feat: add guided exploration and share user-started MCP backend

tests/unit/agents/test_checker.py

@@ -730,6 +730,27 @@ async def test_final_review_reports_no_guidance_explicitly():
     assert "# User guidance received during the run\n(none)\n" in text
 
 
+@pytest.mark.asyncio
+async def test_exploration_review_sees_conversation_and_can_read_earlier_messages():
+    ctx = _mock_ctx()
+    ctx.exploration_id = "exploration"
+    conversation = HumanMessage(content="用户已确认系统设置；示范不代表独立复跑")
+    history = _text_tool("read_exploration_messages", "历史回复")
+    review = AsyncMock(return_value=CheckReport())
+    with (
+        patch("artemis.agents.checker.checker._run_check_loop", review),
+        patch(
+            "artemis.agents.checker.checker._capture_final_screen",
+            AsyncMock(return_value=(None, "elements")),
+        ),
+        patch("artemis.exploration.runtime.conversation_message", return_value=conversation),
+        patch("artemis.exploration.runtime.history_tool", return_value=history),
+    ):
+        await run_final_check(ctx, goal="探索设置", plan_text="", ledger=[], check_items=[])
+    assert conversation in review.call_args.args[1]
+    assert history in review.call_args.args[2]
+
+
 # --- tool results: status is structural, never sniffed from the words -----------------
 
 

tests/unit/agents/test_flash_acceptance.py

@@ -29,6 +29,13 @@ async def test_flash_reuses_checker_with_original_prompt_and_final_report(
     check = AsyncMock(return_value=CheckReport(test_result=result))
     monkeypatch.setattr(flash, "run_final_check", check)
     ctx = context(tmp_path)
+    ctx.on_execution_review = AsyncMock()
+
+    async def review_after_capture_stops(*args, **kwargs):
+        ctx.on_execution_review.assert_awaited_once()
+        return CheckReport(test_result=result)
+
+    check.side_effect = review_after_capture_stops
     execution = {"status": "completed", "explanation": "操作已执行"}
     data = await flash.settle_flash_outcome(ctx, "打开设置，预期看到应用入口", execution)
     assert check.call_count == 1

tests/unit/agents/test_operator_recovery.py

@@ -93,6 +93,57 @@ async def test_prose_retry_keeps_context_and_recovers_once(operator, monkeypatch
     assert any(isinstance(m, SystemMessage) and "No tool call" in m.content for m in messages)
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "scenario", ["complete", "pending", "unfinished", "unchanged", "failed", "continuous", "pro"]
+)
+async def test_completed_exploration_plan_does_not_force_an_extra_action(
+    operator, monkeypatch, scenario
+):
+    node, state = operator
+    node.ctx.exploration_id = None if scenario == "pro" else "exploration"
+    node.ctx.data_engine = None
+    plan = {"text": "- [ ] Return home"}
+    if scenario == "unchanged":
+        plan["text"] = "- [x] Return home"
+    node._read_task_plan_text = lambda: plan["text"]
+    monkeypatch.setattr("artemis.exploration.runtime.publish_response", Mock())
+    monkeypatch.setattr(
+        "artemis.exploration.runtime.needs_exploration_turn", lambda ctx: scenario == "pending"
+    )
+
+    async def save_note(**kwargs):
+        plan["text"] = "- [ ] Return home" if scenario == "unfinished" else "- [x] Return home"
+        if scenario == "continuous":
+            plan["text"] += " [Loop:continuous]"
+        return ToolMessage(
+            content="Saved plan",
+            tool_call_id="plan",
+            status="error" if scenario == "failed" else "success",
+        )
+
+    monkeypatch.setattr("artemis.agents.operator.operator.invoke_tool_with_injection", save_note)
+    response = AIMessage(
+        content="Confirmed home; recorded evidence.",
+        tool_calls=[{"name": "save_note", "args": {}, "id": "plan"}],
+    )
+    invoke = AsyncMock(side_effect=[response, 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":
+        assert invoke.await_count == 1
+        assert result[0] is None
+        assert result[1] == "Confirmed home; recorded evidence."
+        assert result[3] is False
+    else:
+        assert invoke.await_count == 2
+        assert result[0]
+    assert any(isinstance(m, ToolMessage) and m.tool_call_id == "plan" for m in messages)
+
+
 @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")])