address-pr-feedback

model	opus
name	address-pr-feedback
description	Address pending PR review feedback through verified triage, small fix phases, implementation, validation, and evidence-backed replies. Manual invocation only.
disable-model-invocation	true

Address PR Feedback

Work through pending PR review feedback without blindly accepting or rejecting comments. This is a condensed research -> plan -> implement -> validate loop specialized for review feedback.

Load Rules

Read these first:

• ~/.claude/rules/pr-mode-readonly.md
• ~/.claude/rules/question-policy.md
• ~/.claude/rules/tdd-phase.md
• ~/.claude/rules/subagent-contract.md
• ~/.claude/rules/loop-detection.md
• ~/.claude/rules/no-outward-actions.md
• ~/.claude/rules/review-finding-format.md
• ~/.claude/rules/pr-cost-control.md

Use ~/.agents/rules/ when running through Codex.

Always read references/pushback-patterns.md.

Load targeted references as needed:

• references/feedback-triage.md before classifying comments.
• references/fix-planning.md before dispatching implementation phases.
• references/replies-and-publishing.md before drafting or publishing replies.

Flow

1. Resolve the PR from $ARGUMENTS, current branch, or gh pr status.
2. Fetch PR metadata, diff, reviews, inline comments, review bodies, and issue comments using filtered payloads only.
3. Build a pending-feedback index:
   • reviewer
   • location
   • comment text
   • comment ID and type
   • current addressed/resolved status
4. Fetch linked Linear requirements when present and build a requirements map for regression checks.
5. Investigate every pending comment in code context:
   • reproduce or trace the concern
   • verify suggested utilities or patterns
   • check docs for framework/library claims
   • classify with evidence
6. Classify each item:
   • Confirmed Fix
   • Partially Correct
   • Question Requiring Response
   • Valid Deferral
   • Disagree / Push Back
   • Already Addressed
7. Run adversarial challenge on classifications before acting.
8. Present triage and wait for user confirmation.
9. Plan fixes:
   • behavioral fixes -> implementation-executor TDD phases
   • non-behavioral edits -> quick-implement-agent direct-edit phases
10. Dispatch one phase at a time, re-verify each result, and apply loop detection.
11. Run final validation against tests, requirements map, and reviewer concerns.
12. Draft evidence-backed replies. Do not publish unless explicitly asked.

Boundaries

• Do not check out PR branches or treat local changed files as PR truth.
• Do not commit, push, publish replies, or mark threads resolved unless explicitly asked.
• Do not implement behavioral fixes in the main context; dispatch the executor.
• Do not defer fixes under roughly 20 lines unless there is a real scope or product reason.
• Do not push back without specific code, test, docs, or requirement evidence.

Output

Return pending-feedback triage, fixes completed, validation commands/results, unresolved items, and draft replies grouped by comment.