feat(asr): 新增「自定义 OpenAI 兼容」ASR 预设与高级参数（#854）

[H-Chris233]

User description

背景

issue #854：Windows 用户无法稳定把 ASR 指到局域网自建 OpenAI 兼容端点（llama.cpp 等）。现有 Whisper 兼容预设都是命名厂商（whisper / siliconflow / zhipu / groq / stepfun / openrouter），各自带着厂商特异的硬编码开关，对“任意 OpenAI 兼容端点”不够通用。

改动

• 新增 ASR 预设 openai-compatible（UI 名「自定义 OpenAI 兼容」）：endpoint / model / api key 全部由用户填写，无默认值。
• 行为默认最保守：multipart、不发 response_format、不分片，与服务端能力对齐。
• 该预设提供两个高级选项（设置页，仅对通用预设显示）：
  • verbose_json 开关（配合幻听过滤；自建服务不支持时保持关闭）；
  • 分片时长（0 = 不分片，适合长录音或服务端单次请求时长受限）。
• 高级配置按 provider 存于系统凭据 vault（账号 asr.advanced_config，JSON {"verboseJson": bool, "chunkDurationMs": number|null}），缺失/非法回落保守默认。
• 命名厂商预设行为保持不变（whisper/groq 仍发 verbose_json，zhipu/openrouter 仍 30s 分片等）。

验证

• cargo check --lib 通过；新增 4 个单测（默认保守行为 / 高级配置切换 / JSON 解析容错 / vault 写读往返）通过；whisper 相关 36 个回归用例通过。
• 前端 tsc 与 vite build 通过；advancedAsrConfig.test.ts、ProvidersSection.test.ts 通过。
• 说明：本 PR 只含上述改动，不含同一工作区中其它并行任务的未提交改动。

---

PR Type

Enhancement, Bug fix

---

Description

• Add openai-compatible ASR preset for arbitrary OpenAI-compatible endpoints.

• Allow empty API key, omitting Authorization to avoid 401.

• Require endpoint/model for new preset with clear errors.

• Add advanced verbose_json/chunking config with frontend UI and tests.

---

Diagram Walkthrough

flowchart LR
  ASR["Custom OpenAI-compatible ASR"] --> ADV["Advanced options (verbose_json, chunking)"]
  ASR --> NOAUTH["Optional API key, omit Authorization header"]
  ADV --> VAULT["Vault per-provider config"]
  ADV --> UI["Settings UI + i18n"]
  NOAUTH --> LAN["LAN llama.cpp transcription"]

Loading

File Walkthrough

	Relevant files
Enhancement	7 files whisper.rs Optional Authorization header for empty API key                    +62/-12  providers.rs Make ASR API key optional for custom preset                            +11/-8    coordinator.rs Add advanced ASR config parsing and chunking                          +166/-1  asr_wiring.rs Gate openai-compatible credentials and verbose_json            +63/-2    advancedAsrConfig.ts Create advanced config parse/serialize helpers                      +41/-0    ProvidersSection.tsx Add advanced options UI for custom ASR                                      +115/-1  shared.tsx Add openai-compatible preset to ASR presets                            +4/-0
Configuration changes	2 files credentials.rs Register AsrAdvancedConfig credential account                        +2/-0      credentials.rs Persist advancedConfig in ASR provider entries                      +43/-2
I18n	5 files en.ts Add English translations for custom ASR preset                      +6/-0      ja.ts Add Japanese translations for custom ASR preset                    +6/-0      ko.ts Add Korean translations for custom ASR preset                        +6/-0      zh-CN.ts Add Simplified Chinese translations for custom ASR preset +6/-0      zh-TW.ts Add Traditional Chinese translations for custom ASR preset +6/-0
Tests	2 files advancedAsrConfig.test.ts Test advanced config parsing behavior                                        +64/-0    ProvidersSection.test.ts Test custom preset has no defaults                                              +13/-0

---

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 67fbb12)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 854 - Partially compliant Compliant requirements: Added a generic openai-compatible ASR preset with user-supplied endpoint/model/api key. Empty API key is allowed and the Authorization header is omitted to support unauthenticated LAN endpoints. Default request behavior is conservative: multipart, no response_format, no chunking. Added per-provider advanced options for verbose_json and chunk duration, stored in the credentials vault. Added Rust and frontend tests for parsing, serialization, and header omission behavior. Non-compliant requirements: The Foundry long-recording "only recorder callbacks, never transcribe" bug is not addressed. The overwriting of externally edited preferences.json / forced active provider behavior is not investigated or documented. The PR does not explicitly state whether LAN llama-server ASR is supported in the product docs, nor provide a roadmap. No new handling or test was added for the llama.cpp transcript.text.done response shape; compatibility with that shape is not explicitly verified. Requires further human verification: End-to-end verification on Windows against a LAN llama-server/Qwen3-ASR endpoint. UI configuration flow for the new preset, including save/commit order and whether the active provider is persisted reliably. Long-recording behavior with and without chunking against the LAN server. Whether the response parser handles {"type":"transcript.text.done","text":"..."} in practice.

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Possible Issue parse_advanced_asr_config accepts positive fractional chunkDurationMs values below 1 and floors them to Some(0). If the vault JSON contains {"chunkDurationMs":0.9} (for example via direct editing or a future client that does not pre-normalize), batch_asr_chunk_limit_ms returns Some(0), which is passed to split_pcm_by_duration and can cause a divide-by-zero or invalid chunking. The parser should reject values whose floored result is 0, e.g., by requiring ms >= 1.0 or checking ms.floor() > 0. The current UI normalizes before saving, so this only manifests with out-of-band JSON values. chunk_duration_ms: value.get("chunkDurationMs").and_then(|v| { v.as_u64().filter(|ms| *ms > 0).or_else(|| { v.as_f64() .filter(|ms| ms.is_finite() && *ms > 0.0 && *ms <= u64::MAX as f64) .map(|ms| ms.floor() as u64) }) }),

[github-actions]

Persistent review updated to latest commit f95baf4

[github-actions]

Persistent review updated to latest commit 67fbb12