address-follow-up-comments

name

address-follow-up-comments

description

Handle GitHub PR follow-up review loops after automated or human comments arrive. Use when Codex needs to inspect PR comments/review threads, decide which suggestions are in scope and correct, implement fixes, update the PR body, trigger reviewers such as @codex review or /gemini review when requested, poll for follow-up comments, and repeat until there are no further actionable in-scope suggestions.

Address Follow-up Comments

Use this workflow for iterative PR review follow-up. Bias toward live PR state and local source over memory or prior summaries.

Core Workflow

1. Resolve the PR.

   • Use the user-provided PR URL/number when present.
   • Otherwise use the current branch with gh pr view --json number,url,headRefName,baseRefName.
   • Confirm gh auth status if any GitHub command fails.

2. Get enough context before editing.

   • Run git status -sb and inspect the current branch, remote, and local dirt.
   • Fetch thread-aware review context with scripts/fetch_pr_review_context.py.
   • Read the exact commented file ranges and surrounding implementation before deciding whether a suggestion is correct.
   • Treat unresolved or non-outdated review threads as candidates, but still verify against the current head commit because old threads may already be fixed.

3. Classify each suggestion.

   • Implement only suggestions that are both in scope for the PR and technically correct.
   • Ignore or explain suggestions that are stale, already fixed, out of scope, ambiguous, or would regress behavior.
   • If the user asks to fix all feedback, interpret that as all unresolved actionable in-scope feedback.

4. Implement and verify.

   • Keep edits tightly scoped to the confirmed feedback.
   • Run the smallest relevant checks, plus existing project checks when cheap.
   • Commit and push to the PR branch if this workflow is operating on a remote PR.

5. Update the PR body when useful.

   • Add a concise follow-up section with commit SHAs, review points addressed, and checks run.
   • Keep the original PR description intact unless it is stale or misleading.

6. Trigger follow-up reviewers only when requested.

   • Post exactly the trigger comments the user asks for, such as @codex review and /gemini review.
   • When multiple reviewers are requested, post each trigger as its own PR conversation comment, not as a combined message.
   • Do not reply to review threads, submit reviews, or mark threads resolved unless the user explicitly asks for those GitHub write actions.

7. Poll without noisy updates.

   • Immediately after posting triggers, record the trigger time and head commit.
   • Poll with scripts/fetch_pr_review_context.py until every requested reviewer has a new response for that recorded head.
   • Treat chatgpt-codex-connector conversation comments or reviews as Codex responses, and gemini-code-assist reviews as Gemini responses.
   • If a poll fails due to a transient GitHub/API/network error, retry rather than treating the round as complete.
   • During long polling, avoid interval status messages unless the user asks for status.

8. Iterate to a stop condition.

   • For each new review round, repeat context gathering and classify only current-head feedback plus unresolved non-outdated threads that are still applicable to the current code.
   • Treat old unresolved threads as stale when the exact issue is already fixed or the latest reviewer response for the current head says there are no comments to address.
   • If any in-scope, correct, implementable suggestion is found, implement it, verify it, commit and push it to the PR branch, update the PR body with a concise follow-up section, then re-trigger the same requested reviewers and poll again.
   • Stop only when all requested reviewers have responded to the latest head commit and there are no remaining in-scope, correct, implementable suggestions.

Script

Use scripts/fetch_pr_review_context.py to fetch PR comments, reviews, and inline review threads with isResolved and isOutdated:

uv run python3 ~/.codex/skills/address-follow-up-comments/scripts/fetch_pr_review_context.py
uv run python3 ~/.codex/skills/address-follow-up-comments/scripts/fetch_pr_review_context.py --repo OWNER/REPO --pr 123

The script shells out to gh api graphql, so it requires an authenticated GitHub CLI. If the current repository's Python project dependencies make plain uv run fail, use uv run --no-project python3 ... with the same script arguments.

Final Response

Summarize:

• PR URL, branch, and latest commit.
• Which feedback was implemented, skipped, or found stale.
• Checks run.
• Latest reviewer status and whether the stop condition was reached.