running-checks
/running-checks
Określa w Esposter, kiedy uruchamiać pnpm typecheck, lint, test, build, coverage i bench oraz które kontrole zmiana wykonuje przed pushem, a które zostawia CI.
- Autor
- Esposter ↗
- Repozytorium
- Esposter ↗
- Licencja
- Apache-2.0
- Ostatnia zmiana repozytorium
- 26.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 2
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
Co robi i kiedy
Określa w Esposter, kiedy uruchamiać pnpm typecheck, lint, test, build, coverage i bench oraz które kontrole zmiana wykonuje przed pushem, a które zostawia CI.
Wyzwalacze z opisu
Oryginalny opis (EN)
Apply when about to run pnpm typecheck, lint, lint:fix, test, build, coverage or bench in any package, or format from the root, when deciding which checks a change owes before its push, and whenever you find yourself waiting on a check's output. Esposter's split between the checks a session runs and the checks CI runs on the push.
Jak działa
- Settled — do not re-propose
- The tests of what the change touched, in the background
- Never poll a backgrounded check
- One run, after all the edits going out
- Watching the push's run, when asked
- A check CI owns, run locally
- Reading the result
- Reference pages
Podgląd SKILL.md
---
name: running-checks
description: Apply when about to run pnpm typecheck, lint, lint:fix, test, build, coverage or bench in any package, or format from the root, when deciding which checks a change owes before its push, and whenever you find yourself waiting on a check's output. Esposter's split between the checks a session runs and the checks CI runs on the push.
---
# Running Checks
**CI runs the checks; a session runs only the tests of what it touched.** Every push to `ai/queue` runs typecheck,
lint, format and the whole suite. The session does not watch its push's run unless the user asks for verification;
then it fixes a red typecheck or lint job itself, in a commit behind the push (`AGENTS.md`, "Finishing a change", step 6). The pre-commit hook only formats what is staged and
runs no check (`.githooks/pre-commit`); it re-adds each staged file whole from the working copy, so a partial stage of
a file commits all of it, and a commit is split by file or with the other hunks kept out of the working copy. The command and its directory are the `package-scripts` skill's; this page is
which checks a session runs and how it waits on them.
## Settled — do not re-propose
- **Typecheck and lint before a push.** CI runs them on every push at no session cost, while a local run takes
minutes per change and the root type-aware lint needs more memory than a local session should spend. Nothing waits on them
either way: the push already went out beside them, so a local run only moved the finding earlier by the length of
a CI run, at the session's price.
- **Dropping the touched tests with them.** Their cost scales with the change rather than the repository, and they
catch what neither a typecheck nor a lint can — the behaviour the change was for.
- **Lint, a typecheck or any other check in the pre-commit hook.** Checks stay in CI, which runs them on every push; the
pre-commit hook only formats, so no commit waits on a check.
- **A Pulumi apply.** `pnpm typecheck` in `apps/infra` runs before `infra:up`, since an apply reaches Azure and CI runs after it.
## The tests of what the change touched, in the background
The tests are one `Bash` call with `run_in_background: true`, redirecting their output to a log file and appending
their own exit code (`echo "exit $?" >> log`), and the session moves to the next piece of work: the commit message,
the spec sweep, the ledger row. The harness delivers a completion notification; read the log then. The one run that proves a regression test against a reverted fix is the exception to committing while it is in flight; its restore order is the `testing` skill's ("What to Test"). **This holds for
a one-file, "quick" suite too** — the habit that survives is the one applied without weighing it.
Which paths, and from which directory, are the `package-scripts` skill's `references/check-suite.md`: the suites
of the files touched and the workspace tests for a spec edit; a change that moved bytes into a bundle adds that package's own suite (`AGENTS.md`, "Finishing a change", step 4).
## Never poll a backgrounded check
A foreground poll of a backgrounded check is a foreground check: `for i in $(seq 1 40); do grep -q done log && break;
sleep 10; done` spends the turn the flag was meant to save. A backgrounded check announces its own completion. The
wait on a condition in the `context-efficiency` skill ("Wait on a condition, never a sleep") is for an **external**
process the harness cannot see finish — a dev server, a deploy — never for a check.
**Blocking is correct only when the sole remaining step is commit, merge or push.** When it genuinely is the last
step, end the turn with a one-line status and let the notification re-enter.
## One run, after all the edits going out
The tests run once, after **all** edits going out together — not per sub-task, not per unit. Each run re-pays a
fixed startup cost, and nothing is learned at chunk 3 that chunk 7 won't also reveal. Commit per coherent chunk
regardless: commits are cheap and protect against other sessions' resets, checks are not. The run follows the
review's quality lane, since cleanup edits code (`AGENTS.md`, "Finishing a change").
## Watching the push's run, when asked
Watching waits for the user to ask for verification; it does not start on its own after a push. Then
`pnpm ai:ci:watch`, in the background. It finds the run by the tip's commit — never by
`--branch` with `--workflow`, whose per-workflow filter answers `ai/queue` with weeks-old runs — follows a run that
was cancelled while it waited to the newer run that carries the commit too (`concurrency` in `CI.yaml` keeps only the
newest waiting run), waits for it, and exits non-zero on a red with each failed job's error lines, which is what wakes
the session to fix it. A wave of agents landing one after another is watched once, after the last landing.
## A check CI owns, run locally
Only to answer a red: a session whose push turned a typecheck or lint job red, a session asked to fix a failed CI
job, runs that one check as CI runs it, to reproduce the red and prove the fix. The
failed job's log (`gh run view <id> --log-failed`) names the package whose typecheck failed, the root's own
included, and the reproduction is `pnpm typecheck` from that package rather than the whole repository; a lint red is
reproduced by `vp run lint:fix` from the repo root, the one scope matching CI's
(the `package-scripts` skill's `references/check-suite.md`). A check that writes — `lint:fix`, `pnpm format` — owns
the files it covers until its exit line, so nothing is edited under it meanwhile. A worktree installed without the app has no `apps/web/.nuxt`, and every package's ESLint config imports `withNuxt` from the `eslint.config.mjs` Nuxt generates there, so ESLint cannot load until the shared checkout's is linked in: `ln -s "$(dirname "$(git rev-parse --path-format=absolute --git-common-dir)")/apps/web/.nuxt" apps/web/.nuxt` from the worktree's root. Node resolves the generated file's relative imports from the link's real path, so they reach the shared checkout's installs. On Windows, long paths must be enabled once (see `openspec/specs/architecture/monorepo-tooling/design.md`, "Windows long paths"), or `vp lint` fails on a worktree's deep node_modules; a worktree lives at `.agents/worktrees/<lane>`.
## Reading the result
The verdict is the exit line appended to the log — never a grep of the output, never the notification's exit code, never a pipeline's — and a red result is an edit then a fresh background run (`references/reading-results.md`).
## Reference pages
- `references/reading-results.md` — when a backgrounded check finishes and its verdict is read.
Pliki w paczce (2)
- SKILL.md 7 KB
- references/reading-results.md 2 KB
Wersje
W bazie nie ma innych skilli o nazwie running-checks.
Migawka bazy z 13.11.2026 · wersja paczki b6cf169b15d4 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 1 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90package-scripts
Opisuje skrypty pnpm projektu Esposter: który z nich uruchamia które sprawdzenie, build, benchmark i wydanie, skąd go wywołać oraz katalog poleceń ai:* dla agenta. - 95git-preflight
Uruchamia skonfigurowane w projekcie kontrole typów, lintu, formatu, bezpieczeństwa i testów przed commitem i proponuje naprawę tego, co nie przeszło. - 100bench
Prowadzi przy dodawaniu i edycji benchmarków, mierzeniu czasu działania i kosztu typów dla kompilatora oraz zmianach w plikach bench.md i typebench.md w projekcie Esposter. - 100verify-before-done
Dowodzi, że zmiana w kodzie działa, uruchamiając build, typy, lint i testy oraz sprawdza interfejs w przeglądarce. - 100gate
Uruchamia lint, typecheck, build i testy w samonaprawiającej się pętli, która czyta błędy, poprawia je i ponawia do 4 razy. - 100sfnext-quality-gates
Uruchamia i naprawia bramki jakości Storefront Next przed PR lub wdrożeniem: OxLint, formatowanie Biome, lint:a11y, lint:css, typecheck, budżety bundla, Lighthouse CI i listę kontrolną.