بنقرة واحدة
flywheel-swarm
Launch parallel Cursor Task agents to implement multiple beads with Agent Mail.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
القائمة
Launch parallel Cursor Task agents to implement multiple beads with Agent Mail.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Relentless goal/design interview that sharpens framing and writes durable docs (brainstorm artifact, ADRs, glossary) as decisions crystallize. Use when refining a flywheel goal, pressure-testing scope, or "grill with docs".
Start or resume the full agentic coding flywheel. Thin index — observe, then ceremony, then phase skills on demand.
Use during flywheel planning (Step 4.5–5) when mockups, layouts, or diagrams help the user decide before beads. Optional browser companion — not for implement or review.
Run Cursor-native dueling idea wizards inside the flywheel (or legacy NTM). Routes artifacts into discovery, plan, or review. Use for /flywheel-duel and start.md Duel menu.
استنادا إلى تصنيف SOC المهني
| name | flywheel-swarm |
| description | Launch parallel Cursor Task agents to implement multiple beads with Agent Mail. |
Launch a parallel implement swarm. $ARGUMENTS
Cursor-first (default). Fan out with Cursor Task subagents (
run_in_background: true) on a single shared branch, coordinated via Agent Mail file reservations (program: "cursor"). Each Task must setmodelfrom the confirmed implement table (see step 3). Do not use git worktrees.NTM optional. Only when the user explicitly wants tmux panes, set
FW_IMPL_BACKEND=ntmor followskills/legacy/ntm/swarm-ntm.mdfrom disk. Do not silently fall back to NTM.
Preflight (MANDATORY): Ensure Agent Mail is running. If flywheel_observe reports agent-mail unreachable, run flywheel_remediate({ checkName: "agent_mail_liveness", mode: "execute", autoConfirm: true }) or am serve-http. Parallel swarm is blocked without Agent Mail.
Bootstrap: macro_start_session(human_key: <absolute cwd>, program: "cursor", model: <coordinator model>, task_description: "Swarm coordinator").
Call flywheel_approve_beads with action: "start" (or use ready beads from the current wave). If no beads are ready, say "No beads are ready — run planning/start first."
Ask the user: "How many agents in parallel? (Recommended: 2–4)"
Recommend models, user chooses (MANDATORY once per flywheel run):
flywheel_confirm_impl_models({ cwd }).confirmed: false, explain implModelsGate.rationale to the user, show the recommended table (implModelsGate.recommended), then present numbered implModelsGate.options. Wait for their reply.confirmImplModels: "recommended"confirmImplModels: { uniform: "<slug>" }confirmImplModels: { simple, medium, complex }confirmImplModels: "defaults"spawnInstructions from the confirm result when spawning Tasks.Coordinator: git pull on the shared branch before spawning the first wave.
Per ready bead (up to the user's limit), spawn in parallel (~30s stagger):
Task({
model: "<from table for bead complexity>",
subagent_type: "generalPurpose",
run_in_background: true,
description: "Impl <bead-id>",
prompt: "<marching orders — see template below>"
})
Complexity → model mapping is in implModels / flywheel.config.yaml implement: (simple, medium, complex).
Marching orders template (embed in each Task prompt):
## Agent Mail Bootstrap
macro_start_session(human_key: '<absolute cwd>', program: 'cursor', model: '<task model>',
task_description: 'Implementing bead <id>: <title>')
file_reservation_paths (exclusive=true) for every file you will edit — retry on conflict; STOP if blocked.
git pull --rebase before first edit.
## Bead: <id> — <title>
<description + acceptance criteria>
## Pre-completion gate
1. /ubs-workflow on changed files (or justify skip)
2. Scoped tests only: `cd plugins/cursor-orchestrator/mcp-server && npm test -- <paths-for-touched-test-files>` (no watch/background/full suite)
3. Self-review diff
4. Write `.pi-flywheel/completion/<bead-id>.json` (CompletionReport v1)
Coordinator runs one full `npm test` per wave after `acquire_build_slot({ slot: "npm-test", exclusive: true })`.
## On completion
git commit + git push (shared branch); send_message to coordinator with UBS + verify + self-review summary; close bead per AGENTS.md; release_file_reservations.
Monitor: Poll fetch_inbox for completion messages. Nudge stuck agents via send_message. For wave 2+, use flywheel_advance_wave (it reuses confirmed models or shows the same gate if not yet confirmed).
Wave-completion review (MANDATORY — before commit or wrap-up):
flywheel_wave_review_gate({ cwd, beadIds: [...] }).userGate.title, explain userGate.rationale, show numbered userGate.options, wait for the user.coordinatorAction (e.g. option 1 → flywheel_review looks-good per bead).When the bead queue is empty after review:
flywheel_wrap_up_gate({ cwd }) — present full / commit only / skip.confirmWrapUp: "full" | "commit_only" | "skip".skills/start/_wrapup.md for the chosen branch (sub-steps still use numbered choices).When FW_IMPL_BACKEND=ntm or the user insists on tmux panes: invoke /vibing-with-ntm, resolve NTM_PROJECT, spawn via ntm spawn + ntm --robot-send with program: "claude-code" per pane. See skills/legacy/ntm/swarm-ntm.md on disk.