4 min read
Kill switch & loop detection
Stop one run or every local run. Understand what stopped, and resume deliberately.
On this page
Stop a session or activate panic
A session kill prevents new requests and aborts active local streams. Global panic is a persistent latch that survives proxy restarts. The dashboard sends a command to the relevant device, or broadcasts workspace panic to enrolled devices.
leash status
leash kill SESSION_ID
leash kill --allResume explicitly
Clearing panic does not resume every killed session. Inspect the reason first, clear the panic latch if it is set, then resume the individual session. Add budget only when you intend to authorize more spend.
leash resume --all
leash resume SESSION_ID
# Explicitly add $2 to that session cap:
leash resume SESSION_ID --add 2Automatic stops for runaway patterns
The policy engine watches repeated tool/request fingerprints, request volume, subagent count and spend velocity. These signals are deterministic safety rules, not a judgment of whether the task is making useful progress.
| Signal | Default policy |
|---|---|
| Repeated fingerprint | Stop on the 8th matching request within 3 minutes |
| Request count | 800 requests per session |
| Subagent count | 10 allowed; further spawns stop |
| Velocity | 5-minute rate above the configured baseline threshold |
Something unclear?Help us improve this guide