Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Browser & automation

polish-pr

Finish BarefootJS PRs by addressing CI failures and Pullfrog reviews through current-head verification and one final human notification. Use for PR polish or blocker removal, not a read-only review or status request.

piconic-ai · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add piconic-ai/barefootjs --skill polish-pr

About this skill


name: polish-pr description: Finish BarefootJS PRs by addressing CI failures and Pullfrog reviews through current-head verification and one final human notification. Use for PR polish or blocker removal, not a read-only review or status request.

Polish a PR

Drive the requested PR to verified readiness, not merely "ready for review". This is a shared, instruction-only skill for Claude and Codex. Both follow this same file; it does not require Codex app tools or a Claude Workflow runtime. Read root AGENTS.md and .github/pullfrog/review.md completely before acting. Use available GitHub tools or gh; no particular harness or model is required. A status-only or read-only review request authorizes inspection, not this skill's mutations. Finishing does not authorize merging or unrelated fixes.

Host capabilities

  • Read/edit files and run commands using the current agent's own tools. For GitHub operations use an available connector/MCP tool or authenticated gh through its shell tool; the required evidence and completion gates are identical.
  • PR attachment to a chat is optional UI bookkeeping. If the host has no such tool, retain the PR URL in working notes; attachment is never a completion gate.
  • Use review/simplification skills when available (including Claude's existing skills), otherwise perform those passes inline using the repository rubric. Do not depend on a particular tool name, agent launcher, or model setting.
  • Use event/wait tools when available, otherwise inspect GitHub with gh and poll with backoff. If later monitoring is requested but the host cannot schedule it, disclose that limitation; do not pretend a persistent monitor was started.

Establish ownership and state

Resolve the PR, actual base branch, remote head SHA, and local checkout. Record the PR URL and optionally attach it if supported. Preserve unrelated work. For a stack, process bottom-up and propagate lower-layer fixes before reviewing upper diffs.

Keep a compact evidence record in working notes: PR/base/head, local validation commands and results, latest Pullfrog-reviewed SHA, findings/threads, applicable CI and drift checks, mergeability, and whether the final mention was posted. Refresh it after every push; never combine results from different heads.

Review, simplify, and trigger review

  • Read the cumulative diff against the PR's actual base. Review correctness and the repository design/testing contracts; a self-assessment of "clean" is not independent evidence. Available review skills may help, but inline review is a valid fallback. Do not spawn agents or switch models without authorization.
  • Simplify only changed code: shared decisions, clarity, and unnecessary work. Avoid unrelated cleanup. Verify real entry-point assumptions and meaningful tests, changesets, limitation graduation, and generated-artifact diffs.
  • Run the affected full layer suites plus targeted regressions. When a local failure appears unrelated, reproduce it on the base before calling it pre-existing. Missing tools or a skipped runtime are not a passing runtime test; obtain evidence from that backend's CI and disclose local limitations.
  • Push fixes and update the PR description to the actual diff. Undraft when the initial review pass is complete so Pullfrog can run asynchronously. Undrafting is a trigger, not the final completion gate; do not mention the human yet.

Own the CI and Pullfrog loop

Use the host's event/wait mechanism when available; otherwise poll with backoff. Keep waits bounded and give concise progress updates without narrating unchanged polls. Long-running adapter jobs are expected, not evidence of a blocker. Use the provided scheduling mechanism if the user requests later monitoring.

For each actionable CI/review finding:

  1. Re-diff against the current head and verify the finding is still applicable. Inspect failed job logs, not only the summary or an older run's failure.
  2. Add a repro at the correct layer, fix the root decision, verify, commit, and push. New divergences found during this work need executable evidence, not speculative registry entries or broad late-stage cleanup. A finding that names one missed shape usually has siblings: check the list in implement-issues step 3 and cover them in the same push. A drift failure in a fixture ledger (expectedHtml, coverage map, data points, compat or support-matrix lock) is fixed by bun run fixtures:regen, never by regenerating one file by hand. Verify with bun run fixtures:smoke <fixture> before pushing.
  3. Reply with the fix and verification; resolve an inline thread only after its finding is addressed or a reasoned disagreement is documented. Check review bodies as well as threads: a blocking finding can exist without an inline thread.
  4. Wait for review of the NEW head and all applicable checks again. Every push, rebase, or lower-stack fix invalidates stale approval and CI evidence.

Do not blindly rerun failures. If the failure is infrastructure-related, verify that diagnosis, retry the affected job, and inspect the result. Do not weaken tests, alter expectations to match a defect, or silently skip failed checks. If a permission, dependency, product decision, or external coordination is needed, report the precise blocker and ask for direction; persistence does not expand scope.

Final gate — same immutable head

Before notifying the human, establish ALL of the following on the current SHA:

  • Pullfrog has actually reviewed this SHA, with no unresolved blocking findings. A green pullfrog check alone, old review, undraft event, or bot launch comment is not sufficient. "No new issues" also requires earlier blockers to be addressed.
  • All applicable latest CI checks pass. Distinguish intentional path-filter/opt-in skips from missing validation; check job names together with their workflows because many adapters have a job called test.
  • For change
Read the complete source ↗