One turn, many model responses

What looks like a single reply from tCode is often the result of several separate model responses chained together behind the scenes. The model reads a file, gets the contents back, decides what that means for the next step, maybe searches for something else, maybe edits a file, gets the result of that edit, and keeps going.

This chaining is what makes tCode fundamentally different from a one-shot code generator: it is closer to watching someone actually work through a task step by step, checking their own assumptions as they go, than to receiving a single generated answer to a single question.

Because each step genuinely depends on the previous one's real result, not a guess about what that result would probably be, the chain adapts as it goes. See Outputs for what this chain actually looks like printed to the terminal as it happens.

The number of responses chained together in a single turn is not something you set or predict in advance; it falls out entirely of how many steps the model actually needs, which is why the same kind of request can take a very different number of steps depending on what it finds along the way.

This is also why interrupting a turn partway through, covered in Run times, does not simply stop "the response", there may be several more queued up in the chain that would have followed the one currently running.