Loop Killer — Development skill for Claude Code

loop-killer

Claude Code plugin that ends a stop-hook re-prompt loop. When a stop hook (for example an active `/goal`) keeps re-prompting the session and the replies make no tool calls, loop-killer runs `/goal clear` and says so.

How detection works

At the end of every main-session turn the plugin reads the conversation and takes the trailing run of re-prompt turns with no tool work:

• A re-prompt turn is a user message that begins with Stop hook feedback:, Stop hook (optionally preceded by [), or Hook re-prompted, ignoring case.
• The assistant reply to it counts when it made no tool call.
• The run ends at the first pair that is not a re-prompt or whose reply used a tool.

The goal is cleared when the run holds either:

• 3 or more replies that are similar to one another, or
• 6 or more replies, whatever they say.

Two replies are similar when the words they share make up at least 60% of all the distinct words in the two replies (word-set Jaccard overlap). Words are lowercase runs of three or more ASCII letters or digits. Every consecutive pair in the run must be similar.

Threshold	Value	Meaning
SAME_REPLY_TURNS	3	Trailing re-prompt turns with similar replies and no tool work
IDLE_TURNS	6	Trailing re-prompt turns with no tool work, any replies
SIMILAR	0.6	Minimum word overlap between consecutive replies

After a kill, the next one can happen no sooner than 10 main-session turns later. Subagent turns are ignored and not counted. The thresholds are constants in `hooks/register.ts`.

When it fires, the plugin runs `/goal clear`, shows a toast, writes a log line, and appends a one-line notice to the turn's answer. The notice names the rule that fired and whether `/goal clear` succeeded.

Install

claude plugin marketplace add cavi-ai/plugins

Then, in a session:

/plugin install loop-killer@cavi-ai

Uninstall

/plugin uninstall loop-killer@cavi-ai

L