What's new in RambleDesk 0.3.2
Speech recognition
- SenseVoice is now the recommended default model for reliable multilingual transcription. X-ASR remains available as the lower-priority streaming option.
- Existing rc.7 users who only inherited the old X-ASR default are migrated once to SenseVoice; an explicit later X-ASR selection is preserved.
- Settings and onboarding now show the same recommended model, ordering, descriptions, and download actions.
Ramble workflow
- /ramble [task] now consistently starts a task-scoped feedback loop across Pi, DeepSeek Harness, and Generic MCP hosts, while /ramble_on remains the explicit persistent-mode switch.
- A bare /ramble or a generic starter uses the active conversation when possible, otherwise gathering the goal, context, constraints, desired output, priorities, and completion criteria inside RambleDesk.
- Generic MCP and dsh now install one shared capability-aware skill, and onboarding uses natural English and Chinese starters so the agent begins in the user's language.
Feedback reliability
- Pi and DeepSeek Harness waits no longer disconnect because of Node/Undici's response-header timeout; interrupted flows keep the same durable request id for recovery.
- Generic MCP calls are now stateless, so a long human Ramble or stale transport session cannot hide an already-completed feedback request.
- Ramble guidance now explicitly distinguishes the durable request id from disposable MCP transport state and prevents duplicate replacement requests.
中文摘要
- 语音识别：SenseVoice 提升为推荐默认模型，X-ASR 保留为低优先级流式选项；rc.7 继承旧默认值的用户会一次性迁移到 SenseVoice，后续手动选择 X-ASR 不会被覆盖；设置与新手引导同步展示推荐状态和模型顺序。
- Ramble 工作流：Pi、DSH 与通用 MCP 的 /ramble [任务] 统一为任务级反馈循环，/ramble_on 专门开启持续模式；裸 /ramble 或通用开场语会优先利用当前任务，否则在 RambleDesk 内收集完整任务简报；通用 MCP 与 dsh 共用同一份能力自适应 skill，新手引导使用自然的中英文启动语以保持 Agent 回复语言。
- 反馈可靠性：Pi/DSH 的人工等待不再因 Node/Undici 响应头超时而断开；通用 MCP 改为无状态调用，长时间 Ramble 或陈旧 transport session 不再影响使用原 request_id 读取结果。
Full changelog: v0.3.1...v0.3.2