Principles

Projects, terminals and Focus come first

The core of PrumoGrid fits in one sentence: several projects, several terminals in each one, and Focus to read a terminal calmly. Everything else (stages, flows, routines, rules, roles) is an addition. You can spend weeks using only projects, terminals, Grid and Focus, and the app is still useful.

In practice:

• Each project is a folder. Terminals open inside it and appear on the project card. See Projects.
• Selecting a project switches the Grid and the sidebar together. You look at one piece of work at a time, and Attention tells you when another project needs you.
• The Grid is for keeping an eye on several terminals; Focus is for reading or talking to one of them. See Grid and Focus.

When repetition shows up (the same request in several terminals, the same implement-and-review sequence, the same guidance for every agent), that is the time to adopt the feature that solves that repetition. Not before.

The daemon holds the sessions

The terminals do not run inside the window. They run in a background service, the daemon, and the window only connects to it. The status bar shows daemon connected when everything is fine.

That is why the Close this window? dialog appears when there is work that depends on the window being open: it warns you that the flow will not advance until you come back. See Resume after restart.

The agents are your CLIs

PrumoGrid has no AI model and no account of its own. When a stage says Runs on: Claude Code, what runs is the Claude Code installed on your computer, with your account, your plan, your permissions and your cost. The same goes for Codex, Gemini or any other CLI registered in Tools → Agent profiles….

Consequences:

• Installing, authenticating and choosing the default model of each CLI is up to you. An empty model field on a stage means "the CLI's default".
• Each CLI has its own way of asking for permission. PrumoGrid shows the request, but you decide, in the dialog or in the terminal.
• Different agents do not share a conversation. When Codex needs what Claude Code found, the bridge is a file, such as ANALISE.md. See the recipe Two agents, one handoff file.
• The notification that the agent finished its turn comes from the CLI itself. PrumoGrid puts that notification in the command it uses to open the agent, without touching your account's configuration.

PrumoGrid follows the work

The terminal is still yours. You can type in any pane, open the agent by hand with claude or codex, run your scripts and use your own harness. New stage, Flow and Routines are optional: they are for when you want PrumoGrid to deliver the task, wait for the agreed condition and tell you.

What PrumoGrid does is observe and record: which instruction was delivered to which terminal, when the agent signaled the end of its turn, what the verification returned, who approved. What it does not do is decide for you. Permission requests and review approvals are still answered by a person, and a permission request shown in the dialog is read from the terminal screen: if it changes, close and reopen it before answering.

Verification must be able to fail

"The agent said it finished" and "the test passed" are different facts, and PrumoGrid shows them separately. A stage without verification is done when the agent signals the end of its turn. A stage with verification is only done when the chosen command ends as expected.

A useful verification is one that fails while the work is not done. echo ok always passes and proves nothing. For the loja-api coupon bug (double discount with two items), the right verification is npm test with a test that reproduces the case: it fails before the fix and passes after.

Even so, the exit code only proves what the command tested. The quality of the solution, the clarity of the code and whatever was left out of the tests still need a human to read them. See Verification.

Nothing leaves the machine through PrumoGrid

PrumoGrid does not ask for an account, has no data server and does not send the contents of your terminals anywhere. Projects, layout, routines, history and global rules stay in %APPDATA%\TerminalGrid, on your computer.

What can use the network:

• The agent CLIs you run, according to each one's account and configuration. That is how code reaches the model provider.
• The commands you run in the terminals, such as npm install or git push.
• The update check, which queries the published manifest (Check now, or on startup and every 6 hours, if that option is on).
• The plugins you install, which run code with your account's rights.

When PrumoGrid needs to write to your repository (the guidance block in CLAUDE.md or AGENTS.md), it asks first, folder by folder, and shows what it will write. See Rules.