feat: add official Codex skill and harden task execution

Install and update artemis-use with task diagnostics and plugin guidance. Preserve cross-device file receipts, review canonical reports, guard stage transitions and reduce small history summaries. Improve exploration lifecycle, request recovery and browser live connections.

tests/unit/agents/test_operator_recovery.py

@@ -127,23 +127,74 @@ async def test_completed_exploration_plan_does_not_force_an_extra_action(
         content="Confirmed home; recorded evidence.",
         tool_calls=[{"name": "save_note", "args": {}, "id": "plan"}],
     )
-    invoke = AsyncMock(side_effect=[response, click()])
+    final = AIMessage(content="# 完整调查报告\n已回主页，证据见步骤 3。未检查其他页面。")
+    finishing = scenario in {"complete", "unchanged"}
+    invoke = AsyncMock(side_effect=[response, final if finishing else click()])
     monkeypatch.setattr("artemis.agents.operator.operator.acomplete", invoke)
     messages = []
     result = await node._invoke_llm_loop(
         Mock(), messages, [SimpleNamespace(name="save_note")], [], state
     )
-    if scenario == "complete":
+    if finishing:
+        assert invoke.await_count == 2
+        assert result[0] is None
+        assert final.content in result[1]
+        assert state.execution_report == final.content
+        assert any("完整报告正文" in m.content for m in messages if isinstance(m, SystemMessage))
+        assert result[3] is False
+    elif scenario == "pending":
         assert invoke.await_count == 1
         assert result[0] is None
-        assert result[1] == "Confirmed home; recorded evidence."
-        assert result[3] is False
+        assert state.execution_report is None
+        return
     else:
         assert invoke.await_count == 2
         assert result[0]
+        assert state.execution_report is None
     assert any(isinstance(m, ToolMessage) and m.tool_call_id == "plan" for m in messages)
 
 
+@pytest.mark.asyncio
+async def test_final_report_is_discarded_when_guidance_arrives(operator, monkeypatch):
+    node, state = operator
+    state.execution_report = "旧报告"
+    node.ctx.exploration_id = "exploration"
+    node._read_task_plan_text = lambda: "- [x] 调查完成"
+    pending = {"value": False}
+    monkeypatch.setattr(
+        "artemis.exploration.runtime.needs_exploration_turn", lambda ctx: pending["value"]
+    )
+    publish = Mock()
+    monkeypatch.setattr("artemis.exploration.runtime.publish_response", publish)
+
+    async def answer(*args):
+        pending["value"] = True
+        return AIMessage(content="这是过时的最终报告")
+
+    monkeypatch.setattr("artemis.agents.operator.operator.acomplete", answer)
+    result = await node._invoke_llm_loop(Mock(), [], [], [], state)
+    assert result[0] is None
+    assert state.execution_report is None
+    publish.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_empty_final_report_uses_bounded_recovery(operator, monkeypatch):
+    node, state = operator
+    node.ctx.exploration_id = "exploration"
+    node._read_task_plan_text = lambda: "- [x] 调查完成"
+    monkeypatch.setattr("artemis.exploration.runtime.needs_exploration_turn", lambda ctx: False)
+    monkeypatch.setattr("artemis.exploration.runtime.publish_response", Mock())
+    monkeypatch.setattr("artemis.agents.operator.operator.get_llm", Mock(return_value=Mock()))
+    monkeypatch.setattr(settings, "ARTEMIS_TOOL_FALLBACK_MAX_NO_PROGRESS_ATTEMPTS", 2)
+    invoke = AsyncMock(return_value=AIMessage(content="  "))
+    monkeypatch.setattr("artemis.agents.operator.operator.acomplete", invoke)
+    with pytest.raises(OperatorNoProgressError):
+        await node._invoke_llm_loop(Mock(), [], [], [], state)
+    assert state.execution_report is None
+    assert invoke.await_count == 5
+
+
 @pytest.mark.asyncio
 async def test_invalid_target_can_be_corrected_before_limit(operator, monkeypatch):
     result, messages, invoke = await run_loop(operator, monkeypatch, [click("14"), click("13")])