Small local models (qwen2.5-vl-3b, gemma-4-e4b) narrate actions ('Sure,
I'll pause the music') without emitting the tool call once the full
system prompt and 26-tool set are in context. Measured: qwen emitted
tools in 2/8 realistic intents, gemma 5/8; single-tool tests pass 100%,
so this is context-induced drift, not plumbing.
- agent/intent_router.py: deterministic forced-call router for
high-frequency intents — media controls, now_playing/list_library,
'remember that ...' memory saves, explicit YouTube Music searches,
and personal-fact recall (memory_recall/search_notes/skill_recall,
excluded for weather/time questions). Fires only when the tool is
available; guards against how-do-I questions.
- agent.py: generalize _ForcedToolCallStream to any tool; route through
intent_router after the dispatch check. System prompt adds a hard
'call, don't narrate' rule and reconditions 'answer directly'.
- tests/test_intent_router.py: 30 unit tests pinning router behavior.
Verified against the live LLM with the full fidelity harness: combined
forced-or-emitted coverage 8/8 on qwen2.5-vl-3b (was 2/8) and 8/8 on
gemma-4-e4b (was 5/8). Full suite: 79 passed.
- yt-dlp stderr now goes to a temp file so a chatty stderr can never
deadlock the stdout progress loop.
- report/search/seek numeric parsing returns 400 instead of 500.
- now-playing card reflects failed/preparing download states.
- media runtime data (library.json, jobs/) is gitignored; only
.gitkeep is tracked.
Document the media server process, /media/ port map, library layout on
the ./media bind mount, Hope's media MCP toolset, the music/mic
mutual-exclusion rule, and yt-dlp download/proxy conventions.
- web/media_server.py: FastAPI service owning the permanent library
(library.json + ./media volume), yt-dlp YouTube Music search, queued
download jobs with mp3 transcode (ffmpeg), per-room player desired
state, and an HTTPS proxy for external live-stream URLs.
- agent/media_mcp.py: Hope tools — search_music, play_song,
offer_song_choices, pause/resume/stop/skip/seek, now_playing,
list_library, add_stream.
- agent/agent.py: media toolset wiring + conversation rule (music
playing disables the agent mic; re-engaging pauses music).
- web UI: slide-out Music panel (search, suggestions, library, live
stream add, play/pause, ±10s, seek, volume) with a reconciling poller
and media_state data-channel reports.
- token server returns the minted room name; nginx serves /media/files
as static (native Range) and proxies /media/api to the media server.
- supervisord runs media-server; compose mounts ./media and adds
ffmpeg to the image.
- Default LLM_MODEL to qwen2.5-vl-3b across agent, task worker, tests,
compose, workflow, and docs
- Cap min/max length-style constraints in tool schemas to 1000 before
LLM requests; llama.cpp rejects grammar repetitions >= 2000 and the
memory MCP's update_note body maxLength (200000) broke qwen2.5-vl-3b
with 'Failed to initialize samplers: failed to parse grammar'
- Move task_dir assignment before VoiceAssistant construction in
handle_job; UnboundLocalError was crashing every agent job before it
joined the room, leaving the UI stuck on 'Opening a secure room'
- Make settings sheet full-window height (100dvh) to match tasks panel
- Point the agent at OpenWebUI's OpenAI-compatible API (LLM_BASE_URL,
LLM_API_KEY, LLM_MODEL) with GEMMA_* env fallbacks
- Add list_available_models() fetching {base_url}/models, filtering
non-chat ids (embedding/reranker)
- Rename GemmaLLM -> HopeLLM with update_options(model=...) for runtime
model switching (mutates _opts.model) and instance-based compaction
- Data channel: agent publishes {type: models} after connect and on
get_models; set_model switches live and acks model_changed
- Web UI: Brain section in the settings sheet with a model picker that
persists the choice in localStorage
- Update tests, CI workflow env, and docs for the new endpoint
- Updated Memory section to always use session_id='hope' for remember/recall
- This makes memory operations instant (0.3s) instead of 20+ seconds
- Fixed tool_call logging bug (FunctionToolCall has .name, not .function.name)
Reduces tool count from 10 to 8 for the LLM, removing the conflict
between 'use dispatch for everything' and having web_search available.
The task_worker has its own web tools for background tasks.
- Remove memory_mcp.py (stdio markdown-file memory)
- Add Cognee MCP as remote Streamable HTTP toolset at 192.168.86.2:8003/mcp
- Filter to only remember/recall/forget tools via allowed_tools
- Update system prompt: Memory section moved to top priority with
explicit 'call recall FIRST' instructions and examples
- Add COGNEE_MCP_URL env var to docker-compose
- Remove /memory volume mount (no longer needed)
- Rewrite memory tests to use Cognee HTTP client fixture
- 28/28 tests passing
- token_server: generate voice-{uuid8} room name per request so each
browser session creates a fresh room and triggers agent dispatch
- token claim: add jobType JT_ROOM to roomConfig.agents
- livekit.yaml: revert empty_timeout to default (300s)
- agent.py: system prompt now mandates dispatch_task for all non-trivial
tasks (not just research); removed inline web_search instructions;
added tool_call logging in _ReasoningFallbackWrapper
The FastMCP server never started listening on stdio because the file was
missing 'if __name__ == "__main__": mcp.run(transport="stdio")'. The
LiveKit agent's MCP client got 'Connection closed' during initialize,
which killed the entire toolset setup — ALL tools (weather, memory,
skills, dispatch) were unavailable. This is why Hope could hear you but
never called any tools.
Set room.empty_timeout to 86400 (24h) in livekit.yaml. Previously the
default 300s timeout destroyed the room when all participants left,
which killed the agent job and left no worker registered for the next
join — causing 'assistant not ready retry' on tab reopen.
In the container, agent.py is at /opt/voice-agent/agent.py (flat, not a
package). The 'from agent.task_registry import ...' raised ModuleNotFoundError
which killed the job before session.start(), causing 'assistant not ready
retry' in the browser. Added try/except ImportError fallback to bare import.
- agent/task_registry.py: file-based JSONL event registry (cross-process)
- agent/task_worker.py: autonomous LLM loop with weather/time/memory/web tools
- agent/dispatch_mcp.py: MCP tool exposing dispatch_task to the main agent
- agent/agent.py: registers dispatch toolset, polls task events → room data
- web: slide-out task panel (FAB button + badge), live step streaming via
data channel topic 'tasks', status dots (running/completed/failed)
- Dockerfile: copies new task_*.py and dispatch_mcp.py files
The dispatch MCP runs in its own process; events flow through
/tmp/tasks/events.jsonl which the main agent tails every second and
forwards to the browser. Tasks run up to 10 LLM iterations with tool calls.
The 4B model was responding conversationally instead of calling memory_save/
memory_recall. Added imperative language (MUST call), concrete examples of
trigger phrases, and explicit instructions to never skip the tool call.
Verified: model now reliably generates tool_calls for save/recall/list.
When conversation exceeds 24 items (~12 turns), Gemma summarizes the older
messages into a 2-4 sentence recap that replaces them, keeping the last 10
items verbatim. Also triggers after 5 minutes of idle time so returning
users get a compacted context rather than a bloated one.
Falls back to hard truncation if the summary call fails.
Compaction:
- GemmaLLM.chat() truncates ChatContext to last 30 items (~15 turns)
before sending to LLM, preventing context window overflow on long
conversations. Preserves system prompt and removes orphaned tool calls.
Skills:
- agent/skills_mcp.py: MCP server with skill_save, skill_recall,
skill_list, skill_update tools backed by .md files in /skills
- skills/ dir bind-mounted into container, git-trackable
- System prompt instructs Hope to save repeatable procedures as skills
and recall them before performing tasks she's done before
- Distinct from memory (facts) — skills are learned *procedures*
- agent/memory_mcp.py: MCP server with memory_recall, memory_save, memory_list
tools backed by .md files in /memory (simple keyword matching for v1)
- memory/ dir bind-mounted into container, persists across rebuilds,
easily backed up via git
- System prompt instructs Hope to recall on past references and save
personal info/preferences naturally without announcing it
- Dockerfile: copy memory_mcp.py; compose: ./memory:/memory volume
- preemptive_generation was incompatible with MCP tool calls: it starts
generating before the turn finalizes, breaking the tool execution loop
(agent said 'let me check' then hung forever)
- Add get_time(location?) to weather_mcp.py for date/time queries
- Update system prompt with Weather & Time section
- lk.transcription handler now uses addPartial() (in-place update)
instead of addMessage() (new row per word)
- addMessage() clears stale partials before rendering the final
- New agent/weather_mcp.py: stdio MCP server with get_weather(location)
that returns a short conversational summary (temp, conditions, high/low)
- Register weather toolset in build_mcp_toolsets() alongside web-access
- Update system prompt with Weather section
- agent/web_mcp.py: stdio MCP server exposing web_search and web_scrape,
backed by the self-hosted Firecrawl stack on xNAS (no API key needed)
- agent.py: Agent now attaches mcp_servers built from config; EXTRA_MCP_SERVERS
env var allows adding arbitrary HTTP/SSE MCP servers as JSON
- Dockerfile: installs livekit-agents[mcp], copies web_mcp.py
- .env.example: WEB_MCP_ENABLED, FIRECRAWL_BASE, EXTRA_MCP_SERVERS documented
- nginx serves the UI over HTTPS on 8090 with a self-signed cert
(browsers require a secure context for microphone access)
- added /token endpoint (tiny Python HTTP server) that signs LiveKit
JWTs server-side, keeping the API secret out of the browser
- app.js now fetches a signed token from /token and uses wss:// when
the page is served over HTTPS
- supervisord runs the token-server as a fourth process
snowspeeder
Shane