The runtime media-attachment mechanism was broken for streaming channels
(e.g. WebSocket): the _streamed flag caused _send_once to skip the final
OutboundMessage that carried generated media, so images were never delivered.
Rather than adding complex coordination between streaming and media delivery,
delegate image delivery to the LLM: after generate_image returns artifact
paths, the next_step prompt now instructs the LLM to call the message tool
with the paths in the media parameter. This works uniformly across all
channels, streaming or not.
Remove generated_media from TurnContext, _assemble_outbound, and _state_save.
Update prompts in identity.md, SKILL.md, message tool description, and
artifacts.py to reflect the new flow.
* feat(long-task): add LongTaskTool for multi-step agent tasks
Implements a meta-ReAct loop where long-running tasks are broken into
sequential subagent steps, each starting fresh with the original goal
and progress from the previous step. This prevents context drift when
agents work on complex, multi-step tasks.
- Extract build_tool_registry() from SubagentManager for reuse
- Add run_step() for synchronous subagent execution (no bus announcement)
- Add HandoffTool and CompleteTool as signal mechanisms via shared dict
- Add LongTaskTool orchestrator with simplified prompt (8 iterations/step)
- Register LongTaskTool in main agent loop
- Add _extract_handoff_from_messages fallback for robustness
* fix(long-task): add debug logging for step-level observability
* feat(long-task): major overhaul with structured handoffs, validation, and observability
- Structured HandoffState: HandoffTool now accepts files_created,
files_modified, next_step_hint, and verification fields instead of
a plain string. Progress is passed between steps as structured data.
- Completion validation round: After complete() is called, a dedicated
validator step runs to verify the claim against the original goal.
If validation fails, the task continues rather than returning
a false completion.
- Dynamic prompt system: 3 Jinja2 templates (step_start, step_middle,
step_final) selected based on step number. Final steps get tighter
budget and stronger "wrap up" guidance.
- Automatic file change tracking: Extracts write_file/edit_file events
from tool_events and injects them into the next step's context if
the subagent forgot to report them explicitly.
- Budget tracking & adaptive strategy: Cumulative token usage is tracked
across steps. Per-step tool budget drops from 8 to 4 in the last
two steps to force handoff/completion.
- Crash retry with graceful degradation: A step that crashes is retried
once. Persistent crashes terminate the task and return partial progress.
- Full observability hooks for future WebUI integration:
- set_hooks() with on_step_start, on_step_complete, on_handoff,
on_validation_started, on_validation_passed, on_validation_failed,
on_task_complete, on_task_error, and catch-all on_event.
- Readable state properties: current_step, total_steps, status,
last_handoff, cumulative_usage, goal.
- inject_correction() allows external code to send user corrections
that are injected into the next step's prompt.
- run_step() accepts optional max_iterations for dynamic budget control.
All 27 long-task tests and 11 subagent tests pass.
* test(long-task): add boundary tests and fix race conditions
- Add 7 edge-case tests: validation crash resilience, hook exception safety, mid-run correction injection, FIFO correction ordering, explicit file changes overriding auto-detection, final budget for max_steps=1, and dynamic budget switching boundaries
- Fix assertion in test_long_task_completes_after_multiple_handoffs to match exact prompt format
- Remove asyncio timing hack from test_state_exposure
- Add asyncio.sleep(0) yield in test_inject_correction_during_execution to prevent race between signal injection and step continuation
- All 34 tests passing
* fix(long-task): address code review findings
- Declare _scopes = {"core"} explicitly to prevent recursive nesting in subagent scope
- Document fragile coupling in _extract_file_changes: path extraction depends on
write_file/edit_file detail format; add debug log for unexpected formats
- Align final-template threshold (max_steps - 2) with budget switch threshold
- Eliminate hasattr(self, "_state") in _reset_state by initializing in __init__
* fix(long-task): honor final signal and file tracking
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(long-task): improve prompt structure and agent contract
- Expand LongTaskTool.description to instruct parent agent on goal
construction, return value semantics, and how to handle results.
- Expand CompleteTool.description to emphasize that the summary IS the
final answer returned to the parent agent.
- Prefix validated return value with an explicit "final answer" directive
to stop parent agent from re-running work.
- Redesign step_start.md: Step 1 is now explicitly for exploration,
planning, and skeleton-building. complete() is discouraged.
- Remove bulky payload debug logging from _emit(); add targeted
info/warning/error logs at key state transitions instead.
- Add signal_type to HandoffState for cleaner signal detection.
* test(long-task): expect wrapped completion message after validation
Align assertions with LongTaskTool final return shape on main.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(webui): turn timing strip, latency, and session-switch restore
- Agent loop: publish goal_status run/idle for WebSocket turns; attach
wall-clock latency_ms on turn_end and persisted assistant metadata.
- WebSocket channel: forward goal_status and latency fields to clients.
- NanobotClient: track goal_status started_at per chat without requiring
onChat; useNanobotStream restores run strip when returning to a chat.
- Thread UI: composer/shell viewport hooks for run duration and latency;
format helpers and i18n strings.
- MessageBubble: drop trailing StreamCursor (layout artifact vs block markdown).
- Builtin / tests: model command coverage, websocket and loop tests.
Covers multi-session UX and round-trip timing visibility for the WebUI.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: keep message-tool file attachments after canonical history hydrate
- MessageTool records per-turn media paths delivered to the active chat.
- nanobot.utils.session_attachments stages out-of-media-root files and
merges into the last assistant message before save (loop stays a thin call).
- WebUI MediaCell: use a signed URL as a real download link when present.
Fixes attachments flashing then vanishing on turn_end when paths lived
outside get_media_dir (e.g. workspace files).
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(webui): agent activity cluster, stable keys, LTR sheen labels
- Group reasoning and tool traces in AgentActivityCluster with i18n summaries
- Stabilize React list keys for activity clusters (first message id anchor)
- Replace background-clip shimmer with overlay sheen for streaming labels
- ThreadMessages/MessageList integration and locale strings
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): render assistant reasoning with Markdown + deferred stream
- Use MarkdownText for ReasoningBubble body (same GFM/KaTeX path as replies)
- Apply muted/italic prose tokens so thinking stays visually subordinate
- useDeferredValue while reasoningStreaming to ease parser work during deltas
- Preload markdown chunk when trace opens; add regression test with preloaded renderer
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): default-collapse agent activity cluster while Working
Outer fold no longer auto-expands during isTurnStreaming; user opens to see traces.
Header sheen and live summary unchanged.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(long_task): cumulative run history, file union, and prompt tuning
Inject cross-step summaries and merged file paths into middle/final step
templates so chains do not lose early context. Strip the last run-history
block when it duplicates Previous Progress to save tokens. Add optional
cumulative_prompt_max_chars and cumulative_step_body_max_chars parameters
with clamped defaults.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): session switch keeps in-flight thread and replays buffered WS
Save the prior chat message list to the per-chat cache in a layout effect
when chatId changes (before stale writes could corrupt another chat).
Skip one post-switch layout cache tick so we do not snapshot the wrong tab.
Buffer inbound events per chat_id when no onChat subscriber is registered
(e.g. user focused another session) and drain on resubscribe up to a cap,
so streaming deltas are not lost while off-tab.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): snap thread scroll to bottom on session open (no smooth glide)
Use scroll-behavior auto on the viewport, instant programmatic scroll when
following new messages and on scrollToBottomSignal. Keep smooth only for
the explicit scroll-to-bottom button.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): respect manual scroll-up after opening a session
Track when the user leaves the bottom with a ref and skip ResizeObserver
and deferred bottom snaps until they return or the conversation is reset.
Remove the time-based force-bottom window that overrode atBottom.
Multi-frame scrollToBottom honours the same guard unless force (scroll button).
Co-authored-by: Cursor <cursoragent@cursor.com>
* Publish long_task UI snapshots on outbound metadata
- Add OUTBOUND_META_AGENT_UI (_agent_ui) for channel-agnostic structured state
- LongTaskTool publishes {kind: long_task, data: snapshot} on the bus with _progress
- WebSocket send forwards metadata as agent_ui for WebUI clients
- Tests for bus payload, WS frame, and progress assertions
- Fix loop progress tests: ignore _goal_status in streaming final filter and
avoid brittle outbound[-1] ordering after goal status idle messages
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat: WebUI long_task activity card and resilient history merge
Add optional ui_summary to the long_task tool for one-line UI labels. Stream
long_task agent_ui into a dedicated message row with timeline, markdown peek,
and a right sheet for details. Merge canonical history after turn_end while
re-inserting long_task rows before the final assistant reply. Collapse
duplicate task_start/step_start steps in the timeline and extend i18n.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor: align long_task with thread_goal and drop orchestrator UI
- Persist sustained objectives via session metadata (long_task / complete_goal); no subagent wiring or tool-driven agent_ui payloads.\n- Remove WebUI long-task activity UI, types, and translations; history merge preserves trace replay only, with legacy long_task rows normalized to traces.\n- Drop long_task prompt templates and get_long_task_run_dir; add webui thread disk helper for gateway persistence tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(agent): thread goal runtime context, tools, and skill
- Add thread_goal_state helper and mirror active objectives into Runtime Context
- Wire loop/context/memory/events as needed for goal metadata in turns
- Expand long_task / complete_goal semantics (pivot/cancel/honest recap)
- Add always-on thread-goal SKILL.md; align /goal command prompt
- Tests for context builder and thread goal state
- Remove unused webui ChatPane component
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(thread-goal): add websocket snapshot helper and publish goal updates from long_task
Introduce thread_goal_ws_blob for bounded JSON snapshots, attach snapshots to
websocket turn_end metadata in AgentLoop, and let long_task fan-out dedicated
thread_goal frames on the websocket channel after persisting session metadata.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(channels): websocket thread_goal frames, turn_end replay, and session API scrub for subagent inject
Emit thread_goal events and optional thread_goal on turn_end; scrub persisted
subagent announce blobs on GET /api/sessions/.../messages and shorten session
list previews so WebUI does not surface full Task/Summarize scaffolding.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(webui): merge ephemeral traces per user turn when reconciling canonical history
Preserve disk/live trace rows inside the matching user–assistant segment instead
of stacking every trace before the final assistant reply (fixes inflated tool
counts after refresh or session switch).
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(webui): show assistant reply copy only on the last slice before the next user turn
Avoid duplicate copy affordances on intermediate assistant bubbles that precede
more agent activity in the same turn (tools or further assistant text).
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(webui): thread_goal stream plumbing, composer goal strip, sky glow, and client-side subagent scrub projection
Track thread_goal and turn_goal snapshots in NanobotClient, hydrate React state
from thread_goal frames and turn_end, surface objective/elapsed in the composer,
add breathing sky halo CSS while goals are active, mirror server scrub logic on
history hydration and webui_thread snapshots, and extend tests/client mocks.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(channels): add Slack Socket Mode connect timeout with actionable timeout errors
Abort hung websockets.connect handshakes after a bounded wait, log REST-vs-WSS
guidance, surface RuntimeError to channel startup, and log successful WSS setup.
Co-authored-by: Cursor <cursoragent@cursor.com>
* webui: expand thread goal in composer bottom sheet
Add ChevronUp control on the run/goal strip that opens a bottom Sheet
with full ui_summary and objective. Inline preview logic in RunElapsedStrip,
add i18n strings across locales, and a composer unit test.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(webui): widen dedupeToolCallsForUi input for session API typing
fetchSessionMessages types tool_calls as unknown; accept unknown so tsc
build passes when passing message.tool_calls through.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(agent): extract WebSocket turn run status to webui_turn_helpers
* refactor(skills): rename thread-goal to long-task and document idempotent goals
* feat(skills): rename sustained-goal skill to long-goal and tighten long_task guidance
* chore: remove unused subagent/context/router helpers
* feat(session): rename sustained goal to goal_state and align WS/WebUI
- Move helpers from agent/thread_goal_state to session/goal_state:
GOAL_STATE_KEY, goal_state_runtime_lines, goal_state_ws_blob, parse_goal_state.
- Session metadata now uses "goal_state"; still read legacy "thread_goal";
long_task writes drop the legacy key after save.
- WebSocket: event/field goal_state, _goal_state_sync; turn_end carries goal_state;
accept legacy _thread_goal_sync/thread_goal inbound metadata for dispatch.
- WebUI: GoalStateWsPayload, goalState hook/client props, i18n keys goalState*.
- Runtime Context copy uses "Goal (active):" instead of "Thread goal".
* feat(agent): stream Anthropic thinking deltas and fix stream idle timeout
* refactor(webui): transcript jsonl as sole timeline source
* fix(agent): reject mismatched WS message chat_id and stream reasoning deltas
* feat(webui): hydrate sustained goal and run timer after websocket subscribe
* chore(webui,websocket): remove unused fetch helpers and legacy thread_goal WS paths
* Raise default max_tokens and context window in agent schema.
Align AgentDefaults and ModelPresetConfig with typical Claude-scale usage
(32k completion budget, 256k context window) and update migration tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(gateway): bootstrap prefers in-memory model; clarify websocket naming
* fix(websocket): websocket _handle_message passes is_dm; refresh /status test expectations
---------
Co-authored-by: chengyongru <2755839590@qq.com>
Co-authored-by: chengyongru <chengyongru.ai@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Reasoning surfacing was split across three branches in runner.py plus
two separate streaming buffers (loop hook and runner progress stream),
with three independent display-side gates in the CLI. This collapsed
the policy into one source of truth and fixed two real bugs:
- Structured `reasoning_content` was suppressed whenever the answer was
streamed, because the runner gated emission on `streamed_content`.
Providers don't stream `reasoning_content`; it only arrives on the
final response, so the answer stream and the reasoning channel are
independent. Added `streamed_reasoning` to `AgentHookContext` to track
the right bit.
- `channels.showReasoning` was subordinated to `sendProgress`. They are
orthogonal — turning off progress streaming shouldn't silence
reasoning. Reworked the CLI gates accordingly.
Single-helper consolidation:
- `extract_reasoning(reasoning_content, thinking_blocks, content)`
returns `(reasoning_text, cleaned_content)` with a defined fallback
order: dedicated field → Anthropic thinking_blocks → inline
`<think>`/`<thought>` tags. Models that expose none of these
short-circuit to `(None, content)` — zero overhead.
- `IncrementalThinkExtractor` replaces the ad-hoc `emit_incremental_think`
function and its hand-rolled "emitted cursor" state in both the loop
hook and the runner progress stream.
Also documented the new `showReasoning` channel option in
docs/configuration.md and noted its independence from sendProgress.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add extract_think() and emit_incremental_think() helpers to extract thinking content from inline <think> and <thought> tags in the content field. This handles models served via Ollama, self-hosted vLLM, or other compatible endpoints that embed reasoning as inline tags instead of using the dedicated reasoning_content API field.
Also adds Anthropic thinking_blocks support for extended thinking via the thinking content blocks array.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
Replaces PR #3493's blanket fatal abort with a "tell the model + throttle
the bypass loop" policy. Workspace-bound rejections are now ordinary
recoverable tool errors enriched with a structured "this is a hard policy
boundary" instruction; SSRF stays the only marker that aborts the turn.
Why the fatal-abort approach broke
----------------------------------
PR #3493 promoted every shell `_guard_command` and filesystem path-resolution
rejection to a turn-fatal RuntimeError. Two of those messages (`path
outside working dir` and `path traversal detected`) are heuristic substring
scans on the raw command, so legitimate commands like `rm <ws>/x.txt
2>/dev/null` or `find . -type f` killed the user's turn (#3599). On
channels with outbound dedupe (Telegram) the user just saw silence (#3605),
and the noise polluted the LLM's context until it started hallucinating
guard rejections on plain relative paths (#3597).
Why we still need *some* throttle
---------------------------------
The original #3493 pain point was real: the LLM, refused once, would
swap tools and try again -- read_file -> exec cat -> exec cp -> bash -c
-> ln -sf -> python -c open(...). Just removing the fatal escape lets
that loop run wild until max_iterations.
What this commit does
---------------------
- `nanobot/utils/runtime.py`: add `workspace_violation_signature` and
`repeated_workspace_violation_error`. The signature normalizes
filesystem `path` arguments and the first absolute path inside an
exec command, so swapping tools against the same outside target hits
the same throttle bucket. Two soft attempts are allowed; the third
attempt's tool result is replaced with a hard "stop trying to bypass"
message that quotes the target path and tells the model to ask the
user for help.
- `nanobot/agent/runner.py`: split classification into `_is_ssrf_violation`
(still fatal) and `_is_workspace_violation` (now soft). All three
failure branches in `_run_tool` (prep_error / exception / Error
result) route through a shared `_classify_violation` that bumps the
per-turn workspace_violation_counts dict and either keeps the tool's
own message or substitutes the throttle escalation. `_execute_tools`
now threads that dict alongside the existing external_lookup_counts.
- `nanobot/agent/tools/shell.py`: append a structured boundary note to
every workspace-bound guard rejection (`working_dir could not be
resolved`, `working_dir is outside`, `path outside working dir`,
`path traversal detected`). SSRF errors stay short and direct so the
model doesn't try to "phrase around" them. Existing `2>/dev/null`
allow-list and benign device passthrough from the previous commit
remain.
- `nanobot/agent/tools/filesystem.py`: append the same boundary note to
the `outside allowed directory` PermissionError so read_file / write_file
/ list_dir errors give the LLM the same explicit hint.
Tests
-----
- `tests/utils/test_workspace_violation_throttle.py` (new): signature
collapses across read_file/exec/python -c against the same path,
different paths get independent budgets, escalation only fires after
the third attempt.
- `tests/agent/test_runner.py`:
- `test_runner_does_not_abort_on_workspace_violation_anymore` -- v2
contract: filesystem PermissionError is now soft, runner moves to
the next iteration and finalizes cleanly.
- `test_is_ssrf_violation_remains_fatal` + the existing
`test_runner_aborts_on_ssrf_violation` -- SSRF still aborts on the
first attempt.
- `test_runner_lets_llm_recover_from_shell_guard_path_outside` -- end
to end recovery from `path outside working dir`.
- `test_runner_throttles_repeated_workspace_bypass_attempts` -- four
bypass attempts against the same outside target produce at least
one `workspace_violation_escalated` event and the run completes
naturally without aborting the turn.
- The two `_execute_tools` direct-call tests now pass the new
workspace_violation_counts dict.
- `tests/tools/test_tool_validation.py`: relax three `==` assertions
to `startswith` + "hard policy boundary" substring check to match
the new structured error messages.
- `tests/tools/test_exec_security.py` keeps the prior `2>/dev/null`
regression and the `> /etc/issue` negative case from the previous
commit on this branch -- they still pass under the new policy.
Coverage status: full pytest 2648 passed / 2 skipped (was 2638 / 2
on origin/main). Ruff is clean for every file touched in this commit.
Co-authored-by: Cursor <cursoragent@cursor.com>
cmd_restart only persisted channel + chat_id across the os.execv boundary, so
when the new process announced "Restart completed" the OutboundMessage had
no Slack thread_ts and the reply fell back to the channel root.
Serialize msg.metadata into NANOBOT_RESTART_NOTIFY_METADATA, restore it on the
RestartNotice, and forward it to OutboundMessage so the completion message
follows the same routing as the original /restart invocation.
Made-with: Cursor
Some models / Ollama renderers occasionally emit tokenizer-level template
leaks that the existing regexes miss:
1. Malformed opening tags with no closing `>`, running straight into
user-facing content — e.g. `<think广场照明灯目前…` (observed with
Gemma 4 via Ollama). The earlier `<think>[\s\S]*?</think>` and
`^\s*<think>[\s\S]*$` patterns both require `>`, so these leak into
rendered messages.
2. Harmony-style channel markers like `<channel|>` / `<|channel|>` at
the start of a response.
3. Orphan `</think>` / `</thought>` closing tags left behind when only
the opener was consumed upstream.
Handles each case conservatively:
- Malformed `<think` / `<thought` only match when the next char is NOT
a tag-name continuation (`[A-Za-z0-9_\-:>/]`). Explicit ASCII class
instead of `\w` because Python's Unicode `\w` matches CJK and would
defeat the primary fix.
- Orphan closing tags and channel markers are stripped **only at the
start or end of the text**. `strip_think` is also applied before
persisting history (memory.py), so mid-text stripping would silently
rewrite transcripts where the tokens themselves are discussed.
Preserves: `<thinker>`, `<think-foo>`, `<think_foo>`, `<think1>`,
`<think:foo>`, `<thought/>`, literal `` `</think>` `` / `` `<channel|>` ``
inside prose or code blocks.
Adds 16 new regression tests covering both the leak cases and the
preserved-prose cases.
Treat `.git` files the same as `.git` directories so GitStore refuses to initialize inside git worktrees, and add a focused regression test for that checkout shape.
Made-with: Cursor
GitStore.init() now checks if the workspace is already inside a git
repository before calling porcelain.init(). If so, it refuses to create
a nested repo. Additionally, existing .gitignore files are preserved
by appending only missing Dream-specific entries rather than overwriting.
Closes#2980
Three improvements to Dream's memory consolidation:
1. Per-line git-blame age annotations: MEMORY.md lines get `← Nd` suffixes
(N>14) from dulwich annotate. SOUL.md/USER.md excluded as permanent.
LLM uses content judgment, not just age, to decide what to prune.
2. Dedup-aware Phase 1 prompt: reframed as dual-task (extract facts +
deduplicate existing files) with explicit redundancy patterns to scan for.
Validated through 20 experiments (exp-002 prompt + max_iter=15 was best,
averaging -1643 chars/5.4% compression per run).
3. Phase 1 analysis as commit body: dream git commits now include the full
Phase 1 analysis for transparency via /dream-log.
4. max_iterations raised from 10 to 15: 30% improvement over 10 with no
risk; 20 showed diminishing returns (exp-020: -701 vs exp-017: -1643).
The _parse_tavily_usage implementation was updated to use the real
{account: {plan_usage, plan_limit, ...}} structure, but the tests
still used the old flat {used, limit, breakdown} format.
Made-with: Cursor
