fix(scheduler): preserve current Host readback in installed launchers

[loopx-agent]

Goal And Delivered Outcome

After #6259, installed Unix and PowerShell launchers still dispatched bound scheduler follow-ups directly to Node. A Codex ACK could bypass the entrypoint's actual Host readback, and an outer-controlled follow-up could bypass its existing write guard.

This change routes matched Codex ACKs and outer-controlled calls through the established entrypoint before the same typed scheduler commit. An actual 60-minute Host rejects a declared 30-minute ACK; after actual Host update/readback to 30 minutes, the original bound ACK writes once and replays without another write. Other generated follow-ups retain their native fast path.

• Basis: reproduced launcher defect; intended base main, related to fix(scheduler): require current App host readback before ACK #6259.
• Owning boundary: built-in scheduler follow-up transport, existing Python Host IO and permission guard, TypeScript effect/state owner. No new provider or parallel decision owner.

Author Declaration

• Written by: model_agent, GPT-6, OpenAI.

Implemented against

• Specification: automatic-execution-admission-v0.md §5, baseline fb5273002fae9d652a0c21c579bb65ee07755f18; the accompanying update makes the installed-launcher boundary explicit.

Criterion	Disposition	Symbol / path	Evidence
Actual Codex Host read before ACK commit	implemented	Unix/PowerShell routing; existing entrypoint	Real launchers and actual App API, File/SQLite
Missing/mismatched observations reject before persistence	implemented	requestFromArgs; existing commit	Matching, wrong, absent and ambiguous Host cases
Failure preserves its bound target/observation pair	implemented	Existing native failure route	Actual Unix/PowerShell with invalid Python executable
ACK/replay remains bound to original Turn	implemented	Existing receipt fence	One write, original replay, newer guard rejects stale original
Elapsed timer, installed adoption and pre-model admission qualify separately	deferred	Existing RFC boundary	Not claimed by this PR

• Self-check: whole diff and native repository validators; identical final test oracles against immutable baseline. No new receipt schema, state engine, configuration switch, or Cron observer.

Scope And Continuation

• Completed slice: current Host readback and permission enforcement through shipped launcher call paths. The decoder also binds the existing host-match flag to its captured fact and honors an explicit neutral current-RRULE argument for other App adapters; explicit empty observations remain unavailable, while absence preserves their bound observation.
• Future-facing pass: reuse the existing reader/guard rather than introducing another TOML parser or Python decision branch. Native failure, Trae and unbound compatibility paths retain their owners.
• Remaining: maintainer review/merge, then qualification of the adopted installation; actual Windows OS, elapsed timer/backoff and pre-model admission remain separate acceptance. Existing SQLite-default/migration program gates are unchanged. This is an independently reversible scheduler repair.

Validation

• Tested revision: 1b636bc2063a266d987dbbbb16ff697a0dab619e.
• Run state: finished. Input classes: synthetic, public_fixture.

Check kind	Result	Public-safe evidence / limitation
static	passed	Control-plane TypeScript typecheck, repository mypy (19 files), repository ruff scope, Bash syntax, semantic advisory then full inventory smoke, diff/public-boundary checks
real_entrypoint	passed	132 focused Python tests: launchers, interpreter selection, console/module entrypoints, controller scope and actual Host binding; zero failures/skips
unit	passed	49 TypeScript scheduler commit/follow-up/CLI/receipt-fence tests; flag stripping, wrong and explicitly empty neutral observations reject
regression_parity	passed	Identical final launcher test oracle on immutable baseline: 10 failures/17 passes; final source passes all 27. Two old module/controller fixtures independently failed on baseline because they omitted actual post-update readback; fixtures now perform it
real_backend	passed	Disposable canonical File/SQLite stores, actual Codex App API update/readback; wrong 60-minute Host rejected, actual 30-minute Host acknowledged once, original replay idempotent, stale receipt rejected, quota unchanged
real_entrypoint	passed	Proper wheel build after real bundled-frontend build; independent wheel install and archive-layout Unix launcher outside source checkout, source/provenance checked; actual PowerShell 7.6.6 process on macOS
regression_parity	passed	Eight alternating archive/console pairs per provider, identical stale denied ACK workload and complete equal outputs. Median archive/console: File 172/167 ms, SQLite 162/155 ms. Bounded startup/read observation, not a formal latency or capacity qualification
integration	passed	Risk-selected native canary: 15 checks plus 4 direct checks; final exact-scope quality receipt recorded
manual	not_run	Actual Windows OS, global installation adoption, elapsed timer and pre-model admission; not claimed

• Coverage: the real launcher/backend matrix exercises both routing and rejection before effects, with original receipt replay and native default-off paths. Disposable probe automation removed. No PostgreSQL authority-store/refactor or provider promotion is changed; no live Goal is used as a mutation fixture.

Frontend / Visual Evidence

• UI impact: none. No screen, settings editor or configuration capability changed; the existing App API was used for isolated validation. Screenshots and attention/craft review are not applicable.

Type of Change

• Bug fix
• Documentation update
• Test update

LoopX Area

• Control plane
• Host or runtime integration

Technical Direction

• Existing overall-roadmap S7/S4/R2 and automatic-execution-admission §5: shared typed Host follow-up semantics across real entrypoints. A passing schedule ACK is not elapsed-wakeup or SQLite-default qualification.

Shared-authority RFC fixture impact

• No authority-provider routing/promotion or compatibility-projection change. Synthetic canonical File/SQLite conformance arms cover this scheduler boundary; production-scale/legacy/PostgreSQL promotion rehearsal is not claimed.

Boundary Checklist

• Public-safe diff/body; no private state, raw logs, credentials or local machine paths.
• No duplicate benchmark work or new benchmark launch.
• Cohesive reproduced repair scope.
• UI impact none.
• All three commits include DCO sign-off.

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=xhigh

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

精确 head：1b636bc2063a266d987dbbbb16ff697a0dab619e。没有发现阻断这份修复 PR 的问题；这是可独立验收的入口修复，正式安装采用和时间调度仍有边界。

动机

Codex App 用户通过已安装的 LoopX 命令确认定时配置时，会遇到这个问题。 App 定时器已改为30分钟时，旧启动器仍因没有重新读取实际配置而拒绝确认；修复后重新回读并接受一次，重复确认不会再次写入。 真实60分钟配置拒绝声明30分钟的确认；真实30分钟配置确认成功，原回执重放保持单次效果。 这份 PR 不宣称实际时间唤醒、模型启动前准入或新 Goal 默认 SQLite 已完成。 当前正式安装尚未采用此修复，实际 Windows 系统执行、时间唤醒和后续退避仍需原验收。

这里的 ACK 是对实际定时配置的确认，原回执绑定本次执行。旧入口绕过读取或权限守卫，会让后续调度无法取得可信确认；只验证 console 无法证明安装启动器具有相同语义。

改动思路

实际 Host 读取留在已有 Python IO，状态与效果判断留在既有 TypeScript owner；启动器只选择已有入口，避免增加第二份 Host 解析与决策。 本 PR 完整交付启动器到实际回读、拒绝和单次持久化的调用链；合并后的正式安装采用和时间唤醒继续归原验收。

Unix/PowerShell 在已有生成命令中识别绑定的 Codex 确认与 outer-controller 环境，交回已有 entrypoint。它重新读取实际 App 配置，再把观测交给同一 TS commit；其他原生 failure、Trae 和无绑定兼容路径保留。提示生产者生成的 flag 与捕获事实必须一致，读取、决策和效果没有新增平行 owner。相比复制 Node TOML 解析或把所有命令送回 Python，这个边界更小且保留未受影响路径的成本。

具体改动

依据 docs/architecture/rfcs/automatic-execution-admission-v0.md，验收文本固定在 fb5273002fae9d652a0c21c579bb65ee07755f18；先读改动前 §5 Host contracts，而非用本次文档修改证明自身。

• §5 Host contracts: actual schedule：implemented。实际 App60分钟拒绝声明30分钟，实际更新/readback30分钟后接受一次。
• §5 Host contracts: reject before effects：implemented。错、缺、歧义 Host 与明确空观测在迁移/写入前拒绝；丢弃 flag 也不能绕过读取路由。
• §5 Host contracts: failure pair：implemented。失败命令保留原 target/observation，并用无效 Python 路径证明原生路由仍可执行。
• §5 Host contracts: separate qualification：deferred。源代码和独立候选安装已验；正式安装采用、实际 Windows OS、时间唤醒/退避和模型前准入保持原验收。

关键代码讲解

1. scripts/loopx:59 的 native dispatch gate 把 matched Codex ACK 和 outer-controlled 调用留给现有 Python entrypoint，其他绑定 followup 继续 exec node。条件使用现有明确 flag；并不在 shell 决定状态迁移。
2. scripts/loopx.ps1:57 的 needsCodexHostRead 与 isNativeSchedulerFollowup 执行相同选择，保留已有 $LASTEXITCODE 传播。实际 PowerShell7.6.6 进程在 macOS 上验证，不以源码字符串断言替代运行。
3. heartbeat_followup_cli.ts:191 的 requestFromArgs 校验捕获事实与 --host-match-observed，复用原 identity mismatch 错误。显式 neutral current RRULE 也对其他 App生效；Object.hasOwn 区分参数缺席与明确空值，缺席继续保留其绑定观测，空值不拿旧值冒充新读回。

其余三份测试覆盖真实 File/SQLite、原回执重放、控制范围及已有 source fixture 的 post-update读取。RFC只补当前调用链和已有接受规则的适配边界，没有把未测 timer/hook 写成已完成。七个文件整体审查，没有新模块或持久化协议。

对主干的风险

最强反例是 console 已成功但安装脚本仍跳过 Host读取，或 controlled Goal换 actor 就绕过权限。相同最终 oracle 在不可变主干的 launcher矩阵出现10失败/17通过，在 head全部27通过；TS相同新 oracle 主干10通过/5失败，修复后整个相关 TS套件49通过。source scope测试还覆盖同容器其他 Goal、激活后新增独立 Goal、scope escape以及 owner释放后真实恢复写入，避免修成全注册表误拦截。

最终132项 Python与49项TS验证零失败/跳过；typecheck、mypy、ruff、语义 smoke和风险选择 canary15项加4个direct检查通过。首次 canary 的 missing quality receipt是独立执行门，随后经精确范围质量评审补回执再通过，未改预算消除失败。早期 fixtures没有实际读回的失败独立复现后修正，相关原断言保留。按本 Goal当前政策没有读取或等待远端 CI。

正常构建 bundled前端后完成 wheel、独立 wheel install及 archive-layout启动器，并用实际 App API在隔离 File/SQLite验证错误拒绝→匹配写入→原回执幂等重放→后续 guard稳定和 stale原回执拒绝；quota为0。八对相同 stale拒绝负载的 archive/console全输出相同，median约File172/167ms、SQLite162/155ms，仅是有界启动/读取观察，不能替代正式p95或容量资格。正式安装未切换，实际 Windows OS与timer仍未验证；测试没有修改活跃 Goal或正式automation。

语义与 CI 对齐

复用既有 Host surface、flag、neutral参数和错误码；临时路由布尔值仅留在 launcher。状态/效果继续由typed owner执行。当前新错误是硬拒绝，不是可忽略guidance；没有新权限、peer生命周期或默认 SQLite推广。无 packet指令压缩、冻结门放宽或语义登记绕过。文档及重命名原生Trae测试公开披露调用默认变化。

我的整体评价

结论为 APPROVE，交付判断是 justified_increment：long_horizon improved（后续稳定、原回执单次效果和旧回执失效），user_experience improved（原命令自动完成必要回读，无新手工同步步骤）。本 PR 完整交付启动器到实际回读、拒绝和单次持久化的调用链；合并后的正式安装采用和时间唤醒继续归原验收。

有意改变的是原入口缺陷、丢弃绑定flag以及非Codex显式观测被忽略的行为；native失败、无绑定与非Codex无覆盖观测仍保留。已落实面向未来的复用重构，避免第二份 Host解析/决策；保留真实receipt/独立adapter兼容，不把代码减少等同验收。剩余风险是实际Windows与正式采用/时间资格，runtime合并交maintainer。

English verdict: APPROVE - head 1b636bc2063a266d987dbbbb16ff697a0dab619e; real launchers now reuse actual Host readback and scoped guard before the existing typed commit. 132 Python/49 TS tests, native quality/canary and isolated actual App File/SQLite candidate-install validation pass; Windows OS/global adoption and elapsed/pre-model remain separate.