[vinayagg]

Desktop installers currently carry LibreOffice, Sherpa/ONNX speech libraries, and uv. Download these components in the background after first launch, then install managed Python and Hindsight without waiting for a local AI model. Chat startup stays independent, and Node.js remains bundled. This removes approximately 220 MiB of uncompressed files from the Mac ARM installer payload.

Publish a support archive for each platform and pin its URL, size, and checksum inside the app. Verify downloads before extraction, install atomically into private Client storage, reuse cached files offline, and retry failed setup. Office and speech loaders resolve their engines from that cache; development and headless builds retain bundled tools. Mac support packages are signed and notarized by the release build.

Validation:

• npm run check passed, including download corruption, interruption, concurrent installation, cache repair, and offline reuse tests.
• cargo check --manifest-path src-tauri/Cargo.toml passed after staging the reduced runtime.
• Startup and agent workflow checks passed against the reduced packaged runtime.
• A real downloaded LibreOffice engine converted Word to PDF offline.
• The actual speech worker loaded downloaded Sherpa/ONNX libraries offline and resampled test audio.

Production notarization, Windows/Linux installer execution, full voice transcription, and a complete Python/Hindsight installation were not exercised locally. First setup requires internet access.