Transcription fix — MovieStar v0.7.1

Transcription fix — MovieStar v0.7.1

MovieStar v0.7.1 restores transcription for new installations. It contains no
other behavior changes from v0.7.0.

What's fixed

PyAV 19, released on 2026-09-29, removed an argument that faster-whisper 1.2.1
passes when it opens audio. Fresh installations of v0.7.0 resolved PyAV 19, and
every transcription — moviestar load without --no-transcribe,
moviestar retranscribe, and caption generation from new sources — failed
with:

Transcription failed: open() got an unexpected keyword argument 'metadata_errors'

v0.7.1 requires av<19, so installations resolve a compatible PyAV release.
The cap will be lifted once a faster-whisper release supports PyAV 19; upstream
tracks this in
SYSTRAN/faster-whisper#1492.

Upgrade

pip install --upgrade "moviestar==0.7.1"
moviestar doctor

If an existing environment already has PyAV 19, upgrading MovieStar downgrades
PyAV to a compatible release. Environments that installed v0.7.0 before
2026-09-29 were not affected.

Projects, specs, and transcripts from v0.7.0 need no changes. MovieStar v0.7.1
is licensed under the Apache License 2.0.

A video editor for AI agents — MovieStar v0.7.0

MovieStar gives AI agents a deterministic, non-destructive way to turn raw
footage into finished video. Agents can inspect media, make precise edits,
compose scenes, mix audio, add captions and overlays, and verify the result
through one structured command-line interface.

Explore MovieStar at trymoviestar.com.

Watch the MovieStar demo on YouTube →

What's new in v0.7.0

MovieStar is now open source

The complete Python package, test suite, agent-facing documentation, and
portable MovieStar skill now live together in the public
jdilla1277/moviestar repository
under the Apache License 2.0. Agents and developers can inspect the entire
editing pipeline, propose changes, and build directly on the same tools shipped
through PyPI.

Build transitions into the timeline

Authored scene transitions persist in project state and render through the same
deterministic timeline used by screenshots, previews, and exports. An agent can
plan a transition, inspect the resolved edit, and verify the finished output
without switching mental models.

Correct captions without losing intent

Caption rules can replace phrases, split tokens safely, and report which
current captions they affect before an agent commits to a reload. That makes
repeatable cleanup possible without hand-editing generated subtitle files.

Connect an optional account safely

A human can approve a named MovieStar installation through the website while
the CLI stores only its revocable device credential. The agent never receives
the sign-in link, and local editing remains fully available without an account.

Keep the v0.6 editing foundation

The geometry, camera motion, silent-footage activity analysis, environment
diagnostics, and result-time audio verification introduced in v0.6.0 remain
available in this release.

Try MovieStar

pip install --upgrade "moviestar==0.7.0"
moviestar doctor

Then visit the getting-started guide for a first
verified edit, or install the
MovieStar skill to give a compatible coding agent the
recommended workflow directly.

License and upgrade notes

MovieStar v0.7.0 is licensed under the Apache License 2.0. Versions through
v0.6.0 remain under the Business Source License 1.1; the new license does not
retroactively relicense those historical packages.

The public repository begins with the v0.7.0 codebase rather than importing
private development history. Existing v0.6.0 projects remain compatible, but
MovieStar is still pre-1.0, so review the structured output of an edit before
rendering.