release: v0.0.19

CHANGELOG.md

@@ -7,6 +7,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.0.19] - 2026-05-25
+
+### Fixed
+
+- **`claude_code` adapter — assistant turns were being silently dropped from `agent*_traj.json`.** The session-JSONL parser (`parse_session_jsonl` in `src/cooperbench/agents/claude_code/parsers.py`) treated `message.role` as authoritative and rejected any event whose role wasn't in `{user, assistant, system}`. Recent claude-code session writers emit assistant turns with `message.role: None` — the role lives only in the top-level `event.type` — so every LLM turn (text, thinking, tool_use) got filtered out, leaving traj files that contained only `user` tool_result entries. Affected every `claude_code` run since the session-format shift (including all 0.0.17 / 0.0.18 trajectories on disk). Now falls back to `event.type` when `message.role` is missing; on a representative session (`anyhow_task/390/f1_f4/agent2_session.jsonl`) this recovers all 86 assistant events, taking the parsed trajectory from 43 messages (all user) to 129 (43 user + 86 assistant). The underlying `*_session.jsonl` and `*_stream.jsonl` files were always complete — only the derived `*_traj.json` was wrong, so historical runs can be re-parsed by calling `parse_session_jsonl` on the on-disk session file.
+
 ## [0.0.18] - 2026-05-25
 
 ### Removed

src/cooperbench/__about__.py

@@ -1,3 +1,3 @@
 """Version information for CooperBench."""
 
-__version__ = "0.0.18"
+__version__ = "0.0.19"

src/cooperbench/agents/claude_code/parsers.py

@@ -136,6 +136,12 @@ def parse_session_jsonl(text: str) -> list[dict[str, str]]:
 
     Returns a list of ``{"role": ..., "content": ...}`` dicts, sorted by
     ``timestamp``.  ``content`` is always a string.
+
+    Role resolution: prefer ``message.role`` when present, otherwise fall
+    back to ``event.type``.  Recent claude-code session writers emit
+    assistant turns with ``message.role: None`` (the role is only in the
+    top-level ``type`` field), so a strict role-validation check would
+    silently drop every LLM turn.
     """
     events: list[dict[str, Any]] = list(_iter_json_lines(text))
     events.sort(key=lambda e: e.get("timestamp") or "")
@@ -145,7 +151,7 @@ def parse_session_jsonl(text: str) -> list[dict[str, str]]:
         message = event.get("message")
         if not isinstance(message, dict):
             continue
-        role = message.get("role")
+        role = message.get("role") or event.get("type")
         if role not in {"user", "assistant", "system"}:
             continue
         content_text = _content_blocks_to_text(message.get("content"))