| name | swarming |
| description | Use after pulse:validating approves execution and swarm mode is recommended, when the current work slice should be run by coordinated parallel workers. |
| metadata | {"version":"1.3","ecosystem":"pulse","role":"orchestrator","dependencies":[{"id":"beads-cli","kind":"command","command":"br","missing_effect":"unavailable","reason":"Swarming assigns and tracks beads through br."},{"id":"beads-viewer","kind":"command","command":"bv","missing_effect":"unavailable","reason":"Swarming inspects the live bead graph with bv."}]} |
Swarming
If preflight readiness is missing, stale, or blocked (check .pulse/tooling-status.json), stop and invoke pulse:using-pulse before continuing.
Role Boundary — Read First
You are the ORCHESTRATOR. You launch workers, monitor coordination, handle escalations, and keep the swarm moving. You do NOT implement beads. If you find yourself editing source files, stop immediately — that is the pulse:executing skill's job.
- swarming = launches and tends workers (this skill)
- executing = each worker's self-routing implementation loop
Hard Rule — Active Swarm Must Not Ignore Actionable Coordination
If workers are spawned, online, busy, blocked, or expected to report, you are in a tending phase whenever there is actionable coordination work.
While actionable signals exist, keep looping through the active coordination surface and the live bead graph. Do not ignore pending coordination work:
- inspect worker follow-ups
- inspect the live graph
- send reminders
- resolve conflicts
- escalate only when the next move truly requires human judgment
If the runtime has no autonomous wakeup or polling mechanism and there is no current actionable signal, run one complete tend cycle, update .pulse/STATE.md plus owner handoff/resume instructions, report that the swarm is paused awaiting the next actionable signal, and do not claim swarm completion.
User escalation is for real product decisions, unresolved blockers, or persistent worker silence after you have already tried to recover the swarm through the active coordination surface.
Communication Standard
Blocker reports, conflict reports, and handoffs should be written so a busy teammate can understand them in one read.
Prefer:
- what is blocked
- what is happening right now
- one concrete example of the collision or failure
- what needs to happen next
Do not hide the real issue behind labels like reservation conflict, startup drift, or runtime blocker without explaining the practical effect.
When to Use
Invoke only if all are true:
pulse:validating has approved execution
- current-work beads are in
open status and approved for execution
.pulse/tooling-status.json says recommended_mode=swarm
- the active CLI exposes a real native swarm path for this session
- if
.pulse/scripts/pulse_status.mjs exists, run node .pulse/scripts/pulse_status.mjs --json first to confirm onboarding, current work slice, reservations, and any saved handoff before launching the swarm
If preflight recommends single-worker, do not invoke this skill. Invoke pulse:executing directly instead.
Read references/runtime-adapter-spec.md before adapting these instructions to a concrete runtime.
Phase 1: Confirm Swarm Readiness
- Read
.pulse/tooling-status.json
- Read
.pulse/state.json if present, then .pulse/STATE.md
- Identify the current epic or feature bead root
- Run a live graph check:
bv --robot-triage --graph-root <EPIC_ID>
Confirm:
- open executable work exists
- no validation blockers remain
- dependencies are sane
Update .pulse/state.json and .pulse/STATE.md with current swarm intent and epic ID.
Phase 2: Initialize the Coordination Surface
Use the smallest runtime primitives that preserve these behaviors:
- spawn workers
- deliver startup context
- observe worker progress
- send coordinator follow-ups
- receive worker reports
- prevent overlapping file edits through the shared reservation helper
- write and restore owner-scoped handoffs
Adapter mapping:
Claude Code
- use
TeamCreate when explicit teammate coordination helps
- use
Agent to spawn bounded workers
- use
SendMessage for coordinator ↔ worker follow-ups
- use
Task* only as optional runtime metadata, never as the work graph
Codex
- use the native subagent spawn path
- use parent-thread follow-ups as the coordination surface
- keep bead selection and handoffs in the shared Pulse artifacts
Shared rules:
- beads plus
bv stay the source of truth for work selection
.pulse/scripts/pulse_reservations.mjs is the file-coordination layer for every runtime
.pulse/STATE.md, .pulse/state.json, and .pulse/handoffs/ stay authoritative for pause/resume
- do not invent extra registration, inbox, or topic mechanics when the runtime does not use them
Post the swarm start notification on the active coordination surface using references/swarming-appendix.md.
That coordination surface is where workers report startup acknowledgments, completions, blockers, conflicts, handoffs, and receive overseer broadcasts.
Phase 3: Spawn Workers
Spawn bounded workers that immediately load pulse:executing.
Provide each worker:
runtime_identity
coordinator_identity
adapter_name
epic_id
feature_name
- optional
startup_hint
- scoped task-specific context by default; full parent-context inheritance only when explicitly needed
Do not invent worker identities locally. Use the identity returned by the runtime's worker-spawn primitive.
Do not assign workers fixed tracks, fixed waves, or fixed bead lists as the normal case. Workers are expected to:
- read
AGENTS.md and project context
- load
pulse:executing
- post an
[ONLINE] acknowledgment
- run
bv --robot-priority
- reserve bead paths through
.pulse/scripts/pulse_reservations.mjs
- implement and report
- loop
Mark spawned workers in .pulse/STATE.md under ## Active Workers immediately after each spawn result.
Use one line per worker:
- Runtime: <runtime-identity> | Adapter: <adapter-name> | Status: spawned | Current bead: -
The worker startup acknowledgment later updates the same line to online.
Use the worker prompt template in references/swarming-appendix.md.
Phase 4: Monitor + Tend
The swarm is live; now you manage it.
Run a poll-act-repeat loop for as long as any of these are true:
- a worker is
spawned, online, busy, or blocked
- a worker owes a startup acknowledgment, completion report, blocker alert, or handoff
bv --robot-triage --graph-root <EPIC_ID> still shows ready or in-progress work
Every loop cycle must do all of the following:
- Inspect every new worker update on the active coordination surface
- Update
.pulse/STATE.md to reflect the latest worker status
- Reply, remind, or coordinate immediately when a worker is blocked or waiting
- Re-run the live graph check when a bead closes, a blocker clears, a worker goes silent, or the coordination surface looks stale
- Refresh reservation state when a conflict, release, or stalled worker could affect file ownership
Use live graph checks for oversight, not assignment:
bv --robot-triage --graph-root <EPIC_ID>
Do not ignore actionable coordination while the swarm is active. If updates are quiet and the runtime cannot autonomously wake or poll, complete one full tend cycle, persist state/handoff/resume instructions, and pause without claiming completion.
Worker Event Handling
Treat worker events as protocol-driven, not ad hoc. The canonical protocol, required fields, and coordinator message bodies are in references/swarming-appendix.md.
If a required event field is missing, request a corrected event and do not infer missing values.
Minimum coordinator obligations per cycle:
- Validate incoming event shape against the contract.
- Update the worker entry in
.pulse/STATE.md keyed by runtime identity.
- Verify bead-state transitions in
br/bv before acknowledging completion.
- Resolve reservations through
.pulse/scripts/pulse_reservations.mjs before permitting overlapping edits.
- Serialize same-branch commits by granting at most one
COMMIT_SLOT_GRANTED at a time; defer any worker that is ready to commit without the active slot.
- Escalate to the user when blockers require product judgment or a worker stays silent through the appendix silence ladder.
Handoff ownership rule:
- normal handoff path is same-runtime-identity resume only
- a worker must never consume another worker's handoff directly
- if the original runtime identity is unavailable, only the coordinator may reassign the orphaned handoff
Before coordinator reassignment, confirm and record all of the following:
- prior worker inactivity is confirmed
- reservation state for the prior owner is safe for transfer
- same-branch commit queue state is safe for transfer
- reassignment metadata is written to
.pulse/handoffs/manifest.json and the owner handoff file before resume, including previous owner, new owner, reason, and coordinator approval
Context Checkpoint
After each significant event, estimate your own context budget.
If context >65% used:
- Write
.pulse/handoffs/coordinator.json using the shared handoff envelope from ../using-pulse/references/handoff-contract.md.
- Register it in
.pulse/handoffs/manifest.json using the same summary, next_action, and path.
- Broadcast a pause notification on the active coordination surface that includes the rendered handoff summary, resume briefing, and transfer block highlights sourced from that JSON.
- If
.pulse/checkpoints/<feature>/... is in use, capture or refresh the feature checkpoint before leaving the swarm pause boundary.
- Report to the user that the orchestrator paused safely and how to resume.
- Do NOT abandon the swarm without writing the handoff.
The coordinator handoff must follow the same companion contract as planning/executing/validating:
summary -> short orchestrator handoff headline
next_action + read_first -> resume briefing for the next swarm turn
payload.transfer -> detailed transfer block for live worker state, blockers, and restart notes
Do not write the retired global handoff file.
Phase 5: Swarm Complete
When no current-work beads remain in_progress and the graph shows no remaining executable work for the active work slice, treat this as a graph-state signal only and verify review eligibility from approved shape artifacts + .pulse/STATE.md before any handoff:
-
Run final bead verification:
bv --robot-triage --graph-root <EPIC_ID>
-
If orphaned or blocked beads remain:
- report which beads remain and why
- ask the user whether to defer, create cleanup beads, or continue later
-
If all current-work beads are closed:
- run final build/test commands appropriate to the project
- clear
## Active Workers from .pulse/STATE.md
- inspect approved shape artifacts and
.pulse/STATE.md:
- work-shape path:
history/<feature>/work-shape.md
- phase-plan path:
history/<feature>/phase-plan.md
- epic-map path:
history/<feature>/epic-map.md and history/<feature>/current-story-pack.md
- determine review eligibility from those artifacts and current-work completion state, not from graph emptiness alone
- if shape artifact and
.pulse/STATE.md disagree about approved/current work slice or whether later slices remain: stop and route back to planning/state sync before any review handoff
- if more work slices remain:
- keep the whole-feature epic open
Active skill: swarming -> COMPLETE
Swarm: <EPIC_ID> - current work slice complete
Next: planning for next work slice
- if this was the final work slice and the artifacts agree no later work remains:
Active skill: swarming -> COMPLETE
Swarm: <EPIC_ID> - final work slice complete
Next: reviewing
-
Handoff message:
- if more work slices remain:
"Swarm execution complete for the current work slice. The whole-feature epic stays open. Return to pulse:planning to prepare the next work slice."
- if this was the final work slice:
"Swarm execution complete for the final work slice. Invoke pulse:reviewing."
Red Flags
Stop and diagnose before continuing if you see:
- Worker implements multiple beads at once — self-routing does not mean parallelizing within one worker
- Orchestrator edits source files — role violation
- Workers are idle but ready beads exist — inspect the active coordination surface, inspect the graph, and recover the swarm instead of waiting for the user
- No coordination activity for >5 poll cycles while work remains — workers may be stuck, off-thread, or context-exhausted; run the silence ladder
- The same file conflict repeats — bead decomposition may be too coarse; escalate
- Workers stop using
bv --robot-priority and start freelancing — re-broadcast the execution contract
- Build/test failures accumulate without intervention — create fix beads or stop and escalate
- Swarm is attempted while preflight recommended
single-worker — stop and use standalone executing
- An empty epic subtree is treated as proof the whole feature is complete — confirm final-slice review eligibility from approved shape artifact (
work-shape.md | phase-plan.md | epic-map.md) + .pulse/STATE.md before handing off to reviewing
Reference Files
Load when needed:
| File | Load When |
|---|
references/swarming-appendix.md | Worker startup template, message protocol, silence ladder, and coordinator handoff contract |
references/runtime-adapter-spec.md | Adapting canonical swarm behaviors to a concrete runtime |
docs/evaluation/pulse-swarming-hardening.md | Re-running RED/GREEN pressure tests for swarm coordination behavior |