Fix session history retrieval and incomplete final answers

[devhims]

Conversation-history requests could leave direct finalization for an irrelevant video lookup, and schema-valid non-answers such as "The" or a promise to look up history were accepted as completed. This change gathers session context before a separate structured-answer call, prevents history-only requests from escalating to video inspection, and verifies exact-first-message answers against the original stored text.

Known fragment-only and promise-only answers get one repair before failure. These are narrow checks, not a general guarantee of semantic completeness. Saved-frame follow-ups now identify reused images in the activity trace while still allowing analysis for a new visual question.

Context gathering is capped at four model steps, followed by an answer call and at most one repair under the existing deadline and cost limit. Separating the calls can add inference latency. Existing persisted routes and traces remain readable. No new bindings or migrations.

Validation:

• Platform build and 622 unit tests passed; 7 existing skips.
• 66 local Workerd/SQLite/R2 tests passed, including new frame-focus reuse without extraction or retrieval charges.
• Dashboard build, type check, 32 web tests, and Chrome frame-preview regression passed.
• Three live synthetic Fireworks checks passed with GLM classification and DeepSeek finalization on priority: exact first message beyond eight turns, full user-message listing including the current request, and stored-transcript retrieval. No external video provider calls. Temporary live harness and diagnostics are not included.
• Regenerated OpenAPI and verified generated documentation.

[vercel]

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

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Sep 19, 2026 1:24pm UTC