swarm
GitHubSwarm技能用于通过调用者选择的执行器,将明确且写作用域互斥的任务包精确分发一次。它确保并行任务独立运行、结果互不干扰,并提供严格的并发安全保证。
Trigger Scenarios
Install
npx skills add boshu2/agentops --skill swarm -g -y
SKILL.md
Frontmatter
{
"name": "swarm",
"description": "Dispatch explicit disjoint packets exactly once through a caller-selected executor. Triggers: \"swarm\", \"dispatch disjoint packets\", \"parallel explicit tasks\"."
}
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 factual results about N experiments rather than one tangle. Those results are not semantic verdicts.
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 developer-only reference implementation is scripts/dispatch_once.py.
Repository tests exercise this module; the supported native installed path
follows this skill through the caller-selected executor and does not invoke Python.
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. The adapter cannot select AgentOps semantics, issue a binding verdict, or turn factory completion into delivery or validation proof.
Version History
-
8061085
Current 2026-09-09 05:20
更新原生证据助手和基于模型家族的新审查默认值;引入意图快照、严格证据验证及原子判决存储功能。
-
7b07a7d
2026-08-19 22:04
新增对可移植性元数据保留、YAML重复键及符号链接包含的安全校验;修正部分语义描述,强化作用域规范与错误处理机制。
- 3f402e5 2026-07-24 22:10


