fix: bound video agent phases and enforce native output budgets

[devhims]

The video agent now gives classification, research and finalisation separate persisted deadlines, rejects tasks outside supported YouTube synthesis, and uses native structured output and SDK token ceilings for final answers. Compact transcript notes and concise answer guidance reduce generation work while retaining source references and caveats.

Changes

• Bound classification to 20 seconds, research to 40 seconds and finalisation to 40 seconds. Recovery reuses persisted deadlines, and the watchdog respects the active phase. Queueing and the separate 30-second persistence window remain outside these model budgets.
• Let classification reject unsupported tasks, select storyboard access and choose standard or detailed answer budgets. Preserve legacy route compatibility and expose rejected outcomes in the API contract.
• Compact transcript analyst output to five findings and up to three warnings with a 1,200-token ceiling. Keep full-transcript reading and exact excerpt resolution; generate the compatibility summary in the application.
• Use native JSON-schema output for reserved finalisation, with citation requirements visible in the schema. Repair invalid structure, references or token-limit termination once within the existing deadline.
• Apply AI SDK maxOutputTokens of 1,500 for standard answers and 2,500 for detailed answers across answer-capable research generations, reserved finalisation and terminal-tool argument repair. Remove redundant final-answer length/schema instructions.
• Add regression coverage and update OpenAPI and architecture documentation. Keep experiment reports, benchmark fixtures, source snapshots and response captures local and ignored by Git.

Validation

• TypeScript, generated OpenAPI, and Mintlify build/link/accessibility checks passed.
• 132 focused unit tests across six files and 16 AgentRuntimeDO integration tests passed.
• Offline replay of the recovered original Workers AI response confirms the provider adapter and AI SDK preserve its content exactly.
• No full local test suite or new live inference run during PR preparation. Experiment artifacts remain local and are excluded from this PR.

Remaining limitations and deployment

Native token limits and schema validation do not guarantee that every answer fulfills the requested scope or finishes before its deadline. Detailed-report timeouts and schema-valid incomplete answers remain unresolved. The recovered upstream response rules out client-side truncation for the investigated failure, but model generation versus provider serving remains unresolved.

Requested-count enforcement and inference reliability remain follow-up work. An alternative model provider is explicitly deferred.

The latest SDK-budget changes have not been manually deployed. Cloudflare's connected build pipeline watches main, so merging this PR will trigger verification and production deployment. This PR does not claim that every request completes successfully within 100 seconds.

[mintlify]

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

Project	Status	Preview	Updated
video2ctx	🟢 Ready	View Preview	Sep 8, 2026, 2:16 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 8, 2026 2:23pm UTC