steps B.3/B.4/B.5 complete: bidirectional WS handler, _tool_loop verified, clean final frame

AI_REFACTOR_PLAN.md

@@ -158,7 +158,7 @@ Tools must use **camelCase** field names (Drizzle maps them to snake_case intern
 - **Outcome:** All 23 tools query real user data via WS. LLM sees actual rows, not action descriptors.
 
 ### Step B.3 — Bidirectional WebSocket handler
-- [ ] Refactor `app/api/routes/chat.py` WS endpoint:
+- [x] Refactor `app/api/routes/chat.py` WS endpoint:
   - After auth + accept + receive `chat_request`:
     1. Create `execute_on_client` callback closure capturing the websocket:
        ```python
@@ -192,13 +192,13 @@ Tools must use **camelCase** field names (Drizzle maps them to snake_case intern
 - **Outcome:** Full bidirectional WS. Tool calls and text streaming happen concurrently on the same connection.
 
 ### Step B.4 — `_tool_loop` — no changes needed
-- [ ] Verify `app/core/agent_registry.py` works unchanged:
+- [x] Verify `app/core/agent_registry.py` works unchanged:
   - `_tool_loop` calls `tool_fn.ainvoke(args)` → tool awaits `execute_on_client()` (WS round-trip) → returns string → `ToolMessage(content=string)` → LLM sees real data
   - The async WS round-trip happens inside each tool. `_tool_loop` just sees an awaited tool returning a string — same as before, different content.
 - **No code changes.** Just verify + add a log line for tool execution times if desired.
 
 ### Step B.5 — Orchestrator cleanup
-- [ ] Update `app/core/orchestrator.py`:
+- [x] Update `app/core/orchestrator.py`:
   - `orchestrate_stream()`: remove `"actions": []` from final frame. Final becomes: `{"done": true, "response": "..."}`
   - No other changes — `classify_intent` → `call_agent` → chunk response → final frame
 - **Files:** `app/core/orchestrator.py`

app/core/orchestrator.py

@@ -144,14 +144,15 @@ async def orchestrate_stream(
     request: ChatRequest,
     reg: AgentRegistry | None = None,
 ) -> AsyncGenerator[str, None]:
-    """Streaming orchestration — yields text chunks then a final JSON frame.
+    """Streaming orchestration — yields plain text chunks only.
 
-    The final frame is a JSON object:
-    ``{"done": true, "response": "...", "actions": []}``.
+    The WebSocket handler in ``app/api/routes/chat.py`` is responsible for
+    wrapping each chunk in a ``text_chunk`` frame and sending the final
+    ``final`` frame once the generator is exhausted.
 
     Agents do not yet support token-level streaming; the full response is
-    fetched first, then emitted in fixed-size chunks.  Token-level streaming
-    will be wired in Step 6 when agents expose ``astream()``.
+    fetched first (which may involve multiple WS round-trips for tool calls),
+    then emitted in fixed-size chunks.
     """
     if reg is None:
         reg = _default_registry
@@ -163,6 +164,3 @@ async def orchestrate_stream(
     chunk_size = 50
     for i in range(0, len(response_text), chunk_size):
         yield response_text[i : i + chunk_size]
-
-    final = ChatResponse(response=response_text)
-    yield json.dumps({"done": True, **final.model_dump()})