Skip to content

Record current-model Core task recovery evidence - #8777

Merged
cocolord merged 1 commit into
mainfrom
codex/r1-model-recovery-20261114t151637
Nov 14, 2026
Merged

cocolord merged 1 commit into
mainfrom
codex/r1-model-recovery-20261114t151637

Conversation

@cocolord

@cocolord cocolord commented Nov 14, 2026 •

Copy link
Copy Markdown
Member

R1 的既有检查主要使用确定性 Host。此 PR 在双语 RFC 中记录当前交互模型实际生成报告后的中断恢复证据,并明确与 C1 共用完整来源和 Turn owner。

The R1 checkpoint previously relied on deterministic Host fixtures. This bilingual RFC update records recovery of a tool-produced result from the current interactive model while retaining the existing C1 and Turn owners.

  • Pinned sources: main 83809355f84a and C1 C1: unify work sources and validate recovery #8266 at 6a31d83497a6, each installed separately as Core. Six executed entrypoint/owner files match the source bytes.
  • The model produced JSON and Markdown inventory reports from isolated synthetic input. An independent Decimal oracle checked all groups and totals. Both confirmed results survived SIGKILL and resumed the original Turn with one Host invocation and one debit after replay.
  • On C1, a corrupted report failed validation with the Todo open and zero debit. Restoring the checked artifact allowed the original Turn to finish.
  • Only the paired RFC checkpoint changes. The private relay, input, artifacts, runtime state and raw logs remain outside the PR. No product pipeline, schema, default, model service or capability activation changes.

Validation: the current C1 installed Core and context regression passed 39 cases; exact-head bilingual and diff checks passed. The pinned C1 source-contract regression passed 325 cases. Risk-based premerge passed four direct and 11 automated checks with no failures.

Remaining acceptance: installed native model/App adoption, natural-language onboarding, native compaction retention, progressive controls, full D2/D5 closure, comparable complete costs, V1–V5/V10, R1/M3 and C1 default cutover. The model reused its generator between checks; this is not an independent inference or cost experiment. Full original-objective regression is still required before Goal completion.

RFC alignment / RFC 关联:

  • Primary: Small-kernel RFC / 中文 RFC. This is an evidence increment for V4 and the R1 milestone in §11: confirmed current-model result, interruption, validation and original-Turn readback. It does not close the complete V4 or R1 criteria.
  • Remaining original rows: V5 native model/context continuity and V10 comparable complete overhead; the full V1–V5/V10 requirement for R1 stays open.
  • Shared owner: Controller/executor unified contract RFC, C1 C1: unify work sources and validate recovery #8266 at 6a31d83497a67b58f3da981a3869bea4e4b14e49. The probe reuses its complete source, identity, admission, capture, recovery and settlement owners. This PR does not authorize C1 default cutover.

Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com>
Co-authored-by: TRAE CLI <traecli@bytedance.com>

@cocolord cocolord left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewer: model_agent; model=GPT-6; provider=OpenAI; self_reported (host execution observation unavailable).

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

评审精确头:bc88200babe4ea6b6cf13d633ecb1372f9342b19。没有发现阻塞问题。

动机 / Motivation

需要判断最小组合能否在 Agent 中断后可靠返回成果的使用者与维护者。 以前确定性 Host 的检查通过,仍不能判断实际模型的报告能否在进程被终止后读回;现在模型生成的报告经过独立校验,在两个固定版本上恢复并只结算一次。 报告的模型生成、独立校验、原 Turn 恢复与单次扣额已有证据,安装原生 Host 和完整成本仍明确未验收。

不新增执行管线,不改变能力激活或默认值,不完成 R1/M3,也不将两次复用生成器的检查当成推理或成本对照实验。 原生 App/模型接入、自然语言启动、compaction 后的模型保留、渐进控制、完整 D2/D5 闭包及可比完整成本仍未验收。 这个差异是有明确边界的证据增量,不能用合入数量或测试数量替代原目标验收。

改动思路 / Approach

沿现有 Core 的通用 Host 请求和 Turn 执行器验证。模型读取实际请求,再用工具生成报告;独立校验器从原始输入重算,而不是接受模型自报“已验证”。待结果进入持久化日志的 typed_result 阶段后终止拥有该 Turn 的 CLI,再以原键恢复。写回、配额和任务状态始终由既有 owner 决定,私有消息 relay 不获得这些权限。

主干与 C1 的请求、报告、journal、读回和扣额均已经检查。C1 的固定版本移除了 complete planning 的双输入;本 PR 不改变该实现,Git merge-tree 组合也没有冲突。相较新增模型执行器或公共 smoke,直接维护现有 RFC checkpoint 避免了第二套执行管线和重复覆盖。

具体改动 / Changes

关键内容讲解

英文 checkpoint 增加 28 行,中文对应文件增加 21 行。来源、实际成果、恢复、owner 边界和剩余验收逐项对应。公开差异不包含 relay、CSV、报告、运行时状态、原始日志或本地路径;没有生产代码、配置、schema 或默认值变更。

修改前依据为 docs/architecture/rfcs/monorepo-distribution-split-v0.md,spec_revision:83809355f84aac19649924416cd8b90d1ff9e79c,具体是第 9、11 节。按原条目映射:

原标准 本 PR 的处置
V4 deferred:确认后的实际模型成果恢复、独立校验和读回已有证据;原生 Host 接入及完整恢复成本仍需验证。
V5 deferred:已有 protocol context 回归不等于原生 compaction 后的模型保留,继续沿 C1 的原上下文 owner 验证。
V10 deferred:未进行完整进程/IO/IPC/模型 token 成本对照,复用生成器的两次检查不能证明成本收益。
R1 deferred:完整 V1–V5/V10 与 split-package 安装资格不由本 PR 完成,原 R1 和 D2/D5 工作保持开放。

这些原标准没有修改为当前已经通过的较小条件。

对主干的风险 / Risks

主要风险是读者把私有 relay 的结果恢复误认为已安装原生 App/模型接入,或把复用生成器当成独立推理实验。两种语言都明确了 synthetic 输入、当前模型的工具执行、安装 Core 的边界和剩余项。报告的所有组与金额由独立 Decimal oracle 校验;损坏报告的反例直接在 owning Turn 恢复中执行,没有放宽断言或更换身份。

验证 结果
main 与固定 C1 的当前模型报告及 SIGKILL 恢复 PASS;各一次 Host、一次扣额,Todo done,重放不重复效果。
C1 损坏报告、修复和原 Turn 重试 PASS;错误成果拒绝、Todo open、零写回及零扣额,修复后完成。
当前 C1 的独立 Core/context/SSH 回归 PASS;39 项。
双语 staged 与精确 HEAD 检查、DCO、公私扫描、风险 premerge PASS;4 项直接检查、11 项自动检查。验证期间文档字节不变,提交后精确头双语检查再次通过。
GitHub CI PENDING;观测时 5 项排队,Summary 成功,一项旧 PR 取消检查跳过。未把排队标成成功;合入就绪另行读回。

C1 的额外来源合同回归仍在运行,合入前会读回结果。该补充批次与公共文档改动区分记录。整个原目标的回归仍是最终完成前的要求。

我的整体评价 / Verdict

这是 justified_increment:实际模型成果恢复有新增独立证据,双语记录准确保留了已验证、未验证和不适用的范围,能够独立回退。long_horizon 为 preserved:文件差异不改变执行或继续规则,并明确保留原目标的剩余验收,避免把一次成果恢复当成整个长程目标完成。 user_experience 为 not_applicable:两份 RFC 的末尾 checkpoint 不修改 App、Lark、CLI 用户操作或首屏;自然语言启动与产品控制旅程继续作为未验收项。

仅修改两份 RFC 的末尾事实记录,无 App、Lark、CLI 或页面布局改动,未验证任何新增视觉产品旅程。 没有新增语义词表、兼容 decoder 或需要额外迁移的 authority。相邻代码重构在本次文档边界内无必要;继续复用 C1 与现有 Turn owner。最强的缺失验收仍是原生 Host/模型采用、原生 compaction、普通用户入口和可比完整成本,它们不能由本次报告恢复代替。支持合入这个事实 checkpoint;不批准 C1 默认 promotion 或宣告 R1/Goal 完成。

English verdict: APPROVE - bc88200; accurate bilingual current-model task-result recovery checkpoint, with independent positive/negative validation, Core39 and premerge passing. Native Host adoption, full R1 and cost acceptance remain open; CI readiness is separate.

@cocolord

Copy link
Copy Markdown
Member Author

Additional validation at unchanged head bc88200: the C1 complete-source/Turn/context regression at 6a31d83 completed with 325 passing cases. The installed Core/context suite passed 39 cases, and risk-based premerge passed four direct and 11 automated checks. The published exact-head approval remains scoped to this bilingual factual checkpoint. Native model/App adoption, complete R1 acceptance and C1 default promotion remain open. / 补充回归已经完成:当前固定 C1 来源合同 325 项、独立 Core/context 39 项及 premerge 均通过。原精确头评审只支持这份双语事实记录,完整 R1、原生 App/模型采用和 C1 默认切换仍未验收。

@cocolord
cocolord merged commit 80c6949 into main Nov 14, 2026
7 of 11 checks passed
@cocolord
cocolord deleted the codex/r1-model-recovery-20261114t151637 branch November 14, 2026 15:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant