[ampersandru]

Adds approved native computer control and an isolated browser to AI chat, with optional local decision/speech models and settings that make their behavior and resource use visible. Works with configured local or remote chat servers.

Features and behavior

• Computer/browser tools: installed-app discovery/launch, bounded observations, native accessibility actions and browser DOM control. Actions use Allow once / Allow this session / Deny; New Conversation and Stop revoke control. Normal launched apps can remain open after tools stop.
• Browser passwords: fill observed password/current-password/new-password fields and submit them through existing approvals. Use the browser-only password argument. Permission previews and DOM values are masked; known reflected text and active debug transcript values are redacted. Password fills skip Qev scoring while retaining consent and fresh-target checks. Native desktop password fields remain unsupported.
• Qev: optional Local/Remote provider, master enable switch, Off/Suggest/Execute modes and optional vision. Unique grounded targets avoid redundant vision capture. The main chat model still plans the task; Qev classifies targets. Local Qev unloads after the request or app exit.
• Local speech: optional Paradee v2f lite/v2f/v2g on CPU and Kokoro FP16/Q8 on CPU or DirectML GPU, plus retained Remote settings. Read Aloud preloads on submission; immediate or idle unloading and owned-process shutdown apply. Ordinary chat can speak completed sentences during generation, with remote streaming/buffered fallback.
• Readable speech: inline links/emphasis preserve adjacent words and punctuation. Hidden destinations and definition-only chunks are not spoken; streamed reference labels work before definitions arrive.
• Settings: searchable labels/aliases on every panel, current-panel results first, section navigation without editing preferences. Hardware status lists runtime, embedding, Dictation, Speech, encoder, OCR, FFmpeg, PDF and RAW in order. Remote Speech matches Dictation's readiness styling. Local speech reports worker RAM, dedicated VRAM, selected weights and total speech storage, with scoped install/uninstall controls.
• Diagnostics: opt-in per-conversation prompt/response/tool transcript with numeric token/cache/Qev diagnostics. Images are omitted. Transcripts remain private data; chat credentials are still processed by the configured chat provider.
• Capture: restore the prior full/popup AI window after area capture, optionally copy captures to the clipboard, and support image drops into Search, Clipboard and AI chat.

Reliability fixes

Fresh observations and stable target identities are checked before input. Native traversal is paginated and includes a focused editor outside its current page. File-row typing and clipboard operations in active rename editors are rejected, selection receives focus, and actions report performed state consistently. Completion checks reject unverified success/unfinished narration; bounded progress checks limit repeated unchanged actions. Follow-ups retain useful verified context, while diagnostics expose cache changes and orchestration overhead.

Validation and limits

425 unit tests passed (25 optional fixtures ignored). Real Paradee linked-text generation, streamed/buffered linked-text GUI checks, a localhost password login (including Enter submission, autocomplete fields, stale/normal-target rejection and masked readbacks), denied-password handling and transcript redaction passed. Earlier branch checks also cover all five CPU speech choices, both Kokoro DirectML models, resource sampling/unloading, settings ranking/navigation and scoped speech removal.

Windows foreground policy blocked some unattended pointer/Explorer probes; the full real Explorer workflow still needs interactive retesting. Qev did not demonstrate a speedup in the measured local-server comparison: see benchmark. Current-head changes are supplied as test builds; beta.2 is an earlier published checkpoint and its assets remain unchanged.

Credits

• Qev by labqoat, for optional decision classification.
• Paradee by Sahil Mahendrakar, for upstream Kokoro distillation research/code. The optional v2 variants use Mesmer.tools' supplied PDN v2 bundle, distinguished from the upstream v1 releases, with bundle-specific notices retained.
• Kokoro, llama.cpp, ONNX Runtime, Node.js, Misaki and eSpeak NG; notices remain with optional runtime downloads.

See docs/COMPUTER-USE.md, docs/LOCAL_TTS.md, docs/DECISION_ROUTING.md and docs/SETTINGS_SEARCH.md for implementation boundaries and test instructions.

Final 0.3.44 update

• Add all ten local Paradee voices from the October 27 Mesmer release, including the nine Gradium Phonon students. The voice selector downloads one selected voice plus the shared engine, with pinned checksums and upstream notices. Heart retains all three model variants.
• Reload correctly when switching voices, apply publisher pause/loudness settings, and show the selected voice and its storage in Status.
• Verified actual WAV synthesis and immediate unload for all ten voices, and repeated voice switching in a retained worker. 426 unit tests pass; Rust formatting and release-gate regressions pass.
• Prepare stable version 0.3.44; notes describe features and fixes since stable 0.3.42.