openenv.yaml

name: SecureCodeEnv
version: "2.0.0"
description: >
  An RL environment for training LLM agents to write production-ready,
  secure Python code. Agents are graded on correctness, security attack
  resistance (dynamic adversarial payloads), CWE-based static analysis,
  performance, and codebase consistency via a novel CodeGraph memory system.
  No other public OpenEnv environment combines attack simulation + codebase
  consistency grading. All grading is 100% automated and deterministic.

author: Vishal Dhakad
hf_space: vishaldhakad/SecureCodeEnv
license: MIT

action_space:
  type: text
  description: Python source code string submitted by the agent
  fields:
    - name: code
      type: string
      description: The complete Python function(s) to be graded
    - name: filename
      type: string
      description: Logical filename for CodeGraph tracking (e.g. src/auth/validator.py)
    - name: session_id
      type: string
      description: Session ID returned from /reset

observation_space:
  type: structured
  fields:
    - name: total_reward
      type: float
      range: [0.0, 1.0]
      description: Weighted final score across all 7 dimensions
    - name: scores
      type: dict
      description: >
        Per-dimension scores: correctness, attack_resist, static_security,
        consistency, performance, documentation, code_structure
    - name: feedback
      type: dict
      description: Human-readable feedback string per grading dimension
    - name: codegraph
      type: dict
      description: >
        Full codebase context including components, detected conventions,
        dependency list, and natural-language context prompt for the agent
    - name: done
      type: bool
      description: True if episode is complete (reward >= 0.90 or max steps reached)
    - name: step_count
      type: int
      description: Current step number within the episode

reward:
  type: multi_dimensional
  range: [0.0, 1.0]
  dimensions:
    - name: correctness
      weight: 0.30
      description: Fraction of test cases passed (including edge cases)
    - name: attack_resistance
      weight: 0.20
      description: Fraction of randomized adversarial payloads blocked
    - name: static_security
      weight: 0.15
      description: bandit + AST security linter score (CWE-mapped)
    - name: codegraph_consistency
      weight: 0.15
      description: Adherence to conventions from existing codebase components
    - name: performance
      weight: 0.10
      description: Relative efficiency vs naive/optimal baselines (timeit)
    - name: documentation
      weight: 0.05
      description: Docstring + type hint coverage across all functions
    - name: code_structure
      weight: 0.05
      description: Clean code checks (no bare print, no bare except, etc.)

tasks:
  - id: easy_password_validator
    difficulty: easy
    cwe: [CWE-916, CWE-521]
    description: Validate password strength and hash with bcrypt (not MD5)

  - id: easy_input_sanitizer
    difficulty: easy
    cwe: [CWE-20, CWE-116]
    description: Sanitize HTML (XSS prevention) and filenames

  - id: easy_token_generator
    difficulty: easy
    cwe: [CWE-338, CWE-330]
    description: Generate cryptographically secure tokens using secrets module

  - id: medium_sql_query_builder
    difficulty: medium
    cwe: [CWE-89, CWE-20]
    description: Build parameterized SQL queries — never string-format user input

  - id: medium_file_path_handler
    difficulty: medium
    cwe: [CWE-22, CWE-20]
    description: Resolve file paths safely — block path traversal attacks

  - id: medium_rate_limiter
    difficulty: medium
    cwe: [CWE-770, CWE-400]
    description: Thread-safe sliding window rate limiter

  - id: hard_file_upload_handler
    difficulty: hard
    cwe: [CWE-22, CWE-434]
    description: Validate uploads — block traversal filenames, executable extensions, MIME spoofing

  - id: hard_jwt_validator
    difficulty: hard
    cwe: [CWE-347, CWE-613]
    description: Validate JWTs — enforce HS256, block none-alg attack, check expiry

  - id: hard_auth_middleware
    difficulty: hard
    cwe: [CWE-287, CWE-352]
    description: CSRF protection and Bearer auth using hmac.compare_digest (timing-safe)

runtime:
  max_steps_per_episode: 5
  done_reward_threshold: 0.90
  max_inference_time_minutes: 20
  min_vcpu: 2
  min_memory_gb: 8
  port: 7860

endpoints:
  health: GET /health
  reset: POST /reset
  step: POST /step
  state: GET /state
  docs: GET /docs