PR Ready

PR Ready

An open PR/MR isn't mergeable. Two things block it — a stale base or red CI, and open review threads. Same loop for both: fix, push, re-query the forge, repeat. Handle whichever is blocking; usually that's both.

No feature work. No drive-bys.

Rebase + CI

1. Find the worktree for this PR (or create one per repo convention). Prefer an existing worktree the user already has.
2. git fetch origin <default> (usually main).
3. Honor repository history rules before choosing an update strategy. Where force pushes are forbidden or the user requires a merge, merge origin/<default> into the existing branch without rewriting commits. Otherwise rebase onto origin/<default>. Resolve conflicts by reading both sides' intent and regression tests; shared callback code must preserve newer callback-once behavior as well as the PR's guards.
4. Push (--force-with-lease if the rebase rewrote history).
5. Watch CI. Fix failures caused by the rebase or existing breakage on this branch — nothing else.
6. Flaky upstream on the default branch → say so. Don't paper over it with unrelated test deletes.
7. Verify process.execPath in the same execution mode as MR tests, using the repository's package manager. Background login shells can replace an earlier PATH export, so select the project's declared Node version inside the launched command. A newer runtime can shadow happy-dom localStorage or remove util.isDate; rerun unchanged failing tests under the supported runtime before editing application code.
8. For merge commits, old lint-staged hooks can autoformat every upstream file in the index. Run lint on the PR delta and the repository precommit checks explicitly, preserve upstream files, and document any one-command hook bypass used solely to prevent unrelated formatting. Never call that bypass a substitute for verification. Do not assume HUSKY_SKIP_HOOKS=1 is honored: inspect the hook output. Legacy hook git add can also include unstaged edits to already-staged files. Reconcile the resulting commit against the pre-merge PR file set plus intended fixes, and restore unrelated suppression-file pruning rather than publishing incidental changes.

• Unset npm_config_prefix on MR commit processes as well as test processes so Husky’s nvm hook can select Node 18. If legacy lint-staged concurrently runs git add for JS and Vue groups and races on index.lock, first verify the lock has disappeared and review the staged diff. Split the staged extension groups into additive commits if needed; do not delete a live lock, rewrite published history, or bypass the hooks to hide the race.

CI-only failures

• Distinguish a V8 heap-limit abort from a container OOM before changing CI resources. Measure the failing typecheck with the supported runtime and extended diagnostics, then verify a bounded heap override scoped to that job; keep the typecheck enabled.
• Reproduce repository-wide duplicate scans against a tracked git archive snapshot, not a working tree full of local build output. Attribute clones before changing configuration. Machine-generated API artifacts should be regenerated and drift-checked, not manually deduplicated; preserve the authored-code threshold when excluding those artifacts.
• Verify the package script's actual scope before calling a run the full backend suite. A package's test script may run unit tests only, while exec jest --runInBand exercises every configured Jest project. Require the completed project/suite summary, not a partial log.

Interrupted repairs

• Reconcile the current remote head, worktree diff, and saved test evidence before resuming an interrupted repair. Preserve an existing merge and relevant uncommitted fixes; do not repeat a repair already pushed by another worker. A partial test log without a completed runner summary is not a pass.
• When a merge exposes tests for deliberately deleted files, verify the deletion commit and remaining production references before retiring those assertions. Keep active replacement coverage and record the before/after failure evidence; do not restore retired application code or delete unrelated tests merely to turn CI green.

Review threads

1. Commit/push any unpushed work first (ask if the message is unclear).
2. List open/unresolved inline comments + review bodies.
3. Want a bot pass? Request the repo's review bot, then wait for it — don't spin forever; if it's stuck ~10–15m, say so. A Claude comment explaining how to request a manual review is not a completed review. When actual AI review is requested, trigger it for each PR/base and verify the submitted review matches that PR's current head. Inspect the check's title/body: a neutral clean review with an explicit no-issues result differs from a neutral worker result with zero turns and no review. Keep those outcomes separate. A managed Claude review and the legacy Claude GitHub Action can both react to one mention. If the legacy action fails but the managed review completes, reply with both exact run/review links; do not call the failed action successful or blindly rerun both services.
4. Per comment: valid → fix and resolve. Wrong or outdated → short reply and resolve. Nit out of scope → ask once. Never resolve a #1 maintainer comment without fixing it or an explicit override (pr-triage).
5. Push (pr-update if they want the commits split).
6. Repeat until unresolved threads are gone, or only ones they OK'd deferring.