[xarthurx]

What changed

Root cause — Chinese dictation mixed with English (bilingual mode). The bilingual Whisper prompt was a Chinese sentence followed by its English translation. Whisper treats the prompt as the preceding transcript and imitates it, writing every spoken segment twice: as an interleaved English translation (那你就记得。 Then you remember.) or a verbatim duplicate (Thank you for the email. Thank you for the email.). Bilingual mode now conditions on one language's sentence (first slot that has one), with the dictionary after a Latin-script sentence and before a CJK one.

Enhancement. Bilingual cleanup uses a code-switch-preserving rule (never translate; Chinese / Japanese / English punctuation each in its own convention). Bare base codes no longer fall into the forced "respond in X regardless of what the user spoke" template: zh → its tailored instruction; en (en-US/en-GB) → the language-preserving auto instruction, so no spelling convention is imposed.

Provider requests (audit against current provider docs):

• verbose_json only for Whisper models — gpt-transcribe rejects it (HTTP 400).
• gpt-transcribe: languages[] (both languages in bilingual mode) + sanitized keywords[]; detection read leniently from languages[].code.
• gpt-4o-transcribe-diarize: no prompt.
• Mistral: dictionary as context_bias, no prompt/response_format; voxtral-small-latest (not a transcription-endpoint model) removed and migrated to voxtral-mini-latest.
• Refactor: request closures take RequestHints { prompt, vocabulary }; the dictionary-echo retry drops both.

Normalization. CJK small-form marks (﹐﹑﹒﹔﹕﹖﹗) → standard full-width; in a pair that includes Chinese, Han text under an English resolution keeps Chinese punctuation + Simplified (other pairs such as ja+en keep plain English handling, so kanji are never simplified); more Chinese silence-hallucination phrases.

Evidence

• History DB: 0 interleaved translations in ~1,000 Chinese dictations Feb–May 2026; the first appeared 2026-06-04, the day bilingual mode (v0.8.0) shipped.
• 22-clip eval on Groq whisper-large-v3 (TTS + noise/pauses; Chinese, English, code-switched): old prompt 12/22 clean (10 duplicated/translated/echoed); the shipped shapes 22/22.
• Live run of the real Rust request path (transcribe_groq + new prompt builder) against Groq: fields accepted, detection returned, no translation/duplication.

Verification

• cargo test: 255 passed · bun test: 30 passed · bun run typecheck clean · bun run build OK
• cargo clippy --all-targets: 8/9/1 warnings vs baseline 12/13/1 (none in changed code)
• Fresh whole-branch review: 3 Important findings + 1 re-graded Minor fixed, each with a failing-first regression test.

Not verified live

• gpt-transcribe request shape (dev machine's OpenAI key returns 401) and Mistral context_bias (no key). Both follow the providers' docs/SDK encoding; detection parsing is lenient so a shape surprise only loses detection. The Standard default stays gpt-4o-mini-transcribe until a live check (TODO).

Plan: docs/superpowers/plans/2026-10-13-bilingual-prompt-and-provider-requests.md

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Pair-aware cleanup normalization is incorrectly applied to chat responses and can rewrite valid non-Chinese Han text.

2 open findings

• Restrict bilingual normalization to cleanup mode · New
• Correct architecture description of bare language mapping · New

What changed in this PR

Fixes bilingual transcription duplication/translation and aligns provider request payloads with model-specific APIs.

Changes:

• Uses single-language bilingual conditioning and language-preserving cleanup.
• Adds provider-specific vocabulary, language, and response handling.
• Improves CJK normalization, hallucination filtering, migrations, tests, and documentation.

File	Description
tests/​retiredModels.test.ts	Tests Mistral model migration.
tests/​prompts.test.ts	Tests bilingual language instructions.
src/​utils/​languageSupport.ts	Adds bilingual and base-code handling.
src/​services/​tauriApi.ts	Passes language pairs to reasoning.
src/​models/​modelRegistryData.json	Removes unsupported Voxtral Small.
src/​hooks/​useTranscriptionPipeline.ts	Threads bilingual context through cleanup.
src/​config/​retiredModels.ts	Migrates retired Mistral model.
src/​config/​prompts.ts	Applies bilingual cleanup prompts.
src-tauri/​src/​transcription/​normalize.rs	Adds pair-aware CJK normalization.
src-tauri/​src/​transcription/​mod.rs	Revises bilingual Whisper conditioning.
src-tauri/​src/​transcription/​hallucination.rs	Expands Chinese hallucination filtering.
src-tauri/​src/​transcription/​dictionary_echo.rs	Separates prompt and vocabulary hints.
src-tauri/​src/​transcription/​cloud.rs	Adds provider-specific request fields.
src-tauri/​src/​commands/​transcription.rs	Threads languages and request hints.
src-tauri/​src/​commands/​reasoning.rs	Applies pair-aware response normalization.
docs/​TODO.md	Records pending live verification.
docs/​superpowers/​plans/​2026-10-13-bilingual-prompt-and-provider-requests.md	Documents implementation plan.
docs/​PROGRESS.md	Records diagnosis and evaluation results.
docs/​CHANGELOG.md	Documents user-visible fixes.
docs/​ARCHITECTURE.md	Updates transcription architecture.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.