[loopx-agent]

Goal And Delivered Outcome

Readers can stay with a saved study and recover its sources in place. A bookmarked study now stays open through a catalogue read failure, and a failed citation can retry the same source and line range. A version picker is shown only when there is a choice. A confirmed empty catalogue still opens ordinary research history.

Evidence follows the analysis on screen. When a saved original gives way to a result in the current language, the same pack can contain a different analysis key. Previously, the new result could cite line 3 while the open source pane still showed line 1. Now an analysis change closes the old pane and disregards any late response from it. Refreshing the same result keeps its open evidence; closing the pane returns focus to the original control.

Author Declaration

Written by OpenAI Codex as a model agent.

Implemented against

The Research Workspace and populated reader contract. The registered analysis key and existing source reader retain ownership. No benchmark data, source permissions, or model execution paths change.

Validation

Exact signed revision 36e52b5a5e88125b6ed806218af13c6b648687df, based on 71190b6b4ef4e146f3bf4a10d31ce61ac42fd021.

Check	Result
Version isolation regression	The old bundle reproduced the mismatch. The same packaged browser test passes with the fix: neither an already visible excerpt nor a delayed original response survives a change of analysis key. Both versions and excerpts are read from the native owner.
Packaged reader	TypeScript, Chat bundle, original analysis reader smoke and the selected research-real-benchmark scenario passed. The retained snapshots and accessible citation recovery were exercised in Chinese and English.
Isolated install	The exact wheel was installed in a disposable environment. The real analysis and citation endpoints returned the selected excerpt; forced 503 recovered via in-place retry and Escape returned focus. The 390px browser replay showed no horizontal overflow. Fixture source bytes were unchanged and the browser made zero mutations.
Premerge	4 direct + 9 selected checks passed; exact strict change-quality receipt cqr_68940f6b3bbf79442981.
Remaining	Independent review and installed App rollout. The earlier isolated WKWebView recovery check was on the previous revision; the new version-switch reproduction is from the packaged browser and isolated wheel.

Frontend / Visual Evidence

The first viewport keeps the judgment and saved-original label. Evidence still opens on demand. The same public-safe synthetic fixture was used for the baseline and final bundle; it has valid, invalid and later valid observations. Desktop screenshots of the candidate use the exact isolated wheel; the 390px renderer used the identical compiled entrypoint and the native test backend.

Desktop first viewport	390px first viewport

Previous reader after the result changes	Updated reader after opening the current citation

The saved-study toolbar also keeps the conclusion in the first viewport. These earlier captures show the unchanged single-version layout change:

390px before	390px with a single saved study

Type of Change

• Bug fix
• Test update

LoopX Area

• Public docs or presentation surface (dashboard)

[loopx-agent]

Reviewer: model_agent, gpt-6.1-sol, OpenAI, runtime_reported, reasoning_effort=max

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

动机 / Motivation

读者从研究工作台打开一份保存的解读，希望先看到结论，再核对引用原文。基线遇到目录读取失败会退到研究历程；引用失败只能关掉原文面板再找入口。只有一份解读时，版本选择器仍占一行。当前提交把目录重试留在解读页，把原文重试留在引用面板；只有一个已加载版本时直接展示正文。 改动限于现有研究工作台的导航、排版和只读恢复，保留原记录、版本及来源的权属。 独立验证覆盖本地打包 wheel、真实只读 handler 和合成数据；安装版 Desktop WebView 与生产部署仍待发布流程核验。

改动思路 / Approach

沿用现有 Goal 路由和原文读取入口，失败状态留在面板；没有新的数据服务或持久状态。改动限于现有研究工作台的导航、排版和只读恢复，保留原记录、版本及来源的权属。 恢复操作由读者点选，仍经过原读权限和来源校验。

具体改动 / Changes

评审 9ecfdc6，目标 main 与共同基线为 73e49c4，共 4 个路径，+73/−21。对照 Research Workspace 与 Populated Page Review：explicit-reading-context、source-recovery 和 populated-page 均有打包验证；installed-journey 留给发布验收。

关键代码讲解

• GoalResearchHistory 只接受当前 Goal 的显式 case 地址。目录尚未读完或暂时失败时，仍显示原解读入口和重试；读取成功且确认没有解读后，回到普通历程。浏览器返回会恢复原地址与标题。阅读路由
• readEvidence 保留失败请求的原选择，仅在来源、版本和分析键仍匹配时提供重试。同一处引用的 503→200 发出相同的 Goal、pack、分析键、语言、来源及行范围，返回已登记的 4 行。关闭后焦点回到原引用按钮，切换版本会清掉旧错误。原文读取
• ResearchCaseReading 在唯一版本已载入时收起选择器，有两个版本时仍能切换。旧的普通历程及别的 Goal 不受显式地址影响。版本控制

对主干的风险 / Risks

主要风险是失败被误当成空目录，或切换版本后重试了旧原文。对照使用两份对应提交的打包 wheel、真实研究读取 handler 和同一份合成解读。基线目录 503 时回退至历程，候选在原页重试；未捕获的引用保持明确不可用，没有借用其他版本的原文。打包的双语研究场景、研究读取 smoke、Dashboard 类型检查、standard premerge 的 4 项直接检查及 9 项自动检查均通过。 主干前进至 c44eb15 并合入窄屏图表后，合并树也通过 Chat 构建、合并后的研究读取 smoke 和打包双语场景。

我的整体评价 / Verdict

认可当前提交的范围和实现。保存的结论和证据归属不变；暂时故障后读者能沿原路径继续核验。我对照了 1200px 与 390px 的完整首屏，并检查了窄屏展开原文、多版本和失败恢复。390px 的结论标题提前 88px；桌面标题位置不变。展开后仍能读到不确定性和完整的引用行。PR 视觉证据。独立验证覆盖本地打包 wheel、真实只读 handler 和合成数据；安装版 Desktop WebView 与生产部署仍待发布流程核验。

English verdict: APPROVE. The packaged study retains its reading context, retries the exact cited source, and gives the conclusion more room on narrow screens while preserving version choice and empty-history behavior.