Skip to content

Terminal & tool execution

How command and tool execution is gated, and how to see the raw output behind it.

Command and tool execution isn't governed by a separate AgentDock-level permission system — it's gated by whichever permission mode the active agent is running in, using the same interaction-card mechanism described for approvals generally:

  • Claude Code's canUseTool callback blocks or allows a tool call, including shell commands, based on its current mode.
  • Codex's sandbox mode (read-only / workspace-write / danger-full-access / etc.) determines what it can execute without prompting.
  • Antigravity's --mode flag plus terminal-output classification determines what proceeds automatically versus what raises a prompt.

Seeing what's actually happening

AgentDock includes a raw terminal drawer that shows the underlying PTY/CLI output directly — the same stream the structured conversation view is built from — including a trace log. This is a visibility and debugging tool, not a permission control: it lets you see exactly what a command produced without changing what's allowed to run.

Tip

If an agent's behavior seems off, or a permission prompt doesn't look right, the raw terminal drawer is the first place to check what the underlying CLI actually said.

For what each mode allows per agent, see Agent permissions.