[elfcv]

这次改了什么

摘要

Agent 新建的检查任务，没变化就不要再启动模型。以前 silentWhenIdle=true 只是不播报，模型仍会每次被叫醒去判断有没有变化。现在这类新建检查默认带程序变化门：没有脚本 exit 0 或插件放行，就不建会话、不启动模型，run 记 skipped。上网检索这类必须由模型发现变化的检查，创建时显式设 modelDetectsChange=true，仍走原来的「启动模型、无变化不播报」。

变更类型

• fix 缺陷修复

范围

• 关联 Issue / 需求：无。对应「检查空转仍叫醒模型」。
• 本 PR 包含：schedule_create 对检查默认写入 programChangeGate；runner 在没有程序信号时跳过模型；schedule_update 可用 modelDetectsChange 开关这道门；SQLite 列 program_change_gate（默认 false）；产品规则一句。
• 明确不包含：不迁移已有定时任务；桌面表单新建的检查仍按原行为启动模型；提醒和定时发送不变；不改伙伴例行任务的事件门；不把游戏雷达这类检索检查改成静默跳过。
• 用户可见变化：之后经 Agent 新建的检查，没挂前置检查时不会再开模型。已有任务不受影响。
• 是否存在 breaking change：无。旧行默认 false；省略 silentWhenIdle 的创建仍会启动模型。
• 与同期 scheduler PR 的关系：fix(scheduler): 检查类定时任务默认复用同一会话 #8270、fix(scheduler): skip the model when a check has no change signal #8285（本 PR）、fix(scheduler): 派发前撤回的提示直接删除，并定期清掉历史软隐藏 #7828 都改调度器，但语义上互相独立，没有合并依赖。本 PR 与 fix(scheduler): 检查类定时任务默认复用同一会话 #8270 在 create.ts、cindy_schedulerMcpServer.test.ts、bot-routines.md 有重叠改动，建议 fix(scheduler): 检查类定时任务默认复用同一会话 #8270 先合、本 PR 随后 rebase；与 fix(scheduler): 派发前撤回的提示直接删除，并定期清掉历史软隐藏 #7828 只在 scheduler-host/runner.ts 相邻，无语义依赖，任意顺序。

UI 变化

不涉及。

• 引用的设计规范：不涉及

怎么验证的

自动验证

cd apps/desktop && pnpm exec vitest run src/main/scheduler-host/__tests__/runnerPreRunHookCwd.test.ts src/main/scheduler-host/__tests__/storage.test.ts --maxWorkers=2
结果：48 passed（runner 22 + storage 26）。runner 文件约 13s，其中测试体 16ms；耗时在 Electron 收集，不是用例本身。

cd packages/lizi-mcps && pnpm exec vitest run src/__tests__/cindy_schedulerMcpServer.test.ts --maxWorkers=2
结果：75 passed，218ms。

node apps/desktop/scripts/validate-migrations.mjs
结果：0000..0141 通过，无 schema drift。

pnpm --filter @cindy/maker-scheduler exec tsc --noEmit
结果：通过。

测试价值：最小失败输入是 { silentWhenIdle: true } 的 schedule_create，读回 programChangeGate=true；以及 runner 对带门、无 hook、无插件的检查直接 skipped，createSession 未被调用。对照锁住 modelDetectsChange=true、提醒、有 hook、有插件、把检查改回提醒这几条不误伤。

手工验证

不涉及。没有启动 Desktop，也没有改真实定时任务。

未执行的验证

• Desktop 全量 tsc --noEmit 在本机 8GB 堆上 OOM。失败点是堆上限，不是类型报错。定向 vitest 已编译并跑过改动文件。
• 没有在真机上观察一条新建检查的 skipped 记录。

风险

风险分类

• SQLite / migration
• 其他：改了模型可见的 schedule_create / schedule_update 说明。

影响与回滚

• 影响范围：只影响本机、经 schedule_create 新建、且 silentWhenIdle=true、未设 modelDetectsChange=true 的 agent 检查。已有任务列默认 false。提醒、脚本任务、桌面表单不受影响。跨电脑 schedule_update 不接受 modelDetectsChange。
• 回滚 / 降级方式：revert 本提交。新列默认 false，旧客户端忽略未知列；已建的带门任务可在本机用 modelDetectsChange=true 关掉。
• 存量插件影响：无。没选插件的任务由脚本或这道门决定；选了插件的检查仍先问插件，插件放行才会进入模型路径。

提交前检查

• 已 review 完整 diff
• 每个 commit 都带 DCO 签名
• 未提交凭证、令牌或授权文件
• 已核对受影响的文档并同步
• 已确认测试结果或说明未执行原因

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-04T17:02:08.488171Z	0dced01	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The change affects how selected scheduled checks start across agent-created routines.

Fix Continue for interrupted gated checks before merging.

Findings

1.  Continue skips interrupted checks ▶

Fix with agent prompt

### Issue 1
apps/desktop/src/main/scheduler-host/runner.ts:883-887
A check created with `silentWhenIdle=true` and a passing `preRunHook` cannot continue after an app restart. When the user clicks Continue, `fire()` clears `preRunHook` but keeps `programChangeGate=true`. This condition then skips the run instead of continuing its existing task.

The engine also marks the original run as `skipped`, clears its `sessionId`, and sets its cost to zero even though it already started. Exempt `ctx.resumeSessionId` from this gate and add a regression test for continuing a gated check.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

Adds programChangeGate so new agent checks created through schedule_create skip the model when no script or plugin check is configured.

• New agent checks skip the model unless a program signal or model-led check is chosen.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[User clicks Continue] --> B[Load interrupted run]
  B --> C[Keep programChangeGate; remove preRunHook]
  C --> D{No plugin configured?}
  D -->|Yes| E[New gate returns skipped]
  E --> F[Original run loses task link and is marked zero cost]

Loading

_{Reviews (1) · Last reviewed commit: "fix(scheduler): skip the model when a ch..." · Reviewed by Greptile}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f1b2b40c1a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[MagicLizi]

维护者确认请求（自动审查流程）

本 PR 命中维护者确认门，需要维护者确认后才能继续自动审查与合并：

• rules：改动了规则文档 docs/product-rules/bot-routines.md，且为实质规则含义变更（新建检查默认带程序变化门、没变化不再启动模型等新规定），审查判据自身发生变化；
• arch：调度核心路径改动 160 行（≥150 阈值），总 diff 7696 行（≥800 阈值），并改变新建检查的默认行为。

改动摘要：Agent 新建的检查（silentWhenIdle=true）默认带程序变化门，没有脚本 exit 0 或插件放行就不启动模型、run 记 skipped、零 token；变化必须由模型发现时显式设 modelDetectsChange=true；提醒、定时发送、桌面表单和已有任务不迁移。附带数据库 migration（schedules 新增 program_change_gate 列）。

请维护者直接在本 PR 上 Approve 确认（Approve 即视为确认通过，流程会继续自动审查与合并）；如需修改，请 Request Changes 并说明，作者改完后重新审查。

详细说明见关联讨论 issue。

讨论 issue:#8286

[elfcv]

Follow-up on b4e5ca5, after merging current main (conflict was only schedule_create; remote create of a silent check still gets programChangeGate).

Review fixes:

• Continue after exit no longer hits the no-signal gate just because resume clears the hook.
• A gated check bound to a deleted or service-owned archived target pauses before the empty skip, so it does not stay active.

Local checks, not GitHub CI:

• pnpm --dir apps/desktop exec vitest run src/main/scheduler-host/__tests__/runnerPreRunHookCwd.test.ts --maxWorkers=2: 28 passed, 13.4s
• pnpm --dir packages/lizi-mcps exec vitest run src/__tests__/cindy_schedulerMcpServer.test.ts src/__tests__/schedulerDevices.test.ts --maxWorkers=2: 91 passed, 1.1s
• apps/desktop tsc --noEmit exit 0
• eslint on the touched desktop runner files exit 0

CI on this head has not finished. Product gate #8286 is unchanged and still needs a maintainer Approve.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b4e5ca5272

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[elfcv]

Rebased onto current main. The program-change-gate migration is now 0144_brief_elektra (main already owns 0141-0143). Head 0fa2d40543.

Local checks before push, not GitHub CI:

• pnpm --dir apps/desktop db:validate: passed, 0000..0144, drizzle-kit check ok, 1.3s
• runnerPreRunHookCwd 29 passed, 13.3s
• scheduler MCP tests 91 passed, 1.1s

CI is rerunning on this head. Still needs a maintainer Approve on the PR; the author cannot self-approve.

[elfcv]

Merged current main again (a65d671fe5). Main now owns 0144 and 0145, so the program-change-gate migration was regenerated with drizzle-kit as 0146_handy_quentin_quire. The SQL is unchanged (ALTER TABLE schedules ADD program_change_gate ...).

Local: db:validate passes (0000..0146). test:migration-replay was not run: it only executes on macOS/Windows, and it is a manual tier that CI does not run either.

The review finding about promotion to a project automation is fixed in 0dced01893 (see the thread).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b2e3cc2ace

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0dced01893

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

在复制或降级项目检查时保留程序变化门

把 silentWhenIdle=true、programChangeGate=true 且无 hook/插件的项目检查复制或降级为个人自动任务时，SchedulerPage.tsx:489–492,515–516 都会调用 scheduleToUserCreateInput()，但该转换（scheduleFormLogic.ts:675–698）没有携带这两个字段。定向运行生产转换确认，重建输入确实缺少两项；此路径直接走 IPC schedule.create，不会经过 MCP 新增的默认门控逻辑，因此落库后门变为 false，原本无信号时零 token 的检查会重新每轮启动模型，降级路径还会删除原调度。请同步这个反向转换，保留已有门控和静默状态。

Useful? React with 👍 / 👎.