Commit History
-
Expand to 30 tickets, fix Phase 1 grader, add benchmark table
Tarun +1: Claude Sonnet 4.6- openenv/tasks: expand easy/medium/hard from 5 → 10 tickets each (30 total) - openenv/graders/grader.py: Phase 1 now scores 0.998 only when classify + correct issue_type in response (0.55 for classify-only, 0.002 for wrong action) - inference.py + app.py: update PHASE_PROMPTS[1] to require issue_type in response field - app.py: update hero stat (15→30), tier labels, scenarios header; add 6-model benchmark table to Overview with per-phase color-coded scores - server/app.py: synced from app.py Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>868368d -
Redesign environment as proper 3-phase RL MDP
Tarun +1: Claude Sonnet 4.6Previously all action types (classify/assign/refund/escalate) terminated the episode in 1 step, making it a bandit problem rather than an MDP. New episode structure (3 phases, ~3 steps per episode): Phase 1 — Triage: agent must classify the ticket Phase 2 — Route: agent must assign to correct specialist team Phase 3 — Resolve: agent escalates/refunds/responds with quality response Key changes: - openenv/env.py: add self.phase tracking; classify/assign are no longer terminal; observation updates between phases (issue_type revealed after phase 1, team context added after phase 2); episode terminates only after resolution phase or step limit - openenv/graders/grader.py: add grade_phase() — phase 1/2 give binary 0.998/0.002 rewards; phase 3 uses full proportional quality scoring - openenv/reward.py: pass phase to grade_phase() - openenv/tasks/easy.py: all scenarios now include team + resolution action + response keywords for all 3 phases - openenv/tasks/medium.py: chan…c35438f
-
Initial commit with files from Support-Agent repository
Tarun29edd33