Skip to content

test(benchmark): qualify warm receivers with progressive model delivery - #8052

Merged
huangruiteng merged 3 commits into
mainfrom
codex/qualify-receiver-model-transport-1104
Nov 5, 2026
Merged

huangruiteng merged 3 commits into
mainfrom
codex/qualify-receiver-model-transport-1104

Conversation

@loopx-agent

@loopx-agent loopx-agent commented Nov 4, 2026 •

Copy link
Copy Markdown
Collaborator

Problem and change

The joint checkpoint qualification on Linux failed before the first restored model turn when progressive reading returned a capture reference. The fixture indexed that reference as an inline decision. Request the machine representation for its guard assertions, then pass that same decision through the receiver's configured native delivery path. The change is scoped to the qualification fixture.

Verification

On the integration branch with progressive context, the old context-on receiver failed with a missing interaction contract. With this change a fresh synthetic source, an ordinary receiver and a Stride receiver passed using the real native binary in separate Linux containers. On integration commit 4e1f954aea77, both context-on and context-off completed the full chain: fresh source, ordinary and Stride receivers, automatic recapture, another receiver doing work and the ordinary Goal cross-check. The core cross-check also requests a machine-readable decision explicitly. The model endpoint in this qualification is synthetic.

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>

@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-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max.

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

未发现阻塞本次改动的代码问题。必需质量检查和维护者合并资格仍未解除,详见风险。

动机 / Motivation

验证普通与 Stride 接收器续跑流程的贡献者。渐进读取开启时,原测试把供模型导航的引用当成完整决定,尚未执行恢复工作就因缺少字段失败;改后机器先读取原决定,再把同一决定交给原模型交付路径。机器断言与模型交付共用同一原始决定;本地 File/SQLite 的开关对照及普通 Goal 接续验证通过,Linux native joint 整链资格仍未独立重跑。不改生产准入、模型上下文默认值、原断言、冻结实验条件,也不宣称已完成 Linux、真实模型或长期收益资格。

Current localmachine/originalreader boundaryqualified; caller-owned freshLinux/native ordinary/Stride/recapture qualification remains unverified and is not replaced by syntheticlocaltrace.

改动思路 / Approach

复用既有CLI决定与模型投影owner,在测试消费者局部区分机器表示和模型交付;无第二Python策略owner。 两测试文件14增19删;只修资格fixture的机器读取边界,生产loopx整树及原preflight helper Git对象相同。

我对比了既有 read_cli_policy_payload 与原模型交付 helper:前者可解码 host 引用,但此 fixture 要先检查机器字段,再由原 run_native_work 把同一个决定交给模型。局部 raw 参数复用已有 CLI,避免改环境或再造 decoder。正向流程是 raw guard → 原始文件/SHA/Turn → 原 preflight → 实际投影和 reader;on/off 的模型选择留在原 owner。

具体改动 / Changes

审查精确 head 18e9d5ab97bd5073eb9dec3c88aa4e11010ad9b6,实际三点比较基线 d29ec8bf2a105aac24629bcbf3dfb7c97aea51fe。两测试文件14增19删;只修资格fixture的机器读取边界,生产loopx整树及原preflight helper Git对象相同。

关键代码讲解

  • receiver_guard:Raw machine guard at five existing call sites. Explicitoff is argument-local; env remains on.
  • run_native_work:Preserve model-facing current context. Original SHA/Turn handed to configured on/off branch.
  • _ordinary_review_and_work:Ordinary nonbenchmark cross-check. No weakening original acceptance.

规范依据为 benchmark/README.md,固定版本 d29ec8bf2a105aac24629bcbf3dfb7c97aea51fe,按该变更之前已接受的正文判断。

--context-delivery on:机器断言不能替代模型读取;同一原决定的 on preflight 和完整 reader 已独立执行,保持值、类型与 Turn。--source-revision:对照固定 base/head 及实际导入来源,不把旧源结果算成本 head 资格。--reuse-point:本地边界已验证,fresh Linux ordinary/Stride/recapture 整链仍为 UNVERIFIED,保留作者集成结果的诊断身份。

对主干的风险 / Risks

最强反例是把机器 off 意外扩散到模型,使开启资格名存实亡。实际 File/SQLite×off/on 各四个匹配案例中,base 的 on 确实返回无机器字段的导航,head 机器断言全部通过;head on 继续执行未改的真实 preflight、context-project 和每个 reader 页。原文件身份和 SHA/Turn 绑定保留,env on 未变,每例一条 heartbeat、零 spend。reader 返回 capture 的紧凑 JSON,严格保留数值类型、字段和文本;仅规范化 JSON 空白/转义,没有声称等于原 pretty-print 文件的字节。

完整普通 adoption 模块 base/head 均 16 PASS,145 条原断言 AST 相同,Ruff PASS。标准 premerge 的 5 个直接检查与 6 个选定检查 PASS;整体仍 FAIL,保留 benchmark_sensitive 维护者 hold。必需 17-source mypy 在 base/head 同一未改模块 :110 报 no-any-return;完整 208B 输出相同,相关源码/配置、生产整树和生成环境锁相同。此归因只支持代码审查判断,不豁免 required FAIL 或严格 CQR FAIL。

未独立启动 Linux native joint 整链,也未用本地 synthetic journal 证明 native fork、模型采用或效果增益;作者集成资格仍待对应 owner 的完整条件复验。早期本地探针的导入/环境声明及过强 JSON 空白 oracle 失败保留,最终只修私有探针,不修改冻结145断言或产品。

我的整体评价 / Verdict

这是 justified_increment:修复可复现的 fixture 表示边界,long_horizon / user_experience 均 preserved;不把它升级为完整父目标或 Linux/model 资格。复用现有 owner、两文件局部 helper 的成本适当,无需要另加状态/兼容层的重构。全差异仅资格fixture;已检查receiver_guard→quota should-run(raw machine)→run_native_work→preflight_context_instruction→quota context-project→actual context-read pages,无页面或布局变化,视觉评审不适用。 未查询远端CI;未验证真实模型采纳、installed App、完整长期成本。Linux fresh native ordinary/Stride/recapture整链未独立重跑。 严格质量门、required mypy 与维护者 hold 留给合并资格处理;当前结论只针对上述不变 head 的代码。

English verdict: APPROVE this exact-head fixture change; required quality and maintainer merge holds remain.

@huangruiteng
huangruiteng merged commit d145cd9 into main Nov 5, 2026
4 checks passed
@huangruiteng
huangruiteng deleted the codex/qualify-receiver-model-transport-1104 branch November 5, 2026 03:24
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