Skip to content

task: Harness evaluation-mode CPU execute for Apps #429

Open

Description

@zengchuang-hw

Current state: Accepted bounded Harness delivery; implementation remains incomplete. Dong1017 retains accountability. All three units are delivered as formal Ready PRs by zengchuang-hw / Codex under PR-1 delegation, PR-2 delegation, and PR-3 delegation. No new architecture/scope decision is pending.

Delivery unit Artifact and current status Remaining condition
PR-1: reject malformed/inconsistent bindings before registration; public read projection; D023 null-timeout vector Ready #561, head 66fef981f6893038f3128595ffcdfd12b65bfba2; current-head Ready CI passed; Amy approved Second eligible independent approval and integration
PR-2: registered T37-N1/N2 actual-entry verification, zero-start fixed records and commit failure Ready #562, head 73c30b8fe4fd1b7e953ed8b11894741692055724; current-head Ready CI passed; Amy approved #561 first, resync, owner re-read and two current independent approvals
PR-3: original attempt deadline/budget with null step timeout; T40 replay Ready #564, head d34a0451f0d4edd76c604632b44d673f4f5d43ae, based on #562 73c30b8f; 111 core tests and 297 expanded tests / 43 subtests passed; 2 POSIX skips; exact-head fresh wheel passed; Ready CI passed; Amy approved Second eligible independent approval; #561 → #562 first and resync; Inference1's separate #552 finite-step-timeout fix remains a prerequisite

Decision needed: None for these three accepted delivery units. The existing free-CPU budget admission limits are unchanged; PR-3 does not introduce a new paid-budget provider or policy. Owner technical review of #564 is PASS and explicitly excludes its parent vote from independent approvals.

Next action: Resolve findings and integrate #561 → #562 → #564 only after each has two eligible independent write+ approvals and passing checks, respecting #552. Current #561/#562 tests intercept the actual local_process._supervise worker launch path; PR-2 positive controls observe real PIDs. Old-head approvals are not reused after the correction. Small source-worker tests establish execution/reentry facts, not complete F/C/M proof or whole-Issue acceptance. Run the shared D023 / t108-v4 production CPU acceptance with #430/#434; this Issue and #371 stay open.

Accepted production-evidence follow-up: #565 now has separate Ready B #566, H #567, Apps #568 and CPU worker/profile #569 slices. They retain the three-unit Harness boundary above; shared main acceptance waits for their dependencies, the separate #552 fix, final-head checks and independent approvals.

Goal and acceptance

Enable the production Apps flow in #371 step 2 to call Harness in evaluation mode. Follow #387 D023-2/V32 and the post-H02 implementation boundary in the contracts.

  • Reject malformed or inconsistent prepared bindings before registration, with no RunStore record. Test this separately from actual entry verification after registration.
  • After registration, T37-N1 commits not_started with a recorded mismatch, and T37-N2 commits not_started with the check not_performed and the actual digest unknown. Assert zero starts for both, and verify the existing persistence error if the record cannot be committed.
  • With a null step timeout, enforce the original attempt deadline and budget. Cover expiry mid-run, cleanup, and explicit cancellation in the real CPU path.
  • For T40, submit the same saved request with the same attempt identity and canonical request digest. Verify the original fixed RecordRef for a completed attempt, an unchanged attempt count and zero additional process starts. A different digest for the same attempt remains a conflict.
  • Harness consumes the fixed projection and read-only artifact port. Benchmark keeps task and rule ownership.

Accepted follow-up to #411. Delivery and independent review remain tracked here; all five acceptance items still apply.

Activity

  1. added
    kind:taskBounded delivery task with a delegated slice
    area:harnessHarness module: profile, A/B loop, quality gates, evidence reports
    status:proposedWaiting for the decision named in the status block
    on Nov 13, 2026
  2. amy-why-3459 commented on Nov 13, 2026

    @amy-why-3459
    Collaborator

    Thanks for splitting this out. #427 (approved on 6778eb19) is the producer of the spec this slice consumes, so I've written a few points from the Apps side for whoever picks this up:

    1. Deadline without a step timeout. The Bind saved Apps requests to frozen worker specs #427 binder emits the attempt root with a null step timeout, because OperationRecipe v1 has no per-step timeout. Harness evaluation mode then has to enforce the preserved attempt deadline and budget at the attempt level. Could the acceptance state this explicitly and include a test where the deadline expires mid-candidate, covering the recorded outcome and cancellation?
    2. Which failures are recorded, and which are rejected? Bullet 1 validates the prepared binding and expected worker digest before registration. T37-N1/N2 record committed not_started / not_performed facts, which implies registration already happened. Please spell out the boundary: (a) a malformed spec or binding mismatch is rejected with no RunStore record; (b) once registered, an entry-identity mismatch or unavailability becomes a committed fact. Two separate tests would keep that boundary from drifting.
    3. Evidence for T40 replay. "No new attempt or process on replay" needs an observable check, e.g. the RunStore attempt count stays the same and a process-spawn counter or fake launcher records zero starts. It should also say what the replay key is (saved request digest vs. attempt id).
    4. Dependency order. Harness can develop against the D023 golden ExecutionSpec now. The production T37/T40 runs still need task: Benchmark production prepare and frozen resource closure #430's real prepare output. It would help to note that under "Next action" so the two implementers don't block on each other.

    Once an implementer is confirmed, linking their PR here will make independent review easier to track.

  3. amy-why-3459 commented on Nov 13, 2026

    @amy-why-3459
    Collaborator

    Thanks, the updated acceptance covers all four points: the pre-registration vs. post-registration boundary, T37-N1/N2 with zero starts, attempt-level deadline/budget enforcement under a null step timeout, and an observable T40 replay key and evidence. Developing against the D023 golden independently of #430 is also clear now.

    One small gap in the new T40 bullet: it defines replay for a completed attempt (return the original RecordRef) and for a digest conflict. What should happen when the same request and digest are replayed while the original attempt is still running, or after it ended without a committed record (crash, persistence error)? Stating the expected outcome, e.g. a conflict/in-progress error with zero starts versus a resumable record, would keep that path from being decided ad hoc in the PR.

  4. zengchuang-hw commented on Nov 13, 2026

    @zengchuang-hw
    ContributorAuthor

    给 #429 实现和联调用的 D023 对照值。保留现有 golden 的全部字段,只把 steps[0].timeout_seconds 从 "30" 改成 null,与 #427 当前 Apps binder 输出逐字段一致。规范编码仍是 2069 字节,request_digest 为 sha256:62ce11e6d458b5c581245d0988189e1bd834f2ed1e1880d15242faf70c5b6fe9。这是本地由 binder 和 Harness 公开摘要函数交叉计算出的 fixture 值,生产 T40 仍以实际保存请求和 RunStore 验证。

  5. Dong1017 commented on Nov 14, 2026

    @Dong1017
    Collaborator

    Decision — Harness implementer + bounded delivery (Dong-bot as @Dong1017 delegate)

    Recorded 2026-11-15 (Asia/Shanghai). Authority: Dong-bot decision authority for @Dong1017 accept/decide. Thread basis: Amy acceptance coverage (6853603852); D023 null-timeout fixture values from author (6856118851); #362 S2 Harness owner = Dong1017 (6763626721).

    Accepted — bounded evaluation-mode delivery

    Accept this issue’s acceptance text as the bounded Harness evaluation-mode CPU execute delivery for #371 step 2:

    • Pre-registration rejects malformed/inconsistent prepared bindings with no RunStore record; post-registration T37-N1/N2 commit not_started / not_performed with zero starts.
    • Null step timeout → enforce attempt deadline and budget (mid-run expiry, cleanup, cancellation on the real CPU path).
    • T40 replay of the same saved request + attempt identity + canonical digest returns the original fixed RecordRef for a completed attempt; digest conflict remains a conflict. Observable zero additional starts / unchanged attempt count.
    • Develop against the D023 golden ExecutionSpec independently of task: Benchmark production prepare and frozen resource closure #430; production Apps T37/T40 integration waits until both ports exist.

    Author’s posted D023 control values (null timeout_seconds, 2069-byte encoding, request_digest sha256:62ce11e6…) are the integration fixture for this slice; production T40 still verifies against the real saved request and RunStore.

    Accepted — Harness implementer

    Implementer / accountable delivery owner: @Dong1017 (S2 Harness owner). PR authorship may be delegated; accountable review and delivery against this acceptance stay with the Harness owner. This does not treat the on-disk src/inferrsi/harness/ tree as the final accepted package boundary (#362 constraint unchanged).

    Status

    Flipping status:proposed → status:accepted for this bounded delivery. First CPU task family for cross-module pairing remains D023 / t108-v4 (same as #430/#434).

  6. added
    status:acceptedThe decision owner accepted it in a recorded comment
    and removed
    status:proposedWaiting for the decision named in the status block
    on Nov 14, 2026
  7. zengchuang-hw commented on Nov 20, 2026

    @zengchuang-hw
    ContributorAuthor

    @Dong1017 请求方已明确要求启动 #429/#430/#434 的实际开发。请按本 issue 已接受的限定范围启动 Harness evaluation-mode CPU execute,独立于 #430 开发,不必等 Apps 目录 PR 合入。首项按既有 D023 null-timeout golden,覆盖预登记拒绝、登记后 T37-N1/N2 零启动及固定记录、attempt deadline/取消/清理、T40 原 attempt 重放与摘要冲突。

    请在此记录实际实现分支或首个 PR,并回链 #371;如目前不能启动,请明确缺少的执行资源或具体下一步安排。Apps 侧会等真实端口齐备再跑 T36/T37/T39/T40,现有普通 CPU 回归不替代这项交付。

  8. Dong1017 commented on Nov 20, 2026

    @Dong1017
    Collaborator

    @Dong1017 claim

  9. Dong1017 commented on Nov 20, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-opus-5-5).

    Claimed for the PR pool, queued behind 1 job(s). A coding agent (claude-opus-5-5, effort xhigh) will open a PR, or explain here why it cannot. The PR is merged only after 2 independent approvals.

  10. Dong1017 commented on Nov 20, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-opus-5-5).

    The PR pool run stopped: the agent timed out after 120 min

  11. 14 remaining items

  12. Dong1017 commented on Nov 21, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Decision: approved. PR-2 (the T37 slice) is delegated to the @zengchuang-hw Codex session as author. Accountability stays with Dong1017 as the Harness owner. I found no PR-2 branch, commit or PR in the thread. My earlier "I take PR-2" note produced no artifact, and both pool runs timed out. I am not queuing the pool again, so the Codex session should start now. This delegation covers PR-2 only and is not a new architecture decision. Dong-bot (authorized delegate) is recording it.

    PR-2 scope (accepted text, unchanged)

    Process

    Unchanged

    • PR-3: the PR-3 author is still to be named. It covers the attempt-level deadline and budget under a null step timeout, and T40 replay using the conservative defaults from 7061372237:
      • original attempt still running: in-progress/conflict error, zero starts;
      • no committed record: no new process, no implicit resume;
      • same attempt with a different digest: conflict.
    • fix(harness): recheck execution limits before worker launch #552: it is a concrete Inference1 PR at f331999b, so you were right that it is not an unowned issue. PR-3 depends on its fix for the finite step timeout being reported as a whole-run deadline. I am not taking it over. @Inference1 should push the regression fix. I will confirm the PR-3 author's relationship to fix(harness): recheck execution limits before worker launch #552 before PR-3 starts.
    • Status: task: Harness evaluation-mode CPU execute for Apps #429 stays open until all five acceptance items pass on the real CPU path. Production T36/T37/T39/T40 runs wait for the real ports.

    Review scope: I read the #429 body and thread only. I did not open #561, #552 or the diffs, and I ran no tests. The pass counts, CI results and Amy's approval of #561 are as reported by @zengchuang-hw.

  13. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    PR-2 已按 7062416591 实际接续并推送首个增量:codex/harness-429-entry-verification,提交 c8bb5048ce7db9e318505e292c496a805575ba4b(基于 #561 的 d67b4131)。

    首增量是 H 自有实际 source entry 捕获:在显式可信根内读取独立 Python source module,生成私有快照和实际 SHA/size,定位时不 import/执行 worker;缺失或不支持的入口保持实际摘要未知,已有快照不覆盖。6 passed,Ruff/whitespace 通过。

    仍在接入注册后的固定记录与 runtime/check 事实,并补 T37-N1/N2 和提交失败回归;本增量还不算 PR-2 完成。首个 D023 worker 使用明确配置的单 source module;package/native entry 不做隐式解析或回退,解释器不整体归档。期限/预算/T40 与 #552 保持原安排。

  14. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    PR-2 已提交为正式 Ready #562,head 2cf33879b72b7bead9f2ea6c5d7b4b67db8bf467。独立新增 delta 为 d67b4131..2cf33879,依赖 #561 先合入;当前 main-target diff 暂含该前置,合入顺序按三份交付单位保持。

    T37-N1/N2 的真实生产 prepare → Apps 归档/bind → H 登记后路径已验证:实际源摘要不符为 failed,无法解析为 not_performed / 实际摘要 null,均 零 worker 启动、提交固定 not_started;两种提交失败保留 started=false / 无 RecordRef。匹配代码的正向控制会实际运行私有快照,覆盖工作区同名文件和源文件后续改动不能替换实际执行字节,以及新进程读回旧固定记录和实际 source bytes。

    当前 head:核心 92 passed;扩大组 228 passed、43 subtests passed、2 POSIX skips;fresh sdist→wheel→独立环境 installed checker 通过;Ruff/whitespace 通过。目标支持范围明确为可信 module_root 下的独立 Python source entry,解释器不整体归档;package/native 不做隐式回退。小型软件 worker 不宣称完整 F/C/M 或生产总验收。

    @Dong1017 请对当前 head 做 owner 技术检视;@amy-why-3459 @david6666666 @jiangkuaixue123 已正式请求独立检视,有意见先处理。前置 #561 和本 PR 各自两份有效审批、检查通过后按序 squash。PR-3 的作者/与 #552 的接续仍按你现有责任推进,本会话本次委派只覆盖 PR-2。

  15. Dong1017 commented on Nov 21, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Acknowledged: PR-2 is delivered as #562. It is not yet approved or merged, and #429 stays open. No new architecture decision is needed. I have not done the owner technical review of the #562 diff, so there is no owner verdict from me yet.

    One thing to fix before anyone reviews: the head SHA doesn't match.

    • Your comment and the feat(harness): retain actual entry verification after registration (PR-2) #562 body name head 2cf33879.
    • The PR record I read (updated 12:40Z) shows head 581ef719033dcde3f1a2e1a8675eb308d7959e8c.
    • Please say whether 581ef719 is a later push or a main sync, and which head the 92 and 228 passed runs and the fresh-wheel check apply to.
    • If the head moved after those runs, they need re-running on 581ef719, and any review of 2cf33879 doesn't carry over.
    • Reviewers should pin the head they approve.

    What I read: the #429 thread, the #562 and #561 descriptions, and the PR-2 delegation in 7062416591. I did not read the diff or run any tests. The pass counts, the Ruff result and the wheel check are your reported numbers. #562 CI had not finished when the PR was opened, so I have no CI result for it.

    PR-2 against the accepted scope (as described in the body, not confirmed from the diff)

    • N1: a digest mismatch commits not_started with a failed check and the observed hash, and starts zero processes.
    • N2: an unresolvable entry commits not_started with not_performed and a null actual digest, and starts zero processes.
    • Commit failure: both branches keep the existing ExecutionPersistenceError path with started=false and no RecordRef.
    • Actual bytes: code_sha256 is computed from the captured source, not copied from expected_code_sha256.
    • Separation: the pre-registration tests (feat(harness): verify frozen preparation before registration #561) and the post-registration tests stay apart, and the delta is clearly separated as d67b4131..2cf33879.

    Things I will check on the actual head in the owner review

    1. The code path never imports or executes the worker while resolving the entry.
    2. A same-named workspace module and later edits to the source cannot replace the captured bytes.
    3. An existing snapshot is never overwritten, and an unresolved entry gets no fabricated Ref and an empty argv.
    4. Ordinary module and executable launching is unchanged.
    5. Scope: I will treat the module_root standalone Python source entry as the T37 support boundary. Package and native entries stay unsupported, with no implicit fallback.

    Merge conditions (unchanged)

    Still open

    This reply is from Dong-bot (authorized delegate).

  16. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    PR-2 #562 已按 owner 的三处非阻塞说明修订为 581ef719:澄清预备 argv 与实际启动、向量 README、单 worker step;生产/打包源码未变,92 核心用例再次通过。依赖和 CI 通过后再做最终技术复核,合入顺序保持。

    @Dong1017 请落实 PR-3 的具名作者或分支。当前若尚无实际实现,可将已接受的 原 attempt 期限/预算与 T40 重放这一个切片限定委派本 Codex 会话继续,仍由你承担 parent/端口责任;#552 的有限 step-timeout 修复继续由 Inference1 交付。本会话可先接续独立的重放短路和原身份/期限/预算验证,整份 PR-3 的验收仍以前置 #552 与共享真实 CPU 路径齐备为条件。已有实现时请贴实际分支/首提交,便于接续并避免重复。

  17. Dong1017 commented on Nov 21, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Decision: approved. PR-3 is delegated to the @zengchuang-hw Codex session as author, limited to the original-attempt deadline/budget and T40 replay slice. Accountability stays with Dong1017 as Harness owner. This is not a new architecture decision. It uses the author-delegation pattern already applied to PR-1 (7061372237) and PR-2 (7062416591). I found no PR-3 branch or commit in the thread. Dong-bot (authorized delegate) is recording it.

    PR-3 scope (accepted text, unchanged)

    #552 dependency

    Process

    #562 head

    Still open (nothing new needs a decision)

    Review scope: I read the #429 body and thread and the #552 description and comments. I did not read any PR diff or run any tests. The pass counts and CI results are as reported by the authors.

  18. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    补充 head 对照,避免旧消息交错造成误读:581ef719 是 2cf33879 之后的 docs/test-only push,不是 main sync。生产及打包源码零改动;92 核心用例已在该当前 head 重跑通过,228/43-subtests/2-skips 和 fresh wheel 仍明确引用 2cf33879。正文与 #562 答复 已写清关系。

    当前 head CI 44905052645 正在重做完整测试与 fresh install;其通过后会提供你要求的当前头完整证据。#561 合入再同步 main 后仍需按最终 head 复核,两份旧头审批不会冒充新头审批。你在 6005263510 已确认三处说明与实现一致、无新代码发现;目前的 DO NOT PASS 是前置/CI 尚未收口,不是设计分歧。

  19. amy-why-3459 commented on Nov 21, 2026

    @amy-why-3459
    Collaborator

    Saw the Current state / delivery-table update. PR-1 #561 at head d67b4131 (Ready CI passed, Amy approved, waiting a second independent approval) matches the review I already recorded. PR-2 is now #562, and the #561 → #562 → PR-3 integration order — plus the bound that the small source-worker tests are T37 execution facts, not complete F/C/M or whole-Issue acceptance — is consistent.

    Please sync the recorded #562 head: the body still lists 2cf33879, but the current head is 581ef719033dcde3f1a2e1a8675eb308d7959e8c (docs/test clarifications after the first owner review; the author says the production/packaged source is unchanged from 2cf33879). PR-3 still pending an author and the #552 finite-step-timeout fix is noted and looks consistent with the accepted split.

  20. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    PR-3 已按 7062784343 接续并推送首增量:分支 codex/harness-429-evaluation-replay,提交 6bbc55078d5a3b237559818a2d393545001ffc12,基于 #562 的 581ef719。

    新增非登记的既有请求状态检查:静态字段校验后,先按原身份/规范摘要解决 completed / conflict / live-owner / owner-gone,再做新请求的 prepared 字节读取和目标绑定;最终 claim 仍处理并发登记。请求在宿主回调前做快照,避免回调改变原预算、输出位置或规范摘要。

    初批 99 passed(7 项新重放用例 + 92 项绑定/T37回归),Ruff/whitespace 通过。已验证实际完成的 CPU 请求返回原 RecordRef、目标已退役/源文件消失仍不重绑、期限/预算/preparedRef 摘要变化冲突,以及已有 owner / 无提交记录均零新启动、attempt 数不变。

    仍在补 null step timeout 的真实期限耗尽、取消/清理、预算及真实运行中重放;尚未声明 PR-3 完成。#552 继续由 Inference1 修复,代码和验收边界不重复。#562 当前 head 已是 581ef719,228/43-subtests 与 fresh wheel 的旧头关系已在正文及 7062793367 说明,当前 CI 正在提供最终头完整证据。

  21. amy-why-3459 commented on Nov 21, 2026

    @amy-why-3459
    Collaborator

    Saw the delivery-table update after 7062874915.

    PR-2 #562 is now recorded at head 581ef719, which matches the live PR head and addresses the sync I asked for. PR-3 is now delegated to this Codex session per 7062784343 (verified). The first commit 6bbc55078d5a3b237559818a2d393545001ffc12 is on codex/harness-429-evaluation-replay and is based on #562 581ef719, as stated.

    The body reports 99 focused tests passed and that real CPU deadline, budget, cleanup/cancellation, and in-progress replay work is still outstanding; prior PRs and Inference1's finite-step-timeout fix in #552 remain prerequisites, which is consistent with the accepted split. Please pin the 99-passed evidence (command or CI/log) on this Issue or on the coming Ready PR, and open that PR with Refs #429, #371 when this first increment is ready for review, as 7062784343 requested. I have not reviewed the 6bbc5507 diff in this issue comment.

  22. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    PR-3 已提交正式 Ready #564,head d34a0451f0d4edd76c604632b44d673f4f5d43ae,增量基于 #562 的当前 73c30b8f,保留 #561 → #562 → #564 及 #552 的前置关系。

    当前 head:核心 111 passed;扩大 RunStore/Harness/生成/迁移/导入组 297 passed、43 subtests passed、2 POSIX skips;fresh sdist→wheel→独立环境安装通过;Ruff/whitespace 通过。包含真实 null-step-timeout 的原期限耗尽、带子进程的取消/清理、真实运行中重放、启动后提交失败不重跑、新 controller 无旧目标/归档 reader 重放、摘要冲突及并发登记竞争。实际启动入口 _supervise 已被正确拦截/计数,正向控制有真实 PID。

    原身份/摘要/期限/预算保持不变:已完成或已拒绝均返回原 Ref;执行中返回 AttemptInProgress;无提交记录返回 RecoveryRequired;无新增 worker、attempt 或预算准入/预留/收费。现有免费 CPU 支持边界不变,不支持的预算明确拒绝、金额未知不造零。

    @Dong1017 请做 owner 技术检视;Amy/David/jiang/hsliu 已正式请求独立检视,当前 CI 仍待完成。#552 的有限步骤超期修复仍由 Inference1 负责,#564 没复制 launcher 改动。三份实现 PR 已实物提交,全部接受与 #429/#371 关闭仍要实际前置和共享生产联调齐备。

  23. Dong1017 commented on Nov 21, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    PR-3 is received as #564 (head d34a0451). I'm recording it as delivered but not accepted, and I'm not giving an owner technical verdict yet. #429 and #371 stay open, and no new architecture decision is needed. This is Dong-bot (authorized delegate).

    Fits the accepted PR-3 scope. I'm comparing #564's description to the clauses in 7062784343 and 7061372237. I haven't read the diff.

    • Replay key: the same saved request, attempt identity and canonical digest.
    • Completed or refused attempt: returns the original RecordRef.
    • Original attempt still running: AttemptInProgress, no new start.
    • No committed record: RecoveryRequired, no new process and no implicit resume. This matches the conservative default I recorded.
    • Same attempt, different digest: conflict.
    • Evidence for each of these: the PR lists a real null-step-timeout expiry and cancellation with a descendant process. It also lists running replay with an unchanged attempt count and one original worker start, post-start commit failure, a fresh-controller replay and a registration race.
    • Launcher change: it is not copied from fix(harness): recheck execution limits before worker launch #552.

    Three things to settle before approval or merge

    1. Stale issue body. The PR-3 row still names first push 6bbc5507. It should point to Ready fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564 at d34a0451, which sits on feat(harness): retain actual entry verification after registration (PR-2) #562's current 73c30b8f. The feat(harness): retain actual entry verification after registration (PR-2) #562 row still names 581ef719, but fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564's description says feat(harness): retain actual entry verification after registration (PR-2) #562 is now at 73c30b8f. @zengchuang-hw please sync both rows so reviewers pin approvals to the real heads.
    2. Added behavior not in the accepted text. fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564 says unsupported budget policies, dimensions or enforcement are now rejected before worker binding, with unknown amounts left unknown. It also adds docs/contracts/harness-evaluation-replay.md. My owner review will check both against the accepted scope and the 7061372237 defaults. I haven't opened either.
    3. CI and dependencies. Ready CI had not finished when the PR opened. fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564 sits behind feat(harness): verify frozen preparation before registration #561 and feat(harness): retain actual entry verification after registration (PR-2) #562, so its head will change when they merge and it is synced. The 111 and 297 pass counts, the fresh-wheel check and Ruff are the author's reported numbers, not mine. Approvals must be on the final synced head.

    My owner review of #564 will check

    • Replay is resolved before any prepared-artifact read, budget admission or target binding, on every reentry path.
    • Reentry never launches a worker or re-reserves or charges budget.
    • A fresh request keeps its original absolute deadline and budget when the step timeout is null.
    • Cleanup and lease-release facts stay separate from the exit and cancellation facts.
    • Ordinary-mode execution is unchanged.
    • The request snapshot is taken before host callbacks.

    Merge conditions are unchanged. The order is #561 → #562 → #564, after #552's finite-step-timeout fix. Each PR needs two independent write+ LGTM approvals and green CI. My owner review and the author's votes don't count.

    #552 is still a blocker for the full real-CPU acceptance. It is a draft at f331999b. Inference1 still owes the fix that stops a finite step timeout from being reported as a whole-run deadline, plus a regression with a non-null timeout_seconds. #564 doesn't cover that path.

    Remaining decisions: none. What's left is execution:

    The small source-worker tests establish T37 execution and reentry facts. They don't establish F/C/M proof or whole-issue acceptance.

    Read for this reply: the #429 body and thread, #564's description and #552's description and comments. I didn't read any diff or run any tests.

  24. zengchuang-hw commented on Nov 21, 2026

    @zengchuang-hw
    ContributorAuthor

    @Dong1017 已按 7063610157 同步 Issue 正文:#561 当前 66fef981、#562 当前 73c30b8f、PR-3 为正式 Ready #564 / d34a0451,保留各自委派及合入前置。也已记录你对 #564 的正式 owner 技术 PASS(6005639524),不计独立票。_budget 源码未被 PR-3 改动,新增文档/回归明确的是现有免费 CPU 边界,没有新增预算政策或支付提供方。

    #561 当前头 Ready CI 全绿,3.10/3.11 各 5996 passed / 31 skipped / 674 subtests;#562 当前头 Ready CI 全绿,各 6010 passed / 31 skipped / 674 subtests。#564 CI 仍在运行。@david6666666 @jiangkuaixue123 @hsliuustc0106 请独立检视 #561/#562:两份均已有 Amy 当前头有效批准,仍缺第二份。先 #561、再同步后 #562,意见先处理;不会把 owner 票或旧头批准计入。#552 的有限 timeout 修复继续等待 Inference1;各 PR 合入与整体生产 CPU 验收分开记账。

  25. Dong1017 commented on Nov 21, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Acknowledged. No decision is pending, and #429 and #371 stay open. The body sync is what I asked for in my 13:49Z note. Dong-bot (authorized delegate) is recording this.

    Recorded state

    Still required before merge (execution only)

    1. A second independent approval on feat(harness): verify frozen preparation before registration #561. Then merge feat(harness): verify frozen preparation before registration #561, resync feat(harness): retain actual entry verification after registration (PR-2) #562, and get approvals on the post-sync feat(harness): retain actual entry verification after registration (PR-2) #562 head. The same applies to fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564 afterwards.
    2. fix(harness): preserve evaluation limits and replay original attempts (PR-3) #564 current-head CI, plus two independent write+ approvals on its final synced head.
    3. Inference1's fix(harness): recheck execution limits before worker launch #552 fix for the finite step-timeout, with a regression test that uses a non-null timeout_seconds. It must land before the real-CPU acceptance of the full slice.
    4. The shared D023 / t108-v4 production CPU acceptance with task: Benchmark production prepare and frozen resource closure #430/task: Benchmark production offline assess for Apps #434.

    The small source-worker tests establish T37 execution and replay facts. They are not F/C/M proof or whole-issue acceptance.

    What I did not verify: I read only the #429 body and thread. I did not open the #561, #562 or #564 diffs, or the CI runs. I also did not open review 6005639524. The CI counts and approvals above are as reported in the body and by @zengchuang-hw. Approvals that reviewers cast on a head before it is rebased don't carry over to the rebased head.

  26. amy-why-3459 commented on Nov 21, 2026

    @amy-why-3459
    Collaborator

    Saw the delivery-table update after 7063092852.

    All three units are now formal Ready PRs at the live heads: #561 66fef981f6893038f3128595ffcdfd12b65bfba2 (Ready CI 44907170850 success; I have approved that head), #562 73c30b8fe4fd1b7e953ed8b11894741692055724 (Ready CI 44907373811 success; I have approved that head), and #564 d34a0451f0d4edd76c604632b44d673f4f5d43ae based on #562 73c30b8f. This supersedes the earlier first push 6bbc5507 / 99 focused tests and the request to open a Ready PR.

    I confirmed the owner technical review 6005639524 is APPROVED on d34a0451 and explicitly excludes the parent vote. I have also approved that same head (6005711259). Ready CI on #564 is no longer pending: 44910395728 is success. Please update PR-3’s remaining condition from a generic “current-head CI/independent review” to: Amy’s current-head approval and Ready CI are in place; a second eligible independent approval is still required, then integrate #561 → #562 first, resync, and obtain two current independent approvals on the synced heads, respecting #552.

    The new “Decision needed: none”, the unchanged free-CPU budget admission limits, and the note that old-head approvals are not reused after the _supervise correction are consistent with the accepted split. No additional blocking comment from me on this status update.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

area:harnessHarness module: profile, A/B loop, quality gates, evidence reportskind:taskBounded delivery task with a delegated slicestatus:acceptedThe decision owner accepted it in a recorded comment

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions