ruflo 循环工作器实战指南用 Claude Code 原生 /loop 与 ScheduleWakeup 打造缓存感知的后台自动化【免费下载链接】ruflo The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated项目地址: https://gitcode.com/GitHub_Trending/cl/ruflo导读本文围绕 ruflo 生态中的ruflo-loop-workers插件与loop-workerSkill 展开讲解如何利用 Claude Code 原生的/loop会话内调度机制与ScheduleWakeup工具驱动 audit、optimize、consolidate、predict、map、testgaps 等 12 类后台工作器周期性运行。核心要点是缓存感知cache-aware通过min(270, cache_ttl * 0.9)的延迟公式让每次唤醒都能命中 prompt cache从而显著降低重复会话的推理开销。读完本文你将掌握 loop-worker Skill 的完整调用协议、MCP 工作器分发的正确姿势、270s 心跳契约的来龙去脉以及 loop 与 cron 两种调度模式的选择依据。背景ruflo 的后台工作器子strate 插件ruflo-loop-workers是 ruflo 插件家族中负责周期性任务的 substrate 插件README 中明确其定位Cache-aware /loop workers and CronCreate background automation. Substrate plugin for every recurring task in the ruflo familyplugins/ruflo-loop-workers/README.md。它并不自己实现 worker 的具体业务逻辑而是提供调度协议与触发契约真正执行任务的消费者插件如ruflo-security-audit、ruflo-testgen、ruflo-knowledge-graph、ruflo-docs等监听各自对应的 trigger 名称。插件清单 plugins/ruflo-loop-workers/.claude-plugin/plugin.json 将自身描述为Cache-aware /loop workers and CronCreate background automation — wraps 5hooks_worker-*MCP tools (list/dispatch/status/detect/cancel) and exposes 12 background worker triggers插件依赖ruflo-core提供 MCP serverCLI 侧固定使用claude-flow/cliv3.6majorminor pin并由bash plugins/ruflo-loop-workers/scripts/smoke.sh作为契约验证入口。loop-worker Skill会话内周期工作器的正确用法1. Skill 元数据与能力边界loop-workerSkill 的定义位于 plugins/ruflo-loop-workers/skills/loop-worker/SKILL.md其 frontmatter 如下--- name: loop-worker description: Run Ruflo background workers using Claude Code native /loop scheduling argument-hint: worker-name allowed-tools: Bash(npx *) mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch mcp__plugin_ruflo-core_ruflo__hooks_worker-status ScheduleWakeup ---值得注意的三点argument-hint: worker-name调用该 Skill 时必须传入工作器名称即下文的 trigger 名。allowed-tools严格收窄Skill 只被授予三类能力——受限的Bash(npx *)、两个 ruflo-core MCP 工具worker-dispatch、worker-status、以及 Claude Code 原生的ScheduleWakeup。smoke 契约第 12 条专门校验skills 中不得出现allowed-tools: *通配授权见 plugins/ruflo-loop-workers/scripts/smoke.sh这是安全设计的一部分。核心流程只有两步先通过 MCP 分发一次 worker 运行再用ScheduleWakeup排定下一次迭代。2. 缓存感知的延迟公式SKILL.md 给出了 loop-worker 的灵魂公式Delay formula: min(270, cache_ttl * 0.9) -- default 270s含义如下Claude Code 的 prompt cache TTL 约为5 分钟300 秒取 TTL 的 90% 即 270 秒并设置 270s 上限保证下一次ScheduleWakeup唤醒时上下文仍处于缓存命中窗口内延迟超过 300 秒会支付一次 cache miss而恰好取整到 5 分钟则是最差情况——既多等了时间又可能因边界抖动导致缓存失效。这一点在 plugins/ruflo-loop-workers/README.md 中被进一步明确Recommended fallback heartbeat is270 seconds— under the 5-minute prompt-cache TTL so the next wake-up reads conversation context cached. Going past 300s pays a cache-miss; rounding to 5 minutes is the worst-of-both case.该 270s 心跳契约的所有权属于ruflo-autopilot的 ADR-0001而ruflo-loop-workers是承载这一契约的 substrate 层。ADR-0001plugins/ruflo-loop-workers/docs/adrs/0001-loop-workers-contract.md将其列为关键 cross-link 修正项。3. 标准调用序列SKILL.md 定义的标准执行序列只有两步1. Dispatch a worker via MCP: mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch({ trigger: WORKER_NAME }) 2. Then schedule the next iteration: ScheduleWakeup({ delaySeconds: 270, reason: next WORKER_NAME iteration })即分发 → 排程。先触发一次真实运行再安排下一次唤醒形成自驱动的循环。分发走 MCP 工具由 ruflo-core 的 MCP server 提供排程走 Claude Code 原生ScheduleWakeup两者各司其职。对应的 CLI 形式README 中的调用模式为# CLI 分发 worker npx claude-flow/clilatest hooks worker dispatch --trigger document --scope api # MCP 等价调用 mcp tool call hooks_worker-dispatch --json -- {trigger: document, scope: api}注意 MCP 工具名带mcp__plugin_ruflo-core_ruflo__前缀即hooks_worker-dispatch其实现位于 v3/claude-flow/cli/src/mcp-tools/hooks-tools.tshooks_worker-dispatch定义于约第 4502 行hooks_worker-status定义于约第 4641 行。从源码看dispatch 支持守护进程模式当 worker 被排队时会写入 daemon queue 并由 daemon 每 5 秒轮询一次处理后的条目移入.claude-flow/daemon-queue/.processed/随后可通过hooks_worker-status轮询直到status completed——这为长耗时任务提供了异步完成路径。4. 可用工作器清单SKILL.md 内置了 8 个高频工作器loop 模式下WorkerIntervalDescriptionaudit270sSecurity analysisoptimize270sPerformance optimizationconsolidate600sMemory consolidationpredict270sPredictive preloadingmap600sCodebase mappingtestgaps270sTest coverage analysisdocument600sAuto-documentationbenchmark600sPerformance benchmarking而插件整体暴露12 个 trigger。完整 12 项在loop-worker-coordinatoragentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md与/ruflo-loop命令plugins/ruflo-loop-workers/commands/ruflo-loop.md中有更细的优先级划分Worker优先级Triggerloop 间隔auditcriticalaudit270soptimizehighoptimize270sconsolidatelowconsolidate600spredictnormalpredict270smapnormalmap270scoordinator/ 600sSKILLtestgapsnormaltestgaps270sdocumentnormaldocument600sbenchmarknormalbenchmark600sdeepdivenormaldeepdive270srefactornormalrefactor270sultralearnnormalultralearn270spreloadlowpreload600s注map在 SKILL.md 中标注 600s、coordinator 中标注 270s两处为不同粒度的建议值SKILL.md 是会话内使用手册coordinator 则面向统一调度实际间隔可按需调整只要遵守 270s 缓存友好的默认心跳即可。5. MCP 响应提示LOOP_SUGGESTION 与 CRON_SUGGESTIONSKILL.md 末尾明确MCP tool responses include[LOOP_SUGGESTION]and[CRON_SUGGESTION]hints -- follow them.即hooks_worker-dispatch/hooks_worker-status的返回结果中会携带调度建议标记[LOOP_SUGGESTION]提示适合用/loop会话内自驱的排程方式[CRON_SUGGESTION]提示适合改用 CronCreate 持久化调度。调用方应当遵循这些提示而非机械地固定使用某一种模式。5 个 hooks_worker-* MCP 工具ruflo-loop-workers包装的是 ruflo-core 提供的 5 个hooks_worker-*MCP 工具定义于 v3/claude-flow/cli/src/mcp-tools/hooks-tools.tsADR 给出了约略行号3489、3538、3656、3714、3923ToolPurposehooks_worker-list列出可用 worker 及其 triggerhooks_worker-dispatch以--trigger worker-name可选--scope分发一次 worker 运行hooks_worker-status检查正在运行的 worker 状态hooks_worker-detect基于上下文检测当前应触发哪些 workerhooks_worker-cancel取消正在运行的 workerworker-detect特别适合上下文驱动场景不手动指定 trigger而是让 MCP 层根据当前会话上下文判断该跑哪个 worker随后再 dispatch。这一组工具同时被 CLI 侧npx claude-flow/clilatest hooks worker ...命令族封装源码见 v3/claude-flow/cli/src/commands/hooks.ts 与 v3/claude-flow/cli/src/services/worker-daemon.ts。两种调度模式/loop 与 CronCreateruflo-loop-workers同时支持两种执行模式loop-workerSkill 对应/loop模式cron-scheduleSkillplugins/ruflo-loop-workers/skills/cron-schedule/SKILL.md对应持久化模式。模式选择原则cron-scheduleSkill 给出了简明判据/loop会话内、缓存感知、自我节拍self-pacing适合活跃开发期CronCreate持久化、可跨会话存活适合CI / 监控场景。CronCreate 持久化用法CronCreate({ schedule: */15 * * * *, prompt: Run security audit worker via mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch })cron-schedule Skill 推荐的默认表达式WorkerCronDescriptionaudit*/15 * * * *Security scanningoptimize*/30 * * * *Performance optimizationconsolidate0 * * * *Memory consolidationmap*/30 * * * *Codebase mappingtestgaps*/15 * * * *Test coverage analysisdocument0 */2 * * *API documentation/ruflo-schedule命令plugins/ruflo-loop-workers/commands/ruflo-schedule.md封装了这一过程用法为/schedule worker [cron-expression]默认 cronaudit、testgaps →*/15 * * * *optimize、map →*/30 * * * *consolidate、document →0 * * * *示例/schedule audit */15 * * * *等价于创建CronCreate(audit, */15 * * * *, Run security audit worker)。cron 模式下 coordinator 推荐的完整间隔可参见上文 12-worker 表如 audit*/15、predict*/15、ultralearn*/15、deepdive/refactor*/30、preload0 * * * *等。loop-worker-coordinator统一调度的编排者插件附带的loop-worker-coordinatoragentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md推荐model: haiku以控制成本负责跨两种模式的 worker 生命周期管理其职责为通过mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch以正确 trigger 分发 worker通过mcp__plugin_ruflo-core_ruflo__hooks_worker-status监控健康状态并上报失败用ScheduleWakeuploop 模式或CronCreate持久化模式排定下一次迭代遵守缓存 TTL——默认 270s 延迟保持 prompt cache 温暖5 分钟 TTL × 0.9。其工作流为1. 检查当前 worker 状态npx claude-flow/clilatest hooks worker status 2. 分发需要的 worker npx claude-flow/clilatest hooks worker dispatch --trigger WORKER_NAME 3. 按执行模式排定下一次检查此外coordinator 还接入神经学习回路任务完成后通过hooks post-task --task-id TASK_ID --success true --quality 0.9记录成功模式并通过memory search --query TASK_TYPE patterns --namespace patterns检索既有模式实现调度经验的持续积累。worker 触发 → 消费者插件的契约映射ADR-0001 与 README 共同固化了12 worker trigger → consumer plugin的归属表这是调度契约的核心每个 trigger 由哪个插件消费是单一权威来源消费者插件可据此校验自身 trigger 命名TriggerConsumer pluginPurposeultralearnruflo-intelligence从深度代码扫描引导学习语料optimizeruflo-cost-tracker,ruflo-intelligence性能 成本优化建议consolidateruflo-intelligence,ruflo-agentdbEWC 记忆整合predictruflo-intelligence面向后续任务的预测性路由auditruflo-security-audit,ruflo-aidefence安全与合规审计mapruflo-knowledge-graph构建/刷新实体关系知识图谱preloadruflo-core,ruflo-rag-memory高频操作前预热缓存deepdiveruflo-goals(deep-research)多源深度调查documentruflo-docs生成 API 文档 漂移检测refactorruflo-jujutsu差异感知的重构建议benchmarkruflo-cost-tracker,ruflo-iot-cognitum性能基准测试testgapsruflo-testgen覆盖率缺口检测 测试生成命名空间协调worker-history插件声明对 AgentDB 的worker-history命名空间的所有权kebab-case遵循 plugins/ruflo-agentdb/docs/adrs/0001-agentdb-optimization.md 的命名空间约定。该命名空间记录分发事件、耗时与成功/失败判定通过memory_*工具按命名空间路由访问。ADR 同时提醒不得遮蔽保留命名空间pattern、claude-memories、default。安装与契约验证安装/plugin marketplace add ruvnet/ruflo /plugin install ruflo-loop-workersruflo前置依赖ruflo-core插件提供 MCP server。契约验证bash plugins/ruflo-loop-workers/scripts/smoke.sh # Expected: 12 passed, 0 failedsmoke.sh作为smoke-as-contract门禁执行 12 项结构检查见 plugins/ruflo-loop-workers/scripts/smoke.sh覆盖plugin.json版本 0.2.1 且包含mcp、background-workers、cache-aware、schedule-wakeup关键词两个 Skillloop-worker、cron-schedule均含name:/description:/allowed-tools:字段且 agent 与 2 个命令文件存在全部 5 个hooks_worker-*工具在文档中被引用12 个 worker trigger 均在 README 中记录README 固定claude-flow/cliv3.6README 引用 ruflo-agentdb 命名空间约定worker-history命名空间已声明270s 缓存感知 ScheduleWakeup 已有文档存在对 ruflo-autopilot 270s 心跳契约的交叉引用worker-trigger → consumer-plugin 归属表存在含ruflo-docs、ruflo-testgenADR-0001 存在且状态为 Acceptedskills 中无通配工具授权。最佳实践小结默认 270s 心跳所有/loop场景优先使用 270s 延迟min(270, cache_ttl * 0.9)既低于 5 分钟缓存 TTL 又留有抖动余量超过 300s 将付出缓存未命中代价。先 dispatch 再 ScheduleWakeup每次迭代遵循分发 → 排程两步协议形成自驱动循环。事件驱动 安全网对于事件驱动型循环可先 arm 一个Monitor让 270s 唤醒充当兜底安全网README 明确推荐此模式。跟随响应提示MCP 返回中的[LOOP_SUGGESTION]/[CRON_SUGGESTION]是调度模式选择的直接信号应遵循执行。跨会话任务用 cron需要持久化、跨重启存活的 CI/监控类任务走 CronCreate活跃开发中的周期任务走/loop。尊重命名空间边界worker-history属本插件所有勿覆盖pattern、claude-memories、default等保留命名空间。验证契约修改插件文档或结构后运行bash plugins/ruflo-loop-workers/scripts/smoke.sh确保 12 项检查全部通过。延伸阅读插件契约全文plugins/ruflo-loop-workers/docs/adrs/0001-loop-workers-contract.md插件总览与安装说明plugins/ruflo-loop-workers/README.md持久化调度 Skillplugins/ruflo-loop-workers/skills/cron-schedule/SKILL.md协调 Agentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md命令封装plugins/ruflo-loop-workers/commands/ruflo-loop.md、plugins/ruflo-loop-workers/commands/ruflo-schedule.mdMCP 工具实现v3/claude-flow/cli/src/mcp-tools/hooks-tools.ts270s 心跳契约归属方ruflo-autopilotADR-0001plugins/ruflo-autopilot/docs/adrs/0001-autopilot-contract.md【免费下载链接】ruflo The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated项目地址: https://gitcode.com/GitHub_Trending/cl/ruflo创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考 ruflo 循环工作器实战指南：用 Claude Code 原生 /loop 与 ScheduleWakeup 打造缓存感知的后台自动化
ruflo 循环工作器实战指南用 Claude Code 原生 /loop 与 ScheduleWakeup 打造缓存感知的后台自动化【免费下载链接】ruflo The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated项目地址: https://gitcode.com/GitHub_Trending/cl/ruflo导读本文围绕 ruflo 生态中的ruflo-loop-workers插件与loop-workerSkill 展开讲解如何利用 Claude Code 原生的/loop会话内调度机制与ScheduleWakeup工具驱动 audit、optimize、consolidate、predict、map、testgaps 等 12 类后台工作器周期性运行。核心要点是缓存感知cache-aware通过min(270, cache_ttl * 0.9)的延迟公式让每次唤醒都能命中 prompt cache从而显著降低重复会话的推理开销。读完本文你将掌握 loop-worker Skill 的完整调用协议、MCP 工作器分发的正确姿势、270s 心跳契约的来龙去脉以及 loop 与 cron 两种调度模式的选择依据。背景ruflo 的后台工作器子strate 插件ruflo-loop-workers是 ruflo 插件家族中负责周期性任务的 substrate 插件README 中明确其定位Cache-aware /loop workers and CronCreate background automation. Substrate plugin for every recurring task in the ruflo familyplugins/ruflo-loop-workers/README.md。它并不自己实现 worker 的具体业务逻辑而是提供调度协议与触发契约真正执行任务的消费者插件如ruflo-security-audit、ruflo-testgen、ruflo-knowledge-graph、ruflo-docs等监听各自对应的 trigger 名称。插件清单 plugins/ruflo-loop-workers/.claude-plugin/plugin.json 将自身描述为Cache-aware /loop workers and CronCreate background automation — wraps 5hooks_worker-*MCP tools (list/dispatch/status/detect/cancel) and exposes 12 background worker triggers插件依赖ruflo-core提供 MCP serverCLI 侧固定使用claude-flow/cliv3.6majorminor pin并由bash plugins/ruflo-loop-workers/scripts/smoke.sh作为契约验证入口。loop-worker Skill会话内周期工作器的正确用法1. Skill 元数据与能力边界loop-workerSkill 的定义位于 plugins/ruflo-loop-workers/skills/loop-worker/SKILL.md其 frontmatter 如下--- name: loop-worker description: Run Ruflo background workers using Claude Code native /loop scheduling argument-hint: worker-name allowed-tools: Bash(npx *) mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch mcp__plugin_ruflo-core_ruflo__hooks_worker-status ScheduleWakeup ---值得注意的三点argument-hint: worker-name调用该 Skill 时必须传入工作器名称即下文的 trigger 名。allowed-tools严格收窄Skill 只被授予三类能力——受限的Bash(npx *)、两个 ruflo-core MCP 工具worker-dispatch、worker-status、以及 Claude Code 原生的ScheduleWakeup。smoke 契约第 12 条专门校验skills 中不得出现allowed-tools: *通配授权见 plugins/ruflo-loop-workers/scripts/smoke.sh这是安全设计的一部分。核心流程只有两步先通过 MCP 分发一次 worker 运行再用ScheduleWakeup排定下一次迭代。2. 缓存感知的延迟公式SKILL.md 给出了 loop-worker 的灵魂公式Delay formula: min(270, cache_ttl * 0.9) -- default 270s含义如下Claude Code 的 prompt cache TTL 约为5 分钟300 秒取 TTL 的 90% 即 270 秒并设置 270s 上限保证下一次ScheduleWakeup唤醒时上下文仍处于缓存命中窗口内延迟超过 300 秒会支付一次 cache miss而恰好取整到 5 分钟则是最差情况——既多等了时间又可能因边界抖动导致缓存失效。这一点在 plugins/ruflo-loop-workers/README.md 中被进一步明确Recommended fallback heartbeat is270 seconds— under the 5-minute prompt-cache TTL so the next wake-up reads conversation context cached. Going past 300s pays a cache-miss; rounding to 5 minutes is the worst-of-both case.该 270s 心跳契约的所有权属于ruflo-autopilot的 ADR-0001而ruflo-loop-workers是承载这一契约的 substrate 层。ADR-0001plugins/ruflo-loop-workers/docs/adrs/0001-loop-workers-contract.md将其列为关键 cross-link 修正项。3. 标准调用序列SKILL.md 定义的标准执行序列只有两步1. Dispatch a worker via MCP: mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch({ trigger: WORKER_NAME }) 2. Then schedule the next iteration: ScheduleWakeup({ delaySeconds: 270, reason: next WORKER_NAME iteration })即分发 → 排程。先触发一次真实运行再安排下一次唤醒形成自驱动的循环。分发走 MCP 工具由 ruflo-core 的 MCP server 提供排程走 Claude Code 原生ScheduleWakeup两者各司其职。对应的 CLI 形式README 中的调用模式为# CLI 分发 worker npx claude-flow/clilatest hooks worker dispatch --trigger document --scope api # MCP 等价调用 mcp tool call hooks_worker-dispatch --json -- {trigger: document, scope: api}注意 MCP 工具名带mcp__plugin_ruflo-core_ruflo__前缀即hooks_worker-dispatch其实现位于 v3/claude-flow/cli/src/mcp-tools/hooks-tools.tshooks_worker-dispatch定义于约第 4502 行hooks_worker-status定义于约第 4641 行。从源码看dispatch 支持守护进程模式当 worker 被排队时会写入 daemon queue 并由 daemon 每 5 秒轮询一次处理后的条目移入.claude-flow/daemon-queue/.processed/随后可通过hooks_worker-status轮询直到status completed——这为长耗时任务提供了异步完成路径。4. 可用工作器清单SKILL.md 内置了 8 个高频工作器loop 模式下WorkerIntervalDescriptionaudit270sSecurity analysisoptimize270sPerformance optimizationconsolidate600sMemory consolidationpredict270sPredictive preloadingmap600sCodebase mappingtestgaps270sTest coverage analysisdocument600sAuto-documentationbenchmark600sPerformance benchmarking而插件整体暴露12 个 trigger。完整 12 项在loop-worker-coordinatoragentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md与/ruflo-loop命令plugins/ruflo-loop-workers/commands/ruflo-loop.md中有更细的优先级划分Worker优先级Triggerloop 间隔auditcriticalaudit270soptimizehighoptimize270sconsolidatelowconsolidate600spredictnormalpredict270smapnormalmap270scoordinator/ 600sSKILLtestgapsnormaltestgaps270sdocumentnormaldocument600sbenchmarknormalbenchmark600sdeepdivenormaldeepdive270srefactornormalrefactor270sultralearnnormalultralearn270spreloadlowpreload600s注map在 SKILL.md 中标注 600s、coordinator 中标注 270s两处为不同粒度的建议值SKILL.md 是会话内使用手册coordinator 则面向统一调度实际间隔可按需调整只要遵守 270s 缓存友好的默认心跳即可。5. MCP 响应提示LOOP_SUGGESTION 与 CRON_SUGGESTIONSKILL.md 末尾明确MCP tool responses include[LOOP_SUGGESTION]and[CRON_SUGGESTION]hints -- follow them.即hooks_worker-dispatch/hooks_worker-status的返回结果中会携带调度建议标记[LOOP_SUGGESTION]提示适合用/loop会话内自驱的排程方式[CRON_SUGGESTION]提示适合改用 CronCreate 持久化调度。调用方应当遵循这些提示而非机械地固定使用某一种模式。5 个 hooks_worker-* MCP 工具ruflo-loop-workers包装的是 ruflo-core 提供的 5 个hooks_worker-*MCP 工具定义于 v3/claude-flow/cli/src/mcp-tools/hooks-tools.tsADR 给出了约略行号3489、3538、3656、3714、3923ToolPurposehooks_worker-list列出可用 worker 及其 triggerhooks_worker-dispatch以--trigger worker-name可选--scope分发一次 worker 运行hooks_worker-status检查正在运行的 worker 状态hooks_worker-detect基于上下文检测当前应触发哪些 workerhooks_worker-cancel取消正在运行的 workerworker-detect特别适合上下文驱动场景不手动指定 trigger而是让 MCP 层根据当前会话上下文判断该跑哪个 worker随后再 dispatch。这一组工具同时被 CLI 侧npx claude-flow/clilatest hooks worker ...命令族封装源码见 v3/claude-flow/cli/src/commands/hooks.ts 与 v3/claude-flow/cli/src/services/worker-daemon.ts。两种调度模式/loop 与 CronCreateruflo-loop-workers同时支持两种执行模式loop-workerSkill 对应/loop模式cron-scheduleSkillplugins/ruflo-loop-workers/skills/cron-schedule/SKILL.md对应持久化模式。模式选择原则cron-scheduleSkill 给出了简明判据/loop会话内、缓存感知、自我节拍self-pacing适合活跃开发期CronCreate持久化、可跨会话存活适合CI / 监控场景。CronCreate 持久化用法CronCreate({ schedule: */15 * * * *, prompt: Run security audit worker via mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch })cron-schedule Skill 推荐的默认表达式WorkerCronDescriptionaudit*/15 * * * *Security scanningoptimize*/30 * * * *Performance optimizationconsolidate0 * * * *Memory consolidationmap*/30 * * * *Codebase mappingtestgaps*/15 * * * *Test coverage analysisdocument0 */2 * * *API documentation/ruflo-schedule命令plugins/ruflo-loop-workers/commands/ruflo-schedule.md封装了这一过程用法为/schedule worker [cron-expression]默认 cronaudit、testgaps →*/15 * * * *optimize、map →*/30 * * * *consolidate、document →0 * * * *示例/schedule audit */15 * * * *等价于创建CronCreate(audit, */15 * * * *, Run security audit worker)。cron 模式下 coordinator 推荐的完整间隔可参见上文 12-worker 表如 audit*/15、predict*/15、ultralearn*/15、deepdive/refactor*/30、preload0 * * * *等。loop-worker-coordinator统一调度的编排者插件附带的loop-worker-coordinatoragentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md推荐model: haiku以控制成本负责跨两种模式的 worker 生命周期管理其职责为通过mcp__plugin_ruflo-core_ruflo__hooks_worker-dispatch以正确 trigger 分发 worker通过mcp__plugin_ruflo-core_ruflo__hooks_worker-status监控健康状态并上报失败用ScheduleWakeuploop 模式或CronCreate持久化模式排定下一次迭代遵守缓存 TTL——默认 270s 延迟保持 prompt cache 温暖5 分钟 TTL × 0.9。其工作流为1. 检查当前 worker 状态npx claude-flow/clilatest hooks worker status 2. 分发需要的 worker npx claude-flow/clilatest hooks worker dispatch --trigger WORKER_NAME 3. 按执行模式排定下一次检查此外coordinator 还接入神经学习回路任务完成后通过hooks post-task --task-id TASK_ID --success true --quality 0.9记录成功模式并通过memory search --query TASK_TYPE patterns --namespace patterns检索既有模式实现调度经验的持续积累。worker 触发 → 消费者插件的契约映射ADR-0001 与 README 共同固化了12 worker trigger → consumer plugin的归属表这是调度契约的核心每个 trigger 由哪个插件消费是单一权威来源消费者插件可据此校验自身 trigger 命名TriggerConsumer pluginPurposeultralearnruflo-intelligence从深度代码扫描引导学习语料optimizeruflo-cost-tracker,ruflo-intelligence性能 成本优化建议consolidateruflo-intelligence,ruflo-agentdbEWC 记忆整合predictruflo-intelligence面向后续任务的预测性路由auditruflo-security-audit,ruflo-aidefence安全与合规审计mapruflo-knowledge-graph构建/刷新实体关系知识图谱preloadruflo-core,ruflo-rag-memory高频操作前预热缓存deepdiveruflo-goals(deep-research)多源深度调查documentruflo-docs生成 API 文档 漂移检测refactorruflo-jujutsu差异感知的重构建议benchmarkruflo-cost-tracker,ruflo-iot-cognitum性能基准测试testgapsruflo-testgen覆盖率缺口检测 测试生成命名空间协调worker-history插件声明对 AgentDB 的worker-history命名空间的所有权kebab-case遵循 plugins/ruflo-agentdb/docs/adrs/0001-agentdb-optimization.md 的命名空间约定。该命名空间记录分发事件、耗时与成功/失败判定通过memory_*工具按命名空间路由访问。ADR 同时提醒不得遮蔽保留命名空间pattern、claude-memories、default。安装与契约验证安装/plugin marketplace add ruvnet/ruflo /plugin install ruflo-loop-workersruflo前置依赖ruflo-core插件提供 MCP server。契约验证bash plugins/ruflo-loop-workers/scripts/smoke.sh # Expected: 12 passed, 0 failedsmoke.sh作为smoke-as-contract门禁执行 12 项结构检查见 plugins/ruflo-loop-workers/scripts/smoke.sh覆盖plugin.json版本 0.2.1 且包含mcp、background-workers、cache-aware、schedule-wakeup关键词两个 Skillloop-worker、cron-schedule均含name:/description:/allowed-tools:字段且 agent 与 2 个命令文件存在全部 5 个hooks_worker-*工具在文档中被引用12 个 worker trigger 均在 README 中记录README 固定claude-flow/cliv3.6README 引用 ruflo-agentdb 命名空间约定worker-history命名空间已声明270s 缓存感知 ScheduleWakeup 已有文档存在对 ruflo-autopilot 270s 心跳契约的交叉引用worker-trigger → consumer-plugin 归属表存在含ruflo-docs、ruflo-testgenADR-0001 存在且状态为 Acceptedskills 中无通配工具授权。最佳实践小结默认 270s 心跳所有/loop场景优先使用 270s 延迟min(270, cache_ttl * 0.9)既低于 5 分钟缓存 TTL 又留有抖动余量超过 300s 将付出缓存未命中代价。先 dispatch 再 ScheduleWakeup每次迭代遵循分发 → 排程两步协议形成自驱动循环。事件驱动 安全网对于事件驱动型循环可先 arm 一个Monitor让 270s 唤醒充当兜底安全网README 明确推荐此模式。跟随响应提示MCP 返回中的[LOOP_SUGGESTION]/[CRON_SUGGESTION]是调度模式选择的直接信号应遵循执行。跨会话任务用 cron需要持久化、跨重启存活的 CI/监控类任务走 CronCreate活跃开发中的周期任务走/loop。尊重命名空间边界worker-history属本插件所有勿覆盖pattern、claude-memories、default等保留命名空间。验证契约修改插件文档或结构后运行bash plugins/ruflo-loop-workers/scripts/smoke.sh确保 12 项检查全部通过。延伸阅读插件契约全文plugins/ruflo-loop-workers/docs/adrs/0001-loop-workers-contract.md插件总览与安装说明plugins/ruflo-loop-workers/README.md持久化调度 Skillplugins/ruflo-loop-workers/skills/cron-schedule/SKILL.md协调 Agentplugins/ruflo-loop-workers/agents/loop-worker-coordinator.md命令封装plugins/ruflo-loop-workers/commands/ruflo-loop.md、plugins/ruflo-loop-workers/commands/ruflo-schedule.mdMCP 工具实现v3/claude-flow/cli/src/mcp-tools/hooks-tools.ts270s 心跳契约归属方ruflo-autopilotADR-0001plugins/ruflo-autopilot/docs/adrs/0001-autopilot-contract.md【免费下载链接】ruflo The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated项目地址: https://gitcode.com/GitHub_Trending/cl/ruflo创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考