A safety cap on tool calls

There is an upper limit on how many tool calls a single turn is allowed to make. It is a safety limit rather than something you would normally notice in day-to-day use, sized well above what any reasonable request needs.

Its purpose is narrow and specific: catching the case where a model gets stuck reasoning in an unproductive circle, repeatedly trying variations of the same failed approach without making real progress, and making sure that state ends with a report back to you rather than running indefinitely.

If a turn genuinely does hit the cap, tCode tells you plainly that this happened rather than silently cutting the response off partway through, so a capped turn is never mistaken for a normally completed one.

Hitting the cap is rare enough in normal use that most people never encounter it across an entire project. When it does happen, it is usually a signal worth paying attention to on its own, that the request as phrased may have been genuinely ambiguous or was pointing at something that does not actually exist in the codebase the way it was described.

If a turn does hit the limit, the practical next step is the same as correcting any other unsatisfying result: narrow or clarify the request based on what tCode already found, described in Correcting course mid-turn, rather than repeating the identical request and expecting a different outcome.