From c818a09a49c3bfb501ad8655011b75291c905901 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=99=88=E9=93=AD=E8=BD=A9?= <chenmingxuan@eapil.com>
Date: Tue, 29 Sep 2026 00:12:35 +0800
Subject: [PATCH] fix: align Responses messages and video exports

---
 apps/admin_console/services/report_export.py  |  8 +-
 artemis/agents/flash/runner.py                |  5 +-
 artemis/agents/flash/summarizer.py            |  4 +-
 artemis/llm/messages.py                       | 69 +++++++++++++++
 artemis/memory/chunking.py                    |  4 +-
 artemis/services/llm.py                       | 22 +++++
 tests/unit/admin_console/test_automation.py   | 37 ++++++++
 .../unit/agents/test_flash_step_summarizer.py | 31 ++++---
 tests/unit/test_llm_gateway.py                | 27 +++++-
 tests/unit/test_openai_responses.py           | 88 ++++++++++++++++++-
 10 files changed, 277 insertions(+), 18 deletions(-)
 create mode 100644 artemis/llm/messages.py

diff --git a/apps/admin_console/services/report_export.py b/apps/admin_console/services/report_export.py
index 1b58796..0bb2457 100644
--- a/apps/admin_console/services/report_export.py
+++ b/apps/admin_console/services/report_export.py
@@ -632,15 +632,17 @@ def session_videos(report: dict) -> list[tuple[Path, str]]:
     if not candidate:
         return []
     source = Path(str(candidate)).resolve()
-    session_root = (TRACES_PATH / identifier(report["id"])).resolve()
-    if not source.is_relative_to(session_root):
+    trace_root = TRACES_PATH.resolve()
+    # Workers save recordings in their finalized web_<time>_<id>_PASS/FAIL
+    # folder. The recording row already ties that path to this session.
+    if not source.is_relative_to(trace_root):
         return []
     # A resumed recording lives under resume_NNNN; the initial recording is
     # in its parent, with any later segments in sibling resume folders.
     directory = source.parent
     if directory.name.startswith("resume_"):
         directory = directory.parent
-    if not directory.is_dir() or not directory.is_relative_to(session_root):
+    if not directory.is_dir() or not directory.is_relative_to(trace_root):
         return []
     paths = [
         p
diff --git a/artemis/agents/flash/runner.py b/artemis/agents/flash/runner.py
index b336d85..61fba03 100644
--- a/artemis/agents/flash/runner.py
+++ b/artemis/agents/flash/runner.py
@@ -78,6 +78,7 @@ from artemis.controllers.unified_controller import UnifiedMobileController
 from artemis.data_engine.trace import trace
 from artemis.graph.perception import _check_injected_instruction_file
 from artemis.graph.state import State
+from artemis.llm.router import ModelEndpoint
 from artemis.llm.structured import ParseFailure, parse_structured
 from artemis.mcp.action_executor import McpActionExecutor
 from artemis.mcp.observation import observe
@@ -288,7 +289,9 @@ class FlashRunner:
             logger.warning(f"Failed to get operator LLM from config, using default: {e}")
 
             return RobustChatModelWrapper(
-                get_openai_llm(model_name=settings.OPENAI_MODEL), self.ctx
+                get_openai_llm(model_name=settings.OPENAI_MODEL),
+                self.ctx,
+                endpoint=ModelEndpoint(model_name=settings.OPENAI_MODEL),
             )
 
     def _render_system_prompt(self, tools_declaration: list) -> str:
diff --git a/artemis/agents/flash/summarizer.py b/artemis/agents/flash/summarizer.py
index 11ced42..2db238a 100644
--- a/artemis/agents/flash/summarizer.py
+++ b/artemis/agents/flash/summarizer.py
@@ -36,6 +36,7 @@ from jinja2 import Template
 from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage
 
 from artemis.context import ArtemisContext
+from artemis.llm.messages import adapt_responses_messages
 from artemis.memory.step_memory import JobKey, StepMemoryService
 from artemis.services.llm import RobustChatModelWrapper, get_openai_llm, get_llm
 from artemis.services.token_meter import record_llm_usage
@@ -462,7 +463,8 @@ class VisualStepSummarizer(StepMemoryService):
             ]
 
             response = await asyncio.wait_for(
-                self._llm.ainvoke(messages), timeout=self._timeout_seconds
+                self._llm.ainvoke(adapt_responses_messages(messages, self._model_name)),
+                timeout=self._timeout_seconds,
             )
             self._meter_lens_call(response)
             summary_raw = response.content if isinstance(response.content, str) else ""
diff --git a/artemis/llm/messages.py b/artemis/llm/messages.py
new file mode 100644
index 0000000..a3199c1
--- /dev/null
+++ b/artemis/llm/messages.py
@@ -0,0 +1,69 @@
+"""Adapt LangChain messages for Responses-compatible model families."""
+
+from collections.abc import Sequence
+from typing import Any
+
+from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage
+
+
+_USER_ROLE_MODELS = {"mimo-v2.6-pro", "mimo-v2.6-flash"}
+
+
+def adapt_responses_messages(
+    messages: Sequence[BaseMessage], model_name: str | None
+) -> Sequence[BaseMessage]:
+    """Send selected models user-role instructions without changing stored history.
+
+    Full Artemis requests to GPT models can fail with system-role input. The
+    equivalent user-role request works with GPT and the two tested Mimo models.
+    Other models keep their existing message roles.
+    """
+    if not model_name:
+        return messages
+    model = model_name.strip().lower()
+    if not (model.startswith("gpt") or model in _USER_ROLE_MODELS):
+        return messages
+    if not any(isinstance(message, SystemMessage) for message in messages):
+        return messages
+
+    adapted: list[BaseMessage] = []
+    pending_blocks: list[dict[str, Any]] = []
+    for message in messages:
+        if isinstance(message, SystemMessage):
+            content = message.content
+            if isinstance(content, str):
+                pending_blocks.append({"type": "text", "text": content})
+            else:
+                pending_blocks.extend(
+                    {"type": "text", "text": block} if isinstance(block, str) else block
+                    for block in content
+                )
+            continue
+
+        if pending_blocks:
+            if isinstance(message, HumanMessage):
+                content = message.content
+                blocks = (
+                    [{"type": "text", "text": content}]
+                    if isinstance(content, str)
+                    else list(content)
+                )
+                adapted.append(
+                    message.model_copy(
+                        update={
+                            "content": [
+                                *pending_blocks,
+                                *blocks,
+                            ]
+                        }
+                    )
+                )
+                pending_blocks = []
+                continue
+            adapted.append(HumanMessage(content=pending_blocks))
+            pending_blocks = []
+        adapted.append(message)
+
+    if pending_blocks:
+        adapted.append(HumanMessage(content=pending_blocks))
+    return adapted
diff --git a/artemis/memory/chunking.py b/artemis/memory/chunking.py
index 152c001..a61c226 100644
--- a/artemis/memory/chunking.py
+++ b/artemis/memory/chunking.py
@@ -46,6 +46,7 @@ from uuid import uuid4
 
 from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage
 
+from artemis.llm.messages import adapt_responses_messages
 from artemis.memory.step_memory import JobKey, StepLens, StepMemoryService
 from artemis.memory.transcript import format_session_offset
 from artemis.utils.logger import get_logger
@@ -503,7 +504,8 @@ class StepCapsuleLens(StepLens):
         return parsed
 
     async def _invoke(self, llm: Any, messages: list[BaseMessage], model_label: str):
-        response = await asyncio.wait_for(llm.ainvoke(messages), timeout=90.0)
+        request_messages = adapt_responses_messages(messages, model_label)
+        response = await asyncio.wait_for(llm.ainvoke(request_messages), timeout=90.0)
         # Raw-model bypass metering (gateway-wrapped models meter themselves);
         # capsule prompts are small and are kept out of the session meter's
         # last_prompt_tokens (the compaction base itself is operator-owned on
diff --git a/artemis/services/llm.py b/artemis/services/llm.py
index c1266d6..ff88c0f 100644
--- a/artemis/services/llm.py
+++ b/artemis/services/llm.py
@@ -54,6 +54,7 @@ from artemis.llm.reliability import (
     classify_failure,
     retry_policy_for,
 )
+from artemis.llm.messages import adapt_responses_messages
 from artemis.llm.router import ModelEndpoint, ModelFactory, ModelProvider
 from artemis.services.token_meter import record_llm_usage
 from artemis.llm.structured import (
@@ -568,6 +569,25 @@ class RobustChatModelWrapper:
             args, kwargs = _inject_parent_trace_id(trace_id, *args, **kwargs)
         return args, kwargs, trace_id
 
+    def _prepare_model_call(self, args: tuple, kwargs: dict) -> tuple[tuple, dict]:
+        model_name = self.endpoint.model_name if self.endpoint else None
+        if (
+            args
+            and isinstance(args[0], (list, tuple))
+            and all(isinstance(message, BaseMessage) for message in args[0])
+        ):
+            return (adapt_responses_messages(args[0], model_name), *args[1:]), kwargs
+        if (
+            "input" in kwargs
+            and isinstance(kwargs["input"], (list, tuple))
+            and all(isinstance(message, BaseMessage) for message in kwargs["input"])
+        ):
+            return args, {
+                **kwargs,
+                "input": adapt_responses_messages(kwargs["input"], model_name),
+            }
+        return args, kwargs
+
     def _meter_usage(self, response) -> None:
         """Record measured token usage for this call (M0 metering, best-effort)."""
         engine = self.ctx.data_engine if self.ctx else None
@@ -576,6 +596,7 @@ class RobustChatModelWrapper:
         record_llm_usage(engine, response, source=self._endpoint_key())
 
     async def ainvoke(self, *args, **kwargs):
+        args, kwargs = self._prepare_model_call(args, kwargs)
         args, kwargs, _ = self._traced_call(args, kwargs)
         response = await _run_with_recovery(
             functools.partial(self.base_model.ainvoke, *args, **kwargs),
@@ -597,6 +618,7 @@ class RobustChatModelWrapper:
         signals the UI to drop it, and retries the whole call per policy;
         partial or duplicated chunks can never reach message history.
         """
+        args, kwargs = self._prepare_model_call(args, kwargs)
         args, kwargs, _ = self._traced_call(args, kwargs)
         emit_deltas = bool(self.ctx and self.ctx.data_engine)
         response = await _run_with_recovery(
diff --git a/tests/unit/admin_console/test_automation.py b/tests/unit/admin_console/test_automation.py
index c37bf18..0a3285f 100644
--- a/tests/unit/admin_console/test_automation.py
+++ b/tests/unit/admin_console/test_automation.py
@@ -544,6 +544,43 @@ def test_task_material_zip_contains_ordered_evidence_and_reports(
     assert "1.1-all.zip" in response.headers["content-disposition"]
 
 
+def test_video_export_accepts_finalized_worker_trace_and_rejects_external_path(
+    isolated, tmp_path, monkeypatch
+):
+    job = isolated.jobs(service.create_batch(payload())["id"])[0]
+    session_id = job["session_id"]
+    recording_dir = reports.TRACES_PATH / f"web_123_{session_id[:8]}_PASS_2026-09-28"
+    recording_dir.mkdir(parents=True)
+    first = recording_dir / "recording.mp4"
+    first.write_bytes(b"first segment")
+    resumed = recording_dir / "resume_0001"
+    resumed.mkdir()
+    second = resumed / "recording.mp4"
+    second.write_bytes(b"second segment")
+
+    repository = Mock()
+    repository.get_video_recording_for_session.return_value = {"local_video_path": str(first)}
+    repository.get_session_by_id.return_value = {"video_filepath": str(first)}
+    monkeypatch.setattr(reports, "SessionRepository", lambda: repository)
+
+    archive_path = reports.export_session_media(session_id, "video")
+    with zipfile.ZipFile(archive_path) as archive:
+        names = archive.namelist()
+        assert names == [
+            "任务录屏/0001_web_123_" + session_id[:8] + "_PASS_2026-09-28_recording.mp4",
+            "任务录屏/0002_resume_0001_recording.mp4",
+            "文件清单.json",
+        ]
+        assert archive.read(names[0]) == b"first segment"
+        assert archive.read(names[1]) == b"second segment"
+
+    external = tmp_path / "outside" / "recording.mp4"
+    external.parent.mkdir()
+    external.write_bytes(b"unrelated")
+    repository.get_video_recording_for_session.return_value = {"local_video_path": str(external)}
+    assert reports.session_videos({"id": session_id}) == []
+
+
 def test_history_filters_all_persisted_rows_and_escapes_like(client, isolated):
     batch = service.create_batch(payload())
     first, second = isolated.jobs(batch["id"])
diff --git a/tests/unit/agents/test_flash_step_summarizer.py b/tests/unit/agents/test_flash_step_summarizer.py
index ae4a643..572eaff 100644
--- a/tests/unit/agents/test_flash_step_summarizer.py
+++ b/tests/unit/agents/test_flash_step_summarizer.py
@@ -39,6 +39,14 @@ def mock_context():
     return ctx
 
 
+def _user_role_prompt(messages):
+    """Return the instruction text and observation blocks sent to the model."""
+    assert len(messages) == 1
+    assert isinstance(messages[0], HumanMessage)
+    blocks = messages[0].content
+    return blocks[0]["text"], blocks[1:]
+
+
 @pytest.mark.asyncio
 async def test_summarizer_dispatch_and_caching(mock_context):
     """Verify non-blocking dispatch and summary generation."""
@@ -425,8 +433,8 @@ async def test_single_frame_dispatch_uses_single_image_prompt_variant(mock_conte
     await summarizer.flush()
 
     assert summarizer.has_summary(1)
-    system_text = str(captured[0][0].content)
-    human_text = str(captured[0][1].content)
+    system_text, human_blocks = _user_role_prompt(captured[0])
+    human_text = str(human_blocks)
     assert "Exactly ONE screenshot" in system_text
     assert "no independent post-action evidence" in system_text
     assert "DECISION FRAME" in human_text
@@ -464,8 +472,8 @@ async def test_dual_frame_dispatch_keeps_transition_prompt(mock_context):
     await summarizer.flush()
 
     assert summarizer.has_summary(1)
-    system_text = str(captured[0][0].content)
-    human_text = str(captured[0][1].content)
+    system_text, human_blocks = _user_role_prompt(captured[0])
+    human_text = str(human_blocks)
     assert "Before/After screenshots" in system_text
     assert "--- [1] BEFORE ACTION SCREEN (Action Marked Visually in Red) ---" in human_text
     assert "--- [2] AFTER ACTION SCREEN ---" in human_text
@@ -628,8 +636,8 @@ async def test_focus_block_and_provenance_marked_action_phrase_reach_the_lens(mo
     await summarizer.flush()
 
     assert summarizer.has_summary(3)
-    system_text = str(captured[0][0].content)
-    lead_text = captured[0][1].content[0]["text"]
+    system_text, human_blocks = _user_role_prompt(captured[0])
+    lead_text = human_blocks[0]["text"]
     assert "# OPERATOR FOCUS" in system_text
     assert "Absence is a fact" in system_text
     assert lead_text.startswith(
@@ -643,7 +651,7 @@ async def test_focus_block_and_provenance_marked_action_phrase_reach_the_lens(mo
     assert "User instruction at this step: Do not join any network yet." in lead_text
     assert "expect it to read On." in lead_text
     # The focus precedes the frames: context first, evidence second.
-    assert captured[0][1].content[1]["text"].startswith("--- [1] BEFORE ACTION SCREEN")
+    assert human_blocks[1]["text"].startswith("--- [1] BEFORE ACTION SCREEN")
 
 
 @pytest.mark.asyncio
@@ -672,7 +680,8 @@ async def test_dispatch_without_focus_carries_no_focus_block(mock_context):
     )
     await summarizer.flush()
 
-    lead_text = captured[0][1].content[0]["text"]
+    _, human_blocks = _user_role_prompt(captured[0])
+    lead_text = human_blocks[0]["text"]
     assert lead_text.startswith("Step 1 Physical Action: Tapped 'Search' at [500, 600]\n")
     assert "(self-described)" not in lead_text
     assert FOCUS_BLOCK_HEADER not in lead_text
@@ -706,7 +715,8 @@ async def test_single_frame_focus_label_forbids_describing_the_outcome(mock_cont
     )
     await summarizer.flush()
 
-    lead_text = captured[0][1].content[0]["text"]
+    _, human_blocks = _user_role_prompt(captured[0])
+    lead_text = human_blocks[0]["text"]
     assert "the expected outcome is NOT in this frame and must not be described" in lead_text
     assert "what screen change it expected" not in lead_text
     # The focus text is released once the summary landed, like the frames.
@@ -798,7 +808,8 @@ async def test_verbatim_argument_line_appears_only_when_the_phrase_folds_argumen
 
     lead_by_step = {}
     for messages in captured:
-        text = messages[1].content[0]["text"]
+        _, human_blocks = _user_role_prompt(messages)
+        text = human_blocks[0]["text"]
         lead_by_step[text.split(" ")[1]] = text
     launch = lead_by_step["1"]
     assert launch.startswith("Step 1 Physical Action: Launched app 'Settings'\nController Outcome:")
diff --git a/tests/unit/test_llm_gateway.py b/tests/unit/test_llm_gateway.py
index 6699d9f..ad890b8 100644
--- a/tests/unit/test_llm_gateway.py
+++ b/tests/unit/test_llm_gateway.py
@@ -2,10 +2,11 @@
 
 from types import SimpleNamespace
 
-from langchain_core.messages import AIMessage, AIMessageChunk
+from langchain_core.messages import AIMessage, AIMessageChunk, HumanMessage, SystemMessage
 import pytest
 
 from artemis.llm.reliability import LLMExhaustedError, LLMPermanentError
+from artemis.llm.router import ModelEndpoint
 from artemis.services import llm as llm_service
 from artemis.services.llm import (
     RobustChatModelWrapper,
@@ -54,6 +55,30 @@ async def test_mid_stream_failure_retries_without_duplicating_chunks():
     assert result.content == "Hello world"
 
 
+@pytest.mark.asyncio
+async def test_gpt_stream_uses_adapted_messages_without_mutating_history():
+    class RecordingStream:
+        messages = None
+
+        async def astream(self, messages, **kwargs):
+            self.messages = messages
+            yield AIMessageChunk(content="OK")
+
+    base = RecordingStream()
+    wrapper = RobustChatModelWrapper(
+        base, endpoint=ModelEndpoint(model_name="gpt-6-luna", max_retries=0)
+    )
+    original = [SystemMessage(content="Use tools"), HumanMessage(content="Inspect screen")]
+    result = await wrapper.complete(original)
+
+    assert result.content == "OK"
+    assert len(base.messages) == 1
+    assert isinstance(base.messages[0], HumanMessage)
+    assert base.messages[0].content[0]["text"] == "Use tools"
+    assert base.messages[0].content[1]["text"] == "Inspect screen"
+    assert isinstance(original[0], SystemMessage)
+
+
 @pytest.mark.asyncio
 async def test_auth_error_fails_fast_without_retry_or_pause(tmp_path):
     class AuthFails:
diff --git a/tests/unit/test_openai_responses.py b/tests/unit/test_openai_responses.py
index 8e351e9..21454e7 100644
--- a/tests/unit/test_openai_responses.py
+++ b/tests/unit/test_openai_responses.py
@@ -5,10 +5,11 @@ from unittest.mock import MagicMock
 
 import httpx
 import pytest
-from langchain_core.messages import HumanMessage, ToolMessage
+from langchain_core.messages import HumanMessage, SystemMessage, ToolMessage
 from pydantic import BaseModel, SecretStr, ValidationError
 
 from artemis.config.settings import settings
+from artemis.llm.messages import adapt_responses_messages
 from artemis.llm.router import ModelEndpoint, ModelFactory, ModelProvider
 from artemis.services.llm import RobustChatModelWrapper
 
@@ -85,6 +86,91 @@ async def test_text_and_image_use_responses_transport(transport_model):
     assert result.usage_metadata["total_tokens"] == 13
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("model_name", "reasoning_effort"),
+    [
+        ("mimo-v2.6-pro", "high"),
+        ("mimo-v2.6-flash", "high"),
+        ("gpt-6-luna", "max"),
+        ("gpt-6-luna", "high"),
+    ],
+)
+async def test_selected_models_put_instructions_in_user_input(
+    transport_model, model_name, reasoning_effort
+):
+    _, requests, _ = transport_model
+    endpoint = ModelEndpoint(
+        model_name=model_name,
+        reasoning_effort=reasoning_effort,
+        timeout_seconds=23,
+        max_retries=0,
+    )
+    model = ModelFactory.create_model(endpoint)
+    wrapper = RobustChatModelWrapper(model, endpoint=endpoint)
+    tool = {
+        "type": "function",
+        "function": {
+            "name": "inspect_screen",
+            "description": "Inspect the current screen",
+            "parameters": {"type": "object", "properties": {}},
+        },
+    }
+    await wrapper.bind_tools([tool]).ainvoke(
+        [
+            SystemMessage(content="Use the available tools."),
+            HumanMessage(
+                content=[
+                    {"type": "text", "text": "Inspect this screen."},
+                    {"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}},
+                ]
+            ),
+        ]
+    )
+
+    payload = json.loads(requests[-1].content)
+    assert payload["model"] == model_name
+    assert payload["reasoning"]["effort"] == reasoning_effort
+    assert payload["input"][0]["role"] == "user"
+    assert payload["input"][0]["content"][0]["text"] == "Use the available tools."
+    assert payload["input"][0]["content"][2]["type"] == "input_image"
+    assert payload["tools"][0]["name"] == "inspect_screen"
+    assert all(item.get("role") != "system" for item in payload["input"])
+
+
+@pytest.mark.parametrize("model_name", ["gpt-6-luna", "GPT-5.6-luna"])
+def test_gpt_message_adaptation_keeps_history_and_images(model_name):
+    image = {"type": "image_url", "image_url": {"url": "data:image/png;base64,AA=="}}
+    original = [
+        SystemMessage(content="First instruction"),
+        SystemMessage(content="Second instruction"),
+        HumanMessage(content=[{"type": "text", "text": "Observation"}, image]),
+        SystemMessage(content="Follow-up instruction"),
+    ]
+    adapted = adapt_responses_messages(original, model_name)
+
+    assert len(adapted) == 2
+    assert isinstance(adapted[0], HumanMessage)
+    assert adapted[0].content[0]["text"] == "First instruction"
+    assert adapted[0].content[1]["text"] == "Second instruction"
+    assert adapted[0].content[3] == image
+    assert isinstance(adapted[1], HumanMessage)
+    assert adapted[1].content == [{"type": "text", "text": "Follow-up instruction"}]
+    assert isinstance(original[0], SystemMessage)
+    assert len(original[2].content) == 2
+
+
+@pytest.mark.parametrize("model_name", ["mimo-v2.6-pro", "mimo-v2.6-flash"])
+def test_mimo_messages_use_same_user_role_format(model_name):
+    messages = [SystemMessage(content="Instructions"), HumanMessage(content="Task")]
+    adapted = adapt_responses_messages(messages, model_name)
+    assert len(adapted) == 1
+    assert isinstance(adapted[0], HumanMessage)
+    assert adapted[0].content[0]["text"] == "Instructions"
+    assert adapted[0].content[1]["text"] == "Task"
+    assert isinstance(messages[0], SystemMessage)
+
+
 @pytest.mark.asyncio
 async def test_max_reasoning_is_sent_unchanged_to_responses(transport_model):
     model, requests, _ = transport_model