I build small tools for the coding-agent ecosystem, mostly the unglamorous part: checking that the things people drop into their agents actually work.
skillprobe lints agent skills — will it trigger, what does it cost you on every turn, does it collide with another one. Three things that fail silently, so nobody notices until something behaves oddly and there's nothing in the logs to explain it.
hookprobe does the same for Claude Code hooks, against the real event schema: events that don't exist, tool-name casing that quietly never matches, matchers on the events that ignore them.
Both got most of their accuracy from being run against real shipped configs rather than my own fixtures — which is how I found out my fixtures agreed with my bugs.
The rest of the time I send fixes upstream. Recently to langchain-ai/openwiki, Tencent/BrowserSkill, apache/maka and debpalash/VoiceStudio.