feat(web): pipeline filters, drift KPIs, and trace prompt inspector

[cursor]

Summary

Adds four product improvements to the LLM debate research dashboard, grounded in existing patterns and data already present in run artifacts.

Features

1. Pipeline insight filters (Agents + Timing)

• /agents and /timing now use InsightFilterCard with the same search/model/preset/fast/date filters as other insight pages
• Stats and timing tables respect the filtered run subset, with clear empty states when filters match nothing

2. Preset leaderboard filter parity

• /presets upgraded from fast-mode-only filtering to full InsightFilterCard support via applyIndexFilters
• "Filter runs" links preserve active filter context (matching model leaderboard behavior)

3. Overview confidence drift KPIs

• Overview dashboard now shows all three aggregate confidence drift means: solver→revision, revision→synth, and calibrated−synth
• Reorganizes KPI rows to avoid duplication and surface evidence-plan risk alongside drift metrics

4. Trace prompt & retry inspector

• Run trace steps expose collapsible LLM request (prompt) and Parse retries panels when artifact steps include request and rawAttempts
• Human-readable summary of model, temperature, schema, and message previews before raw JSON

Testing

• pnpm test — 172 tests passed
• pnpm typecheck — passed
• pnpm web:typecheck — passed
• pnpm web:build — passed
• pnpm format:check — passed

Manual verification

1. Visit /agents or /timing and apply model/preset filters — table counts should update
2. Visit /presets and compare filter behavior with /leaderboard
3. Visit / with analysis index loaded — confirm three drift KPI cards
4. Open any run trace (/runs/[id]) and expand "LLM request (prompt)" on agent steps

  

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
llm-debate-research	Ready	Preview, Comment	Jun 12, 2026 8:08am