Feat/english thinking when hidden

[cmyyy]

Problem

When show_thinking is disabled, thinking blocks are hidden from the
UI (HistoryCell::Thinking is filtered out in history.rs), but the
API still generates reasoning_content (controlled separately by
reasoning_effort). The ## Language rule in base.md forces the
thinking chain to match the user's input language — so a Chinese user
pays for Chinese thinking they never see.

The token-count difference between Chinese and English reasoning is
modest (DeepSeek's tokenizer handles CJK efficiently), but there is
still waste: the surrounding system prompt is English, mixed-language
reasoning fragments the token stream, and invisible content has no
reason to be localized at all.

Solution

Inject a ## Thinking Language override into the system prompt when
show_thinking is false. The model is told:

The user has disabled thinking display in settings — they will never
see your reasoning_content. Therefore, your internal thinking MUST
be in English regardless of the user's language. Your final reply
must STILL match the user's language.

Benefits:

• Prompt-cache locality — English reasoning sits better inside an
  English system prompt, improving prefix-cache hit rates.
• Mixed-language overhead — reasoning often interleaves code
  identifiers, file paths, and API names with natural language.
  Switching between CJK and ASCII mid-stream creates unnecessary
  token-boundary breaks.
• Aligns cost with intent — when the user hides thinking, there is
  no user-facing reason to localize it.

Changes (6 files, +53 lines)

File	Change
crates/tui/src/prompts.rs	Add show_thinking to PromptSessionContext; inject ## Thinking Language override

|
| crates/tui/src/core/ops.rs | Add show_thinking to Op::SendMessage |
| crates/tui/src/core/engine.rs | Add to EngineConfig; thread through handle_send_message and
refresh_system_prompt |
| crates/tui/src/tui/ui.rs | Wire app.show_thinking through build_engine_config and prompt construction |
| crates/tui/src/main.rs | Pass show_thinking in CLI exec path |
| crates/tui/src/runtime_threads.rs | show_thinking: false for background agent threads |

Data flow: Settings::show_thinking → App::show_thinking →
Op::SendMessage → EngineConfig::show_thinking →
PromptSessionContext::show_thinking → override injection.

Behavior

show_thinking	Reasoning language	Reply language
true (default)	Matches user input	Matches user input
false	English (forced)	Matches user input

Testing

• cargo test --all-features — 3120 passed; 6 pre-existing
  failures unrelated to this change (API key / temp directory /
  AGENTS.md path)
• cargo fmt --all -- --check — clean
• cargo clippy --all-targets --all-features — clean (2
  pre-existing needless_return warnings unrelated to this change)

Checklist

• Updated docs or comments as needed
• Added or updated tests where relevant (existing prompt tests
  updated for the new PromptSessionContext field)
• Verified TUI behavior manually if UI changes (no UI logic
  change; the entire path is a backend-only data flow)

[gemini-code-assist]

Code Review

This pull request introduces a show_thinking configuration option across the TUI engine, prompts, and UI components. When disabled, the system prompt is updated to instruct the model to perform its internal reasoning in English to save tokens, while still responding in the user's preferred language. Feedback was provided regarding the formatting of the multi-line system prompt string in crates/tui/src/prompts.rs, where the use of line continuation characters might introduce unintended whitespace.

[gemini-code-assist]

The use of \ for line continuation in this string literal will include all the leading whitespace from the subsequent lines. This results in a single long line with multiple spaces between words, which is likely not the intended formatting for the prompt. Using string literal concatenation is a cleaner way to format this multi-line string while maintaining readability in the code.

        full_prompt.push_str(
            "\n\n## Thinking Language\n\n"
            "The user has disabled thinking display in settings — they will "
            "never see your `reasoning_content`. Therefore, your internal "
            "thinking MUST be in English regardless of the user's language. "
            "This directive overrides the `## Language` section above for "
            "reasoning_content only. Your final reply must STILL match the "
            "user's language."
        );

[Hmbown]

Independent review:

PR threads a new show_thinking: bool from settings through EngineConfig → PromptSessionContext → the system prompt builder, and emits a ## Thinking Language block when !show_thinking instructing the model to keep reasoning_content in English while replying in the user's locale. Wiring is correct and the test coverage is reasonable.

This is fully superseded by v0.8.48 (pr-2256):

• crates/tui/src/prompts.rs already declares pub show_thinking: bool on PromptSessionContext (line 40) with default true.
• v0.8.48 emits the override under the header ## Hidden Thinking Language (line 118-121 of prompts.rs) with essentially the same semantics: "keep that hidden internal thinking in English regardless of..." — the test even asserts text.contains("## Hidden Thinking Language") and text.contains("reasoning_content") && text.contains("English").
• runtime_threads.rs in v0.8.48 already reads Settings::load().show_thinking (line 1614) and propagates it.
• All ~12 show_thinking: ... initializers across prompts.rs tests in v0.8.48 confirm the field is fully integrated.

Merge sim vs main: conflicts in core/engine.rs, main.rs, prompts.rs, runtime_threads.rs, tui/ui.rs. Vs pr-2256: same five files conflict — v0.8.48 already owns the surface this PR wants to add.

Header naming differs slightly (## Thinking Language here vs ## Hidden Thinking Language in v0.8.48); v0.8.48's is clearer.

v0.8.48 (#2256) compatibility: superseded — v0.8.48 implements the same feature with the same show_thinking field name and equivalent prompt-injection semantics. Recommend close as superseded by #2256.

[Hmbown]

@cmyyy — thank you for this. The "English thinking when hidden" feature you proposed is now fully present in v0.8.48 (#2256) — same show_thinking: bool field on PromptSessionContext, the same ## Hidden Thinking Language system-prompt block, and the same Settings::load().show_thinking wiring through runtime_threads.rs. Your design landed.

If there's a specific test scenario or edge case from your branch you'd like included in #2256 before it merges — particularly around the locale interaction (zh/ja preserving final-reply language while reasoning stays English) — flag it and I'll help carry it forward. Otherwise this gets credited in the v0.8.48 release notes.

Genuinely good idea — thanks for proposing it.