fix: background tasks follow the live model selection

TaskWorker resolved its model from LLM_MODEL env (frozen at MCP spawn),
so after a live set_model switch in the UI, dispatched background tasks
still asked for the old model. On local-vl-models (llama-swap
--models-max 1) that forces a swap against the loaded conversation
model; a force-kill mid-load surfaces as 400 'model failed to load'.

The agent now writes the selected model to <task_dir>/model.txt at
session start and on every switch; TaskWorker reads it with the env
value as fallback.

AGENTS.md

@@ -216,6 +216,7 @@ curl -s "https://eastus.tts.speech.microsoft.com/cognitiveservices/v1" \
 - **Azure TTS uses SSML, not JSON.** The REST endpoint requires `Content-Type: application/ssml+xml`. The LiveKit Azure plugin handles this internally.
 - **Voice changes are live.** The web UI sends a data message to the agent; the agent calls `tts.update_options(voice=...)` without restarting. The agent filters data messages by topic ("voice-control") before processing.
 - **Model changes are live too.** The agent fetches the chat-capable model list (`HopeLLM` filters out embedding/reranker ids) and publishes `{type: "models", models, selected}`; the UI's settings sheet renders a picker and sends `{type: "set_model", model}`, which calls `HopeLLM.update_options(model=...)` (mutates `self._opts.model`) and acks with `{type: "model_changed", model}`. The choice persists in the browser's localStorage.
+- **Background tasks follow the live model.** The dispatch MCP subprocess inherits its env at spawn, so it cannot see later `set_model` switches. The agent writes the selected model to `<task_dir>/model.txt` at session start and on every switch; `TaskWorker` reads that file (falling back to `LLM_MODEL` env) when starting a task. Without this, a switched session (e.g. gemma-4-e4b) dispatches tasks against the stale env model (qwen2.5-vl-3b), which llama-swap on `local-vl-models` (`--models-max 1`) cannot load without killing the conversation model mid-swap.
 - **All supervisord programs log to /dev/stdout** so `docker compose logs -f voice` shows everything. Individual process logs are no longer written to files.
 - **Hope Music stays external.** `hope-media-music` owns yt-dlp, ffmpeg, the
   library volume, streaming endpoints, and the player UI. Do not add those

agent/agent.py

@@ -750,6 +750,20 @@ def build_mcp_toolsets(user_profile: dict, task_dir: str) -> list[mcp.MCPToolset
     return toolsets
 
 
+def _sync_task_model(task_dir: str, model: str) -> None:
+    """Record the live model choice where background task workers can read it.
+
+    The dispatch MCP subprocess inherits its environment at spawn time, so it
+    can never see later UI model switches. The worker instead reads
+    <task_dir>/model.txt (falling back to LLM_MODEL env) when it starts a task.
+    """
+    try:
+        os.makedirs(task_dir, exist_ok=True)
+        with open(os.path.join(task_dir, "model.txt"), "w") as f:
+            f.write((model or "").strip())
+    except OSError as e:
+        logger.warning("failed to record selected model for background tasks: %s", e)
+
 class VoiceAssistant(Agent):
     """The conversational agent. LLM is the brain; STT/TTS are senses."""
 
@@ -824,6 +838,10 @@ async def handle_job(ctx: JobContext) -> None:
     # into it for live model switching.
     voice_assistant = VoiceAssistant(user_profile, task_dir)
 
+    # Background task workers (dispatch MCP subprocess) read this file so
+    # they use the same model as the conversation, even after live switches.
+    _sync_task_model(task_dir, voice_assistant.llm.model)
+
     # Publish the chat-capable model list so the web UI can offer a picker.
     async def _publish_models() -> None:
         models = await list_available_models()
@@ -1009,6 +1027,7 @@ async def handle_job(ctx: JobContext) -> None:
                             ack, reliable=True, topic="voice-control"
                         )
                     )
+                    _sync_task_model(task_dir, llm.model)
             except Exception as e:  # noqa: BLE001
                 logger.warning("Failed to update model: %s", e)
 

agent/task_worker.py

@@ -45,6 +45,22 @@ def _html_to_text(html: str) -> str:
     return text.strip()
 
 
+def _live_model(default: str) -> str:
+    """Follow the conversation's live model selection when it is fresher.
+
+    The main agent writes the UI-selected model to the session's task dir
+    (shared with this MCP subprocess) on every set_model. Env is the
+    fallback for tasks dispatched before any switch happened.
+    """
+    env = os.environ.get("TASKS_DIR", "/tmp/tasks")
+    try:
+        with open(os.path.join(env, "model.txt")) as f:
+            chosen = f.read().strip()
+    except OSError:
+        return default
+    return chosen or default
+
+
 class TaskWorker:
     """Runs a background research/lookup task using the LLM + tools."""
 
@@ -65,7 +81,7 @@ class TaskWorker:
             or os.environ.get("GEMMA_BASE_URL")
             or "http://192.168.86.2:3005/api"
         )
-        model = (
+        model = _live_model(
             os.environ.get("LLM_MODEL")
             or os.environ.get("GEMMA_MODEL")
             or "qwen2.5-vl-3b"