[loopx-agent]

Problem and result

新生成的研究说明需要按界面语言直接解释样本、重复测量和数量，减少生硬术语、编码说明和重复英文。新的默认生成指导区分样本组与执行队列、重复测量与独立输入、相对百分比与百分点；保留原文、否定、未知和全部数量约束。沿用现有分析任务，无需增加译文编辑器或修订流程。

New default requests receive clearer localized prose guidance. This proposal also includes the runtime cleanup fix from #7460 so the complete generation journey can be qualified against one combined revision.

Implementation and compatibility

• New preparations with an omitted display encoding select v4 and persist that choice before analysis. Explicit encodings, configurations and frozen requests retain their original prompts, keys, schemas and artifacts.
• Instruction assembly stays in the existing TypeScript assessment owner. Every prior fidelity clause remains; literal checking and original-text fallback are unchanged. Generation guidance does not verify model meaning or comprehension.
• A direct child may exit between the first nonblocking reap and process-group observation. Cleanup now reuses the existing reap helper after observing exit. Authentication, drain, deadlines, non-child zombies and ordinary serving handoff keep their existing behavior. No extra signal, wait loop or authority is introduced.
• fix(runtime): reap children after shutdown group observation #7460 remains an independently reviewable cleanup proposal. This branch contains its exact commit 20fc9706b; it can be reviewed as part of this combination or merged first. The PR base remains main, integrated at 0c75dacf0.

Validation

Final head: 05cd41c46f684e361fd82b29dc672b55d64d4e89. The earlier combined revision 8961e83c passed the full owning suites; all seven proposed file blobs are identical in the final head. The new main dependency and packaged journey were then revalidated.

Check	Result
Typed assessment owner	35 tests passed at 8961e83c, unchanged TypeScript tree; default/explicit/frozen behavior and corrupted numeric fallback
Source integration	142 passed at 8961e83c; final 05cd41c additionally passed 84 research/source HTTP and configuration tests after integrating main’s new configuration readback behavior
Frozen compatibility	19 complete preparations equal the retained baseline, including an original saved request
Isolated built wheel	CLI/default storage identity, language isolation, guidance reaching one synthetic provider, reuse without another process and source refusal/recovery passed; 73 product modules loaded from the wheel
Packaged runtime	Two real-process tests passed, including the deterministic exit race and Node replacement/shutdown
Static and public boundary	TypeScript, Ruff, semantic advisory, bilingual diff, DCO and all seven outgoing files passed
Exact-scope quality / standard premerge	Passed: exact-scope quality receipt valid; 6 direct and 11 selected standard premerge checks passed, no manual holds. Merge authority is separate.

The guarded final-head package build used the normal frontend build and succeeded. An initial premerge had passing automatic checks but a stale quality receipt after the shared target ref advanced; the new base was integrated and qualification was repeated against its captured commit. That initial gate remains a recorded failure. No real model was called. The synthetic provider is an integration fixture, not translation-quality evidence. Added generation instructions consume context; actual token cost, comprehension and installed-App adoption remain unverified. Native Windows/Linux were not run; PostgreSQL authority behavior is unchanged.

Historical failure and remaining scope

An earlier HTTP run failed an immediate runtime-PID cleanup assertion; immutable base also reproduced it. The old head and its failed quality receipt remain historical failures. #7460 demonstrates and repairs a specific real-child race, and the current combined source/package passes. Old PIDs were not retained, so this does not claim every historical intermittent failure had a uniquely proven cause or retroactively pass the earlier receipt.

Actual model-output semantics and installed use still need their own evidence. Browsing and switching language never regenerate analysis. This author proposal grants no merge, installation, paid execution or benchmark authority.

Author and visual boundary

Written by: model_agent — OpenAI Codex, GPT-6 family. Contract: docs/reference/trajectory-assessment.md and the existing research-workspace design rules. Reviewed the complete seven-file diff; reused existing owners and kept the change reversible. No paired RFC changed.

UI impact: none in this slice. No component, layout, navigation or rendered candidate changed, so viewport screenshots and a preview are not applicable. No private source, raw trace, local path or screenshot is published. Whole-page readability is not certified by these checks.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · effort=max

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

本精确版本未发现阻塞问题，同意作为有边界的默认生成指导与清理修复交给维护者评审。这是作者自评，不是独立验收，也不表示实际译文质量已通过。

动机 / Motivation

阅读研究轨迹并明确发起分析的用户会遇到样本、重复测量与百分比表达含混的问题。此前新请求采用 v3；本提案让新请求采用更自然的 v4 指导，旧请求与原文保持不变。

已验证新默认被保存并送达同一分析子进程，以及旧版准备和拒绝回退保持；尚未验证真实模型语言质量。本次不增加译文编辑器、修订流程、自动重生成或新的模型执行权限。真实 v4 输出语义、正式 App 采用和完整实际输入成本仍待验证。

同时核对了完整分析链路所需的进程清理修复：直接子进程可能在第一次非阻塞回收后退出，原生进程组观测已确认退出，但父进程仍未收取其退出状态。不能把这类“仍可观察到的进程”靠延长超时或放宽断言消掉。

改动思路 / Approach

生成策略保留在既有 TypeScript 分析 owner；退出回收复用既有进程 helper，不建立第二份执行或翻译状态。当前边界是新默认指导及其完整进程交付的前置修复；历史译文、模型语义认证、界面改版与合并安装均不在本次交付结论内。

普通用户仍在研究页明确发起既有分析，语言和已知输入沿原路径传递，没有新增必填信息、编辑器或确认轮次。缓存、原文、取消、重试和失败读回继续由原任务负责。新指导实际送达子进程已经验证，子进程使用合成 provider；它不能证明真实模型会正确理解。

输入/状态	审阅后的行为
新建默认展示请求、未指定编码	保存 v4，再通过既有任务执行
已存请求、显式旧编码、普通读取	保持原提示、键、结构与产物，不升级或执行
已退出的直接子进程	进程组确认退出后再次非阻塞回收，再返回停止结果

具体改动 / Changes

本次检查的是完整七文件差分（+152/-69），包括三个生产文件、一个合同文档和三个既有测试模块。主干基线为 0c75dacf053b6caf41284f86b41641c2be429ae0，精确 head 为 05cd41c46f684e361fd82b29dc672b55d64d4e89。没有新增生产模块、CLI、调度器或模型调用入口。

规范依据：docs/reference/trajectory-assessment.md，规范修订 0c75dacf053b6caf41284f86b41641c2be429ae0；按修改前已接受的“Whole-history presentation language”合同逐项核对：

• display-language：完整快照叙事字段随界面语言解释；指导仍在既有分析 owner 内。
• frozen-identity：19 份完整准备结果与基线相等，包含提示、键、结构和输入；不是只比较编码或哈希字段。
• explicit-job：隔离 wheel 中正常准备、保存、明确执行、复用与来源拒绝/恢复走原任务；复用没有再启动子进程。
• original-fallback：v3/v4 的符号、精度或百分号被破坏时保留完整原文，指标判断不被改写；词法通过仍不证明语义。

关键代码讲解

1. parseTrajectoryInput 只在新写入且未指定编码时选择 v4；显式旧值保留，非法值仍拒绝。Python 的准备和登记入口复用同一 TypeScript 决策。
2. displayGenerationInstructions 把原有完整含义、否定、失败、未知、上下文来源和数值限制条款原样集中。新 v4 再补自然术语、相对百分比/百分点、输入/重复测量的区别；同一字符串用于提示和分析身份，避免两处组装漂移。这是有界的同域重构。
3. _stop_effect_runtime 在 POSIX 进程组观测成功后复用 _reap_exited_runtime_child；它只以 WNOHANG 收取已退出的直接子进程，不加信号、等待循环或权限。普通 serving 交接和无法证实退出时的 pending 路径保持。

对主干的风险 / Risks

最强的剩余反例是：模型保留了全部原始标记，却把数量归到错误对象，或者仍写出难读的中文。现有机械校验不能排除它，所以本提案只交付生成指导，未宣称语义认证。新增指导还会占用输入上下文，实际 token 成本和净效率收益未知。

PASS：复用并逐项校验既有验证产物及源码身份；初次组合有 35 项 TypeScript、142 项来源/HTTP/runtime 检查，最终 head 在合入基线后另有 84 项来源/配置检查。隔离构建 wheel 加载 73 个产品模块、同任务合成生成/复用/语言隔离/来源拒绝恢复通过，两个真实进程检查通过；标准预检 6 项直接检查和 11 项选中检查通过。此次复核确认七个提案文件的 blob 与初始验证相同，31 份验证产物的记录哈希一致，随后主干变更没有改动相关分析或 runtime owner；因此未重复运行相同矩阵。没有查询或等待 CI。

反例使用真实 OS 子进程及原生进程组观测，固定其在第一次回收后退出：不可变基线会在“退出状态仍未回收”断言处失败，所含修复及最终 wheel 通过。认证请求部分被隔离以精确安排竞态；真实 Node 替换/关闭另有打包检查。不能据此把每次历史间歇性清理失败都归为同一原因。早期失败和基线前移造成的 stale quality receipt 保留为历史，未补标通过。后续主干不是此 head 的已验证组合，当前合并资格须单独检查。

语义与 CI 对齐

v4 扩展既有 translation_encoding 类型，不另建状态或 authority。旧版保留有实际持久化/重放依据；没有新增 prose 分类、领域化核心义务或暗中执行。数值完整性是已有强制校验，自然语言质量是生成指导，两者在文档和结果中明确区分。独立 proposal #7460 的准确清理 commit 已包含，未以不冲突代替行为比较。没有成对 RFC 变更；DCO 与公共边界检查保留。

我的整体评价 / Verdict

long_horizon=preserved：旧输入、结果、失败与复用继续稳定；精确修复了一个可复现的清理竞态，没有扩展执行权限。user_experience=preserved：原明确分析操作和原文读回保持，新增必需交互为零，不把指导变更冒充已经观察到的理解提升。机制范围和版本兼容成本与当前问题相称；没有理由引入译文修订产品。

七个变更文件未修改组件、布局、导航或已存展示产物；本次没有可归因的像素变化，不以已有截图认证新译文质量。真实 v4 译文整页阅读和正式 App 采用未验；无本次新增界面变化需要视觉验收。

仍需真实输出和正式采用证据才能判断完整语言体验。原生 Windows/Linux、实际新 PostgreSQL 运行和模型净成本没有在本次验证；此变更没有改写 PostgreSQL authority 路径。作者结论不代替独立验收、维护者授权、最新合并资格或安装读回。本次未合并、安装、重跑模型或启动 benchmark。

English verdict: APPROVE — 05cd41c46f684e361fd82b29dc672b55d64d4e89. No blocking finding for the bounded default-guidance and child-reap change; frozen-request parity and real process failure/recovery are covered. Actual model semantics and installed adoption remain unverified; merge authority is separate.