name: ticketmelt
version: "0.1"
description: >
  Multi-agent RL environment: four on-call engineers coordinate incident response
  across two production servers. Trains LLMs to break convergent reasoning and
  adapt to heterogeneous urgency simultaneously.
interface: gym-http
endpoints:
  health: GET /health
  reset: POST /reset
  step: POST /step
  state: GET /state
action_space:
  type: discrete
  choices: [DEPLOY_PROD_A, DEPLOY_PROD_B, MONITOR]
observation_space:
  type: dict
  description: >
    Current round, own service state (name, fix_rounds_remaining, deadline, urgency_flag),
    peer progress (public: service, rounds_remaining, deadline, completed), episode history.
reward:
  type: binary
  values: [0.0, 1.0]
  description: >
    End-of-episode binary GRPO reward. Continuous components (R1-R4) in info["reward_breakdown"].
    goodness_threshold: 0.45
agents: 4
scripted_peers: 3
episodes_per_training_prompt: 1