refactor local directory agent to two-phase LLM-with-tools architecture

Replace the single-pass FE-driven agent_run/agent_data flow with a
BE-orchestrated two-phase execution using LangChain tool-calling:
- Phase 1 (Triage): explores directory via new filesystem tools, matches
  files to existing projects using PROJECT_TOOLS
- Phase 2 (Processing): reads files and performs CRUD per project group
  with clean LLM context windows

Key changes:
- Add filesystem_agent.py with list_directory, read_file_content,
  get_file_metadata tools using execute_on_client()
- Move setup journey from REST to WebSocket (journey_start/message frames)
- Add batch_runs_per_day billing limit and enforce in /trigger
- Remove deprecated agent_data/agent_complete frame handlers and queues

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

app/core/device_manager.py

@@ -3,20 +3,15 @@
 Maintains in-memory state for all active Electron → backend WebSocket
 connections.  One connection per user (latest replaces previous).
 
-The manager participates in two interaction patterns:
+The manager handles the **tool-call round-trip** pattern:
+  - Backend sends ``tool_call`` frame → Electron executes the action →
+    returns ``tool_result`` frame.
+  - ``create_pending_call`` registers a Future keyed by ``call_id``.
+  - ``resolve_pending_call`` fulfils the Future; callers awaiting it
+    receive the result dict from Electron.
 
-1. **Tool-call round-trip** (bidirectional CRUD):
-   - Backend sends ``tool_call`` frame → Electron executes CRUD → returns
-     ``tool_result`` frame.
-   - ``create_pending_call`` registers a Future keyed by ``call_id``.
-   - ``resolve_pending_call`` fulfils the Future; callers awaiting it
-     receive the result dict from Electron.
-
-2. **Agent-data streaming** (local directory agent runs):
-   - Backend sends ``agent_run`` frame → Electron reads files and sends
-     back a stream of ``agent_data`` frames followed by ``agent_complete``.
-   - ``get_agent_data_queue`` returns (or creates) an asyncio.Queue for
-     a specific ``run_id`` so the agent runner can iterate frames.
+This pattern is used by all tools (CRUD, file-system, etc.) via
+``execute_on_client()`` in ``ws_context.py``.
 
 The ``device_manager`` module-level singleton is imported by both the
 device WS route and the agent runner.
@@ -42,8 +37,6 @@ class DeviceConnection:
     device_id: str
     # Futures indexed by tool_call id — resolved when tool_result arrives.
     pending_calls: dict[str, asyncio.Future[dict]] = field(default_factory=dict)
-    # Per-run queues for agent_data / agent_complete frames.
-    agent_data_queues: dict[str, asyncio.Queue[dict | None]] = field(default_factory=dict)
 
 
 class DeviceConnectionManager:
@@ -153,31 +146,6 @@ class DeviceConnectionManager:
         if fut is not None and not fut.done():
             fut.set_result(result)
 
-    # ── Agent-data queue ──────────────────────────────────────────────
-
-    def get_agent_data_queue(
-        self, user_id: str, run_id: str
-    ) -> asyncio.Queue[dict | None]:
-        """Return (creating if absent) the queue for *run_id* agent frames.
-
-        The agent runner reads from this queue.  The device WS handler writes
-        to it.  ``None`` is the sentinel that signals the stream is finished.
-        """
-        conn = self._connections.get(user_id)
-        if conn is None:
-            raise RuntimeError(
-                f"get_agent_data_queue: user {user_id!r} is not connected"
-            )
-        if run_id not in conn.agent_data_queues:
-            conn.agent_data_queues[run_id] = asyncio.Queue()
-        return conn.agent_data_queues[run_id]
-
-    def cleanup_agent_data_queue(self, user_id: str, run_id: str) -> None:
-        """Remove the queue for *run_id* once a run has completed."""
-        conn = self._connections.get(user_id)
-        if conn:
-            conn.agent_data_queues.pop(run_id, None)
-
 
 # Module-level singleton — import this everywhere.
 device_manager = DeviceConnectionManager()