8ad221b41f
Phase 1 — Run Quality Scoring: - Auto-score every run in background via qwen2.5 judge (1-10) - Thumbs up/down vote buttons on output cards - POST /api/runs/<id>/score for user feedback - run_saved SSE event enables vote buttons after run completes - User votes override auto-scores (race-condition safe) - DB: quality_score, score_method, score_metadata on team_runs Phase 1 — Analytics Dashboard: - GET /api/admin/analytics: score-by-mode, score-by-model, heatmap, trend - New Analytics tab on Admin page with bar charts, heatmap table, trend sparkline - Scoring coverage tracker (scored vs total runs) - Model × Mode heatmap with color-coded cells Phase 2 — Reactive Pipeline: - _assess_stage(): orchestrator evaluates each stage's output mid-run - _reactive_decide(): can insert/skip stages based on assessment - Dynamic stage loop replaces fixed iteration in run_refine() - Budget tracking prevents infinite loops (max_stages hard cap) - Reactive decisions render as dashed notification bars between cards - Pipeline adjusts in real-time: "Inserting VALIDATE — high severity gaps found" Phase 3 — Cross-Run Learning: - _build_routing_table(): queries historical scores for model×mode performance - Best stage sequences per content_type from pipeline_runs - Routing table cached with 30-min TTL - Auto-Refine strategist prompt augmented with historical data - GET /api/suggest-models?mode=X returns top 3 models for that mode Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>