Very cool how this is currently spoofed with tool calls :) Just found this when working on some course materials, though looks like it'll probably will be solved by #866.
library(ellmer)
colors <- type_string(required = FALSE)
chat <- chat_github()
chat$chat_structured("What's my favourite colour?", "My favourite food is cake", type = colors)
#> Using model = "gpt-4.1".
#> NULL
chat <- chat_anthropic()
chat$chat_structured("What's my favourite colour?", "My favourite food is cake", type = colors)
#> Using model = "claude-sonnet-4-5-20250929".
#> [1] "I don't have information about your favourite colour. You mentioned that your favourite food is cake, but you haven't told me what your favourite colour is. Would you like to tell me your favourite colour?"
It's a bit trickier when the type is, say, an integer not a string as it'll still return the string so breaks the type "guarantee" now.
library(ellmer)
fave_num <- type_integer(required = FALSE)
chat <- chat_anthropic()
#> Using model = "claude-sonnet-4-5-20250929".
chat$chat_structured("What's my favourite number?", "My favourite food is cake", type = fave_num)
#> [1] "<UNKNOWN>"