fix: isolate session loading

core/framework/llm/litellm.py

@@ -328,6 +328,20 @@ class LiteLLMProvider(LLMProvider):
                         f"Full request dumped to: {dump_path}"
                     )
 
+                    # finish_reason=length means the model exhausted max_tokens
+                    # before producing content. Retrying with the same max_tokens
+                    # will never help — return immediately instead of looping.
+                    if finish_reason == "length":
+                        max_tok = kwargs.get("max_tokens", "unset")
+                        logger.error(
+                            f"[retry] {model} returned empty content with "
+                            f"finish_reason=length (max_tokens={max_tok}). "
+                            f"The model exhausted its token budget before "
+                            f"producing visible output. Increase max_tokens "
+                            f"or use a different model. Not retrying."
+                        )
+                        return response
+
                     if attempt == retries:
                         logger.error(
                             f"[retry] GAVE UP on {model} after {retries + 1} "
@@ -621,6 +635,20 @@ class LiteLLMProvider(LLMProvider):
                         f"Full request dumped to: {dump_path}"
                     )
 
+                    # finish_reason=length means the model exhausted max_tokens
+                    # before producing content. Retrying with the same max_tokens
+                    # will never help — return immediately instead of looping.
+                    if finish_reason == "length":
+                        max_tok = kwargs.get("max_tokens", "unset")
+                        logger.error(
+                            f"[async-retry] {model} returned empty content with "
+                            f"finish_reason=length (max_tokens={max_tok}). "
+                            f"The model exhausted its token budget before "
+                            f"producing visible output. Increase max_tokens "
+                            f"or use a different model. Not retrying."
+                        )
+                        return response
+
                     if attempt == retries:
                         logger.error(
                             f"[async-retry] GAVE UP on {model} after {retries + 1} "
@@ -903,6 +931,7 @@ class LiteLLMProvider(LLMProvider):
             tool_calls_acc: dict[int, dict[str, str]] = {}
             input_tokens = 0
             output_tokens = 0
+            stream_finish_reason: str | None = None
 
             try:
                 response = await litellm.acompletion(**kwargs)  # type: ignore[union-attr]
@@ -938,6 +967,7 @@ class LiteLLMProvider(LLMProvider):
 
                     # --- Finish ---
                     if choice.finish_reason:
+                        stream_finish_reason = choice.finish_reason
                         for _idx, tc_data in sorted(tool_calls_acc.items()):
                             try:
                                 parsed_args = json.loads(tc_data["arguments"])
@@ -992,6 +1022,24 @@ class LiteLLMProvider(LLMProvider):
                         for event in tail_events:
                             yield event
                         return
+
+                    # finish_reason=length means the model exhausted
+                    # max_tokens before producing content. Retrying with
+                    # the same max_tokens will never help.
+                    if stream_finish_reason == "length":
+                        max_tok = kwargs.get("max_tokens", "unset")
+                        logger.error(
+                            f"[stream] {self.model} returned empty content "
+                            f"with finish_reason=length "
+                            f"(max_tokens={max_tok}). The model exhausted "
+                            f"its token budget before producing visible "
+                            f"output. Increase max_tokens or use a "
+                            f"different model. Not retrying."
+                        )
+                        for event in tail_events:
+                            yield event
+                        return
+
                     wait = _compute_retry_delay(attempt)
                     token_count, token_method = _estimate_tokens(
                         self.model,

core/framework/server/agent_manager.py

@@ -61,7 +61,7 @@ async def _extract_subgraph_steps(nodes: list, llm: Any) -> None:
 
             response = await llm.acomplete(
                 messages=[{"role": "user", "content": prompt}],
-                max_tokens=1000,
+                max_tokens=4096,
                 json_mode=True,
             )
 
@@ -172,13 +172,6 @@ class AgentManager:
             if runner._agent_runtime is None:
                 await loop.run_in_executor(None, runner._setup)
 
-            # Extract subgraph steps for frontend visualization (non-critical)
-            if runner.graph and runner._llm:
-                try:
-                    await _extract_subgraph_steps(runner.graph.nodes, runner._llm)
-                except Exception as e:
-                    logger.warning(f"Subgraph extraction skipped: {e}")
-
             runtime = runner._agent_runtime
 
             # Start runtime on event loop
@@ -224,6 +217,9 @@ class AgentManager:
         - **Judge**: timer-driven background GraphExecutor (silent monitoring)
         - **Worker**: the existing AgentRuntime (unchanged)
         """
+        import uuid
+        from datetime import datetime
+
         from framework.graph.executor import GraphExecutor
         from framework.monitoring import judge_goal, judge_graph
         from framework.runner.tool_registry import ToolRegistry
@@ -238,6 +234,12 @@ class AgentManager:
             event_bus = runtime._event_bus
             llm = runtime._llm
 
+            # Generate a shared session ID for queen, judge, and worker.
+            # All three use the same ID so conversations are scoped to this
+            # agent load and start fresh each time.
+            ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+            session_id = f"session_{ts}_{uuid.uuid4().hex[:8]}"
+
             # 1. Monitoring tools — standalone registry, NOT merged into worker
             monitoring_registry = ToolRegistry()
             register_worker_monitoring_tools(
@@ -247,14 +249,15 @@ class AgentManager:
                 worker_graph_id=runtime._graph_id,
             )
 
-            # 2. Storage dirs
+            # 2. Storage dirs — scoped by session_id so each agent load
-            judge_dir = storage_path / "graphs" / "worker_health_judge" / "session"
+            #    gets fresh queen/judge conversations.
+            judge_dir = storage_path / "graphs" / "judge" / "session" / session_id
             judge_dir.mkdir(parents=True, exist_ok=True)
-            queen_dir = storage_path / "graphs" / "queen" / "session"
+            queen_dir = storage_path / "graphs" / "queen" / "session" / session_id
             queen_dir.mkdir(parents=True, exist_ok=True)
 
             # 3. Health judge — background task, fires every 2 minutes
-            judge_runtime = Runtime(storage_path / "graphs" / "worker_health_judge")
+            judge_runtime = Runtime(storage_path / "graphs" / "judge")
             monitoring_tools = list(monitoring_registry.get_tools().values())
             monitoring_executor = monitoring_registry.get_executor()
 
@@ -272,7 +275,7 @@ class AgentManager:
                             tools=monitoring_tools,
                             tool_executor=monitoring_executor,
                             event_bus=event_bus,
-                            stream_id="worker_health_judge",
+                            stream_id="judge",
                             storage_path=judge_dir,
                             loop_config=judge_graph.loop_config,
                         )
@@ -282,7 +285,7 @@ class AgentManager:
                             input_data={
                                 "event": {"source": "timer", "reason": "scheduled"},
                             },
-                            session_state={"resume_session_id": "persistent"},
+                            session_state={"resume_session_id": session_id},
                         )
                     except Exception:
                         logger.error("Health judge tick failed", exc_info=True)
@@ -300,6 +303,7 @@ class AgentManager:
                 worker_runtime=runtime,
                 event_bus=event_bus,
                 storage_path=storage_path,
+                session_id=session_id,
             )
             register_worker_monitoring_tools(
                 queen_registry,
@@ -365,7 +369,7 @@ class AgentManager:
                         graph=queen_graph,
                         goal=queen_goal,
                         input_data={"greeting": "Session started."},
-                        session_state={"resume_session_id": "persistent"},
+                        session_state={"resume_session_id": session_id},
                     )
                     logger.warning("Queen executor returned (should be forever-alive)")
                 except Exception:

core/framework/tools/queen_lifecycle_tools.py

@@ -36,9 +36,14 @@ def register_queen_lifecycle_tools(
     worker_runtime: AgentRuntime,
     event_bus: EventBus,
     storage_path: Path | None = None,
+    session_id: str | None = None,
 ) -> int:
     """Register queen lifecycle tools bound to *worker_runtime*.
 
+    Args:
+        session_id: Shared session ID so the worker uses the same session
+                    scope as the queen and judge.
+
     Returns the number of tools registered.
     """
     from framework.llm.provider import Tool
@@ -55,7 +60,12 @@ def register_queen_lifecycle_tools(
         """
         try:
             # Get session state from any prior execution for memory continuity
-            session_state = worker_runtime._get_primary_session_state("default")
+            session_state = worker_runtime._get_primary_session_state("default") or {}
+
+            # Use the shared session ID so queen, judge, and worker all
+            # scope their conversations to the same session.
+            if session_id:
+                session_state["resume_session_id"] = session_id
 
             exec_id = await worker_runtime.trigger(
                 entry_point_id="default",

core/framework/tui/app.py

@@ -468,6 +468,8 @@ class AdenTUI(App):
         into the worker runtime.  The worker is completely untouched.
         """
         import asyncio
+        import uuid
+        from datetime import datetime
         from pathlib import Path
 
         from framework.graph.executor import GraphExecutor
@@ -486,6 +488,10 @@ class AdenTUI(App):
             llm = self.runtime._llm
             agent_loop = self.chat_repl._agent_loop
 
+            # Generate a shared session ID for queen, judge, and worker.
+            ts = datetime.now().strftime("%Y%m%d_%H%M%S")
+            session_id = f"session_{ts}_{uuid.uuid4().hex[:8]}"
+
             # 1. Monitoring tools (health summary, emit ticket, notify operator).
             #    Registered on a standalone registry — NOT merged into the worker.
             monitoring_registry = ToolRegistry()
@@ -496,11 +502,11 @@ class AdenTUI(App):
                 worker_graph_id=self.runtime._graph_id,
             )
 
-            # 2. Storage dirs — under worker's base path but completely owned
+            # 2. Storage dirs — scoped by session_id so each agent load
-            #    by the judge/queen.  Worker never writes here.
+            #    gets fresh queen/judge conversations.
-            judge_dir = storage_path / "graphs" / "judge" / "session"
+            judge_dir = storage_path / "graphs" / "judge" / "session" / session_id
             judge_dir.mkdir(parents=True, exist_ok=True)
-            queen_dir = storage_path / "graphs" / "queen" / "session"
+            queen_dir = storage_path / "graphs" / "queen" / "session" / session_id
             queen_dir.mkdir(parents=True, exist_ok=True)
 
             # ---------------------------------------------------------------
@@ -542,7 +548,7 @@ class AdenTUI(App):
                             input_data={
                                 "event": {"source": "timer", "reason": "scheduled"},
                             },
-                            session_state={"resume_session_id": "persistent"},
+                            session_state={"resume_session_id": session_id},
                         )
                     except Exception:
                         log.error("Health judge tick failed", exc_info=True)
@@ -584,6 +590,7 @@ class AdenTUI(App):
                 worker_runtime=self.runtime,
                 event_bus=event_bus,
                 storage_path=storage_path,
+                session_id=session_id,
             )
             register_worker_monitoring_tools(
                 queen_registry,
@@ -596,9 +603,6 @@ class AdenTUI(App):
             queen_tool_executor = queen_registry.get_executor()
 
             # Build worker identity to inject into the queen's system prompt.
-            # This must be in the system prompt (not input_data) because
-            # persistent sessions restore the old conversation and skip
-            # _build_initial_message — the queen would lose context.
             worker_graph_id = self.runtime._graph_id
             worker_goal_name = getattr(self.runtime.goal, "name", worker_graph_id)
             worker_goal_desc = getattr(self.runtime.goal, "description", "")
@@ -657,7 +661,7 @@ class AdenTUI(App):
                         graph=queen_graph,
                         goal=queen_goal,
                         input_data={"greeting": "Session started."},
-                        session_state={"resume_session_id": "persistent"},
+                        session_state={"resume_session_id": session_id},
                     )
                     # Should never reach here — queen is forever-alive.
                     log.warning(

core/frontend/src/pages/workspace.tsx

@@ -569,7 +569,7 @@ export default function Workspace() {
       const streamId = event.stream_id;
 
       // Suppress judge events (silent background monitoring)
-      if (streamId === "worker_health_judge") return;
+      if (streamId === "judge") return;
 
       // Determine if this is a queen event
       const isQueen = streamId === "queen";