[sunrioa]

Summary

Refs #6110.

Addresses the three additional display cases reported in the issue’s follow-up:

• Show selected excerpts as Maka’s existing quote chips, with the user’s comments in the message body.
• Show historical questions as quote chips and their answers separately, without restoring an interactive selector.
• Hide recognized, automatically supplied browser-context envelopes from the displayed message while retaining the user’s request.

The importer preserves the original message text for model replay and produces a separate displayText. Quote chips are projected only in the UI; no new components, styles, dependencies, or persisted quote references are introduced. Unrecognized or malformed envelopes remain unchanged.

Recall search, session-title generation input, and CLI transcripts use the cleaned display text. URLs present only in hidden browser context are therefore excluded from Recall’s searchable message text. The existing Edit-and-resend action also rejects these transformed imported messages (displayText !== text) to avoid discarding their original context.

Image attachment import is outside this PR’s scope. Images are not migrated, and their file metadata remains visible. #6110 should remain open until the image-attachment problem is addressed.

Verification

• Build, typecheck, lint, formatting, ASF headers, and UI/Desktop Knip checks passed.
• Full Storage suite: 1,648 passed, 8 platform-specific skips.
• Full UI suite: 791 passed.
• Targeted import, model-history, and revision tests: 113 passed.
• Regression controls reproduced the missing display projections without the corresponding changes.
• Read-only checks against actual Codex records confirmed that original message content and model replay remain unchanged.
• I manually checked all three display cases in the local Desktop build.
• Full Electron E2E, Windows validation, and packaging were not run.

Before / after screenshots · 前后对比截图

Historical question reply / 历史问答回复

Before / 修改前

After / 修改后

Selected-text annotation / 文本批注

Before / 修改前

After / 修改后

Automatically supplied browser context / 自动浏览器上下文

Before / 修改前

After / 修改后

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex assisted with investigation, implementation, regression tests, code review, and drafting this description. Manual Desktop verification was performed by me.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

简体中文

概述

关联 #6110。

处理该 issue 跟进评论中补充的三类显示问题：

• 使用 Maka 现有引用标签展示选中的原文，将用户评论单独放在消息正文中。
• 使用引用标签展示历史问题，将回答单独显示，不恢复可操作的选择器。
• 在显示层隐藏已识别的自动浏览器上下文包装，保留用户的实际请求。

导入器保留原始消息文字供模型回放使用，并生成独立的 displayText。引用标签仅在 UI 投影中生成，不新增组件、样式、依赖，也不向存储消息中添加引用字段。无法识别或格式异常的包装保持原样。

Recall 检索、会话标题生成的输入和 CLI 对话展示均使用清洗后的显示文字。因此，仅存在于已隐藏浏览器上下文中的 URL 不再计入 Recall 的消息检索文本。现有“编辑并重发”操作也会拒绝这些已转换的导入消息（displayText !== text），以免丢弃其原始上下文。

图片附件导入不在本 PR 范围内。本次没有迁移图片，相关文件元数据仍然保留显示。图片附件问题解决前，#6110 应继续保持开启。

验证

• 构建、类型检查、lint、格式检查、ASF 许可证头检查，以及 UI/Desktop Knip 检查均通过。
• 完整 Storage 测试：1,648 项通过，8 项因平台条件跳过。
• 完整 UI 测试：791 项通过。
• 导入、模型历史和消息修订相关定向测试：113 项通过。
• 对照验证确认，移除对应修改后，回归测试能够复现显示投影缺失。
• 对实际 Codex 记录进行只读核对，确认原始消息内容和模型回放保持不变。
• 已在本地 Desktop 开发版中亲自确认三类显示均正常。
• 未运行完整 Electron E2E、Windows 验证和打包测试。

三组前后对比截图见上方验证部分，展开对应场景即可查看，标题和修改前后标识均包含中英对照。

AI 使用说明

• 生成式工具未作出实质性贡献
• 生成式工具作出了实质性贡献

工具及范围：OpenAI Codex 协助进行了调查、实现、回归测试、代码审查和本描述起草。Desktop 手工验证由我完成。

检查清单

• 测试覆盖本次修改，且缺少对应修改时能够失败
• Lint、格式检查、类型检查及受影响测试均在本地通过

本 PR 是否改变行为？

• 是——已在上方概述中说明
• 否

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Reviewed: full first review of 1751c39..8dc7e2a (merge-base with main is the current main tip; 6 files, +388/-4). There are no prior reviews.

Summary. The Codex adapter now keeps text exactly as written and only adds a displayText when it recognises one of these envelopes: a whole-message <send_user_message_question_reply> JSON array; or the exact Response-annotations header, then an optional Files block, then an optional ambient-ui-state browser block, followed by \n\n## My request:. The UI then re-parses the raw text to build display-only quote chips, and it does so only when the displayText prefix exactly matches what the importer would have produced. Model replay is unchanged because text is untouched. Edit still reads the stored message and refuses transformed text, so the display-only quotes never reach a resend.

False positives and information loss. I found no P0-P2 defects. The checks are anchored and conservative:

• Every envelope has to start the message (after leading newlines), and the reply envelope has to be the entire message.
• Prose before an envelope, envelopes inside code fences, an unknown source=, unknown context blocks, invalid JSON, empty arrays, non-string comments, and missing or null answers all fall back to the raw text. The tests cover each of these.
• A user body that contains a literal ## My request: heading or leading indentation is kept.
• File/image metadata stays visible.

I also checked a few cases outside the test suite. Pretty-printed annotation JSON works. A ## My request: marker inside a quoted excerpt does not split the message, because JSON escapes the newlines. A trailing CRLF on the body is handled. Text before a reply envelope makes the importer return no displayText. All behaved as intended.

Findings (all P3, none blocking):

1. P3 - The browser block is only checked at its edges (codex-session-adapter.ts:1624). It is recognised by startsWith(header) && endsWith('</in-app-browser-context>') and then the whole remaining context is dropped. Anything between the first opening tag and the last closing tag is hidden from display, for example a different envelope sitting between two browser blocks. This is display-only, since text keeps everything, and it is unlikely with current Codex output. Matching exactly one block (the first closing tag must be the end of the context) would be tighter.

2. P3 - Quote/comment pairing is lost in the UI when there are several annotations (materialize.ts:1179). The chips show the excerpts and the bubble shows the comments joined in order, but chips do not show comments. With [A: c1, B: (none), C: c3] the bubble reads c1\n\nc3, and the reader cannot tell which comment belongs to which excerpt. The stored displayText does keep the pairing, and this matches how native commented quotes render, so it is a UX note only.

3. P3 - Side effects of the new displayText, worth stating in the PR:

   • recallSearchableText, session-title derivation and the CLI transcript all read displayText ?? text, so the hidden browser URL is no longer searchable via recall for imported sessions.
   • Imported messages that now have a displayText can no longer be edited. They hit revisionTransformedTextUnsupported, whereas before the raw envelope was editable.

   Both look acceptable, but they are behaviour changes.

4. P3 - The envelope grammar is now parsed in two places (the storage importer and UI projectImportedCodexQuotes), and the UI's annotation regex accepts any one-line header. The exact-prefix check against displayText keeps the two consistent today, so this is a maintenance note. Not persisting quotes is the right call, because QuoteRef.comment is model-facing and would duplicate context on replay.

Tests. I reverted only codex-session-adapter.ts and materialize.ts to the merge-base and kept the new tests:

• The storage codex-session-adapter display test fails for both formats (legacy and item_completed).
• All five materialize imported Codex context display tests fail.

With the fix, storage gives 51/51 and UI gives 64/64 on the targeted files. The chat-turn-answer-identity change only strengthens an assertion and passes either way, as expected.

Protocol/epoch: no protocol files are touched.

CI and mergeability. CI on 8dc7e2a: label and test both pass. The PR is MERGEABLE and merges cleanly into current main. Its state is BLOCKED only because a review is still required.

[Astro-Han]

P3: This drops everything from the first browser opening header to the last </in-app-browser-context>. A different envelope between two browser blocks would be hidden from display too (text is still intact). Consider requiring that the first closing tag ends the context, or iterating over whole blocks.

[Astro-Han]

P3: With several annotations, the comments are joined into the bubble while the chips show only the excerpts, so the reader loses which comment goes with which excerpt (e.g. [A: c1, B: none, C: c3] renders c1\n\nc3). This is display-only and matches how native quotes render, so just noting it.

[Astro-Han]

Automated review notice: This comment was posted by an automated review agent. It is not an independent human review and does not replace one.

Incremental review: 8dc7e2a2ca4c8f18a32e03cffc05ae2e3a270c78..fbef74609f167015800e6bd94f0dce1cccdb7c5b, one commit (fbef74609) touching packages/storage/src/codex-session-adapter.ts and its test.

The earlier P3 about the browser block being matched only at its edges is addressed: the block is now hidden only when the first </in-app-browser-context> closing tag is the final line, so a context that contains a second block or other content after an earlier close is left visible. The new test case (two browser blocks around an unknown context) covers exactly that and keeps the raw text. Text inside a single well-formed block is still treated as browser state, which is the intended behaviour. The other three P3 notes from the previous review are unchanged and remain non-blocking.

No new findings. No protocol files touched (epoch unaffected). The PR is MERGEABLE; current-head CI is still running, so this is not a claim of green CI or an approval.

[sunrioa]

Thanks for the review. Following up on the four P3 notes:

1. Browser-context boundary: fixed in fbef74609. Only a single complete browser block is hidden. Content after an earlier closing tag remains visible, with regression coverage for both Codex record formats.

2. Multiple quote/comment pairs: I’ve kept Maka’s existing quote UI. Each quote retains its corresponding comment in the hover card, so the pairing is preserved, although it is less obvious when viewing several collapsed chips alongside the combined message body. I agree with the UX concern; improving that presentation would also affect native quotes, so I’m leaving it outside this importer fix.

3. displayText effects: the PR description now documents the effects on Recall search, title-generation input, CLI transcripts, and the existing Edit-and-resend restriction. The original message text remains unchanged.

4. Parsing in storage and UI: I’ve left this split unchanged for now. Storage produces the cleaned display text; the UI derives display-only quotes and checks them against the importer’s exact display prefix. Persisting these quotes would risk duplicating model-facing context. A shared-parser refactor could be considered separately, but this PR does not introduce one.

简体中文

感谢审查。针对四项 P3，说明一下处理结果：

1. 浏览器上下文边界： 已在 fbef74609 中修复。现在只隐藏单个完整浏览器上下文块；如果较早的闭合标签后还有内容，则保留显示。两种 Codex 记录格式均补充了回归覆盖。

2. 多组引用与评论的对应关系： 本次保留 Maka 现有引用界面。每个引用的悬浮卡片仍包含对应评论，所以对应关系没有丢失，但同时查看多个折叠引用标签和合并后的消息正文时，确实不够直观。我认可这项体验问题；改善这部分展示也会影响原生引用，因此本次导入器修复暂不调整。

3. displayText 的影响： 已在 PR 描述中说明其对 Recall 检索、标题生成输入、CLI 对话展示，以及现有“编辑并重发”限制的影响。原始消息文字保持不变。

4. 存储层和 UI 分别解析： 暂时保留目前的分工。存储层生成清洗后的显示文字；UI 生成仅用于展示的引用，并核对是否与导入器产生的显示前缀完全一致。如果把这些引用写入存储，可能重复加入面向模型的上下文。共享解析器可以另行考虑，本 PR 暂不引入这项重构。

[Astro-Han]

Thanks @sunrioa for the contribution, and for working through the review feedback. Merging now.

[Astro-Han]

Approved at @Astro-Han's explicit request: no blocking findings in the automated review of this head, and CI is green.