loop-address-pr-feedback（循环地址 pr 反馈）

Loop: Address PR Feedback

You are a PR author addressing review feedback. Reviewers give feedback, you fix and respond.

Core Concept

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  AI/Human       │────▶│  Claude fixes   │────▶│  Reply & push   │
│  reviews PR     │     │  valid issues   │     │  request re-rev │
└─────────────────┘     └─────────────────┘     └─────────────────┘
         │                                               │
         └─────────────── loop ──────────────────────────┘

The loop:

1. Gather — Collect review comments from all sources
2. Triage — Valid → fix, invalid → explain, unclear → ask
3. Fix — Make changes, commit, restack dependent PRs
4. Respond — Reply with fix SHA or rationale, resolve threads
5. Summon — Tag reviewers for re-review
6. Loop — Until all threads resolved and approvals obtained

Relationship to loop-codex-review

Same decision procedure, different interface.

Phase: Initialize

Detect the PR stack and get full context.

1. Get stack overview with full context: gt log # Full detail (recommended) gt ls # Short form (just branch names) gt log -s # Current stack only (if multiple stacks) gt log shows everything in one shot:

- PR numbers, titles, and status (Draft, Needs approvals, Approved) - Local changes needing submit - Commit history per branch - Graphite links

1. If not using Graphite, fall back to: gh pr list --author "@me" --state open --json number,headRefName,reviewDecision
2. Determine review order — address base PRs before children (changes propagate down via restack)

Args:

/loop-address-pr-feedback              # All PRs in current stack
/loop-address-pr-feedback --pr 123     # Single PR only
/loop-address-pr-feedback --skip 456   # Skip a PR (e.g., deferred)

Phase: Gather

Fetch all review feedback for each PR.

GitHub has three places reviews live:

# Issue comments
gh api repos/{owner}/{repo}/issues/{num}/comments \
  --jq '.[] | {id, user: .user.login, body}'

# Review comments
gh api repos/{owner}/{repo}/pulls/{num}/comments \
  --jq '.[] | {id, user: .user.login, path, line, body}'

# Review threads (for resolution status)
gh api graphql -f query='
  query($owner: String!, $repo: String!, $num: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $num) {
        reviewThreads(first: 50) {
          nodes { id isResolved comments(first: 1) { nodes { body } } }
        }
      }
    }
  }' -f owner=OWNER -f repo=REPO -F num=NUM

Phase: Triage

Categorize each issue before acting.

Considerations:

• Verify before fixing — especially AI reviews
• False positives signal unclear code — consider adding comments
• Check if later commits already addressed the issue
• Some "issues" are intentional design tradeoffs

Phase: Fix

Spawn agents to fix valid issues.

Task(
  description: "Fix: <issue summary>",
  prompt: "Fix the issue found by code review: ...",
  subagent_type: "general-purpose",
  run_in_background: true
)

Group by file or PR. Parallelize independent fixes.

Phase: Commit

git add -A && git commit -m "$(cat <<'EOF'
Address review feedback

- Fixed <issue 1>
- Fixed <issue 2>

Co-Authored-By: Claude <noreply@anthropic.com>
EOF
)"

Phase: Restack

If you modified a base PR, child PRs need restacking.

gt restack

Handle conflicts:

• If conflicts occur in skipped/deferred PRs, abort and push what succeeded
• Push each branch individually: git push origin <branch> --force-with-lease

Phase: Respond

Reply, react, and resolve.

# Reply to issue comment
gh api repos/{owner}/{repo}/issues/{num}/comments \
  -X POST -f body="Fixed in {sha}."

# Reply to review thread
gh api graphql -f query='
  mutation($threadId: ID!, $body: String!) {
    addPullRequestReviewThreadReply(input: {pullRequestReviewThreadId: $threadId, body: $body}) {
      comment { id }
    }
  }' -f threadId="PRRT_..." -f body="Fixed in {sha}."

# React to comment (👍 valid, 👎 invalid)
gh api repos/{owner}/{repo}/pulls/comments/{id}/reactions \
  -X POST -f content="+1"

# Resolve thread
gh api graphql -f query='
  mutation($threadId: ID!) {
    resolveReviewThread(input: {threadId: $threadId}) {
      thread { isResolved }
    }
  }' -f threadId="PRRT_..."

Phase: Summon

Request re-review by tagging the reviewer.

gh api repos/{owner}/{repo}/issues/{num}/comments \
  -X POST -f body="@{reviewer} Please re-review. Addressed:
- Fixed X
- Fixed Y"

Be specific — reviewers should know what to verify.

Phase: Wait

Reviews arrive asynchronously. Either:

1. Exit — Tell user to re-invoke when reviews arrive
2. Poll — Check for new comments periodically

Addressed all current feedback. Requested re-reviews.
Run /loop-address-pr-feedback again when new reviews arrive.

Fixed Point

A PR is done when:

• All review threads resolved
• No pending comments
• Required approvals obtained
• CI passing

A stack is done when all PRs are done.

┌───────────────────────────────────────────┐
│  STACK COMPLETE                           │
├───────────────────────────────────────────┤
│  PR #1: ✓ resolved, approved              │
│  PR #2: ✓ resolved, approved              │
│  PR #3: ⏸ skipped                         │
├───────────────────────────────────────────┤
│  Ready for merge queue: #1, #2            │
└───────────────────────────────────────────┘

State Tracking

Persist in task descriptions for compaction survival:

prs:
  - num: 1
    branch: feature-a
    status: done # addressing | waiting | done | skipped
  - num: 2
    branch: feature-b
    status: addressing
iteration: 1

Anti-patterns

• Fixing before triaging (some issues are false positives)
• Ignoring false positives (they signal unclear code)
• Responding without fixing (empty acknowledgment)
• Pushing without restacking (breaks child PRs)
• Resolving without replying (no audit trail)
• Summoning without summary (reviewer doesn't know what changed)
• Addressing children before parents (causes conflicts)

Resumption

After compaction:

1. TaskList — find tracking task
2. Read task description for state
3. gt ls — check stack state
4. Resume from appropriate phase

Enter loop-address-pr-feedback mode. Detect PR stack, gather feedback, triage, fix, respond, summon. Address base PRs first.