fix: 防止 thinking 泄漏到 operator-visible 投影

[jolestar]

变更概述

• 让 Web GUI 的 newest brief、timeline brief 与 agent lastBrief 统一以持久化 BriefRecord.text 为 canonical content，不再从 transcript 拼接正文
• verbose assistant_round_recorded 只投影 assistant text block，避免 thinking、reasoning 与 tool block 泄漏到 operator-visible 文本
• 强化主模型响应语言契约，明确语言选择优先级，并覆盖工具发布或 artifact 携带的 review/comment/PR body 等 human-facing prose
• 保持 best-effort：不增加额外 LLM 调用、语言检测、重试或 brief promotion 阻断

根因

GUI 的多条 brief 展示路径优先或允许使用关联 transcript；通用 transcript 文本提取又拼接了所有带 text/content 的 block，导致 thinking 与最终答复混合。与此同时，弱模型可能在工具参数中生成英文的人类可读 review 文本，并通过 command preview 回灌后续上下文。

验证

• cargo fmt --all -- --check
• RUSTFLAGS="-D warnings" cargo check --all-targets
• cargo test
• cargo test --lib response_language
• Web GUI runtime 定向测试：63 项通过
• Web GUI typecheck
• GPT 5.5 与 Qwen 3.7 Max 多场景 live probe

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
holon	Ready	Preview, Comment	Jul 10, 2026 12:06pm

[jolestar]

LGTM. Well-scoped fix that correctly addresses thinking/reasoning leakage into operator-visible projections.

Summary of review:

1. Brief text canonical source (client.ts, runtime-store.ts) — Correctly makes persisted BriefRecord.text the single source of truth for brief display, removing the transcript-based text extraction that was concatenating all block types including thinking. Clean removal of transcriptEntryText() from both files.

2. Verbose assistant round filtering (session-reducer.ts) — The type === "text" filter in transcriptEntryText() is the minimal correct fix: only operator-visible text blocks are projected, while thinking/reasoning/tool_use blocks are excluded.

3. Response language contract (prompt/mod.rs) — Moving response_language_section() to the last position gives it proper recency weight. The rewritten contract with 4-level precedence and explicit coverage of tool-published prose (PR reviews, comments, issue bodies) is a meaningful improvement.

4. Test coverage — Excellent: 3 new test suites covering all three change areas with targeted assertions. Rust tests updated for new ordering.

No blocking issues found.