| name | swarm-shaped-work |
| description | Use when planning a multi-piece task to decide whether to swarm (spawn parallel coordinator + worker sessions) or do the work serially in one session. Covers the heuristic, the topology, and the spin-up sequence. |
Swarm-Shaped Work
A "swarm" is a small set of opencode sessions on the same machine that cooperate on one outcome. One session is the coordinator (holds shared context, single point of contact for the human); the others are workers (each focused on one slice of the work).
This skill answers: when is a task swarm-shaped, what is each role for, and how do you spin one up.
When To Reach For A Swarm
The strongest signal is multiple repos with dependencies between them that require timing/coordination. Concrete shape from real work:
- Backend in repo A, frontend in repo B, proto definitions in repo C, BigQuery views in repo D — each repo has its own build, tests, conventions, and devloop. A change touches all four.
- Dependencies between them: proto defs must land before consumers; backend must deploy before the frontend can pull schema; etc.
- Each slice is too large to comfortably hold in one session alongside the others (separate context windows pay off).
Other signals that often accompany the above:
- The work has clear hand-off points between slices (worker-to-worker notifications matter).
- The total wall-clock time is long enough that parallelism is worth the coordination tax (rule of thumb: > ~1 hour of total work).
- Some slices need human-in-the-loop decisions you'd rather field one at a time through a single point of contact (the coordinator).
When NOT To Swarm
- Single-repo, single-subsystem changes. A bug fix in one codebase, even if it touches multiple files, is usually faster sequentially.
- You don't have a clear decomposition. If you can't write down "Worker A does X, Worker B does Y, here are the integration points", you don't have a swarm-shaped task — you have an exploratory task. Do that solo first; swarm later if a shape emerges.
- The slices race on the same files. Workers must own disjoint surface area or they'll fight for git locks and you'll spend more time merging than implementing.
- The coordination overhead exceeds the parallel speedup. A 4-worker swarm where each worker takes 5 minutes is probably not worth the spin-up + envelope traffic.
Roles
Coordinator
The coordinator is the shared-memory + human-interface role. Its jobs:
- Hold shared context. The Jira ticket, the design doc, the cross-cutting state. Workers don't all need to load the full ticket; they get the slice relevant to them.
- Single point of contact for the human. You message the coordinator; it distributes (or aggregates) for the swarm.
The coordinator is NOT a chokepoint for worker-to-worker traffic. Workers can swarm-message each other directly when only those two need to know. They loop the coordinator in only when the coordinator's shared-context role actually matters (cross-cutting state, human-facing decisions, integration timing across more than two parties).
Workers
Each worker owns one slice of the work — typically one repo or one subsystem. Its job:
- Plan and execute that slice, in its own session, with its own context window.
- Notify peers directly when it hits a hand-off point ("BE: my API is deployed; FE: pull schema now").
- Notify the coordinator when shared-context-relevant things happen ("Done with my slice", "Blocked on a decision I need the human to weigh in on").
Spin-Up Sequence
The mechanism is the existing tools — there's no swarm spawn command. You orchestrate it from a single planning session (or from a shell):
1. Decompose
Write down (in the planning session, on paper, in a doc — somewhere durable):
- The coordinator's purpose: "hold the design for PROJ-1234, field human decisions, broker integration timing".
- For each worker: which dir, what slice, what hand-off points it owns.
- The communication graph: who needs to know what from whom. (E.g. "worker A → worker B at proto-published, worker B → worker C at deploy-complete".)
2. Launch the coordinator
opencode-launch <coordinator-dir> "$(cat <<'PROMPT'
You are the coordinator for PROJ-1234. Read the design doc at docs/plans/...
You will spawn workers and broker decisions. Workers will message you via
swarm IPC. The human will reach you via Telegram.
Workers (you'll receive their session ids shortly):
- BE: <repo>
- FE: <repo>
- protos: <repo>
- dbt: <repo>
Plan: ...
PROMPT
)"
Capture the coordinator's session id from the output.
3. Launch each worker
For each worker, launch from its own dir with a prompt that includes:
- Its slice of the work.
- The coordinator's session id (so it knows where to escalate).
- Optional: the other workers' session ids if it has a known direct hand-off (e.g. BE → FE for "schema is live").
Workers are writable — launch them with --worktree <slug>. A swarm worker
edits code, so it must NOT start in a repo's primary root (in mono that's the
read-only trunk the guard protects). --worktree <slug> runs work <slug> in
<worker-dir> and lands the session in a fresh .worktrees/<slug> off the
local trunk, so the worker is isolated and the read-only-main guard is bypassed
by construction. Give each worker a distinct slug (its role/ticket) to avoid
collisions. The worktree is reclaimed automatically once its branch merges (the
nightly reset-workspace runs work --prune-merged).
opencode-launch --worktree be-proj-1234 <worker-dir> "$(cat <<'PROMPT'
You are the BE worker for PROJ-1234. Your slice: implement the GraphQL
endpoints for X. You are in a fresh worktree off trunk — commit here and open
a PR from this branch when done.
Coordinator: ses_<coordinator-id>
Other workers: FE=ses_<fe-id>, protos=ses_<protos-id>
When your API is deployed, use the `swarm_send` tool (to=ses_<fe-id>, kind=status.update, message="API live at /v2/foo") so the FE can proceed.
Plan: ...
PROMPT
)"
Capture each worker's session id.
The coordinator (previous step) is read-only — it holds context and
brokers decisions, it doesn't edit — so launch it WITHOUT --worktree, at the
project root. Only writable workers get a worktree.
4. Tell the coordinator the worker ids
After launches, send the worker ids to the coordinator so it can name them in its mental model. From your planning session, call the swarm_send tool (to=ses_) with a message like:
Workers spawned:
- BE: ses_<be-id>
- FE: ses_<fe-id>
- protos: ses_<protos-id>
- dbt: ses_<dbt-id>
5. Kick off the work
Either:
- Tell the coordinator to begin orchestrating ("the swarm is live; please plan and dispatch initial assignments with the
swarm_send tool, kind=task.assign"), OR
- Tell each worker to begin its slice directly ("you may now start; report status to the coordinator").
The first style centralizes orchestration; the second parallelizes earlier. Pick based on how much the coordinator needs to gate vs. just observe.
Communication Patterns
Once the swarm is live, all cross-session messaging goes through the swarm_send tool (see the swarm-messaging skill for the full protocol). from is filled in automatically from the calling session.
Useful conventions (all are swarm_send calls):
- Coordinator broadcasts assignments:
to=ses_<worker>, kind=task.assign, priority=urgent, message="...".
- Workers report status to coordinator:
to=ses_<coordinator>, kind=status.update, message="Done with <X>; PR at <url>".
- Workers coordinate directly:
to=ses_<peer-worker>, kind=status.update, message="API deployed" — no need to bounce off the coordinator if only the peer needs to know.
- Threading replies: set
reply_to=<their-msg-id> so receivers can follow conversation chains.
- Discovery: don't have a peer's id? Call
swarm_list.
- Backlog / replay: receivers can call
swarm_read to fetch their inbox if they suspect they missed a message.
Tear-Down
When the work is done:
curl -sf -X DELETE http://127.0.0.1:4096/session/ses_<coordinator-id>
curl -sf -X DELETE http://127.0.0.1:4096/session/ses_<worker-id>
/kill <session-id>
Or just let the session reaper expire them after 1 week of inactivity.
Old swarm messages stay in the daemon's swarm_messages table. They aren't auto-cleaned yet (see pigeon's swarm-operations skill for the manual cleanup query). Stale/queued messages targeting deleted sessions will exhaust their retry budget and terminally fail — this is fine.
Anti-Patterns
- Spinning up workers without giving them each other's ids. They can't coordinate directly and end up funneling everything through the coordinator, defeating the parallelism.
- Using the coordinator as a message broker. If a message only matters to two workers, send it directly. The coordinator's job is shared context, not packet routing.
- Spawning more workers than there are decoupled slices. Two workers fighting over the same code is worse than one worker doing both.
- Forgetting
--reply-to in chained conversations. Without threading, the receiver has to reconstruct context from prose. Cheap to set; expensive to omit.
- Not telling workers about the coordinator. Workers that don't know how to escalate will silently get stuck or make decisions that should have been the human's.
See Also
opencode-launch — spawn headless sessions.swarm-messaging — sender + receiver protocol; the swarm_send/swarm_read/swarm_list tools; envelope format; kinds; replay.- pigeon repo
swarm-architecture / swarm-operations skills — daemon internals if you need to debug delivery.
