feat(i18n): localize system prompt to Chinese for zh-Hans users

[bistack]

Problem

The system prompt that drives reasoning is ~90% English (constitution.md, mode prompts, tool descriptions, approval policies, output rules). When a Chinese-speaking user writes in Chinese, the model's reasoning_content defaults to English, then translates back to the final reply — producing unnatural, machine-translated Chinese that is hard to read.

Current mitigations are instruction-level patches that cannot outweigh the English context mass:

• LOCALE_PREAMBLE_ZH_HANS (crates/tui/src/prompts.rs:795) — a few Chinese paragraphs at the top of the system prompt
• LOCALE_CLOSER_ZH_HANS (crates/tui/src/prompts.rs:831) — reinforcement at the bottom
• Translation mode — English thinking + display-time translation (adds latency, produces awkward output)

These were added in response to #732 and #1118, but the core prompt layers themselves were never localized. The scattered instructions are fighting thousands of tokens of English context; the model naturally follows the dominant language.

Proposed solution

When ## Environment.lang is zh-Hans, select a Chinese variant of each prompt layer instead of the English default:

crates/tui/src/prompts/
├── constitution.md          → constitution.zh-Hans.md
├── language.md              → language.zh-Hans.md
├── output.md                → output.zh-Hans.md
├── core_execution.md        → core_execution.zh-Hans.md
├── compact.md               → compact.zh-Hans.md
├── memory_guidance.md       → memory_guidance.zh-Hans.md
├── subagent_output_format.md → subagent_output_format.zh-Hans.md
├── modes/
│   ├── agent.md             → agent.zh-Hans.md
│   ├── operate.md           → operate.zh-Hans.md
│   ├── yolo.md              → yolo.zh-Hans.md
│   └── plan.md              → plan.zh-Hans.md
└── approvals/
    ├── never.md             → never.zh-Hans.md
    ├── suggest.md           → suggest.zh-Hans.md
    └── auto.md              → auto.zh-Hans.md

The prompt builder in prompts.rs already has the locale tag available (PromptSessionContext.locale_tag). Each include_str! layer picks its file variant based on the resolved locale. Missing locale variants fall back to the English default — partial localizations are safe.

Once all layers are localized, LOCALE_PREAMBLE_ZH_HANS and LOCALE_CLOSER_ZH_HANS become redundant and can be removed in a follow-up.

Other locales (ja, pt-BR, vi, etc.) can adopt the same file-convention pattern — no new code paths, just new .md files.

Review requirement

All Chinese prompt files must be reviewed by a native Chinese speaker before merge. Machine-translated prompts defeat the purpose of this feature — the translations must read as natural, professional Chinese that a developer would write to another developer.

Acceptance criteria

• A zh-Hans user sees an entirely Chinese system prompt (no English preamble/closer hacks needed).
• reasoning_content and final reply are both in Chinese without requiring translation mode.
• English-only users are unaffected (default path unchanged).
• Adding a new locale follows the same file-convention pattern.

Related

• for Chinese,show thinking use Chinese please #732 — original request for Chinese thinking
• Language has been configured to Chinese, but "thinking" outputs are still in English #1118 — thinking still in English despite Chinese config
• Deduplicate project, environment, route, locale, and skill context #4709 — context dedup (locale layers should not duplicate content across languages)

[Hmbown]

I've worked on this - are you seeing a difference?

[bistack]

'v0.9.3' reasons in English while the chat is in Chinese.

[bistack]

Latest main branch:

[bistack]

Key observation

LLM reasoning language is a model statistic, not a harness-controlled output. Each reasoning_content token is sampled from P(token | all preceding context), so the language of the output tracks the language composition of the context — not the language of the instructions in it. Two variables dominate:

1. First-turn language: the session's opening reasoning latches the language, and the model self-reinforces it across turns by conditioning on its own prior reasoning.
2. Chinese share of the context: as English code, logs, and tool output accumulate, Chinese becomes a shrinking minority of the conditioning distribution, and reasoning drifts to English — even with explicit Chinese-thinking instructions.

This is a model characteristic: even a Chinese-native model (DeepSeek V4 is Chinese-first) adapts to the language mixture it is conditioned on. The harness cannot change the model's prior; instructions are only a few dozen tokens of evidence against tens of thousands of English tokens. So under an English-majority context, English reasoning is the expected outcome, and no amount of prompt wording reliably prevents it.

[bistack]

The Chinese translated from English thought chains is painful to read.