Expand video agent API research breadth with parallel analysis

[devhims]

Summary

Recommendation queries such as “Research the best design skills for frontend developers using Claude Code” previously had a fixed two-video transcript budget. Classification now selects focused research targeting two videos or comparative research targeting four. New classifications use a required structured tool call; older persisted routes retain the two-video behavior.

Selected transcript calls run together with at most two transcript or visual analysts active at once. Queued analyses check cancellation before starting. Research planning uses low reasoning effort to leave more time for evidence collection. The one-search limit, 60-second total deadline, and 20-second finalization reserve remain in place. Answers disclose a RESEARCH_COVERAGE_SHORTFALL when fewer distinct videos provide usable transcript evidence.

Live testing also exposed a finalization failure when a comparison block exceeded five evidence references. The bounded limit is now twelve, with persisted-reference validation unchanged. Schema failure logs include validation paths and codes without model text.

Validation

• TypeScript build and documentation generation/check passed.
• 324 unit tests and 10 local runtime integration tests passed.
• Final local GLM research run: four distinct transcript analyses, 12 citations, approximately 57 seconds, six credits, no partial-answer or coverage-shortfall warning. Run: a3d69728-6e34-4f4d-84af-0d6dd28773d8.
• Final local GLM inspect run: transcript and storyboard evidence, 11 citations, approximately 33 seconds, three credits. Run: 5568af05-e2d3-4d6f-9ee0-c818093d3970.
• The live smoke test now requires four analyzed videos for its default recommendation query.

Limits

Four videos is a target, not a guarantee. Earlier local runs missed the target because provider/model latency consumed the collection budget. The coverage warning reports this explicitly. Earlier testing also caught classifier JSON-output issues and the citation limit failure addressed here. The final smoke run passed both routes; it does not establish a reliability rate.

The non-admin live check was skipped because no non-admin test token was supplied. No production deployment or production test was performed for these changes.

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated
video2ctx	🟢 Ready	View Preview	Sep 5, 2026, 2:48 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

[vercel]

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

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 5, 2026 2:47pm UTC