Skip to content

v0.42.0.0: multicam, moment search, and categorized cleanup

Choose a tag to compare

@tomaslau tomaslau released this 13 Jul 16:50
· 94 commits to main since this release
69ee44a

Highlights

  • Contextual cam switch (multicam): ingest per-speaker cam files (openklip cam-add, up to 8 cams, speaker/wide roles, manual --offset), automatic speaker ID from per-track audio energy (no ML or cloud diarization), and mix down to a switched program with follow-speaker or LLM auto scene modes (openklip cam-mix), a synthetic wide shot when no physical wide cam exists, and locked manual overrides (openklip cam-override) that survive re-mix. CLI cam-add / cams / cam-set / cam-mix / cam-override; MCP cam_add / list_cams / cam_set / cam_mix / cam_override; GUI Config → Project Cameras section; templates/cam-mix/skill.md. Transcript words gain an optional speaker field (the attributed cam id) once a project has been through cam-mix.
  • Moment search: text + scene: find moments in the source footage by transcript text or visual content. Ingest indexes sample frames with a local CLIP model into working/moment-index.json (lazy backfill on older projects, or openklip index); a fourth left-rail Search tab (Mod+Shift+F) shows text and scene results as thumbnail cards, click seeks, and dragging a card (or its hover Keep button) restores any cut words in that span. Same engine on openklip search <slug> "query" and the MCP moment_search tool.
  • Categorized cleanup (Cutback parity): cleanup candidates now carry a category (hesitation, hedging, repeat, dead-air) via a new deterministic repeated-n-gram detector for immediate repeats and false starts; cleanup-config persists category toggles and thresholds to project.cuts.cleanup; cleanup-apply (CLI --apply-enabled) applies enabled categories at any risk plus all dead-air at the configured threshold. A dedicated Cleanup tab in the Config panel adds a real waveform silence card, per-category cards with bulk apply and undo, and an "AI false-start pass" through the subscription agent CLI.
  • Verification: bun run typecheck, bun run check, bun test --isolate (2342 pass, 8 skip); 98 MCP tools; 54 capabilities; 46 registry actions; bun run multicam-acceptance.

Known gaps

See TODO.md for the current gaps and known issues. Multicam programmatic acceptance (tests/multicam-acceptance.test.ts) satisfies the v0.42 machinery gate; optional human review on real per-speaker recordings remains deferred.

Full changelog: see CHANGELOG.md