A tiny coding agent with zero runtime dependencies. Development checks use Ruff and mypy via uv.

WARNING: this was created for educational purposes, not meant for actual use. Agent can run shell commands and has no sandbox.

• Single invocation, not interactive: prompt in, answer out, no REPL scaffolding. The agent may still loop internally on tool calls until it reaches a final answer.
• Per-command approval kept simple (yes/no); there is no auto-approve mode.

Run:

OPENROUTER_API_KEY="..." MODEL="minimax/minimax-m2.7" python3 -m choom "your prompt"

Use AGENT_STREAM=1 to show generated text as it arrives. By default shell command output is silent; set AGENT_VERBOSE=1 to print the full $ command\nexit N\n<stdout> block as each command runs. Use AGENT_REASONING=auto to request and show reasoning summaries when the model supports them. Shell commands require per-command approval (y/n); there is no auto-approve mode.

With local llama.cpp:

URL="http://127.0.0.1:8080/v1/chat/completions" MODEL="gemma-4-12b" python3 -m choom "your prompt"

Inspect sessions:

cat agent_sessions.json

It logs prompts, exact API request payloads, responses, tool results, errors, and final answers.

Checks:

uv run ruff check .
uv run ruff format --check .
uv run mypy
uv run --script scripts/check_source.py [paths...]

Override source limits with SOURCE_LINE_LIMIT and SOURCE_COMPLEXITY_LIMIT.

1. Add Config Keep all runtime settings in one dataclass: API key, model, working dir, token limits.

2. Add the spinner Show activity states in the terminal.

3. Add local context collect_files() loads README.md, AGENTS.md, and .agents/skills/**/SKILL.md.

4. Build the system prompt Include cwd, OS, Python version, shell, and paths to discovered context files.

5. Add shell execution execute_shell() runs commands with approval, cwd validation, timeout, visible output, and output truncation.

6. Add the model call call_model() sends a minimal Responses API request with urllib.request.

7. Add function tools The model can call execute_shell with command, description, cwd, timeout, and env.

8. Add the agent loop Call the model, run requested tool calls, append results, repeat until final or limits.

9. Add session logs Write the full run to ./agent_sessions.json for debugging.

10. Add live trace Print reads, API calls, and shell commands as they happen.