emrg: fix task_vibe_check LLM 400 from window-boundary orphan tool messages

Agent.md

@@ -118,7 +118,7 @@ Community needs voiced in HN agent-UI discussions map directly to EMRG's design:
 pkill -f "emrg.server"; rm -f ~/.emrg/emrgd.port; python -m emrg
 ```
 
-Python: `uv run pytest tests/ -v` (977) — import check: `uv run python -c "from emrg.client.app import run_client"`
+Python: `uv run pytest tests/ -v` (983) — import check: `uv run python -c "from emrg.client.app import run_client"`
 GUI: `cd emrg/gui && npm test` (260: 45 daemon_client + 19 conn-manager + 22 app-commands + 131 renderer smoke + 16 i18n + 7 integration + 3 commands + 8 build-config + 7 gui-state + 2 tool-group) — syntax: `node --check main.js preload.js daemon_client.js renderer/js/*.js`
 CI: `uv run pytest` (ubuntu + **windows-2025 matrix** — Windows pytest 回归在 PR CI 即失败，v0.2.29 教训 #725) + GUI tests + **actionlint workflow lint** (`rhysd/actionlint@v1.7.12` gate, #444 — workflow 解析错误在 PR CI 即失败，如 `if:` secrets 上下文)
 Re-trigger: `scripts/re-trigger-ci.sh [branch]` (workflow_dispatch, #527 — 替代空 commit 重触发：Actions outage 会整段丢弃 push 事件，dispatch 走 API 路径不受影响)

emrg/server/daemon.py

@@ -50,7 +50,7 @@
  ServerPong,
  TaskRequest,
 )
-from emrg.session import Session
+from emrg.session import Session, last_n_messages
 
 # ── 日志脱敏（rant 2026-08-06T10:21:26）────────────────────────────
 # tool call 参数可能含 api_key/token/authorization/password 等敏感字段，
@@ -1259,7 +1259,12 @@ async def _task_vibe_check(self, task_name: str, session_id: str, cwd: str,
  session = Session.load(session_id, Path(cwd))
  history = session.get_messages_for_llm()
  if history:
- messages.extend(history[-100:])
+ # Rant 2026-08-19T19:25:56 (root cause): slicing the
+ # validated list can orphan a leading role:"tool" message
+ # whose matching assistant(tool_calls) lies before the
+ # window → LLM 400 "tool must follow tool_calls". Use
+ # last_n_messages to drop window-boundary orphans.
+ messages.extend(last_n_messages(history, 100))
  except Exception:
  logger.warning(
  "task_vibe_check: session history load failed (%s/%s)",

emrg/session.py

@@ -569,6 +569,24 @@ def list_sessions(cwd: Path) -> list[dict]:
  return results
 
 
+def last_n_messages(messages: list[dict], n: int) -> list[dict]:
+ """Take the last ``n`` messages of a validated LLM message list.
+
+ The list must already be OpenAI-valid (e.g. produced by
+ ``_validate_tool_messages``). Slicing a valid list can still orphan a
+ leading ``role: "tool"`` message: its matching assistant message with
+ ``tool_calls`` lies just before the window boundary, so the API rejects
+ it with "Messages with role 'tool' must be a response to a preceding
+ message with 'tool_calls'" (observed on task_vibe_check, 2026-08-19).
+ Leading orphaned tool messages are dropped; the remaining window keeps
+ its original order.
+ """
+ window = messages[-n:]
+ while window and window[0].get("role") == "tool":
+ window.pop(0)
+ return window
+
+
 def _validate_tool_messages(messages: list[dict]) -> list[dict]:
  """Post-process messages to ensure OpenAI API validity.
 

tests/test_session.py

@@ -6,7 +6,12 @@
 
 import pytest
 
-from emrg.session import Session, _validate_tool_messages, generate_session_id
+from emrg.session import (
+ Session,
+ _validate_tool_messages,
+ generate_session_id,
+ last_n_messages,
+)
 
 
 class TestValidateToolMessages:
@@ -167,6 +172,82 @@ def test_mixed_block_one_valid_one_orphaned(self):
 # ── Session core operations ────────────────────────────────────
 
 
+class TestLastNMessages:
+ """Tests for last_n_messages — window slicing that drops boundary
+ orphans (rant 2026-08-19T19:25:56: task_vibe_check LLM 400)."""
+
+ def test_plain_window_unchanged(self):
+ """Normal window with no leading tool message passes through."""
+ msgs = [
+ {"role": "user", "content": "a"},
+ {"role": "assistant", "content": "b"},
+ {"role": "user", "content": "c"},
+ ]
+ assert last_n_messages(msgs, 2) == [
+ {"role": "assistant", "content": "b"},
+ {"role": "user", "content": "c"},
+ ]
+
+ def test_leading_tool_orphan_dropped(self):
+ """A window starting with role:'tool' (its assistant lies before the
+ boundary) has the orphan dropped so the API accepts the request."""
+ msgs = [
+ {"role": "user", "content": "start"},
+ {"role": "assistant", "content": None,
+ "tool_calls": [{"id": "x", "type": "function",
+ "function": {"name": "bash", "arguments": "{}"}}]},
+ {"role": "tool", "tool_call_id": "x", "content": "out"},
+ {"role": "user", "content": "next"},
+ {"role": "assistant", "content": "final"},
+ ]
+ window = last_n_messages(msgs, 3)
+ assert window == [
+ {"role": "user", "content": "next"},
+ {"role": "assistant", "content": "final"},
+ ]
+
+ def test_consecutive_leading_tool_orphans_all_dropped(self):
+ """Multiple consecutive leading tool orphans are all dropped."""
+ msgs = [
+ {"role": "assistant", "content": None,
+ "tool_calls": [{"id": "x", "type": "function",
+ "function": {"name": "bash", "arguments": "{}"}}]},
+ {"role": "tool", "tool_call_id": "x", "content": "1"},
+ {"role": "tool", "tool_call_id": "x", "content": "2"},
+ {"role": "user", "content": "after"},
+ ]
+ window = last_n_messages(msgs, 3)
+ assert [m["role"] for m in window] == ["user"]
+
+ def test_all_tool_window_empties(self):
+ """Window of only tool orphans returns empty list (caller guards)."""
+ msgs = [
+ {"role": "tool", "tool_call_id": "x", "content": "1"},
+ {"role": "tool", "tool_call_id": "y", "content": "2"},
+ ]
+ assert last_n_messages(msgs, 2) == []
+
+ def test_reproduced_real_session_case(self):
+ """Regression: a 172-message valid history whose [-100:] slice starts
+ with role:'tool' (real aitokenpool session, 2026-08-19) must no longer
+ produce a leading tool message."""
+ hist = []
+ for i in range(172):
+ if i % 2 == 0:
+ hist.append({"role": "user", "content": f"u{i}"})
+ else:
+ hist.append({"role": "assistant", "content": None,
+ "tool_calls": [{"id": f"t{i}", "type": "function",
+ "function": {"name": "bash", "arguments": "{}"}}]})
+ hist.append({"role": "tool", "tool_call_id": f"t{i}", "content": "r"})
+ # force the 100-window to start with a tool message: insert a trailing
+ # tool cluster right before position len-100 like the real history
+ hist = hist[:72] + [{"role": "tool", "tool_call_id": "orphan", "content": "o"}] + hist[72:]
+ window = last_n_messages(hist, 100)
+ assert len(window) <= 100
+ assert window[0]["role"] != "tool"
+
+
 class TestSessionCreate:
  """Tests for Session.create() and Session.create_with_id()."""
 

tests/test_ws_e2e.py

@@ -297,6 +297,74 @@ async def fake_chat(messages, tools=None):
  await cleanup()
  asyncio.run(_test())
 
+ def test_vibe_check_long_session_window_leading_tool_stripped(self):
+ """Rant 2026-08-19T19:25:56 (root cause): slicing a validated session
+ history to the last 100 messages can orphan a leading role:'tool'
+ message whose matching assistant(tool_calls) lies before the window →
+ the LLM rejects the request with 400 "tool must follow tool_calls"
+ (observed on every task_vibe_check for long sessions). The daemon
+ must strip window-boundary orphans before calling the LLM."""
+ async def _test():
+ with tempfile.TemporaryDirectory() as tmp:
+ tmp = Path(tmp)
+ sess_dir = tmp / ".emrg" / "sessions" / "emrg-evolution-emrg-task"
+ sess_dir.mkdir(parents=True, exist_ok=True)
+ lines = []
+ # 60 user → assistant(tool_calls) → tool_result triples:
+ # 180 llm messages → [-100:] starts at an index ≡ 2 (mod 3) = tool.
+ for i in range(60):
+ lines.append(json.dumps({
+ "type": "message", "role": "user", "content": f"cycle {i}",
+ }, ensure_ascii=False))
+ lines.append(json.dumps({
+ "type": "message", "role": "assistant", "content": None,
+ "tool_calls": [{
+ "id": f"x{i}", "type": "function",
+ "function": {"name": "bash", "arguments": "{}"},
+ }],
+ }, ensure_ascii=False))
+ lines.append(json.dumps({
+ "type": "tool_result", "tool_call_id": f"x{i}",
+ "content": "ok",
+ }, ensure_ascii=False))
+ (sess_dir / "history.jsonl").write_text("\n".join(lines) + "\n", encoding="utf-8")
+
+ server, _, cleanup = await _boot_server(tmp)
+ try:
+ seen = {}
+
+ async def fake_chat(messages, tools=None):
+ seen["messages"] = messages
+ return {"content": '{"meaningful": false, "recommend_slowdown": false, "reason": "nt", "done": ""}'}
+ server.llm.chat = fake_chat
+
+ ws = await connect_to_server()
+ try:
+ await ws.send(json.dumps({
+ "type": "task_vibe_check",
+ "session_id": "emrg-evolution-emrg-task",
+ "cwd": str(tmp),
+ "task_name": "emrg-task",
+ "prompt": "run cycle",
+ "completion_summary": "aux",
+ }, ensure_ascii=False))
+ frame = await asyncio.wait_for(ws.recv(), timeout=10)
+ data = json.loads(frame)
+ assert data.get("ok") is True, data
+ msgs = seen.get("messages", [])
+ assert msgs[0]["role"] == "system"
+ assert msgs[-1]["role"] == "user"
+ # no leading tool orphan after system; history portion
+ # must start with a non-tool role
+ first_hist = next(m for m in msgs[1:] if m["role"] != "system")
+ assert first_hist["role"] != "tool", [m["role"] for m in msgs[:6]]
+ assert len(msgs) <= 2 + 100, len(msgs)
+ finally:
+ await ws.close()
+ finally:
+ await cleanup()
+ asyncio.run(_test())
+
  def test_vibe_check_missing_done_field_is_compatible(self):
  """Old models / old parsing omit 'done' → empty string, no crash."""
  async def _test():