AGENT SKILL

addressing-review-feedback

How to work a task returned with review, QA or UAT findings. Use when a task is in need_revision or PR review comments are in your context.

License 许可
Apache-2.0
最近更新
2026-11-15
Stars
★ 122
Forks
18

01 前言 Front Matter

name
addressing-review-feedback
source
obra/superpowers (MIT), adapted
category
workflow
description
How to work a task returned with review, QA or UAT findings. Use when a task is in need_revision or PR review comments are in your context.

03 SKILL.md 正文

Addressing Review Feedback

Overview

A need_revision task is somebody else's evidence that something is wrong, not a request to improvise. Treat it the way you would a bug report: verify, fix at the root, prove it with a test, and answer every point — silence on one is how a task bounces a second time for something you could have agreed or disagreed with up front.

The Process

  1. Read everything before touching code. Every numbered point on the task, plus the PR review comments when the round happened there (list_task_comments, get_task_pull_request with include_diff: false when you don't need the diff itself). If a point is unclear, re-read the code it's about before acting on a guess. Points can be related — a root cause can explain two of them at once.
  2. Verify each claim against the code and, where you can, the running product — a reviewer can be wrong about where a line lives or what a function does. Classify each point: fix (the claim holds), disagree (with evidence — a file:line or a run that shows the current behavior is correct), or needs a human decision (a product call neither of you can settle from the code).
  3. Fix in order: blocking → simple → complex. Blocking = crash, security hole, or an acceptance criterion left unmet. Each fix gets its own guard test that fails before the fix and passes after (root-cause-debugging) — a fix with no failing-first test is not verified, it's hoped.
  4. Push back technically, never performatively. A point you disagree with gets a reasoned answer with evidence, not silent compliance and not silent skipping. Reply in the PR thread (comment_on_pull_request with reply_to_comment_id) when it came from a review comment — a new top-level comment leaves the reviewer's thread looking unanswered; otherwise add_task_comment for a task-comment point. No "You're absolutely right," no thanks, no preamble — state the finding.
  5. A YAGNI check when a reviewer asks you to "implement it properly" or "handle the general case": grep_code for actual callers first. Build the general version only if more than one caller needs it; otherwise say so and keep the fix scoped to what broke.
  6. Closing message: a point map. One line per numbered point — point → root cause → fix (file:line) → guard test, or point → why you disagree, with evidence. The architect's review of your revision reads this map against the diff; a revision with no map is a second bounce waiting to happen.

Red Flags

  • A partial fix — three of four points addressed and the fourth left unmentioned.
  • A fix with no guard test that failed first.
  • Weakening or deleting the test that caught the problem instead of fixing the code.
  • Silently agreeing with a wrong claim because arguing felt slower — a wrong "fix" based on a wrong claim bounces again, further down the line.
  • A closing message that says "addressed all feedback" without the per-point map.

04 同仓库更多 Skills