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): decode OpenRouter cache writes (#141)

nachiketb-nvidia9a4393065837
diff --git a/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs b/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
index 6c0ad424f..bc3688b37 100644
--- a/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/openai_chat/buffered.rs
@@ -1072,7 +1072,11 @@ pub(crate) fn decode_openai_usage(value: Option<&Value>) -> Usage {
.and_then(Value::as_u64);
let cache_creation_input_tokens = value
.get("prompt_tokens_details")
- .and_then(|details| details.get("cache_creation_tokens"))
+ .and_then(|details| {
+ details
+ .get("cache_write_tokens")
+ .or_else(|| details.get("cache_creation_tokens"))
+ })
.and_then(Value::as_u64);
Usage {
input_tokens: value
diff --git a/crates/switchyard-translation/src/codecs/openai_chat/stream.rs b/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
index c74774046..b2dbca718 100644
--- a/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
+++ b/crates/switchyard-translation/src/codecs/openai_chat/stream.rs
@@ -244,7 +244,11 @@ fn openai_usage(usage: &Map<String, Value>) -> Usage {
.and_then(Value::as_u64);
let cache_creation_input_tokens = usage
.get("prompt_tokens_details")
- .and_then(|details| details.get("cache_creation_tokens"))
+ .and_then(|details| {
+ details
+ .get("cache_write_tokens")
+ .or_else(|| details.get("cache_creation_tokens"))
+ })
.and_then(Value::as_u64);
Usage {
input_tokens: usage
diff --git a/crates/switchyard-translation/tests/response_translation.rs b/crates/switchyard-translation/tests/response_translation.rs
index 6935467a9..3e61f0151 100644
--- a/crates/switchyard-translation/tests/response_translation.rs
+++ b/crates/switchyard-translation/tests/response_translation.rs
@@ -159,38 +159,42 @@ fn openai_chat_cache_usage_translates_to_responses_usage_details() -> TestResult
Ok(())
}
-// Verifies OpenAI cache usage is expressed in Anthropic's standard sibling fields.
+// Verifies OpenRouter's cache-write field and the legacy alias normalize identically.
#[test]
-fn openai_chat_cache_usage_translates_to_anthropic_usage_fields() -> TestResult {
+fn openai_chat_cache_write_aliases_translate_to_anthropic_usage_fields() -> TestResult {
let engine = TranslationEngine::default();
- let body = json!({
- "id": "chatcmpl-test",
- "model": "gpt-cached",
- "choices": [{
- "index": 0,
- "message": {"role": "assistant", "content": "Cached answer"},
- "finish_reason": "stop"
- }],
- "usage": {
- "prompt_tokens": 100,
- "completion_tokens": 5,
- "total_tokens": 105,
- "prompt_tokens_details": {"cached_tokens": 80}
- }
- });
-
- let output = engine
- .translate_response(
- WireFormat::OpenAiChat,
- WireFormat::AnthropicMessages,
- &body,
- &TranslationPolicy::default(),
- )?
- .body;
-
- assert_eq!(output["usage"]["input_tokens"], 20);
- assert_eq!(output["usage"]["cache_read_input_tokens"], 80);
- assert_eq!(output["usage"]["output_tokens"], 5);
+ for cache_write_field in ["cache_write_tokens", "cache_creation_tokens"] {
+ let mut body = json!({
+ "id": "chatcmpl-test",
+ "model": "gpt-cached",
+ "choices": [{
+ "index": 0,
+ "message": {"role": "assistant", "content": "Cached answer"},
+ "finish_reason": "stop"
+ }],
+ "usage": {
+ "prompt_tokens": 100,
+ "completion_tokens": 5,
+ "total_tokens": 105,
+ "prompt_tokens_details": {"cached_tokens": 70}
+ }
+ });
+ body["usage"]["prompt_tokens_details"][cache_write_field] = json!(10);
+
+ let output = engine
+ .translate_response(
+ WireFormat::OpenAiChat,
+ WireFormat::AnthropicMessages,
+ &body,
+ &TranslationPolicy::default(),
+ )?
+ .body;
+
+ assert_eq!(output["usage"]["input_tokens"], 20);
+ assert_eq!(output["usage"]["cache_read_input_tokens"], 70);
+ assert_eq!(output["usage"]["cache_creation_input_tokens"], 10);
+ assert_eq!(output["usage"]["output_tokens"], 5);
+ }
Ok(())
}
diff --git a/crates/switchyard-translation/tests/stream_translation.rs b/crates/switchyard-translation/tests/stream_translation.rs
index b0cbbc538..4cf05b69f 100644
--- a/crates/switchyard-translation/tests/stream_translation.rs
+++ b/crates/switchyard-translation/tests/stream_translation.rs
@@ -310,6 +310,41 @@ fn openai_chat_stream_cache_usage_translates_to_responses_usage_details() -> Tes
Ok(())
}
+// Verifies OpenRouter's cache-write field and the legacy alias normalize identically.
+#[test]
+fn openai_chat_stream_cache_write_usage_is_normalized() -> TestResult {
+ for cache_write_field in ["cache_write_tokens", "cache_creation_tokens"] {
+ let mut state = StreamTranslationState::new(WireFormat::OpenAiChat, WireFormat::OpenAiChat);
+ let mut event = json!({
+ "id": "chatcmpl-test",
+ "object": "chat.completion.chunk",
+ "model": "gpt-cached",
+ "choices": [],
+ "usage": {
+ "prompt_tokens": 100,
+ "completion_tokens": 5,
+ "total_tokens": 105,
+ "prompt_tokens_details": {"cached_tokens": 70}
+ }
+ });
+ event["usage"]["prompt_tokens_details"][cache_write_field] = json!(10);
+
+ let chunks = decode_stream_event(&mut state, WireFormat::OpenAiChat, &event);
+ let Some(usage) = chunks.iter().find_map(|chunk| match chunk {
+ LlmResponseChunk::Usage(usage) => Some(usage),
+ _ => None,
+ }) else {
+ return Err(format!("expected usage chunk for {cache_write_field}").into());
+ };
+
+ assert_eq!(usage.input_tokens, Some(20));
+ assert_eq!(usage.cached_input_tokens(), Some(70));
+ assert_eq!(usage.cache_creation_input_tokens(), Some(10));
+ assert_eq!(usage.output_tokens, Some(5));
+ }
+ Ok(())
+}
+
// Verifies the streamed total-token fallback keeps cached tokens when upstream omits total_tokens.
#[test]
fn responses_stream_usage_without_total_keeps_cached_tokens_in_total() -> TestResult {