPE"):
+            signal.signal(signal.SIGPIPE, signal.SIG_IGN)
+
+    async def _run_classic_loop(self, channel: str, chat_id: str, *, markdown: bool) -> None:
+        await self.mcp.connect()
+        bus_task = asyncio.create_task(self.loop.run())
+        turn_done = asyncio.Event()
+        turn_done.set()
+        turn_response: list[OutboundMessage] = []
+        renderer: StreamRenderer | None = None
+        reasoning_buffer = cli_terminal._ReasoningBuffer()
+
+        async def consume_outbound() -> None:
+            while True:
+                try:
+                    msg = await asyncio.wait_for(self.bus.consume_outbound(), timeout=1.0)
+                    event = outbound_event_from_message(msg)
+                    if isinstance(event, StreamDeltaEvent):
+                        if renderer:
+                            await renderer.on_delta(msg.content)
+                        continue
+                    if isinstance(event, StreamEndEvent):
+                        if renderer:
+                            await renderer.on_end(resuming=event.resuming)
+                        continue
+                    if isinstance(event, StreamedResponseEvent):
+                        if msg.content and renderer and not renderer.streamed:
+                            await renderer.close()
+                            cli_terminal._print_agent_response(
+                                msg.content,
+                                render_markdown=markdown,
+                                metadata=msg.metadata,
+                                **({"show_header": False} if renderer.header_printed else {}),
+                            )
+                        turn_done.set()
+                        continue
+                    if await cli_terminal._maybe_print_interactive_progress(
+                        msg,
+                        None,
+                        self.loop.channels_config,
+                        renderer,
+                        reasoning_buffer,
+                    ):
+                        continue
+                    if not turn_done.is_set():
+                        if msg.content:
+                            turn_response.append(msg)
+                        turn_done.set()
+                    elif msg.content:
+                        await cli_terminal._print_interactive_response(
+                            msg.content,
+                            render_markdown=markdown,
+                            metadata=msg.metadata,
+                        )
+                except asyncio.TimeoutError:
+                    continue
+                except asyncio.CancelledError:
+                    break
+
+        outbound_task = asyncio.create_task(consume_outbound())
+        try:
+            while True:
+                try:
+                    cli_terminal._flush_pending_tty_input()
+                    if renderer:
+                        renderer.stop_for_input()
+                    user_input = sanitize_surrogates(
+                        await cli_terminal._read_interactive_input_async()
+                    )
+                    command = user_input.strip()
+                    if not command:
+                        continue
+                    if cli_terminal._is_exit_command(command):
+                        cli_terminal._restore_terminal()
+                        cli_terminal.console.print("\nGoodbye!")
+                        break
+
+                    turn_done.clear()
+                    turn_response.clear()
+                    reasoning_buffer.clear()
+                    renderer = self.renderer(markdown)
+                    await self.bus.publish_inbound(
+                        InboundMessage(
+                            channel=channel,
+                            sender_id="user",
+                            chat_id=chat_id,
+                            content=user_input,
+                            metadata={"_wants_stream": True},
+                        )
+                    )
+                    await turn_done.wait()
+                    if turn_response:
+                        response = turn_response[0]
+                        if response.content and not isinstance(
+                            response.event, StreamedResponseEvent
+                        ):
+                            if renderer:
+                                await renderer.close()
+                            cli_terminal._print_agent_response(
+                                response.content,
+                                render_markdown=markdown,
+                                metadata=response.metadata,
+                                **(
+                                    {"show_header": False}
+                                    if renderer and renderer.header_printed
+                                    else {}
+                                ),
+                            )
+                    elif renderer and not renderer.streamed:
+                        await renderer.close()
+                except (KeyboardInterrupt, EOFError):
+                    cli_terminal._restore_terminal()
+                    cli_terminal.console.print("\nGoodbye!")
+                    break
+        finally:
+            self.loop.stop()
+            outbound_task.cancel()
+            await asyncio.gather(bus_task, outbound_task, return_exceptions=True)
+            await self.close()

nanobot/cli/commands.py

@@ -49,6 +49,8 @@ from nanobot import __logo__, __version__  # noqa: E402
 from nanobot import optional_features as feature_support  # noqa: E402
 from nanobot.agent.hooks import create_file_edit_activity_hook  # noqa: E402
 from nanobot.agent.loop import AgentLoop  # noqa: E402
+from nanobot.agent.tools.mcp import MCPProvider  # noqa: E402
+from nanobot.agent.tools.registry import ToolRegistry  # noqa: E402
 from nanobot.cli import terminal as cli_terminal  # noqa: E402
 from nanobot.cli.agent import agent  # noqa: E402
 from nanobot.cli.gateway import create_gateway_app  # noqa: E402
@@ -351,12 +353,15 @@ def serve(
     sync_workspace_templates(runtime_config.workspace_path)
     bus = MessageBus()
     session_manager = SessionManager(runtime_config.workspace_path)
+    tools = ToolRegistry()
+    mcp_provider = MCPProvider.from_config(runtime_config, tools)
     try:
         agent_loop = AgentLoop.from_config(
             runtime_config, bus,
             session_manager=session_manager,
             image_generation_provider_configs=image_gen_provider_configs(runtime_config),
             hook_factories=[create_file_edit_activity_hook],
+            tool_registry=tools,
         )
     except ValueError as exc:
         console.print(f"[red]Error: {exc}[/red]")
@@ -378,13 +383,17 @@ def serve(
     api_app = create_app(
         agent_loop, model_name=model_name, request_timeout=timeout,
         api_key=api_key,
+        prepare_agent=mcp_provider.connect,
     )
 
     async def on_startup(_app: Any) -> None:
-        await agent_loop._connect_mcp()
+        await mcp_provider.connect()
 
     async def on_cleanup(_app: Any) -> None:
-        await agent_loop.close_mcp()
+        try:
+            await agent_loop.aclose()
+        finally:
+            await mcp_provider.aclose()
 
     api_app.on_startup.append(on_startup)
     api_app.on_cleanup.append(on_cleanup)
@@ -431,6 +440,44 @@ app.add_typer(
 app.command(name="agent")(agent)
 
 
+# ============================================================================
+# Session Commands
+# ============================================================================
+
+
+sessions_app = typer.Typer(help="Manage persisted session history")
+app.add_typer(sessions_app, name="sessions")
+
+
+@sessions_app.command("restore-workspace")
+def sessions_restore_workspace(
+    config: str | None = typer.Option(None, "--config", "-c", help="Path to config file"),
+    workspace: str | None = typer.Option(None, "--workspace", "-w", help="Workspace directory"),
+) -> None:
+    """Copy sessions back into the workspace before downgrading nanobot."""
+    from nanobot.session.manager import SessionManager
+
+    runtime_config = _load_runtime_config(config, workspace)
+    data_dir = runtime_config.runtime_data_dir
+    manager = SessionManager(
+        runtime_config.workspace_path,
+        sessions_root=data_dir / "sessions" if data_dir is not None else None,
+    )
+    result = manager.restore_sessions_to_workspace()
+    console.print(
+        f"Restored {result.restored} session file(s) to "
+        f"{escape(str(runtime_config.workspace_path / 'sessions'))}; "
+        f"{result.unchanged} already matched."
+    )
+    if result.conflicts:
+        console.print(
+            "[red]Rollback is incomplete: existing or invalid files require manual review.[/red]"
+        )
+        for path in result.conflicts:
+            console.print(Text(f"- {path}", style="red"))
+        raise typer.Exit(1)
+
+
 # ============================================================================
 # Channel Commands
 # ============================================================================

nanobot/cli/gateway.py

@@ -14,8 +14,9 @@ from rich.console import Console
 
 from nanobot.config.schema import Config
 from nanobot.gateway import (
+    GatewayAlreadyRunningError,
+    GatewayInstance,
     GatewayRuntime,
-    GatewayRuntimePaths,
     GatewayStartOptions,
     GatewayStatus,
 )
@@ -35,6 +36,15 @@ GatewayServiceFactory = Callable[[], Any]
 WebUIBundlePreparer = Callable[[Config, BuildMode], None]
 
 
+def _resolved_config_selector(config: str | None) -> Path:
+    """Return the one canonical config identity used by every local client."""
+    if config:
+        return Path(config).expanduser().resolve(strict=False)
+    from nanobot.config.loader import get_config_path
+
+    return get_config_path().resolve(strict=False)
+
+
 def create_gateway_app(
     *,
     console: Console,
@@ -69,19 +79,21 @@ def create_gateway_app(
             filter=lambda record: record["extra"].setdefault("channel", "-") or True,
         )
 
+    def instance_for_selectors(
+        *,
+        workspace: str | None = None,
+        config: str | None = None,
+    ) -> GatewayInstance:
+        return GatewayInstance.resolve(
+            config_path=_resolved_config_selector(config),
+            workspace=workspace,
+        )
+
     def runtime_for_instance(*, workspace: str | None = None, config: str | None = None):
         if runtime_factory is not None:
             return runtime_factory(workspace=workspace, config=config)
-        config_path = str(Path(config).expanduser().resolve(strict=False)) if config else None
-        workspace_path = str(Path(workspace).expanduser().resolve(strict=False)) if workspace else None
-        data_dir = Path(config_path).parent if config_path else None
-        return GatewayRuntime(
-            paths=GatewayRuntimePaths.for_instance(
-                data_dir=data_dir,
-                workspace=workspace_path,
-                config_path=config_path,
-            )
-        )
+        instance = instance_for_selectors(workspace=workspace, config=config)
+        return GatewayRuntime(paths=instance.paths)
 
     def service_installer():
         return service_factory() if service_factory is not None else GatewayServiceInstaller()
@@ -100,13 +112,12 @@ def create_gateway_app(
         loaded_config: Config | None = None,
     ) -> GatewayStartOptions:
         cfg = loaded_config or load_runtime_config(config, workspace)
-        resolved_config = str(Path(config).expanduser().resolve()) if config else None
-        resolved_workspace = str(Path(workspace).expanduser().resolve(strict=False)) if workspace else None
-        return GatewayStartOptions(
+        return instance_for_selectors(
+            workspace=workspace,
+            config=config,
+        ).start_options(
             port=port if port is not None else cfg.gateway.port,
             verbose=verbose,
-            workspace=resolved_workspace,
-            config_path=resolved_config,
         )
 
     def print_status(status: GatewayStatus) -> None:
@@ -118,6 +129,10 @@ def create_gateway_app(
             console.print(f"Port: {status.port}")
         if status.started_at is not None:
             console.print(f"Started At: {status.started_at}")
+        if status.running:
+            console.print(f"Launch Mode: {status.launch_mode}")
+            console.print(f"Lifetime: {status.lifetime}")
+            console.print(f"Clients: {status.clients}")
         console.print(f"State: {status.state_path}")
         console.print(f"Logs: {status.log_path}")
 
@@ -166,9 +181,55 @@ def create_gateway_app(
                     loaded_config=cfg,
                 )
             )
+            if (
+                result.message == "gateway_already_running"
+                and result.status.launch_mode == "foreground"
+            ):
+                console.print(
+                    "[yellow]Gateway is already running in the foreground; "
+                    "an attached process cannot be detached in place.[/yellow]"
+                )
+                console.print(
+                    "[dim]Stop it in its current terminal, then run "
+                    "`nanobot gateway --background`.[/dim]"
+                )
+                print_status(result.status)
+                raise typer.Exit(1)
+            if (
+                result.message == "gateway_already_running"
+                and result.status.launch_mode == "unknown"
+                and result.status.lifetime == "explicit"
+            ):
+                console.print(
+                    "[yellow]Gateway is already running, but this older process did "
+                    "not record whether it is attached or detached.[/yellow]"
+                )
+                console.print(
+                    "[dim]Stop it first, then rerun `nanobot gateway --background` "
+                    "to establish an unambiguous lifecycle.[/dim]"
+                )
+                print_status(result.status)
+                raise typer.Exit(1)
             if result.ok:
                 console.print("[green]Gateway started in the background.[/green]")
-                print_status(result.status)
+                print_status(runtime.status())
+                return
+            if result.message == "gateway_already_running":
+                if result.promoted:
+                    console.print(
+                        "[green]Existing on-demand gateway promoted to persistent "
+                        "background mode.[/green]"
+                    )
+                    console.print(
+                        "[dim]It will keep running after all local clients exit; "
+                        "use `nanobot gateway stop` to stop it.[/dim]"
+                    )
+                else:
+                    console.print(
+                        "[yellow]Gateway is already running in persistent "
+                        "background mode.[/yellow]"
+                    )
+                print_status(runtime.status())
                 return
             console.print(f"[yellow]Gateway was not started: {result.message}[/yellow]")
             print_status(result.status)
@@ -176,18 +237,22 @@ def create_gateway_app(
 
         configure_logging(verbose)
         cfg = load_runtime_config(config, workspace)
+        instance = instance_for_selectors(workspace=workspace, config=config)
         unconfigured_provider_error = None
         if validate_startup_config is not None:
             unconfigured_provider_error = validate_startup_config(cfg)
-        if unconfigured_provider_error is None:
-            run_gateway(cfg, port=port, webui_bundle_mode=interactive_build_mode())
-        else:
+        try:
             run_gateway(
                 cfg,
                 port=port,
                 webui_bundle_mode=interactive_build_mode(),
                 unconfigured_provider_error=unconfigured_provider_error,
+                gateway_instance=instance,
             )
+        except GatewayAlreadyRunningError as exc:
+            console.print("[yellow]Gateway is already running.[/yellow]")
+            print_status(exc.status)
+            raise typer.Exit(1) from None
 
     @gateway_app.command("status")
     def gateway_status(  # pyright: ignore[reportUnusedFunction]
@@ -222,7 +287,8 @@ def create_gateway_app(
         config: str | None = typer.Option(None, "--config", "-c", help="Path to config file"),
     ) -> None:
         """Stop the background gateway."""
-        result = runtime_for_instance(workspace=workspace, config=config).stop(timeout_s=timeout)
+        runtime = runtime_for_instance(workspace=workspace, config=config)
+        result = runtime.stop(timeout_s=timeout)
         if result.ok:
             console.print("[green]Gateway stopped.[/green]")
         else:
@@ -260,6 +326,24 @@ def create_gateway_app(
             console.print("[green]Gateway restarted in the background.[/green]")
             print_status(result.status)
             return
+        if result.message == "gateway_not_running":
+            console.print("[yellow]Gateway is not running; there is nothing to restart.[/yellow]")
+            console.print(
+                "[dim]Start a persistent gateway with `nanobot gateway --background`.[/dim]"
+            )
+            print_status(result.status)
+            raise typer.Exit(1)
+        if result.message == "gateway_foreground_restart_required":
+            console.print(
+                "[yellow]Gateway is attached to a foreground terminal and cannot "
+                "be restarted as a background process.[/yellow]"
+            )
+            console.print(
+                "[dim]Restart it in that terminal, or stop it and run "
+                "`nanobot gateway --background`.[/dim]"
+            )
+            print_status(result.status)
+            raise typer.Exit(1)
         console.print(f"[red]Gateway restart failed: {result.message}[/red]")
         print_status(result.status)
         raise typer.Exit(1)

nanobot/cli/gateway_runtime.py

@@ -14,6 +14,8 @@ from rich.console import Console
 from nanobot import __logo__, __version__
 from nanobot.agent.hooks import create_file_edit_activity_hook
 from nanobot.agent.loop import AgentLoop
+from nanobot.agent.tools.mcp import MCPProvider
+from nanobot.agent.tools.registry import ToolRegistry
 from nanobot.cli import terminal as cli_terminal
 from nanobot.cli.runtime_config import _migrate_cron_store
 from nanobot.cli.webui_support import (
@@ -30,6 +32,7 @@ from nanobot.cli.webui_support import (
 )
 from nanobot.config.paths import is_default_workspace
 from nanobot.config.schema import Config
+from nanobot.gateway.runtime import GatewayInstance
 from nanobot.security.network import is_loopback_host
 from nanobot.session.keys import UNIFIED_SESSION_KEY, last_channel_from_metadata
 from nanobot.utils.evaluator import evaluate_response, resolve_evaluator_prompt
@@ -233,6 +236,7 @@ def _print_gateway_health_endpoint(host: str, port: int) -> None:
 
 async def _close_gateway_runtime(
     agent: AgentLoop,
+    mcp_provider: MCPProvider,
     channels: Any,
     tasks: list[asyncio.Task[Any]],
     runtime_tasks: asyncio.Future[list[Any]] | None,
@@ -240,18 +244,13 @@ async def _close_gateway_runtime(
     task_wait_timeout: float = 15.0,
     close_timeout: float = 15.0,
 ) -> None:
-    """Cancel runtime tasks, then deterministically close agent resources.
+    """Cancel runtime tasks, then deterministically close application resources.
 
     Order matters: runtime tasks (including the agent loop and any in-flight
-    turn) are cancelled and awaited -- bounded -- before exec sessions,
-    subagents, and MCP servers are torn down, so no active turn is using a
-    shared resource when it closes. The final close is bounded and idempotent:
-    the agent loop's own finally also calls ``close_mcp()``, so this runs again
-    as a no-op when that path already completed, and as the guaranteed final
-    close when it was skipped or cut short (which previously left asyncio
-    subprocess transports alive past ``loop.close()``, producing
-    "RuntimeError: Event loop is closed" noise and potentially orphaned
-    processes at interpreter exit).
+    turn) are cancelled and awaited -- bounded -- before the loop-owned resources
+    and the application-owned MCP provider are torn down. The final close is
+    bounded and idempotent, so it also covers a cancelled or incomplete loop
+    cleanup without leaving subprocess transports alive past ``loop.close()``.
     """
     # Some SDKs swallow task cancellation while attempting to reconnect.
     # Close channel transports before waiting for their runners to exit.
@@ -272,10 +271,14 @@ async def _close_gateway_runtime(
             task.cancel()
     if runtime_tasks is not None and not runtime_tasks.done():
         runtime_tasks.cancel()
+    for label, close in (
+        ("agent", agent.aclose),
+        ("MCP provider", mcp_provider.aclose),
+    ):
         try:
-        await asyncio.wait_for(agent.close_mcp(), timeout=close_timeout)
+            await asyncio.wait_for(close(), timeout=close_timeout)
         except BaseException as exc:  # noqa: BLE001 - shutdown must proceed
-        logger.warning("Gateway shutdown: agent resource cleanup incomplete: {}", exc)
+            logger.warning("Gateway shutdown: {} cleanup incomplete: {}", label, exc)
     # Retrieving an already-finished gather prevents noisy unhandled exceptions,
     # but never wait for it here: its children were bounded individually above.
     if runtime_tasks is not None and runtime_tasks.done():
@@ -296,6 +299,7 @@ def _run_gateway(
     health_server_enabled: bool = True,
     unconfigured_provider_error: str | None = None,
     webui_dev_server: WebUIDevServer | None = None,
+    gateway_instance: GatewayInstance | None = None,
 ) -> None:
     """Shared gateway runtime; ``open_browser_url`` opens a tab once channels are up."""
     from nanobot.agent.model_presets import load_model_preset_catalog
@@ -385,19 +389,20 @@ def _run_gateway(
             raise typer.Exit(1) from exc
     session_manager = SessionManager(config.workspace_path)
 
-    # Self-heal the gateway state file with the current PID after any restart.
+    # Use the same runtime identity for foreground and managed gateway processes.
     from nanobot.config.loader import get_config_path
-    from nanobot.gateway.runtime import GatewayRuntime, GatewayRuntimePaths
+    from nanobot.gateway.runtime import (
+        GatewayClientLease,
+        GatewayRuntime,
+        monitor_gateway_clients,
+    )
 
-    config_path = str(get_config_path().resolve(strict=False))
-    GatewayRuntime.refresh_state_pid(
-        paths=GatewayRuntimePaths.for_instance(
-            workspace=str(config.workspace_path)
-            if not is_default_workspace(config.workspace_path)
-            else None,
-            config_path=config_path,
-        )
+    instance = gateway_instance or GatewayInstance.resolve(
+        config_path=get_config_path(),
     )
+    config_path = str(instance.config_path)
+    gateway_runtime = GatewayRuntime(paths=instance.paths)
+    gateway_start_options = instance.start_options(port=port)
 
     # Preserve existing single-workspace installs, but keep custom workspaces clean.
     if is_default_workspace(config.workspace_path):
@@ -414,6 +419,9 @@ def _run_gateway(
         route_policy=WebuiTurnRoutePolicy(session_manager),
     )
 
+    tools = ToolRegistry()
+    mcp_provider = MCPProvider.from_config(config, tools)
+
     # Create agent with cron service
     agent = AgentLoop.from_config(
         config, bus,
@@ -431,6 +439,7 @@ def _run_gateway(
         hooks=[TokenUsageHook(timezone_name=config.agents.defaults.timezone)],
         local_trigger_store=trigger_store,
         hook_factories=[create_file_edit_activity_hook],
+        tool_registry=tools,
     )
     def _schedule_webui_background(awaitable: Awaitable[None]) -> None:
         agent.schedule_background(cast(Coroutine[Any, Any, None], awaitable))
@@ -512,6 +521,7 @@ def _run_gateway(
                 prompt, last_cursor = result
                 key = dream_session_key()
                 dream_runtime = agent.dream_runtime()
+                await mcp_provider.connect()
                 resp = await agent.process_direct(
                     prompt,
                     session_key=key,
@@ -559,7 +569,7 @@ def _run_gateway(
                 if sha:
                     logger.info("Dream commit: {}", sha)
                 store.compact_history()
-                prune_dream_sessions(agent.sessions.sessions_dir)
+                prune_dream_sessions(agent.sessions)
             return None
 
         # Heartbeat is a system job that checks HEARTBEAT.md for active tasks.
@@ -589,6 +599,7 @@ def _run_gateway(
             if isinstance(message_tool, MessageTool):
                 suppress_token = message_tool.set_suppress_delivery(True)
             try:
+                await mcp_provider.connect()
                 resp = await agent.process_direct(
                     prompt,
                     session_key="heartbeat",
@@ -649,6 +660,9 @@ def _run_gateway(
     def _webui_runtime_model_name() -> str | None:
         return agent.model.strip() or None
 
+    def _webui_refresh_runtime_config() -> None:
+        agent.refresh_runtime_config()
+
     def _webui_skill_state_action(disabled_skills: set[str]) -> None:
         config.agents.defaults.disabled_skills = sorted(disabled_skills)
         agent.context.skills.disabled_skills = set(disabled_skills)
@@ -663,12 +677,16 @@ def _run_gateway(
         cron_service=cron,
         local_trigger_store=trigger_store,
         webui_runtime_model_name=_webui_runtime_model_name,
+        webui_refresh_runtime_config=_webui_refresh_runtime_config,
         webui_cron_pending_job_ids=agent.pending_cron_job_ids_for_session,
         webui_local_trigger_pending_ids=agent.pending_local_trigger_ids_for_session,
         webui_static_dist=webui_static_dist,
         webui_runtime_surface=webui_runtime_surface,
         webui_runtime_capabilities=webui_runtime_capabilities,
+        webui_mcp_runtime_status=mcp_provider.runtime_status,
+        webui_mcp_reload=mcp_provider.reload,
         webui_skill_state_action=_webui_skill_state_action,
+        config_path=Path(config_path),
     )
 
     def _pick_heartbeat_target() -> tuple[str, str]:
@@ -842,6 +860,21 @@ def _run_gateway(
             await cron.start()
             # Re-read once on first admission to close the watcher subscription window.
             agent.runtime_resolver.invalidate()
+            async def _run_agent() -> None:
+                try:
+                    await mcp_provider.connect()
+                    await agent.run()
+                finally:
+                    await mcp_provider.aclose()
+
+            async def _monitor_local_clients() -> None:
+                orphaned = await monitor_gateway_clients(
+                    GatewayClientLease(gateway_runtime, kind="gateway-monitor"),
+                    shutdown_event,
+                )
+                if orphaned:
+                    logger.info("Last local client disappeared; stopping on-demand gateway")
+
             tasks = [
                 asyncio.create_task(
                     watch_config_file(
@@ -850,7 +883,7 @@ def _run_gateway(
                     ),
                     name="nanobot-config-watcher",
                 ),
-                asyncio.create_task(agent.run(), name="nanobot-agent-loop"),
+                asyncio.create_task(_run_agent(), name="nanobot-agent-loop"),
                 asyncio.create_task(channels.start_all(), name="nanobot-channels"),
                 asyncio.create_task(
                     run_local_trigger_queue(
@@ -860,6 +893,10 @@ def _run_gateway(
                     ),
                     name="nanobot-local-triggers",
                 ),
+                asyncio.create_task(
+                    _monitor_local_clients(),
+                    name="nanobot-gateway-client-monitor",
+                ),
             ]
             if health_server_enabled:
                 tasks.append(asyncio.create_task(
@@ -908,7 +945,13 @@ def _run_gateway(
                 agent.stop()
                 # Cancel runtime tasks first, then deterministically close
                 # exec/MCP resources while the event loop is still alive.
-                await _close_gateway_runtime(agent, channels, tasks, runtime_tasks)
+                await _close_gateway_runtime(
+                    agent,
+                    mcp_provider,
+                    channels,
+                    tasks,
+                    runtime_tasks,
+                )
                 # Flush all cached sessions to durable storage before exit.
                 # This prevents data loss on filesystems with write-back
                 # caching (rclone VFS, NFS, FUSE mounts, etc.).
@@ -918,4 +961,5 @@ def _run_gateway(
             finally:
                 restore_shutdown_handlers()
 
+    with gateway_runtime.foreground_instance(gateway_start_options):
         asyncio.run(run())

nanobot/cli/onboard.py

@@ -1594,7 +1594,6 @@ def _pause(message: str = "Press Enter to continue...") -> None:
 def _set_primary_quick_start_preset(config: Config, provider_name: str, model: str) -> None:
     """Store the primary preset used by Quick Start."""
     config.model_presets["primary"] = ModelPresetConfig(
-        label="Primary",
         model=model,
         provider=provider_name,
     )

nanobot/cli/tui_launcher.py

@@ -0,0 +1,497 @@
+"""Launch the TypeScript terminal client against the local gateway."""
+
+from __future__ import annotations
+
+import hashlib
+import io
+import json
+import os
+import platform
+import shutil
+import subprocess
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+import zipfile
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, cast
+
+from nanobot import __version__
+from nanobot.cli.runtime_config import _model_display
+from nanobot.cli.webui_support import (
+    _gateway_health_ready,
+    _webui_browser_url,
+    _webui_endpoint_reachable,
+    webui_bootstrap_secret,
+)
+from nanobot.config.paths import get_data_dir
+from nanobot.config.schema import Config
+
+if TYPE_CHECKING:
+    from nanobot.gateway import GatewayClientLease
+
+
+class TuiUnavailableError(RuntimeError):
+    """Raised when the native TypeScript TUI cannot run on this installation."""
+
+
+class TuiSessionError(ValueError):
+    """Raised when a session selector cannot be opened by the native TUI."""
+
+
+_TUI_RELEASE_FILES = (
+    "THIRD_PARTY_NOTICES.txt",
+    "RELINKING.md",
+    "SOURCE_OFFER.md",
+    "LICENSE",
+    "BUN-1.3.13-LICENSE.md",
+    "LGPL-2.0.txt",
+    "LGPL-2.1.txt",
+    "nanobot-tui-source.tar.gz",
+)
+_TUI_RELEASE_LIMITS = {
+    "THIRD_PARTY_NOTICES.txt": 4 * 1024 * 1024,
+    "RELINKING.md": 256 * 1024,
+    "SOURCE_OFFER.md": 256 * 1024,
+    "LICENSE": 256 * 1024,
+    "BUN-1.3.13-LICENSE.md": 1024 * 1024,
+    "LGPL-2.0.txt": 256 * 1024,
+    "LGPL-2.1.txt": 256 * 1024,
+    "nanobot-tui-source.tar.gz": 20 * 1024 * 1024,
+    "MANIFEST.sha256": 64 * 1024,
+}
+
+
+@dataclass(frozen=True)
+class _GatewayHandle:
+    base_url: str
+    lease: GatewayClientLease | None = None
+
+
+def launch_tui(
+    config: Config,
+    *,
+    config_path: Path,
+    workspace_override: str | None,
+    session_id: str | None,
+    theme: str,
+) -> int:
+    """Run the native TUI against the shared local gateway."""
+    state_path = config_path.parent / "tui" / "state.json"
+    chat_id = _initial_tui_chat_id(session_id, state_path)
+    command = _resolve_tui_command()
+    gateway = _ensure_gateway(
+        config,
+        config_path=config_path,
+        workspace_override=workspace_override,
+    )
+    try:
+        bootstrap = _fetch_bootstrap(
+            gateway.base_url,
+            secret=webui_bootstrap_secret(config),
+        )
+        env = os.environ.copy()
+        env.update(
+            {
+                "NANOBOT_TUI_WS_URL": _authenticated_ws_url(bootstrap),
+                "NANOBOT_TUI_API_URL": gateway.base_url,
+                "NANOBOT_TUI_API_TOKEN": str(bootstrap.get("api_token") or ""),
+                "NANOBOT_TUI_MODEL": _model_display(config)[0],
+                "NANOBOT_TUI_MODEL_PRESET": config.agents.defaults.model_preset or "default",
+                "NANOBOT_TUI_WORKSPACE": str(config.workspace_path),
+                "NANOBOT_TUI_VERSION": __version__,
+                "NANOBOT_TUI_ACCESS": (
+                    "workspace access" if config.tools.restrict_to_workspace else "full access"
+                ),
+                "NANOBOT_TUI_THEME": theme,
+            }
+        )
+        env["NANOBOT_TUI_STATE_PATH"] = str(state_path)
+        if chat_id:
+            env["NANOBOT_TUI_CHAT_ID"] = chat_id
+        else:
+            env.pop("NANOBOT_TUI_CHAT_ID", None)
+        return subprocess.run(command, env=env, check=False).returncode
+    except OSError as exc:
+        raise TuiUnavailableError(f"could not start the native TUI: {exc}") from exc
+    finally:
+        lease = getattr(gateway, "lease", None)
+        if lease is not None:
+            lease.release()
+
+
+def _resolve_tui_command() -> list[str]:
+    override = os.environ.get("NANOBOT_TUI_BIN", "").strip()
+    if override:
+        executable = Path(override).expanduser().resolve(strict=False)
+        if not executable.is_file():
+            raise TuiUnavailableError(f"NANOBOT_TUI_BIN does not exist: {executable}")
+        return [str(executable)]
+
+    suffix = ".exe" if os.name == "nt" else ""
+    system = {"Windows": "win32", "Darwin": "darwin", "Linux": "linux"}.get(
+        platform.system(),
+        platform.system().lower(),
+    )
+    machine = {"x86_64": "x64", "AMD64": "x64", "aarch64": "arm64"}.get(
+        platform.machine(),
+        platform.machine().lower(),
+    )
+    if system == "win32" and machine == "arm64":
+        raise TuiUnavailableError(
+            "the native TUI is not available on Windows ARM64 because Bun FFI is disabled "
+            "on that platform; use the classic prompt until the upstream runtime supports it"
+        )
+    asset = f"nanobot-tui-{system}-{machine}{suffix}"
+    source_dir = _source_checkout_tui_dir()
+    if source_dir is not None:
+        bun = shutil.which("bun")
+        if not bun:
+            raise TuiUnavailableError(
+                "this source checkout requires Bun to run its matching TUI; "
+                "install Bun, then run `nanobot agent` again"
+            )
+        return _resolve_source_tui_command(source_dir, bun)
+
+    packaged = Path(__file__).resolve().parents[1] / "tui" / "bin" / asset
+    if packaged.is_file():
+        return [str(packaged)]
+
+    downloaded = _download_release_tui(asset)
+    if downloaded is not None:
+        return [str(downloaded)]
+
+    raise TuiUnavailableError(
+        f"no native TUI archive is published for nanobot {__version__} on this platform; "
+        "current source installs must be editable and keep their checkout and Bun available, "
+        "while released packages need a matching GitHub release archive; use "
+        "`nanobot agent --classic` if intentional"
+    )
+
+
+def _source_checkout_tui_dir() -> Path | None:
+    """Return this checkout's TUI source, never a neighboring unrelated directory."""
+    return _tui_source_dir(Path(__file__).resolve().parents[2])
+
+
+def _tui_source_dir(project_root: Path) -> Path | None:
+    project_root = project_root.resolve(strict=False)
+    source_dir = project_root / "tui"
+    if (project_root / "pyproject.toml").is_file() and (source_dir / "package.json").is_file():
+        return source_dir
+    return None
+
+
+def _resolve_source_tui_command(source_dir: Path, bun: str) -> list[str]:
+    dependency = source_dir / "node_modules" / "@opentui" / "core"
+    try:
+        install = subprocess.run(
+            [bun, "install", "--frozen-lockfile"],
+            cwd=source_dir,
+            capture_output=True,
+            text=True,
+            check=False,
+        )
+    except OSError as exc:
+        raise TuiUnavailableError(f"could not install TUI dependencies: {exc}") from exc
+    if install.returncode != 0 or not dependency.is_dir():
+        detail = (install.stderr or install.stdout).strip().splitlines()
+        suffix = f": {detail[-1]}" if detail else ""
+        raise TuiUnavailableError(f"could not install TUI dependencies{suffix}")
+    return [bun, str(source_dir / "src" / "index.ts")]
+
+
+def _download_release_tui(asset: str) -> Path | None:
+    """Install the complete, version-matched TUI release bundle."""
+    if os.environ.get("NANOBOT_TUI_NO_DOWNLOAD") == "1":
+        return None
+    version = __version__.strip()
+    if not version or version.endswith((".dev0", "+dev")):
+        return None
+
+    target_dir = get_data_dir() / "bin" / "tui" / version
+    cached = _cached_release_tui(target_dir, asset)
+    if cached is not None:
+        return cached
+
+    base = f"https://github.com/HKUDS/nanobot/releases/download/v{version}"
+    archive_name = f"{asset}.zip"
+    try:
+        checksum = _read_release_asset(f"{base}/{archive_name}.sha256", max_bytes=1024)
+        expected = _release_checksum(checksum, archive_name)
+        if expected is None:
+            return None
+        archive = _read_release_asset(f"{base}/{archive_name}", max_bytes=200 * 1024 * 1024)
+    except (OSError, TimeoutError, urllib.error.URLError, urllib.error.HTTPError):
+        return None
+    if hashlib.sha256(archive).hexdigest() != expected:
+        raise TuiUnavailableError("downloaded TUI archive failed checksum verification")
+    files = _verified_release_archive(archive, asset)
+
+    temporary: dict[str, Path] = {}
+    try:
+        target_dir.mkdir(parents=True, exist_ok=True)
+        for name, content in files.items():
+            path = target_dir / name
+            pending = path.with_name(f"{path.name}.tmp-{os.getpid()}")
+            pending.write_bytes(content)
+            if name == asset and os.name != "nt":
+                pending.chmod(0o755)
+            temporary[name] = pending
+        for name in _release_bundle_names(asset):
+            temporary[name].replace(target_dir / name)
+    except OSError:
+        for path in temporary.values():
+            path.unlink(missing_ok=True)
+        _clear_cached_release(target_dir, asset)
+        return None
+    return target_dir / asset
+
+
+def _release_bundle_names(asset: str) -> tuple[str, ...]:
+    return (asset, *_TUI_RELEASE_FILES, "MANIFEST.sha256")
+
+
+def _release_checksum(raw: bytes, archive_name: str) -> str | None:
+    try:
+        parts = raw.decode("utf-8").split()
+    except UnicodeDecodeError:
+        return None
+    if len(parts) != 2 or parts[1] != archive_name:
+        return None
+    digest = parts[0].lower()
+    if len(digest) != 64 or any(character not in "0123456789abcdef" for character in digest):
+        return None
+    return digest
+
+
+def _release_manifest(raw: bytes, asset: str) -> dict[str, str]:
+    expected_names = set(_release_bundle_names(asset)[:-1])
+    try:
+        lines = raw.decode("utf-8").splitlines()
+    except UnicodeDecodeError as exc:
+        raise TuiUnavailableError("TUI release manifest is not valid UTF-8") from exc
+    checksums: dict[str, str] = {}
+    for line in lines:
+        digest, separator, name = line.partition("  ")
+        digest = digest.lower()
+        if (
+            separator != "  "
+            or name not in expected_names
+            or name in checksums
+            or len(digest) != 64
+            or any(character not in "0123456789abcdef" for character in digest)
+        ):
+            raise TuiUnavailableError("TUI release manifest is malformed")
+        checksums[name] = digest
+    if set(checksums) != expected_names:
+        raise TuiUnavailableError("TUI release manifest is incomplete")
+    return checksums
+
+
+def _verified_release_archive(raw: bytes, asset: str) -> dict[str, bytes]:
+    expected_names = set(_release_bundle_names(asset))
+    files: dict[str, bytes] = {}
+    try:
+        with zipfile.ZipFile(io.BytesIO(raw)) as archive:
+            entries = archive.infolist()
+            names = [entry.filename for entry in entries if not entry.is_dir()]
+            if len(names) != len(entries) or len(names) != len(set(names)):
+                raise TuiUnavailableError("TUI release archive contains invalid entries")
+            if set(names) != expected_names:
+                raise TuiUnavailableError("TUI release archive is incomplete")
+            for entry in entries:
+                limit = 150 * 1024 * 1024 if entry.filename == asset else _TUI_RELEASE_LIMITS[
+                    entry.filename
+                ]
+                if entry.file_size == 0 or entry.file_size > limit:
+                    raise TuiUnavailableError(
+                        f"TUI release file has an invalid size: {entry.filename}"
+                    )
+                files[entry.filename] = archive.read(entry)
+    except zipfile.BadZipFile as exc:
+        raise TuiUnavailableError("downloaded TUI archive is not a valid ZIP file") from exc
+
+    checksums = _release_manifest(files["MANIFEST.sha256"], asset)
+    for name, expected in checksums.items():
+        if hashlib.sha256(files[name]).hexdigest() != expected:
+            raise TuiUnavailableError(f"TUI release file failed verification: {name}")
+    return files
+
+
+def _cached_release_tui(target_dir: Path, asset: str) -> Path | None:
+    target = target_dir / asset
+    manifest = target_dir / "MANIFEST.sha256"
+    if not target.is_file() and not manifest.exists():
+        return None
+    try:
+        checksums = _release_manifest(manifest.read_bytes(), asset)
+        for name, expected in checksums.items():
+            if hashlib.sha256((target_dir / name).read_bytes()).hexdigest() != expected:
+                raise OSError("cached release checksum mismatch")
+        if os.name != "nt":
+            target.chmod(0o755)
+    except (OSError, TuiUnavailableError):
+        _clear_cached_release(target_dir, asset)
+        return None
+    return target
+
+
+def _clear_cached_release(target_dir: Path, asset: str) -> None:
+    for name in _release_bundle_names(asset):
+        try:
+            (target_dir / name).unlink(missing_ok=True)
+        except OSError:
+            pass
+
+
+def _read_release_asset(url: str, *, max_bytes: int) -> bytes:
+    request = urllib.request.Request(url, headers={"User-Agent": f"nanobot/{__version__}"})
+    with urllib.request.urlopen(request, timeout=5) as response:
+        content_length = response.headers.get("Content-Length")
+        if content_length and int(content_length) > max_bytes:
+            raise OSError("release asset exceeds size limit")
+        body = response.read(max_bytes + 1)
+    if len(body) > max_bytes:
+        raise OSError("release asset exceeds size limit")
+    return body
+
+
+def _ensure_gateway(
+    config: Config,
+    *,
+    config_path: Path,
+    workspace_override: str | None,
+) -> _GatewayHandle:
+    from nanobot.gateway import (
+        GatewayClientLease,
+        GatewayInstance,
+        GatewayRuntime,
+    )
+
+    base_url = _webui_browser_url(config).split("/#/", 1)[0].rstrip("/")
+    instance = GatewayInstance.resolve(
+        config_path=config_path,
+        workspace=workspace_override,
+    )
+    runtime = GatewayRuntime(paths=instance.paths)
+    lease = GatewayClientLease(runtime, kind="tui")
+    lease.acquire()
+    try:
+        status = runtime.status()
+        endpoint_reachable = _webui_endpoint_reachable(base_url)
+        if status.running:
+            if status.port not in {None, config.gateway.port}:
+                raise TuiUnavailableError(
+                    "the matching gateway instance is running on a different port; "
+                    "restart it or use `nanobot agent --classic`"
+                )
+            if endpoint_reachable:
+                return _GatewayHandle(base_url=base_url, lease=lease)
+        elif endpoint_reachable:
+            raise TuiUnavailableError(
+                "the configured gateway port belongs to a different nanobot instance; "
+                "stop that instance or use `nanobot agent --classic`"
+            )
+
+        result = lease.ensure_on_demand_gateway(
+            instance.start_options(port=config.gateway.port)
+        )
+        if not result.ok and result.message != "gateway_already_running":
+            raise TuiUnavailableError(
+                f"could not start the local gateway ({result.message}); "
+                f"logs: {result.status.log_path}"
+            )
+
+        deadline = time.monotonic() + 20
+        while time.monotonic() < deadline:
+            if _webui_endpoint_reachable(base_url):
+                current = runtime.status()
+                if current.running and current.port in {None, config.gateway.port}:
+                    return _GatewayHandle(base_url=base_url, lease=lease)
+                break
+            if not runtime.status().running and not _gateway_health_ready(
+                config.gateway.host,
+                config.gateway.port,
+            ):
+                break
+            time.sleep(0.1)
+
+        raise TuiUnavailableError(
+            f"local gateway did not become ready; logs: {result.status.log_path}"
+        )
+    except BaseException:
+        lease.release(timeout_s=5)
+        raise
+
+
+def _fetch_bootstrap(base_url: str, *, secret: str) -> dict[str, Any]:
+    headers = {"X-Nanobot-Auth": secret} if secret else {}
+    request = urllib.request.Request(f"{base_url}/webui/bootstrap", headers=headers)
+    try:
+        with urllib.request.urlopen(request, timeout=5) as response:
+            raw_payload: Any = json.loads(response.read().decode("utf-8"))
+    except (OSError, TimeoutError, urllib.error.URLError, json.JSONDecodeError) as exc:
+        raise TuiUnavailableError(
+            f"could not authenticate with the local gateway: {exc}"
+        ) from exc
+    if not isinstance(raw_payload, dict):
+        raise TuiUnavailableError("gateway bootstrap response is missing ws_path")
+    payload = cast(dict[str, Any], raw_payload)
+    if not payload.get("ws_path"):
+        raise TuiUnavailableError("gateway bootstrap response is missing ws_path")
+    return payload
+
+
+def _authenticated_ws_url(bootstrap: dict[str, Any]) -> str:
+    raw_url = str(bootstrap.get("ws_url") or "").strip()
+    if not raw_url:
+        raise TuiUnavailableError("gateway bootstrap response is missing ws_url")
+    parsed = urllib.parse.urlsplit(raw_url)
+    query = urllib.parse.parse_qsl(parsed.query, keep_blank_values=True)
+    token = str(bootstrap.get("token") or "").strip()
+    if token:
+        query.append(("token", token))
+    query.append(("client_id", f"tui-{os.getpid()}"))
+    return urllib.parse.urlunsplit(
+        (parsed.scheme, parsed.netloc, parsed.path, urllib.parse.urlencode(query), parsed.fragment)
+    )
+
+
+def _websocket_chat_id(session_id: str) -> str | None:
+    """Map the CLI selector to the WebSocket namespace used by the native TUI."""
+    if session_id.startswith("websocket:"):
+        return session_id.split(":", 1)[1] or None
+    if ":" in session_id:
+        raise TuiSessionError(
+            "the native TUI can open only WebSocket sessions; use --classic to resume "
+            f"{session_id!r}"
+        )
+    return session_id or None
+
+
+def _initial_tui_chat_id(session_id: str | None, state_path: Path) -> str | None:
+    """Resume the last TUI chat, while keeping an explicit selector authoritative."""
+    if session_id is not None:
+        return _websocket_chat_id(session_id)
+    return _read_tui_chat_id(state_path)
+
+
+def _read_tui_chat_id(path: Path) -> str | None:
+    """Read the last attached chat without making launch depend on optional state."""
+    try:
+        raw_payload: Any = json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return None
+    if not isinstance(raw_payload, dict):
+        return None
+    payload = cast(dict[str, Any], raw_payload)
+    value = payload.get("chat_id")
+    if not isinstance(value, str):
+        return None
+    value = value.strip()
+    if not value or len(value) > 256 or any(character in value for character in "\r\n"):
+        return None
+    return value

nanobot/cli/webui.py

@@ -7,7 +7,6 @@ from pydantic import ValidationError
 from rich.console import Console
 
 from nanobot.cli import terminal as cli_terminal
-from nanobot.cli.gateway_runtime import _run_gateway
 from nanobot.cli.runtime_config import (
     _load_runtime_config,
     _print_config_error,
@@ -27,7 +26,6 @@ from nanobot.cli.webui_support import (
     _open_webui_browser,
     _prepare_webui_bundle_for_gateway,
     _print_foreground_port_conflict,
-    _print_webui_foreground_lifecycle,
     _resolve_webui_config_path,
     _run_quick_start_for_webui,
     _tcp_endpoint_reachable,
@@ -84,7 +82,7 @@ def webui(
     background: bool = typer.Option(
         False,
         "--background",
-        help="Keep the gateway running after this command exits",
+        help="Deprecated; use `nanobot gateway --background`",
     ),
     dev: bool = typer.Option(
         False,
@@ -101,13 +99,31 @@ def webui(
 ) -> None:
     """Prepare the local WebUI, start the gateway, and open the browser workbench."""
     from nanobot.config.loader import resolve_config_env_vars, save_config
-    from nanobot.gateway import GatewayRuntime, GatewayRuntimePaths, GatewayStartOptions
+    from nanobot.gateway import (
+        GatewayClientLease,
+        GatewayInstance,
+        GatewayRuntime,
+    )
 
     cli_terminal._ensure_interactive_tty_mode()
-    if dev and background:
-        console.print("[red]Error: --dev cannot be combined with --background.[/red]")
-        raise typer.Exit(1)
     config_path = _resolve_webui_config_path(config)
+    if background:
+        import shlex
+
+        command = ["nanobot", "gateway", "--background", "--config", str(config_path)]
+        if workspace:
+            command.extend(
+                ["--workspace", str(Path(workspace).expanduser().resolve(strict=False))]
+            )
+        console.print(
+            "[red]`nanobot webui --background` no longer owns gateway lifecycle.[/red]"
+        )
+        console.print("Start the persistent gateway explicitly, then open the WebUI:")
+        console.print("  [cyan]" + " ".join(shlex.quote(part) for part in command) + "[/cyan]")
+        console.print(
+            "  [cyan]nanobot webui --config " + shlex.quote(str(config_path)) + "[/cyan]"
+        )
+        raise typer.Exit(1)
     created_config = not config_path.exists()
     if created_config:
         console.print(f"[yellow]No config found at {config_path}.[/yellow]")
@@ -131,12 +147,6 @@ def webui(
     if settings_setup_error:
         console.print(f"[yellow]Model setup is incomplete: {provider_error}[/yellow]")
         console.print("Configure a provider and model in WebUI Settings → Models.")
-        if background:
-            console.print(
-                "[red]First-time WebUI setup must run in the foreground. "
-                "Run `nanobot webui` without --background.[/red]"
-            )
-            raise typer.Exit(1)
     elif provider_error:
         console.print(f"[dim]Provider check: {provider_error}[/dim]")
         setup_config = _run_quick_start_for_webui(
@@ -207,25 +217,21 @@ def webui(
             )
 
     webui_bundle_mode = _webui_build_mode_for_interactive(yes=yes)
-
-    config_arg = str(config_path)
-    workspace_arg = str(Path(workspace).expanduser().resolve(strict=False)) if workspace else None
-    runtime = GatewayRuntime(
-        paths=GatewayRuntimePaths.for_instance(
-            data_dir=config_path.parent,
-            workspace=workspace_arg,
-            config_path=config_arg,
-        )
-    )
-    start_options = GatewayStartOptions(
-        port=effective_gateway_port,
-        workspace=workspace_arg,
-        config_path=config_arg,
+    _prepare_webui_bundle_for_gateway(
+        runtime_config,
+        mode="skip" if dev else webui_bundle_mode,
     )
 
-    if background:
-        _prepare_webui_bundle_for_gateway(runtime_config, mode=webui_bundle_mode)
-        result = runtime.start_background(start_options)
+    instance = GatewayInstance.resolve(
+        config_path=config_path,
+        workspace=workspace,
+    )
+    runtime = GatewayRuntime(paths=instance.paths)
+    start_options = instance.start_options(port=effective_gateway_port)
+
+    def ensure_shared_gateway(*, client_lease: GatewayClientLease) -> None:
+        """Start or refresh the one managed gateway shared by local clients."""
+        result = client_lease.ensure_on_demand_gateway(start_options)
         restarted = False
         restart_attempted = False
         if not result.ok and result.message == "gateway_already_running" and changed_webui:
@@ -244,6 +250,8 @@ def webui(
             console.print("[green]Gateway started in the background.[/green]")
         else:
             console.print("[yellow]Gateway is already running in the background.[/yellow]")
+
+    def print_shared_gateway_controls() -> None:
         console.print(
             "Manage this instance: "
             f"[cyan]{_gateway_instance_command('status', config_path=config_path, workspace=workspace)}[/cyan]"
@@ -257,14 +265,26 @@ def webui(
             "Stop nanobot: "
             f"[cyan]{_gateway_instance_command('stop', config_path=config_path, workspace=workspace)}[/cyan]"
         )
-        if not no_open:
-            _open_webui_browser(webui_url)
-        return
 
     gateway_ready = _gateway_health_ready(runtime_config.gateway.host, effective_gateway_port)
     webui_ready = _webui_endpoint_reachable(webui_url)
     if gateway_ready and webui_ready:
-        console.print("[yellow]Gateway is already running; attaching to the existing WebUI.[/yellow]")
+        lease = GatewayClientLease(runtime, kind="webui")
+        lease.acquire()
+        try:
+            if changed_webui and runtime.status().running:
+                ensure_shared_gateway(client_lease=lease)
+                gateway_ready = _gateway_health_ready(
+                    runtime_config.gateway.host,
+                    effective_gateway_port,
+                )
+                webui_ready = _webui_endpoint_reachable(webui_url)
+                if not gateway_ready or not webui_ready:
+                    console.print("[red]Gateway did not become ready after the config update.[/red]")
+                    raise typer.Exit(1)
+            console.print(
+                "[yellow]Gateway is already running; attaching to the existing WebUI.[/yellow]"
+            )
             if not dev:
                 console.print(
                     "Restart the gateway if you need it to pick up local source changes: "
@@ -305,6 +325,11 @@ def webui(
                 console.print(f"[red]Error: {exc}[/red]")
                 raise typer.Exit(1) from exc
             return
+        finally:
+            if lease.release():
+                console.print(
+                    "[dim]Last local client exited; the on-demand gateway was stopped.[/dim]"
+                )
 
     gateway_port_taken = gateway_ready or _tcp_endpoint_reachable(
         _host_for_local_browser(runtime_config.gateway.host),
@@ -319,7 +344,11 @@ def webui(
         )
         raise typer.Exit(1)
 
-    _print_webui_foreground_lifecycle(attached=False)
+    lease = GatewayClientLease(runtime, kind="webui")
+    lease.acquire()
+    try:
+        ensure_shared_gateway(client_lease=lease)
+        print_shared_gateway_controls()
         if dev_browser_url:
             dev_proxy_target = webui_dev_proxy_target(webui_url)
             try:
@@ -328,25 +357,20 @@ def webui(
                     browser_url=dev_browser_url,
                     output=lambda message: console.print(f"[green]✓[/green] {message}"),
                 ) as dev_server:
-                _run_gateway(
-                    runtime_config,
-                    port=effective_gateway_port,
-                    open_browser_url=None if no_open else dev_browser_url,
-                    open_browser_ready_url=f"{dev_proxy_target}/webui/bootstrap",
-                    webui_static_dist=False,
-                    webui_bundle_mode="skip",
-                    unconfigured_provider_error=settings_setup_error,
-                    webui_dev_server=dev_server,
+                    if not no_open:
+                        _open_webui_browser(dev_browser_url)
+                    _attach_to_background_gateway(
+                        runtime,
+                        poll_hook=dev_server.ensure_running,
                     )
             except WebUIDevError as exc:
                 console.print(f"[red]Error: {exc}[/red]")
                 raise typer.Exit(1) from exc
             return
 
-    _run_gateway(
-        runtime_config,
-        port=effective_gateway_port,
-        open_browser_url=None if no_open else webui_url,
-        webui_bundle_mode=webui_bundle_mode,
-        unconfigured_provider_error=settings_setup_error,
-    )
+        if not no_open:
+            _open_webui_browser(webui_url)
+        _attach_to_background_gateway(runtime)
+    finally:
+        if lease.release():
+            console.print("[dim]Last local client exited; the on-demand gateway was stopped.[/dim]")

nanobot/cli/webui_support.py

@@ -24,6 +24,7 @@ from nanobot.webui.build import (
     BuildMode,
     WebUIBuildError,
     ensure_webui_bundle,
+    inspect_webui_bundle,
 )
 
 if TYPE_CHECKING:
@@ -49,6 +50,7 @@ __all__ = [
     "_validate_gateway_startup",
     "_warn_webui_bind_scope",
     "_webui_browser_url",
+    "webui_bootstrap_secret",
     "_webui_build_mode_for_interactive",
     "_webui_channel_enabled",
     "_webui_display_url",
@@ -190,6 +192,11 @@ def _prepare_webui_bundle_for_gateway(
         return typer.confirm(message, default=True)
 
     try:
+        # A source checkout is the development product. Every gateway entrypoint
+        # keeps its browser client in lockstep with Python; only Vite mode skips
+        # the production bundle intentionally.
+        if mode != "skip" and inspect_webui_bundle().source_available:
+            mode = "auto"
         ensure_webui_bundle(
             mode=mode,
             confirm=_confirm if mode == "prompt" else None,
@@ -224,7 +231,8 @@ def _gateway_health_bind_note(host: str) -> str:
     return "" if is_loopback_host(host) else f" [dim](listening on {host})[/dim]"
 
 
-def _webui_bootstrap_secret(config: Config) -> str:
+def webui_bootstrap_secret(config: Config) -> str:
+    """Return the shared local bootstrap credential for WebUI protocol clients."""
     ws_cfg = _webui_config_dict(config)
     return str(ws_cfg.get("tokenIssueSecret") or ws_cfg.get("token") or "").strip()
 
@@ -236,7 +244,7 @@ def _webui_browser_url(config: Config) -> str:
     host = _host_for_local_browser(str(ws_cfg.get("host") or "127.0.0.1"))
     port = int(ws_cfg.get("port") or 8765)
     base_url = f"http://{host}:{port}"
-    secret = _webui_bootstrap_secret(config)
+    secret = webui_bootstrap_secret(config)
     if not secret:
         return base_url
     return f"{base_url}/#/?bootstrapSecret={quote(secret, safe='')}"
@@ -382,15 +390,25 @@ def _print_foreground_port_conflict(
     gateway_host: str,
     gateway_port: int,
 ) -> None:
+    gateway_running = _gateway_health_ready(gateway_host, gateway_port)
+    if gateway_running:
         console.print(
-        "[red]Error: nanobot cannot start because one of its local ports is already in use.[/red]"
+            "[yellow]A nanobot gateway is already running for this local instance.[/yellow]"
         )
-    console.print(f"  WebUI: [cyan]{webui_url}[/cyan]")
+    else:
+        console.print(
+            "[red]Error: nanobot cannot start because one of its local ports "
+            "is already in use.[/red]"
+        )
+    console.print(f"  WebUI: [cyan]{_webui_display_url(webui_url)}[/cyan]")
     console.print(
         f"  Gateway health: "
         f"[cyan]http://{_host_for_local_browser(gateway_host)}:{gateway_port}/health[/cyan]"
     )
     console.print()
+    if gateway_running:
+        console.print("Use the existing instance, or stop it first:")
+    else:
         console.print("If this is an existing nanobot instance, use it or stop it first:")
     console.print("  [cyan]nanobot gateway status[/cyan]")
     console.print("  [cyan]nanobot gateway stop[/cyan]")
@@ -418,33 +436,31 @@ def _print_webui_foreground_lifecycle(*, attached: bool) -> None:
     """Explain how the browser and gateway lifecycles differ."""
     console.print()
     if attached:
-        console.print("[green]nanobot is attached to the existing gateway.[/green]")
+        console.print("[green]WebUI is attached to the shared gateway.[/green]")
     else:
-        console.print("[green]nanobot is running in this terminal.[/green]")
+        console.print("[green]WebUI is attached to the shared gateway.[/green]")
     console.print("[dim]Closing the browser does not stop channels or automations.[/dim]")
-    console.print("[dim]Press Ctrl+C here to stop nanobot.[/dim]")
+    console.print(
+        "[dim]Press Ctrl+C to detach; the gateway stops only when the last local client exits.[/dim]"
+    )
 
 
 def _attach_to_background_gateway(
     runtime: "GatewayRuntime",
     *,
     poll_hook: Callable[[], None] | None = None,
+    sleep: Callable[[float], None] = time.sleep,
 ) -> None:
-    """Keep a foreground WebUI command attached to a managed gateway."""
+    """Keep a WebUI launcher attached without taking ownership of the gateway."""
     _print_webui_foreground_lifecycle(attached=True)
     try:
         while runtime.status().running:
             if poll_hook is not None:
                 poll_hook()
-            time.sleep(0.5)
+            sleep(0.5)
     except KeyboardInterrupt:
-        console.print("\n[yellow]Stopping nanobot...[/yellow]")
-        result = runtime.stop()
-        if result.ok or result.message == "gateway_not_running":
-            console.print("[green]Gateway stopped.[/green]")
+        console.print("\n[yellow]WebUI launcher detached.[/yellow]")
         return
-        console.print(f"[red]Gateway could not be stopped: {result.message}[/red]")
-        raise typer.Exit(1)
 
     console.print("[yellow]Gateway stopped.[/yellow]")
 

nanobot/command/builtin.py

@@ -12,7 +12,7 @@ from dataclasses import dataclass
 from typing import TYPE_CHECKING, Any, Literal, cast
 
 from nanobot import __version__
-from nanobot.bus.events import OutboundMessage
+from nanobot.bus.events import INBOUND_META_USER_SHELL, OutboundMessage
 from nanobot.command.router import CommandContext, CommandRouter, normalize_command_text
 from nanobot.utils.helpers import build_status_content
 from nanobot.utils.restart import set_restart_notice_to_env
@@ -37,6 +37,8 @@ CommandLifecycle = Literal[
     "agent_turn_with_args",
 ]
 
+USER_SHELL_COMMAND = "/__shell"
+
 
 @dataclass(frozen=True)
 class BuiltinCommandSpec:
@@ -302,6 +304,7 @@ async def cmd_new(ctx: CommandContext) -> OutboundMessage:
     """Stop active task and start a fresh session."""
     loop = ctx.loop
     await loop._cancel_active_tasks(ctx.key)  # pyright: ignore[reportPrivateUsage]
+    loop.discard_session_file_state(ctx.key)
     session = ctx.session or loop.sessions.get_or_create(ctx.key)
     snapshot = session.messages[session.last_consolidated:]
     runtime = None
@@ -374,16 +377,7 @@ async def cmd_model(ctx: CommandContext) -> OutboundMessage:
             metadata=metadata,
         )
 
-    parts = args.split()
-    if len(parts) != 1:
-        return OutboundMessage(
-            channel=ctx.msg.channel,
-            chat_id=ctx.msg.chat_id,
-            content="Usage: `/model [preset]`",
-            metadata=metadata,
-        )
-
-    name = parts[0]
+    name = args
     try:
         runtime = loop.set_session_model_preset(ctx.key, name)
     except (KeyError, ValueError) as exc:
@@ -490,7 +484,7 @@ async def cmd_dream(ctx: CommandContext) -> OutboundMessage:
                 if sha:
                     content += f" (commit {sha})"
             store.compact_history()
-            prune_dream_sessions(loop.sessions.sessions_dir)
+            prune_dream_sessions(loop.sessions)
         await loop.bus.publish_outbound(OutboundMessage(
             channel=msg.channel, chat_id=msg.chat_id, content=content,
         ))
@@ -1007,6 +1001,30 @@ async def cmd_help(ctx: CommandContext) -> OutboundMessage:
     )
 
 
+async def cmd_user_shell(ctx: CommandContext) -> OutboundMessage:
+    """Run a trusted local ``!command`` through nanobot's exec policy."""
+    metadata = dict(ctx.msg.metadata or {})
+    if (
+        ctx.msg.channel != "websocket"
+        or metadata.get("webui") is not True
+        or metadata.get(INBOUND_META_USER_SHELL) is not True
+    ):
+        return OutboundMessage(
+            channel=ctx.msg.channel,
+            chat_id=ctx.msg.chat_id,
+            content="Shell commands are only available from a trusted local client.",
+            metadata={**metadata, "render_as": "text"},
+        )
+    if not ctx.args.strip():
+        return OutboundMessage(
+            channel=ctx.msg.channel,
+            chat_id=ctx.msg.chat_id,
+            content="Type a command after `!`, for example `!pwd`.",
+            metadata={**metadata, "render_as": "text"},
+        )
+    return await ctx.loop.execute_user_shell_command(ctx)
+
+
 def build_help_text() -> str:
     """Build canonical help text shared across channels."""
     lines = ["🐈 nanobot commands:"]
@@ -1046,3 +1064,5 @@ def register_builtin_commands(router: CommandRouter) -> None:
     router.exact("/help", cmd_help)
     router.exact("/pairing", cmd_pairing)
     router.prefix("/pairing ", cmd_pairing)
+    router.exact(USER_SHELL_COMMAND, cmd_user_shell)
+    router.prefix(f"{USER_SHELL_COMMAND} ", cmd_user_shell)

nanobot/config/loader.py

@@ -75,6 +75,7 @@ def load_config(config_path: Path | None = None) -> Config:
                 summary="Environment-based configuration is invalid.",
                 issues=validation_issues(exc),
             ) from exc
+        config.bind_source_path(path)
         _apply_ssrf_whitelist(config)
         return config
 
@@ -130,6 +131,7 @@ def load_config(config_path: Path | None = None) -> Config:
             issues=issues,
         ) from exc
 
+    config.bind_source_path(path)
     _apply_ssrf_whitelist(config)
     return config
 

nanobot/config/schema.py

@@ -4,7 +4,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, ClassVar, Literal, cast
 
-from pydantic import AliasChoices, ConfigDict, Field, field_validator, model_validator
+from pydantic import AliasChoices, ConfigDict, Field, PrivateAttr, field_validator, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 from nanobot.config.timezone import detect_system_timezone
@@ -97,7 +97,6 @@ FallbackCandidate = str | InlineFallbackConfig
 class ModelPresetConfig(Base):
     """A named set of model + generation parameters for quick switching."""
 
-    label: str | None = None
     model: str
     provider: str = "auto"
     max_tokens: int = 8192
@@ -262,6 +261,7 @@ class ProvidersConfig(Base):
     anthropic: ProviderConfig = Field(default_factory=ProviderConfig)
     openai: ProviderConfig = Field(default_factory=ProviderConfig)
     openrouter: ProviderConfig = Field(default_factory=ProviderConfig)
+    orcarouter: ProviderConfig = Field(default_factory=ProviderConfig)  # OrcaRouter API gateway
     assemblyai: ProviderConfig = Field(default_factory=ProviderConfig)  # AssemblyAI voice transcription
     huggingface: ProviderConfig = Field(default_factory=ProviderConfig)
     skywork: ProviderConfig = Field(default_factory=ProviderConfig)  # Skywork / APIFree API gateway
@@ -373,6 +373,7 @@ class MCPServerConfig(Base):
     """MCP server connection configuration (stdio or HTTP)."""
 
     type: Literal["stdio", "sse", "streamableHttp"] | None = None  # auto-detected if omitted
+    auth: Literal["oauth"] | None = None  # Remote MCP OAuth; tokens are stored outside config
     command: str = ""  # Stdio: command to run (e.g. "npx")
     args: list[str] = Field(default_factory=list)  # Stdio: command arguments
     env: dict[str, str] = Field(default_factory=dict)  # Stdio: extra env vars
@@ -430,6 +431,8 @@ class ToolsConfig(Base):
 class Config(BaseSettings):
     """Root configuration for nanobot."""
 
+    _source_path: Path | None = PrivateAttr(default=None)
+
     agents: AgentsConfig = Field(default_factory=AgentsConfig)
     channels: ChannelsConfig = Field(default_factory=ChannelsConfig)
     transcription: TranscriptionConfig = Field(default_factory=TranscriptionConfig)
@@ -448,8 +451,20 @@ class Config(BaseSettings):
             _resolve_tool_config_refs()
         super().__init__(**values)
 
+    def bind_source_path(self, path: Path) -> None:
+        """Record the config file that owns instance-level runtime data."""
+        self._source_path = path.expanduser().resolve(strict=False)
+
+    @property
+    def runtime_data_dir(self) -> Path | None:
+        """Return the active instance data directory when loaded from a config path."""
+        return self._source_path.parent if self._source_path is not None else None
+
     @model_validator(mode="after")
     def _validate_model_preset(self) -> "Config":
+        # Keep persisted names accepted by previous releases loadable. New
+        # names are normalized and checked case-insensitively at mutation
+        # boundaries, where conflicts can be reported without breaking startup.
         if "default" in self.model_presets:
             raise ValueError("model_preset name 'default' is reserved for agents.defaults")
         name = self.agents.defaults.model_preset

nanobot/cron/service.py

@@ -170,6 +170,7 @@ class CronService:
         self._timer_task: asyncio.Task[None] | None = None
         self._running = False
         self._active_executions = 0
+        self._store_dirty = False
         self.max_sleep_ms = max_sleep_ms
 
     def _should_persist_store(self) -> bool:
@@ -305,6 +306,11 @@ class CronService:
           load (during ``start``) can return ``None`` to signal an unrecoverable
           state to the caller.
         """
+        # Never replace state that a previous save failed to persist.  Reloading
+        # the older on-disk snapshot here could make an already executed job due
+        # again and repeat its side effect.
+        if self._store_dirty and self._store:
+            return self._store
         if self._active_executions > 0 and self._store and not reload_during_execution:
             return self._store
         loaded = self._load_jobs()
@@ -347,6 +353,9 @@ class CronService:
         if not self._store:
             return
 
+        # Set this before serialization/write so every exceptional exit keeps
+        # the in-memory snapshot authoritative until a later save succeeds.
+        self._store_dirty = True
         self.store_path.parent.mkdir(parents=True, exist_ok=True)
 
         data = {
@@ -399,6 +408,7 @@ class CronService:
         }
 
         self._atomic_write(self.store_path, json.dumps(data, indent=2, ensure_ascii=False))
+        self._store_dirty = False
 
     @staticmethod
     def _atomic_write(path: Path, content: str) -> None:
@@ -514,12 +524,18 @@ class CronService:
         reload_store = self._active_executions == 0
         self._active_executions += 1
         try:
+            # A prior tick may have completed external side effects but failed
+            # to persist their advanced schedule.  Persist that exact snapshot
+            # before reloading or executing anything else; otherwise the older
+            # disk state can replay the same job.
+            if self._store_dirty:
+                self._save_store()
+                return
+
             store = self._load_store(reload_during_execution=reload_store)
-            # If a hot reload found a corrupt store on disk, ``self._store`` may
-            # still hold the previous, known-good in-memory snapshot.  Keep using
-            # it rather than crashing the timer or wiping live jobs.
+            # If a hot reload found a corrupt store on disk, ``self._store``
+            # may still hold the previous, known-good in-memory snapshot.
             if store is None:
-                self._arm_timer()
                 return
 
             now = _now_ms()
@@ -532,8 +548,20 @@ class CronService:
                 await self._execute_job(job)
 
             self._save_store()
+        except Exception:
+            # A load/persist failure must not kill the scheduler: keep the
+            # in-memory store and retry on the next tick.  This mirrors the
+            # read-path defense in ``_load_jobs`` (``.corrupt-<ts>`` backups);
+            # ``_load_store`` may also persist (agent-binding migrations).
+            logger.exception(
+                "Cron: tick failed ({}); "
+                "keeping in-memory state and retrying on next tick",
+                self.store_path,
+            )
         finally:
             self._active_executions -= 1
+            # Always re-arm the timer, even on unexpected failures, so a
+            # single bad tick cannot silently stop all future jobs.
             self._arm_timer()
 
     async def _execute_job(self, job: CronJob) -> None:
@@ -797,6 +825,11 @@ class CronService:
         reload_store = self._active_executions == 0
         self._active_executions += 1
         try:
+            # A manual run is another side-effecting entrypoint.  Do not start
+            # it while the result of a previous timer execution is still only
+            # in memory.
+            if self._store_dirty:
+                self._save_store()
             store = self._require_store(reload_during_execution=reload_store)
             for job in store.jobs:
                 if job.id == job_id:

nanobot/gateway/__init__.py

@@ -1,6 +1,9 @@
 """Lightweight background runtime for the nanobot gateway."""
 
 from nanobot.gateway.runtime import (
+    GatewayAlreadyRunningError,
+    GatewayClientLease,
+    GatewayInstance,
     GatewayRuntime,
     GatewayRuntimePaths,
     GatewayStartOptions,
@@ -10,6 +13,9 @@ from nanobot.gateway.runtime import (
 )
 
 __all__ = [
+    "GatewayAlreadyRunningError",
+    "GatewayClientLease",
+    "GatewayInstance",
     "GatewayRuntime",
     "GatewayRuntimePaths",
     "GatewayStartOptions",

nanobot/gateway/runtime.py

@@ -1,14 +1,25 @@
 """Gateway-specific configuration for the shared background process runtime."""
 
+# pyright: reportPrivateUsage=false
+
 from __future__ import annotations
 
+import asyncio
 import hashlib
+import json
+import os
 import subprocess
+import tempfile
 import time
+import uuid
 from collections.abc import Callable
+from contextlib import contextmanager
 from dataclasses import dataclass
+from datetime import UTC, datetime
 from pathlib import Path
-from typing import Any
+from typing import Any, Generator, Literal, cast
+
+from filelock import FileLock
 
 from nanobot.config.paths import get_data_dir
 from nanobot.process_runtime import (
@@ -17,11 +28,51 @@ from nanobot.process_runtime import (
     ProcessRuntimePaths,
     ProcessStartOptions,
     ProcessStatus,
+    process_identity_record,
+    process_is_running,
 )
 
 GatewayStartOptions = ProcessStartOptions
-GatewayStatus = ProcessStatus
-RuntimeResult = ProcessResult
+
+GatewayLaunchMode = Literal["foreground", "background", "unknown"]
+GatewayLifetime = Literal["explicit", "on_demand"]
+
+
+def _default_config_path() -> Path:
+    return (Path.home() / ".nanobot" / "config.json").resolve(strict=False)
+
+
+@dataclass(frozen=True)
+class GatewayStatus(ProcessStatus):
+    """Observable lifecycle state for one shared local gateway."""
+
+    launch_mode: GatewayLaunchMode = "unknown"
+    lifetime: GatewayLifetime = "explicit"
+    clients: int = 0
+
+
+@dataclass(frozen=True)
+class GatewayLeaseSnapshot:
+    """Live local clients and the gateway lifetime they imply."""
+
+    auto_stop: bool
+    clients: int
+
+
+@dataclass(frozen=True)
+class RuntimeResult(ProcessResult):
+    """Result of a gateway lifecycle operation."""
+
+    status: GatewayStatus
+    promoted: bool = False
+
+
+class GatewayAlreadyRunningError(RuntimeError):
+    """Raised when a foreground gateway tries to replace a live instance."""
+
+    def __init__(self, status: GatewayStatus) -> None:
+        super().__init__("gateway_already_running")
+        self.status = status
 
 
 def build_gateway_command(python_executable: str, options: GatewayStartOptions) -> list[str]:
@@ -69,6 +120,56 @@ class GatewayRuntimePaths(ProcessRuntimePaths):
         )
 
 
+@dataclass(frozen=True)
+class GatewayInstance:
+    """One stable local gateway identity and its child-process selectors."""
+
+    config_path: Path
+    workspace: str | None
+    paths: GatewayRuntimePaths
+
+    @classmethod
+    def resolve(
+        cls,
+        *,
+        config_path: str | Path,
+        workspace: str | None = None,
+    ) -> "GatewayInstance":
+        resolved_config = Path(config_path).expanduser().resolve(strict=False)
+        resolved_workspace = (
+            str(Path(workspace).expanduser().resolve(strict=False)) if workspace else None
+        )
+        # The released default instance used gateway.json. Keep that identity stable
+        # across upgrades while still namespacing explicit configs and workspaces.
+        config_selector = (
+            None if resolved_config == _default_config_path() else str(resolved_config)
+        )
+        return cls(
+            config_path=resolved_config,
+            workspace=resolved_workspace,
+            paths=GatewayRuntimePaths.for_instance(
+                data_dir=resolved_config.parent,
+                workspace=resolved_workspace,
+                config_path=config_selector,
+            ),
+        )
+
+    def start_options(
+        self,
+        *,
+        port: int,
+        verbose: bool = False,
+    ) -> GatewayStartOptions:
+        return GatewayStartOptions(
+            port=port,
+            verbose=verbose,
+            workspace=self.workspace,
+            config_path=(
+                None if self.config_path == _default_config_path() else str(self.config_path)
+            ),
+        )
+
+
 class GatewayRuntime(ManagedProcessRuntime[ProcessStartOptions]):
     """Manage a background ``nanobot gateway`` process."""
 
@@ -96,6 +197,453 @@ class GatewayRuntime(ManagedProcessRuntime[ProcessStartOptions]):
     def _build_child_command(self, options: ProcessStartOptions) -> list[str]:
         return build_gateway_command(self.python_executable, options)
 
+    def _transition_lock(self) -> FileLock:
+        """Serialize long lifecycle transitions without blocking child cleanup."""
+        return FileLock(f"{self.paths.state_path}.transition.lock")
+
+    def start_background(self, options: ProcessStartOptions) -> RuntimeResult:
+        """Start the gateway detached from the current terminal."""
+        lease = GatewayClientLease(self, kind="gateway-background")
+        while True:
+            lease.wait_for_shutdown()
+            with self._transition_lock(), self._lifecycle_lock():
+                promoted = lease._try_mark_persistent_locked()
+                if promoted is None:
+                    continue
+                result = self._start_background(options)
+                return RuntimeResult(result.ok, result.message, result.status, promoted)
+
+    def start_on_demand(self, options: ProcessStartOptions) -> RuntimeResult:
+        """Atomically reuse a gateway or start one owned by local client leases."""
+        lease = GatewayClientLease(self, kind="gateway-start")
+        while True:
+            lease.wait_for_shutdown()
+            with self._transition_lock(), self._lifecycle_lock():
+                if lease._shutdown_pending_locked():
+                    continue
+                status = self.status()
+                if status.running:
+                    return RuntimeResult(False, "gateway_already_running", status)
+                lease._mark_ephemeral_locked()
+                return self._start_background(options)
+
+    def _start_background(self, options: ProcessStartOptions) -> RuntimeResult:
+        result = super()._start_background(options)
+        if not result.ok:
+            return self._result(result)
+        state = self._read_state()
+        if state and result.status.pid == state.get("pid"):
+            state["launch_mode"] = "background"
+            state["pending_pid_handoff"] = True
+            self._write_state(state)
+        return RuntimeResult(True, result.message, self.status())
+
+    def stop(self, *, timeout_s: int = 20) -> RuntimeResult:
+        """Stop the gateway recorded by this runtime."""
+        with self._transition_lock():
+            result = self._stop(timeout_s=timeout_s)
+            with self._lifecycle_lock():
+                if result.ok or result.message in {
+                    "gateway_not_running",
+                    "gateway_state_stale",
+                }:
+                    GatewayClientLease(self, kind="gateway-stop")._clear_locked()
+            return self._result(result)
+
+    def status(self, *, reason: str | None = None) -> GatewayStatus:
+        """Return process, launch, and client lifetime state in one snapshot."""
+        process = super().status(reason=reason)
+        state = self._read_state() if process.running else None
+        raw_mode = state.get("launch_mode") if state else None
+        launch_mode: GatewayLaunchMode = (
+            raw_mode if raw_mode in {"foreground", "background"} else "unknown"
+        )
+        lease = GatewayClientLease(self, kind="gateway-status").snapshot()
+        return GatewayStatus(
+            running=process.running,
+            pid=process.pid,
+            state_path=process.state_path,
+            log_path=process.log_path,
+            started_at=process.started_at,
+            port=process.port,
+            command=process.command,
+            reason=process.reason,
+            launch_mode=launch_mode,
+            lifetime="on_demand" if lease.auto_stop else "explicit",
+            clients=lease.clients,
+        )
+
+    @contextmanager
+    def foreground_instance(self, options: ProcessStartOptions) -> Generator[None]:
+        """Publish this foreground gateway while it is available to local clients."""
+        self._claim_current_process(options)
+        try:
+            yield
+        finally:
+            self._release_current_process()
+
+    def _claim_current_process(self, options: ProcessStartOptions) -> GatewayLaunchMode:
+        lease = GatewayClientLease(self, kind="gateway-foreground")
+        pid = os.getpid()
+        while True:
+            lease.wait_for_shutdown()
+            with self._transition_lock(), self._lifecycle_lock():
+                current = self.status()
+                state = self._read_state() or {}
+                pid_handoff = (
+                    self.platform_name == "Windows"
+                    and current.running
+                    and current.pid != pid
+                    and current.pid == os.getppid()
+                    and state.get("pid") == current.pid
+                    and state.get("launch_mode") == "background"
+                    and state.get("pending_pid_handoff") is True
+                )
+                if current.running and current.pid != pid and not pid_handoff:
+                    raise GatewayAlreadyRunningError(current)
+                if lease._shutdown_pending_locked():
+                    continue
+                launch_mode: GatewayLaunchMode = (
+                    "background"
+                    if state.get("launch_mode") == "background"
+                    and (state.get("pid") == pid or pid_handoff)
+                    else "foreground"
+                )
+                state.update(
+                    {
+                        "pid": pid,
+                        "started_at": datetime.now(UTC).isoformat(),
+                        "platform": self.platform_name,
+                        "port": options.port,
+                        "workspace": options.workspace,
+                        "config_path": options.config_path,
+                        "command": self._build_child_command(options),
+                        "log_path": str(self.paths.log_path),
+                        "launch_mode": launch_mode,
+                    }
+                )
+                state.pop("pending_pid_handoff", None)
+                state.pop("stable_identity", None)
+                state.update(self.process_identity_record(pid))
+                self._write_state(state)
+                if launch_mode == "foreground":
+                    lease._try_mark_persistent_locked()
+                return launch_mode
+
+    def _release_current_process(self) -> None:
+        with self._lifecycle_lock():
+            state = self._read_state()
+            if state and self._record_matches_process(state, os.getpid()):
+                self._clear_state()
+            GatewayClientLease(
+                self,
+                kind="gateway-exit",
+            )._finish_shutdown_locked()
+
+    def restart(self, options: ProcessStartOptions, *, timeout_s: int = 20) -> RuntimeResult:
+        """Restart an existing gateway without creating a new persistent instance."""
+        with self._transition_lock():
+            with self._lifecycle_lock():
+                status = self.status()
+                if not status.running:
+                    return RuntimeResult(False, "gateway_not_running", status)
+                if status.launch_mode == "foreground":
+                    return RuntimeResult(
+                        False,
+                        "gateway_foreground_restart_required",
+                        status,
+                    )
+            stop_result = self._stop(timeout_s=timeout_s)
+            if not stop_result.ok:
+                return self._result(stop_result)
+            with self._lifecycle_lock():
+                return self._start_background(options)
+
+    def _result(self, result: ProcessResult) -> RuntimeResult:
+        status = result.status
+        gateway_status = status if isinstance(status, GatewayStatus) else self.status()
+        return RuntimeResult(result.ok, result.message, gateway_status)
+
+
+class GatewayClientLease:
+    """Reference-count an on-demand gateway shared by local interactive clients."""
+
+    def __init__(
+        self,
+        runtime: GatewayRuntime,
+        *,
+        kind: str,
+        pid: int | None = None,
+        token: str | None = None,
+    ) -> None:
+        self.runtime = runtime
+        self.kind = kind
+        self.pid = pid or os.getpid()
+        self.token = token or uuid.uuid4().hex
+        state_path = runtime.paths.state_path
+        self.state_path = state_path.with_name(
+            f"{state_path.stem}.clients{state_path.suffix}"
+        )
+        self.transition_lock = FileLock(f"{state_path}.transition.lock")
+        self.lifecycle_lock = FileLock(f"{state_path}.lock")
+        self.lock = FileLock(f"{self.state_path}.lock")
+        self._acquired = False
+
+    def acquire(self) -> None:
+        """Register this client before it starts or attaches to the gateway."""
+        while True:
+            self.wait_for_shutdown()
+            with self.transition_lock, self.lifecycle_lock, self.lock:
+                state = self._live_state()
+                if state.get("stopping"):
+                    continue
+                self._register(state)
+                return
+
+    def ensure_on_demand_gateway(self, options: GatewayStartOptions) -> RuntimeResult:
+        """Atomically reuse a gateway or start one owned by local client leases."""
+        if not self._acquired:
+            raise RuntimeError("gateway client lease must be acquired before startup")
+        return self.runtime.start_on_demand(options)
+
+    def mark_ephemeral(self) -> None:
+        """Mark a gateway started by a client for last-client shutdown."""
+        with self.transition_lock, self.lifecycle_lock:
+            self._mark_ephemeral_locked()
+
+    def _mark_ephemeral_locked(self) -> None:
+        with self.lock:
+            state = self._live_state()
+            state["auto_stop"] = True
+            self._write_state(state)
+
+    def mark_persistent(self) -> bool:
+        """Keep an explicitly backgrounded gateway alive; return whether it was promoted."""
+        while True:
+            self.wait_for_shutdown()
+            with self.transition_lock, self.lifecycle_lock:
+                promoted = self._try_mark_persistent_locked()
+                if promoted is not None:
+                    return promoted
+
+    def _try_mark_persistent_locked(self) -> bool | None:
+        with self.lock:
+            state = self._live_state()
+            if state.get("stopping"):
+                return None
+            promoted = bool(state.get("auto_stop"))
+            state["auto_stop"] = False
+            self._write_or_clear(state)
+            return promoted
+
+    def clear(self) -> None:
+        """Forget leases after an explicit gateway stop."""
+        with self.transition_lock, self.lifecycle_lock:
+            self._clear_locked()
+
+    def _clear_locked(self) -> None:
+        with self.lock:
+            self.state_path.unlink(missing_ok=True)
+
+    def snapshot(self) -> GatewayLeaseSnapshot:
+        """Prune dead clients and return current lifetime state."""
+        with self.lock:
+            state = self._live_state()
+            self._write_or_clear(state)
+            return GatewayLeaseSnapshot(
+                auto_stop=bool(state.get("auto_stop")),
+                clients=len(self._clients(state)),
+            )
+
+    def begin_orphan_shutdown(self) -> bool:
+        """Commit shutdown only while an on-demand gateway still has no clients."""
+        with self.transition_lock, self.lifecycle_lock, self.lock:
+            state = self._live_state()
+            if not bool(state.get("auto_stop")) or self._clients(state):
+                self._write_or_clear(state)
+                return False
+            state["stopping"] = True
+            self._write_state(state)
+            return True
+
+    def release(self, *, timeout_s: int = 20) -> bool:
+        """Release this client and stop an ephemeral gateway when it was the last."""
+        if not self._acquired:
+            return False
+        while True:
+            self.wait_for_shutdown()
+            with self.transition_lock:
+                with self.lifecycle_lock, self.lock:
+                    state = self._live_state()
+                    if state.get("stopping"):
+                        continue
+                    clients = self._clients(state)
+                    clients.pop(self.token, None)
+                    self._acquired = False
+                    should_stop = not clients and bool(state.get("auto_stop"))
+                    self._write_or_clear(state)
+                if not should_stop:
+                    return False
+                result = self.runtime._stop(timeout_s=timeout_s)
+                stopped = result.ok or result.message in {
+                    "gateway_not_running",
+                    "gateway_state_stale",
+                }
+                with self.lifecycle_lock:
+                    if stopped:
+                        self._clear_locked()
+                    else:
+                        self._mark_ephemeral_locked()
+                return stopped
+
+    def wait_for_shutdown(self, *, timeout_s: float = 20) -> None:
+        """Wait until a committed orphan shutdown can no longer accept clients."""
+        deadline = time.monotonic() + timeout_s
+        while True:
+            with self.lifecycle_lock:
+                with self.lock:
+                    state = self._live_state()
+                    if not state.get("stopping"):
+                        return
+                if not self.runtime.status().running:
+                    self._finish_shutdown_locked()
+                    return
+            if time.monotonic() >= deadline:
+                raise RuntimeError("gateway is still shutting down; try again shortly")
+            time.sleep(0.05)
+
+    def _shutdown_pending_locked(self) -> bool:
+        with self.lock:
+            return bool(self._live_state().get("stopping"))
+
+    def _finish_shutdown_locked(self) -> None:
+        with self.lock:
+            state = self._live_state()
+            state.pop("stopping", None)
+            if not self._clients(state):
+                self.state_path.unlink(missing_ok=True)
+            else:
+                self._write_state(state)
+
+    def _register(self, state: dict[str, object]) -> None:
+        clients = self._clients(state)
+        record: dict[str, object] = {
+            "pid": self.pid,
+            "kind": self.kind,
+        }
+        record.update(process_identity_record(self._process_identity(self.pid), lease=True))
+        clients[self.token] = record
+        self._write_state(state)
+        self._acquired = True
+
+    def _live_state(self) -> dict[str, object]:
+        state = self._read_state()
+        clients = self._clients(state)
+        stale: list[str] = []
+        for token, value in clients.items():
+            if not isinstance(value, dict):
+                stale.append(token)
+                continue
+            record = cast(dict[str, object], value)
+            pid = record.get("pid")
+            identity = record.get("stable_identity")
+            if identity is None:
+                identity = record.get("identity")
+            if not isinstance(pid, int) or not self._process_is_running(pid):
+                stale.append(token)
+                continue
+            if self._process_identity_match(identity, pid) == "mismatch":
+                stale.append(token)
+        for token in stale:
+            clients.pop(token, None)
+        return state
+
+    def _process_identity(self, pid: int) -> str | int | None:
+        resolver = getattr(self.runtime, "process_identity", None)
+        value = resolver(pid) if callable(resolver) else None
+        return value if isinstance(value, (str, int)) else None
+
+    def _process_identity_match(
+        self,
+        recorded: object,
+        pid: int,
+    ) -> Literal["match", "mismatch", "unknown"]:
+        matcher = getattr(self.runtime, "process_identity_match", None)
+        if callable(matcher):
+            result = matcher(recorded, pid)
+            if result in {"match", "mismatch", "unknown"}:
+                return cast(Literal["match", "mismatch", "unknown"], result)
+        if recorded is None:
+            return "match"
+        current = self._process_identity(pid)
+        if current is None:
+            return "unknown"
+        return "match" if recorded == current else "mismatch"
+
+    def _process_is_running(self, pid: int) -> bool:
+        checker = getattr(self.runtime, "process_is_running", None)
+        return bool(checker(pid)) if callable(checker) else process_is_running(pid)
+
+    @staticmethod
+    def _clients(state: dict[str, object]) -> dict[str, object]:
+        value = state.get("clients")
+        if isinstance(value, dict):
+            return cast(dict[str, object], value)
+        clients: dict[str, object] = {}
+        state["clients"] = clients
+        return clients
+
+    def _read_state(self) -> dict[str, object]:
+        try:
+            payload: object = json.loads(self.state_path.read_text(encoding="utf-8"))
+        except (OSError, json.JSONDecodeError, ValueError):
+            return {"auto_stop": False, "clients": {}}
+        if isinstance(payload, dict):
+            return cast(dict[str, object], payload)
+        return {"auto_stop": False, "clients": {}}
+
+    def _write_or_clear(self, state: dict[str, object]) -> None:
+        clients = state.get("clients")
+        if not clients and not bool(state.get("auto_stop")):
+            self.state_path.unlink(missing_ok=True)
+            return
+        self._write_state(state)
+
+    def _write_state(self, state: dict[str, object]) -> None:
+        self.state_path.parent.mkdir(parents=True, exist_ok=True)
+        fd, temporary_name = tempfile.mkstemp(
+            prefix=f"{self.state_path.name}.",
+            suffix=".tmp",
+            dir=self.state_path.parent,
+        )
+        temporary = Path(temporary_name)
+        try:
+            with os.fdopen(fd, "w", encoding="utf-8") as handle:
+                json.dump(state, handle, indent=2, ensure_ascii=False)
+                handle.write("\n")
+                handle.flush()
+                os.fsync(handle.fileno())
+            temporary.replace(self.state_path)
+        finally:
+            temporary.unlink(missing_ok=True)
+
+
+async def monitor_gateway_clients(
+    lease: GatewayClientLease,
+    shutdown_event: asyncio.Event,
+    *,
+    poll_interval_s: float = 1.0,
+) -> bool:
+    """Stop waiting when an on-demand gateway loses every live client."""
+    while not shutdown_event.is_set():
+        try:
+            await asyncio.wait_for(shutdown_event.wait(), timeout=poll_interval_s)
+        except TimeoutError:
+            if lease.begin_orphan_shutdown():
+                shutdown_event.set()
+                return True
+    return False
+
 
 def _instance_suffix(*, workspace: str | None, config_path: str | None) -> str | None:
     raw = "|".join(value for value in (workspace, config_path) if value)

nanobot/nanobot.py

@@ -10,6 +10,8 @@ from typing import Any
 from nanobot.agent.hook import AgentHook, SDKCaptureHook
 from nanobot.agent.hooks import create_file_edit_activity_hook
 from nanobot.agent.loop import AgentLoop
+from nanobot.agent.tools.mcp import MCPProvider
+from nanobot.agent.tools.registry import ToolRegistry
 from nanobot.config.schema import Config
 from nanobot.providers.image_generation import image_gen_provider_configs
 from nanobot.sdk.clients import MemoryClient, RuntimeClient, SessionClient
@@ -71,9 +73,16 @@ class Nanobot:
         print(result.content)
     """
 
-    def __init__(self, loop: AgentLoop, *, config: Config | None = None) -> None:
+    def __init__(
+        self,
+        loop: AgentLoop,
+        *,
+        config: Config | None = None,
+        mcp_provider: MCPProvider | None = None,
+    ) -> None:
         self._loop = loop
         self._config = config
+        self._mcp_provider = mcp_provider
         self.sessions = SessionClient(loop)
         self.memory = MemoryClient(loop)
         self.runtime = RuntimeClient(loop)
@@ -120,12 +129,15 @@ class Nanobot:
         elif model_preset is not None:
             config.agents.defaults.model_preset = model_preset
 
+        tools = ToolRegistry()
+        mcp_provider = MCPProvider.from_config(config, tools)
         loop = AgentLoop.from_config(
             config,
             image_generation_provider_configs=image_gen_provider_configs(config),
             hook_factories=[create_file_edit_activity_hook],
+            tool_registry=tools,
         )
-        return cls(loop, config=config)
+        return cls(loop, config=config, mcp_provider=mcp_provider)
 
     async def run(
         self,
@@ -178,6 +190,8 @@ class Nanobot:
         )
         if runtime is not None:
             kwargs["runtime"] = runtime
+        if self._mcp_provider is not None:
+            await self._mcp_provider.connect()
         response = await self._loop.process_direct(
             message,
             **kwargs,
@@ -259,6 +273,8 @@ class Nanobot:
             if override_runtime is not None:
                 kwargs["runtime"] = override_runtime
             try:
+                if self._mcp_provider is not None:
+                    await self._mcp_provider.connect()
                 response = await self._loop.process_direct(
                     message,
                     **kwargs,
@@ -327,8 +343,12 @@ class Nanobot:
                 await run.aclose()
 
     async def aclose(self) -> None:
-        """Release resources held by this instance (MCP connections, etc.)."""
-        await self._loop.close_mcp()
+        """Release resources held by this instance."""
+        try:
+            await self._loop.aclose()
+        finally:
+            if self._mcp_provider is not None:
+                await self._mcp_provider.aclose()
 
     async def __aenter__(self) -> Nanobot:
         return self

nanobot/process_runtime.py

@@ -5,17 +5,21 @@ from __future__ import annotations
 import ctypes
 import json
 import os
+import re
 import signal
+import struct
 import subprocess
 import sys
 import tempfile
 import time
 from collections.abc import Callable
 from contextlib import suppress
+from ctypes import wintypes
 from dataclasses import dataclass
 from datetime import UTC, datetime
+from functools import lru_cache
 from pathlib import Path
-from typing import Any, Generic, TypeVar, cast
+from typing import Any, Generic, Literal, TypeVar, cast
 
 from filelock import FileLock
 
@@ -87,6 +91,10 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
         self._popen = popen
         self._subprocess_run = subprocess_run
         self._sleep = sleep
+        # Keep the handle for children spawned by this runtime. On POSIX an
+        # exited child remains visible to kill(pid, 0) until its parent reaps
+        # it; poll() both reaps it and reports the real lifecycle state.
+        self._owned_process: Any | None = None
 
     @classmethod
     def refresh_state_pid(cls, *, paths: ProcessRuntimePaths) -> None:
@@ -99,7 +107,8 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
             return
         state["pid"] = os.getpid()
         runtime = cls(paths=paths)
-        state["identity"] = runtime._process_identity(os.getpid())
+        state.pop("stable_identity", None)
+        state.update(runtime.process_identity_record(os.getpid()))
         state["started_at"] = _utc_now()
         runtime._write_state(state)
 
@@ -125,16 +134,15 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
                 stderr=subprocess.STDOUT,
                 **self._popen_platform_kwargs(),
             )
+        self._owned_process = process
 
         pid = int(process.pid)
         self._sleep(0.2)
         if not self._is_pid_running(pid):
             return ProcessResult(False, self._message("exited_during_startup"), self.status())
 
-        self._write_state(
-            {
+        state: dict[str, object] = {
             "pid": pid,
-                "identity": self._process_identity(pid),
             "started_at": _utc_now(),
             "platform": self.platform_name,
             "port": options.port,
@@ -143,7 +151,8 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
             "command": command,
             "log_path": str(self.paths.log_path),
         }
-        )
+        state.update(self.process_identity_record(pid))
+        self._write_state(state)
         return ProcessResult(True, self._message("started_background"), self.status())
 
     def stop(self, *, timeout_s: int = 20) -> ProcessResult:
@@ -157,7 +166,14 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
             return ProcessResult(False, self._message("not_running"), status)
 
         state = self._read_state()
-        if not self._record_matches_process(state, status.pid):
+        identity_match = self._process_identity_match(state, status.pid)
+        if identity_match == "unknown":
+            return ProcessResult(
+                False,
+                self._message("identity_unavailable"),
+                status,
+            )
+        if identity_match == "mismatch":
             self._clear_state()
             return ProcessResult(
                 False,
@@ -200,7 +216,8 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
             )
         assert state is not None
 
-        if not self._is_pid_running(pid) or not self._record_matches_process(state, pid):
+        identity_match = self._process_identity_match(state, pid)
+        if not self._is_pid_running(pid) or identity_match == "mismatch":
             self._clear_state()
             return ProcessStatus(
                 running=False,
@@ -219,7 +236,9 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
             started_at=_as_str(state.get("started_at")),
             port=_as_int(state.get("port")),
             command=tuple(cast(list[str], command)) if isinstance(command, list) else (),
-            reason=reason or "running",
+            reason=reason or (
+                "identity_unavailable" if identity_match == "unknown" else "running"
+            ),
         )
 
     def read_log_tail(self, *, tail: int = 200) -> list[str]:
@@ -250,6 +269,50 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
         except KeyboardInterrupt:
             return 130
 
+    def process_identity(self, pid: int) -> str | int | None:
+        """Return an identity that changes when an operating-system PID is reused."""
+        return self._process_identity(pid)
+
+    def process_identity_record(
+        self,
+        pid: int,
+        *,
+        lease: bool = False,
+    ) -> dict[str, str | int | None]:
+        """Serialize an identity without breaking pre-upgrade macOS readers."""
+        return process_identity_record(self._process_identity(pid), lease=lease)
+
+    def process_identity_match(
+        self,
+        recorded: object,
+        pid: int,
+    ) -> Literal["match", "mismatch", "unknown"]:
+        """Compare a recorded identity with the current process safely."""
+        if recorded is None:
+            return "match"
+        current = self._process_identity(pid)
+        if current is None:
+            return "unknown"
+        if recorded == current:
+            return "match"
+        # Older POSIX state files stored only the process group id.
+        if (
+            isinstance(recorded, int)
+            and isinstance(current, str)
+            and (
+                current.startswith(f"{recorded}:")
+                or current.startswith(f"darwin:{recorded}:")
+            )
+        ):
+            return "match"
+        if self.platform_name == "Darwin":
+            return _darwin_identity_match(recorded, current)
+        return "mismatch"
+
+    def process_is_running(self, pid: int) -> bool:
+        """Return whether the recorded operating-system process is still live."""
+        return self._is_pid_running(pid)
+
     def _message(self, event: str) -> str:
         return f"{self.service_name}_{event}"
 
@@ -295,26 +358,18 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
         return self._wait_for_exit(pid, 2)
 
     def _terminate_windows(self, pid: int, *, timeout_s: int) -> bool:
-        ctrl_break = getattr(signal, "CTRL_BREAK_EVENT", None)
-        if ctrl_break is not None:
-            ctrl_break_sent = False
-            try:
-                os.kill(pid, ctrl_break)
-            except ProcessLookupError:
-                return True
-            except OSError:
-                pass
-            else:
-                ctrl_break_sent = True
-            if ctrl_break_sent and self._wait_for_exit(pid, timeout_s):
-                return True
+        # ``os.kill(pid, CTRL_BREAK_EVENT)`` delegates to
+        # GenerateConsoleCtrlEvent.  That API targets a console process group,
+        # not an individual process, and can interrupt the caller when a
+        # detached/no-window child has no addressable console group.  Keep
+        # termination scoped to the recorded PID tree instead.
         self._subprocess_run(
             ["taskkill", "/PID", str(pid), "/T"],
             check=False,
             stdout=subprocess.DEVNULL,
             stderr=subprocess.DEVNULL,
         )
-        if self._wait_for_exit(pid, 2):
+        if self._wait_for_exit(pid, timeout_s):
             return True
         self._subprocess_run(
             ["taskkill", "/PID", str(pid), "/T", "/F"],
@@ -335,33 +390,65 @@ class ManagedProcessRuntime(Generic[_StartOptionsT]):
     def _is_pid_running(self, pid: int) -> bool:
         if pid <= 0:
             return False
-        if self.platform_name == "Windows":
-            return _windows_process_identity(pid) is not None
+        owned_process = self._owned_process
+        if owned_process is not None and getattr(owned_process, "pid", None) == pid:
+            poll = getattr(owned_process, "poll", None)
+            if callable(poll):
                 try:
-            os.kill(pid, 0)
-        except ProcessLookupError:
-            return False
-        except PermissionError:
-            return True
+                    return poll() is None
                 except OSError:
-            return False
-        return True
+                    pass
+        return process_is_running(pid, platform_name=self.platform_name)
 
     def _process_identity(self, pid: int) -> str | int | None:
-        if self.platform_name == "Windows":
+        # Process inspection must follow the host API even when tests inject a
+        # target platform.  On Windows, falling through to POSIX calls is not
+        # merely unsupported: ``os.kill(pid, 0)`` broadcasts CTRL_C_EVENT.
+        host_platform = _platform_name()
+        if host_platform == "Windows" or self.platform_name == "Windows":
             return _windows_process_identity(pid)
+        if self.platform_name == "Darwin":
+            birth = _darwin_process_birth(pid)
+            if birth is None:
+                return None
+            process_group, started_at_seconds, started_at_microseconds = birth
+            return (
+                f"darwin:{process_group}:{started_at_seconds}:"
+                f"{started_at_microseconds}"
+            )
         try:
-            return os.getpgid(pid)
+            process_group = os.getpgid(pid)
         except OSError:
             return None
+        started_at = self._posix_process_started_at(pid)
+        return f"{process_group}:{started_at}" if started_at else process_group
+
+    def _posix_process_started_at(self, pid: int) -> str | None:
+        if self.platform_name == "Linux":
+            try:
+                stat = Path(f"/proc/{pid}/stat").read_text(encoding="utf-8")
+            except OSError:
+                return None
+            closing_paren = stat.rfind(")")
+            fields = stat[closing_paren + 2 :].split() if closing_paren >= 0 else []
+            # /proc/<pid>/stat fields after comm begin at field 3; starttime is field 22.
+            return fields[19] if len(fields) > 19 else None
+        return None
 
     def _record_matches_process(self, state: dict[str, Any] | None, pid: int) -> bool:
+        return self._process_identity_match(state, pid) == "match"
+
+    def _process_identity_match(
+        self,
+        state: dict[str, Any] | None,
+        pid: int,
+    ) -> Literal["match", "mismatch", "unknown"]:
         if not state:
-            return False
-        recorded = state.get("identity")
+            return "mismatch"
+        recorded = state.get("stable_identity")
         if recorded is None:
-            return True
-        return recorded == self._process_identity(pid)
+            recorded = state.get("identity")
+        return self.process_identity_match(recorded, pid)
 
     def _read_state(self) -> dict[str, Any] | None:
         try:
@@ -401,6 +488,52 @@ def _platform_name() -> str:
     return "Linux"
 
 
+def process_is_running(pid: int, *, platform_name: str | None = None) -> bool:
+    """Probe a PID without delivering a control event on Windows."""
+    if pid <= 0:
+        return False
+    host_platform = _platform_name()
+    if host_platform == "Windows" or platform_name == "Windows":
+        # On Windows ``os.kill(pid, 0)`` sends CTRL_C_EVENT (whose value is 0)
+        # instead of performing the harmless POSIX existence probe.
+        return _windows_process_identity(pid) is not None
+    try:
+        os.kill(pid, 0)
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+    except OSError:
+        return False
+    return _posix_process_state(pid, platform_name=host_platform) != "Z"
+
+
+def _posix_process_state(pid: int, *, platform_name: str) -> str | None:
+    """Return the host process state when available; zombies are not live clients."""
+    if platform_name == "Linux":
+        try:
+            stat = Path(f"/proc/{pid}/stat").read_text(encoding="utf-8")
+        except OSError:
+            return None
+        closing_paren = stat.rfind(")")
+        fields = stat[closing_paren + 2 :].split() if closing_paren >= 0 else []
+        return fields[0] if fields else None
+    if platform_name == "Darwin":
+        try:
+            result = subprocess.run(
+                ["ps", "-o", "stat=", "-p", str(pid)],
+                check=False,
+                capture_output=True,
+                text=True,
+                timeout=1,
+            )
+        except (OSError, subprocess.SubprocessError):
+            return None
+        value = getattr(result, "stdout", "").strip()
+        return value[:1].upper() or None
+    return None
+
+
 def _utc_now() -> str:
     return datetime.now(UTC).isoformat().replace("+00:00", "Z")
 
@@ -420,6 +553,145 @@ def _as_str(value: object) -> str | None:
     return value if isinstance(value, str) else None
 
 
+def _darwin_identity_match(
+    recorded: object,
+    current: object,
+) -> Literal["match", "mismatch", "unknown"]:
+    """Compare the new numeric identity with a pre-upgrade ``ps`` identity."""
+    if not isinstance(recorded, str) or not isinstance(current, str):
+        return "mismatch"
+    current_identity = _parse_darwin_identity(current)
+    if current_identity is None:
+        return "mismatch"
+    current_group, current_seconds, _ = current_identity
+    recorded_group, separator, recorded_started_at = recorded.partition(":")
+    if not separator or not recorded_group.isdigit():
+        return "mismatch"
+    if int(recorded_group) != current_group:
+        return "mismatch"
+    legacy_epoch = _legacy_darwin_started_at(recorded_started_at)
+    if legacy_epoch is None:
+        # The PID is alive and its process group still matches, but an older
+        # locale produced a date we cannot safely parse. Keep the record until
+        # the owning client exits instead of killing a live gateway.
+        return "unknown"
+    return "match" if legacy_epoch == current_seconds else "mismatch"
+
+
+def _parse_darwin_identity(value: object) -> tuple[int, int, int] | None:
+    if not isinstance(value, str):
+        return None
+    match = re.fullmatch(r"darwin:(\d+):(\d+):(\d+)", value)
+    if match is None:
+        return None
+    return int(match.group(1)), int(match.group(2)), int(match.group(3))
+
+
+def process_identity_record(
+    identity: str | int | None,
+    *,
+    lease: bool = False,
+) -> dict[str, str | int | None]:
+    """Serialize an identity without breaking pre-upgrade macOS readers."""
+    darwin = _parse_darwin_identity(identity)
+    if darwin is None:
+        return {"identity": identity}
+    process_group, _, _ = darwin
+    # Old process-state readers understand a PGID-only integer. Old lease
+    # readers raw-compare identities, so ``None`` asks them to rely on the
+    # still-live PID while upgraded readers use the stable native value.
+    return {
+        "identity": None if lease else process_group,
+        "stable_identity": identity,
+    }
+
+
+def _legacy_darwin_started_at(value: str) -> int | None:
+    """Parse the English and numeric macOS ``ps lstart`` formats we released."""
+    english = re.fullmatch(
+        r"[A-Za-z]{3}\s+([A-Za-z]{3})\s+(\d{1,2})\s+"
+        r"(\d{2}):(\d{2}):(\d{2})\s+(\d{4})",
+        value.strip(),
+    )
+    months = {
+        "Jan": 1,
+        "Feb": 2,
+        "Mar": 3,
+        "Apr": 4,
+        "May": 5,
+        "Jun": 6,
+        "Jul": 7,
+        "Aug": 8,
+        "Sep": 9,
+        "Oct": 10,
+        "Nov": 11,
+        "Dec": 12,
+    }
+    if english is not None:
+        month = months.get(english.group(1))
+        if month is None:
+            return None
+        day, hour, minute, second, year = map(int, english.groups()[1:])
+    else:
+        numeric = re.fullmatch(
+            r"\S+\s+(\d{1,2})/(\d{1,2})\s+"
+            r"(\d{2}):(\d{2}):(\d{2})\s+(\d{4})",
+            value.strip(),
+        )
+        if numeric is None:
+            return None
+        month, day, hour, minute, second, year = map(int, numeric.groups())
+    try:
+        return int(time.mktime((year, month, day, hour, minute, second, -1, -1, -1)))
+    except (OSError, OverflowError, ValueError):
+        return None
+
+
+@lru_cache(maxsize=1)
+def _darwin_proc_pidinfo() -> Any | None:
+    if sys.platform != "darwin":
+        return None
+    try:
+        proc_pidinfo = ctypes.CDLL(
+            "/usr/lib/libproc.dylib",
+            use_errno=True,
+        ).proc_pidinfo
+    except (AttributeError, OSError):
+        return None
+    proc_pidinfo.argtypes = [
+        ctypes.c_int,
+        ctypes.c_int,
+        ctypes.c_uint64,
+        ctypes.c_void_p,
+        ctypes.c_int,
+    ]
+    proc_pidinfo.restype = ctypes.c_int
+    return proc_pidinfo
+
+
+def _darwin_process_birth(pid: int) -> tuple[int, int, int] | None:
+    """Read PGID and microsecond process birth time from ``proc_bsdinfo``."""
+    proc_pidinfo = _darwin_proc_pidinfo()
+    if proc_pidinfo is None:
+        return None
+    # ``proc_bsdinfo`` is 136 bytes on supported macOS versions. These stable
+    # field offsets come from ``sys/proc_info.h``: pid=12, pgid=100,
+    # start_tvsec=120, and start_tvusec=128.
+    buffer = ctypes.create_string_buffer(136)
+    try:
+        written = proc_pidinfo(pid, 3, 0, buffer, len(buffer))
+    except (OSError, ValueError):
+        return None
+    if written != len(buffer) or struct.unpack_from("=I", buffer, 12)[0] != pid:
+        return None
+    process_group = struct.unpack_from("=I", buffer, 100)[0]
+    started_at_seconds = struct.unpack_from("=Q", buffer, 120)[0]
+    started_at_microseconds = struct.unpack_from("=Q", buffer, 128)[0]
+    if started_at_seconds <= 0:
+        return None
+    return process_group, started_at_seconds, started_at_microseconds
+
+
 def _windows_process_identity(pid: int) -> str | None:
     if os.name != "nt":
         return None
@@ -432,7 +704,21 @@ def _windows_process_identity(pid: int) -> str | None:
             return (int(self.high) << 32) | int(self.low)
 
     process_query_limited_information = 0x1000
-    kernel32 = ctypes.windll.kernel32
+    kernel32 = ctypes.WinDLL("kernel32", use_last_error=True)
+    kernel32.OpenProcess.argtypes = [wintypes.DWORD, wintypes.BOOL, wintypes.DWORD]
+    kernel32.OpenProcess.restype = wintypes.HANDLE
+    kernel32.GetProcessTimes.argtypes = [
+        wintypes.HANDLE,
+        ctypes.POINTER(FileTime),
+        ctypes.POINTER(FileTime),
+        ctypes.POINTER(FileTime),
+        ctypes.POINTER(FileTime),
+    ]
+    kernel32.GetProcessTimes.restype = wintypes.BOOL
+    kernel32.GetExitCodeProcess.argtypes = [wintypes.HANDLE, ctypes.POINTER(wintypes.DWORD)]
+    kernel32.GetExitCodeProcess.restype = wintypes.BOOL
+    kernel32.CloseHandle.argtypes = [wintypes.HANDLE]
+    kernel32.CloseHandle.restype = wintypes.BOOL
     handle = kernel32.OpenProcess(process_query_limited_information, False, pid)
     if not handle:
         return None
@@ -450,7 +736,7 @@ def _windows_process_identity(pid: int) -> str | None:
         )
         if not ok:
             return None
-        exit_code = ctypes.c_uint32()
+        exit_code = wintypes.DWORD()
         if not kernel32.GetExitCodeProcess(handle, ctypes.byref(exit_code)):
             return None
         if exit_code.value != 259:

nanobot/providers/anthropic_provider.py

@@ -782,11 +782,15 @@ class AnthropicProvider(LLMProvider):
         idle_timeout_s = resolve_stream_idle_timeout_s()
         try:
             async with self._client.messages.stream(**kwargs) as stream:
-                if on_content_delta or on_thinking_delta or on_tool_call_delta:
                 # Idle timeout must track *any* SSE chunk (thinking_delta,
                 # tool JSON deltas, etc.), not only text_stream tokens.
                 # Otherwise extended thinking can stall text_stream for minutes
                 # while the connection is healthy (e.g. MiniMax Anthropic).
+                # Drain the whole stream with per-chunk idle waits so the
+                # timeout measures inactivity, not total generation time: a
+                # long but continuously-active stream must never be killed.
+                # The SDK accumulates the final message snapshot during
+                # iteration, so get_final_message() below returns instantly.
                 tool_blocks: dict[int, dict[str, str]] = {}
                 while True:
                     try:
@@ -839,10 +843,7 @@ class AnthropicProvider(LLMProvider):
                                 "name": state.get("name", ""),
                                 "arguments_delta": partial,
                             })
-                response = await asyncio.wait_for(
-                    stream.get_final_message(),
-                    timeout=idle_timeout_s,
-                )
+                response = await stream.get_final_message()
             return self._parse_response(response)
         except asyncio.TimeoutError:
             return LLMResponse(

nanobot/providers/base.py

@@ -258,6 +258,12 @@ class LLMResponse:
     tool_calls: list[ToolCallRequest] = field(default_factory=list)
     finish_reason: str = "stop"
     usage: dict[str, int] = field(default_factory=dict)
+    # Locally measured streaming telemetry. ``generation_ms`` excludes time to
+    # first token and provider retry gaps; ``ttft_ms`` measures the first
+    # streamed reasoning/content delta from request start. They stay separate
+    # from provider usage because providers do not report these consistently.
+    generation_ms: int | None = None
+    ttft_ms: int | None = None
     retry_after: float | None = None  # Provider supplied retry wait in seconds.
     reasoning_content: str | None = None  # Kimi, DeepSeek-R1, MiMo etc.
     thinking_blocks: list[dict[str, Any]] | None = None  # Anthropic extended thinking

nanobot/providers/factory.py

@@ -24,6 +24,7 @@ class ProviderSnapshot:
 @dataclass(frozen=True)
 class _ProviderSetup:
     model: str
+    provider_name: str
     provider_config: ProviderConfig | None
     spec: ProviderSpec | None
     backend: str
@@ -48,6 +49,30 @@ def _provider_extra_headers(
     return headers or None
 
 
+def _provider_spec_for_config(
+    provider_name: str,
+    provider_config: ProviderConfig | None,
+) -> ProviderSpec | None:
+    spec = find_by_name(provider_name)
+    if (
+        spec is not None
+        and spec.name == "orcarouter"
+        and provider_config is not None
+        and provider_config.api_base
+        and provider_config.api_base.rstrip("/").lower()
+        != spec.default_api_base.rstrip("/").lower()
+    ):
+        # Before OrcaRouter became a built-in provider, this name was valid for a
+        # dynamic custom provider. Preserve that provider's model-prefix behavior
+        # when an existing config points the name at a different endpoint.
+        return create_dynamic_spec(
+            provider_name,
+            display_name=provider_config.display_name or "",
+            thinking_style=provider_config.thinking_style or "",
+        )
+    return spec
+
+
 def _resolve_provider_setup(
     config: Config,
     *,
@@ -60,7 +85,7 @@ def _resolve_provider_setup(
     p = config.get_provider(model, preset=preset)
     if not provider_name:
         raise ValueError(f"No provider is configured for model '{model}'.")
-    spec = find_by_name(provider_name)
+    spec = _provider_spec_for_config(provider_name, p)
     if not spec and p:
         if not p.api_base:
             raise ValueError(f"Provider '{provider_name}' requires api_base in config.")
@@ -99,6 +124,7 @@ def _resolve_provider_setup(
 
     return _ProviderSetup(
         model=model,
+        provider_name=provider_name,
         provider_config=p,
         spec=spec,
         backend=backend,
@@ -134,6 +160,7 @@ def _make_provider_core(
         model=model,
     )
     model = setup.model
+    provider_name = setup.provider_name
     p = setup.provider_config
     spec = setup.spec
     backend = setup.backend
@@ -198,7 +225,7 @@ def _make_provider_core(
             extra_headers=_provider_extra_headers(spec, p),
             spec=spec,
             extra_body=p.extra_body if p else None,
-            api_type=p.api_type if p else "auto",
+            api_type=p.api_type if p and provider_name == "openai" else "auto",
             extra_query=p.extra_query if p else None,
             proxy=p.proxy if p else None,
         )

nanobot/providers/openai_compat_provider.py

@@ -49,13 +49,15 @@ from nanobot.providers.openai_responses import (
 if TYPE_CHECKING:
     from openai import AsyncOpenAI as AsyncOpenAIType
 
-    from nanobot.providers.registry import ProviderSpec, ResponsesCapabilities
+    from nanobot.providers.registry import ProviderSpec
 
 # Module-level placeholder — set lazily by _ensure_client on first real
 # use, or replaced by tests via ``patch(...)``.  Kept as a plain name so
 # that ``unittest.mock.patch`` can find and replace it.
 AsyncOpenAI: Any = None
 
+_GEMINI_SKIP_THOUGHT_SIGNATURE = "skip_thought_signature_validator"
+
 
 def _is_hosted_web_search_type(value: object) -> bool:
     return isinstance(value, str) and (
@@ -447,6 +449,28 @@ def _merge_unique_list(base: object, override: object) -> object:
     return result
 
 
+def _merge_chat_extra_body(
+    kwargs: dict[str, Any],
+    extra_body: dict[str, Any],
+) -> dict[str, Any]:
+    """Merge configured Chat Completions fields without clobbering tools."""
+    regular_extra = {key: value for key, value in extra_body.items() if key != "tools"}
+    merged = dict(kwargs)
+    if regular_extra:
+        existing = kwargs.get("extra_body", {})
+        merged["extra_body"] = _deep_merge(existing, regular_extra)
+
+    if "tools" in extra_body:
+        current_tools = kwargs.get("tools")
+        configured_tools = extra_body["tools"]
+        if isinstance(current_tools, list) and isinstance(configured_tools, list):
+            merged["tools"] = [*current_tools, *configured_tools]
+        else:
+            merged["tools"] = configured_tools
+
+    return merged
+
+
 def _merge_responses_extra_body(
     body: dict[str, Any],
     extra_body: dict[str, Any],
@@ -496,19 +520,11 @@ class OpenAICompatProvider(LLMProvider):
         self.extra_headers = extra_headers or {}
         self._spec = spec
         self._extra_body = dict(extra_body or {})
-        responses = spec.responses if spec is not None else None
-        self._api_type = (
-            api_type
-            if responses is not None and responses.allows_api_type_override
-            else "auto"
-        )
+        self._api_type = api_type if spec and spec.name == "openai" else "auto"
         self._extra_query = extra_query or {}
         self._proxy = proxy or None
         self._native_compaction_available = True
 
-        if api_key and spec and spec.env_key:
-            self._setup_env(api_key, api_base)
-
         effective_base = api_base or (spec.default_api_base if spec else None) or None
         self._effective_base = effective_base
         self._default_headers = {"x-session-affinity": uuid.uuid4().hex}
@@ -601,20 +617,6 @@ class OpenAICompatProvider(LLMProvider):
                 raise RuntimeError("OpenAI client initialization did not produce a client")
             return self._client
 
-    def _setup_env(self, api_key: str, api_base: str | None) -> None:
-        """Set environment variables based on provider spec."""
-        spec = self._spec
-        if not spec or not spec.env_key:
-            return
-        if spec.is_gateway:
-            os.environ[spec.env_key] = api_key
-        else:
-            os.environ.setdefault(spec.env_key, api_key)
-        effective_base = api_base or spec.default_api_base
-        for env_name, env_val in spec.env_extras:
-            resolved = env_val.replace("{api_key}", api_key).replace("{api_base}", effective_base)
-            os.environ.setdefault(env_name, resolved)
-
     @classmethod
     def _apply_cache_control(
         cls,
@@ -690,6 +692,8 @@ class OpenAICompatProvider(LLMProvider):
         if strip_reasoning:
             for msg in sanitized:
                 msg.pop("reasoning_content", None)
+        if self._spec and self._spec.name == "gemini":
+            sanitized = self._ensure_gemini_thought_signatures(sanitized)
 
         def map_id(value: Any) -> Any:
             if not isinstance(value, str):
@@ -767,6 +771,81 @@ class OpenAICompatProvider(LLMProvider):
                 clean["content"] = self._coerce_content_to_string(clean.get("content"))
         return self._enforce_role_alternation(sanitized)
 
+    @staticmethod
+    def _gemini_thought_signature(tool_call: dict[str, Any]) -> str | None:
+        """Return Gemini's thought signature attached to a tool call, if any.
+
+        Gemini's OpenAI-compatible endpoint returns tool calls with an
+        ``extra_content`` field: ``{"google": {"thought_signature": "..."}}``.
+        nanobot preserves it through the parse -> serialize round-trip so
+        replayed calls stay valid. Calls produced by other providers (e.g.
+        after a mid-conversation model switch) carry no signature.
+        """
+        extra = tool_call.get("extra_content")
+        if not isinstance(extra, dict):
+            return None
+        google = cast(dict[str, Any], extra).get("google")
+        if not isinstance(google, dict):
+            return None
+        signature = cast(dict[str, Any], google).get("thought_signature")
+        if isinstance(signature, str) and signature:
+            return signature
+        return None
+
+    def _ensure_gemini_thought_signatures(
+        self, messages: list[dict[str, Any]]
+    ) -> list[dict[str, Any]]:
+        """Keep migrated tool history wire-valid without losing tool context.
+
+        Gemini requires the first call in each function-call step to carry a
+        thought signature. Native parallel calls intentionally leave later
+        calls unsigned, so they must remain in their original order. For a
+        fully unsigned step imported from another provider, Google documents
+        ``skip_thought_signature_validator`` as a last-resort migration value.
+        """
+        kept: list[dict[str, Any]] = []
+        for msg in messages:
+            role = msg.get("role")
+            calls = msg.get("tool_calls")
+            if role != "assistant" or not isinstance(calls, list) or not calls:
+                kept.append(msg)
+                continue
+
+            call_values = cast(list[object], calls)
+            typed_calls = [
+                cast(dict[str, Any], tool_call)
+                for tool_call in call_values
+                if isinstance(tool_call, dict)
+            ]
+            if not typed_calls:
+                if msg.get("content"):
+                    clean = dict(msg)
+                    clean.pop("tool_calls", None)
+                    kept.append(clean)
+                continue
+
+            clean_calls = typed_calls
+            if self._gemini_thought_signature(typed_calls[0]) is None:
+                first = dict(typed_calls[0])
+                extra_value = first.get("extra_content")
+                extra = dict(cast(dict[str, Any], extra_value)) if isinstance(
+                    extra_value, dict
+                ) else {}
+                google_value = extra.get("google")
+                google = dict(cast(dict[str, Any], google_value)) if isinstance(
+                    google_value, dict
+                ) else {}
+                google["thought_signature"] = _GEMINI_SKIP_THOUGHT_SIGNATURE
+                extra["google"] = google
+                first["extra_content"] = extra
+                clean_calls = [first, *typed_calls[1:]]
+
+            if clean_calls != call_values:
+                msg = dict(msg)
+                msg["tool_calls"] = clean_calls
+            kept.append(msg)
+        return kept
+
     # ------------------------------------------------------------------
     # Build kwargs
     # ------------------------------------------------------------------
@@ -973,14 +1052,11 @@ class OpenAICompatProvider(LLMProvider):
                 if msg.get("role") == "assistant" and "reasoning_content" not in msg:
                     msg["reasoning_content"] = ""
 
-        # Merge user-configured extra_body last so it can override or
-        # extend provider-specific defaults (e.g. chat_template_kwargs,
-        # guided_json, repetition_penalty).  Uses recursive merge so
-        # nested dicts like {"chat_template_kwargs": {"enable_thinking": false}}
-        # do not clobber sibling keys already set by thinking-style logic.
+        # Merge user-configured extra_body last so ordinary fields can override
+        # provider defaults. Keep configured tools at the top level: the SDK
+        # otherwise lets extra_body.tools replace nanobot's generated functions.
         if self._extra_body:
-            existing = kwargs.get("extra_body", {})
-            kwargs["extra_body"] = _deep_merge(existing, self._extra_body)
+            kwargs = _merge_chat_extra_body(kwargs, self._extra_body)
 
         return kwargs
 
@@ -992,33 +1068,35 @@ class OpenAICompatProvider(LLMProvider):
         """Choose Responses for providers/models that explicitly support it."""
         if self._api_type == "chat_completions":
             return False
-        capabilities = self._responses_capabilities()
-        if capabilities is None:
-            return False
+        spec_name = self._spec.name if self._spec is not None else None
         model_name = self._request_model_name(model or self.default_model).lower()
-        if self._api_type == "responses":
-            # Explicit configuration means Responses is mandatory; do not
+        supported_models = {
+            supported.lower()
+            for supported in getattr(self._spec, "responses_models", ())
+        }
+        model_responses = any(
+            model_name == supported or model_name.endswith(f"/{supported}")
+            for supported in supported_models
+        )
+        provider_responses = spec_name in ("openai", "github_copilot")
+        if not provider_responses and not model_responses:
+            return False
+        if self._responses_is_required():
+            # Explicit Responses-only request fields are mandatory; do not
             # consult the circuit breaker or fall back to Chat Completions.
             return True
-
-        explicitly_supported = capabilities.matches_model(model_name)
-        if self._hosted_web_search_enabled() and (
-            capabilities.auto_route or explicitly_supported
-        ):
-            # Provider-hosted tools require Responses on models that the
-            # capability profile declares eligible for that transport.
-            return True
-        if (
-            capabilities.requires_direct_openai_base
-            and not _is_direct_openai_base(self._effective_base)
-        ):
+        if provider_responses and (self._spec is None or self._spec.name != "github_copilot"):
+            if not _is_direct_openai_base(self._effective_base):
                 return False
 
-        wants_auto_route = capabilities.auto_route and (
-            (reasoning_effort is not None and reasoning_effort.lower() != "none")
-            or any(token in model_name for token in ("gpt-5", "o1", "o3", "o4"))
-        )
-        if not explicitly_supported and not wants_auto_route:
+        wants = False
+        if model_responses:
+            wants = True
+        elif reasoning_effort and reasoning_effort.lower() != "none":
+            wants = True
+        elif any(token in model_name for token in ("gpt-5", "o1", "o3", "o4")):
+            wants = True
+        if not wants:
             return False
 
         return self._responses_circuit_allows_probe(model, reasoning_effort)
@@ -1042,9 +1120,6 @@ class OpenAICompatProvider(LLMProvider):
             )
         )
 
-    def _responses_capabilities(self) -> ResponsesCapabilities | None:
-        return self._spec.responses if self._spec is not None else None
-
     def _responses_state_provider(self) -> str:
         spec_name = self._spec.name if self._spec is not None else "custom"
         effective_base = self._effective_base or "https://api.openai.com/v1"
@@ -1067,20 +1142,14 @@ class OpenAICompatProvider(LLMProvider):
     def supports_native_compaction(self, model: str | None = None) -> bool:
         """Enable server compaction only on direct OpenAI Responses endpoints."""
         _ = model
-        capabilities = self._responses_capabilities()
         if (
             not self._native_compaction_available
             or self._api_type == "chat_completions"
-            or capabilities is None
-            or not capabilities.supports_native_compaction
         ):
             return False
-        if (
-            capabilities.requires_direct_openai_base
-            and not _is_direct_openai_base(self._effective_base)
-        ):
+        if self._spec is not None and self._spec.name != "openai":
             return False
-        return True
+        return _is_direct_openai_base(self._effective_base)
 
     def _responses_circuit_allows_probe(
         self,
@@ -1168,10 +1237,8 @@ class OpenAICompatProvider(LLMProvider):
                     self._sanitize_empty_content(sanitized_state.pending_messages)
                 )
             )
-        capabilities = self._responses_capabilities()
-        preserve_reasoning = (
-            capabilities is not None and capabilities.reasoning_replay == "plaintext"
-        )
+        is_deepseek = bool(self._spec and self._spec.name == "deepseek")
+        preserve_reasoning = is_deepseek
         instructions, input_items, replayed = prepare_responses_input(
             sanitized_messages,
             state=sanitized_state,
@@ -1202,17 +1269,12 @@ class OpenAICompatProvider(LLMProvider):
                 "compact_threshold": compact_threshold,
             }]
 
-        supports_temperature = self._supports_temperature(model_name, reasoning_effort)
-        if supports_temperature:
+        if self._supports_temperature(model_name, reasoning_effort):
             body["temperature"] = temperature
 
-        if (
-            not supports_temperature
-            and capabilities is not None
-            and capabilities.reasoning_replay == "encrypted"
-        ):
+        if not self._supports_temperature(model_name, reasoning_effort) and not preserve_reasoning:
             body["include"] = ["reasoning.encrypted_content"]
-        if reasoning_effort and reasoning_effort.lower() != "none":
+        if reasoning_effort and (reasoning_effort.lower() != "none" or is_deepseek):
             body["reasoning"] = {"effort": reasoning_effort}
         if replayed and "gpt-5.6" in model_name.lower():
             body.setdefault("reasoning", {})["context"] = "all_turns"
@@ -1860,8 +1922,10 @@ class OpenAICompatProvider(LLMProvider):
                     self._record_responses_success(model, reasoning_effort)
                     return result
                 except Exception as responses_error:
-                    capabilities = self._responses_capabilities()
-                    if capabilities is not None and not capabilities.allows_chat_fallback:
+                    if self._spec and self._spec.name == "github_copilot":
+                        # Copilot gateway exposes GPT-5/o-series only via /responses;
+                        # falling back to /chat/completions cannot succeed and would
+                        # hide the real error.
                         raise
                     if self._responses_is_required():
                         raise
@@ -1954,8 +2018,10 @@ class OpenAICompatProvider(LLMProvider):
                         )
                     return result
                 except Exception as responses_error:
-                    capabilities = self._responses_capabilities()
-                    if capabilities is not None and not capabilities.allows_chat_fallback:
+                    if self._spec and self._spec.name == "github_copilot":
+                        # Copilot gateway exposes GPT-5/o-series only via /responses;
+                        # falling back to /chat/completions cannot succeed and would
+                        # hide the real error.
                         raise
                     if self._responses_is_required():
                         raise

nanobot/providers/registry.py

@@ -13,7 +13,7 @@ Every entry writes out all fields so you can copy-paste as a template.
 from __future__ import annotations
 
 from dataclasses import dataclass
-from typing import Any, Literal
+from typing import Any
 
 from pydantic.alias_generators import to_snake
 
@@ -28,32 +28,6 @@ class ProviderModelSpec:
     context_window: int | None = None
 
 
-@dataclass(frozen=True)
-class ResponsesCapabilities:
-    """Provider capabilities for the shared OpenAI Responses execution path.
-
-    ``reasoning_replay`` selects whether multi-turn reasoning is retained as
-    encrypted server content, plaintext local history, or not requested.
-    """
-
-    models: tuple[str, ...] = ()
-    auto_route: bool = False
-    requires_direct_openai_base: bool = False
-    allows_api_type_override: bool = False
-    reasoning_replay: Literal["none", "encrypted", "plaintext"] = "none"
-    supports_native_compaction: bool = False
-    allows_chat_fallback: bool = True
-
-    def matches_model(self, model: str) -> bool:
-        """Return whether *model* is explicitly routed through Responses."""
-        model_name = model.lower()
-        return any(
-            model_name == supported.lower()
-            or model_name.endswith(f"/{supported.lower()}")
-            for supported in self.models
-        )
-
-
 @dataclass(frozen=True)
 class ProviderSpec:
     """One LLM provider's metadata. See PROVIDERS below for real examples.
@@ -137,8 +111,9 @@ class ProviderSpec:
     # Substring match against the wire model name (lowercased).
     implicit_reasoning_models: tuple[str, ...] = ()
 
-    # Capabilities for providers/models served through the shared Responses path.
-    responses: ResponsesCapabilities | None = None
+    # Models that expose the OpenAI Responses wire format.  This is model-level
+    # because providers may add Responses support incrementally.
+    responses_models: tuple[str, ...] = ()
 
     # Provider-hosted Responses tools sent unless extraBody.tools explicitly
     # supplies the hosted-tool selection. Values are raw Responses tool types.
@@ -224,6 +199,18 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         supports_prompt_caching=True,
         gateway_reasoning_style="reasoning_effort",
     ),
+    # OrcaRouter: global gateway, keys start with "sk-orca-"
+    ProviderSpec(
+        name="orcarouter",
+        keywords=("orcarouter",),
+        env_key="ORCAROUTER_API_KEY",
+        display_name="OrcaRouter",
+        backend="openai_compat",
+        is_gateway=True,
+        detect_by_key_prefix="sk-orca-",
+        detect_by_base_keyword="orcarouter",
+        default_api_base="https://api.orcarouter.ai/v1",
+    ),
     # Eden AI: OpenAI-compatible gateway. Models use the "provider/model"
     # naming scheme (e.g. "anthropic/claude-sonnet-4-5"); the full id is sent upstream.
     ProviderSpec(
@@ -413,13 +400,6 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         display_name="OpenAI",
         backend="openai_compat",
         supports_max_completion_tokens=True,
-        responses=ResponsesCapabilities(
-            auto_route=True,
-            requires_direct_openai_base=True,
-            allows_api_type_override=True,
-            reasoning_replay="encrypted",
-            supports_native_compaction=True,
-        ),
     ),
     # OpenAI Codex: OAuth-based, dedicated provider
     ProviderSpec(
@@ -503,11 +483,6 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         strip_model_prefix=True,
         is_oauth=True,
         supports_max_completion_tokens=True,
-        responses=ResponsesCapabilities(
-            auto_route=True,
-            reasoning_replay="encrypted",
-            allows_chat_fallback=False,
-        ),
     ),
     # DeepSeek: OpenAI-compatible at api.deepseek.com
     ProviderSpec(
@@ -518,10 +493,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
         backend="openai_compat",
         default_api_base="https://api.deepseek.com",
         thinking_style="thinking_type",
-        responses=ResponsesCapabilities(
-            models=("deepseek-v4-flash",),
-            reasoning_replay="plaintext",
-        ),
+        responses_models=("deepseek-v4-flash", "deepseek-v4-pro"),
         responses_default_tools=("web_search",),
     ),
     # Gemini: Google's OpenAI-compatible endpoint

nanobot/sdk/clients.py

@@ -138,6 +138,7 @@ class SessionClient:
 
     def clear(self, session_key: str) -> SessionSnapshot:
         """Clear one session and persist the empty session."""
+        self._loop.discard_session_file_state(session_key)
         session = self._loop.sessions.get_or_create(session_key)
         session.clear()
         self._loop.sessions.save(session)