Turn lifecycle

How one message becomes a reply.

Your message reaches the Hub, with the chat, the open folder (by an id and a name, never its path), any attachments and your time zone.
The context block. The Hub builds everything the brain needs to know for this turn and signs it. A forged or oversized block is refused and nothing runs.
The agent works. The brain's task agent reasons and calls tools, in rounds.
Hub calls. What needs credentials (an app action, an automation, a search, a schedule) goes back to the Hub, which checks your permissions, the organization's policy and Stop all, and runs it, or turns it into an approval card. A turn has a few rounds of these.
Client calls. What needs your computer (read a file, run a command, take a screenshot) goes to your client on the live stream. Your client runs it, one at a time and in order, under your permission mode and the sandbox, and sends the result back.
Checks. After writes in a folder, the project check runs in the same round; after a change to a screen, the agent looks at it. Before the reply goes out, the verifier checks its claims; if one doesn't hold, the agent gets one more try, continuing where it stopped.
The reply streams to you as it is written: text, tool rows, approval cards, tables, charts and files. The turn ends done, stopped or failed.

Bounds

A number of tool rounds per turn, more with a folder open.
About five minutes per turn; past it the agent is told to stop and say what's left.
A tool that fails three times is blocked for the rest of the turn.
Long folder work keeps its tool results within the context by shortening earlier results first, oldest first.

A turn that proposes a write ends on the card. Your answer starts the next turn, which resumes the proposal and runs exactly what was approved, once.