Execution Runtime

Execution behavior lives in executor/* and provisioner/*.

Base Images

Auto-built once if missing:

• validtr-agent-base:latest
• validtr-test-runner:latest

Agent base image includes:

• Python 3.12
• Node.js 20
• SDKs: anthropic, openai, google-genai, httpx, pydantic

Per-Run Files

ComposeGenerator creates run artifacts under temp storage (/tmp/validtr-runs/<run_id> by default):

• docker-compose.yml
• Dockerfile.agent
• entrypoint.py
• agent_loop.py
• workspace/task.json
• workspace/stack.json

Agent Execution

• container mounts run workspace and scripts
• output files are expected under /workspace/output
• artifact collector reads files recursively from output directory

Tooling in Agent Loop

Anthropic/OpenAI agent loops can call:

• write_file(path, content)
• run_command(command)

Gemini loop currently uses a simplified file-block parsing flow.

Path Traversal Guard

write_file blocks writes that escape /workspace/output via realpath checks.

Streamable HTTP MCP Services

For streamable-http MCP entries, compose creates sidecar services named mcp-<server-name>.