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

tests/unit/test_openai_responses.py

@@ -40,9 +40,9 @@ def text_output(text):
 
 @pytest.fixture
 def transport_model(monkeypatch):
-    import langchain_openai
+    from langchain_openai import ChatOpenAI
+    import openai
 
-    original = langchain_openai.ChatOpenAI
     requests = []
     outputs = []
 
@@ -52,16 +52,16 @@ def transport_model(monkeypatch):
             200, json=response_body(outputs.pop(0) if outputs else text_output("OK"))
         )
 
-    client = httpx.AsyncClient(transport=httpx.MockTransport(handle))
     monkeypatch.setattr(
-        langchain_openai,
-        "ChatOpenAI",
-        lambda **kwargs: original(**kwargs, http_async_client=client),
+        openai,
+        "DefaultAsyncHttpxClient",
+        lambda **kwargs: httpx.AsyncClient(transport=httpx.MockTransport(handle), **kwargs),
     )
     monkeypatch.setattr(settings, "OPENAI_API_KEY", SecretStr("test-secret"))
     monkeypatch.setattr(settings, "OPENAI_BASE_URL", "https://responses.example/v1")
     monkeypatch.delenv("ARTEMIS_FAKE_LLM", raising=False)
     model = ModelFactory.create_model(ModelEndpoint(model_name="test-model", timeout_seconds=23))
+    assert isinstance(model, ChatOpenAI)
     return model, requests, outputs
 
 
@@ -284,7 +284,10 @@ def test_video_text_fallback_extracts_responses_blocks():
 
 
 @pytest.mark.asyncio
-async def test_wrapper_structured_output_uses_responses_function_call(transport_model):
+@pytest.mark.parametrize("mode", ["parsed", "raw", "pipeline"])
+async def test_wrapper_structured_output_uses_responses_function_call(
+    transport_model, monkeypatch, mode
+):
     class Verdict(BaseModel):
         ok: bool
 
@@ -302,9 +305,53 @@ async def test_wrapper_structured_output_uses_responses_function_call(transport_
         ]
     )
     wrapper = RobustChatModelWrapper(model, endpoint=ModelEndpoint(model_name="test-model"))
-    result = await wrapper.with_structured_output(Verdict).ainvoke("Check")
-    assert result.ok is True
+    metered = []
+    monkeypatch.setattr(
+        "artemis.services.llm.record_llm_usage",
+        lambda engine, response, **kw: metered.append(response),
+    )
+    structured = wrapper.with_structured_output(Verdict, include_raw=mode == "raw")
+    if mode == "pipeline":
+        from langchain_core.runnables import RunnableLambda
+
+        result = await (structured | RunnableLambda(lambda value: value.ok)).ainvoke("Check")
+        assert result is True
+    else:
+        result = await structured.complete("Check")
+        assert (result["parsed"] if mode == "raw" else result).ok is True
+        if mode == "raw":
+            assert result["parsing_error"] is None
+            assert result["raw"].tool_calls
     payload = json.loads(requests[0].content)
     assert requests[0].url.path == "/v1/responses"
     assert payload["tools"][0]["name"] == "Verdict"
     assert payload.get("text", {}).get("format", {}).get("type") != "json_schema"
+    from artemis.llm.usage import usage_from_message
+
+    assert len(metered) == 1
+    assert usage_from_message(metered[0])["total_tokens"] == 13
+
+
+@pytest.mark.asyncio
+async def test_cache_key_is_stable_for_exploration_but_separates_roles(
+    transport_model, monkeypatch
+):
+    from types import SimpleNamespace
+
+    model, requests, _ = transport_model
+    engine = SimpleNamespace(current_session_id="run-1")
+    ctx = SimpleNamespace(data_engine=engine, exploration_id="exploration-1")
+    monkeypatch.setattr("artemis.services.llm.record_llm_usage", lambda *a, **k: None)
+    wrapper = RobustChatModelWrapper(
+        model, ctx, endpoint=ModelEndpoint(model_name="test-model"), role="operator"
+    )
+    await wrapper.ainvoke([HumanMessage(content="first")])
+    engine.current_session_id = "resumed-run-2"
+    await wrapper.ainvoke([HumanMessage(content="first"), HumanMessage(content="next")])
+    wrapper.role = "checker"
+    await wrapper.ainvoke([HumanMessage(content="verify")])
+    payloads = [json.loads(r.content) for r in requests]
+    assert payloads[0]["prompt_cache_key"] == payloads[1]["prompt_cache_key"]
+    assert payloads[1]["prompt_cache_key"] != payloads[2]["prompt_cache_key"]
+    assert "exploration-1" not in payloads[0]["prompt_cache_key"]
+    assert payloads[1]["input"][:1] == payloads[0]["input"]