Option to specify that model supports Prompt Caching with OpenAI Compatible provider

[dleen]

We are using a gateway to access models (similar to OpenRouter) and we are using anthropic.claude-3-7-sonnet-20250219-v1:0 which does support prompt caching. It would be great to have a checkbox like "Computer Use" to tell Roo that this model supports prompt caching.

[mrubens]

Makes sense @dleen! I assume you'd also want to be able to set the prices for the cached input/output?

[dleen]

Absolutely! Is it similar to Cline testing where I checkout the branch and run the local build?

[mrubens]

Yes, I think https://github.com/RooVetGit/Roo-Code?tab=readme-ov-file#local-setup--development is up to date but please let me know if you run into any issues.

[dleen]

Confirm that the UI works fine, but I realised it won't do anything without a provider code change since OpenAI does the prompt caching by default on their backend. The good news is that the logic is the exact same as OpenRouter: https://github.com/RooVetGit/Roo-Code/blob/main/src/api/providers/openrouter.ts#L61-L98

My suggestion is: if the OpenAI compat model from the UI has the prompt caching option as True then add the anthropic style cache control like OpenRouter does. If needed we can update the UI to say "Prompt Caching (Anthropic Compatible)" or something to indicate that we will use the anthopic style cache param vs others (e.g. the bedrock style).

I can take a stab at the change if you want

[mrubens]

Sounds great - go for it!

[dleen]

Done! #1587