From harness-floor-cursor
Cursor port of /agent-all (intent → plan → wave-dispatch → gate → PR pipeline). Prompt-template approach — there is no programmatic runner. Cursor delegates to `.cursor/agents/agent-all-*.md` subagents via description-matching and parallelizes implementers/reviewers with `is_background: true`. See plugins/harness-floor/skills/agent-all/SKILL.md for the source-of-truth pipeline.
How this skill is triggered — by the user, by Claude, or both
Slash command
/harness-floor-cursor:agent-all-cursorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Cursor's subagent dispatch is *implicit*: a parent agent invokes a child by
lib/ask-user-adapter.mjslib/break-resolver.mjslib/config-loader.mjslib/host-invoker.mjslib/plan-parser.mjslib/state-rw.mjsphases/0-preflight.mdphases/1-intent.mdphases/2-plan.mdphases/3-dispatch.mdphases/4-gate.mdphases/5-pr.mdphases/6-loop.mdreferences/porting-notes.mdtemplates/agent-all.config.json.hbstemplates/agents/agent-all-coordinator.md.hbstemplates/agents/agent-all-implementer.md.hbstemplates/agents/agent-all-reviewer.md.hbstemplates/pr-body.md.hbstemplates/rules/agent-all.mdc.hbsCursor's subagent dispatch is implicit: a parent agent invokes a child by
matching the child's description frontmatter — there is no dispatch()
call to write. Cursor's planner handles routing automatically. That means
the orchestrator is not a .mjs file but a rule + subagent kit that
this skill installs into the target project. The pipeline lives in
phases/*.md for the parent (coordinator) agent to read sequentially.
.agent-all.json — config (same shape as Claude Code; platform-agnostic)..cursor/rules/agent-all.mdc (alwaysApply: true) — gives every Cursor
chat in this workspace the pipeline rules..cursor/agents/agent-all-coordinator.md — parent agent. Reads phases
1–6 and routes work to implementer/reviewer subagents..cursor/agents/agent-all-implementer.md (is_background: true) — fans
out per wave task. Cursor invokes multiple in parallel..cursor/agents/agent-all-reviewer.md (is_background: true) — fans
out per wave for spec + quality review.@agent-all-coordinator run /agent-all for "add user signup form"
@agent-all-coordinator run /agent-all using docs/tasks/12-fix-login.md --loop --max-iter=5
@agent-all-coordinator run /agent-all using docs/tasks/x.md --no-pr --wave-size=large
Same as Claude Code: --loop, --max-iter=<N>, --max-cost=<USD>,
--wave-size=small|medium|large, --no-pr, --no-brainstorm,
--resume, --force, --yes,
--break-condition=<spec>, --reconfigure.
When --loop is set, Phase 0 prompts the user interactively for the
break-condition preset (test-auto / visual-qa / Custom shell / Composite)
and offers to save the choice to .agent-all.json. Use
--break-condition=<spec> (JSON object or plain shell string) to skip
the prompt for one invocation, or --reconfigure to re-prompt even when
a non-default value already lives in config.
| Phase | File | Purpose |
|---|---|---|
| 0 | phases/0-preflight.md | git + .cursor/agents/ + config + input checks |
| 1 | phases/1-intent.md | brainstorming (chat-driven) OR load task file |
| 2 | phases/2-plan.md | draft plan into docs/superpowers/plans/ |
| 3 | phases/3-dispatch.md | fan out to agent-all-implementer (parallel via is_background) |
| 4 | phases/4-gate.md | fan out to agent-all-reviewer for spec + quality review |
| 5 | phases/5-pr.md | branch push + gh pr create (or skip) |
| 6 | phases/6-loop.md | breakCondition shell loop (manual re-invoke) |
.agent-all-state.json. Same shape as Claude port.agent-all-implementer.agent-all-implementer with this task block" — Cursor's planner fans out to background agents.--max-iter clamped to 50; --max-cost is best-effort — Cursor doesn't expose per-turn cost in the chat surface, so the coordinator records cost only when the user pastes it.| Aspect | Claude Code (/agent-all) | Cursor (agent-all-cursor) |
|---|---|---|
| Phase runner | Read each phase file, execute steps | Read each phase file, write chat output |
| Dispatch | superpowers:subagent-driven-development (Task tool) | Cursor planner matches agent-all-implementer.description |
| Parallelism | One Task per wave task | is_background: true per subagent file |
| State writes | .agent-all-state.json (atomic) | Coordinator writes via Cursor's edit surface |
| Plan writer | superpowers:writing-plans skill | Coordinator drafts inline OR user supplies file |
| breakCondition loop | loop-evaluator.mjs reruns the pipeline | User re-invokes @agent-all-coordinator per iter |
git stash instruction..cursor/agents/ missing agent-all-* → run cursor-init --theme=floor (future flag) or install this skill via node plugins/harness-floor-cursor/bin/init.mjs (future flag)..agent-all.json missing → warn + use built-ins from templates/agent-all.config.json.hbs.--max-cost best-effort only on Cursor.Coordinator prints summary: phases completed, iters, PR URL, cost (if tracked).
references/porting-notes.md — design rationale for the prompt-template approachplugins/harness-floor/skills/agent-all/SKILL.md — source-of-truth Claude Code orchestratornpx claudepluginhub kim-song-jun/agent-skill --plugin harness-floor-cursorProvides a checklist for code reviews covering functionality, security, performance, maintainability, tests, and quality. Use for pull requests, audits, team standards, and developer training.