// Evaluate pending (unsubmitted) review comments on the current branch's PR and, after user confirmation, address each in a separate sub-agent and separate commit.
| name | pr-pending-feedback |
| description | Evaluate pending (unsubmitted) review comments on the current branch's PR and, after user confirmation, address each in a separate sub-agent and separate commit. |
| user-invocable-only | true |
| allowed-tools | Bash, Read, Grep, Glob, Edit, Write, Agent |
The goal: look at every comment on the pending (unsubmitted) review the current user has drafted on the PR for the current branch, evaluate each one independently, recommend which to address (and which to skip), and — only after the user confirms — dispatch a sub-agent per comment to fix it, one commit per comment.
Pending reviews are private to their author until submitted, so this skill only sees comments authored by the currently authenticated gh user.
pr_number=$(gh pr view --json number --jq .number)
pr_url=$(gh pr view --json url --jq .url)
repo=$(gh repo view --json nameWithOwner --jq .nameWithOwner)
me=$(gh api user --jq .login)
If gh pr view fails, stop and tell the user there is no open PR for the current branch.
List reviews on the PR and find one in PENDING state authored by the current user:
gh api "repos/$repo/pulls/$pr_number/reviews" --jq \
".[] | select(.state==\"PENDING\" and .user.login==\"$me\") | {id, body, submitted_at, html_url}"
If there is no pending review, stop and tell the user there is no pending review on this PR for them. If there are multiple pending reviews (rare), list them and ask which to use.
Capture the id as review_id.
gh api "repos/$repo/pulls/$pr_number/reviews/$review_id/comments"
Each comment includes path, line (or original_line), body, diff_hunk, and html_url. Also note the pending review's top-level body from Step 2 — it may contain general feedback worth addressing too.
For each pending comment, read the referenced file(s) at the relevant lines to understand the current code. Then form an independent judgment on:
address, skip, or discuss.
address — clearly actionable and worth doing.skip — stale, already done, out of scope, or you disagree on reflection (explain briefly).discuss — needs user input before acting (ambiguous, subjective, or a design call).Since these are the user's own comments, frame skip/discuss recommendations as "you may want to reconsider this" rather than disagreeing with a third party.
Evaluate each comment on its own merits; do not let one comment's recommendation influence another's.
Output a compact markdown table or list. For each comment include:
#1, #2) used for confirmation.End with a prompt like:
Reply with which to address (e.g. "all", "1,3,4", "all except 2"). I'll run one sub-agent per comment and commit each fix separately.
Stop and wait for the user's reply. Do not proceed to Step 6 without explicit confirmation.
For each confirmed comment, dispatch a separate sub-agent via the Agent tool. Run them sequentially, not in parallel — each creates a commit on the same branch, and parallel edits would conflict.
Prompt for each sub-agent should include:
Example skeleton:
Agent({
subagent_type: "general-purpose",
description: "Address pending comment #<n>",
prompt: "Working directory: <cwd>.\n\nAddress this pending PR review comment (drafted by the user themself, not yet submitted):\n\n<comment body>\n\nFile: <path>:<line>\nComment URL: <url>\nPR: <pr_url>\n\nDiff hunk for context:\n<diffHunk>\n\nMake the smallest change that resolves the feedback. Run the relevant tests for the changed file. Create exactly one git commit with a concise message (no ticket prefix — this isn't the first commit of the branch, per the repo's git conventions). Do not push. Do not address any other feedback. Report back with the commit SHA and a one-line summary."
})
After each sub-agent returns, verify with git log --oneline -1 that a new commit landed, then proceed to the next.
If a sub-agent reports it could not address the comment (e.g. tests failed, unclear intent), stop and surface the problem to the user before continuing with remaining comments.
After all sub-agents finish, output a summary:
gh-authenticated user. Pending reviews by others are not visible via the API.CLAUDE.md (e.g. Chore: prefix for non-user-facing changes, no ticket code on subsequent commits, one thing per commit).Review and merge open Dependabot pull requests
Evaluate unresolved review comments on the current branch's PR and, after user confirmation, address each in a separate sub-agent and separate commit.
Walk through PR changes from the user's perspective. Traces each UI change through its full vertical slice — what the user sees, what it triggers, and how the server handles it. Use when asked to walk through what changed, review a PR, or summarize branch changes.
Run bin/claude-review --print and automatically fix all reported issues, committing each fix individually.
Amend a git commit further back in the history.
Summarize work done since the last standup across the user's configured repos — merged PRs, open PR reviews, PRs/commits authored, and a per-repo summary of code changes.