fix: harden L3 runtime continuity and tool execution

Align the L3 graph, agent service, and sync tool shims on one canonical continuity contract so clarification resumes and persisted snapshots behave consistently. Add targeted regressions and hardening notes covering system-message coalescing, async bridge usage, and continuity rehydration.

backend/app/agents/state.py

@@ -1,6 +1,6 @@
-from dataclasses import dataclass, field
-from typing import TypedDict, Annotated, Sequence
+from dataclasses import dataclass
 from enum import Enum
+from typing import Annotated, Any, TypedDict
 
 from langchain_core.messages import BaseMessage
 from langgraph.graph.message import add_messages
@@ -23,40 +23,65 @@ class ConversationTurn:
 
 
 class AgentState(TypedDict):
-    # Core message history with add_messages reducer
     messages: Annotated[list[BaseMessage], add_messages]
-    
-    # Session identifiers
     user_id: str
     conversation_id: str
 
-    # Agent routing state
     current_agent: str | None
-    next_step: str | None  # For explicit graph routing
-    
-    # Traceability
+    next_step: str | None
+    active_agents: list[AgentRole]
+    current_sub_commander: str | None
+    active_sub_commanders: list[str]
+    sub_commander_trace: list[dict[str, Any]]
     agent_trace: list[str]
-    
-    # Task & Entity Tracking (Business Logic)
-    pending_tasks: list[dict]
-    completed_tasks: list[dict]
-    created_entities: list[dict]
 
-    # Context summaries (for long-term or cross-agent context)
+    pending_tasks: list[dict[str, Any]]
+    completed_tasks: list[dict[str, Any]]
+    tool_calls: list[dict[str, Any]]
+    last_tool_result: str | None
+    action_results: list[dict[str, Any]]
+    created_entities: list[dict[str, Any]]
+    tool_outcomes: list[dict[str, Any]]
+
+    tool_strategy_used: str | None
+    tool_round_count: int
+    max_tool_rounds: int
+    retry_count: int
+    max_retries: int
+    iteration_count: int
+    max_iterations: int
+    routing_hops: int
+    max_routing_hops: int
+    terminated_due_to_loop_guard: bool
+    retrieval_trace: list[dict[str, Any]]
+    stop_reason: str | None
+
+    clarification_needed: bool
+    clarification_question: str | None
+    fallback_parse_error: str | None
+    should_respond: bool
+
     knowledge_context: str | None
+    graph_context: str | None
     schedule_context_summary: str | None
+    plan: str | None
+    plan_steps: list[dict[str, Any]]
     analysis_report: str | None
-
-    # Output control
     final_response: str | None
-    
-    # Memory & Environment
+
     memory_context: str | None
     current_datetime_context: str | None
-    
-    # Configuration
-    user_llm_config: dict | None
-    provider_capabilities: dict | None
+    current_datetime_reference: dict[str, str] | None
+
+    turn_context: dict[str, Any] | None
+    routing_decision: dict[str, Any] | None
+    continuity_state: dict[str, Any] | None
+    pending_action: dict[str, Any] | None
+    last_completed_action: dict[str, Any] | None
+    clarification_context: dict[str, Any] | None
+
+    user_llm_config: dict[str, Any] | None
+    provider_capabilities: dict[str, Any] | None
 
 
 def initial_state(user_id: str, conversation_id: str) -> AgentState:
@@ -66,16 +91,50 @@ def initial_state(user_id: str, conversation_id: str) -> AgentState:
         conversation_id=conversation_id,
         current_agent=AgentRole.MASTER.value,
         next_step=None,
+        active_agents=[AgentRole.MASTER],
+        current_sub_commander=None,
+        active_sub_commanders=[],
+        sub_commander_trace=[],
         agent_trace=[AgentRole.MASTER.value],
         pending_tasks=[],
         completed_tasks=[],
+        tool_calls=[],
+        last_tool_result=None,
+        action_results=[],
         created_entities=[],
+        tool_outcomes=[],
+        tool_strategy_used=None,
+        tool_round_count=0,
+        max_tool_rounds=2,
+        retry_count=0,
+        max_retries=1,
+        iteration_count=0,
+        max_iterations=3,
+        routing_hops=0,
+        max_routing_hops=2,
+        terminated_due_to_loop_guard=False,
+        retrieval_trace=[],
+        stop_reason=None,
+        clarification_needed=False,
+        clarification_question=None,
+        fallback_parse_error=None,
+        should_respond=True,
         knowledge_context=None,
+        graph_context=None,
         schedule_context_summary=None,
+        plan=None,
+        plan_steps=[],
         analysis_report=None,
         final_response=None,
         memory_context=None,
         current_datetime_context=None,
+        current_datetime_reference=None,
+        turn_context=None,
+        routing_decision=None,
+        continuity_state=None,
+        pending_action=None,
+        last_completed_action=None,
+        clarification_context=None,
         user_llm_config=None,
         provider_capabilities=None,
     )