VectorDash 0.3.44-beta.2

VectorDash 0.3.44-beta.2

This prerelease adds optional local speech and decision models, expands control diagnostics, and improves computer/browser control. It is built from PR #21; the PR remains open.

New features

• Local speech: optional Paradee v2f lite, v2f and v2g downloads for CPU, plus Kokoro 82M FP16/Q8 for CPU or DirectML GPU. Models and runtime downloads remain optional.
• Speech streaming: ordinary AI chat can begin read-aloud from completed sentences while the answer is generated. Remote audio streaming is used where supported, with buffered fallback. Streaming can be disabled. Computer/browser responses retain completion validation before speech.
• Model lifecycle: preload local speech when sending with Read Aloud selected; choose immediate unloading or an idle timeout. Owned local workers shut down with VectorDash.
• TTS resource status: Hardware → Status shows the active model/device, worker RAM, Windows-reported dedicated VRAM, selected model-weight size, and storage for all installed speech files. Shared runtime storage is included once; unavailable GPU counters are labelled.
• Qev decision model: optional local or remote target scoring for computer/browser use, with Off, Suggest and Execute modes and optional vision. A separate enable switch hides mode selectors and bypasses Qev when disabled.
• Local/Remote dropdowns: decision and voice settings retain their remote connection details when switched to Local.
• Combined control debugging: opt-in per-conversation files include prompts, responses, actual tool steps and per-step token/cache/Qev diagnostics.

Fixes

• Keep focused editors available outside the current accessibility page, including long Explorer search and rename views.
• Reject typing directly into file/list rows. File renaming requires an actual editor; clipboard shortcuts in active inline rename fields are rejected rather than acting on filename text.
• Focus selected list items for subsequent keyboard actions, distinguish typing from committed renaming/navigation, and support F5 refresh and Ctrl+E search.
• Report browser actions consistently to completion checks and reject short unfinished control plans as successful answers.
• Avoid redundant vision capture for uniquely identified Qev targets while retaining probability, consent and observation checks.
• Fall back from unsupported remote speech formats, including plain-text errors, and keep streaming/buffered capability caches separate.

Validation and limits

416 unit tests passed (24 optional tests ignored). Actual audio generation was checked with all five CPU model choices and both Kokoro DirectML models, together with cancellation/unload/shutdown behavior and resource sampling. Disposable native-control and explicit Chrome browser checks passed. Ordinary-chat streaming was checked against a delayed localhost answer, including buffering and format fallback.

Qev is a target-scoring assistant, not a replacement for the main chat planner. The current local-server comparison did not demonstrate an end-to-end speedup; see measured results. Windows foreground restrictions prevented a complete automated retest of the real Explorer copy workflow; test file operations in a disposable folder.

Credits

• Qev by labqoat, for the optional decision classifier.
• Paradee by Sahil Mahendrakar, for upstream Kokoro distillation research and code.
• Mesmer.tools, for the supplied Paradee v2 PDN web bundle used for the v2f lite/v2f/v2g options. Bundle-specific license and frontend notices are retained.
• Kokoro, llama.cpp, ONNX Runtime, Node.js, Misaki and eSpeak NG, with their upstream notices retained in the optional runtimes.

PR #21 · Local speech details