菜单
Coordinator-driven parallel implementation — dispatch one implementer subagent per task in its own worktree, per-task review, fix loop, integrate. Load when the plan has independent tasks ready to execute.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/weselben/RooForge --skill subagent-driven-development命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | subagent-driven-development |
| source | https://raw.githubusercontent.com/obra/superpowers/main/skills/subagent-driven-development/SKILL.md |
| description | Coordinator-driven parallel implementation — dispatch one implementer subagent per task in its own worktree, per-task review, fix loop, integrate. Load when the plan has independent tasks ready to execute. |
You are the coordinator. Dispatch a fresh implementer per task — each in its own worktree — review every task (spec compliance + code quality), integrate branches locally. Your context stays clean for coordination.
Leading word: dispatch + report. Every implementer reports back with the four-item contract (branch, commits, tests, summary). You integrate; subagents never merge.
dev, feat/*, fix/*), merge back INTO it only. Never main/master — main changes only via PR..worktrees/<task-slug>/, branch <task-slug>, off the integration branch.using-git-worktrees + conventional-commits + caveman-commit + verification-before-completion. Reviewer: add caveman-review.dispatching-parallel-agents "Prompt rules". Every subagent prompt includes broader context (plan, why, what came before, what comes after) AND task context (exact files, exact commands, exact output). STE100 prose: one meaning per word, short sentences, active voice. The subagent must never need to ask a clarifying question — that signal means the prompt was incomplete.dev/feat/*/fix/*, never main/master). If unclear, ask before any work.using-git-worktrees:
git worktree add ".worktrees/$TASK_SLUG" -b "$TASK_SLUG" "$INTEGRATION_BRANCH"
Done when: integration branch confirmed, worktrees created with green baselines.
Dispatch all implementers in ONE AgentSwarm call (see dispatching-parallel-agents):
prompt_template with {{item}} placeholderimplementer-prompt.md — fill in the worktree, branch, and per-task spec where indicatedsubagent_type: coderKeep moving without check-ins. As each implementer returns, handle its report, dispatch its reviewer, run the fix loop until review is clean/parked/BLOCKED.
Implementers report one of four statuses:
Never rush an implementer asking questions — answer clearly and completely.
After each DONE, record BASE (integration-branch head the task branched from — never HEAD~1, truncates multi-commit tasks). Generate the diff:
git log --oneline <BASE>..<HEAD> > /tmp/review-<task-slug>.diff
git diff --stat <BASE>..<HEAD> >> /tmp/review-<task-slug>.diff
git diff -U10 <BASE>..<HEAD> >> /tmp/review-<task-slug>.diff
Dispatch a task reviewer (read-only explore) with: task spec, implementer report, diff path, binding global constraints verbatim. Reference template: task-reviewer-prompt.md — fill in TASK_SPEC, REPORT, BASE_SHA, and the diff path. Both verdicts required: spec compliance AND task quality. Implementer self-review never replaces this.
⚠️ "cannot verify from diff" items don't block — resolve each yourself before completion.
Triggers: spec ❌, Critical/Important finding, or confirmed ⚠️ gap. Two exits:
Everything else loops, 5 rounds max per task:
Agent(resume=...)) with open findings verbatim — context intact.re-review-prompt.md) over the fix range only (FIX_BASE = previous review's head).Breaker. Round 5 still has findings? Stop dispatching, adjudicate each:
Adjudicate only at the cap. Every ruling recorded; silent discard forbidden. Never fix findings yourself — coordinator fixes skip review and pollute context.
After all tasks complete, dispatch ONE final reviewer on the most capable available model over the full branch range. Point it at deferred-minor and parked findings for triage.
If findings return, dispatch ONE fix subagent with the complete list. Then one scoped re-review of the fix wave; adjudicate residuals. No second fix wave — residual load-bearing findings surface to the user at handoff.
git checkout <integration-branch> && git merge --squash <task-slug> && git commit -m "<conventional-commit>". One commit per task on the integration branch, each a natural Conventional Commit (e.g. feat(api): add user profile endpoint). Never main/master.finishing-a-development-branch for cleanup and PR creation.Before declaring integration complete, load verification-before-completion — check each subagent's claimed state against git status and full suite run, not against the summary it returned.
| Excuse | Reality |
|---|---|
| "Close enough on spec" | Spec gaps = not done. Fix, or hit the cap and adjudicate. |
| "I'll fix it myself, dispatching is overhead" | Coordinator fixes skip review. Resume the implementer. |
| "One more round will converge" | Past the cap, rounds don't converge. Adjudicate and route. |
| "This finding is obviously wrong" | Adjudicate only at the cap, record every ruling. |
| "The fix was small, skip re-review" | Unreviewed fixes = regressions. Every round ends with re-review. |
| "Subagents can share one worktree" | Parallel agents in one tree collide. One per task. |
| "Subagents should push their branches" | Worktrees share object store. Local merge. |
| "Merge everything to main when done" | Main changes only via PR. |