Fix video agent API evidence collection and answer quality

[devhims]

Problem

A production request for “suggest the top use cases of gpt 6 astra” exhausted the research window without usable transcript evidence and returned promotional discovery text as a partial answer. Local reproductions also exposed extra search planning, transcript selection in separate waves, and invalid structured model output.

Changes

• Execute the classifier's search query immediately, preserving the one-search budget and compatibility with persisted routes.
• Batch selected transcripts with independent persistence and billing. Allow two concurrent analysts for focused research and four for comparative research; inspection retains two.
• Return an explicit NO_CONTENT_EVIDENCE fallback when content analysis fails, instead of synthesizing recommendations from titles and descriptions. Track transcript attempts synchronously so deadline aborts cannot race failure recording.
• Guide analysts and synthesis toward three practical, source-attributed recommendations, with relevant examples and caveats.
• Expose classifier and answer tool parameters as top-level objects while retaining conditional route and citation validation.
• Document scheduling, fallback behavior, and local quality checks.

The 60-second deadline and 20-second finalization reserve remain. The proposed public response redesign is separate from this fix.

Validation

• 334 unit tests across 50 files passed, including batch concurrency, partial failures, all-transcript timeout recovery, one-search enforcement, and conditional schema validation.
• 10 runtime integration tests passed.
• TypeScript and generated documentation checks passed; diff whitespace checks passed.
• Local live GLM research on the exact reported query analyzed four videos and returned a three-option answer in approximately 51 seconds, without partial-evidence or coverage-shortfall warnings.
• Local live GLM inspect regression completed in approximately 39 seconds with transcript and storyboard evidence.

The final timeout race guard was validated with a regression test after the live runs. These timings demonstrate successful local runs, not a latency guarantee. Source claims remain attributed rather than independently verified. Production verification follows merge and deployment.

[mintlify]

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

Project	Status	Preview	Updated
video2ctx	🟢 Ready	View Preview	Sep 5, 2026, 5:44 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 5:43pm UTC