update-pr-review

update-pr-review

Use this skill to turn human feedback on bot PR reviews into concise updates to repo-local review companion skills.

This skill owns only the self-evolution logic: how to turn aggregated human feedback into local review guidance. The GitHub Actions runner owns data collection, write-surface validation, commits, pushes, and PR creation. When run inside GitHub Actions, .agents may be read-only. Write proposed changes only to update-pr-review-output/; the runner applies them.

Workflow

1. Read the aggregated feedback JSON provided by the runner.
2. Identify repeated human-feedback patterns or stable repo preferences.
3. Convert those patterns into concise repo-specific guidance.
4. Write proposed local companion skill content to update-pr-review-output/.
5. Stop; the runner validates and publishes the result.

Output Contract

Always write update-pr-review-output/status.json:

{
  "status": "changed",
  "reason": "Brief evidence summary.",
  "updated_files": [".github/skills/review-pr-repo/SKILL.md"]
}

Allowed statuses:

• changed when guidance should be updated
• no_change when evidence is insufficient or already covered
• error when feedback cannot be interpreted safely

Use no_change when there is no human evidence. Agent-only feedback is not enough to update guidance.

For changed, write the complete replacement content for each updated file:

• update-pr-review-output/review-pr-repo/SKILL.md
• update-pr-review-output/review-spec-repo/SKILL.md

Do not edit .agents directly.

Learn And Edit

Read the JSON and look for patterns worth adding to local review rules:

• humans clearly said an agent comment was wrong
• the finding was directionally right, but severity, scope, or line target was wrong
• the comment was not actionable
• reviewers repeatedly emphasized a repo-specific check
• human-only review threads reveal stable repo preferences
• a pattern belongs in the top-level summary instead of inline comments