Context

The Chinese tech Twitter audience is a real userbase right now. We already have:

• `crates/tui/src/localization.rs` — Chinese-Simplified (`zh-Hans`) translations for 27 `MessageId` constants (composer placeholder, history search, config modal, help modal). Full coverage for the surface that's wired through `MessageId`.
• `README.zh-CN.md` — Chinese README, kept in sync with English.
• `Settings.locale` accepts `zh-Hans`, with `zh-CN` as an alias (`localization.rs:536`).

What's not translated yet — most of what a Chinese-speaking user actually sees during a session:

1. Slash command descriptions and `usage:` strings — `crates/tui/src/commands/mod.rs:109-375` has ~50 `CommandInfo` entries with English-only `description` and `usage` fields. `/help` renders these directly.
2. Slash command output — `tokens`, `cost`, `cache` (new in feat(debug): add /cache command surfacing per-turn DeepSeek cache hit/miss #278), `models`, `subagents`, etc. all return literal English strings.
3. Status messages and toasts — `app.status_message = Some("...".to_string())` call sites are English-only. Same for `StatusToast`.
4. Error messages surfaced to the user — `anyhow!` contexts (`failed to read directory`, `shell manager lock poisoned`, …), `ToolError` (`path_escape`, `missing_field`), and `format_tool_error` in the engine.
5. Doctor / setup / onboarding output — `deepseek doctor`, the first-run setup wizard, the missing-companion-binary error message (`bb88ab91`, English-only).
6. Footer text and labels — mode indicators ("AGENT", "YOLO", "PLAN"), mode hints ("Tab to switch mode"), and the cost/context summary line.

Why this is worth a milestone-sized effort

Without this expansion, a `zh-Hans`-locale user gets a Chinese composer hint and a Chinese config modal, then English everything else once they actually start working. The locale switch is currently more cosmetic than functional. Two consequences:

• Bug reports won't translate. Error messages like "path escapes workspace" are unreadable to a non-English speaker, so they screenshot and we need to translate by hand.
• Pointless first-impression cost. A Chinese-speaking user runs `deepseek` for the first time, gets a localized welcome, and then sees "Token Usage: ─── Active context: ~123 / 1000000" — sets the expectation that the rest of the tool is half-translated.

Scope of this issue

Concrete scope to land in v0.8.5 (or later — pick your milestone):

Phase 1 — high-traffic user-facing surfaces

• Slash command help text — add `description_id: MessageId` next to `description: &str` in `CommandInfo`. Render via `tr(locale, id)`. Affects ~50 entries; mostly mechanical.
• `/tokens`, `/cost`, `/cache` (the debug commands) — these are paste-into-bug-report surfaces. Translate the labels ("Active context", "Last API input", "Cache hit/miss", "Approx session cost", "Σ in", "avg hit ratio", …). Numbers stay numeric.
• `/help` — already routes through MessageId; verify the Chinese strings match the new command set after Phase 1 lands.
• Footer mode labels — "AGENT/YOLO/PLAN" can stay (they're capitalized acronyms now and a recognizable brand), but the hint strings ("Tab=switch mode", etc.) should localize.
• Doctor output — `deepseek doctor` is the first thing a frustrated user runs. Translate the section headings and the verdict lines.

Phase 2 — onboarding and setup

• Setup wizard (`crates/tui/src/...`/setup) — first-run, gets read carefully.
• Missing-companion-binary error (`bb88ab91`) — already a long localized message in English, deserves zh-Hans.
• `deepseek init` prompts — the project-context initialization wizard.

Phase 3 — error messages

• `format_tool_error` (`core/engine.rs`) — every tool failure that surfaces to the user.
• Sandbox-denial messages (`SandboxManager::denial_message`) — the macOS Seatbelt / Linux Landlock / Windows messages.
• `approval_*` messages — what users see when the approval prompt fires.

Out of scope intentionally

• Tool descriptions sent to the model (`tool.description()`) — these stay English so DeepSeek's training data alignment isn't disturbed and the prefix cache doesn't fork into per-locale variants.
• The base system prompt — same reason. The model speaks the user's language because `zh-Hans` users ask in Chinese, not because we translated the prompt.

Implementation notes

• Add MessageIds in alphabetical groups so the diff stays reviewable.
• Add a CI test that asserts every locale function returns `Some(_)` for every `MessageId` in `ALL_MESSAGE_IDS` — there's already a partial check in `localization.rs:524+`; tighten it to fail on any missing id.
• Consider hiring a native zh-Hans speaker for review pass after the mechanical translation. The current 27 strings are clean (verified by reading them), but the 200-300 strings this expansion adds are too many to vibe-check without a native speaker.

Related

• Existing translations: `crates/tui/src/localization.rs:449-483` (`chinese_simplified`)
• Locale aliases: `crates/tui/src/localization.rs:536` (`zh-CN` → `zh-Hans`)
• README: `README.zh-CN.md`