I was working on a tool for identifying bugs and noticed that sometimes you have to manually remove parts of a broken request and rerun it to obtain a minimal reproducible example. I decided to automate this process and created an open-source CLI tool called ‘toolcall doctor’. It takes a failing request, the error condition and the parts that must not be lost, and then attempts to minimise the request, checking each time that the specified error still occurs. In one of the tested examples, the request size was reduced from 583 to 185 bytes. The project is still in its early stages: it does not automatically diagnose the root cause, and the failure condition and ‘keepers’ are specified by the developer. My top priority right now is to find people with genuine tool-calling bugs and to understand whether the tool helps them.
I was working on a tool for identifying bugs and noticed that sometimes you have to manually remove parts of a broken request and rerun it to obtain a minimal reproducible example. I decided to automate this process and created an open-source CLI tool called ‘toolcall doctor’. It takes a failing request, the error condition and the parts that must not be lost, and then attempts to minimise the request, checking each time that the specified error still occurs. In one of the tested examples, the request size was reduced from 583 to 185 bytes. The project is still in its early stages: it does not automatically diagnose the root cause, and the failure condition and ‘keepers’ are specified by the developer. My top priority right now is to find people with genuine tool-calling bugs and to understand whether the tool helps them.