fix: improve cross-device evidence, request diagnostics and prompt efficiency

tests/unit/agents/test_operator_transcript.py

@@ -141,6 +141,74 @@ def _transcript_state(**overrides):
     return state
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("exploration", [False, True])
+@pytest.mark.parametrize("transcript", [False, True])
+async def test_plugin_result_is_delivered_with_pc_image_and_separate_phone_observation(
+    tmp_path, exploration, transcript, monkeypatch
+):
+    from langchain_openai import ChatOpenAI
+    from PIL import Image
+
+    ctx = _transcript_ctx()
+    ctx.exploration_id = "explore" if exploration else None
+    monkeypatch.setattr(
+        "artemis.exploration.runtime.exploration_messages",
+        lambda _: [SystemMessage(content="Explore"), HumanMessage(content="Guidance")],
+    )
+    ctx.data_engine = MagicMock()
+    ctx.data_engine.base_dir = str(tmp_path)
+    pc = tmp_path / "pc.jpg"
+    Image.new("RGB", (200, 100), "blue").save(pc)
+    ctx.data_engine.get_image_path.return_value = pc
+    node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=transcript))
+    ledger = node._ensure_transcript_ledger()
+    tool = "plug_test_get_app_state"
+    ledger.stage_turn(
+        [
+            AIMessage(content="", tool_calls=[{"name": tool, "args": {}, "id": "pc_call"}]),
+            ToolMessage(content="Action Recorded", tool_call_id="pc_call"),
+        ]
+    )
+    report = {
+        "status": "dispatched",
+        "execution": [
+            {
+                "action": tool,
+                "tool_call_id": "pc_call",
+                "plugin_id": "test",
+                "plugin_status": "success",
+                "result": "PC Address bar",
+                "post_image_name": "pc",
+            }
+        ],
+    }
+    state = _transcript_state(structured_decisions="[]", last_execution_result=report)
+    messages = await node._build_prompt(state, SCREENSHOT_B64, [], "PHONE_ELEMENTS", 2, [], "")
+    if transcript:
+        output = next(m for m in messages if isinstance(m, ToolMessage))
+        assert output.tool_call_id == "pc_call" and "PC Address bar" in str(output.content)
+    else:
+        assert "PC Address bar" in str(messages)
+    assert base64.b64encode(pc.read_bytes()).decode() in str(messages)
+    assert SCREENSHOT_B64 not in str(messages) and "PHONE_ELEMENTS" in str(messages)
+    wire = ChatOpenAI(model="fixture", api_key="fixture", use_responses_api=True, store=False)
+    payload = wire._get_request_payload(messages)["input"]
+    assert len([x for x in payload if x.get("type") == "function_call_output"]) == int(transcript)
+    guided = await node._build_prompt(
+        _transcript_state(
+            structured_decisions="[]", last_execution_result=report, injected_instruction="回到手机"
+        ),
+        SCREENSHOT_B64,
+        [],
+        "PHONE_ELEMENTS",
+        3,
+        [],
+        "",
+    )
+    assert SCREENSHOT_B64 in str(guided)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
     "exploration,plugins", [(False, False), (True, False), (False, True), (True, True)]
@@ -415,3 +483,38 @@ async def test_flag_off_keeps_legacy_two_message_build_and_no_ledger():
     assert isinstance(captured[0][0], SystemMessage)
     assert isinstance(captured[0][1], HumanMessage)
     assert ctx.transcript_ledger is None
+
+
+@pytest.mark.asyncio
+async def test_exploration_guidance_appends_without_rewriting_static_prefix(monkeypatch):
+    ctx = _transcript_ctx()
+    ctx.exploration_id = "explore"
+    guidance = ["initial guidance"]
+    monkeypatch.setattr(
+        "artemis.exploration.runtime.exploration_messages",
+        lambda _: [SystemMessage(content="rules"), HumanMessage(content=guidance[0])],
+    )
+    node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=True))
+    state = _transcript_state()
+    first = await node._build_prompt_transcript(state, "", "", 1, [], "")
+    prefix = first[0].content
+    assert "initial guidance" not in prefix
+    assert "initial guidance" in str(first[-1].content)
+    node._stage_transcript_turn(first)
+    second = await node._build_prompt_transcript(state, "", "", 2, [], "")
+    assert second[0].content == prefix
+    assert "initial guidance" not in str(second[-1].content)
+    node._stage_transcript_turn(second)
+    guidance[0] = "new direction"
+    third = await node._build_prompt_transcript(state, "", "", 3, [], "")
+    assert third[0].content == prefix
+    assert "new direction" in str(third[-1].content)
+    # An interrupted/uncommitted build must not acknowledge away the guidance.
+    retried = await node._build_prompt_transcript(state, "", "", 3, [], "")
+    assert "new direction" in str(retried[-1].content)
+    node._stage_transcript_turn(retried)
+    await node._build_prompt_transcript(state, "", "", 4, [], "")
+    ctx.transcript_ledger.freeze_turns(3, [HumanMessage(content="compressed history")])
+    compacted = await node._build_prompt_transcript(state, "", "", 4, [], "")
+    assert compacted[0].content == prefix
+    assert "new direction" in str(compacted[-1].content)