fix-pr
>
Works with
Claude CodeCursorCodex CLIGitHub CopilotGemini CLI
--- name: fix-pr description: > license: MIT --- # Fix PR feedback ## Flags | Flag | Default | Effect | |---|---|---| | `--pr <n\|url>` | current branch PR | Target PR | | `--no-push` | off | Keep commits local | | `--no-reply` | off | Do not post review replies | Missing values are `BLOCKED`. ## 1. Resolve and synchronize Resolve owner, repo, number, URL, base, head branch, and remote head SHA. Block on auth failure, missing/closed PR, dirty tree, or unsafe head checkout. Fetch, check out the head, and fast-forward to the remote SHA. Never reset, force, or discard local work. Record: `PR/head SHA | hunt counts | current finding | verdicts | commit/push | replies | terminal`. Done when local HEAD equals the PR head SHA and the ledger identifies the PR. ## 2. Hunt before editing **REQUIRED SUB-SKILL:** Read `../gh/SKILL.md` before any GitHub I/O. A script path is not a substitute. Hunt through that skill: surfaces 1–2 with `pr-threads.ts --json --open --complete`; CI snippets with `ci-failures.ts --json --pr N --sha <head>`. Surfaces: 1. unresolved review threads, including outdated ones 2. every comment page inside each thread 3. review-comment API chains reconciled to thread roots 4. actionable top-level review bodies 5. actionable PR conversation comments 6. PR CI on the head SHA: terminal non-success required/blocking checks and annotations Every hunt reconciles review-comment chains unless every root is proved present. Always load recipe 3 in `./REFERENCE.md` (REST reconcile); a successful script is not proof REST roots are present. Skip 4–5 only when JSON `moreReviews`/`moreComments`/`moreConvo` are false after `--complete`; else load them there. Recipe 6 always SHA-pins required/blocking checks and annotations here; `ci-failures` is drilldown only. Load remaining recipes only after `run` exits following every runtime, or a cap marker. Record counts and page markers. No triage or edit before all six passes finish. Normalize one finding per claim (source, URL/ID, reply target, author, path/line, rule ID, body, replies). Drop acknowledgments and status noise. Deduplicate only identical stable keys from `./REFERENCE.md`; preserve every native reply target. Done when pagination is exhausted and every finding is in the ledger. ## 3. Triage every finding Read surrounding code and trace the claimed path. Reproduce with the narrowest test, type check, or call trace when possible. Assign exactly one verdict: `fix`, `reject`, `clarify`, or `already-fixed`, with one evidence line. If reproduction is skipped, record why. No edits until every finding has a verdict. Done when none are untriaged and rejects have concrete evidence. ## 4. Fix and verify Apply only `fix` verdicts in focused edits. Run the narrowest covering checks for each fixed cluster. A failed required check is `BLOCKED`; if none exists, record that evidence. Do not change code for rejected or clarification findings. Done when every fix has a verified diff or no code fix was needed. ## 5. Commit and push When a diff exists, discard every pre-drafted subject (ledger, teammate, manager, branch, "review follow-up"). Read `../commit/SKILL.md` and run it once with `--unstaged` so that skill alone drafts from the locked diff as `type: <concrete code action proved by dominant hunks>`. Trailers default deny per Commit clean-room in `./REFERENCE.md` (identity trailers and harness footers such as `Made with Cursor`). Require no ban-list token, a passing conversation-only test, and rejection of every excuse there. Canonical rejects: `fix: address review feedback on agent files` and `fix: address ... review findings`. Skip commit on a clean tree. Unless `--no-push`, push and verify remote SHA; never force push. With `--no-push`, fixed findings become `AWAITING_PUSH` with no "fixed" reply. Remote movement or push rejection is `BLOCKED`. Re-read `git log -1 --format=%B`; ban-list, conversation-only text, or banned identity or harness trailers is `BLOCKED` (do not push). Apply Trailer hygiene in `../commit/REFERENCE.md` when banned trailers or harness footers appear, then re-check the push gate. Done when there is no diff, or one verified clean-room commit is local and pushed. ## 6. Re-hunt until stable Repeat all six hunt passes after the last code or remote mutation. Normalize and triage arrivals, then repeat Steps 4–6. After any hunt whose actionable set changed, require two consecutive hunts with the same set. Done when no finding is new, untriaged, or waiting on a local fix and consecutive stable hunts have completed. ## 7. Reply and report Unless `--no-reply`, skip targets whose existing replies already satisfy the verdict, draft remaining replies using `./REFERENCE.md`, then apply `./references/unslop-reply-drafts.md`. Preserve bot command prefixes. Consolidate shared targets. Post each remaining reply through the loaded gh skill (`pr-reply.ts`); do not invent raw `gh` for a covered reply. Re-hunt after replies; new findings return to Step 3. Before retrying a failed reply, refetch its target; if the intended reply is present, mark it posted, otherwise retry once. A second failure is `BLOCKED`. Do not resolve threads unless asked. Report `source | finding | verdict | action | evidence`, PR URL, commit/push state, hunt counts, and unreplied items. Terminals: `SUCCESS`, `NO_CODE_CHANGE` (no commit), `AWAITING_PUSH`, `BLOCKED`. Never `SUCCESS` with an unreplied required target. After interruption, restart at Step 1 and re-hunt before trusting the ledger. Does not fix merge conflicts or stacked-branch order; never invokes `make-pr`.
More General & Other skills
find-skills
vercel-labs/skills
Helps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can...", or express interest in extending capabilities. This skill should be used when the user is looking for functionality that might exist as an installable skill.
1.5M
grill-me
mattpocock/skills
A relentless interview to sharpen a plan or design.
972.7k
grill-with-docs
mattpocock/skills
A relentless interview to sharpen a plan or design, which also creates docs (ADR's and glossary) as we go.
828.8k
