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

Use agent_comments only as context for what the agent said. Do not update guidance from agent_comments alone. Require human evidence from human_review_comments, human conversation_comments, or human-authored review bodies/comments.

Turn concrete feedback into repo-specific guidance:

1. Start from the specific human feedback.
2. Find repeated patterns or stable repo preferences.
3. Abstract the pattern into guidance that applies to future reviews.
4. Merge it into the most relevant existing section.
5. Keep the final wording concise.

Acceptable edits:

• add a bullet to an existing section
• add a small section when no existing section fits
• rewrite an existing rule to make it more accurate
• make no change when the evidence is too weak

Do not:

• paste raw JSON into skill files
• write a chronological summary of PR feedback
• add a rule for one reviewer's one-off preference
• add a rule based only on agent comments
• weaken correctness, security, or data-loss checks from one disagreement
• override the core review contract

Write Routing

Use each PR's review_type:

• code -> .github/skills/review-pr-repo/SKILL.md
• spec -> .github/skills/review-spec-repo/SKILL.md

Skip a file when that feedback type has no useful pattern.

Boundaries

Allowed write surface:

• .github/skills/review-pr-repo/
• .github/skills/review-spec-repo/

Forbidden write surface:

• .github/skills/review-pr/SKILL.md
• .github/skills/review-spec/SKILL.md
• other core skills or product code

Local companion skills may specialize repo preferences, but must not change the core review contract: output schema, severity labels, diff-line targeting, snapshot rules, validation rules, or safety rules.

Handoff

After writing output files, re-read the proposed skill content and keep it concise. Do not run Git commands, push branches, create PRs, edit workflow files, or edit .agents directly as part of this skill. The runner must apply output files and validate the write surface before publishing.