Write conversation logs from the in-loop benchmark runner

CHANGELOG.md

@@ -22,6 +22,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   pinned b9066 builds predate MTP, upstream PR #22673). Example profile:
   `scripts/settings_local_llm_qwen38_27b.yaml` (unvalidated; numbers in it are
   labelled sourced / estimated).
+- **fix(benchmarks): the in-loop runner now writes conversation logs** — with
+  `log_conversations: true` (the setting the CLI already honoured) `experiments/swebench_lite/
+  run_in_loop.py` created no logger, so every benchmark trajectory was thrown away. It now writes
+  `<global_dir>/training/<session>.conversation.jsonl` including the terminal `session_end` record
+  (exit reason and code), so benchmark runs build the trajectory corpus any later tuning needs.
+
 - **feat(llm): Qwen3.5+ thinking control** — `qwen3.5`..`qwen3.9` models get
   `chat_template_kwargs` (`enable_thinking`, `reasoning_effort`) built from
   `reasoning_effort` (`none|low|medium|high`) or `thinking_budget` (mapped to

experiments/swebench_lite/run_in_loop.py

@@ -63,7 +63,7 @@ async def _run_one_async(
     # Imports kept local so this module stays cheap to import from run.py
     from godspeed.agent.conversation import Conversation
     from godspeed.agent.loop import agent_loop
-    from godspeed.agent.result import AgentMetrics, ExitReason
+    from godspeed.agent.result import EXIT_REASON_TO_CODE, AgentMetrics, ExitCode, ExitReason
     from godspeed.audit.trail import AuditTrail
     from godspeed.cli import _ensure_ollama
     from godspeed.config import GodspeedSettings
@@ -175,8 +175,19 @@ def evaluate(self, tool_call: Any) -> PermissionDecision:
         llm_client=llm_client,  # type: ignore[arg-type]
     )
 
+    # Benchmark runs generate exactly the trajectories a tuning corpus needs; honour the same
+    # log_conversations switch the CLI does (this runner used to drop them silently).
+    conversation_logger = None
+    if settings.log_conversations:
+        from godspeed.training.conversation_logger import ConversationLogger
+
+        conversation_logger = ConversationLogger(
+            session_id=session_id, output_dir=settings.global_dir / "training"
+        )
+
     conversation = Conversation(
         system_prompt=system_prompt,
+        conversation_logger=conversation_logger,
         model=effective_model,
         max_tokens=settings.max_context_tokens,
         compaction_threshold=settings.compaction_threshold,
@@ -234,6 +245,22 @@ def on_tool_call(name: str, _args: dict) -> None:
         timed_out = True
         final_text = f"(session exceeded wall-clock timeout of {timeout_s}s)"
         metrics.finalize(ExitReason.TIMEOUT)
+    except BaseException:
+        if conversation_logger is not None:
+            conversation_logger.close()
+        raise
+
+    if conversation_logger is not None:
+        conversation_logger.log_session_end(
+            exit_reason=metrics.exit_reason.value,
+            exit_code=int(EXIT_REASON_TO_CODE.get(metrics.exit_reason, ExitCode.SUCCESS)),
+            iterations_used=metrics.iterations_used,
+            tool_call_count=metrics.tool_call_count,
+            tool_error_count=metrics.tool_error_count,
+            duration_seconds=round(metrics.duration_seconds, 3),
+            cost_usd=llm_client.total_cost_usd,
+        )
+        conversation_logger.close()
 
     audit_trail.record(
         event_type="session_end",

tests/test_run_in_loop_wiring.py

@@ -64,3 +64,55 @@ async def _fake_agent_loop(**_kwargs: Any) -> str:
     assert payload["final_text"] == "done"
     assert captured["reasoning_effort"] == "medium"
     assert captured["model"] == "openai/qwen3.8-27b"
+
+
+def _run_with_settings(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path, extra_yaml: str, stop_with: str = "done"
+) -> Path:
+    """Run one in-loop session with a stub agent loop; return the global settings dir."""
+
+    async def _fake_agent_loop(**_kwargs: Any) -> str:
+        return stop_with
+
+    monkeypatch.setattr("godspeed.agent.loop.agent_loop", _fake_agent_loop)
+    global_dir = tmp_path / "global"
+    global_dir.mkdir()
+    (global_dir / "settings.yaml").write_text(
+        f"global_dir: {global_dir.as_posix()}\n{extra_yaml}", encoding="utf-8"
+    )
+    monkeypatch.setattr("godspeed.config.DEFAULT_GLOBAL_DIR", global_dir)
+    asyncio.run(
+        run_in_loop._run_one_async(
+            model="openai/qwen3.8-27b",
+            prompt="fix the bug",
+            project_dir=tmp_path,
+            instance_id="acme__widget-1",
+            split="dev",
+            timeout_s=30,
+            verify_workdir=tmp_path,
+            max_iterations=1,
+        )
+    )
+    return global_dir
+
+
+def test_conversation_is_logged_when_enabled(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+    """Benchmark runs produce the trajectories a tuning corpus needs; they must not be dropped."""
+    import json
+
+    global_dir = _run_with_settings(monkeypatch, tmp_path, "log_conversations: true\n")
+    files = list((global_dir / "training").glob("*.conversation.jsonl"))
+    assert len(files) == 1
+    records = [json.loads(line) for line in files[0].read_text(encoding="utf-8").splitlines()]
+    assert records[0]["role"] == "system"
+    end = [r for r in records if r.get("role") == "session_end" or "exit_reason" in r]
+    assert end, records
+    assert end[-1]["exit_reason"] == "stopped"
+    assert end[-1]["exit_code"] == 0
+
+
+def test_no_conversation_log_when_disabled(monkeypatch: pytest.MonkeyPatch, tmp_path: Path) -> None:
+    global_dir = _run_with_settings(monkeypatch, tmp_path, "log_conversations: false\n")
+    assert not list((global_dir / "training").glob("*.conversation.jsonl"))