fix: streamline research finalization and trace rejected answers

[devhims]

Problem / Motivation

Users can wait through redundant context preparation after research has already collected the needed analyses. Supported answers can also be rejected because optional identity metadata differs across videos. Missing rejected drafts make those failures difficult to diagnose.

What changed

Research now ends with a report of request parts, supporting evidence references and unresolved questions. With a valid report, finalization skips the general preparation model loop. Application checks identify missing subjects, invalid references, stale evidence and coverage limitations. At most four targeted saved-asset or history reads address identified gaps. Repair reuses those reads.

Private agent traces retain the handoff decision, targeted reads and rejected answers, including detailed validation errors and the responding model. Existing access and deletion controls apply.

Grounding rejection focuses on numerical support. Optional names are advisory; measurements, citations and answer structure retain their checks.

%%{init: {"themeVariables": {"signalColor": "#64748b", "sequenceNumberColor": "#ffffff"}}}%%
sequenceDiagram
    autonumber
    participant R as Research
    participant P as Preparation model
    participant A as Application
    participant F as Answer model
    rect rgba(128,128,128,0.08)
        Note over R,F: Before
        R->>P: Collected evidence
        P->>A: General context reads
        A->>F: Generate answer
    end
    rect rgba(128,128,128,0.08)
        Note over R,F: After a valid handoff
        R->>A: Findings and explicit gaps
        A->>A: Validate coverage and perform targeted reads
        A->>F: Generate answer
    end

Loading

The preparation model loop is removed for valid research handoffs; application checks and answer generation remain.

Tests

All 1,560 unit tests and 347 runtime/session integration tests pass. Type checking passes. Tests cover four available analyses, explicit and missing-subject reads, repair reuse, cancellation, billing admission, time/cost reserves, cross-video names and rejection traces.

Compatibility and deployment

No migration is required. History/context requests and research without a valid report retain existing preparation. Recovery does not restore reports from traces. Semantic completeness still requires model judgment. Platform deployment and live latency verification remain outstanding.

[vercel]

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

Project	Deployment	Actions	Updated
video2ctx-web	Ready	Preview	Oct 7, 2026 11:55am UTC