v0.30.2: Improve background live transcription and current-speech reads

Minutes 0.30.2 reduces unnecessary live transcription work and hardens the local current-speech relay used by connected agents. It also keeps background live work declared to macOS while preserving the existing recording path and audio files.

What changed

• Keep live transcription work bounded. Complete audio jobs no longer run repeated recognition passes whose results were discarded. Stem bursts are split into 100 ms chunks before speech detection and utterance boundaries. The existing Whisper-Silero VAD uses one thread, with the same model and detection settings. #1144 #1146
• Declare active background live work on macOS. Standalone and recording-sidecar live processing hold a scoped user-initiated activity lease. The lease allows idle system sleep and ends when that live work exits. It does not change system power settings.
• Harden current-speech relay reads. Accepted Unix relay connections use an explicit I/O mode and idle bounds. Socket writes no longer hold the shared relay-state lock, and fragmented UTF-8 is retained until a complete JSON frame arrives. Authentication, owner/session identity, freshness, permissions and draft invalidation remain enforced.
• Report live evidence more clearly. Status retains the owned session ID and actual age of the latest final. minutes transcript --include-current --format json includes transcript status and relay errors so readers can distinguish available finalized text from an unavailable current draft. Delayed-final guidance is scoped to the same verified session.
• Use the matching local engine in Recall. New Recall children prefer the CLI bundled with the app. Input-stage diagnostics report exact sample coverage and elapsed stages without recording speech in those diagnostics. Stem resampling preserves fractional position across uneven poll boundaries.

Who should care

People using live transcription during macOS calls, Recall or other connected agents that read current speech, and CLI consumers of the include-current JSON output.

CLI / MCP / desktop impact

• Desktop: background live-work declaration on macOS and bundled CLI resolution for new Recall sessions. Windows desktop remains experimental and unsigned.
• CLI: no new commands. Include-current JSON adds transcript status and an optional relay error field.
• MCP and agent integrations: no new tools or engine/language changes. Current drafts remain provisional and must pass the existing session, age, gap and revision checks before use.

Breaking changes or migration notes

No configuration or stored-meeting migration is required. JSON clients should tolerate the additional include-current fields. Update the desktop app manually, then reopen Recall to start a child with the updated bundled CLI.

Known issues

The client does not yet have a complete end-to-end handshake/replay deadline. The new Unix server limits are idle I/O bounds. An unavailable current draft remains unavailable evidence even while recording and finalized live text continue.

A signed development-app test processed a controlled 35-minute public-audio run while the app was hidden, with 11 healthy relay checkpoints and no dropped/skipped utterances or recognition failures. This does not establish natural-call, multilingual accuracy or battery-endurance acceptance, and does not prove App Nap caused the earlier stalls. Recording preservation and live reader availability remain separate checks.

Install or update

• macOS desktop: download the DMG below, or run brew update followed by brew upgrade --cask silverstein/tap/minutes.
• Command line: cargo install minutes-cli --version 0.30.2.
• MCP server: npx minutes-mcp@0.30.2 with the updated local Minutes engine.

Minutes remains free and open source. Meeting files stay local; content retrieved by a connected cloud assistant becomes context for that assistant under its settings.