Simplify single-video transcript inspection and validate storyboard selections

[devhims]

Single-video inspection could retrieve the same captions repeatedly with different focus questions, while isolated transcript analysis omitted details needed for the complete answer. Inspection now retrieves timed captions once per video and language within a run and lets the main model read them directly. Multi-video research keeps its existing isolated analysts.

Storyboard image calls now require metadata for the same video and validate sheet indexes, timestamps, and selection budgets against its persisted manifest before invoking the provider. Invalid selections return correction guidance without extraction or vision calls.

Complete transcript excerpts are preserved through recovery and finalization when they fit the existing 40,000-character finalization evidence budget. Oversized transcripts receive an explicit truncation warning. This does not add cross-turn transcript caching or change visual analysis.

Validation:

• Regression cases reproduced six failures against the previous implementation and passed with these changes.
• Platform unit suite: 604 passed, 7 skipped; subsequent storyboard coverage: 29 passed, including an additional unrelated-video manifest case.
• Durable Object integration: 41 passed, including persisted reuse and no duplicate retrieval charge.
• Processor tests: 29 passed. Frame tests: 16 passed.
• Platform TypeScript build, generated-doc consistency, and diff whitespace checks passed.

Tests use synthetic fixtures and mocked inference; no new production inference benchmark is included.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 18, 2026 3:20pm UTC