proposed_action patch. The agent proposes it, the run pauses in awaiting_outcome, and waits for you to say what happened.
This is the mirror image of a confirmation: there we ask permission before acting, here you act and then report.
Finding a pending action
GET /v1/loom/threads returns pending_action on the thread:
args is what the agent proposed. Carry it out however your product does.
Reporting
detail is optional, recorded on the outcome, and passed to the model. A short structured reason produces a better next turn than a bare failed.
The response is 202 and the run resumes.