agents.yml

# Agent Configuration for MCode Benchmark
# Pin versions and models for reproducibility

claude-code:
  name: "Claude Code"
  cli_version: "2.0.76"
  model: "claude-opus-4-5-20251101"
  api_key_env: "ANTHROPIC_API_KEY"
  install_cmd: "npm install -g @anthropic-ai/claude-code@{cli_version}"
  cmd_template: "cat prompt.md | IS_SANDBOX=true claude -p --model {model} --dangerously-skip-permissions --output-format stream-json --verbose"
  # Each iteration starts fresh (no session resumption)

gemini-cli:
  name: "Gemini CLI"
  cli_version: "0.22.5"
  model: "gemini-3-flash-preview"
  # Uses Vertex AI with Application Default Credentials (ADC)
  auth_type: "vertex_ai_adc"
  env_vars:
    GOOGLE_CLOUD_PROJECT: "mcode-dev"
    GOOGLE_CLOUD_LOCATION: "us-central1"
    GOOGLE_GENAI_USE_VERTEXAI: "true"
  # Run this before first use to set up ADC credentials
  auth_cmd: "gcloud auth application-default login"
  install_cmd: "npm install -g @google/gemini-cli@{cli_version}"
  cmd_template: "cat prompt.md | gemini -m {model} --yolo --output-format stream-json"

codex-cli:
  name: "Codex CLI"
  cli_version: "0.77.0"
  model: "gpt-5.2"
  api_key_env: "OPENAI_API_KEY"
  install_cmd: "npm install -g @openai/codex@{cli_version} && echo {api_key_env} | codex login --with-api-key"
  cmd_template: "cat prompt.md | codex exec - --model {model} --json --dangerously-bypass-approvals-and-sandbox --skip-git-repo-check"

opencode-claude:
  name: "OpenCode (Claude)"
  cli_version: "1.1.4"
  model: "anthropic/claude-opus-4-5-20251101"
  api_key_env: "ANTHROPIC_API_KEY"
  # Create auth.json for Anthropic authentication and opencode.json for auto-approve permissions
  install_cmd: "npm install -g opencode-ai@{cli_version} && mkdir -p ~/.local/share/opencode ~/.config/opencode && echo '{{\"anthropic\":{{\"type\":\"api\",\"key\":\"'$ANTHROPIC_API_KEY'\"}}}}' > ~/.local/share/opencode/auth.json && echo '{{\"permission\":{{\"*\":\"allow\"}}}}' > ~/.config/opencode/opencode.json"
  cmd_template: "opencode run \"$(<prompt.md)\" --model {model} --format json"

opencode-openai:
  name: "OpenCode (OpenAI)"
  cli_version: "1.1.4"
  model: "openai/gpt-5.2"
  api_key_env: "OPENAI_API_KEY"
  # Create auth.json for OpenAI authentication and opencode.json for auto-approve permissions
  install_cmd: "npm install -g opencode-ai@{cli_version} && mkdir -p ~/.local/share/opencode ~/.config/opencode && echo '{{\"openai\":{{\"type\":\"api\",\"key\":\"'$OPENAI_API_KEY'\"}}}}' > ~/.local/share/opencode/auth.json && echo '{{\"permission\":{{\"*\":\"allow\"}}}}' > ~/.config/opencode/opencode.json"
  cmd_template: "opencode run \"$(<prompt.md)\" --model {model} --format json"

# OpenCode with OpenAI GPT-5.2 and high reasoning effort
opencode-openai-reasoning:
  name: "OpenCode (OpenAI GPT-5.2 High)"
  cli_version: "1.1.4"
  model: "openai/gpt-5.2"
  api_key_env: "OPENAI_API_KEY"
  # Create auth.json for OpenAI authentication and config for high reasoning + auto-approve permissions
  install_cmd: "npm install -g opencode-ai@{cli_version} && mkdir -p ~/.local/share/opencode ~/.config/opencode && echo '{{\"openai\":{{\"type\":\"api\",\"key\":\"'$OPENAI_API_KEY'\"}}}}' > ~/.local/share/opencode/auth.json && echo '{{\"permission\":{{\"*\":\"allow\"}},\"provider\":{{\"openai\":{{\"options\":{{\"reasoningEffort\":\"high\"}}}}}}}}' > ~/.config/opencode/opencode.json"
  cmd_template: "opencode run \"$(<prompt.md)\" --model {model} --format json"