The supervisor LLM now embeds <type>[id1,id2]</type> entity tags in its
response text. The HomeFormatter buffers streamed tokens, detects complete
tags across chunk boundaries, and emits WsStreamBlock with entity type +
specific IDs. This replaces the old approach of emitting blocks for every
tool_end event, which dumped ALL entities regardless of relevance.
Also fixes:
- NoneType guard on metadata in _run_graph_stream (metadata can be None)
- Updated _HOME_SYSTEM prompt with entity tag instructions
- Updated all affected tests
- Add app/core/deep_agent.py with Home and Floating supervisor graphs
using LangGraph create_react_agent (hierarchical pattern)
- Strip ChatAgent classes from all 4 agent files, keep @tool functions
- Rewrite output_formatter.py for event-based (token/tool_end/mutations) stream
- Update device_ws.py to use run_home_stream/run_floating_stream
- Rewrite chat.py REST route to use run_home
- Add update_core_memory tool to both supervisors
- Add langgraph>=0.3.0 to requirements.txt
- Remove orchestrator.py, execution_plan.py, agent_registry.py, plans.py
- Remove PlanAction, PlanStep, ExecutionPlan, execution_mode from schemas
- Update all affected tests to match new API
- Remove 6 deprecated test files for deleted modules
- Clean up stale docstrings referencing removed orchestrator
- device_ws.py: dispatch home_request/popup_request to HomeFormatter/PopupFormatter
via async tasks; each request gets a UUID request_id for frame correlation
- chat.py: remove chat_stream WS endpoint (superseded by unified device WS);
keep POST /chat REST fallback unchanged
- 5 new integration tests pass; all 22 existing device_ws tests still pass
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
