メニュー
Scaffold a new Forge tool — implementation, registration, permission/risk metadata, sandbox routing, and a unit test. Use when adding anything to src/tools/.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | add-tool |
| description | Scaffold a new Forge tool — implementation, registration, permission/risk metadata, sandbox routing, and a unit test. Use when adding anything to src/tools/. |
Four places change. Do all four or the tool is half-registered.
src/tools/<name>.ts — the tool module.
zod schemas for input and output.sideEffect (none|read|write|network|exec) and risk
(low|medium|high|critical). Be honest — these drive permission
prompts and auto-approval.src/sandbox/fs.ts helpers.src/sandbox/shell.ts +
classifyCommandRisk.requestPermission
(src/permissions/manager.ts) before the side effect.{ ok, data } / { ok: false, error }. Throw only for
programmer errors, using ForgeRuntimeError.Registration — wire into the registry following the pattern in
an existing tool like src/tools/read-file.ts.
test/unit/<name>.test.ts — at minimum:
Use vi.mock for the permission manager and sandbox boundaries.
Docs — if user-visible, update the tools table in
docs/ARCHITECTURE.md and the README "At a glance" count.
Finish with the verify skill.
Deep review — random exploration + cross-agent review, alternating. Finds integration bugs that self-review misses. Run every 30–60 min during active implementation.
Overview of Forge's Agentic Coding Flywheel methodology. Use to route between plan, bead, and code space. See FLYWHEEL.md at the repo root.
End-of-session ritual. Close finished beads, file new ones for remaining work, run verify, commit in logical groupings, push. Work is not complete until git push succeeds.
Start a new Flywheel planning round. Produce initial markdown plan, route to competing frontier models, synthesize hybrid, refine 4–5×. Use before any coding begins on a new feature.
Translate a polished markdown plan into self-contained beads in .beads/beads.jsonl. Every bead must be executable without reopening the plan. Use after the plan has converged.
Polishing pass over all open beads. Run 4–6 times. "Check your beads N times, implement once." Use after plan-to-beads and between refinement rounds.