Fix classification repair for conversation history requests

[devhims]

Summary

Asking to list the user messages in an existing conversation failed during classification, before the context finalizer ran. The classifier used automatic tool selection for both attempts, so a response without a routing call could repeat on repair and fail with Invalid fields: tool call.

Require the classify_request tool on repair when the first response lacks exactly one routing call. Keep automatic selection for the initial attempt and argument-only repairs because forcing every call previously produced incomplete arguments. Continue validating the full routing schema; ordinary prose is never accepted as a decision.

Explicitly support requests to list, quote, summarize, or correct conversation messages. Tell the finalizer to list available user messages chronologically, including the current request, and disclose the eight-completed-turn context boundary. Add a distinct diagnostic for missing or multiple routing calls.

Validation

• Regression test reproduced the exact error before the fix and passes afterward, exercising a prose response followed by a required routing call.
• 594 local tests passed.
• Platform TypeScript build and git diff --check passed.
• Added an opt-in live classifier regression. The live attempt was blocked by automatic approval review before execution, so provider behavior has not been verified.

The two production failures were confirmed in Chrome. Raw production model outputs were unavailable, so the test reproduces the missing-tool-call failure mode rather than claiming to replay a captured provider response. Changes are not 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 11:40am UTC