Agent-Skills.md

Agent Skills: polish-implementation

Iterative code review loop against a GitHub PR — atk-code-review posts findings as self-review threads, this skill fixes, replies, and resolves them, up to 10 passes.

UncategorizedID: rstacruz/agentic-toolkit/polish-implementation

Repository

rstacruz/agentic-toolkit
rstacruzLicense: MIT
132

Install this agent skill to your local

pnpm dlx add-skill https://github.com/rstacruz/agentic-toolkit/tree/HEAD/skills/polish-implementation

Skill Files

Browse the full folder contents for polish-implementation.

Download Skill

Loading file tree…

skills/polish-implementation/SKILL.md

Skill Metadata

Name
polish-implementation
Description
Iterative code review loop against a GitHub PR — atk-code-review posts findings as self-review threads, this skill fixes, replies, and resolves them, up to 10 passes.

Run an iterative review loop on an implementation, fixing and replying to self-review threads on GitHub each pass.

This skill pushes commits and posts/resolves GitHub PR comments autonomously across passes — same posture as atk-pr-autofix. Invoking it is the opt-in; it won't ask before each push or comment.

Workflow

  1. Ensure the PR is up to date.

    • Commit any staged changes, then push: git push (first run on a new branch: git push -u origin HEAD).
    • Check for an existing PR: gh pr view --json number -q .number 2>/dev/null.
    • If none: gh pr create --draft.
    • The PR is now the change set — atk-code-review reviews it directly, no separate diff-scoping step needed here.

  2. Subagent review. Spawn a subagent against the PR from step 1 and the plan (if available), using the atk-code-review skill. It gathers PR context, dedups against existing threads (any author), and posts a single batched self-review to the PR — returning the new thread IDs it created (empty if nothing new was found).

  3. If no new threads: done — go to step 6 (report).

  4. Triage and fix. For each thread ID returned by step 2:

    • Triage using atk-pr-autofix's categories: must fix / should fix / won't fix / needs discussion.
    • Apply must-fix and should-fix changes — one commit per item.
    • Reply to the thread via GraphQL addPullRequestReviewThreadReply (takes threadId + body directly, no comment-id lookup needed): quote the finding, state the action taken and commit hash (or the reason for not fixing).
    • Resolve fixed and won't-fix threads: bash <SKILL_DIR>/scripts/resolve-review-threads.sh [number] --thread-id <id>[,<id>...]
    • Leave "needs discussion" threads open — that's a signal for the human reviewer, not something to resolve automatically.
    • Never touch a thread not returned by step 2 — those belong to a human or Copilot, and are atk-pr-autofix's job to handle later.

  5. Push, then return to step 2. Repeat up to 10 passes.

  6. Report.

Notes:

  • If using Claude Code: use general-purpose subagent with model=Opus
  • If oracle type is available: for Agent tool, use subagent_type=oracle max_turns=35

Reporting:

  • Summarise fixes applied and threads left open for discussion (link them).
  • If the pass cap (10) was hit and new threads were still appearing, recommend another $polish-implementation run — there may be more to find.