Przejdź do treści
Biblioteka skilli
Programowanie · Planowanie

forge

/forge

Planuje, buduje i testuje krzyżowo nietrywialną zmianę przed bezpieczną realizacją, a dla przeglądu planu, pojedynczego zlecenia albo odzyskiwania kieruje do innych skilli.

Repozytorium
forge ↗
Licencja
MIT
Ostatnia zmiana repozytorium
28.10.2026

Ocena i bezpieczeństwo

90/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść25 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
22
Skrypty
4
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 28.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co sprawdzić przed instalacją
  • 4 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.

Skrypty: context-policy.test.sh, dispatch-guard.sh, dispatch-guard.test.sh, sol-liveness.sh

Co robi i kiedy

Planuje, buduje i testuje krzyżowo nietrywialną zmianę przed bezpieczną realizacją, a dla przeglądu planu, pojedynczego zlecenia albo odzyskiwania kieruje do innych skilli.

Oryginalny opis (EN)

Use when the user asks to forge, plan and build, or cross-model stress-test a non-trivial change before guarded execution. Use forge-review for plan review only, forge-build for one frozen work order, forge-orchestrate for a multi-unit build, and forge-recovery only for a stalled or stranded run. Do not use for trivial edits or review of already-written code.

Jak działa

  1. Pick the execution profile first
  2. Routing
  3. Safety invariants — binding at every profile
  4. Where the detail lives — load on demand
  5. Gates
  6. What NOT to do
Sekcje z nagłówków SKILL.md, w oryginale · 116 linii, 870 słów
Podgląd SKILL.md
---
name: forge
description: Use when the user asks to forge, plan and build, or cross-model stress-test a non-trivial change before guarded execution. Use forge-review for plan review only, forge-build for one frozen work order, forge-orchestrate for a multi-unit build, and forge-recovery only for a stalled or stranded run. Do not use for trivial edits or review of already-written code.
---

# FORGE — Grill, Cross-Model Review the PLAN, Then Build at the Active Cadence

*(Act-2 cross-model review is of the **plan** and runs at both profiles. Per-**unit**
code review is cadence, and is `assured`-only — see the profile table below.)*

Four acts, one artifact (`PLAN-REVIEW-LOG.md`):

1. **Grill** (you ↔ Claude) — kills *building the wrong thing*. Interrogate until intent
   is locked, then write `PLAN.md`.
2. **Adversarial review** (Claude ↔ another model) — kills *a plan that sounds right but
   breaks*. Read-only, cross-model, ends `VERDICT: APPROVED` or `REVISE`.
3. **Build** (simple) / 4. **Orchestrated build** (complex) — kills *the right plan built
   wrong, or uncontrolled concurrency eating the machine*.

Acts 1–2 write **no product code**. Product code starts only after `APPROVED`.

## Pick the execution profile first

| Source, in order | |
|---|---|
| 1 | Explicit owner instruction in chat |
| 2 | The repo's `.ai/EXECUTION_PROFILE.md` |
| 3 | Otherwise `assured` |

- **[`profiles/sprint.md`](./profiles/sprint.md)** — smallest relevant proof once,
  **no per-unit review, no orchestrator proof rerun**, immediate integration, max three
  disjoint lanes, one comprehensive gate at sprint end. A failed proof means **fix and
  re-run that same targeted proof** — escalation to `assured` requires an actual
  wider-risk finding, not a red result.
- **[`profiles/assured.md`](./profiles/assured.md)** — per-unit cross-model review with
  receipts, orchestrator re-runs the proof itself, integrated gate before the closing commit.

The profile controls **cadence only**. It never relaxes the safety invariants below.

**Cross-model unit review, orchestrator proof rerun, fresh-database rebuilds and full
per-unit gates are `assured`-only.** Opening an Act 3/4 reference does not re-enable
them — the active profile still governs.

## Routing

Governed by **[`ROUTING-POLICY.md`](./ROUTING-POLICY.md)** — canonical, owner-set. Where
anything here conflicts with it, that file wins and this one is wrong.

**Under `sprint`:** apply the owner overlay — or the default executor if none —
**automatically**, and record one line saying which and why. No cross-model routing
debate, no Gate-2 wait.

**Under `assured`:** routing is **proposed per unit and argued** by both models, Claude
arbitrates, and the agreed split is logged with a one-line reason per unit; Gate 2
applies to complex builds.

At both profiles an explicit owner project overlay binds and is passed mechanically to
the guard. Absent a preference the defaults are **Sol 5.6 @ ultra** (Codex side) and
**Fable 5** (Claude side), leaning toward Sol for token-heavy units on cost.

## Safety invariants — binding at every profile

- **Owner instructions win**, always.
- **No lane launches without `dispatch-guard.sh` printing `ADMIT:`.** A `REJECT` is
  stop-the-line — never overridden, never "just this once".
- **Three active lanes total** (not per model); **one heavy/gate lease** machine-wide.
- **One writer per worktree**; parallel units hold **disjoint declared file sets**.
  Declared sets are **binding during parallel writing**. Widen one only via a single
  guarded re-admission through the guard (`--admission-action expand`), continuing in
  the same lane;
  never write an undeclared path that overlaps another active lane.
- **The executor never commits or merges.** The orchestrator owns the commit. Under
  `sprint` it checks only the admitted paths, overlap/safety state, and the executor's
  passing proof receipt; under `assured` it reads the complete diff and reruns proof.
  This is a git-safety property, not a vendor rule.
- **Additive integration only.** Destructive Git — `reset --hard` over landed commits,
  force-push, `--amend` of pushed work, branch deletion, `clean` — needs explicit owner
  approval.
- **No secrets** in diffs, logs, commits or reports.
- Codex is **read-only** through Acts 1–2 whatever the routing says.

## Where the detail lives — load on demand

| Need | File |
|---|---|
| Grill + adversarial-review procedure | [`references/acts-1-2.md`](./references/acts-1-2.md) |
| Assured-only Act-3/4 procedure | [`references/acts-3-4.md`](./references/acts-3-4.md) (`sprint` uses its profile file instead) |
| Engine, routes, preflight | [`ENGINE.md`](./ENGINE.md) |
| Exact `codex exec` templates | [`references/invocation-templates.md`](./references/invocation-templates.md) |
| Assured-only argued routing mechanics | [`references/routing-detail.md`](./references/routing-detail.md) (`sprint` applies overlay/default directly) |
| Usage limits, credit policy | [`references/usage-limits.md`](./references/usage-limits.md) |
| Liveness — is it hung? | [`references/liveness.md`](./references/liveness.md) · `sol-liveness.sh` |
| Admission declaration fields | [`references/admission.md`](./references/admission.md) · `dispatch-guard.sh` |
| Why these rules exist | [`references/incidents.md`](./references/incidents.md) |

Do not reconstruct `codex exec` commands from memory — use the templates verbatim.

## Gates

**Gate 1** — plan `APPROVED` (after Act 2), both profiles.
**Gate 2** — routing approval, **`assured` + complex builds only**; `sprint` applies the
overlay/default automatically and does not wait.
There is no owner commit-approval gate.

**Sprint commit check** — the whole of it: changed paths match the admitted set · no
overlap or safety breach · the executor's proof exited 0. Then the orchestrator commits
and pushes. No complete-diff review, no proof rerun, no unit gate.
**Assured commit check** — complete-diff review + the orchestrator's own proof rerun +
the integrated gate.

## What NOT to do

- Don't review already-written code — that's `/codex:review`.
- Don't skip Act 1; the grill is half the value.
- Don't let a same-model subagent review count as cross-model.
- Don't dispatch around the guard, or ask the owner for the removed commit gate.
- Don't run `assured` ceremony on a `sprint` unit — no unit review, no proof rerun, no
  per-unit gate — whatever a loaded reference describes.
- Don't promote a unit to `assured` just because its proof went red; fix and re-run it.
Pliki w paczce (22)
  • CONTROL-PLANE-CHANGES.md 2 KB
  • DESIGN.md 8 KB
  • ENGINE.md 5 KB
  • LICENSE 1 KB
  • ROUTING-POLICY.md 4 KB
  • SKILL.md 6 KB
  • THIRD-PARTY-NOTICES.md 1 KB
  • context-policy.test.sh 3 KB
  • dispatch-guard.sh 68 KB
  • dispatch-guard.test.sh 49 KB
  • sol-liveness.sh 4 KB
  • profiles/assured.md 2 KB
  • profiles/sprint.md 3 KB
  • references/acts-1-2.md 4 KB
  • references/acts-3-4.md 10 KB
  • references/admission.md 10 KB
  • references/engine-resolution.md 3 KB
  • references/incidents.md 3 KB
  • references/invocation-templates.md 7 KB
  • references/liveness.md 6 KB
  • references/routing-detail.md 5 KB
  • references/usage-limits.md 14 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · 3orionkings/forge90MIT
aimi-so/aimi-engineering-plugin100MIT
claudemods/context-forge100MIT
cwinvestments/memstack100MIT
MohammedAtya/guardrail-forge100MIT
faizanio/FORGE100MIT
thecrazybob/claude-code-plugins100MIT
floka-as/floka-marketplace100MIT

Migawka bazy z 13.11.2026 · wersja paczki c975aee0d196 (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

  • 90forge-with-docs
    Planuje nietrywialną zmianę w repozytorium, w którym słownik domenowy, model kontekstu i ADR-y są wiążące i wymagają aktualizacji.
  • 90forge-orchestrate
    Porządkuje zatwierdzony plan z kilkoma ograniczonymi jednostkami, ustala kolejność zależności i uruchamia je równolegle w izolowanych drzewach roboczych.
  • 100forge-onboarding
    Prowadzi początkującego twórcę aplikacji Forge przez budowę Forge Guru, agenta Rovo odpowiadającego na pytania o Forge, i kieruje inne zadania do wyspecjalizowanych skilli.
  • 100forge-design
    Wspiera budowę interfejsów z komponentami webowymi Tyler Forge (@tylertech/forge), kierując przepływem pracy, drabinkami decyzji i spójnymi konwencjami dla forge-card, forge-table, forge-dialog i innych.
  • 100forge-implement
    Implementuje funkcję lub poprawkę ze zgłoszenia, pliku planu albo opisu, zgodnie ze standardami projektu, i kończy na otwarciu pull requesta.
  • 95forge-app-review
    Sprawdza aplikację Forge przed wdrożeniem pod kątem bezpieczeństwa, architektury, kosztów, wydajności i zbędnych wyzwalaczy.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.