Implement Phase 1: LangGraph backend MVP

Sets up the full backend foundation for CouncilOS:

- CouncilState TypedDict with all required fields and LangGraph reducers
- Three agent nodes: master_agent (drafts), critic_agent (scores + routes),
  writer_agent (final polish)
- LangGraph graph with cyclic rework loop: Master → Critic → (score < 8:
  back to Master | score ≥ 8: Writer → END)
- Safety valve: MAX_ITERATIONS=5 prevents infinite loops
- FastAPI app with REST endpoints (POST /api/councils/run, GET /api/councils/run/{id})
  and WebSocket endpoint (/ws/council/{run_id}) for real-time agent status events
- In-memory RunStore for Phase 1 (PostgreSQL-backed in Phase 3)
- pytest test suite: state, routing logic, critic parser, agent nodes, API endpoints
- .env.example, .gitignore, docker-compose.yml, Dockerfile

https://claude.ai/code/session_01RfMpt3TbMjZEtK3CAyP5iQ

backend/state.py

@@ -0,0 +1,47 @@
+"""
+CouncilState — the central data structure passed between all agents in LangGraph.
+
+All agents must read from and write to this TypedDict. Agents must not store
+state internally; everything passes through CouncilState.
+"""
+
+from typing import Annotated, List, Optional
+import operator
+from typing_extensions import TypedDict
+
+
+class CouncilState(TypedDict):
+    """
+    The global state shared across all agents in a council run.
+
+    Fields:
+        input_topic:        The user's original prompt or uploaded PDF content.
+        current_draft:      The document currently being worked on.
+        feedback_history:   All critic feedback accumulated across loop iterations.
+                            Agents append here — never overwrite.
+        route_decision:     Routing signal used by conditional edges.
+                            Values: "rework" | "approve" | custom strings.
+        messages:           LLM message history (system prompts + responses).
+                            Uses operator.add reducer so messages accumulate.
+        iteration_count:    Tracks how many rework loops have occurred.
+        critic_score:       The numeric score (0–10) assigned by the critic agent.
+        run_id:             Unique identifier for this council run (for WebSocket events).
+        active_node:        Name of the currently executing agent node (for UI updates).
+    """
+
+    input_topic: str
+    current_draft: str
+    feedback_history: Annotated[List[str], operator.add]
+    route_decision: str
+    messages: Annotated[list, operator.add]
+    iteration_count: int
+    critic_score: Optional[float]
+    run_id: str
+    active_node: str
+
+
+# Approval threshold: critic score must reach this value to exit the loop
+APPROVAL_THRESHOLD = 8.0
+
+# Safety limit: maximum number of rework iterations before forcing approval
+MAX_ITERATIONS = 5