fix: preserve task board interventions

frontier_agent/components/observers/task_board.py

@@ -16,7 +16,11 @@
 class TaskBoardObserver(BaseObserver):
     """Re-inject a non-empty task board after a configurable cooldown."""
 
-    critical = False
+    # This observer returns an Intervention.  Non-critical observer hooks are
+    # fire-and-forget and their return values are deliberately not collected by
+    # notify_observers, so this must remain critical for reminders to reach the
+    # next LLM turn.
+    critical = True
 
     def __init__(
         self,

tests/test_task_board_intervention.py

@@ -0,0 +1,115 @@
+from __future__ import annotations
+
+import asyncio
+import json
+
+from frontier_agent.components.observers.task_board import TaskBoardObserver
+from frontier_agent.core.llm import LLMResponse
+from frontier_agent.core.loop_types import (
+    LoopConfig,
+    LoopPolicy,
+    TurnContext,
+    notify_observers,
+)
+from frontier_agent.core.runtime.loop.agent_loop import run_agent_loop
+from frontier_agent.core.tool import tool
+
+_REMINDER = (
+    "Current task board (keep it current via add_task / update_task; "
+    "finalize only once every task is resolved):\nboard"
+)
+
+
+def _observer() -> TaskBoardObserver:
+    return TaskBoardObserver(
+        board_size=lambda _task_id: 1,
+        render_board=lambda _task_id, _bus_task_id: "board",
+        resolve_bus_task_id=lambda _scope: "bus-task",
+    )
+
+
+def _context(turn: int = 1) -> TurnContext:
+    return TurnContext(
+        turn=turn,
+        max_turns=3,
+        task_id="task",
+        role_id="coordinator",
+        ai_text="",
+        thinking="",
+        tool_calls=[],
+        messages=[],
+        usage=None,
+        metadata={},
+    )
+
+
+def test_dispatcher_collects_task_board_intervention() -> None:
+    """Exercise the dispatcher contract, not the observer hook in isolation."""
+    observer = _observer()
+
+    interventions = asyncio.run(
+        notify_observers([observer], "on_turn_end", _context()),
+    )
+
+    assert observer.critical is True
+    assert len(interventions) == 1
+    assert interventions[0].inject_messages == [_REMINDER]
+
+
+@tool
+async def _probe() -> str:
+    """Keep the first loop turn productive so a second LLM turn runs."""
+    return "ok"
+
+
+class _ReminderCapturingLLM:
+    model = "stub"
+
+    def __init__(self) -> None:
+        self.requests: list[list[dict]] = []
+
+    async def chat(self, messages: list[dict], **_kwargs: object) -> LLMResponse:
+        self.requests.append([dict(message) for message in messages])
+        if len(self.requests) == 1:
+            return LLMResponse(
+                content="",
+                tool_calls=[{
+                    "id": "call-probe",
+                    "type": "function",
+                    "function": {
+                        "name": "_probe",
+                        "arguments": json.dumps({}),
+                    },
+                }],
+                finish_reason="tool_calls",
+            )
+        return LLMResponse(content="done", finish_reason="stop")
+
+
+def test_task_board_intervention_reaches_next_llm_request() -> None:
+    """End to end: on_turn_end -> dispatcher -> history -> next LLM call."""
+    llm = _ReminderCapturingLLM()
+
+    result = asyncio.run(
+        run_agent_loop(
+            system_prompt="system",
+            user_message="coordinate this task",
+            llm=llm,
+            tools=[_probe],
+            config=LoopConfig(
+                max_turns=3,
+                task_id="task",
+                role_id="coordinator",
+                loop_policy=LoopPolicy(no_tool_behavior="stop"),
+            ),
+            observers=[_observer()],
+        ),
+    )
+
+    assert len(llm.requests) == 2
+    assert any(
+        message.get("role") == "user" and message.get("content") == _REMINDER
+        for message in llm.requests[1]
+    ), "the task-board Intervention was swallowed before the next LLM call"
+    assert result.stopped_by == "no_tool"
+    assert result.final_content == "done"