Install and get started
Start in the root of your Git repository. An existing Trellis installation should follow the migration guide.
Requirements
- Node.js 18.17 or later
- Python 3.9 or later (
python3on macOS/Linux,pythonon Windows) - Git and a supported coding agent
Install the CLI
npm install -g mini-trellis
mini-trellis --version
Set up the project
Choose the agent you use, and replace your-name with your developer name:
mini-trellis init -u your-name --claude
mini-trellis doctor
You can combine flags:
mini-trellis init -u your-name --claude --codex
v0.3.8 supports --claude, --codex, --cursor, --opencode, --pi, --omp, --kiro, --dsh, --codebuddy, and --zcode. See Agent setup for trust prompts and platform differences.
Initialization creates .trellis/, the managed instruction block, and the selected agent’s integration files. Existing project knowledge is retained. A developer journal lives under .trellis/workspace/<name>/. Initialization writes files; review and commit the changes with your usual Git workflow.
Save one useful decision
Open a new agent session. For this example, assume your team has agreed on the following callback behavior:
Record this project rule: every callback carries an idempotency key;
repeated callbacks return the first result. Update the existing spec for
callbacks, or create a short spec if none exists. Then use
mini-trellis-remember to link this decision from the session journal.
In Claude Code, the recording command is /mini-trellis:remember. The command reference lists each agent’s entry point.
Check that the topic contains the rule and that the journal links to it. Example projects can use .trellis/spec/backend/callbacks.md; an existing README or docs page is also a valid home.
Find it in a new session
mini-trellis mem recall "callback idempotency"
The result contains a source, a short excerpt, and a ref. Expand the returned reference when you need more detail:
mini-trellis mem read "<ref-from-recall>"
You now have a complete first use: record a decision, start another session, and retrieve the source. Continue with day-to-day use, or use troubleshooting if the expected result is missing.