Interim Anthropic Prompt Caching Workaround

[ronakrm]

"Enables" prompt cache control through the system prompt for Anthropic models.

It's basically just a hack that -very roughly- checks if your system prompt string looks like a list, assumes that its been json.dumped, json.loads it, and converts everything to Anthropic's format.

Also adds the returned cache usage values to pydantic-ai's Usage.details, so you can see if it's working or not.

The way I'm using this now is by:

system_prompt_json = [
    ...
    {
        "type": "text",
        "text": my_string_prompt,
        "cache_control": {"type": "ephemeral"},
    }
    ...
]
system_prompt_string = json.dumps(system_prompt_json)
agent = Agent[AgentConfig](
    system_prompt=system_prompt_string,
    ...

This shouldn't be merged, but I thought I'd put it up in case others found this useful. Really happy with pydantic-ai and this was one thing that would have otherwise forced me to write/wrap a bunch of other stuff just to get around this.