| name | experiment-review |
| description | Use when reviewing experiment topics, run.py files, experiment registries, GPU/JAX environment ownership, notebook artifacts, or experiment README/report readiness. |
Experiment Review
Tool Commands
Use the command packet before applying this skill's workflow:
python3 tools/agent_tools/skill_tool_commands.py show --skill experiment-review --format text
Execute the required and task-matching conditional commands that the packet prints.
- Read
agents/skills/experiment-review.md.
- Review from the registered experiment entry before reading implementation detail:
experiments/registry.toml -> topic README.md -> config.yaml -> run.py -> notebook.
- Confirm the topic entrypoint is not confused with setup tooling:
/usr/bin/python experiments/<topic>/run.py with no CLI options is the canonical run command.
- Confirm the topic code and checked-in config do not set GPU visibility, JAX
platform, allocator, preallocation,
max_workers: 1, or equivalent serial
throttles unless the user explicitly requested an environment-contract change.
- Confirm caller-owned environment is preserved by topic-created subprocesses:
notebook execution and workers should inherit
os.environ.copy() or default
inheritance instead of replacing GPU/JAX runtime settings.
- Confirm registered commands, when present, call the topic
run.py entrypoint
directly. Confirm the direct run
writes summary.json, cases.jsonl, config snapshot, case artifacts, and
notebook output under experiments/<topic>/result/<run_name>/.
- Confirm the notebook reads run artifacts and has a Japanese Markdown
explanation immediately above each visualization cell.
- Report findings first, grouped by severity. Treat registered commands that
bypass topic
run.py, environment hard-coding in topic code, or child
subprocess environment reset as fix-now findings.
