Stop runaway runs¶

clawtop kill stops whole run stacks: an agent run's process and everything it started (MCP servers, helpers), together.

Pick interactively¶

clawtop kill

It lists the runs with numbers. Answer with row numbers (1 3), stale, or all. Press Enter to cancel. It then shows what will be killed and how much memory that frees, and asks for confirmation.

Pick by rule¶

clawtop kill 4102 2211          # specific stacks, by the PID in the table
clawtop kill --idle 30          # no transcript activity for 30 minutes (runtimes with a transcript)
clawtop kill --older-than 120   # runs older than 2 hours (any runtime)
clawtop kill --orphans          # gateway-spawned stacks whose parent is gone

Options combine: a stack matching any selector is included.

Option	Effect
`--dry-run`	Show the selection and stop
`--yes`, `-y`	Skip the confirmation
`--grace SECONDS`	Wait this long after SIGTERM before SIGKILL (default 5)

The exit code is non-zero if any process survived.

What makes it safe¶

Only gateway-spawned processes can be selected. A process counts only if it is in the gateway's process tree (or carries OpenClaw's gateway markers in its environment) and runs in an agent workspace. A claude, gemini or npx you started from your own shell is never listed. See what counts as a gateway run.
Whole stacks die together: SIGTERM first, then SIGKILL for survivors. No half-dead MCP helpers are left holding memory.
PID reuse is guarded. The command line of every process is re-checked just before it is signalled. If the PID now belongs to a different program, it is skipped.

What happens to the session¶

Killing a stack stops the process. The gateway will see the run end; clawtop does not edit the session itself. To close sessions afterwards, use sessions prune.

Warning

You cannot tell from the table which Discord channel or cron job a run belongs to. clawtop shows the runtime and the agent. Use IDLE and CTX (transcript size) where they are available. Runs that are working right now show a small IDLE. For a runtime with no activity source IDLE is -, so check AGE and MEM and use --older-than or name the PID.