agent-all-cursor

/agent-all (Cursor port)

Why a prompt template instead of an orchestrator

Cursor'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.

What this skill installs

1. .agent-all.json — config (same shape as Claude Code; platform-agnostic).
2. .cursor/rules/agent-all.mdc (alwaysApply: true) — gives every Cursor chat in this workspace the pipeline rules.
3. .cursor/agents/agent-all-coordinator.md — parent agent. Reads phases 1–6 and routes work to implementer/reviewer subagents.
4. .cursor/agents/agent-all-implementer.md (is_background: true) — fans out per wave task. Cursor invokes multiple in parallel.
5. .cursor/agents/agent-all-reviewer.md (is_background: true) — fans out per wave for spec + quality review.

Usage (from a Cursor chat)

@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

Flags

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.

Pipeline

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)

Rules

1. Coordinator reads phases sequentially. Do not skip; do not parallelize phases.
2. State lives in .agent-all-state.json. Same shape as Claude port.
3. Delegate, don't reimplement. Phase 1 = brainstorming chat; Phase 2 = the user (or coordinator) writes the plan; Phase 3 = dispatch to agent-all-implementer.
4. Cursor handles parallelism implicitly. Coordinator says "for each wave task, invoke agent-all-implementer with this task block" — Cursor's planner fans out to background agents.
5. Hard caps: --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.

Differences from the Claude Code orchestrator

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

On error

• Dirty git tree → abort with 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.
• Wave task BLOCKED 3× → mark wave incomplete, surface to user.
• --max-cost best-effort only on Cursor.

When done

Coordinator prints summary: phases completed, iters, PR URL, cost (if tracked).

References

• references/porting-notes.md — design rationale for the prompt-template approach
• plugins/harness-floor/skills/agent-all/SKILL.md — source-of-truth Claude Code orchestrator