name: incident-commander
version: "2.0.0"
description: >
  SRE incident response RL environment with real Kubernetes backend,
  self-escalating curriculum, LLM-designed adversarial scenarios, and
  a 3-persona LLM judge. An AI agent plays an on-call Site Reliability
  Engineer at AcmeCorp, diagnosing and mitigating production incidents
  on a live e-commerce platform. Observations come from Prometheus
  metrics, Loki logs, Alertmanager alerts, and distributed traces.
  Write actions execute against a mock cluster by default and can be
  routed to real kubectl commands via REAL_K8S=true. Features
  context-gated rewards, phase-aware bonuses (triage/investigate/
  fix/verify), repeat-command penalties, multi-fault scenarios on
  advanced/expert tiers, and holistic per-episode grading.

author: IncidentCommander Team
license: MIT

tasks:
  - id: task1
    name: Redis Connection Pool Exhaustion
    difficulty: easy
    target_score: 0.80
    description: >
      Network latency injection between inventory-service and Redis
      causes connection pool exhaustion. Agent must identify the
      Chaos Mesh experiment and delete it.

  - id: task2
    name: Cascading Failure via Payments OOM
    difficulty: medium
    target_score: 0.45
    description: >
      Memory stress on payments-api causes OOM kills, which cascades
      to Kafka consumer lag on inventory-service. The root cause is
      payments-api, not inventory-service. Agent must rollback
      payments-api deployment.

  - id: task3
    name: Silent Decimal Corruption
    difficulty: hard
    target_score: 0.20
    description: >
      A bad deployment (v2.3.2) silently truncates order totals from
      NUMERIC(12,4) to NUMERIC(12,2). A concurrent postgres VACUUM
      ANALYZE creates a red herring CPU spike. Agent must identify
      the decimal bug, rollback, and recommend a data audit.

  - id: task4
    name: Kafka Broker Network Partition
    difficulty: easy
    target_score: 0.80
    description: >
      Network partition isolates Kafka broker from application pods,
      causing consumer lag >5000 messages across order-worker and
      notification-service. Agent must find and delete the Chaos Mesh
      network partition experiment.

  - id: task5
    name: DNS Resolution Failure
    difficulty: medium
    target_score: 0.45
    description: >
      DNS chaos injection causes NXDOMAIN and resolution timeouts
      across all services. checkout-frontend shows 502 errors.
      Health check "connection refused" errors are secondary symptoms,
      not the root cause. Agent must delete the DNS chaos experiment.

  - id: task6
    name: TLS Certificate Expiry Cascade
    difficulty: hard
    target_score: 0.20
    description: >
      Expired mTLS certificate breaks the payments-api to postgres
      connection path. All database queries fail with x509 errors.
      ECONNRESET and upstream timeout alerts across other services
      are symptoms of the cascade. Agent must apply a config patch
      to rotate the certificate.

  - id: task7
    name: ConfigMap Hot-Reload Race Condition
    difficulty: hard
    target_score: 0.20
    description: >
      A ConfigMap update triggers hot-reload in inventory-service
      but a race condition causes 2 of 4 pods to load new config
      while 2 retain stale values. Redis pool saturation and GC
      pause alerts are unrelated red herrings. Agent must restart
      inventory-service pods to force consistent config.

observation_space: structured_dict
action_space: discrete_with_typed_params
reward_range: [-2.0, 1.0]
score_range: [0.001, 0.999]

endpoints:
  reset: POST /reset
  step: POST /step
  state: GET /state
  tasks: GET /tasks
  grader: POST /grader
  baseline: POST /baseline
  dashboard: GET /dashboard
  curriculum: GET /curriculum
  curriculum_reset: POST /curriculum/reset
  adversarial_design: POST /adversarial/design
  judge_config: POST /judge/config

features:
  curriculum_tiers: [warmup, beginner, intermediate, advanced, expert]
  adversarial_designer: true
  llm_judge_personas: [junior, senior, principal]
  phase_aware_rewards: true
  repeat_command_penalty: true
  multi_fault_scenarios: true
  real_k8s_backend: optional
  grpo_training_script: rl-agent/training/train_grpo.py
  eval_script: rl-agent/eval.py
  colab_notebook: notebooks/incident_commander_colab.ipynb

tags:
  - openenv
  - sre
  - incident-response
  - kubernetes
  - reinforcement-learning
  - curriculum
  - adversarial
  - llm-judge
  - grpo