Control characters in message contents result in request bodies that cause 400s returned from APIs

[preston-m-price]

Consider these messageParams:

messageParams := []anthropic.MessageParam{
  anthropic.NewUserMessage(anthropic.NewTextBlock("What is this control character: \u0002?")),
}

// invoke with streaming, or non-streaming and get a 400 back.

This results in a JSON payload like:

{\"messages\":[{\"content\":[{\"text\":\"What is this control character: \\x02.\",\"type\":\"text\"}],\"role\":\"user\"}]}

Which results in a 400 returned by Anthropic's APIs.

This JSON cannot be parsed in all cases, for example:

anthropicBody := "{\"messages\":[{\"content\":[{\"text\":\"What is this control character: \\x02.\",\"type\":\"text\"}],\"role\":\"user\"}]}"
var data map[string]interface{}
err := json.Unmarshal([]byte(anthropicBody), &data)

results in an error:

invalid character 'x' in string escape code

When testing other model SDKs the resulting JSON payload for a message like this is of the form:

{\"messages\":[{\"content\":\"What is this control character: \\u0002.\",\"role\":\"user\"}]}

Which can be parsed using the strategy above.