Skip to content

Switchyard lets LLM applications route traffic across models and providers while preserving native OpenAI and Anthropic API compatibility - enabling flexible model selection, benchmarking, and cost/performance optimization.

Rust

fix(protocol): normalize Anthropic's refusal stop reason when streaming (#941)

Dmitry Voropaev8e7d72be3847
diff --git a/crates/protocol/src/stream.rs b/crates/protocol/src/stream.rs
index 428d01a0f..04dfbfe01 100644
--- a/crates/protocol/src/stream.rs
+++ b/crates/protocol/src/stream.rs
@@ -505,7 +505,9 @@ fn stop_reason_from_str(reason: Option<&str>) -> StopReason {
match reason {
Some("length" | "max_tokens") => StopReason::MaxTokens,
Some("tool_calls" | "function_call" | "tool_use") => StopReason::ToolUse,
- Some("content_filter") => StopReason::ContentFilter,
+ // Anthropic spells a content-policy stop `refusal`; the buffered codec
+ // already maps it (anthropic/buffered.rs, map_anthropic_stop_reason).
+ Some("content_filter" | "refusal") => StopReason::ContentFilter,
Some("stop" | "end_turn" | "stop_sequence") | None => StopReason::EndTurn,
Some(_) => StopReason::Unknown,
}
@@ -562,6 +564,33 @@ mod tests {
);
}
+ #[test]
+ fn stop_reason_spellings_normalize_the_same_streamed_as_buffered() {
+ // A stop reason has to mean the same thing whether the response arrived
+ // buffered or was folded out of a stream. The expectations here are the
+ // ones the codecs apply to a buffered body.
+ for (reason, expected) in [
+ ("length", StopReason::MaxTokens),
+ ("max_tokens", StopReason::MaxTokens),
+ ("tool_calls", StopReason::ToolUse),
+ ("tool_use", StopReason::ToolUse),
+ ("content_filter", StopReason::ContentFilter),
+ ("refusal", StopReason::ContentFilter),
+ ("stop", StopReason::EndTurn),
+ ("end_turn", StopReason::EndTurn),
+ ] {
+ let agg = fold(vec![LlmResponseChunk::MessageStop {
+ reason: Some(reason.to_string()),
+ }]);
+
+ assert_eq!(
+ agg.outputs.first().and_then(|output| output.stop_reason),
+ Some(expected),
+ "stop reason {reason:?}"
+ );
+ }
+ }
+
#[test]
fn aggregates_normalized_chunks_inside_stream_event() {
let response = LlmResponse::Stream(Box::pin(stream::iter([Ok(