paddock-crew
/paddock-crew
Zakłada zespół projektowy T3 z koordynatorem, budowniczym, recenzentem kodu i niezależnym krytykiem instrukcji i procesu.
Ocena i bezpieczeństwo
- Pliki w paczce
- 10
- Skrypty
- 2
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 09.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 2 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: assets/project/unittest_receipt.py, scripts/paddock.py
Co robi i kiedy
Zakłada zespół projektowy T3 z koordynatorem, budowniczym, recenzentem kodu i niezależnym krytykiem instrukcji i procesu.
Oryginalny opis (EN)
Set up and run a T3-native project crew with a coordinator, builder, code reviewer, and separate instruction/process adversary. Use for Paddock onboarding, new project setup, or tasks in a repo that has adopted Paddock. Single-company, approved-provider workflow; no herdr or local model required.
Jak działa
- Setup or a new project
- Every adopted task
- Boundaries
Podgląd SKILL.md
--- name: paddock-crew description: Set up and run a T3-native project crew with a coordinator, builder, code reviewer, and separate instruction/process adversary. Use for Paddock onboarding, new project setup, or tasks in a repo that has adopted Paddock. Single-company, approved-provider workflow; no herdr or local model required. --- # Paddock crew The current project session is the coordinator and owns integration. Roles are not products: one approved provider can supply separate sessions for all roles. Qwen and multi-company routing are optional extensions, disabled in this core package. Resolve packaged paths relative to this skill. Python 3.11+ and Git are required. The portable CLI is `scripts/paddock.py`; use `python` on Windows and `python3` on macOS/Linux. It never starts models or edits T3's private settings/database. ## Setup or a new project Read [onboarding](references/onboarding.md). Ask which company-approved resources are available: Claude, ChatGPT/Codex, Gemini, LiteLLM, Cursor; allow other answers. Distinguish available subscriptions from approved resources and working T3 adapters. Use supported host capability tools, map explicit instances/models, and leave unavailable resources unbound. Do not buy access, change credentials, enable a provider, download a model, or select a different organization as a fallback. Generate a local crew configuration and adopt/create the requested repo through the CLI. Preview changes first; apply within the user's setup authorization without asking again for routine reversible writes. Preserve existing project instructions and verification commands. The new repo starts with verification **unconfigured**; configure actual project checks, never substitute a passing scaffolding check. T3 project registration uses supported project tools: reuse an existing registration for the same canonical path, otherwise register the explicit existing folder. Do not edit its database or launch a top-level thread unless the user asks for one. Set or merge project actions only after reading existing actions. Users can also add the folder through T3. Skill discovery is provider-specific; the project instruction files supply the shared process even where the host cannot discover this skill. ## Every adopted task Read `.paddock/RULES.md`, the repo's existing instructions, and [orchestration](references/orchestration.md). Use this sequence proportionately: 1. Capture the original request and accepted contract (scope, exclusions, acceptance checks, authorization) using `task`. Run a fresh instruction adversary before implementing a substantial or ambiguous plan. Its remit is instruction quality, contradictions and evidence requirements, not redesigning the user's request. 2. Work directly for small tasks; otherwise use a bounded builder brief. Default to one builder; use host-created separate worktrees for concurrent edits. Explicitly account for uncommitted parent changes, which child worktrees do not inherit. 3. Integrate and execute the project's configured checks using `verify`. Preserve failures, skipped work, command output, and missing data. Never convert them to a green report. An exit code alone does not establish that any tests ran. 4. Freeze the source. Write `claims.md` as the proposed user-facing claims and `host.json` with observed controls plus references to actual host/tool receipts. Run `packet`. Give its paths and digest to a fresh code reviewer and a separate fresh process adversary. Read [review contracts](references/reviews.md). 5. Address findings with evidence. Source or policy changes invalidate final reviews; request new reviews of the new packet. At most two automatic repair/review rounds. Escalate unresolved findings with a concrete explanation. Never approve a worker's host permission request yourself or circumvent a rejected action. 6. Run `audit`. Report verification, code review, adversary review, and host controls separately. `evidence_complete` checks records/freshness; it is not proof that claims are true, controls are enforced, or changes are safe to merge. Report CI, deployment, and Windows results only when actually executed. For trivial work, explain why a review is omitted and leave its state `pending` or `not_run`; do not manufacture a fully audited completion. When delegation is absent, work sequentially and provide review packets for a separate session or human. Never impersonate an independent reviewer. Host routing instructions take precedence over examples here; follow the current tool schemas and returned continuation instructions. ## Boundaries Use only configured, approved, currently usable provider/model bindings; run `route` against a fresh capability snapshot before a host dispatch. A provider label does not prove account identity. This is a preflight policy check, not a host-wide block on other tools. T3/provider permissions supply actual runtime controls. No role receives extra authority to commit, publish, merge, deploy, send messages, or change credentials simply because it is a worker or reviewer. Keep within the user's authorization. Reviewers inspect, cite, and report; they do not fix code. Records are local and editable, not signed attestations. Never claim that digests, worktrees, review agreement or this skill enforce credential/security isolation. Raw logs and prompts can contain company data: keep `.paddock/runs` local and inspect any selected packet before sending it to an approved reviewer. Do not include these records or local crew bindings in distributable ZIPs.
Pliki w paczce (10)
- SKILL.md 5 KB
- agents/openai.yaml 1 KB
- assets/project/RULES.md 3 KB
- assets/project/project.json 1 KB
- assets/project/unittest_receipt.py 1 KB
- references/evidence.md 4 KB
- references/onboarding.md 4 KB
- references/orchestration.md 3 KB
- references/reviews.md 4 KB
- scripts/paddock.py 28 KB
Wersje
W bazie nie ma innych skilli o nazwie paddock-crew.
Migawka bazy z 13.11.2026 · wersja paczki d457b038d1fa (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
- 95paddock-agents
Odczytuje inne agenty w otwartych czatach i panelach tmux oraz przekazuje im zadania przez polecenie paddock-agents. - 91paddock-images
Generuje lub edytuje obrazy przez polecenie paddock-images i pokazuje je w rozmowie, na prośbę użytkownika. - 100crew
Deleguje ograniczone zadania z koordynatora Claude do pomocników w herdr na tym samym koncie, z plikami wymiany, żeby lider czytał tylko krótkie wyniki. - 100crew
Planuje i realizuje zmianę w pionowych plasterkach z czterema podagentami: zwiadowcą, wyrocznią, wykonawcą i recenzentem. - 90agentworks-inspect-crew
Sprawdza zespół agentów Crew w szczegółach: instrukcje, pliki, zakładki, harmonogramy, funkcje i połączenia. - 90ask-crew
Przeszukuje wspólny zasób wiedzy agentów Crew w poszukiwaniu tego, co inni zauważyli w podobnej sytuacji.