Skip to content

kill-switches.md

Latest commit

 

History

History
94 lines (60 loc) · 3.94 KB

kill-switches.md

File metadata and controls

94 lines (60 loc) · 3.94 KB
Raw

Kill Switches

Every loop you start tonight needs a way to die. Know yours before you start it, not while it's committing garbage at 2 AM.

Runaway signs

Pull the switch when you see any of these:

  • The same commit message, over and over
  • Cost ticking up, but git log shows nothing new in the last 20 minutes
  • The same error in the output every iteration, same fix attempted every iteration
  • Your verifier has never exited 0 and the loop has no iteration cap
  • Hundreds of malformed tool calls in minutes — the 400-broken-tool-calls failure mode, see ../verifiers/anti-patterns.md

Ordered from gentle to nuclear:

1. Esc — inside Claude Code

Esc interrupts the current response or tool call mid-turn. Work done so far is kept; you get the prompt back and can redirect. Ctrl+C also interrupts; a second press exits Claude Code entirely. If a /loop is driving the session, interrupt it the same way, then tell it to stop.

Ctrl+X Ctrl+K (press twice within 3 seconds) stops all background subagents running in the session. /tasks shows what's running.

2. Ctrl-C — the foreground bash loop

A Ralph loop running in your terminal dies with Ctrl+C. The signal hits the whole foreground job — the while true and the claude -p inside it. If an iteration is mid-flight and the loop respawns, press it again.

3. Kill a background or tmux loop

Find it first:

pgrep -fl claude          # list matching processes with their PIDs

Kill headless loops:

pkill -f 'claude -p'      # kills every claude running in print mode — all of them

If it lives in tmux:

tmux ls                        # find the session name
tmux kill-session -t <name>    # kill it, and everything inside it

If the loop keeps respawning children faster than you can kill them, kill the process group, not the process:

ps -o pgid= -p <PID>      # get the group id
kill -TERM -- -<PGID>     # note the leading minus: the whole group
kill -KILL -- -<PGID>     # if TERM was ignored

4. Claude Code background agents

Sessions started with claude --bg are managed from the agent view:

claude agents             # open the background session manager; stop sessions from there
claude agents --json      # list active sessions for scripting

5. Cloud sessions

Sessions running on Anthropic's servers don't care about your terminal. Stop them from the session view at claude.ai/code. This works from your phone at breakfast.

6. Revoke the API key

The true kill switch. It works when you've lost track of what's running where — tmux on a box you can't reach, a background session you forgot, three worktrees deep.

  • Anthropic: console.anthropic.com → API Keys → disable or delete the key.
  • OpenAI: platform.openai.com → API keys → revoke.

Every in-flight request starts failing. No loop survives it. If you authenticated with a subscription instead of a key, run /logout in Claude Code or revoke the session from your claude.ai account settings.

This is why tonight's loop should run on its own key: revoking it kills the loop without killing everything else you own.

Cleaning up after

The loop is dead. Now look at what it left behind:

git status                     # uncommitted changes?
git log --oneline -20          # what did it commit?
git stash -u                   # park the uncommitted mess without losing it
git worktree list              # did it leave worktrees behind?
gh pr list --author "@me"      # did it open PRs?

Do not git reset --hard until you've read the diff. Runaway loops sometimes leave good work in the wreckage — the runaway was the stopping, not necessarily the writing. Close or mark draft any PRs you don't trust yet.

Then fix the reason it ran away — 90% of the time it's a verifier that never returns 0. See ../verifiers/good-vs-bad.md before you relaunch.