🐛 fix(orchestration): surface reflection failures

- Record reflection-driven revisions as last_error and emit tool_result errors.

- Synthesize fallback output when only last_error is available.

Why: avoid silent incomplete completions when reflection blocks progress.

Tests: uv run pytest tests/test_orchestration.py::test_run_action_plan_reflection_blocks_progress tests/test_orchestration.py::test_response_synthesis_helpers -q

Dependencies: None

Author: bearlike

packages/meeseeks_core/src/meeseeks_core/action_runner.py

@@ -103,6 +103,9 @@ def run(self, task_queue: TaskQueue) -> TaskQueue:
                 continue
 
             if outcome.reflection is not None and outcome.reflection.status != "ok":
+                reason = outcome.reflection.notes or f"Reflection requested: {outcome.reflection.status}"
+                self._record_reflection_failure(action_step, reason, task_queue)
+                self._emit_tool_result(action_step, None, error=reason)
                 if outcome.reflection.revised_argument:
                     action_step.action_argument = outcome.reflection.revised_argument
                 action_step.result = None
@@ -166,7 +169,7 @@ def _ensure_permission(self, action_step: ActionStep) -> bool:
         if decision == PermissionDecision.DENY:
             mock = get_mock_speaker()
             message = (
-                "Permission denied for " f"{action_step.action_consumer}:{action_step.action_type}."
+                f"Permission denied for {action_step.action_consumer}:{action_step.action_type}."
             )
             action_step.result = mock(content=message)
             if not decision_logged:
@@ -221,6 +224,14 @@ def _record_failure(self, step: ActionStep, reason: str, task_queue: TaskQueue)
             mock = get_mock_speaker()
             step.result = mock(content=f"ERROR: {reason}")
 
+    def _record_reflection_failure(
+        self, step: ActionStep, reason: str, task_queue: TaskQueue
+    ) -> None:
+        note = f"{step.action_consumer} ({step.action_type}) needs revision"
+        if reason:
+            note = f"{note}: {reason}"
+        task_queue.last_error = note
+
     def _emit_tool_result(
         self, action_step: ActionStep, result: str | None, *, error: str | None = None
     ) -> None:

packages/meeseeks_core/src/meeseeks_core/orchestrator.py

@@ -343,7 +343,9 @@ def _collect_tool_outputs(task_queue: TaskQueue) -> list[str]:
                 continue
             content = getattr(step.result, "content", step.result)
             outputs.append(str(content))
-        return outputs
+        if outputs or not task_queue.last_error:
+            return outputs
+        return [f"ERROR: {task_queue.last_error}"]
 
     @staticmethod
     def _should_synthesize_response(task_queue: TaskQueue) -> bool:

tests/test_orchestration.py

@@ -736,6 +736,8 @@ def test_response_synthesis_helpers(monkeypatch):
         ]
     )
     assert Orchestrator._collect_tool_outputs(queue) == []
+    queue.last_error = "tool failed"
+    assert Orchestrator._collect_tool_outputs(queue) == ["ERROR: tool failed"]
     assert Orchestrator._should_synthesize_response(TaskQueue(action_steps=[])) is True
 
     def _fake_model(_inputs):
@@ -1314,6 +1316,8 @@ class DummyReflection:
         model_name="gpt-3.5-turbo",
     )
     assert task_queue.action_steps[0].result is None
+    assert task_queue.last_error is not None
+    assert "needs revision" in task_queue.last_error
 
 
 def test_orchestrate_session_auto_compact(monkeypatch, tmp_path):