prove-e2e-maintain
/prove-e2e-maintain
Sprawdza skill weryfikacji i mapę funkcji względem kodu, uruchamia każdą funkcję i proponuje poprawki w jednym pull requeście.
Ocena i bezpieczeństwo
- Pliki w paczce
- 2
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Sprawdza skill weryfikacji i mapę funkcji względem kodu, uruchamia każdą funkcję i proponuje poprawki w jednym pull requeście.
Oryginalny opis (EN)
Check a repo's verification skill and its feature map against the code. Read every feature from source with parallel subagents, drive every feature live in one session, and ship the proven corrections as at most one PR. Use when the verification skill from h2xd:prove-e2e-create may have drifted, after user-facing changes, or when asked to audit the verify skill.
Jak działa
- Outcomes
- Edit scope
- Steps
- Report
- Do
- Don't
Podgląd SKILL.md
--- name: prove-e2e-maintain description: Check a repo's verification skill and its feature map against the code. Read every feature from source with parallel subagents, drive every feature live in one session, and ship the proven corrections as at most one PR. Use when the verification skill from h2xd:prove-e2e-create may have drifted, after user-facing changes, or when asked to audit the verify skill. disable-model-invocation: true --- # Maintain verification skill A recipe that clicks a button the app renamed last month fails mid-task, for an agent with no memory of the rename. This pass catches it. Every feature file of a skill from `h2xd:prove-e2e-create` gets a source read and a live drive, and one drive can cover several bullets. Run notes go in a scratchpad file, called `$NOTES` below. They hold the features covered, unreachable prerequisites, confirmed drift and the outcome. Never commit them. ## Outcomes End with exactly one, and name it: - **clean**: every feature got a source read and a drive. Nothing changes. - **changed**: one PR ships proven fixes to the map, skill text or helpers. - **blocked**: coverage could not finish, or a proven fix could not ship safely. Name the exact blocker. ## Edit scope Edit only the verification skill's folder: `SKILL.md`, `features/` and `scripts/`. Never edit product code in this pass. A wrong map is doc drift, so fix the map. A broken app is a product gap, so report it. ## Steps ### 1. Locate the target The target is `<skills dir>/verify-<app>/SKILL.md`, where `<skills dir>` is the CLI's project skill directory, such as `.claude/skills/`. It has the steps Doctor, Launch, Drive, Capture evidence and Clean up, and `features/` beside it. Search with `find . -path '*/verify-*/SKILL.md' -not -path '*/node_modules/*'`. When several match, ask which one. When none match, stop and point to `/h2xd:prove-e2e-create`, which the user runs. Done when you have one skill folder, called `$SKILL` below. ### 2. Fix the index Read `$SKILL/features/README.md` and list the files in `$SKILL/features/`. Fix index entries that are missing, extra, duplicated or link to nothing. Done when every feature file has exactly one index entry and every entry links to a file that exists. ### 3. Read each feature from source Start one read-only subagent per feature file, all at once, through the CLI's subagent tool, such as Agent. Run them on a fast model, such as Sonnet. Give each the brief in [`source-reader.md`](source-reader.md) and its feature file's path. Done when every feature file has a returned summary in `$NOTES`. ### 4. Reconcile - Merge overlapping recipes so the live pass needs as few app states as possible. - Check each cited drift against the source yourself. Do not re-check claims a reader called clean. - Read the changes since the map last changed: `git log --stat "$(git log -1 --format=%H -- "$SKILL")"..HEAD`. Call a feature missing from the map only with a source path that proves it. Done when `$NOTES` holds the drive plan, the confirmed drift and the missing features, each with a source path. ### 5. Drive every feature live Run this step even when the source looks clean. You drive, never a subagent. Drive every feature at least once. Follow the Launch step. It picks one long-lived instance driven in order, or a fresh isolated session per drive. Hold these rules whatever fails: 1. **Doctor before you drive.** Run the Doctor step before the first drive, on each fresh session, and after any failed drive. When Doctor cannot see the fault, such as a stuck UI on a healthy process, reset to a known state or relaunch. 2. **Evidence survives cleanup.** After each Clean up, check that the evidence is still where the Capture evidence step says. Do not assume it. 3. **Nothing outlives its drive.** Run Clean up after every failed attempt, whether the session hung, exited or is shared. On a shared instance, remove what the drive created and leave the instance running. When Doctor fails because the skill is wrong, that is drift. Fix it, restart only what the fix invalidated, and retry once. A second failure means blocked. A feature counts as `verified-unreachable` only with the route you tried and the prerequisite that blocked it, such as auth, an entitlement, the OS or external state. When the map does not mention that prerequisite, that is drift too. Done when every feature is `verified` or `verified-unreachable` in `$NOTES`, with its evidence path. ### 6. Triage Sort each finding into one kind: | Kind | Sign | Action | |---|---|---| | Doc drift | The map describes the user path wrongly or leaves it out | Fix the feature file or the skill text | | Harness gap | The app works, but the skill's helpers cannot drive it | Fix the helper. It stays executable and the skill body shows how to call it | | Product gap | The app is broken | Record it for the report. Keep it out of the PR | Drive every harness fix live again before it ships. After the last drive, run the Clean up step once more. Done when every finding has a kind, every harness fix passed a re-drive, and nothing this pass started still runs. ### 7. Ship or stop For **changed**: re-read every changed file and create a branch. Commit with `h2xd:git-commit`. Open one PR with `h2xd:git-pr-body`. For **clean** or **blocked**: no branch, no PR. Offer to file each product gap with the **Create** operation in `docs/agents/issue-tracker.md`. Done when the report is out, with the PR URL for changed. ## Report ``` <outcome>: <n> features, <n> verified, <n> verified-unreachable 🔴 <feature>: <product gap or blocker, and the next action> 🟡 <feature>: <drift fixed, unreachable prerequisite, or missing feature> 🟢 <feature>: verified, evidence at <path> PR: <url or "none"> ``` ## Do - Ask for a source path before calling any drift or missing feature real. - Name the prerequisite and the route for every unreachable feature. - Keep the PR to proven corrections inside the edit scope. ## Don't - Don't skip the live drive because the source looks clean. Drive every feature. - Don't let a subagent drive the app. Keep all driving in your own session. - Don't drive an instance that failed or surprised you. Run Doctor or relaunch. - Don't change the map to hide broken behavior. Report it as a product gap. - Don't kill processes by name. Run the target skill's Clean up step. - Don't commit `$NOTES`. Keep it in the scratchpad.
Pliki w paczce (2)
- SKILL.md 6 KB
- source-reader.md 2 KB
Wersje
W bazie nie ma innych skilli o nazwie prove-e2e-maintain.
Migawka bazy z 13.11.2026 · wersja paczki 757e80fd9831 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100maintain-verification-skill
Okresowo sprawdza skill weryfikacji aplikacji i mapę funkcji: równoległe czytanie kodu, jedna sesja na żywo i jeden zestaw poprawek. - 100maintain-verification-skill
Cyklicznie weryfikuje skill weryfikacji projektu i mapę funkcji: równoległe czytanie źródeł, jedna sesja na żywo obejmująca każdą funkcję i najwyżej jeden pull request z udowodnionymi poprawkami. - 100maintain-verification-skill
Cyklicznie weryfikuje skill weryfikacji projektu i mapę funkcji: równoległe czytanie źródeł, jedna sesja na żywo obejmująca każdą funkcję i najwyżej jeden pull request z udowodnionymi poprawkami. - 100maintain-verification-skill
Cyklicznie weryfikuje skill weryfikacji projektu i mapę funkcji: równoległe czytanie źródeł, jedna sesja na żywo obejmująca każdą funkcję i najwyżej jeden pull request z udowodnionymi poprawkami. - 100maintain-verification-skill
Cyklicznie weryfikuje skill weryfikacji projektu i mapę funkcji: równoległe czytanie źródeł, jedna sesja na żywo obejmująca każdą funkcję i najwyżej jeden pull request z udowodnionymi poprawkami. - 100maintain-verification-skill
Cyklicznie weryfikuje skill weryfikacji projektu i mapę funkcji: równoległe czytanie źródeł, jedna sesja na żywo obejmująca każdą funkcję i najwyżej jeden pull request z udowodnionymi poprawkami.