| name | bones-tasks-swarm-coordination |
| description | Use when dispatching 2+ parallel subharnesses for related work; replaces orchestrator-side file-disjoint task partitioning with a claim-and-close work queue |
Bones Tasks Swarm Coordination
Overview
When dispatching parallel implementers, the orchestrator usually has to partition tasks across them by file boundary so they do not trample each other. That gets fragile fast: tasks shift, file boundaries blur, and a partition mistake creates merge conflicts.
The bones tasks queue replaces orchestrator-side partition with worker-side claim arbitration. Each implementer pulls one open task, locks it, works it, closes it. The next worker picks the next open one. No partition decision required.
When to Use
- Dispatching 3+ parallel subharnesses for related but independent work (Wave B, Wave C, bug-batch)
- Tasks could overlap files but the work itself is independent
- New backlog items arriving while a swarm is in flight (orchestrator can drop a task; live swarm picks it up next claim)
- Replanning a stalled or partially-failed swarm (untouched tasks remain claimable)
Skip when:
- Single task, single agent (no coordination needed)
- Tasks have hard ordering dependencies (use sequential dispatch instead)
- One-off ad-hoc work where queue setup overhead exceeds the dispatch overhead
Setup (one-time per pipeline)
mkdir -p .bones && bones up -R .bones/<pipeline-name>.bones
bones up brings up:
- Fossil server at http://127.0.0.1:8765 (artifact storage + claim arbitration)
- NATS at nats://127.0.0.1:4222 (event stream — workers can subscribe)
Authoring Tasks
REPO=.bones/<pipeline-name>.bones
bones tasks create "<title>" -R $REPO \
--context priority=high \
--context category=substrate \
--context source=<where-this-came-from>
Title is searchable. Use --context k=v (repeatable) for filterable metadata. --files declares files-touched if the implementer should know up front.
Worker Brief (paste into every swarm dispatch)
After spawn:
bones tasks list -R <REPO-path> # see open tasks
Loop:
ID=$(bones tasks list -R <REPO> --format json | jq -r 'first(.[]|select(.claimed==null)) | .id')
[ -z "$ID" ] && exit 0 # queue empty, stop
bones tasks claim "$ID" -R <REPO>
... work, commit, push ...
bones tasks close "$ID" -R <REPO>
The orchestrator does NOT decide which worker gets which task. The queue does.
Recorded Failures (RED — what we did instead before this skill)
- v0.1.17 Phase 2 (parallel impl-2b/c/d/e dispatch): orchestrator hand-partitioned 16 tasks across 4 implementers by file. Two impls got tasks on
cmd/darken/doctor.go (B3+B5+B7) — they had to be the SAME impl to avoid trampling. Required careful manual planning and the partition was still imperfect (impl-2c/d/e merge commits show pulls from each other).
- Cost: ~30 min planning + risk of partition mistakes. With bones tasks queue: just fire 4 impls with the brief above; queue handles ordering.
Quick Reference
| Op | Command |
|---|
| Bootstrap | bones up -R <path> |
| Add task | bones tasks create "<title>" -R <path> |
| List | bones tasks list -R <path> |
| Claim | bones tasks claim <id> -R <path> |
| Close | bones tasks close <id> -R <path> |
| Watch | bones tasks watch -R <path> (events from NATS) |
Anti-Patterns
- Authoring tasks too granular (every commit a task) — keep tasks at logical-unit granularity
- Letting workers run after queue empty — they will idle, costing budget
- Forgetting to close — claims orphan, future workers will not pick them up
- Mixing claim/close with file-disjoint partitioning — pick one model, do not double-coordinate
