Ascend/msverl-daily-regression-triage

MSVerl Daily Regression Triage

Use this skill when a fixed daily verl + MindSpeed training job has run and Codex needs to decide whether the result is healthy, whether there is a training failure or an accuracy regression, and which recent commit is the most likely cause.

Defaults

• Baseline comparison log: /home/st_daily_verl/msverl.log
• Training log pattern: /home/st_daily_verl/logs/msverl_YYYYMMDD.log
• verl repo: https://github.com/verl-project/verl.git on main
• MindSpeed repo: https://gitcode.com/Ascend/MindSpeed.git on master
• Cache root for temporary clones: /tmp/msverl-skill-cache
• Time window: from local previous day 00:00:00 to the task execution time

Hard Stop Rules

• Read the comparison log first.
• If it contains mean abs diff: and the parsed value is exactly 0, stop and report success.
• If it contains mean abs diff: and the value is non-zero, classify as accuracy_regression.
• If it contains error, please check log, classify as train_error.
• If the comparison log is ambiguous, report unknown and explain what evidence is missing before doing expensive work.

Workflow

1. Run parse_result_log.py on the comparison log.
2. Stop immediately on pass.
3. For train_error, run extract_failure_tail.py against the daily training log and keep only the final high-signal error block.
4. For accuracy_regression, use the parsed reward lists and mean abs diff as the primary evidence.
5. Sync lightweight local clones with sync_repos.py.
6. Collect recent commits with list_recent_commits.py for both repositories inside the default time window unless the user gives a different one.
7. Rank suspects with rank_candidate_commits.py.
8. Inspect diffs only for the top few commits when titles and touched files are not enough to explain a plausible fix direction.

Cost Controls

• Never load the whole training log unless the tail-based extractor fails twice.
• Start with the log tail only; prefer the last traceback or last ERROR block.
• Rank commits using title and touched files before reading diffs.
• Limit deep diff reading to the top 3 candidates per repository unless the evidence is still weak.

Expected Output

Return a compact report with:

• status: pass, train_error, accuracy_regression, or unknown
• time_window
• evidence_summary
• candidate_repo
• candidate_commits
• confidence: high, medium, or low
• fix_direction

When evidence is weak, say so clearly instead of forcing a single-commit claim.

References

• Run triage_msverl_regression.py for an end-to-end local workflow.
• Use parse_result_log.py and extract_failure_tail.py separately when validating logs by hand.
• Use list_recent_commits.py when you need a raw recent-commit inventory without ranking.