A tiny, local, read-only Claude Code plugin. Run one command and see how your agent actually behaved across every session you've ever run — how often it woke up and did the same thing again:
- how often it repeated work it already did,
- which file it re-touched the most,
- how many no-op commands it fired (
true,echo,pwd…), - and — the interesting one — how often, when it looked at something twice, the world had moved underneath it.
Everything runs on your machine. Nothing is uploaded. It never edits anything.
/plugin marketplace add fraqtl-ai/groundhog
/plugin install groundhog@fraqtl-ai
Then in any Claude Code session:
/groundhog:stats
This is the friendly, single-developer front end to a bigger piece of work: measuring tool-call reproducibility in coding agents. On one person's interactive history there's almost no money to save — you're in the loop correcting the agent constantly. The number that is interesting is how often a repeated observation comes back different: on large autonomous agent runs (no human in the loop) that rate is high, and it's why naively caching agent tool-outputs is unsafe. Groundhog lets you see a small version of that on your own history.
Read-only. Local. bytes/4 token proxy. No cost claims — on purpose.
Built by fraQtl.