Cookbook

Each recipe here is a whole program. Copy the file, set one API key, run it. No placeholders, no ..., no imaginary services — the tools are standard library, so nothing needs an account beyond your model provider.

Recipe	Solves	Shows off
Changelog writer	Turn a release’s commits into a written changelog entry	Tools over subprocess, returning errors as strings, metadata for traceability
Codebase explainer	Explain an unfamiliar repo without blowing the context window	Ad-hoc sub-agents, secondary_model, parallel tools
Support triage	Classify and route inbound tickets, with an audit trail	Named specialists, structured routing, metadata you can query later

Before you start

pip install minion-ai
export OPENAI_API_KEY="sk-..."

Every recipe uses openai/gpt-4o. Any Tier 1 or Tier 2 model works — change the model string. Every recipe also turns tracing on, so after running one you can open the dashboard and read what actually happened:

minion serve

That is the point of running them, honestly. The code is short; the trace is where you see the agent think, delegate, and spend.

How these are written

A few conventions worth stealing, each of which is explained in the recipe that first uses it:

• Tools return errors as strings, so the model can read the failure and retry instead of the run dying. See Tools.
• Tools are built per call, not shared at module scope, so parallel_tools=True is safe. See Thread safety.
• Every run carries metadata you’d actually want to filter on later — a version, an id, an environment.
• Nothing writes outside the working directory, and nothing runs a command the agent composed freely. Tools are narrow on purpose.