agent-fleet

name	agent-fleet
description	三省六部军团管理：实例生命周期监控、三层预算治理、熔断器控制、卡死检测。当需要查看运行中的智能体实例状态或管理军团资源时使用此技能。
tools	spawn_agent, handoff_to_agent
metadata	{"category":"operations","emoji":"🪖","tree_id":"subagents/spawn_agent","tree_group":"subagents","min_tier":"task_write","intent_priority":15,"intent_keywords":{"zh":["查看实例状态","智能体军团","智能体监控","智能体管理","终止实例","军团指标","预算控制","熔断器","卡死检测","agent fleet","实例运行状态"],"en":["agent fleet status","instance lifecycle","budget control","circuit breaker","stall detection","agent instances","terminate agent"]},"scene_hint":"查看智能体实例运行状态用此→多智能体协作编排用agent-orchestration→智能体定时触发用agent-scheduler"}

三省六部军团管理技能

核心概念

军团管理负责所有子智能体实例（由工厂从蓝图创建）的运行时生命周期，包括状态监控、预算控制、熔断保护和卡死检测。

实例状态机

running → completed (正常完成)
running → failed    (执行失败)
running → stalled   (卡死检测标记)
running → terminated (手动终止)
running → paused    (暂停, 预留)

RPC 操作

状态查询

agent.fleet.status → { blueprintsTotal, activeInstances, byBlueprint, byState }

实例终止

agent.fleet.terminate(instanceId="inst-xxx")

军团指标

agent.fleet.metrics → { budget: {...}, circuitBreaker: {...} }

军团审计

agent.fleet.audit → { events: [...], total: N }

三层预算体系

层级	配置来源	限制维度
全局	config	MaxTokensPerDay, MaxSessionsPerDay
蓝图级（每类蓝图的治理策略）	blueprint.budget	MaxSessionsPerDay, MaxConcurrent
实例级	contract.MaxTokens	单次会话 token 上限

预算耗尽时 → 拒绝 spawn + 返回原因。每日 UTC 0 点自动重置。

熔断器

每类蓝图的熔断保护（治理策略），复用 Media Patrol 的 ConsecutiveFailures 模式：

closed → 连续 5 次失败 → open → 冷却 10 分钟 → half_open → 成功 → closed
                                              → 失败 → open (重新熔断)

手动重置: agent.fleet.metrics 或管理面板操作。

卡死检测

基于心跳的 TTL 检测，复用 plan_confirmation TTL 清理循环模式：

• 每 30 秒扫描全部监控实例
• 超过蓝图的 stall_threshold 无活动 → 触发升级链
• 升级链: self(重试) → parent(通知父智能体) → human(远程审批)
• 升级链耗尽 → 标记卡死 + 移除监控

CLI 命令

crabclaw fleet status                # 查看军团状态概览
crabclaw fleet terminate <id>        # 终止指定实例
crabclaw fleet metrics               # 查看预算/熔断指标
crabclaw fleet audit                 # 查看审计日志