Minimal coding agent

A small tool-calling agent built with LiteLLM. It runs a loop: the model proposes run_shell or finish tool calls; commands execute in a working directory and results are fed back until the model finishes or a step limit is hit.

Contents

File	Role
agent.py	Agent loop, LiteLLM integration, tool execution
config.py	MODEL_NAME, API keys, timeouts, paths (from env)
prompts_v1.py / prompts_v2.py	System prompts (v1 shorter, v2 more guidance)
prepare_env.py	Clone repos for SWE-bench (prepare_SWE_env) or generic Git (prepare_custom_env)
run_in_container.py	Run agent.py inside Docker (custom Ubuntu image or SWE-bench image)

Requirements

• Python 3.10+
• Git (for prepare_env.py)
• Docker (only if you use run_in_container.py)

Setup

Environment variables

The agent reads configuration from the environment. At minimum you need a model name and the matching API key for that provider.

• MODEL_NAME — LiteLLM-style id, e.g. openai/gpt-5-mini-2025-08-07, anthropic/claude-…, openrouter/…, bedrock/…, pcss/… (see config.py for supported providers and key names).
• Provider API key — e.g. OPENAI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY, etc., depending on the provider prefix in MODEL_NAME.

Optional:

• AGENT_WORKING_DIR — where the agent runs shell commands (default ./working_dir)
• AGENT_MAX_STEPS — max LLM steps (default 20)
• AGENT_COMMAND_TIMEOUT — subprocess timeout in seconds (default 30)
• AGENT_PROMPT_VERSION — v1 or v2 (default v2)

config.py validates keys on import.