waker
/waker
Nadzoruje długie zadanie programistyczne, aż każdy punkt listy kontrolnej zostanie potwierdzony plikami i przechodzącymi testami.
- Autor
- khoacat12345 ↗
- Repozytorium
- waker-ai ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 10.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Nadzoruje długie zadanie programistyczne, aż każdy punkt listy kontrolnej zostanie potwierdzony plikami i przechodzącymi testami.
Wyzwalacze z opisu
Oryginalny opis (EN)
Supervise a long coding task until every checklist item has on-disk evidence and tests pass. Use when the user says "use waker", "watch this", "finish it", or asks whether the agent is still working.
Jak działa
- If a run is already active
- Start a new run
- Stop
- Blind evaluation (Kaggle mode)
- Hard rules
Podgląd SKILL.md
---
name: waker
description: Supervise a long coding task until every checklist item has on-disk evidence and tests pass. Use when the user says "use waker", "watch this", "finish it", or asks whether the agent is still working.
---
# waker — supervisor
You must **not** declare a task done yourself. Only the daemon decides, based on
`git diff ∩ scope` and `verify` exit codes. Your words are not evidence.
## If a run is already active
Every time the user sends a message (even if they do not ask for progress),
before answering:
1. Run `waker status --format md --repo .`
2. Insert the **full table** at the top of your answer.
3. If status is `running`: do not speculate or promise on the daemon's behalf.
4. If status is not `running`: read `waker report --repo .` and summarize the
exact numbers (tasks done/total, tests, kicks, cost).
## Start a new run
1. **Identify the target repo.** If the prompt mentions a path that is not in
the current repo (for example `.agents/v2/...` while in `waker-ai`), ask the
user for the target repo before planning. Do not guess.
2. Take the user's prompt **verbatim** (drop a leading `@waker` if present).
Do not summarize or reinterpret.
3. `waker plan --prompt "<verbatim>" --out .waker/tasks.yaml --repo <target repo>`
4. Read the lint result:
- **If there is an ERROR** (missing scope/verify, duplicate scope, empty
verify): **DO NOT ask "OK to run".** Show the errors and offer exactly two
options:
a) `waker plan --via opencode --model <model> --prompt "<verbatim>" --repo <repo>`
to let a model fill scope/verify, then lint again;
b) the user merges/fixes tasks manually.
- Continue only when lint has **0 errors**.
5. When lint is clean: print `.waker/tasks.yaml` + any warnings, then ask
exactly one question: `OK to run? (ok / edit)`.
6. When the user agrees, run:
```bash
waker start --detach --repo . --tasks .waker/tasks.yaml \
--test "<repo test command>" \
--adapter opencode --model <provider/model> [--max-kicks 5] [--max-cost 1.0]
```
If not OpenCode, use the generic adapter:
```bash
waker start --detach --repo . --tasks .waker/tasks.yaml \
--test "<test command>" --adapter generic \
--agent "claude -p --resume {session} {prompt}"
```
7. Answer immediately, do not go silent: "Running in the background. You can
ask me for progress in chat, or open a terminal and run
`waker watch --repo .` for the live view."
> Prompts that look like an instruction (approval, decision, numbered lines)
> are usually not tasks. Use `--via opencode` to group them into 3–6 tasks with
> scope/verify instead of turning each line into an empty task.
## Stop
`waker stop --repo .`
## Blind evaluation (Kaggle mode)
When the user says "blind grading", "hidden test", or "Kaggle mode":
1. Write hidden tests to separate files, **not** in the working repo.
2. `waker vault init --repo .` → get `vault_id`.
3. `waker vault add --vault <id> --suite <name> --task T1 <file>`
4. `waker vault validate --vault <id> --repo .` — if FAIL, fix the suite and
rerun; **do not** start until the suite validates.
5. Start: `waker start --detach --vault <id> --tasks .waker/tasks.yaml ...`
6. When reporting progress: only read `hidden x/y` from
`waker status --format md`. Do not guess, do not print test names, do not
print source.
7. **Never** paste hidden test content into the worker chat; do not put the
vault path in the prompt.
## Hard rules
- If lint has an ERROR, **do not start** and **do not ask "OK to run"**.
- Do not edit `tasks.yaml` after starting (the hash will report TAMPER).
- Do not edit test files. To change tests, stop the run, edit, and restart.
- Do not say "done" — only read `status`/`report`.
- If the user asks "what is missing", answer from the task list with
`status != done` in `waker status --format json`; do not guess.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · khoacat12345/waker-ai | 95 | MIT |
| khoacat12345/waker-ai | 95 | MIT |
| khoacat12345/waker-ai | 95 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 71694bf887d3 (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
Brak podobnych skilli w bazie.