forge
Forge an issue, ticket, or PR into a shipped fix: propose a plan, get it approved, then implement, review, and refactor. Use when the user runs /forge or asks to investigate, fix, resolve, triage, or solve one. See Parameters for modifier flags.
radimsem · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add radimsem/forge-skills --skill forge技能说明
name: forge description: "Forge an issue, ticket, or PR into a shipped fix: propose a plan, get it approved, then implement, review, and refactor. Use when the user runs /forge or asks to investigate, fix, resolve, triage, or solve one. See Parameters for modifier flags."
Forge
Forge an issue into a shipped fix: heat it (implement), hammer it (review loop), then temper it (refactor).
Overview
Turn an issue or ticket reference into a verified, branch-correct, user-approved plan. Then, only after the user confirms (or immediately under automode), run the implementation lifecycle.
One numbered workflow (Steps 1–12), split by a single gate:
Part 1 — Gate (Steps 1–6): classify → fetch → branch → context → propose → [GATE] approve
Part 2 — Lifecycle (7–12): implement → review-loop → refactor → spin-off → self-evolve → close
gate held: nothing touches the codebase until "yes, implement" (automode = only sanctioned bypass)
- Part 1 — The gate. The contract; do not weaken it.
- Part 2 — Lifecycle. Runs autonomously after approval, stopping only at substantive gates (the Autonomy section).
"The agent" means whatever agent runs this skill. Adapt every reference (config directory, agent guide, interview UI) to your runtime. Nothing is hardcoded to one assistant.
Parameters
Parse the invocation as /forge <ref> [automode] [docs] [tdd] [worktree] [lookup] [secure] [changelog] [ci-watch] [compress] [implement] [codex | codex challenge | codex impl [challenge]] [coderabbit] [code-review], or /forge pr <N> [...], or /forge plan <path>[#<slug>] [...]. The words can appear anywhere in the request. A <ref> resolves to a git-host issue or a Jira ticket per the grammar below, and the pr <N> form selects PR-review entry mode. Flags are orthogonal and compose freely; some modes have their own allowed or ignored flags, listed in references/modes/pr-entry.md for PR mode.
Full flag matrix (effects, composition rules, conflicts): references/flags.md.
Target grammar — <ref>
The target is the first reference-shaped token, or the token right after a ticket or pr keyword. Routing works by keyword or by key-shape:
| Form | Routes to | Examples |
|---|---|---|
bare number, #N, issue N | git-host issue (Step 1b) | forge 42, fix #123, issue 7 |
matches [A-Z]+-\d+ | Jira ticket OR Linear issue (disambiguate if both configured — see references/trackers/linear.md) | forge PROJ-123, solve ENG-42, forge AB12-9 |
after ticket keyword | Jira ticket; number-only → ask project key | forge ticket PROJ-123, solve ticket 42 |
after linear keyword | Linear issue (forces Linear routing, skips disambiguation) | forge linear ENG-42 |
after pr keyword | PR-review entry mode (skips Steps 4–7; enters at Step 8 against the PR diff) | forge pr 47, forge pr #123 |
after plan keyword | plan-entry mode (replaces Steps 1, 2, 4–6 with a single Step P validation pass; Step 3 still runs; enters at Step 7 against the plan slice) | forge plan docs/superpowers/plans/x.md#task-3 |
A bare number routes to a tracker only if ticket or linear precedes it, and PR mode requires the pr keyword. A key-shaped ref is either Jira or Linear: when both trackers are configured, the agent disambiguates per the Jira/Linear disambiguation section in references/trackers/linear.md. Single-tracker setups skip the prompt.
Read references/modes/pr-entry.md for the PR-entry mode step modifications, the allowed and ignored flags, and the PR-specific Step 12 closing menu.
Read references/modes/plan-entry.md for the plan-entry step modifications and slice addressing.
Modifier flags
Each line says what the flag does and where it acts; the linked file owns the details.
- (none) — interview the user if a required field is missing, propose, then wait at the gate. The review uses the project reviewer agents plus
superpowers:requesting-code-review. automode— runs with no user gates and the agent decides Steps 6, 9, and 11. It never auto-commits, auto-pushes, or writes back to Jira; that is a hard floor. See references/autonomy.md.docs— works from documentation: the proposal goes toCONTEXT.mdand the plan is sourced from it, and grilling uses/grill-with-docsinstead of/grill-me.tdd— Step 7 writes the failing test first, observes it fail, then implements. See references/modes/tdd.md.worktree— Step 3 creates a sibling worktree instead of switching branch in place. See references/modes/worktree.md.lookup— Step 4 fetches current docs for every library the issue names. See references/modes/lookup.md.secure— adds asecurity-reviewpass at Step 8 once the regular review converges, before Step 9. See references/modes/secure.md.changelog— Step 12 drafts a changelog entry in the repo's existing format. See references/modes/changelog.md.ci-watch— after a Step 12 push, polls CI; a red result reopens Step 8. See references/modes/ci-watch.md.compress— sources a token-saving output skill for the whole session before Step 1:ponytailif installed, elsecaveman; if neither is installed, the flag is ignored with a one-line note. See references/modes/compress.md.implement— delegates Step 7 implementation to/implement, scoped to the approved Step 5 plan; its built-in/code-reviewcloseout is skipped because forge's Step 8 owns review. Conflicts withcodex impl(two engines for one step — stop and report); makestddredun