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(translation): preserve reasoning order in mixed stream chunks (#387)

Blakeolson216babb3b5d4e7
diff --git a/crates/switchyard-translation/src/codecs/openai_chat/stream.rs b/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
index a50dcded1..6972850e9 100644
--- a/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
+++ b/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
@@ -98,14 +98,6 @@ fn decode_openai_chat_stream(
continue;
};
if let Some(delta) = choice.get("delta").and_then(Value::as_object) {
- if let Some(text) = delta.get("content").and_then(Value::as_str)
- && !text.is_empty()
- {
- out.push(LlmResponseChunk::TextDelta {
- index: 0,
- text: text.to_string(),
- });
- }
if let Some(details) = delta
.get("reasoning_details")
.and_then(Value::as_array)
@@ -134,6 +126,14 @@ fn decode_openai_chat_stream(
}
}
}
+ if let Some(text) = delta.get("content").and_then(Value::as_str)
+ && !text.is_empty()
+ {
+ out.push(LlmResponseChunk::TextDelta {
+ index: 0,
+ text: text.to_string(),
+ });
+ }
if let Some(tool_calls) = delta.get("tool_calls").and_then(Value::as_array) {
for tool_call in tool_calls {
if let Some(tool_call) = tool_call.as_object() {
diff --git a/crates/switchyard-translation/tests/stream_translation.rs b/crates/switchyard-translation/tests/stream_translation.rs
index a938a3094..c726bee06 100644
--- a/crates/switchyard-translation/tests/stream_translation.rs
+++ b/crates/switchyard-translation/tests/stream_translation.rs
@@ -5,8 +5,10 @@
pub mod common;
+use std::collections::HashMap;
+
use pretty_assertions::assert_eq;
-use serde_json::json;
+use serde_json::{Value, json};
use switchyard_protocol::{LlmResponseStreamEvent, ResponseAccumulator, StopReason};
use switchyard_translation::{
LlmResponseChunk, StreamTranslationState, TranslationEngine, WireFormat, decode_stream_event,
@@ -16,6 +18,48 @@ use common::{REASONING_MODEL, text_and_encrypted_reasoning_details};
type TestResult = std::result::Result<(), Box<dyn std::error::Error + Send + Sync>>;
+// Reduces Anthropic stream events to ordered labels (`<block>_start`, `<delta>`,
+// `<block>_stop`) so ordering assertions stay readable without restating each payload.
+fn event_labels(events: &[Value]) -> Vec<String> {
+ let mut block_types: HashMap<u64, String> = HashMap::new();
+ events
+ .iter()
+ .map(|event| {
+ let index = event
+ .get("index")
+ .and_then(Value::as_u64)
+ .unwrap_or_default();
+ match event
+ .get("type")
+ .and_then(Value::as_str)
+ .unwrap_or_default()
+ {
+ "content_block_start" => {
+ let block_type = event
+ .get("content_block")
+ .and_then(|block| block.get("type"))
+ .and_then(Value::as_str)
+ .unwrap_or_default()
+ .to_string();
+ block_types.insert(index, block_type.clone());
+ format!("{block_type}_start")
+ }
+ "content_block_delta" => event
+ .get("delta")
+ .and_then(|delta| delta.get("type"))
+ .and_then(Value::as_str)
+ .unwrap_or_default()
+ .to_string(),
+ "content_block_stop" => {
+ let block_type = block_types.get(&index).cloned().unwrap_or_default();
+ format!("{block_type}_stop")
+ }
+ other => other.to_string(),
+ }
+ })
+ .collect()
+}
+
// Same-format replay returns the same parsed JSON value, including provider-specific fields.
#[test]
fn preserved_same_format_events_replay_unknown_fields_exactly() -> TestResult {
@@ -356,6 +400,48 @@ fn openai_chat_stream_event_translates_to_anthropic_message_events() -> TestResu
Ok(())
}
+// A mixed chunk must emit reasoning before text, matching the buffered decoder.
+#[test]
+fn openai_chat_mixed_reasoning_and_content_stream_in_reasoning_first_order() -> TestResult {
+ let engine = TranslationEngine::default();
+ let mut state =
+ StreamTranslationState::new(WireFormat::OpenAiChat, WireFormat::AnthropicMessages);
+ let chunk = json!({
+ "id": "chatcmpl-test",
+ "object": "chat.completion.chunk",
+ "model": "nvidia/nvidia/nemotron-3-ultra-nvfp4",
+ "choices": [{
+ "index": 0,
+ "delta": {
+ "reasoning_content": ".",
+ "content": "Hello"
+ },
+ "finish_reason": null
+ }]
+ });
+
+ let events = engine.translate_event(
+ &mut state,
+ WireFormat::OpenAiChat,
+ WireFormat::AnthropicMessages,
+ &chunk,
+ )?;
+
+ assert_eq!(
+ event_labels(&events),
+ vec![
+ "message_start",
+ "thinking_start",
+ "thinking_delta",
+ "signature_delta",
+ "thinking_stop",
+ "text_start",
+ "text_delta",
+ ]
+ );
+ Ok(())
+}
+
// Verifies Anthropic usage and stop events become terminal OpenAI chunks.
#[test]
fn anthropic_stream_usage_and_stop_translate_to_openai_chunks() -> TestResult {