The AI assistant

The builder has an AI assistant built into the right dock. Press Cmd+? to open it. It chats about the program you have open in the editor. Ask it to explain the structure, critique the loading, or make changes. It works directly against that program.

It needs something to look at. Open a program in the editor before chatting. The panel tells you the same if you forget.

Reads are automatic, writes wait for you

The assistant's ground rule: it can look freely, but it can't touch anything without your sign-off.

• Reads run automatically. Listing your exercises, tags, and protocols, or walking the program's structure, happens as part of answering. Reads need no approval, because nothing changes.
• Writes become approval cards. Every edit the assistant proposes — adding a session, changing loads, applying a protocol — appears as a card with Accept and Reject. Nothing lands in the program until you accept it.

When the assistant proposes a batch of changes, Accept all (n) approves the lot at once. If you find yourself accepting everything, the auto-accept toggle skips the cards. Flip it back off when the stakes go up.

Accepted changes are ordinary edits. They show in the grid immediately. Program History records them. Cmd+Z undoes them like anything else.

Conversations and providers

• Provider — a dropdown chooses which model backs the panel: OpenAI or Anthropic.
• History — the panel keeps past conversations. Start a new conversation when you change topics, so the context stays clean.