Stabilize live chat tool loop

tests/smoke/test_api_stream_chat.py

@@ -113,6 +113,7 @@ def test_stream_chat_forwards_reasoning_toggle_to_thinker(tmp_path, monkeypatch)
 
 def test_stream_chat_runs_memory_policy_and_reflection_after_completion(tmp_path, monkeypatch):
     monkeypatch.setenv("DUCK_DB_PATH", str(tmp_path / "duck.sqlite3"))
+    monkeypatch.setenv("DUCK_ENABLE_REFLECTION", "1")
 
     async def fake_chat(self, role, messages, temperature=None, max_output_tokens=None, response_format=None):
         if role == "action":

tests/smoke/test_runtime_tools.py

@@ -176,6 +176,41 @@ class FakeMalformedActionModelClient:
         )
 
 
+class FakeRepeatingActionModelClient:
+    async def chat(self, role, messages):
+        if role == "action":
+            return ModelResponse(
+                role=role,
+                model="local-main",
+                content=json.dumps(
+                    {
+                        "kind": "action_directive",
+                        "intent": "repeat same action",
+                        "risk_level": "low",
+                        "actions": [
+                            {
+                                "tool": "file_read",
+                                "args": {"path": "note.txt"},
+                                "reason": "Read requested file",
+                            }
+                        ],
+                    }
+                ),
+                reasoning_content=None,
+                raw={},
+                latency_ms=5.0,
+            )
+        assert role == "thinker"
+        return ModelResponse(
+            role=role,
+            model="local-main",
+            content="Final answer from first observation.",
+            reasoning_content=None,
+            raw={},
+            latency_ms=12.0,
+        )
+
+
 @pytest.mark.asyncio
 async def test_runtime_executes_action_directive_tool_and_finishes_with_observation(tmp_path):
     (tmp_path / "note.txt").write_text("hello from tool")
@@ -253,6 +288,48 @@ async def test_runtime_rejects_malformed_action_directive_before_tools(tmp_path)
     assert not any(event.event_type == "tool_call_started" for event in events)
 
 
+def test_runtime_compacts_large_tool_observations_for_model_context(tmp_path):
+    db_path = str(tmp_path / "duck.sqlite3")
+    task_store = TaskStore(db_path)
+    event_store = EventStore(db_path)
+    loop = RuntimeLoop(task_store, event_store, FakeToolModelClient())
+
+    compact = loop.format_tool_observations_for_model([
+        {
+            "tool": "shell_exec_safe",
+            "result": {
+                "ok": True,
+                "output": "A" * 2500 + "KEEP_TAIL",
+                "metadata": {"command": "ls /tmp"},
+            },
+        }
+    ])
+
+    assert "tool_observations" in compact
+    assert "truncated" in compact
+    assert "KEEP_TAIL" in compact
+    assert len(compact) < 2300
+
+
+@pytest.mark.asyncio
+async def test_runtime_skips_duplicate_action_within_same_task(tmp_path):
+    (tmp_path / "note.txt").write_text("hello once")
+    db_path = str(tmp_path / "duck.sqlite3")
+    task_store = TaskStore(db_path)
+    event_store = EventStore(db_path)
+    loop = RuntimeLoop(task_store, event_store, FakeRepeatingActionModelClient())
+
+    result = await loop.run_chat("read note.txt", str(tmp_path), debug=True)
+    events = await event_store.list_events(result.task_id)
+    finished_tools = [event for event in events if event.event_type == "tool_call_finished"]
+    skipped_tools = [event for event in events if event.event_type == "tool_call_skipped"]
+
+    assert result.status == "completed"
+    assert len(finished_tools) == 1
+    assert len(skipped_tools) == 1
+    assert skipped_tools[0].payload["reason"] == "duplicate_action"
+
+
 class FakeApprovalModelClient:
     async def chat(self, role, messages):
         if role == "action":