[lanyue-llk]

Adds a video-gen skill for script, storyboard and narrated-video work, with guidance on source evidence, visual relationships, speech alignment and playback review. A small, renderer-independent timeline validator checks scene and caption timing, named focus targets and retained source material.

Film-specific narration, diagrams, Remotion projects and production tests belong with each film's external production sources. The PR now contains only the reusable skill, its discovery entry and synthetic behavior tests under the existing tests/ directory. The normal Python CI runs these tests without a dedicated video workflow or additional dependencies.

Validation: 35 targeted tests passed, covering timeline behavior, CLI exit status, vendored skill integrity and the existing base CLI. Static checks and skill validation passed. No published video or report was changed.

[lanyue-llk]

Review conclusion

Fail: the visual review command can succeed after detecting an invalid render.

The PR adds a narrated Bright Data explainer with a shared clock for captions, visual emphasis and chapter navigation. Keeping those consumers on one timeline is appropriate. The changes are isolated from the existing benchmark runtime; the defects below affect the new production and validation path.

Findings and recommendations

• [P1] research/brightdata/video/remotion/review.mjs:172-185: detected overflow and missing emphasis do not fail the review command. See inline reproduction and recommendation.
• [P2] research/brightdata/video/timing.py:30-31: normalization makes 1.5 and 15 identical, allowing changed quantities through validation.
• [P2] research/brightdata/video/timing.py:114-116: omitted English words can pass both clause and anchor validation.

Local validation: 32 Python tests discovered, 29 passed and 3 production-input tests skipped; TypeScript compilation passed. Additional probes reproduced both alignment defects, and real Chrome reproduced the visual check failure. These are code-level findings, not claims that the published film contains these mistakes.

For the smallest maintainable change, repair the existing alignment and visual checks and add negative controls for these scenarios. Names and directory placement do not need a separate refactor. One useful cleanup, consistent with AGENTS.md, is to move report-specific dates, customer copy and numeric claims out of Film.tsx into the existing external content model while keeping diagram geometry and rendering components in code.

[lanyue-llk]

Review conclusion

The film-specific production code discussed in the earlier review has been removed from the PR's final tree in d2bb269. The PR now adds the reusable video-gen skill, its timeline validator and standard pytest coverage, with six files in the final diff.

The three scope comments are addressed: the dedicated video workflow is removed; generic tests live in tests/test_video_timeline.py; the Bright Data renderer, speech pipeline and report-specific sources are kept outside the maintained repository tree. Earlier production sources remain recoverable at 16a7e43, and published media is unchanged.

Validation: all 13 synthetic timeline cases passed without private inputs or skips. Together with existing skill-integrity and base CLI checks, 35 targeted tests passed. Static checks and skill validation passed.