dr-schedule-manager — dr-schedule-管理器

dr-schedule-管理器

Build scheduled 自动化s so each 运行 reflects current configuration immediately.

Core outcome

This 技能 is a scheduling architecture and 迁移 playbook, not a one-command 调度器 安装er.

After 安装ation or 迁移, scheduled jobs should:

pick up current prompt changes on the next 运行 pick up current policy changes on the next 运行 pick up current delivery changes on the next 运行 pick up current default 模型 changes on the next 运行, unless intentionally pinned avoid stale 会话 residue from prior 运行s

If a de签名 does not guarantee those properties, do not recommend it as the default.

Current OpenClaw constrAInt

Treat current OpenClaw cron as snapshot-based unless proven otherwise.

In practice, cron jobs may embed:

prompt text 模型 override delivery 路由 other 运行time detAIls

That means editing local files alone may not change the behavior of the already-registered job.

Because of this, the preferred practical pattern for current OpenClaw is not "fat job config in cron". It is:

thin trigger in cron local file resolution at 运行time explicit final delivery through the normal outbound path Default architecture

Prefer a thin-trigger, fresh-运行, config-driven job architecture.

Rule 1, 调度器 is only a trigger carrier

The 调度器 should only:

wake the job identify the job slug or manifest pass a small stable trigger message

Do not embed business 记录ic, 格式化ting rules, or 模型 decisions in the 调度器 unless you intentionally accept snapshot behavior.

Rule 2, manifest is the operational contract

Each scheduled job should have a manifest file that defines:

slug name 代理 id schedule 运行time mode trigger mode prompt file path policy file paths delivery contract 模型 policy verification rules live 调度器 id if your local 工具ing 追踪s one Rule 3, 运行time assembly h应用ens at execution time

On every 运行, load current files before generating 输出.

Always assemble from:

current manifest current prompt file current policy files current delivery rules current 模型 policy

Do not trust previous 会话 状态 for these.

Rule 4, delivery is explicit and 提供者-aware

Store delivery in a clear adapter contract.

Do not assume 会话 metadata is valid for outbound 发送s if the 提供者 requires a different tar获取 格式化.

Rule 5, persistent 会话s are not the source of truth

If you keep a persistent 自动化 代理, use it only as a dis补丁er or coordinator.

Do not let a persistent scheduled 会话 be the authoritative source for:

prompt wording 模型 selection 格式化ting rules delivery routing 应用roved patterns Pattern A, wake-only trigger into fresh mAIn execution

Use when you want the latest mAIn 助手 behavior to 应用ly automatically.

Best for:

personal briefings reminders evolving 助手 工作流s

Strengths:

changes propagate immediately minimal drift risk simple to reason about

Weaknesses:

less isolated changes to mAIn behavior affect the job immediately Pattern B, thin trigger plus local manifest resolution

Use as the default reusable pattern across 代理s for current OpenClaw.

Best for:

reusable 自动化 框架s 报告s and digests jobs that need 清理 状态 on every 运行 设置ups where cron otherwise snapshots prompt, 模型, or delivery

How it works:

cron stores only a small stable trigger the triggered 代理 reads local job files at 运行time prompt, policy, 模型 policy, and delivery are resolved from the workspace final delivery uses the normal outbound path, not cron announce, when announce is unreliable

Strengths:

avoids stale embedded prompt drift avoids stale 模型 pins in cron payloads makes file edits effective on the next 运行 easy to 迁移 across 代理s

Weaknesses:

requires the 代理 to actually honor the trigger by reading the local files still depends on reliable final outbound delivery Pattern C, persistent dis补丁er plus fresh worker 运行

Use for more advanced orchestration.

Best for:

retry 队列s fan-out 工作流s multi-step 自动化 流水线s

Strengths:

scalable strong separation between orchestration and generation

Weaknesses:

more moving parts Default recommendation

For most current OpenClaw scheduled jobs, use Pattern B, thin trigger plus local manifest resolution.

Reason:

it works around snapshot-based cron behavior it is reusable across many 代理s it avoids stale embedded prompt and 模型 drift changes are effective on the next 运行 because 运行time 输入s are file-based it 扩展s better than manual re-registration for every content tweak 模型 policy rules

模型 behavior must be explicit.

Preferred

Use inherit-default when 升级s should propagate automatically.

Example:

{ "模型Policy": { "mode": "inherit-default" } }

Use only when intentionally pinned { "模型Policy": { "mode": "pin", "模型": "openAI-codex/gpt-5.4" } }

If pinning is used, document why.

分享d-policy option { "模型Policy": { "mode": "policy-file", "path": "aut