...and actually some things were not true: only OpenAI APIs do not support loading files dynamically, and updating our MCP accordingly does give us the flexibility we were looking for!
kudos to opencode for documenting this!
i'd been thinking about sandboxing in the way the ramp team described, but this was way too much of an endeavour for us, and it meant we'd also have to rebuild "observability" and evals from scratch
so anyway, codemode is awesome, the starter repo is really good to get a sense of it, kudos!
the example looks silly, there's only two documents and it looks like there's only ever going to be two of them, but our usecases are much more dynamic, e.g. load that document if this - meaning we would have had to manage pipelines on our end, which I desperately wanted to avoid
once the "provider agent" gets the publicly available (signed) url, they can't (won't?) fetch it and use it in the way they'd do if we did another api call with a "file" message part
google, openai & anthropic behave the same here
...in that they all refuse to dynamically load documents in their context, e.g. this use case (we expose a `get_download_url` through our MCP)
```
load these two files
- s3://bucket/file1
- s3://bucket/file2
and tell me if the billing address is the same
```
just want to thank @kentonvarda.com @threepointone.bsky.social and the cloudflare ai teams for codemode 🙏
i've been meaning to try it for some time, and paired with sonnet it's just opened up really cool stuff for us we couldn't do with major providers...
*Please repost* @sjgreenwood.bsky.social and I just launched a new personalized feed (*please pin*) that we hope will become a "must use" for #academicsky. The feed shows posts about papers filtered by *your* follower network. It's become my default Bluesky experience bsky.app/profile/pape...
