Approvals

Approvals pause a proposed action when your Auto-Approve settings require you to accept or reject it.

EngineForge Cockpit sidebar with Auto-Approve expanded: Read Files, Web Pages, Write Files, Run Commands, Game Engine, Execute Code and Asset Generation on, Auto Playtest off.
Auto-Approve categories, with Auto Playtest off by default.

When a request appears

The agent shows what it wants to do and waits for a decision. For work a specialist does for Maestro, the approval card shows the command and why it asks. Review that scope rather than relying on the button label alone.

Review procedure

  1. Confirm the active project, session and agent.
  2. Inspect the proposed action and its target.
  3. Deny or refine a request whose purpose or scope is unclear.
  4. Allow the action only when it matches the intended task.
  5. Inspect the resulting changes, diff and evidence.

Auto-Approve categories

Auto-Approve is on by default so the Crew can build a game end to end; every turn is checkpointed and its file changes can be undone. Turn a category off to be asked first:

  • Read Files: read and search files.
  • Web Pages: open web pages. Pages from the agent’s own searches and links you write never ask.
  • Write Files: create and edit files.
  • Run Commands: run terminal commands.
  • Game Engine: modify scenes and objects in the engine.
  • Execute Code: compile and run code the agent writes in the game engine. On by default. Turn it off to approve each run yourself.
  • Asset Generation: generate game assets such as sound effects, music and synthetic character voice lines. Each generation uses credits.
  • Auto Playtest: off by default; turn it on here. Sends the Playtester after a change to how the game plays instead of asking. A playtest takes a few minutes and uses credits.

Command boundaries

Command approval can also be narrowed with allowed and denied command rules. A denied rule still blocks the command, and an unknown or unmatched action can still require a decision.

Limitations

Approval confirms permission to attempt the action, not that its output is correct. Auto-Approve never gives an agent tools outside its role or makes Plan mode writable.

Related documentation