# Approvals

*How-to: put a human between an agent and anything irreversible.*

Mark a tool `approval: "always"` (see [Defining agents](defining-agents.md)). When the model calls it, the run **parks** durably at zero compute until someone decides:

```bash
toren jobs list --dir my-crew
#  r_9f2c1a  my-crew  waiting_approval  (send_report)

toren jobs show r_9f2c1a --dir my-crew
#  pending approval: send_report {"to":"board@fund.com"}
#  → toren jobs approve r_9f2c1a w1t0 s4

toren jobs approve r_9f2c1a w1t0 s4                     # or:
toren jobs approve r_9f2c1a w1t0 s4 --deny --comment "wrong list"
```

Approve → the tool executes exactly once with the originally recorded arguments and the run continues. Deny → the model receives your comment as a tool error and decides how to proceed (typically it adapts or reports back).

Because parking is durable, an approval can arrive seconds or days later. Nothing polls, nothing bills, and the run survives restarts while it waits. Programmatic equivalents: `listPendingApprovals()` / `resolveApproval()` in `@toren-run/core`.