The same loop for remote-triggered turns

A turn sent from a paired browser tab through the tAI web app runs through this exact same loop, unchanged. There is no separate, lighter version of the mechanism reserved for remote use.

The one and only thing that differs for a remote-triggered turn is where its permission prompts are answered from, the browser rather than the local terminal, since a request that originated remotely does not carry the same trust as someone physically at that machine's keyboard.

Everything else, the tool headers, the reasoning chain, the closing summary, is identical either way. See Pairing a machine for how a machine gets set up to accept these turns in the first place.

This equivalence is deliberate rather than incidental: building and maintaining two separate execution paths, one for local turns and one for remote ones, would be exactly the kind of divergence that eventually causes the two to drift apart in subtle ways. Sharing one path avoids that risk entirely.

From the model's own perspective inside a given turn, there is no signal distinguishing a remote-triggered request from a local one beyond where the permission answer ultimately comes from; the reasoning itself proceeds identically either way.