name: cannon-and-wall
version: 1.1.0
author: Jairaj S
theme: self-improvement+multi-agent
description: >
  A Red vs Blue adversarial RL environment for training LLMs on cybersecurity reasoning.
  Two agents (Cannon = attacker, Wall = defender) compete across escalating curriculum
  stages of a seeded Flask app with OWASP Top 3 vulnerabilities. Rewards are fully
  deterministic — no LLM-as-judge.
base_class: Environment
entry: environment/server.py
client: client/client.py
port: 7860
tags:
  - security
  - red-team
  - self-play
  - rl

action_space:
  cannon:
    type: object
    properties:
      agent:         {type: string, const: cannon}
      vuln_type:     {type: string, enum: [sqli, xss, broken_auth]}
      line_number:   {type: integer}
      explanation:   {type: string}
      proof_of_concept: {type: string}
  wall:
    type: object
    properties:
      agent:         {type: string, const: wall}
      patched_code:  {type: string}
      explanation:   {type: string}

observation_space:
  type: object
  properties:
    source_code:    {type: string, description: "Full Flask app source for the current stage"}
    round:          {type: integer, description: "Current round number (1-3)"}
    stage:          {type: integer, description: "Curriculum stage (1-3)"}
    scores:
      type: object
      properties:
        cannon: {type: number}
        wall:   {type: number}
    last_cannon_report: {type: [object, "null"]}