[xiaobaifly7]

Goal And Delivered Outcome

• Outcome basis: repair the DSH continuation Driver's consumption of the existing scheduler hint contract.
• Goal/source and gap: readQuota() requests --include-detail scheduler, but schedulerWait() looked for cadence under unchanged_poll.local_scheduler. Supported CLIs put it under cold_path_detail.local_scheduler, so wait decisions produced no timer and could remain idle until an external event.
• Observable before → after: a CLI-shaped wait receipt now schedules 30, 60 and 120 minute re-evaluations, then performs the final guard read without scheduling another unchanged poll. A new reset token restores the first interval. A supplied recheck_after_seconds caps the wait. A fresh should_run=true decision still queues the canonical heartbeat.
• Issue/task and intended base: self-contained bug fix targeting main; no issue closure is claimed.

Author Declaration

• Written by: model_agent — OpenAI Codex (GPT-6).

Implemented against

• Specification and revision: loopx/control_plane/scheduler/hint.ts (loopDetails() and the scheduler_hint_v0 / scheduler_hint_detail_v0 projection), docs/status-data-contract.md scheduler hint contract, and packages/dsh-loopx-plugin/README.md Driver retry boundary, at 82867e40597771ab36eaf3ca367657c646f9903e.
• Criteria:

Criterion (spec clause)	Disposition	Symbol / path	Test or command
Local scheduler cadence lives in cold_path_detail.local_scheduler	implemented	schedulerWait() in src/driver.ts	Driver test for detailed wait scheduling
Use the compact unchanged_poll.limits.local_scheduler; null means no unchanged-poll cap	implemented	schedulerWait()	Compact-limit cases for 0, 1 and null, including precedence over the detail copy
Apply the published progression and reset on reset_policy.reset_token change	implemented	schedulerWait() and existing Driver state	30 → 60 → 120 minutes, final unchanged read, token change → 30 minutes
Honor cold_path_detail.local_scheduler.recheck_after_seconds	implemented	schedulerWait()	45-second cap and a later deadline that does not extend the cadence
Preserve runnable admission and finite transport retries	implemented	Existing buildPlan() / evaluate()	Wait-to-runnable transition plus the complete Driver suite

• Self-check before submission: read the scheduler producer, its protocol documentation, the Driver callers/timer flow, and the minimum-version release source; ran the checks below and reviewed the three-file diff. LoopX v1.2.4 already emits the detail/limits shape, and MINIMUM_LOOPX_VERSION is 1.2.4, so no compatibility branch for the older unchanged_poll.local_scheduler object was added. The schema and supported version floor remain unchanged.

Scope And Continuation

• Completed scope and remaining work: the scheduler field-path fix, regression coverage, and retry-boundary documentation are complete within this scope.
• Slice boundary / successor: this changes the existing DSH host adapter only. No new scheduler authority, frontend control, setting, or successor task is needed. A broader refactor was unnecessary: the behavior already has one schedulerWait() owner.

Validation

• Tested revision: ee5226a9bc (the single commit in this PR).
• Run state: finished.
• Input classes: synthetic.
• Environment: native Windows, Node.js 24.18.0, pnpm 12.14.0, frozen dependency lockfile.
• Commands below run from packages/dsh-loopx-plugin unless stated otherwise.

Check kind	Result	Public-safe evidence / limitation
regression_parity	passed	Before the production fix, pnpm test -- --reporter=verbose with the updated CLI-shaped fixtures reported 276 passed / 23 failed. Nine failures were Driver scheduling cases, including the new wait-receipt regression.
static	passed	pnpm typecheck: host, client and test TypeScript configurations passed.
unit	passed	pnpm test tests/driver.spec.ts: 97 passed. Covers recommended delay, progression, final unchanged read, compact limits, reset, recheck bounds and runnable admission.
integration	failed	Final pnpm test: 285 passed / 14 failed, 10 files passed / 5 files failed. All 14 failure names also occurred before the production fix; no new failure names appeared.
static	passed	git diff --check from the repository root. No RFC pair or shared semantic vocabulary changed.

The unchanged full-suite failures are:

Test file	Failures	Observed limitation
tests/admission-closeout.integration.spec.ts	7	Existing progressive/context-reader count assertions fail in this environment.
tests/cli.spec.ts	2	Existing child-close timing assertions fail on native Windows.
tests/init-command.spec.ts	3	Existing path and launcher assertions fail on native Windows.
tests/native-operation.test.ts	1	Existing POSIX socket fixture times out on native Windows.
tests/observer.spec.ts	1	Existing POSIX ledger-path expectation differs from Windows path normalization.

• Coverage and gaps: this validates the source Driver through its public evaluation/timer flow using synthetic CLI-shaped receipts derived from the producer. It does not qualify an installed Harness session or a packaged live-backend journey. The existing full-suite failures are disclosed and were not expanded into this scheduler-path fix.

Frontend / Visual Evidence

• UI impact: none. The existing continuation timer is repaired; no rendered controls or layouts change.
• Before / after / first viewport and preview: N/A.
• States and viewports shown: N/A.
• Source data: none.
• Attention review / craft review: N/A; no visual change.

Type of Change

• Bug fix
• Documentation update
• Test update

LoopX Area

• Host or runtime integration

Technical Direction

• Direction / acceptance reference: existing local scheduler hint consumption in the DSH continuation Driver.

Shared-authority RFC fixture impact

N/A. No authority store, migration, provider routing, compatibility projection, or shared-authority RFC claim changes.

Boundary Checklist

• The diff and this draft contain no private state, credentials, raw traces, internal links, or local machine paths.
• No maintainer-owned benchmark work is duplicated.
• The change is scoped to the scheduler hint consumer defect.
• No paired RFC changed.
• UI impact is none.
• Every commit includes a DCO Signed-off-by trailer.

🤖 Generated with Claude Code

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · max

本次评审对应提交 ee5226a9bcb14e047bd6372ed4dbabd5d661bf0d。

动机 / Motivation

已激活 DSH LoopX Session，处于依赖或定时等待的操作者。DSH 已经向 Core 请求 scheduler detail，等待分支却读取 compact unchanged_poll 下不存在的 cadence 对象，导致已激活 Session 没有下一次定时重查。同一真实 CLI 返回的 resume_at 等待以 45 秒进入定时器。基线没有定时器，当前提交发出一次匹配的 timer。Core 仍决定何时可工作。同一真实 CLI 返回的 resume_at 等待以 45 秒进入定时器。基线没有定时器，当前提交发出一次匹配的 timer。Core 仍决定何时可工作。实际使用本机原生 CLI 和源码 Driver，Session 与时钟由测试提供。正式 DSH 安装、真实模型调用和托管 CI 未验证。

改动思路 / Approach

保留现有 wait plan 和同一 Session 的 timer。Core 发布等待、上限和 reset token。Driver 在已有等待分支采用这些字段。 继续从同一份 quota 决策读取 scheduler detail。上限使用 compact projection，回看时间取已有间隔和 Core 截止时间中的较短值。

具体改动 / Changes

关键代码讲解

以 docs/status-data-contract.md 的 scheduler hint 为准。LOCAL-DETAIL、BOUNDED-DEADLINE 和 LIMIT-RESET-STOP 的行为均已核对。 等待分支 读取 detail 和 limit。 核对了最低支持版本 v1.2.4 的生产者，已发出这两处字段。真实定时等待使用现有 Core 路径产生的原 JSON。原生等待建议为一分钟，deadline 为 45 秒。Driver 把后者交给现有 timer，随后仍需重新查 quota 才可提交 heartbeat。没有 local detail 的 Stop 不会走计数和调度。计数由同一 reset token 串联，读到新 token 后从首档开始。

对主干的风险 / Risks

用同一份真实 CLI 结果驱动基线和当前源码。基线没有定时器，新提交按 45 秒设置。实际 Stop 在两边都没有 timer，也没有发起 heartbeat。有限次重查、reset 和重新准入由 Driver 套件覆盖。

语义与 CI 对齐

97 项 Driver 测试通过，两个原生输入的配对探针通过。TS host/test 类型检查、diff、DCO 与双语检查通过。本机 macOS 全包测试在精确基线通过 293 项，当前 head 通过 299 项。作者报告的 Windows 全包失败未在本机复跑。实际使用本机原生 CLI 和源码 Driver，Session 与时钟由测试提供。正式 DSH 安装、真实模型调用和托管 CI 未验证。

我的整体评价 / Verdict

同意当前 head 的等待路径修复，现有 Session 和 Core 准入边界核验通过。本次评审覆盖现有宿主与 CLI 行为。实际使用本机原生 CLI 和源码 Driver，Session 与时钟由测试提供。正式 DSH 安装、真实模型调用和托管 CI 未验证。

English verdict: APPROVE. The Driver consumes the published scheduler detail and compact limit. A real deferred decision schedules its 45-second recheck, while Stop remains idle.