Set up your project's checks

Your coordinator is the agent you talk to, which starts the other agents, hands out work and reports back to you. Your coordinator checks every change before it merges. Out of the box, it has no rules of your own to check a change against. Three setup skills fix that: an agent reads your project and writes your rules, your hot paths (the folders where your performance-critical code lives) and your build-and-test commands. From then on, your coordinator is set up to check merges against them.

Note

These skills don’t generate a fixed, perfect setup. You run them with a capable agent, which reads your project and builds a version of the checks that have worked well while building Thrum, sized to your code. A small repository gets a small setup. Review what it proposes and change it as you go. A runtime is the AI coding tool an agent runs in, such as Claude Code or Codex. Run them on your runtime’s top model at high effort: start your coordinator on it, or use it in your own Claude Code session.

What each skill sets up

• project-philosophy writes .thrum/philosophy.md.

  Your project’s rules: the mistakes to avoid (a few universal ones, plus one for each reliability risk the agent finds in your code), red flags, and a link to your security notes. Your coordinator checks changes against it.

• project-hotpath-gate writes .thrum/hotpath-gate.json.

  The folders where your performance-critical code lives, and what your coordinator looks for when a change touches them.

• project-gate-setup writes .thrum/gate-config.json.

  The build and test commands to run on the merged code.

Set them up

Do them in this order, one at a time, because each one asks you questions. When it says the file is written, ask for the next.

Ask your coordinator, in your coordinator’s terminal (the one you started it in):

1. Set up the project philosophy for this repository.
2. Set up the hot-path gate for this repository.
3. Set up the merge gate commands for this repository.

Thrum’s web UI is the page you open in your browser to see your agents and their messages. My Inbox is your inbox in Thrum’s web UI.

Answer in that terminal, not My Inbox

The agent asks you questions as it goes: known mistakes your team has hit, things that must never happen, and deploy or release rules. Answer them in that terminal. They don’t come to My Inbox. On Free, answer there. With Pro, you can also answer from the Terminal tab in the web UI. On Claude Code the questions appear as a list you choose from with the arrow keys. On other tools the agent can ask them in plain text, or with the tool’s own picker.

Command approvals are different. Some tools also ask you to approve the commands the agent runs while it reads your project. They show in the same terminal, and can also reach My Inbox; answer in either. See Permission prompts.

If you use Claude Code, you can also run the three skills yourself in a Claude Code session in the same repository folder your coordinator works in:

/thrum:project-philosophy
/thrum:project-hotpath-gate
/thrum:project-gate-setup

Other tools name skills differently; see your tool’s page under Integrations.

Check it worked. The three files exist in your repository’s .thrum/ folder. Or ask your coordinator: Which checks will you run on the next merge? It should name your project philosophy, your hot-path folders and your build and test commands.

What happens at merge

• Once the files exist, your coordinator is set up to check merges against them. There’s nothing to switch on.
• It checks each change against your project philosophy. It skips this only for a one-line config change, a typo fix, or a change with no production code.
• It runs the hot-path review only when a change touches one of your hot-path folders.
• It’s set up to build and test the merged code with the commands you chose.

A gate agent is an agent that runs your project’s checks on the merged result. If you add a gate agent later, it runs these checks for your coordinator.

Change them later

• Edit the files directly, or ask your coordinator to run the skill again: Update the project philosophy; the code has changed.
• Run a skill again and it compares the file with your project. If nothing has changed, it says so and writes nothing. If something has, it shows you the proposed changes and writes only what you approve.

Next steps

• Grow your team of agents, one step at a time: what to add as your team grows.
• Keep track of who’s doing what (Pro): how your coordinator knows what each agent has taken on.
• Keep context when an agent restarts: your agents pick up where they left off.