harness-monitor-agent-batch
Skill0 starsChanged 8 months agoMITView on GitHub
主会话批量监控子代理状态(合并原 harness-monitor-subharness-agent),检测子代理完成,触发结果合并
In this repository33 other files it gives its agents.
- harness-10x-productivity-boostSkill
- harness-adjust-granularitySkill
- harness-analyze-project-scaleSkill
- harness-auto-full-executionSkill
- harness-brainstormSkill
- harness-build-context-indexSkill
- harness-build-core-manifestSkill
- harness-build-deterministic-workflowSkill
- harness-check-clean-stateSkill
- harness-configSkill
- harness-context-clearSkill
- harness-determine-granularitySkill
- harness-diagnose-and-improveSkill
- harness-enable-circuit-breakerSkill
- harness-enforce-architecture-guardrailsSkill
- harness-escalate-to-humanSkill
- harness-evaluate-performanceSkill
- harness-evolveSkill
- harness-executing-plansSkill
- harness-generate-feature-listSkill
- harness-goalSkill
- harness-handle-failureSkill
- harness-initSkill
- harness-merge-agent-resultsSkill
- harness-merge-subharness-resultSkill
- harness-monitor-subharness-agentSkill
- harness-planSkill
- harness-recover-clean-stateSkill
- harness-schedule-parallel-agentsSkill
- harness-session-handoffSkill
- harness-spawn-subharness-agentSkill
- harness-track-feature-progressSkill
- harness-validate-outputSkill
The file
Files
harness-monitor-agent-batch
Core Capabilities
主会话批量监控子代理状态(合并原 harness-monitor-subharness-agent),检测子代理完成,触发结果合并。
核心职责:
- 读取
GLOBAL_STATE.md(获取 active_subagents) - 批量检查子代理状态(增量式状态摘要)
- 检测子代理完成(status: COMPLETED)
- 调用
harness-merge-agent-results合并结果 - 更新
feature_list.json(passes: false → true) - 移除已完成的子代理(active_subagents)
- 输出状态报告
- Sleep 30 秒(避免频繁轮询)
- 循环监控(直到所有子代理完成)
Execution Steps
Step 1: Read GLOBAL_STATE.md(获取 active_subagents)
工具: Read
文件: .EnjoyHarness/GLOBAL_STATE.md
Token 消耗: ~200 tokens
目的: 获取 active_subagents 字段(运行中的子代理列表)Step 2: 批量检查子代理状态(增量式状态摘要)
工具: Bash + grep
命令: grep -E "^status:|^clean_state:" .subharness/*/STATUS.md
Token 消耗: ~50 tokens per 子代理(性能优化版)
示例输出:
.subharness/FEAT-001/STATUS.md:status: COMPLETED
.subharness/FEAT-001/STATUS.md:clean_state: true
.subharness/FEAT-002/STATUS.md:status: RUNNING
.subharness/FEAT-002/STATUS.md:clean_state: false
.subharness/FEAT-003/STATUS.md:status: FAILED
.subharness/FEAT-003/STATUS.md:clean_state: false
性能优化:
- 仅读取关键字段(status, clean_state)
- 不读取完整 STATUS.md
- Token 消耗降低 90%Step 3: 处理每个子代理状态
3.1 如果 status == COMPLETED
处理:
1. 调用 harness-merge-agent-results(合并结果)
2. 更新 feature_list.json(passes: false → true)
3. 移除 active_subagents(从 GLOBAL_STATE.md)
4. 清理子代理目录(可选)
5. 记录到 EVENT_LOG.md(SUBAGENT_COMPLETE)
Token 消耗: ~1000 tokens per 子代理3.2 如果 status == FAILED
处理:
1. 检查 error_count(从 STATUS.md)
2. 如果 error_count < 3 → 触发重试(harness-handle-failure)
3. 如果 error_count >= 3 → 先触发熔断与自动恢复;仅真实阻塞时升级 harness-escalate-to-human
4. 记录到 EVENT_LOG.md(SUBAGENT_FAILED)
Token 消耗: ~500 tokens3.3 如果 status == RUNNING
处理:
1. 检查运行时间(started_at 字段)
2. 如果超过 estimated_completion → 记录警告(可能超时)
3. 继续监控
Token 消耗: ~100 tokensStep 4: 输出状态报告
输出格式:
📊 子代理状态监控
总子代理数: 3
已完成: 1 (FEAT-001)
运行中: 1 (FEAT-002)
失败: 1 (FEAT-003, error_count: 2/3)
下一步:
- FEAT-001: 合并结果
- FEAT-002: 继续监控
- FEAT-003: 重试(失败次数 < 3)
Token 消耗: ~200 tokensStep 5: Sleep 30 秒(避免频繁轮询)
工具: Bash
命令: sleep 30
Token 消耗: ~0 tokens
性能优化:
- 轮询间隔:30 秒(而非 5 秒)
- 降低 Token 消耗 6 倍
- 降低 CPU 占用Step 6: 返回 Step 1(循环监控)
循环条件:
- active_subagents 不为空
- 或者有子代理正在重试
退出条件:
- 所有子代理已完成(active_subagents 为空)
- 或者触发熔断(error_count >= 3)Prerequisites
必须满足:
- ✅ 已执行
harness-schedule-parallel-agents(子代理已生成) - ✅
GLOBAL_STATE.md的active_subagents不为空
如果前置条件不满足:
- 如果无运行中的子代理 → 提示"无运行中的子代理"
- 如果子代理已全部完成 → 输出"所有子代理已完成"
Success Criteria
成功标准:
- ✅ 成功读取
GLOBAL_STATE.md(获取 active_subagents) - ✅ 成功批量检查子代理状态(增量式状态摘要)
- ✅ 正确处理不同状态(COMPLETED/FAILED/RUNNING)
- ✅ 成功合并已完成的子代理结果
- ✅ 成功更新
feature_list.json和GLOBAL_STATE.md - ✅ 输出详细的状态报告
失败情况:
- ❌ GLOBAL_STATE.md 读取失败 → 记录错误,退出监控
- ❌ 子代理状态检查失败 → 记录错误,继续监控其他子代理
- ❌ 结果合并失败 → 先触发自动失败处理
Failure Recovery
错误场景 1: 子代理失败(error_count < 3)
检测: status: FAILED, error_count < 3
处理:
1. 记录到 EVENT_LOG.md(WARNING | SUBAGENT_FAILED)
2. 触发重试(harness-handle-failure)
3. error_count +1
4. 继续监控错误场景 2: 子代理失败 3 次(error_count >= 3)
检测: status: FAILED, error_count >= 3
处理:
1. 记录到 EVENT_LOG.md(ERROR | SUBAGENT_FAILED_3_TIMES)
2. 触发熔断与自动失败处理
3. 移除 active_subagents
4. 退出监控错误场景 3: 子代理超时
检测: 运行时间超过 estimated_completion(默认 30 分钟)
处理:
1. 记录到 EVENT_LOG.md(WARNING | SUBAGENT_TIMEOUT)
2. 先触发 harness-handle-failure;仅真实阻塞时升级人工介入错误场景 4: 结果合并失败
检测: harness-merge-agent-results 返回错误
处理:
1. 记录到 EVENT_LOG.md(ERROR | MERGE_FAILED)
2. 触发 harness-handle-failureRelationships
Triggers (触发下游)
子代理完成:
- 触发 harness-merge-agent-results(合并结果)
子代理失败:
- 触发 harness-handle-failure(默认)
- 仅真实阻塞时触发 harness-escalate-to-human
所有子代理完成:
- 返回主会话(继续选择下一批功能)Triggered By (被谁触发)
触发时机:
- harness-schedule-parallel-agents 完成后
- 用户显式调用:"监控子代理状态"Dependencies (前置依赖)
必须依赖:
- harness-schedule-parallel-agents(子代理生成)
- harness-merge-agent-results(结果合并)Token 消耗分析
单次轮询(3个子代理):
- Read GLOBAL_STATE.md: ~200 tokens
- 增量式状态摘要: ~150 tokens(3 * 50 tokens)
- 处理状态: ~100 tokens
- 输出报告: ~200 tokens
总计: ~650 tokens
每小时轮询(30秒间隔):
- 轮询次数: 120 次
- Token 消耗: 650 * 120 = 78,000 tokens
性能对比:
- v1.0(5秒间隔,完整 STATUS.md): 360,000 tokens/小时
- v2.0(30秒间隔,增量式摘要): 78,000 tokens/小时
- 性能提升: 降低 78% ✅Examples
示例 1: 批量监控子代理(标准场景)
输入:
active_subagents: [FEAT-001, FEAT-002, FEAT-003]
子代理状态:
- FEAT-001: COMPLETED, clean_state: true
- FEAT-002: RUNNING, clean_state: false
- FEAT-003: RUNNING, clean_state: false
处理:
1. Read GLOBAL_STATE.md
2. 批量检查状态(Bash + grep)
3. 处理 FEAT-001(COMPLETED):
- 调用 harness-merge-agent-results
- 更新 feature_list.json
- 移除 active_subagents
4. 处理 FEAT-002, FEAT-003(RUNNING):
- 继续监控
5. 输出状态报告
6. Sleep 30 秒
7. 返回 Step 1(继续监控)
输出:
📊 子代理状态监控
总子代理数: 3
已完成: 1 (FEAT-001) ✅
运行中: 2 (FEAT-002, FEAT-003)
已处理:
- FEAT-001: 结果已合并,功能标记为完成
下一步:继续监控 FEAT-002, FEAT-003示例 2: 子代理失败并重试
输入:
active_subagents: [FEAT-001]
子代理状态:
- FEAT-001: FAILED, error_count: 1/3
处理:
1. 检测到 FAILED 状态
2. 检查 error_count: 1(< 3)
3. 触发重试(harness-handle-failure)
4. error_count +1 → 2
5. 记录到 EVENT_LOG.md
6. 继续监控
输出:
⚠️ 子代理失败检测
功能ID: FEAT-001
失败次数: 1/3
处理: 触发重试(harness-handle-failure)
下一步:继续监控重试结果示例 3: 所有子代理完成
输入:
active_subagents: []
所有子代理已完成
处理:
1. Read GLOBAL_STATE.md(active_subagents 为空)
2. 输出"所有子代理已完成"
3. 返回主会话
输出:
🎉 所有子代理已完成!
已完成功能数: 3
总功能数: 10
完成率: 30%
下一步:返回主会话,选择下一批功能示例 4: 子代理失败 3 次触发熔断
输入:
active_subagents: [FEAT-001]
子代理状态:
- FEAT-001: FAILED, error_count: 3/3
处理:
1. 检测到 FAILED 状态
2. 检查 error_count: 3(>= 3)
3. 触发熔断与自动恢复
4. 记录到 EVENT_LOG.md(ERROR | SUBAGENT_FAILED_3_TIMES)
5. 移除 active_subagents
6. 退出监控
输出:
🔴 子代理失败 3 次,触发熔断
功能ID: FEAT-001
失败次数: 3/3
处理: 先触发 harness-handle-failure;真实阻塞时再升级 harness-escalate-to-human
下一步:进入自动恢复或阻塞升级流程Implementation Notes
关键设计决策
- 增量式状态摘要
- 仅读取关键字段(status, clean_state)
- 使用 Bash + grep 提取
- Token 消耗降低 90%
- 批量合并检查
- 轮询间隔:30 秒(而非 5 秒)
- 降低轮询频率 6 倍
- Token 消耗降低 78%
- 失败重试机制
- 失败次数 < 3 → 自动重试
- 失败次数 >= 3 → 触发熔断
- 避免无限重试
性能优化
Token 优化:
- 增量式状态摘要: 降低 90%
- 批量合并检查: 降低 82%
- 总体 Token 优化: 降低 78%
轮询优化:
- 轮询间隔: 30 秒(而非 5 秒)
- CPU 占用: 降低 83%
- I/O 频率: 降低 83%Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

