One turn is a loop, not a single request: the model is asked, its tool calls are run, the results are sent back, and this repeats until it answers without asking for a tool. Every turn starts with an announced task and, when the work has more than a few steps, a plan.

Nothing caps the number of requests in a turn: the loop ends when the model answers without asking for a tool. A turn that never gets there goes on until something interrupts it: Esc at the terminal, or the stop a sub-agent's caller asks for.

Tools run on the host in this process; the ones that start a program (bash, the searches) carry the deadline, output bound and process-group cleanup described in Tools and Security and state. Esc stops the loop and its running tools, see Interface.

The main model judges completion itself. A final answer without tool calls ends the turn directly (exit code 0), without a status marker, independent checker request or automatic repair round. The CLI prints Turn finished to indicate that the turn ended, not that the requested work necessarily succeeded.

The model must validate its work before claiming success and explain partial work, failures, missing permissions/dependencies or required user input in its answer. Answer text is not parsed as a completion protocol. Blocked, failed, waiting-for-input and paused turns each read differently on the screen, and none of them prints Task completed.

Transport errors and truncation still return errors.

Transient network errors (HttpConnectionClosing, a reset connection, a broken pipe, a timeout, an unreachable network) and HTTP 408/429/500/502/503/504 responses are retried, up to three attempts with a one and then two second pause, each attempt on a fresh connection. A retry is only sent while no answer text has been shown: once the reply is streaming onto the screen, a failure ends the turn with its error rather than printing the answer twice.

Output that hits the token budget is continued, and eventually reported as ResponseTruncated; the budget and its growth are described in Configuration.

A sub-agent runs the same loop on the two clocks its caller set: every self-check interval it reports its progress on the terminal, and at its deadline it stops and writes the final report the caller reads. Its tokens are added to the calling turn.