Swarm Agent Skill

Swarm

Swarm exposes one optional factory port:

dispatch_once(explicit_disjoint_packets, executor)
  -> per-packet candidate | evidence | error

The caller supplies every complete packet, proves their write scopes disjoint,

and chooses the executor. Swarm dispatches each packet once, preserves packet and

context identities, collects results, and stops.

Write scopes must be workspace-relative and canonical — symlink-resolved and

already normalized. The disjointness check is lexical: it case-folds prefixes so

scopes differing only by case are treated as a collision (safe on

case-insensitive filesystems), but it cannot see a symlink that aliases two

scopes onto one target. Supplying non-canonical or symlinked scopes forfeits the

disjointness guarantee; a non-empty write_scope.exclude is rejected, not

silently ignored, because the proof cannot honor it.

Exactly-once dispatch over proven-disjoint scopes is why parallel failures stay

independent: no packet can observe, block, or corrupt another, so N packets

yield N verdicts about N experiments rather than one tangle.

Named failure mode — partial-batch launch: dispatching valid packets before

discovering an invalid one, leaving the batch half-run; validate the entire

batch before the first call.

Anti-pattern: re-dispatching a packet whose executor returned an error.

Corrective: return the error as that packet's factual result; retry is the

caller's decision, not the dispatcher's.

The reference implementation is scripts/dispatch_once.py.

It validates the entire explicit batch before the first call, invokes the supplied

executor exactly once for each packet, and returns executor exceptions as factual

per-packet errors.

Swarm's own effect is invoking the selected executor once per packet; the real

blast radius rides on the packets. Each packet's transitive effects — whatever

its executor writes, runs, or reaches — are the caller's to declare on the

packet, not Swarm's to bound.

Swarm does not select work, create packets, schedule from a backlog, persist a

queue, claim ownership, retry, validate, integrate, close, use Git, or deliver.

Executor failures remain executor evidence and cannot become core phase or

verdict state.