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.
Updates docs/image-generation.md and skills/image-generation/SKILL.md to
include MiniMax configuration examples, supported aspect ratios, and
troubleshooting references. Also updates the supported provider list to
include minimax alongside openrouter, aihubmix, and gemini.
Adds ImageGenerationProvider ABC with shared __init__, _http_post(), and
_require_images(). Introduces _IMAGE_GEN_PROVIDERS registry with
register/get/image_gen_provider_configs() helpers.
Four existing providers (OpenRouter, AIHubMix, Gemini, MiniMax) now inherit
from the base class and self-register. Adding a new provider only requires
writing one class + one registration line.
Eliminates if/else chains in the tool dispatch and hardcoded provider config
dicts in commands.py (3 sites) and nanobot.py (1 site). Fixes the agent CLI
command missing image_generation_provider_configs entirely.
Also simplifies test monkeypatch targets to patch the registry lookup.
Adds GeminiImageGenerationClient covering both Imagen 4 (:predict) and
Gemini Flash (:generateContent), wires the gemini ProviderConfig through
the SDK, API server, and gateway entry points, and updates the
image-generation docs and skill. Errors from the Gemini endpoints are
logged and surface with the HTTP status and parsed message instead of an
empty string.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add MiniMaxImageGenerationClient with support for:
- Text-to-image generation via MiniMax image-01 model
- Reference image support (subject_reference)
- Aspect ratio selection
- Proper error handling aligned with existing providers
Wire up MiniMax provider config in ImageGenerationTool, gateway,
serve, and Nanobot class.
Extract the [M] Model Presets interactive CRUD screen from PR #3696
and adapt it to the current main branch schema (fallback_models
instead of fallback_presets). Adds preset cache, field handlers for
model_preset/provider/fallback_models, and 9 new tests.
Add remark-breaks plugin so that single newlines in assistant messages
(such as /help output) render as line breaks instead of being collapsed
into a single paragraph by standard markdown behavior.
Replace AutoCompact._archive() direct session mutation with delegation
to Consolidator.compact_idle_session(). Remove _split_unconsolidated()
method since that logic now lives inside compact_idle_session.
All session mutation for idle compaction now goes through the
Consolidator's lock, eliminating the race condition between
background token consolidation and idle TTL compaction.
Changes:
- autocompact.py: rewrite _archive() to call compact_idle_session,
remove _split_unconsolidated(), clean up unused imports
- test_autocompact_unit.py: replace TestArchive/TestSplitUnconsolidated
with TestArchiveDelegates that verifies delegation behavior
- test_auto_compact.py: convert all consolidator.archive mocks to
consolidator.compact_idle_session mocks via _make_fake_compact helper
When background consolidation runs with a stale session reference (captured
before AutoCompact replaced the session via compact_idle_session), it could
operate on outdated data. Now, after acquiring the per-session lock, the
method refreshes its session reference from SessionManager.get_or_create().
If the session was replaced, it swaps in the fresh reference before doing
any consolidation work.
This prevents a race where AutoCompact truncates an idle session while a
background maybe_consolidate_by_tokens call is in flight with the old
session object.
Add Consolidator.compact_idle_session(session_key, max_suffix=8) that
performs hard-truncation of idle sessions under the per-session
consolidation lock. This is the single lock-protected path for AutoCompact
to use instead of modifying session state directly, fixing the race
condition between AutoCompact and Consolidator.
Behavior:
- Acquires per-session consolidation lock
- Invalidates cache and reloads fresh from disk
- Splits unconsolidated tail into archive prefix and retained suffix
- Archives prefix via LLM (with raw_archive fallback on failure)
- Persists _last_summary in session metadata on success
- Returns summary text, None on LLM failure, or '' if nothing to archive
Tests: 6 new tests covering prefix archival, empty session timestamp
refresh, (nothing) summary exclusion, LLM failure fallback,
last_consolidated offset, and lock acquisition verification.
The `docker run` example for `gateway` in `docs/deployment.md` had drifted from
the canonical configuration in `docker-compose.yml`:
- It omitted the security flags that `docker-compose.yml` already declares
(`cap_drop: ALL` + `cap_add: SYS_ADMIN` + unconfined apparmor/seccomp).
These are required whenever `tools.exec.sandbox: "bwrap"` is enabled, because
bwrap needs CAP_SYS_ADMIN for user namespaces; without them bwrap exits with
`clone3: Operation not permitted` and exec tools silently fail.
- It omitted `-p 8765:8765`, even though both the bundled `docker-compose.yml`
and `Dockerfile` (`EXPOSE 18790 8765`) already expose the WebSocket channel
/ WebUI port; users following the docs would get a reachable gateway health
endpoint but an unreachable WebUI.
This change keeps the two paths in sync so anyone reading deployment.md and
using `docker run` directly gets the same security posture and port surface
as the Compose path.
Also adds a short `!IMPORTANT` note documenting that `gateway.host` and
`channels.websocket.host` default to `127.0.0.1` (set in
`nanobot/config/schema.py:GatewayConfig`). Docker `-p` cannot forward to the
container's loopback interface, so the user must set both binds to `0.0.0.0`
in `config.json` for the published ports to actually be reachable. This is
the symptom reported as items 2 + 3 of #3873; items 1 + 4 of that issue are
already resolved on `main` (`Dockerfile` line 49 already exposes both ports,
and README.md lines 218-220 already reflect that the WebUI ships in the wheel).
Docs only, no code changes.
Signed-off-by: voidborne-d <258577966+voidborne-d@users.noreply.github.com>
- Note that any string field supports ${VAR_NAME} and resolved values are
never written back to disk.
- Document the failure mode for unset variables.
- Add MCP (stdio env + HTTP headers) and web-search examples.
- Add Docker, direnv, and secret-manager (1Password / pass / Bitwarden)
delivery patterns alongside the existing systemd example.
- Replace plaintext apiKey values in tools.web.search examples (Brave,
Tavily, Jina, Kagi, Olostep) with ${PROVIDER_API_KEY} placeholders so
the docs stop modelling the anti-pattern.
- Cross-link from the Security section.
Refs: HKUDS/nanobot#2172
Batch stream deltas, window long transcripts, lazy-load syntax highlighting, and refine activity/composer interactions.
Add title refresh retries plus tests for streaming, windowing, code blocks, and live activity behavior.
The xiaomi_mimo ProviderSpec carries thinking_style="thinking_type", but
gateway providers (OpenRouter etc.) route MiMo under their own spec
which has no thinking_style. As a result, `reasoning_effort="none"` was
silently ignored: `{"thinking": {"type": "disabled"}}` was never
injected and responses still contained reasoning_content.
Mirror the Kimi pattern that already handles the same problem: add an
explicit _MIMO_THINKING_MODELS allowlist (mimo-v2.5-pro, mimo-v2.5,
mimo-v2-pro, mimo-v2-omni — per Xiaomi docs), an _is_mimo_thinking_model
helper that strips publisher prefixes ("xiaomi/mimo-v2.5-pro" matches),
and a sibling branch in _build_kwargs that injects the thinking payload
by model name. mimo-v2-flash is intentionally excluded — it has no
thinking mode.
Also include MiMo in the explicit_thinking predicate so the
reasoning_content backfill (#3554, #3584) covers the gateway path
consistently with the direct path.
Tests cover the gateway disable/enable signals, bare-slug fallback,
flash exclusion, and a non-MiMo sanity check.
_drain_pending injected a full runtime context block (including goal
state) into every injected user message, but the initial message already
carries runtime context via build_messages(). This caused goal state to
appear multiple times in the LLM context window within a single turn,
wasting tokens (up to 4000 chars per duplicate).
Now _drain_pending only passes the raw user content without runtime
context. The initial turn message remains the sole carrier.
The previous regex r"(?:^|[;&|]\s*)format\b" incorrectly blocked
commands containing URL parameters like &format=json. Added negative
lookahead (?!=) so format= (URL param key=value) is allowed while
standalone format commands (e.g. ;format, &format, |format) remain
blocked. Added test cases for both blocking and allowing scenarios.
Centralize runner_wall_llm_timeout_s in session goal_state metadata helpers so
spawned subagents inherit the same policy as AgentLoop without coupling to
long_task. Pass optional resolver into SubagentManager and add tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
The Development Setup block instructs new contributors to run
`ruff format nanobot/`, but the tree predates the formatter and many
lines exceed the configured 100-char limit (E501 is ignored). Running
the command as documented produces an ~80-file unrelated diff that
buries real changes. Document this and recommend formatting only the
files actually touched.
