Context

• Current branch: !git branch --show-current

• PR info: !gh pr view --json number,title,url,state,author 2>/dev/null || echo "none"

• Repository: !gh repo view --json nameWithOwner --jq '.nameWithOwner' 2>/dev/null

Your task

Collect, classify, and plan responses to GitHub PR review comments.

Step 1: Detect PR

• Check the PR info from the context above

• If the user passed a PR number as an argument, use that instead

• If no PR is found, inform the user and stop

Extract {owner} , {repo} , and {number} from the PR info for subsequent API calls.

Step 2: Collect review data

Fetch review feedback and commit history using gh api . Run all four calls in parallel. Each REST call MUST include --paginate to retrieve all pages.

• Reviews — top-level verdicts (APPROVED, CHANGES_REQUESTED, etc.) gh api --paginate repos/{owner}/{repo}/pulls/{number}/reviews
  --jq '[.[] | {id, user: .user.login, state, body, submitted_at}]'

• Inline comments — file/line-level code comments gh api --paginate repos/{owner}/{repo}/pulls/{number}/comments
  --jq '[.[] | {id, user: .user.login, body, path, line, diff_hunk, created_at, in_reply_to_id}]'

• Conversation comments — general PR discussion gh api --paginate repos/{owner}/{repo}/issues/{number}/comments
  --jq '[.[] | {id, user: .user.login, body, created_at}]'

• Commits — timestamps for filtering gh api --paginate repos/{owner}/{repo}/pulls/{number}/commits
  --jq '[.[] | {sha: .sha, date: .commit.committer.date}]'

Step 3: Filter by timestamp

Determine the latest commit timestamp from the commits list. Classify each comment:

• Needs action: created_at (or submitted_at ) is after the latest commit — not yet addressed

• Likely addressed: timestamp is before the latest commit — a subsequent commit likely handled it

Carry forward only "Needs action" comments. Record the "Likely addressed" count for the summary.

Step 4: Cross-reference resolution status

Fetch thread resolution via GraphQL:

gh api graphql -f query=' query($owner: String!, $repo: String!, $number: Int!) { repository(owner: $owner, name: $repo) { pullRequest(number: $number) { reviewThreads(first: 100) { nodes { isResolved comments(first: 1) { nodes { databaseId } } } } } } }' -f owner='{owner}' -f repo='{repo}' -F number='{number}'

A comment requires action only when both conditions are true:

• Timestamp says "Needs action" (Step 3)

• Thread is unresolved (Step 4)

Early exit: If zero comments remain after this filter, inform the user that all feedback is addressed and stop.

Step 5: Classify actionable comments

Classify each remaining comment as one of:

Category Criterion Plan treatment

Actionable feedback Requires a code change Plan the change

Question Requires a reply, no code change Draft reply text

Praise / acknowledgment No action needed Omit from output

Step 6: Present summary

Display a filtering funnel followed by the actionable detail:

• Funnel: Total comments → noise filtered → likely addressed → resolved → actionable

• Actionable detail, grouped by file:

• File path

• Each comment: reviewer, brief excerpt, category (actionable / question)

Step 7: Read source files and enter plan mode

Before entering plan mode, read each source file referenced by inline comments using the Read tool. The diff_hunk shows only diff context — you MUST read the current file content to plan accurate changes.

Then call EnterPlanMode .

In plan mode, create a plan that:

• Addresses each actionable comment, prioritized by file

• Specifies concrete changes (what to modify, where) referencing current source code

• Uses file_path:line_number format for all source references

• Groups related comments that can be addressed together

• Marks "question" items separately with suggested reply text