Day 4 of Ask TanStack Query: the model stopped guessing
On Day 1, a model with no docs in front of it answered TanStack Query questions from memory. Two of those answers were wrong, and both looked like real code. Today the same kind of question had to be answered from the cards we stored, and each answer had to name the page it used.
The short version
I asked three questions in the terminal.
One came back right. One came back honest and incomplete. One came back as a refusal, which was the correct answer.
The question it got right
"What is the difference between staleTime and gcTime?"
On Day 1 the model told me to use cacheTime, a name TanStack Query v5 renamed. Today it said gcTime is that rename, and it pointed at the migration page. staleTime is how long data stays fresh. gcTime is how long unused data stays in memory.
The first try stopped after the staleTime section. The model had spent its whole length budget thinking. I raised the ceiling, ran it again, and the gcTime section appeared.
The question it could not find
"How do I run code after a query succeeds?"
On Day 1 the model invented onSuccess on useQuery. Version 5 removed that. Today it did not invent it. It said, "I couldn't find this in the docs."
That is only half a win. The docs do explain the replacement. The search brought five other pages, and the answer was not on them. The model followed the rule: if the pages do not contain the answer, do not guess. The miss was the search.
The question that should fail
"What does queryClient.teleport() do?"
There is no such method. The model said it could not find it in the docs. On Day 1, a question like that often got a confident, invented answer.
What changed
With the pages in front of it, the model stops filling gaps from memory. A wrong answer now usually means the search brought the wrong cards. That is a better failure. You can see which pages it used.
Frequently asked questions
Did the assistant get every Day 1 question right?
No. It corrected cacheTime to gcTime. It stopped inventing onSuccess, and it also failed to find the page that explains what to do instead.
Why is "I couldn't find this" a good answer?
Because the other choice is a confident guess. The rule is: use only the pages you were given, and say so when they do not contain the answer.
Why did the first answer get cut off?
The model thinks before it writes, and that thinking counts against a length limit. The limit was too low. Raising it let the answer finish.
Next up
The next job is the quiz: real questions with answers a maintainer marked correct, so these three examples stop being a feeling and start being a score.
Related: Day 2 of Ask TanStack Query: I went and got the books and the quiz