Fix an issue with retry counting

[dmontagu]

This is related to #734 and came from my investigations of that issue, though I'll note that I don't think merging this will resolve the issue.

Right now, when processing function tools, we increase the retry count for each unknown tool call received, which I think is incorrect behavior — I think we should increment the retry count at most once per response.

This PR currently makes that change, but that change causes a new issue — with the implementation in this PR, raising a ModelRetry contributes to the maximum model retry count, rather than just to the tool's retries. (The test I've left failing demonstrates the issue.)

I'm not sure if I have surfaced a bug, and we should require you to increase the number of retries on Agent to avoid this raising the UnexpectedModelBehavior exception, or if the pre-existing behavior was preferred.

If the pre-existing behavior is preferred, I personally think the current distinction between treating unknown tool calls as incrementing the retry count, but tool calls that raise a ModelRetry not incrementing the retry count is rather confusing for end users (I mean, I was confused by it trying to understand what was happening in the example code). But even if you disagree with that, I definitely think the current implementation is confusing for developers — I've had to reverse engineer exactly what scenarios increment the retry count vs. not, and I think if I can get confirmation of the desired behavior, I can refactor the code to have that behavior (whatever it is) but also be easier to understand.

[cloudflare-workers-and-pages]

Deploying pydantic-ai with    Cloudflare Pages

Latest commit:	b48558c
Status:	✅  Deploy successful!
Preview URL:	https://cf236733.pydantic-ai.pages.dev
Branch Preview URL:	https://dmontagu-fix-retry-counting.pydantic-ai.pages.dev

View logs