agent-workspace-starter

A minimal, working scaffold for a governed agent workspace in Claude Code. Click "Use this template", open a terminal in your copy, and you have session discipline (orient → work → wrap), a self-improvement loop, and two safety hooks running on day one.

This is the runnable companion to agent-workspace-architecture — a documented production workspace published as a reference. The reference explains the patterns and the reasoning; this template is the smallest honest subset you can actually start from. Take the interactive tour of the full system to see where a workspace like this can go.

Quickstart

If step 2 prints anything else, both guards stay off. The hooks are wired to python3 in .claude/settings.json, the name macOS and most Linux distributions ship, and Claude Code treats a hook it can't start as a non-blocking error, so the tool call goes through and at most you see a notice. On Windows, the Python Install Manager and the Microsoft Store build provide python3. The python.org installer and conda provide only python and py. Change "command": "python3" to "python" or "py" in both hook entries, then re-run step 2 with the name you chose.

New to how agent context works? Read docs/before-your-first-session.md first: what the model can see, why long sessions drift, and the five habits the loop below assumes. Five minutes.