Session cockpit

Open any agent — from the attention queue, the Sessions tab, or the active field — and you land in its cockpit: everything you need to unblock or steer that one agent.

What you see

• Status — whether the agent is working, waiting for input, or blocked on a decision.
• Context — the workspace and host the agent belongs to, with host telemetry.
• Actions — Allow / Reject for pending prompts, a composer for the next instruction, and a shortcut into the live terminal for the agent's pane.

Opening a cockpit is gated behind your device biometric/passcode — a pocket unlock can't approve anything.

Answering permission prompts

When an agent asks "may I run this command / edit this file?", the cockpit's Allow and Reject buttons answer it.

NOTE — > Under the hood these send keystrokes — Enter to accept, Esc to dismiss — to the agent's terminal, exactly as you would at the keyboard. C3 does not parse the prompt or verify what you're approving. For anything consequential, open the terminal first: it's one tap away and shows precisely what the agent is asking.

Sending the next instruction

The composer sends a prompt to the agent — course-correct, answer its question, or queue up the next task, up to 64 KiB of text. Delivery is live-or-not-at-all: if your phone's hub connection drops mid-send, the command fails visibly instead of executing later (messages expire in seconds). Nothing you type is visible to the hub or push infrastructure — see the security model.

When the cockpit isn't enough

The cockpit is a decision surface, not a full REPL. For scrollback, reading diffs, or interacting below the agent layer, use the terminal view — same pane, live.