flipback: a Stop hook that sends Claude Code back when a turn ends before the work is done

[fainir]

I kept ending Claude Code turns with a message that sounded finished when it was not: one of three parts done, or "Now let me run the tests." and then nothing. So I built flipback, a Claude Code plugin (a Stop hook plus a skill) that sends the agent back when a turn ends before the request is done.

How it works:

• A free local check reads the end of the transcript: your latest request, the tools used since, and the final message. It looks for signs of an early stop (a failed last tool call, announced but not done work, a question into an empty room, a TODO, edits with nothing run).
• Only if something looks off, claude -p --model haiku judges whether the work is really unfinished. If so, the agent gets one short note naming the missing piece and keeps going.
• It stays quiet when you asked a question, said stop, or the agent needs something only you can give (a credential, a payment, a decision). One nudge per turn at most.

Measured on a month of my own sessions (823 turn endings): 86% never reached a model; it sent the agent back on 36 turns and was right on 30 of them. It also misses some by design, so for a must-finish job /goal is still the tool. The README has the method and its limits.

Install, in a Claude Code session:

/plugin marketplace add fainir/flipback
/plugin install flipback@flipback

Needs Node 18+ and the claude CLI; no API key or config. MIT.

Repo: https://github.com/fainir/flipback?utm_source=github-anthropics-skills

Feedback on false nudges is very welcome, those are the cases I tune against. I'm the author.