Skip to content

docs(rfc): qualify first-Goal result adoption and recovery - #8788

Merged
cocolord merged 1 commit into
mainfrom
codex/r1-task-artifact-20261114t172916
Nov 14, 2026
Merged

cocolord merged 1 commit into
mainfrom
codex/r1-task-artifact-20261114t172916

Conversation

@cocolord

Copy link
Copy Markdown
Member

Summary

The minimal-composition RFC now distinguishes first-Goal reply/readback evidence from the automatic governed entry that remains unqualified. Matching English and Chinese ledger entries record the original Task's source recovery qualification and its limits.

RFC alignment: R1/V4 checkpoint, Chinese mirror, and C1 contract / #8266.

The qualification uses real Chat/action/store on isolated state, a current Host model supplied through temporary stdio, independent CSV/Decimal report checks, actual Chat restart and original-action replay. The same initial Goal/Agent/Task then adopts that reply through the existing Core + SQLite + generic Host route with an explicitly declared fixture validator. Confirmed-result SIGKILL recovery, corrupted-result refusal, completion and replay retain one Core Host invocation and one debit.

Author Declaration

  • Written by: model_agent, OpenAI GPT family.

Implemented against

Criterion (spec clause) Disposition Symbol / path Test or command
Section 5: reuse governance, Host and Task owners implemented Paired RFC checkpoint, Governed adoption and Remaining entry gap Existing Task validator, Core, SQLite and generic Host exercised on the same initial Task
Section 9 V4: distinguish recovery evidence from whole-journey acceptance implemented Paired RFC checkpoint, Observed product path and Governed adoption Independent report oracle; real Chat restart/action replay; confirmed-result SIGKILL, corrupted reply rejection and same-Turn recovery/replay
Section 9 V1–V5/V10 and section 11 R1/M3: preserve remaining qualification deferred Paired RFC checkpoint, Boundary Full native entry, compaction retention, complete selected closure and comparable costs remain open
Section 11: source composition does not qualify split-package installation implemented Paired RFC checkpoint, Source and Boundary Source/transport evidence explicitly scoped; no package publication or C1 cutover claim
  • Self-check before submission: read both complete changed ledger entries against the cited clauses and private source evidence; checked translation equivalence, DCO and public boundary. No native installed-model, full-cost or full-objective pass is assumed.

Scope And Continuation

  • Completed scope and remaining work: complete bilingual evidence checkpoint for the real first-result readback and original-Task recovery boundary. The first product Chat still starts a read-only planning conversation; automatic governed adoption remains to integrate and qualify through the registered managed-session, Task validation/admission and C1 owners.
  • Slice boundary / successor: continue the existing R1 work against all retained criteria. This ledger update makes the next integration gap reviewable and can be reverted independently; it introduces no execution pipeline.

The first product Chat request still uses ordinary read-only planning; the fixture's validator declaration and governed adoption are explicit test-caller steps. This does not qualify native App/model onboarding, compaction retention, full D2/D5 closure or comparable costs, full R1/M3, or C1 cutover. The recorded 24,372 input bytes are one observation, not a performance comparison. Runtime/schema/activation/authority and evidence policy are unchanged; temporary scripts, captures and raw results stay outside the PR.

loopx check: the public-file scan passes for both changed RFCs. The overall registry check still fails on six unrelated pre-existing Goals without canonical authority; that failure is retained and is not reported as a pass.

Validation

  • Tested revision: f2177ac1154df75b576e73889b64674e3e1598df; runtime qualification uses its unchanged source base 74ca883d5bd3e4bc64671bac54210eac9e97ed05. The combined C1 source uses 2c099be41e6878298ae273de28b983fc2c205700.
  • Run state: finished.
  • Input classes: synthetic.
Check kind Result Public-safe evidence / limitation
static passed Diff-scoped premerge: four direct checks including committed bilingual RFC diff, and thirteen automated catalog/risk-smoke/public-boundary checks
real_entrypoint / real_backend passed Isolated real Chat/action/store and SQLite; independent CSV/Decimal oracle, original action/Session restart and replay; original Core Task validator/recovery/replay. Temporary stdio uses the current Host model; this is not installed native App/model qualification
integration passed Exact base + C1 combined source: 287 native-session, work-context, first-read and settlement tests, isolated state and synthetic provider/model fixtures; no C1 approval or default cutover
static passed Current source Chat bundle build/verify and staged bilingual check; public-file scan passes both changed RFCs and this description
static failed Overall project-registry check encounters six pre-existing unrelated Goals without canonical authority. Its separate public-file scan passes; unrelated authority migration is outside this diff
manual passed Actual packaged Goal conversation renders the original report; private capture retained, no UI change or public screenshot required
regression_parity not_run Full original-objective, installed native-host/model, compaction and comparable-cost qualification remains open; scoped tests do not establish these claims
  • Coverage and gaps: this PR changes only two non-normative RFC entries. The decisive checks cover the facts and limitations those entries state. Runtime and backend evidence supports the recorded boundary; the overall registry failure and full-objective gaps remain explicit. No source code, schema, activation or authority is modified.

Frontend / Visual Evidence

  • UI impact: none.
  • Before / After / First viewport and preview / States and viewports shown: N/A, documentation ledger only.
  • Source data: none in public attachments; qualification inputs are synthetic.
  • Attention review / Craft review: N/A, no product presentation change.

Type of Change

  • Documentation update

LoopX Area

  • Public docs or presentation surface

Technical Direction

  • Direction / acceptance reference: minimal-composition section 11 R1, section 9 V4; C1 controller/executor and unified-work-contract owner reuse. Remaining V1–V5/V10 and R1/M3 acceptance is retained.

Shared-authority RFC fixture impact

N/A: no TypeScript/shared-authority promotion, routing or compatibility-projection change is proposed. SQLite checks qualify only the recorded recovery boundary.

Boundary Checklist

  • Diff and public text contain no private state, credentials, raw logs, internal links or local paths.
  • No maintainer-owned benchmark work is duplicated or executed.
  • Change is scoped to the linked RFC checkpoint and existing continuation.
  • Both existing English/Chinese RFC files change together; committed bilingual check passes.
  • UI impact is none.
  • Commit has a DCO Signed-off-by trailer.

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; GPT; OpenAI; self_reported.

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

动机 / Motivation

评审最小内核入口的维护者需要判断首轮有用结果是否已被原任务可靠采用。 此前看到库存报告仍无法从公开 checkpoint 判断原任务能否在中断后完成;现在条目说明原任务恢复、损坏拒绝和重放结果,并保留自动入口缺口。 两份 RFC 同步记录真实结果读回、显式校验器、原任务的一次调用/扣额恢复,以及各项未验收边界。 本 PR 不交付自动接管、原生 App/模型安装采用、完整默认闭包、成本收益或 C1 切换。 首轮仍是只读规划对话;自动受治理采用、模型 compaction 保留、完整闭包和可比成本仍需验证。

做完一份报告和让原任务可靠完成是两个需要分别验证的结果。如果继续沿用此前的公开记录,下一轮评审无法区分已验证的恢复能力和仍需人工声明校验器的入口集成。本次更新现有双语记录,补齐这一判断依据。

改动思路 / Approach

沿现有 Chat、Task 和 Core/SQLite owner 逐段核验:真实首轮回复先通过独立库存输入校验,再由测试调用方显式声明校验器,原 Task 采用同一回复。持久确认结果后中断,损坏回复被拒绝;恢复原回复后完成并重放,保留一次 Core Host 调用和一次扣额。

这次 PR 只记录已观察到的来源和边界,没有增加执行框架、状态字段或默认策略。临时 stdio 转接和校验脚本不进入源码或 CI。无需画新的运行架构;分段文字已经清楚说明哪一步是实际服务、哪一步是测试调用方,以及缺失的自动接管。最小可行做法是复用既有双语记录,继续集成时仍使用注册的 managed-session、独立 Task 校验/准入和 C1 owner。

具体改动 / Changes

评审 head:f2177ac1154df75b576e73889b64674e3e1598df。规范依据:docs/architecture/rfcs/monorepo-distribution-split-v0.md,接受修订 74ca883d5bd3e4bc64671bac54210eac9e97ed05 的第 5、9、11 节;本次新增记录不改写这些验收。

关键内容讲解

  • 英文新增条目:来源固定到 main/C1,分别说明真实报告读回、显式受治理采用、剩余入口和非验收范围。
  • 中文对应条目:保留相同的合成输入、临时转接、损坏拒绝、一次调用/扣额、单次字节观察和未验收条件;逐条语义核对通过。
规范条款 本 head 的处理
Section 5 implemented:复用原 owner,未修改运行时或授权
V1、V2、V3 deferred:未声称完整闭包、两种 preset 的权限对照或完整命令兼容验收
V4 deferred:补充实际报告和原 Task 的来源恢复证据,显式 fixture 步骤仍不能验收自动原生入口和完整恢复成本
V5 deferred:会话/回复保留不能证明模型 compaction 或 lost-ACK context 采用
V10 deferred:24,372 UTF-8 字节是单次观察,未证明成对总成本收益
R1、M3 deferred:原 R1 继续;没有 split-package 发布或冷安装验收

验证

  • PASS:独立 CSV/Decimal oracle 验证八组及总计;真实 Chat 重启和原操作 replay 保留原结果。Core confirmed-result 中断、损坏拒绝、恢复完成和 replay 断言与读回一致。
  • PASS:准确 base+C1 组合树的四套会话、work-context、first-read、settlement 测试共 287 项;使用隔离状态和合成 provider/model,不能据此批准 C1 或其默认切换。
  • PASS:exact-head diff premerge 的 4 项直接检查和 13 项自动检查,包含成对 RFC、risk smokes 和公共边界;DCO 存在。两份 RFC 及 PR 文本扫描通过。
  • FAIL,既有独立诊断:同一完整项目检查在不可变 base 与本 head 返回逐项相同的六个 canonical-authority 错误,公共文件扫描均通过。没有将整体检查记成成功,也不要求此文档 PR 修改其他 Goal。
  • UNVERIFIED:初次 CI 观察为 1 项成功、6 项 pending;完整原目标、原生已安装模型、compaction 和成对成本回归仍未验收。合入 readiness 单独检查。

对主干的风险 / Risks

没有阻塞性发现。主要风险是把临时转接和显式校验器解读成产品自动接管,或把 287 项合成测试与一个字节数理解成完整入口/成本验收。两种语言均明确阻止这些推论,并保留原条款和历史未验收范围。

本 diff 的所有 tracked 路径都是两份 RFC;其余运行时代码逐字节保持 base 一致。损坏回复的实际失败仍保留,恢复通过没有改写旧失败。需要回滚时只撤销这两个条目。没有新增公开协议词汇、默认启用、权限或持久化兼容分支;同域源码重构在本次 docs 范围没有必要,C1 作者分支未被修改。

我的整体评价 / Verdict

APPROVE:这是有用且可独立审阅的证据增量,原最小内核目标保持开放。long_horizon=preserved:没有更改后续运行、预算或恢复语义,明确缺口有助于继续正确的集成工作。user_experience=preserved:没有增加用户操作或把显式测试步骤写成自动产品行为。

视觉评审不适用:只修改两份 RFC 的执行记录,没有产品页面、文档导航或展示控件变更。本 PR 不验收已安装原生界面或模型体验。下一步仍需原 R1 的自动入口集成、完整默认闭包、context 和可比成本验证,并在完整目标完成前跑原范围回归。评审归属是 self_reported;运行时模型观察不可用,不据此声称具体构建或独立模型验证。

English verdict: APPROVE f2177ac1154df75b576e73889b64674e3e1598df. No blocking finding in the paired RFC checkpoint; exact-head documentation, bilingual, public-boundary and scoped recovery evidence pass. Native automatic onboarding, full costs, R1/M3 and C1 cutover remain unqualified; CI and merge readiness are separate.

@cocolord

cocolord commented Nov 14, 2026 •

Copy link
Copy Markdown
Member Author

Authorized merge decision for reviewed head f2177ac1154df75b576e73889b64674e3e1598df.

The exact-head COMMENTED approval review is published and read back; its complete body passes the current capability checker. Approval closeout is clear, and review threads are complete with zero unresolved findings. Both changed paths are paired non-normative RFC entries; no runtime, permission, schema, activation, benchmark or evidence-policy behavior changes.

Required local validation passes: committed bilingual RFC check, diff hygiene, DCO, four direct and thirteen automated premerge checks, public-file/text scans and claim-by-claim evidence/translation review. The recorded original-Task corruption/recovery/replay uses isolated real Chat/SQLite and a current-model temporary relay; C1's conflict-free combined source passes 287 scoped synthetic-provider/model tests. Native automatic entry, full model/context/closure/cost qualification and R1/M3 remain open.

Final readiness returns ready=true: the unchanged exact head has a valid approval, a mergeable PR, complete review threads with no unresolved findings, and all 27 remote checks successful. GitHub still requires review policy for this author-owned PR and reports the README-only target advance. Under the existing explicit maintainer/admin merge authorization, use bypass for those GitHub policy holds only; no pending CI is bypassed. No unresolved review, conflict or required local validation is bypassed. The overall project-registry check still fails on six identical pre-existing canonical-authority errors on base/head; it is not reported as a pass.

The target has advanced from the qualified runtime source 74ca883d5bd3 to 2592d4a2a5e8 through a separate bilingual README-only commit. Inspection confirms no changed runtime or RFC owner; the committed bilingual check against the new target passes. This does not change the source-bound claims in this checkpoint or authorize C1 cutover. Merge only the unchanged reviewed head, then read back the merged tree and sync local main.

@cocolord
cocolord merged commit 58cc581 into main Nov 14, 2026
27 checks passed
@cocolord
cocolord deleted the codex/r1-task-artifact-20261114t172916 branch November 14, 2026 18:29
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