- docs/deploy-architecture.md: full deployment topology, agent API, file transfer via shared volume
- .planning/HANDOFF.json + .continue-here.md: session state for Phase 05 planning
Task 4: stale room blocking + agent_id binding
- MatrixBot._check_agent_routing: blocks normal messages when user has no
selected agent or room is bound to a different agent
- agent_routing_enabled flag on MatrixRuntime activates the check only
in real multi-agent mode (RoutedPlatformClient)
- make_handle_new_chat now writes agent_id into new room metadata when
user already has a selected agent
Task 5: durable restart state tests
- test_restart_persistence.py proves selected_agent_id, room agent_id,
platform_chat_id, and the sequence counter all survive SQLiteStore
close/reopen; also covers clean startup with no prior state
Users can now list available agents with !agent and select one by
number. Selection persists in user metadata (selected_agent_id). If the
current room has no agent binding yet, selecting an agent binds it
immediately so the user can start messaging without !new.
Also updates the dispatcher test to reflect that real-mode platform is
now RoutedPlatformClient, not a bare RealPlatformClient.
Upstream AgentApi responses can return attachment objects that don't
implement the Attachment dataclass. _to_core_attachments coerces them
via duck-typing so OutgoingMessage always carries typed Attachment
instances regardless of the upstream response shape.
Instead of calling a /reset endpoint on platform-agent, !reset now
generates a new thread_id (platform_chat_id) for the room. The old
WebSocket connection is closed and the next message creates a fresh
context automatically. No platform changes required.
Add step-by-step setup for running Matrix surface with real platform-agent,
document all available commands, and clearly list what works vs what is
blocked (StateBackend cross-chat load, hardcoded tokens, missing /reset,
no file upload API).
