Skip to content

fix: qualify Codex 0.169 native context replay - #9044

Merged
huangruiteng merged 1 commit into
mainfrom
codex/c1-native-replay-writer-0169-20261119
Nov 19, 2026
Merged

huangruiteng merged 1 commit into
mainfrom
codex/c1-native-replay-writer-0169-20261119

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Goal And Delivered Outcome

Codex 0.169.0-alpha.11.1 writes paginated rollout history, but main has yet to qualify that writer. After a real App Server restart, the retained-read checkpoint is discarded and the selected context is delivered again. Qualifying this exact writer in the existing check allows the restarted thread to reuse the verified read and preserve the original heartbeat receipt.

Related to #8916, #8942, and #8266. Target main.

Author Declaration

Criterion Disposition Owner and verification
Qualified paginated writer and retained read evidence implemented Existing _snapshot check and source plus Core App Server restart
Matching session, workspace, model and permission mode with complete original output implemented Existing replay gate and extended direct and nested-output regression cases
Conservative recovery for unknown versions, changed continuity and invalid output implemented Existing invalidation cases with the qualified writer and an unqualified neighboring version

Scope And Continuation

The independently testable change is the single writer qualification and its reference and regression coverage. #8942 owns reuse across multiple selected pages. The clean composition keeps the current receipt and reconstructs the same input from the surviving native history. C1 still needs installed App adoption and the matched single-session outcome and complete-cost comparison.

Validation

  • Tested revision 7b7797df89db20cadaa1cd0983440f76aab9c9c6
  • Run state finished
  • Input classes synthetic, authorized_private_read_only
Check kind Result Public-safe evidence
real_entrypoint passed Codex 0.169.0-alpha.11.1 App Server restart against a synthetic loopback provider and immutable archived inputs. Main reused zero selected sections. The exact source proposal and an isolated Core wheel each reused the verified section and preserved the input and receipt. The 11 reader pages stayed constant.
regression_parity passed With #8942 at dca761a9e5e0, the clean Git composition restored all 1,152 omitted literals and kept the current receipt. It reused 19 sections, reduced reader pages from 11 to 4 and tool-output bytes from 239,474 to 80,540 for this replay.
unit passed 146 focused cases in tests/control_plane/test_codex_context_resume.py, including the neighboring writer, failed outputs, changed identity, damaged history, compaction and rollback.
static passed Ruff, configured mypy, bilingual and whitespace checks.
integration passed Standard local premerge at 29c312949b6c with an exact-scope quality receipt, 6 direct checks, and 11 selected checks. The maintainability advisory reports the same two unreviewed findings on main and on the proposal.
manual not_run Installed App adoption and the matched single-session outcome and complete-cost comparison are C1 program acceptance. The replay measures delivered bytes with zero paid model calls.

The broad archived replay uses private inputs and the public focused tests reproduce the compatibility and fallback contract.

Frontend / Visual Evidence

UI impact none

Type of Change

  • Bug fix
  • Documentation update
  • Test update

LoopX Area

  • Control plane (goals, todos, quota, scheduler, registry, runtime)

Direction and acceptance Controller/Executor Unified Contract, C1.

Shared-authority RFC fixture impact N/A.

Boundary Checklist

  • The diff and this description contain public-safe evidence.
  • Existing benchmark work is referenced within its original scope.
  • The change and validation are scoped to the exact writer qualification.
  • The bilingual diff check passes and the reference update covers both languages.
  • UI impact is recorded.
  • The commit includes a DCO sign-off.

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator 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-6.2-sol · OpenAI (runtime_reported, max)

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

动机 / Motivation

使用 Codex App Server 的长程 Agent 重启后续作,需要识别原线程已完整读取的上下文。实际 .11.1 paginated rollout 能完整回放选读,目标主干会清除 LoopX 的选读复用记录。候选在同一作用域内保留完整选读和原 Turn。身份变化、失败输出、截断和相邻版本继续走恢复路径。本次复核指定 writer 的续作资格、故障和 #8942 组合。

改动思路 / Approach

沿用 _snapshot 的物理文件身份检查和 _matches 的读取区间核验。SessionStart 在 native replay 后检查当前线程。现有 owner 判定有效输出能否用于后续交付。一行生产改动限定已复测的 writer 版本,配套双语参考和正反例。让同一合成会话通过真实 App Server、原生 shell reader 和生产 hooks。分别在目标主干、精确候选和组合树上重启,用实际模型输入复核原选择正文、Turn 收据和 LoopX 保留摘要。

具体改动 / Changes

精确 head 7b7797df89db20cadaa1cd0983440f76aab9c9c6,diff base 29c312949b6c26e3b27288234881eea2958ead94,3 个文件 +12/-7。逐项复核 exact-writer-and-identity、complete-and-failed-output、companion-replay。验收约定

关键代码讲解

  • 生产改动只将 0.169.0-alpha.11.1 加入精确名单。已有文件归属、session、cwd、paginated 模式和完整记录条件照常执行。源码
  • 真实 App Server 重启配对复原了 1,346 字节选读和原 Turn。主干清除 LoopX 复用记录,候选保留选读摘要和同一 epoch。源码
  • 146 个源码用例覆盖失败输出、权限和身份改变、截断、替换、丢失及相邻 .11.2。失败时使用原 capture 恢复。源码
  • 使用 #8942 的 dca761a 与当前精确 head。真实重启后,两组都可从原输出恢复完整四页 12,002 字节。#9044 单独运行时清除多页复用记录,组合树保留它。组合树另有 350 个 resume/lifecycle 测试通过。源码

对主干的风险 / Risks

独立验证使用真实 App Server、隔离源码与确定性回环 fixture。对照组保留原 Turn,候选另有 4 个真实压缩恢复用例通过。作者的 Core wheel、成本数字和安装版 Desktop/Lark 的实际采用未纳入本轮复测。

语义与 CI 对齐

版本字段描述原 rollout writer。资格仍依赖匹配的物理来源和完整输出。失配时从当前 capture 恢复所缺事实。原权限和接续凭据适用。

我的整体评价 / Verdict

批准当前精确提交。真实重启保留原完整选读及 Turn 收据,失败输出和相邻版本仍触发恢复。#8942 组合通过独立验证,实际安装采用和完整成本仍待 C1 验收。

English verdict: APPROVE. Exact-writer resume preserves a verified selection and the original Turn receipt. Unknown writers and incomplete replay fall back safely. The #8942 composition passed independent validation.

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
@huangruiteng
huangruiteng force-pushed the codex/c1-native-replay-writer-0169-20261119 branch from 7b7797d to 685cbea Compare November 19, 2026 06:44

@huangruiteng huangruiteng left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Reviewer: model_agent · GPT-5 · OpenAI · self_reported

审批结论 / Approval conclusion

动机 / Motivation

Codex App users restarting a thread written by Codex 0.169.0-alpha.11.1 previously lost retained-read eligibility and received selected context again. The candidate extends the existing exact-writer qualification only; all source identity and complete-output checks remain.

  • affected_caller_or_operator: Codex App users restarting threads written by Codex 0.169.0-alpha.11.1.
  • before_after_scenario: On the baseline, the 0.169 writer reused zero selected sections after App Server restart; the candidate reuses the verified section while preserving the original heartbeat receipt.
  • observable_outcome: The exact 0.169 writer is accepted, neighboring and malformed versions remain rejected, and 146 focused replay tests pass.
  • non_goals: Does not claim installed App adoption, matched single-session production outcome, or reuse across multiple selected pages owned by #8942.

改动思路 / Approach

reason: Exact allowlisting preserves explicit compatibility control. current_pr_boundary: One writer version and its documentation/test oracle.

_snapshot continues to own physical-file, session, workspace, model and permission checks. This PR adds one exact session_meta.cli_version value to the same allowlist; it does not widen version matching or create another replay path. R-QUALIFIED-WRITER, R-COMPLETE-TURN and R-FALLBACK are implemented at this head.

具体改动 / Changes

关键代码讲解

  • _snapshot accepts 0.169.0-alpha.11.1 while retaining all surrounding refusals and bounded streaming.
  • The bilingual required-work-context reference records the exact qualified writer and conservative fallback.
  • test_codex_context_resume.py exercises exact and neighboring writer versions, changed continuity, damaged history, complete output and restart behavior.

规格依据为 docs/reference/required-work-context.md revision 29c312949b6c26e3b27288234881eea2958ead94。

对主干的风险 / Risks

最强反例是一个版本字符串相近但语义不兼容的 writer 被误接受。精确 allowlist 和 unchanged identity/byte gates 阻止该情况;146 项 replay tests 与 quick premerge 通过,仅保留已有 maintainability advisory。

我的整体评价 / Verdict

该改动只扩展明确受支持的 writer,不改变回退原则,也不宣称模型输入保留已得到证明。它减少有效重放的无谓全量投递,同时把生产 App 采用留给后续验收。

English verdict: APPROVE for 685cbea. The exact Codex 0.169 paginated writer now replays under the unchanged complete-Turn and conservative-fallback contract.

@huangruiteng
huangruiteng merged commit 1e2c7a4 into main Nov 19, 2026
17 of 21 checks passed
@huangruiteng
huangruiteng deleted the codex/c1-native-replay-writer-0169-20261119 branch November 19, 2026 06:56
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.

2 participants