beo-swarming

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Swarming

集群协作（Swarming）

Role Boundary: Read First

角色边界：请先阅读

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

beo-executing

skill's job.

beo-swarming = launches and tends workers (this skill)
beo-executing = each worker's self-routing implementation loop

The orchestrator launches the swarm, then tends it. Workers decide what to do next by using the live bead graph against the approved current phase.

<HARD-GATE> If Agent Mail is unavailable, do NOT attempt swarming. Degrade to `beo-executing`. </HARD-GATE>

If Agent Mail becomes unavailable mid-run, stop launching new workers, let active workers finish if possible, and degrade remaining work to

beo-executing

你是编排器（ORCHESTRATOR）。你负责启动工作节点、监控协作流程、处理升级事件，保证集群持续运转。你不能实现bead任务。如果你发现自己在编辑源文件，请立即停止，这是

beo-executing

技能的职责。

beo-swarming = 启动并管理工作节点（即本技能）
beo-executing = 每个工作节点的自路由实现循环

编排器启动集群后对其进行运维管理。工作节点会基于实时bead图谱，对照已获批的当前阶段来决定下一步执行内容。

<HARD-GATE> 如果Agent Mail不可用，不要尝试启动集群。降级使用`beo-executing`。 </HARD-GATE>

如果运行过程中Agent Mail不可用，停止启动新的工作节点，尽可能让活跃工作节点完成当前任务，剩余工作降级使用

beo-executing

执行。

Active Swarm Never Idles

活跃集群永不闲置

If workers are spawned, online, busy, blocked, or expected to report, you are not in a waiting phase. You are in a tending phase.

While the swarm is active, keep looping through Agent Mail and the live bead graph. Do not stop and wait for user direction just because the thread is quiet. Silence is work for the orchestrator.

如果工作节点处于已生成、在线、忙碌、阻塞或待上报状态，你就不处于等待阶段，而是处于运维阶段。

集群活跃期间，持续循环检查Agent Mail和实时bead图谱。不要仅仅因为线程没有消息就停下来等待用户指示。静默期是编排器的工作时间。

Minimal Swarm Checklist

最小集群检查清单

Confirm the epic is approved and at least 3 independent ready current-phase tasks exist
Confirm Agent Mail is working before spawning any workers
Register the coordinator and announce the swarm
Spawn bounded workers with explicit current-phase scope
Monitor completions, blockers, idle workers, and file conflicts
Reassign, rescue, or degrade when workers stall
Hand off to
```
beo-planning
```
or
```
beo-reviewing
```
only after the approved current execution scope is complete

确认epic已获批，且当前阶段至少存在3个独立就绪任务
生成任何工作节点前确认Agent Mail运行正常
注册协调器并公告集群启动
生成带有明确当前阶段范围的受限工作节点
监控完成情况、阻塞项、空闲工作节点和文件冲突
当工作节点停顿时进行重分配、救援或降级处理
仅当已获批的当前执行范围全部完成后，才流转至
```
beo-planning
```
或
```
beo-reviewing
```

Worker Count Heuristic

工作节点数量启发式规则

Start with one worker per independent ready track, capped by coordination safety rather than optimism.

3 ready tracks -> usually 2-3 workers
4-6 ready tracks -> usually 3-4 workers
Shared-file hotspots or fragile infrastructure -> bias downward
Never spawn more workers than independent ready tracks

If unsure, start smaller and expand only after the first monitor loop is healthy.

初始按每个独立就绪路径分配1个工作节点，上限由协调安全性而非乐观预期决定。

3条就绪路径 -> 通常2-3个工作节点
4-6条就绪路径 -> 通常3-4个工作节点
共享文件热点或基础设施不稳定 -> 适当降低数量
生成的工作节点数量永远不要超过独立就绪路径数

如果不确定，先从小规模开始，只有在第一轮监控循环确认状态健康后再扩容。

Rescue and Degrade Rules

救援与降级规则

If a worker is silent for >10 poll cycles, ping once and inspect Agent Mail, reservations, and graph state
If the worker is still silent after one rescue attempt, stop assigning it new work and re-route the bead
If the same file conflict happens twice, pause that lane and escalate decomposition or ownership clarity
If 2 or more workers stall, or coordination costs exceed progress, stop expanding the swarm and degrade remaining work to
```
beo-executing
```

如果工作节点超过10个轮询周期无响应，发送一次ping并检查Agent Mail、预留任务和图谱状态
如果一次救援尝试后工作节点仍无响应，停止为其分配新任务并重新路由对应bead
如果同一文件冲突发生两次，暂停该链路，升级申请分解任务或明确所有权
如果有2个或更多工作节点停顿，或协调成本超过进度收益，停止集群扩容，剩余工作降级使用
```
beo-executing
```
执行

Phase 1: Confirm Swarm Readiness

阶段1：确认集群就绪

Load

references/swarming-operations.md

for the exact readiness checks, planning-aware scope verification, epic-claim step, and scheduling cascade.

加载

references/swarming-operations.md

查看准确的就绪检查、感知规划的范围验证、epic认领步骤和调度级联规则。

Phase 2: Initialize Agent Mail

阶段2：初始化Agent Mail

Load

references/swarming-operations.md

for the exact Agent Mail setup sequence, coordinator registration, and thread bootstrap pattern. Use

references/message-templates.md

for message bodies.

加载

references/swarming-operations.md

查看准确的Agent Mail设置顺序、协调器注册和线程启动模式。使用

references/message-templates.md

获取消息正文模板。

Phase 3: Spawn Workers

阶段3：生成工作节点

Load

references/swarming-operations.md

for the exact worker-spawn contract, worker input shape, worker startup acknowledgment expectations, and

STATE.md

tracking expectations. Use

references/worker-template.md

when building worker context.

加载

references/swarming-operations.md

查看准确的工作节点生成约定、工作节点输入格式、工作节点启动确认要求和

STATE.md

跟踪要求。构建工作节点上下文时使用

references/worker-template.md

。

Phase 4: Monitor + Tend

阶段4：监控 + 运维

Load

references/swarming-operations.md

for the exact monitor/tend loop, event types, progress heuristics, and checkpoint mechanics. Use

references/message-templates.md

for mail content.

加载

references/swarming-operations.md

查看准确的监控/运维循环、事件类型、进度启发式规则和检查点机制。邮件内容使用

references/message-templates.md

。

Phase 5: Swarm Complete

阶段5：集群完成

Load

references/swarming-operations.md

for the exact completion checks, planning-aware route decision,

STATE.md

update, and Agent Mail completion announcement.

加载

references/swarming-operations.md

查看准确的完成检查、感知规划的路由决策、

STATE.md

更新和Agent Mail完成公告规则。

Red Flags

危险信号

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 current-phase beads exist: check mail, reservations, or startup drift
No Agent Mail activity for >10 poll cycles: workers may be stuck or context-exhausted
The same file conflict repeats: bead decomposition may be too coarse; escalate
Workers stop using the live graph and start freelancing: re-broadcast the execution contract
Build/test failures accumulate without intervention: create fix beads or stop and escalate
Current phase completes but later phases remain and you route directly to final review: planning-aware routing failure

如果出现以下情况，请停止操作排查问题后再继续：

工作节点同时实现多个bead：自路由不代表单个工作节点内部并行执行
编排器编辑源文件：角色违规
工作节点空闲但存在当前阶段就绪bead：检查邮件、预留任务或启动漂移
超过10个轮询周期无Agent Mail活动：工作节点可能卡住或上下文耗尽
相同文件冲突重复出现：bead分解可能粒度过粗，需升级处理
工作节点不再使用实时图谱开始自行执行：重新广播执行约定
构建/测试失败持续累积无人干预：创建修复bead或停止操作升级处理
当前阶段完成但后续阶段仍存在，你直接路由至最终评审：感知规划的路由失败

Reference Files

参考文件

Load when needed:

File	Load When
`references/worker-template.md`	Spawning any worker (Phase 3)
`references/message-templates.md`	Posting or parsing Agent Mail messages

按需加载：

文件	加载时机
`references/worker-template.md`	生成任何工作节点时（阶段3）
`references/message-templates.md`	发布或解析Agent Mail消息时