feat(cli): post the review body bilingually when the PR description is Chinese

[wenshao]

What this PR does

When the PR author writes Chinese, /review now posts bilingually. On GitHub — which renders <details> — the English body leads and the complete Chinese version rides collapsed under it, the same shape this repo's own PR descriptions use:

• Detection is the CLI's own. fetch-pr fetches body in the same gh pr view --json call it already makes, detects Han characters (\p{Script=Han}), and stamps prDescriptionHasHan into the plan report. compose-review/submit read the flag from the plan — the caller cannot toggle the register of a certified body, and a local plan has no field, so terminal-only reviews are unchanged.
• The posted body renders from en/zh pairs end to end. Every deterministic fragment carries both languages: compose-review's clause templates (openers including the zero-certified warning, suggestion clauses, downgrade sentence, LGTM), describeChunkGap phrases ("the entire diff" / "整个 diff", counts, file lists), the coverage disclosures (reasons, role subjects via a new Brief.publicLabelZh, the path-free unread-brief reason) and the Step 4/5 gap texts including the combined same-shape sentence. Fragments with no deterministic translation — model-written findings, caller echoes, interpolated errors — ride verbatim in both halves. The footer stays outside the fold, once; a body where nothing translated publishes no empty fold.
• verificationGaps returns structural entries ({subject, reason, subjectZh, reasonZh}), which also removes compose-review's last recover-the-boundary-from-prose parse.
• Inline comments follow the same format, via SKILL.md: the English finding first — severity marker and ```suggestion block stay in the English half, since tooling (autofix's **[Suggestion]** filter) keys on them — then the full Chinese translation collapsed in <details><summary>中文说明</summary>, then the model footer.

Why it's needed

qwen-code's contributors largely write PR descriptions in Chinese, and the review bot answered in English only. The body is composed deterministically by compose-review (the model cannot supply it — submit recomputes and refuses a payload carrying body), so serving Chinese readers requires the composer itself to render both languages from the same structural data; a model-side "translation" of a certified body is exactly the tampering surface the composer exists to close.

A zero-certified run on a Chinese-description PR now posts:

⚠️ This run could not certify that any of this diff was reviewed. Suggestions are inline. Not reviewed: 11 of the diff's 49 sections — launched with a prompt that is not the one the CLI built. …

<details><summary>中文说明</summary>

⚠️ 本次运行无法证明这个 diff 的任何部分经过了审查。 建议见行内评论。 未审查：diff 49 个片段中的 11 个——启动时使用的 prompt 不是 CLI 构建的那一份。 …

</details>

— qwen3.7-max via Qwen Code /review

Reviewer Test Plan

How to verify

1. Compose against a plan stamped prDescriptionHasHan: true: the body carries the English clauses first, the complete Chinese version inside <details><summary>中文说明</summary>, and one footer after </details>. Without the flag (or on any local plan), bodies are byte-identical to before — every pre-existing test passes unmodified on the English path.
2. Check the translated pairs render from the same structural data: role gaps say "the whole-diff test-coverage check" / "全 diff 测试覆盖检查", chunk gaps say "the entire diff" / "整个 diff", the Step 4/5 merged sentence says "verification and reverse audit" / "验证与反向审计". Caller-authored text (e.g. a cannotTellCriticals entry) appears verbatim once per half.
3. npx vitest run src/commands/review/ from packages/cli — 793 tests pass (5 new bilingual tests; the describeChunkGap and verificationGaps assertions now pin the Chinese strings too). eslint --max-warnings 0 passes on the changed files; tsc -p packages/cli reports no errors under commands/review.
4. For inline comments, the instruction lands in SKILL.md Step 7 (and the Step 1 field list) — model-authored text, so it is prompt-level by nature.

Evidence (Before & After)

Before: a Chinese-description PR received an English-only review body and English-only inline comments.

After: the body above — English first, complete Chinese collapsed, footer outside the fold. No terminal/UI change otherwise; screenshots N/A (the rendered <details> is standard GitHub markdown).

Tested on

OS	Status
🍏 macOS	⚠️ not tested
🪟 Windows	⚠️ not tested
🐧 Linux	✅ tested

Risk & Scope

• Main risk or tradeoff: the en/zh string pairs must be maintained together; they live side by side at every site (a Bi pair or …Zh field beside its English twin) so an edit that touches one cannot miss the other silently, and untranslated fragments degrade to verbatim English rather than dropping content. Han detection also fires on Japanese kanji descriptions — rare here, and the cost is an extra collapsed translation, not a wrong verdict.
• Not validated / out of scope: no re-verification with a live posted PR review (rendering verified through composed bodies and unit tests); machine translation of model-written finding text (left verbatim by design — the SKILL instruction covers it at authoring time).
• Breaking changes / migration notes: VerificationReport.gaps changes shape from string[] to structural entries (internal API; both consumers updated in this PR). Brief gains a required publicLabelZh. Plans gain an optional prDescriptionHasHan; old plans render English-only.

中文说明

本 PR 做了什么

当 PR 作者用中文写描述时，/review 现在发布中英双语评论。在 GitHub 这类支持 <details> 折叠的平台上，英文正文在前，完整中文版本折叠在下方——与本仓库 PR 描述的惯例一致：

• 检测由 CLI 自己完成：fetch-pr 在既有的 gh pr view --json 调用里顺带取 body，用 \p{Script=Han} 检测汉字，把 prDescriptionHasHan 盖进 plan 报告。compose-review/submit 从 plan 读取该标志——调用方无法切换认证 body 的语域；本地 plan 无此字段，纯终端 review 不受影响。
• 发布的 body 端到端用中英文对渲染：所有确定性片段都携带双语——compose-review 的句式模板（含零认证警告的开场白、建议句、降级句、LGTM）、describeChunkGap 短语（"the entire diff"/"整个 diff"、计数、文件列表）、coverage 披露（reason、经新增 Brief.publicLabelZh 的 role 主语、去路径的 unread-brief reason）以及 Step 4/5 gap 文案（含同形态合并句）。无确定性翻译的片段——模型写的发现、caller 回显、报错插值——在两个半区原样出现。footer 在折叠块外只出现一次；没有任何翻译内容的 body 不会发布空折叠块。
• verificationGaps 返回结构化条目（{subject, reason, subjectZh, reasonZh}），顺带消灭了 compose-review 里最后一处从渲染文本反解析边界的逻辑。
• inline 评论走同样的格式（经 SKILL.md）：英文发现在前——严重级别标记和 ```suggestion 块只留在英文半区，因为工具链（autofix 的 **[Suggestion]** 过滤）以它们为键——然后是折叠在 <details><summary>中文说明</summary> 里的完整中文翻译，最后是模型 footer。

为什么需要它

qwen-code 的贡献者大多用中文写 PR 描述，而 review bot 只用英文回复。body 由 compose-review 确定性生成（模型不能提供 body——submit 会重新计算并拒绝携带 body 的 payload），因此服务中文读者必须由 composer 自己从同一份结构化数据渲染双语；让模型"翻译"认证过的 body 恰恰是 composer 要堵住的篡改面。

Reviewer 测试计划

如何验证

1. 用盖了 prDescriptionHasHan: true 的 plan 走 compose：body 英文句在前，完整中文在 <details><summary>中文说明</summary> 内，</details> 之后只有一个 footer。没有该标志（或任何本地 plan）时，body 与之前逐字节一致——所有既有测试在英文路径上原样通过。
2. 检查翻译对来自同一结构化数据：role gap 为 "the whole-diff test-coverage check"/"全 diff 测试覆盖检查"，chunk gap 为 "the entire diff"/"整个 diff"，Step 4/5 合并句为 "verification and reverse audit"/"验证与反向审计"；caller 文本（如 cannotTellCriticals 条目）在每个半区各出现一次、原样。
3. 在 packages/cli 下 npx vitest run src/commands/review/——793 个测试通过（新增 5 个双语测试；describeChunkGap 与 verificationGaps 的断言现在把中文串也钉住）；改动文件通过 eslint --max-warnings 0；tsc -p packages/cli 在 commands/review 下无错误。
4. inline 评论的指令落在 SKILL.md Step 7（及 Step 1 字段清单）——模型手写文本，本质上是 prompt 级约束。

证据（Before & After）

Before：中文描述的 PR 收到纯英文 review body 和纯英文 inline 评论。

After：英文部分引用的 body——英文在前、完整中文折叠、footer 在折叠外。无其他终端/UI 变更；<details> 是标准 GitHub markdown，截图 N/A。

测试平台

OS	状态
🍏 macOS	⚠️ 未测试
🪟 Windows	⚠️ 未测试
🐧 Linux	✅ 已测试

风险与范围

• 主要风险或权衡：中英文串必须成对维护；它们在每个产生点并排存在（Bi 对或紧邻英文的 …Zh 字段），改动一处很难悄悄漏掉另一处；未翻译片段退化为原样英文而不是丢内容。Han 检测对日文汉字描述也会触发——在本仓库很罕见，代价只是多一个折叠翻译，不影响 verdict。
• 未验证或范围外：未在真实 PR 上发布验证（通过 composed body 与单测验证渲染）；模型写的发现文本不做机器翻译（设计上原样保留——SKILL 指令在撰写时约束双语）。
• Breaking changes / migration notes：VerificationReport.gaps 从 string[] 变为结构化条目（内部 API；两个消费方本 PR 内同步更新）；Brief 新增必填 publicLabelZh；plan 新增可选 prDescriptionHasHan，旧 plan 渲染纯英文。

[qwen-code-ci-bot]

Thanks for the PR!

Template looks good ✓ (minor: no ## Linked Issues section, but that's fine for a feature PR with no linked issue)

Problem: this is a feature, not a bug fix — no reproduction needed. The gap is observable: qwen-code's contributors largely write PR descriptions in Chinese, but the review bot posts English-only bodies and inline comments. The PR description makes the case clearly.

Direction: aligned. The repo already posts bilingual comments in triage (English first, Chinese in <details>), and PR descriptions follow the same pattern. Extending this to the review bot's deterministic body is a natural next step. The detection signal is the CLI's own (fetch-pr stamps prDescriptionHasHan from the PR body it already fetches) — the caller cannot toggle the register of a certified body, which is the right trust boundary.

Size: one core-path file touched — packages/core/src/skills/bundled/review/SKILL.md (+3 −1), a markdown instruction file, not architectural code. Production logic: 467 lines across 4 TypeScript files; test: 136 lines across 2 files. Under the 500-line threshold.

Approach: the scope feels right. Every deterministic fragment needs both languages, so the breadth is inherent to the goal. The verificationGaps change from string[] to structural {subject, reason, subjectZh, reasonZh} entries is a genuine improvement — it removes the last prose-reparsing boundary recovery in compose-review (the gap.indexOf(' — ') cut), which was a latent bug surface. No unrelated changes or drive-by refactors spotted.

Moving on to code review. 🔍

中文说明

感谢贡献！

模板完整 ✓（小问题：缺少 ## Linked Issues 章节，但功能 PR 没有关联 issue 是正常的）

问题：这是一个功能 PR，不是 bug 修复——不需要复现。差距是可观察的：qwen-code 的贡献者大多用中文写 PR 描述，但 review bot 只发布纯英文的 body 和行内评论。PR 描述清楚地说明了这一点。

方向：对齐。仓库已经在 triage 中发布双语评论（英文在前，中文在 <details> 内），PR 描述也遵循同样的格式。将此扩展到 review bot 的确定性 body 是自然的下一步。检测信号由 CLI 自己产生（fetch-pr 从已获取的 PR body 中盖入 prDescriptionHasHan）——调用方无法切换认证 body 的语域，这是正确的信任边界。

规模：触及一个核心路径文件——packages/core/src/skills/bundled/review/SKILL.md（+3 −1），是 markdown 指令文件，不是架构代码。生产逻辑：4 个 TypeScript 文件共 467 行；测试：2 个文件共 136 行。低于 500 行阈值。

方案：范围合理。每个确定性片段都需要双语，所以广度是目标本身决定的。verificationGaps 从 string[] 变为结构化 {subject, reason, subjectZh, reasonZh} 条目是一个真正的改进——它消除了 compose-review 中最后一处从渲染文本反解析边界的逻辑（gap.indexOf(' — ') 切割），那是一个潜在的 bug 面。未发现无关改动或顺手重构。

进入代码审查 🔍

— Qwen Code · qwen3.8-max-preview

_{Reviewed at 0f0f33787b76274faa198b127b61457c2a8af375 · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Code Review

Independent proposal: I would detect the author's language in fetch-pr (it already fetches PR metadata), stamp a flag into the plan, and render each deterministic fragment in compose-review bilingually using the repo's existing <details> pattern. For inline comments (model-written), I'd add SKILL.md instructions. The verificationGaps prose-reparsing (gap.indexOf(' — ')) would need to go — it can't survive bilingual text.

The PR's approach matches this exactly. No simpler path missed.

No critical blockers or AGENTS.md violations found. Specific observations:

• Trust boundary is right. prDescriptionHasHan is computed by fetch-pr from the PR body it already fetches — the caller cannot toggle it. bilingualFromPlan reads the plan file (CLI-written), defaults to false on any error. A local plan has no such field, so terminal-only reviews are unchanged.
• Bi interface and render function are clean. The zh === en guard prevents empty folds. The footer stays outside the fold, once.
• verificationGaps structural change (string[] → {subject, reason, subjectZh, reasonZh}) removes the last prose-reparsing boundary recovery — a genuine improvement, not just bilingual plumbing.
• publicLabelZh is added as a required field on Brief and consistently provided for all 18+ brief definitions. No missing entries.
• reasonZhOf map correctly handles entries sharing the same English reason — they share the Chinese one by construction.
• Untranslatable fragments (model-written findings, caller echoes, interpolated errors) ride verbatim in both halves — correct degradation.
• SKILL.md changes are minimal and well-placed: Step 1 field list gains prDescriptionHasHan, Step 7 gains the bilingual inline comment instruction. The severity marker and ```suggestion block stay English-only (tooling filters on them).

Testing

tmux is not available in this CI environment. Direct verification was performed instead:

Unit tests (PR branch, 793 tests):

 Test Files  31 passed (31)
      Tests  793 passed (793)
   Duration  11.15s

TypeScript: tsc -p packages/cli --noEmit — clean, no errors.

ESLint: --max-warnings 0 on all 4 changed production files — clean.

describeChunkGap bilingual output (built from PR branch):

=== describeChunkGap: entire diff ===
{
  "phrase": "the entire diff",
  "phraseZh": "整个 diff",
  "plural": false
}
=== describeChunkGap: narrow gap ===
{
  "phrase": "the diff section covering src/b.ts",
  "phraseZh": "涉及 src/b.ts 的 diff 片段",
  "plural": false
}
=== describeChunkGap: count gap ===
{
  "phrase": "the diff sections covering src/a.ts, src/b.ts",
  "phraseZh": "涉及 src/a.ts、src/b.ts 的 diff 片段",
  "plural": true
}
=== describeChunkGap: no plan ===
{
  "phrase": "1 section of the diff",
  "phraseZh": "diff 中的 1 个片段",
  "plural": false
}

Han detection regex (/\p{Script=Han}/u):

✓ Chinese description: true (expected true)
✓ English description: false (expected false)
✓ Mixed description: true (expected true)
✓ Japanese kanji: true (expected true)
✓ Empty body: false (expected false)
✓ Null body: false (expected false)

中文说明

代码审查

独立方案： 我会在 fetch-pr 中检测作者语言（它已经获取 PR 元数据），把标志盖进 plan，然后在 compose-review 中用仓库现有的 <details> 格式为每个确定性片段渲染双语。对于行内评论（模型手写），我会添加 SKILL.md 指令。verificationGaps 的文本反解析（gap.indexOf(' — ')）必须去掉——它无法在双语文本中存活。

PR 的方案与此完全一致。 没有遗漏更简路径。

未发现阻断问题或 AGENTS.md 违规。具体观察：

• 信任边界正确。 prDescriptionHasHan 由 fetch-pr 从已获取的 PR body 计算——调用方无法切换。bilingualFromPlan 读取 plan 文件（CLI 写入），任何错误时默认 false。本地 plan 无此字段，纯终端 review 不受影响。
• Bi 接口和 render 函数 干净。zh === en 守卫防止空折叠。footer 在折叠外只出现一次。
• verificationGaps 结构化变更（string[] → {subject, reason, subjectZh, reasonZh}）消除了最后一处从渲染文本反解析边界的逻辑——是真正的改进，不只是双语管道。
• publicLabelZh 作为 Brief 的必填字段添加，所有 18+ 个 brief 定义都一致提供。无遗漏。
• reasonZhOf map 正确处理共享同一英文 reason 的条目——它们按构造共享中文翻译。
• 不可翻译片段（模型写的发现、caller 回显、报错插值）在两个半区原样出现——正确的降级。
• SKILL.md 变更 最小且位置恰当：Step 1 字段清单增加 prDescriptionHasHan，Step 7 增加双语行内评论指令。严重级别标记和 ```suggestion 块只留在英文半区（工具链以它们为过滤键）。

测试

本 CI 环境无 tmux。改用直接验证：

单元测试（PR 分支，793 个测试）：全部通过。TypeScript 和 ESLint 均无错误。describeChunkGap 双语输出和 Han 检测正则均验证正确。

— Qwen Code · qwen3.8-max-preview

_{Reviewed at 0f0f33787b76274faa198b127b61457c2a8af375 · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

Confidence: 5/5 — clean across every stage; would merge without hesitation.

This is a well-crafted PR. The approach is the obvious right one — detect the author's language where the metadata is already fetched, render the deterministic body bilingually from the same structural data, and let the model handle inline comments via a SKILL.md instruction. No over-engineering, no missed simpler path.

What stands out: the verificationGaps structural change (string[] → {subject, reason, subjectZh, reasonZh}) isn't just bilingual plumbing — it removes the last prose-reparsing boundary recovery in compose-review, which was a latent bug surface. The en/zh pairs sit side by side at every production site, so an edit that touches one cannot silently miss the other. Untranslatable fragments degrade to verbatim English rather than dropping content. The trust boundary is right: the caller cannot toggle the register of a certified body.

793 tests pass, TypeScript and ESLint clean, the built output produces correct bilingual phrases. The only thing I couldn't do is a live tmux run (not available in this environment), but the compose pipeline is deterministic and thoroughly unit-tested — the live rendering is standard GitHub <details> markdown.

中文说明

信心：5/5 —— 每个阶段都干净；毫不犹豫地合并。

这是一个精心制作的 PR。方案是显而易见的正确选择——在已获取元数据的地方检测作者语言，从同一份结构化数据渲染确定性 body 的双语版本，让模型通过 SKILL.md 指令处理行内评论。没有过度工程，没有遗漏更简路径。

亮点：verificationGaps 的结构化变更（string[] → {subject, reason, subjectZh, reasonZh}）不只是双语管道——它消除了 compose-review 中最后一处从渲染文本反解析边界的逻辑，那是一个潜在的 bug 面。中英文对在每个产生点并排存在，改动一处很难悄悄漏掉另一处。不可翻译片段退化为原样英文而不是丢内容。信任边界正确：调用方无法切换认证 body 的语域。

793 个测试通过，TypeScript 和 ESLint 无错误，构建输出产生正确的双语短语。唯一没做的是 tmux 实际运行（本环境不可用），但 compose 管道是确定性的且有充分的单测覆盖——实际渲染是标准 GitHub <details> markdown。

— Qwen Code · qwen3.8-max-preview

_{Reviewed at 0f0f33787b76274faa198b127b61457c2a8af375 · re-run with @qwen-code /triage}

[qwen-code-ci-bot]

LGTM, looks ready to ship. ✅

[yiliang114]

LGTM

[yiliang114]

Review Summary

Clean bilingual rendering design. Detection (\p{Script=Han}) lives in fetch-pr, boolean flag in plan JSON, compose-review reads and renders — good separation of concerns. The Bi interface unifies all body fragments, zh !== en guard avoids empty folds for English-only content. Tests cover fold structure, English fallback, LGTM translation, disclosure translation, and untranslatable passthrough.

Suggestions (non-blocking):

(1) \p{Script=Han} also matches Japanese kanji and Korean hanja — a Japanese PR author would trigger the Chinese collapsed block. Consider excluding pure Japanese (check for Hiragana/Katakana presence), or document as accepted tradeoff.

(2) downgradedFrom === 'Approve' ? '批准' : '请求修改' is a hardcoded binary. If a new downgradeable event type is added, Chinese text defaults to '请求修改'. Consider a map with English fallback.

(3) bilingualFromPlan does a second readFileSync + JSON.parse of the plan file that composeReview already parsed upstream. Pass prDescriptionHasHan from the existing plan object to avoid redundant I/O.

(4) Inline comment bilingual formatting relies on the model following natural-language instructions in SKILL.md — no code-level guarantee of <details> format consistency (body side is code-rendered, inline side is model-dependent).

(5) GapText type — confirm gapZh constraint is present so new gap types without Chinese translations fail at compile time.

— qwen3.8-max-preview via Qwen Code /review

[gwinthis]

Review: APPROVE (C=0)

Summary

Well-architected feature (+484/-123, 7 files) that makes /review post bilingually when the PR description contains Chinese. The structural Bi pair approach ({en, zh}) separates translation (per-fragment) from rendering policy (centralized), and the CLI-owned detection signal prevents callers from toggling the language register of a certified body.

Verification Report (tmux)

✓ 31 test files, 793 tests — all passed (16.96s)
✓ tsc -p packages/cli --noEmit: 0 errors in review files

5 new bilingual tests + updated describeChunkGap/verificationGaps assertions pinning Chinese strings.

Key Findings

#	Severity	Finding
1	Minor	Han detection (\p{Script=Han}) matches Japanese kanji — acceptable heuristic, false positive produces a harmless collapsed block
2	Minor	No direct unit test for prDescriptionHasHan detection in fetch-pr.test.ts (rendering side thoroughly covered)
3	Suggestion	publicLabelZh required on Brief — breaking but internal; compiler enforces completeness
4	Suggestion	VerificationReport.gaps shape change (string[] → structural) well-migrated; eliminates prose reparsing

Architecture Insights

1. Bi pair pattern: Every body fragment carries {en, zh} as data. A single render() function decides at the end whether to fold Chinese. Translation is distributed; rendering policy is centralized. zh === en suppresses empty folds.

2. CLI-owned signal: prDescriptionHasHan is stamped by fetch-pr into the plan — the CLI's own artifact. The caller cannot toggle the register. Unreadable plan defaults to English-only: "the language must never take the review down."

3. Structural gaps replace prose reparsing: Old gaps: string[] forced compose-review to recover subject/reason boundaries via indexOf(' — '). New {subject, reason, subjectZh, reasonZh} eliminates this — when a consumer needs to decompose what a producer rendered, emit structure instead.

4. Untranslatable text rides verbatim: Model-written findings set zh = en — quoted as-is in both halves. Avoids silent omission and hallucinated translation.

5. SKILL.md as inline-comment contract: Body is deterministic (compose-review); inline comments are model-written. SKILL.md Step 7 tells the model the bilingual format — severity markers and suggestion blocks stay English-only (tooling filters on them).

中文说明

评审：APPROVE (C=0)

概要

架构良好的功能（+484/-123，7 文件），让 /review 在 PR 描述含中文时发布双语评论。结构化 Bi 对（{en, zh}）将翻译（逐片段）与渲染策略（集中式）分离，CLI 自有的检测信号防止调用方切换认证 body 的语域。

验证报告（tmux）

• 31 个测试文件，793 个测试全部通过（16.96s）
• typecheck：review 文件 0 错误

架构洞察

1. Bi 对模式 — 翻译分布式，渲染策略集中式
2. CLI 自有信号 — 调用方不能切换语域
3. 结构化 gaps 替代文本反解析 — 生产者发射结构，消费者不再拆解渲染文本
4. 不可翻译文本原样保留 — 避免静默省略和幻觉翻译
5. SKILL.md 作为行内评论契约 — body 确定性生成，行内评论由模型按 SKILL 指令撰写

— qwen3.7-max via Qwen Code /review