PR Sweep

Triage unresolved GitHub PR review threads: auto-resolve comments already addressed by current code, reply with context, and create actionable tasks for remaining items.

Prerequisites

Verify before starting:

A PR exists for the current branch
The gh CLI is authenticated

gh pr view --json number,url,title,baseRefName,headRefName 2>/dev/null || echo "NO_PR"

If no PR exists, report this and stop.

Workflow

Step 1 - Gather Context

Collect repository and PR metadata:

gh repo view --json nameWithOwner -q .nameWithOwner
gh pr view --json number,url,title,baseRefName,headRefName

Extract owner, repo, and pr_number from the output.

Step 2 - Fetch Unresolved Review Threads

Run this GraphQL query and pipe through jq to return only the fields needed for triage — keeps response lean and avoids flooding context with metadata:

gh api graphql -F owner='OWNER' -F repo='REPO' -F pr=PR_NUMBER \
  -f query="$(cat << 'GRAPHQL'
query($owner: String!, $repo: String!, $pr: Int!) {
  repository(owner: $owner, name: $repo) {
    pullRequest(number: $pr) {
      reviewThreads(first: 100) {
        nodes {
          id
          isResolved
          path
          line
          comments(first: 1) {
            nodes {
              body
              author { login }
            }
          }
        }
      }
    }
  }
}
GRAPHQL
)" | jq '[.data.repository.pullRequest.reviewThreads.nodes[]
  | select(.isResolved == false)
  | {id, path, line, author: .comments.nodes[0].author.login, comment: .comments.nodes[0].body}]'

The result is a compact array — one object per unresolved thread with only id, path, line, author, and comment.

Step 3 - Analyze Each Unresolved Thread

For each unresolved thread:

Read the file at path with context around line
Check staged/unstaged changes for the file: git diff <baseRefName>...HEAD -- <path> (where baseRefName is from Step 1) and git diff --cached <path>
Determine if addressed by checking whether:
- Current code implements the suggested fix
- Code was refactored in a way that makes the suggestion moot
- Style, naming, or structural issues are now corrected
- Missing tests or documentation were added

Resolution criteria: only mark as addressed when the feedback is fully resolved. When in doubt, leave unresolved and add to the task list.

Step 4 - Resolve Addressed Comments

For each comment determined to be addressed:

Reply to the thread explaining what was done:

gh api graphql \
  -f query="$(cat << 'GRAPHQL'
mutation($threadId: ID!, $body: String!) {
  addPullRequestReviewThreadReply(input: {
    pullRequestReviewThreadId: $threadId,
    body: $body
  }) {
    comment { id }
  }
}
GRAPHQL
)" -F threadId='THREAD_NODE_ID' -F body='REPLY_BODY' \
  | jq -r '.data.addPullRequestReviewThreadReply.comment.id'

Resolve the thread:

gh api graphql \
  -f query="$(cat << 'GRAPHQL'
mutation($threadId: ID!) {
  resolveReviewThread(input: { threadId: $threadId }) {
    thread { isResolved }
  }
}
GRAPHQL
)" -F threadId='THREAD_ID' \
  | jq -r '.data.resolveReviewThread.thread.isResolved'

Keep replies concise - one or two sentences explaining the fix.

Step 5 - Plan Remaining Comments

For comments still needing work:

Group by file
Create a task per unresolved item using TaskCreate
Include the comment text, author, file path, and line number in the task description
Add a suggested implementation approach based on the comment feedback and current code structure

Output Format

## Auto-Resolved Comments (X)

### [file:line] - Thread ID
**Comment:** <summary>
**Reply:** <what was posted>
**Resolved because:** <reasoning>

---

## Remaining Comments (Y)

### [file:line] - Thread ID
**Comment:** <full comment text>
**Author:** <username>
**Suggested fix:** <implementation suggestion>

---

## Created Tasks
<task list summary>

Important Notes

Only resolve comments where the feedback is fully addressed
Always reply before resolving so reviewers see the context
For comments about missing tests or documentation, verify those were actually added
For refactoring suggestions, confirm the new code achieves the same goal
Batch-resolve carefully: a false positive erodes reviewer trust

pr-sweep