问题描述： 在多语言（特别是中文）开发环境下，即便在 Global Rules / System Instructions 中明确要求使用全中文思考链路（Strict Chinese Chain of Thought），模型在 <thought> 推理块中仍频繁退化为全英文思考

[pai-pixel]

Problem Statement

Describe the problem this feature would solve.
Title: [Bug & Feature Request] Native Support for Reasoning Language & Strict Multilingual CoT (Preventing English Priming in Thinking Blocks)

Component: Antigravity IDE / Antigravity 2.0 / Agent Reasoning Engine
Severity: Medium (UX / Multilingual Developer Experience)

--- Description (中文说明) ---
问题描述：
在多语言（特别是中文）开发环境下，即便在 Global Rules / System Instructions 中明确要求使用全中文思考链路（Strict Chinese Chain of Thought），模型在 推理块中仍频繁退化为全英文思考。

根本原因分析：

1. 语言启动偏置（Language Priming）：系统级注入的英文 System Prompt 以及工具回传（终端输出、代码、API 响应）严重污染了 Attention 权重，导致首词预测偏向英文。
2. 缺少采样级约束：当前系统仅依赖 Prompt 软约束，缺乏类似 reasoning_language 的强制解码前缀或采样约束。
3. UI 状态条本地化缺失：Thinking 折叠组件与工具调用状态显示缺乏多语言本地化配置。

诉求与建议（Feature Request）：

1. 建议在 Settings 中提供 原生思考语言配置（e.g., "reasoningLanguage": "follow_ui" | "zh" | "en"）。
2. 在底座推理的 标签起始处，支持根据用户配置自动注入语言锚点（Prefix Injection），压制英文偏置。
3. 优化上下文管理，降低系统级英文 Prompt 与工具输出对思考链语言的负面干扰。

--- English Version for Google Engineering Team ---
Issue Summary:
In non-English (specifically Chinese) programming environments, the reasoning model frequently defaults back to English within the (Chain of Thought) block, despite explicit and strict system instructions / global rules enforcing Chinese reasoning.

Root Cause Analysis:

1. Contextual Priming: The heavy presence of English in base system instructions, MCP schemas, and tool outputs (compiler logs, stdout/stderr, docs) introduces a strong English priming bias into the attention window.
2. Lack of Decoding-Level Constraint: Relying solely on soft prompt guidance fails against the model's base RLHF distribution. There is no decoding-level enforcement (such as prefix forcing or logit biasing) for reasoning tokens.
3. UI Localization Gaps: The Thinking accordion and tool action displays lack seamless locale-aware integration.

Requested Features / Solutions:

1. Dedicated Reasoning Language Setting: Introduce an explicit configuration option (e.g., agent.reasoningLanguage: "zh" | "auto" | "en").
2. First-Token / Anchor Forcing: Allow the engine to inject an initial native-language thought anchor immediately after <thought> to guide subsequent token probability.
3. Decouple System Prompt Priming: Mitigate the English priming effect introduced by underlying runtime injection.

Proposed Solution

Describe the solution you'd like.

Alternatives Considered

Any alternative approaches you've considered.

Additional Context

Any other relevant information.