samuelgursky/davinci-resolve-mcp

v2.61.0

Perception strata — a timecoded track model over every analyzed clip, plus the query layer that turns it into editorial answers. Ten new media_analysis actions, all local compute (ffmpeg + numpy; the face tier detects opencv + mediapipe and degrades honestly when absent). Everything measures, compares, flags, or aims — nothing decides; the cut is the editor's.

Added

• Schema v13: perception strata — two generic track shapes instead of a table per sensor: events (pause, breath, hesitation, blink, beat, downbeat, …) and curves (pitch, vocal_energy, speech_rate, motion_energy, face curves as float32 series), plus two promotions out of the report blob: transcript_words (per-word timestamps as queryable rows) and story_beats (units of meaning over the transcript). Machine re-runs replace their own rows; source='human' rows are append-only and always win.
• Local analyzers (strata_run) — prosody, beat grid, motion energy, and face strata (blink/gaze/expression via opencv + mediapipe when installed). strata_status reports the per-clip track inventory and what this machine can run; backfill_words promotes word timestamps already sitting in stored report blobs without re-analysis.
• take_diff(clip_a, clip_b, text?) — align two deliveries of the same line on transcript words and diff pace, pauses, pitch, and energy. Deltas only; it never picks a winner.
• cut_candidates(clip_id, time_seconds) — the joint solver: scores every frame around an intended cut on the cut-point grammar (cut on the blink, don't cut mid-word, don't bisect a breath, pauses are doors, land on the beat, cut inside movement) with human-readable reasons and honest reporting of missing tracks.
• strata_query — windowed cross-track bundles per clip, or project-wide word find with a joined ±context bundle per hit. timeline_strata projects clip strata through a versioned timeline's recorded placements.
• Story beats (host-LLM pass) — plan_story_beats assembles a timecoded digest + JSON schema (the server never calls an LLM), the host chat produces the beats, commit_story_beats validates and persists them append-only; list_story_beats reads them back.
• Schema v14: timeline timebase snapshot — timeline_versions records the timeline's fps and start frame at archive time, so timeline_strata returns timeline-relative frames/seconds instead of assuming 24fps against absolute (start-timecode-inclusive) snapshot frames.

Fixed

Pre-release review (multi-agent: 8 finder angles + adversarial verification) caught these before anything shipped:

• Re-ingest cascade wipe — the ingest INSERT OR REPLACE on the clips row cascade-deleted every clip-child table under PRAGMA foreign_keys=ON; invisible while all children were rebuilt during ingest, destructive once the strata tables (not rebuilt by ingest, human annotations included) joined the cascade. The clips row is now a true upsert and is never deleted.
• Words-less re-analysis wiped word rows — ingest_transcript_words deleted before validating; it now parses first, preserves existing rows when the incoming report has no words, and falls back per-segment to the top-level words list for mixed reports.
• Machine story-beat commits could replace human rows — the content-hash beat_uuid let INSERT OR REPLACE overwrite a human beat sharing the same span+label. Beat rows now get random UUIDs and plain INSERTs; duplicate spans within one commit are skipped and reported.
• Dispatch hardening — stringified numbers no longer crash strata_query or silently redirect timeline_strata to the latest version; cut_candidates rejects non-positive fps instead of ZeroDivisionError; word matches escape SQL LIKE metacharacters; clip bundles fetch every recorded track instead of a hardcoded subset; analyzer ffmpeg calls go through the stdio-safe proc.safe_run; a broken mediapipe/protobuf install degrades to "face unavailable" instead of crashing strata_status.

Validation

• Full offline suite: 1,418 tests green (17 new regression tests covering the fixes above). Static checks, drift guards, API-parity audit, and npm pack all pass.
• No live Resolve behavior changed beyond two defensive getter reads (GetSetting("timelineFrameRate"), GetStartFrame()) at archive time, both try/except-guarded and unit-tested against mock timelines.