The workspace assistant failed on every question with "The workspace assistant could not reach OpenAI. Try again." The assistant sends its list of available tools with each question, and the two tools that take no arguments described their arguments in a shape the model service rejects. One malformed entry fails the whole request, so no question got through, whatever it asked about. Both tools are now described correctly, and a test checks the shape of every tool the assistant publishes so a new one cannot repeat it. Assistant failures are also recorded with the reason the request was rejected, which had previously been discarded.
The workspace assistant answers again
Every assistant question was failing straight away. The tool list it sends with each question contained a malformed entry, which caused the whole request to be rejected before the question was read.