agentskills.codes
FE

feedback-handler

Resolves pull-request review comments on a Proxyfan PR systematically — works through the queue, addresses one item at a time, and marks it done. Use when the user asks to "address PR feedback", "go through review comments", or names a specific PR number and asks for changes.

Install

mkdir -p .claude/skills/feedback-handler && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/15063" && unzip -o skill.zip -d .claude/skills/feedback-handler && rm skill.zip

Installs to .claude/skills/feedback-handler

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Resolves pull-request review comments on a Proxyfan PR systematically — works through the queue, addresses one item at a time, and marks it done. Use when the user asks to "address PR feedback", "go through review comments", or names a specific PR number and asks for changes.
276 chars✓ has a “when” triggerlonger than Claude Code's old 250-char listing cap (fine on current versions)

About this skill

Feedback handler

Drives the loop that turns external GitHub review comments on a Proxyfan PR into a sequence of focused, verifiable code changes — across as many agent sessions as the work takes — without losing track of what is done.

When to invoke

The user has asked you to address feedback on a specific PR. They will name the PR by number, URL, or by being currently checked out on the topic branch. If the user has not named a PR, do not invoke this skill; ask them which PR first.

What you OWN end-to-end

  1. Hydrate the queue for the named PR from GitHub.
  2. Walk the queue in priority order (MAX → HIGH → MEDIUM → LOW, oldest first within each tier).
  3. For each item: read the comment, decide whether to apply, implement the change, validate locally, mark the comment resolved in the queue.
  4. Report progress between items so the user can interrupt the loop.

What you do NOT own

  • Deciding whether to ignore a comment — see the Disagreeing with a comment section; you must surface disagreement to the user before skipping.

You DO own the rest of the lifecycle: implementing the fix, committing, pushing, resolving the GitHub review threads, and merging the PR once the queue is clean and the checks are green. Hand the PR back unmerged only when something concrete blocks the merge (failing checks, conflicts you can't resolve, an unresolved disagreement).

Tooling contract

This skill is built on .tools/Get-PrCommentQueue.ps1. The queue is persisted under ~/.copilot/pr-queues/pr-<owner>-<repo>-<number>.json and is authoritative across sessions for the status of each comment (pending / in_progress / resolved). Comment bodies always come from GitHub (refresh before starting work).

The process is documented in PROCESS.md. Read it once at the start of the session and follow it step by step.

Disagreeing with a comment

You may not silently skip a review comment. If you disagree:

  1. Surface the comment to the user with your reasoning in two or three lines.
  2. Wait for explicit acknowledgement before marking it resolved.
  3. If acknowledged, mark it resolved with a note in the journal entry at end of session explaining the reasoning so a future agent does not redo the same analysis.

Exit

When the queue has zero pending items, summarise:

  • How many items were addressed and how many were marked resolved without changes.
  • Files touched (grouped by module).
  • Suggested validation commands (build, test, markdown gate as relevant).
  • The user's next step — typically git add -p followed by git commit, then git push.

Search skills

Search the agent skills registry