Unify agent finalization and preserve conversation history

[devhims]

Summary

Follow-up requests could lose earlier conversation messages during finalization, analysis, and tool repair. A request to correct a previous answer could therefore claim that the earlier answer was unavailable.

Classification now chooses research, inspection, or direct finalization. Direct requests and completed research use the same finalizer to answer from existing context, ask for clarification, or decline. The finalizer receives prior messages and can cite retained source evidence from earlier turns.

• Keep the latest eight completed turns without the 64,000-character history cap, and pass history to analysts, finalization retries, and tool repairs.
• Reuse previously cited evidence within the selected conversation ancestry, with citation validation and no additional provider charges for reuse.
• Preserve session prompt-cache affinity for the finalizer and keep retry instructions stable. Cache hit rates have not been measured against the live provider.
• Retain compatibility with persisted clarification and rejection routes, and update the API schema and architecture documentation.

Validation

• Platform suite: 593 passed, 6 opt-in live tests skipped.
• Durable Object integration suite: 40 passed.
• Platform TypeScript build passed.
• Generated API documentation check passed.
• git diff --check passed.

Regression coverage includes direct finalization, research finalization and repair, analyst history, eight long conversation turns, retained evidence citations, user isolation, and finalizer cache affinity.

These changes have not been deployed.

[vercel]

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

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 18, 2026 11:03am UTC