What actually ends a turn

There is no fixed number of tool calls that defines when a turn is "done." What ends it is the model itself judging that it has accomplished what your message actually asked for, whatever that took.

A one-line fix might genuinely need exactly two tool calls, a read and an edit. A broad investigation might need a dozen, several searches followed by several reads, before anything is even proposed as a change. Both are equally normal; neither is padded or artificially shortened to hit some expected count.

This is also why tCode does not report a time or step estimate up front for a request, described further in Run times: it genuinely does not know how much a task will need until it has started working through it.

This judgment is made fresh on every single turn rather than being influenced by how the previous turn in the same session happened to end. A simple two-step turn followed immediately by a genuinely complex twelve-step one is completely ordinary, not a sign of inconsistency.

The signal the model actually uses to decide it is done is whether it can now give a complete, grounded answer to what you asked, not a fixed checklist or a target number of actions to reach before stopping.