v0.4.3

New features

• add-transition / batch-add-transitions accept audio tracks —
  trackType: "audio" was rejected on the assumption that no audio
  transition API existed; the QE DOM has addTransition on audio clips
  (resolved via getAudioTransitionByName), so a crossfade can now go on
  every cut of an audio track. list-available-transitions /
  list-available-audio-transitions also now return the real lists (110
  video, 3 audio) instead of zero — the QE lists are plain string arrays.
• add-background command — fills given time spans of a video track
  with a looped background clip, trimming the final repeat so each span is
  covered exactly. Repeats are quantised to whole frames on both the
  sequence's and the source's frame grid, so mixed 25/30/50 fps material
  tiles gap-free. --dry-run reports the plan; the source item's in/out is
  restored afterwards. Loud warning: a background with an audio stream
  destroys the audio underneath it — strip the audio from the file first.
• sync-assets command — imports every media file under an assets/
  tree into bins mirroring its subdirectories. Existing items with the same
  name are relinked rather than re-imported; hidden files and AppleDouble
  ._* sidecars are skipped.
• insert-clips command — places every clip in a bin whose name begins
  MM-SS-FF onto a given track at that timecode; items without a leading
  timecode are reported, not guessed, and an out-of-range frame field is
  rejected.
• delete-all-markers command — clears every marker on a sequence or
  project item in one round trip, instead of one delete-marker call per
  guid.

Bug fixes

• Transition durations were mis-sized. The QE addTransition duration
  argument is SS.FF (seconds.frames), not decimal seconds, so 0.08 gave
  an 8-frame transition at 25 fps. Durations now go through the
  seconds.frames form, video and audio alike.

Documentation

• remove-track-intervals silently desyncs linked A/V clips — on
  linked pairs it slips audio source in-points while leaving timeline
  positions, duration and coverage correct, so no health metric sees it.
  Unlink → apply → relink, verify every pair's
  audio.inPointSeconds − video.inPointSeconds against the sync offset,
  and repair a slipped clip in place with trim-clip --in-point-seconds.
• There is no API way to place video without its audio — the -1
  track index does not do it on Premiere 2026; the source's audio lands on
  A1 regardless of targeting and overwrites what was there.

v0.4.2

Bug fixes

• desktop-take-screenshot captured its own status overlay — the overlay is a real always-on-top window pinned top-right of the screen, and screencapture -x captures the whole screen including it, so its "driving" banner ended up baked into the corner of every screenshot. take_screenshot now hides the overlay immediately before capturing (instead of showing "driving") and only re-shows it afterward, once the frame is already safely written.

v0.4.1

Bug fixes

• desktop-* commands could capture/act mid window-swap — desktop-take-screenshot (and the other desktop-* primitives) confirmed Premiere was frontmost via the window server and immediately proceeded, but the actual window-swap animation/redraw can still be in flight at that instant — a screenshot taken right then could show a sliver of the previously-frontmost app. _require_frontmost now waits 0.5s after confirming activation before proceeding, but only when Premiere actually needed to be brought to front — skipped when it was already frontmost, so repeated calls while Premiere stays active aren't slowed down.

v0.4.0

New features

• Generic desktop-* UI-automation commands (macOS only, optional
  macos-desktop extra): desktop-take-screenshot, desktop-press-key,
  desktop-enter-text, desktop-enter-text-with-validate,
  desktop-move-mouse, desktop-click-mouse — raw primitives (screenshot,
  key press, text entry, mouse move/click) for driving Premiere's native UI
  via the Accessibility API + synthetic key/mouse events, for actions
  ExtendScript can't do on this build. Each confirms Premiere is frontmost
  before sending input. Also desktop-notify / desktop-dismiss- notifications — a persistent, always-on-top notification window via a
  separate daemon process, so it can be updated in place or dismissed by
  later, separate CLI invocations. See
  docs/DESKTOP_DRIVER_NOTES.md.

  (An earlier desktop-set-input-lut command, built the same day on
  hand-written AX-tree navigation specific to Premiere's Lumetri "Input
  LUT" control, was removed before release in favor of rebuilding that
  capability on top of these generic primitives plus AI vision-language
  screen understanding.)

v0.3.1

Bug fixes

• export-frame was silently exporting the wrong frame — qe.project.getActiveSequence().exportFramePNG ignores its position argument on this Premiere build regardless of arg-order (ticks string, Time object, output-path-as-position, ...), always exporting whatever frame happened to be currently rendered in the Program Monitor, while still writing a real file and returning success. The command now exports exclusively via Adobe Media Encoder's exportAsMediaDirect(), narrowing the sequence's in/out points to a single frame around the requested timecode — verified live via pixel diff that exports at different timecodes actually differ and match the requested position.

v0.3.0

New features

• init-project command — creates a fresh empty Premiere Pro project from a bundled template

Improvements

• Renamed the CEP extension ID to com.stefanwebb.premierecli

Infrastructure / Documentation

• Moved the Claude Code plugin into the separate premiere-ai-skills repo