When the agent stops
In short
A stopped turn is usually governance working as designed: the agent needs an answer, needs your approval on a //!SAFE diff, hit a failing check, or reached an action that requires a verified expert's verdict.
Four things stop a turn. Only one of them is a failure.
1. It asked you a question
The agent pauses when an assumption would be load-bearing and it cannot verify it — a missing business rule, two plausible readings of your request. Answer in the thread and it continues with the answer as context.
2. A //!SAFE change is waiting for approval
The proposal is a diff. Approve it, reject it, or reply with the correction. Nothing lands in that zone without you, by construction — see ownership zones.
3. A check failed
Generated changes are verified before they land — type checks, tests, automated review. When verification fails, the turn rolls back rather than leaving a half-applied change, and the failure is preserved so it can be inspected. Read what failed and re-brief; a rolled-back turn usually means the request was under-specified, not that the agent is stuck.
4. The action needs an expert
Some actions are gated on human expertise rather than on your approval — the categories where a confident wrong answer is expensive, such as security posture, production data, and architecture decisions with long consequences. The agent stops and prepares a review brief describing what was asked, what it was about to do, and why it stopped.
You choose whether to send it. A verified expert returns a verdict — approve, approve with changes, reject, or need more information — and an approval clears the gate exactly once, for exactly the reviewed action. Expert time is priced by job size and quoted before work starts.
What not to do
Do not work around a gate by moving the code to //!AUTO and re-asking. The gate is a
judgement about the action’s risk, not about the file.