openenv.yaml

version: "0.2"
name: "cicd-debugger-env"
description: "RL environment for CI/CD debugging with deterministic, hidden, and quality-aware scoring"
metadata:
  domain: "devops"
  real_world_task: "ci-cd pipeline debugging"
  deployment: "huggingface-space-docker"

environment:
  entry_point: "env.environment:CICDDebuggerEnvironment"

interface:
  observation_type: "json"
  action_type: "text"
  max_steps: 30

action_space:
  tools:
    - read_file
    - read_logs
    - analyze_error
    - edit_config
    - run_pipeline_stage
    - run_tests
    - validate_fix
    - submit_solution

tasks:
  - id: "easy-command-typo"
    difficulty: "easy"
    failure_stage: "test"
    graders: ["deterministic"]
  - id: "easy-missing-checkout"
    difficulty: "easy"
    failure_stage: "build"
    graders: ["deterministic"]
  - id: "easy-yaml-indentation"
    difficulty: "easy"
    failure_stage: "build"
    graders: ["deterministic"]
  - id: "medium-python-version"
    difficulty: "medium"
    failure_stage: "build"
    graders: ["deterministic", "llm_judge"]
  - id: "medium-cache-key"
    difficulty: "medium"
    failure_stage: "test"
    graders: ["deterministic", "llm_judge"]
  - id: "medium-artifact-permissions"
    difficulty: "medium"
    failure_stage: "deploy"
    graders: ["deterministic", "llm_judge"]
  - id: "hard-matrix-logic"
    difficulty: "hard"
    failure_stage: "test"
    graders: ["deterministic", "llm_judge"]
  - id: "hard-conditional-deploy"
    difficulty: "hard"
    failure_stage: "deploy"
    graders: ["deterministic", "llm_judge"]
  - id: "hard-needs-order"
    difficulty: "hard"
    failure_stage: "deploy"
    graders: ["deterministic", "llm_judge"]