Use human_review for a general checkpoint. Approval-gated and manual tools populate run.pause.pending_tool_calls with exact call IDs.
Continue through Agent#continue_run with:
ToolDecision values in approvals or rejections
ToolOutput values for manual work
human_input for a general review pause
- the prior run, or
run_id with a state accessor
Decisions use exact call IDs, so concurrent calls to the same tool remain distinct. A compatible state object implements load(run_id) and save(result).
Last modified on July 26, 2026