A Stop hook for unfinished Claude Code turns

This walkthrough was created using an AI model from Flipback's published code, replay results, and offline tests. When coding tasks are given, sometimes the agent provides one task and a message indicating all three tasks along with a reminder. The assistant added a "slugify()" function to "src/text.ts" and acknowledged testing it.

However, the reminder regarding tests and an example in the README was never followed, as the turn ended without any further action. Flipback is a Claude Code plugin that detects unfinished work at the end of a turn. If Flipback identifies a clear unfinished step, it can prompt the agent to continue. In the example scenario, the reminder could be: The request had three parts and you did one.

Your last message announced the rest, but no tool calls followed. The reminder could be issued with the "/goal" command to address a job with a checkable finish line. Flipback functions as a Stop hook, which reads the final message of a transcript after a turn ends. It allows the agent to stop if a Stop hook has already been sent, if a subagent is due back, the final message is short, and nothing failed, or if the user explicitly says "stop."

A local check performed at this point looks for signs of an early stop, such as a failed tool call, a statement about running work, asking for assistance, mentioning a TODO or placeholder, unchecked checkboxes, or edited files without running. If no match is found, the agent stops, and a small model judges if the work was incomplete.

Claude Code resets its count of consecutive Stop continuations after each tool call. If multiple hooks are used in an unattended print-mode run, a finite --max-turns limit should be set. The measurement is based on real sessions from 823 turn endings across 262 interactive Claude Code sessions. Flipback's local check selected 86% of turn endings to bypass the Haiku judge, while the second model, Claude Opus 5.5, reviewed the remaining 14% and marked 30 of them as unfinished (83%).