3b4fa449f1
When viewing any past run in History, click "Optimize" to trigger an automated workflow that: 1. Analyzes the original prompt + responses + score 2. Identifies improvement strategies (clarity, depth, specificity, etc.) 3. Generates 3-5 improved prompt variations 4. Tests each variation across original mode + brainstorm 5. Auto-scores all results via background judge 6. Ranks results and highlights the winner 7. "Use This" button loads winning prompt into composer Architecture: - _run_optimize(job_id, run_id): background thread, 5-phase engine - POST /api/runs/<id>/optimize: starts optimization job - GET /api/optimize/<job_id>/stream: SSE for live progress - Budget-capped at 15 model calls per optimization - Child runs saved as real team_runs (source: "optimize") - Auto-scored → feeds into analytics + routing table automatically - Results saved to pipeline_runs (pipeline: "optimize") Frontend: - "Optimize" button in history detail panel (accent-colored) - startOptimize(runId): replaces detail view with live optimization stream - Phase cards: Analysis → Variations → Testing → Ranked Results - Score bars with color coding (green/amber/red) - Winner row highlighted with star + "Use This" button Closes the learning loop: system studies its own history → generates better prompts → tests them → scores results → routing table improves. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>