fix: improve video agent research, finalization, and diagnostics

[devhims]

Video research can lose useful evidence or return an overly broad partial label when extraction, validation, or synthesis does not complete. This change improves provider configuration, source selection, transcript grounding, and run diagnostics while keeping the classification, research, and finalization phases separate.

• Default GLM classification, research, transcript analysis, and visual analysis to Fireworks, with DeepSeek finalization on every completion path. Retain the Workers AI provider option and native SDK output/reasoning controls.
• Have classification select research video count and storyboard availability. Distinguish an internal research target from an explicit user requirement, report coverage, and return the original stored request message.
• Improve source-specific names, numeric grounding, citation validation, and preservation of supported findings. Persist transcript validation and repair diagnostics, including bounded rejected output and relevant source windows, through authenticated run diagnostics. Keep captured content out of ordinary logs and default compact responses.
• Let the agent inspect storyboard metadata and choose sheets or timestamps instead of using a fixed two-sheet selection. Build the processor's extraction dependency from this checkout, with an allowlisted Docker context and corresponding CI setup. Cloudflare deployment does not require an npm package release.

Validation: 115 focused tests and 17 Durable Object integration tests passed on this exact working tree, along with TypeScript, generated OpenAPI checks, and git diff checks. The move onto the latest main preserved all changed files byte-for-byte. Earlier component and live-provider checks informed the implementation; their captures and reports remain local and are excluded from this PR.

Known limitation: a fetched transcript still becomes accepted evidence only after analysis succeeds. A repair canceled at the research deadline can therefore still return insufficient evidence. These diagnostics make future rejections inspectable; they do not recover missing diagnostics from historical runs or establish that all inference reliability issues are resolved. The latest classification, storyboard-selection, and diagnostic changes have not been deployed to production. Production must have FIREWORKS_API_KEY configured, and the processor image must be rebuilt for the storyboard changes.

[vercel]

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

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 11, 2026 5:19am UTC