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): always emit Responses usage detail objects (#144)

Lin Jia0a03235ba9ca
diff --git a/crates/switchyard-translation/src/codecs/responses/buffered.rs b/crates/switchyard-translation/src/codecs/responses/buffered.rs
index 0c57cd837..cf7342da0 100644
--- a/crates/switchyard-translation/src/codecs/responses/buffered.rs
+++ b/crates/switchyard-translation/src/codecs/responses/buffered.rs
@@ -1210,21 +1210,18 @@ fn encode_responses_usage(usage: &Usage) -> Value {
let input_tokens = usage.input_tokens.unwrap_or(0)
+ usage.cached_input_tokens().unwrap_or(0)
+ usage.cache_creation_input_tokens().unwrap_or(0);
- let mut value = json!({
+ // The detail objects are required by the Responses schema, so both are always present. An
+ // upstream that reports no cache or reasoning breakdown means zero, not "unknown"; omitting
+ // them instead yields a payload that fails to deserialize into the OpenAI SDK's
+ // ResponseUsage, which types both as non-optional.
+ json!({
"input_tokens": input_tokens,
"output_tokens": usage.output_tokens.unwrap_or(0),
"total_tokens": usage
.total_tokens
.or_else(|| Some(input_tokens + usage.output_tokens.unwrap_or(0)))
.unwrap_or(0),
- });
- if let Some(cached_tokens) = usage.cached_input_tokens() {
- value["input_tokens_details"] = json!({"cached_tokens": cached_tokens});
- }
- if let Some(reasoning_tokens) = usage.reasoning_tokens {
- value["output_tokens_details"] = json!({
- "reasoning_tokens": reasoning_tokens,
- });
- }
- value
+ "input_tokens_details": {"cached_tokens": usage.cached_input_tokens().unwrap_or(0)},
+ "output_tokens_details": {"reasoning_tokens": usage.reasoning_tokens.unwrap_or(0)},
+ })
}
diff --git a/crates/switchyard-translation/src/codecs/responses/stream.rs b/crates/switchyard-translation/src/codecs/responses/stream.rs
index bf4e884bb..60610e0db 100644
--- a/crates/switchyard-translation/src/codecs/responses/stream.rs
+++ b/crates/switchyard-translation/src/codecs/responses/stream.rs
@@ -555,22 +555,17 @@ fn responses_usage_value(usage: &Usage) -> Value {
let input_tokens = usage.input_tokens.unwrap_or(0)
+ usage.cached_input_tokens().unwrap_or(0)
+ usage.cache_creation_input_tokens().unwrap_or(0);
- let mut value = json!({
+ // Both detail objects are always present, for the same reason as the buffered encoder: the
+ // Responses schema types them as required, so a missing breakdown serializes as zero.
+ json!({
"input_tokens": input_tokens,
"output_tokens": usage.output_tokens.unwrap_or(0),
"total_tokens": usage.total_tokens.unwrap_or_else(|| {
input_tokens + usage.output_tokens.unwrap_or(0)
}),
- });
- if let Some(cached_tokens) = usage.cached_input_tokens() {
- value["input_tokens_details"] = json!({"cached_tokens": cached_tokens});
- }
- if let Some(reasoning_tokens) = usage.reasoning_tokens {
- value["output_tokens_details"] = json!({
- "reasoning_tokens": reasoning_tokens,
- });
- }
- value
+ "input_tokens_details": {"cached_tokens": usage.cached_input_tokens().unwrap_or(0)},
+ "output_tokens_details": {"reasoning_tokens": usage.reasoning_tokens.unwrap_or(0)},
+ })
}
// Converts any upstream message ID into a Responses-looking response ID.
diff --git a/crates/switchyard-translation/tests/response_translation.rs b/crates/switchyard-translation/tests/response_translation.rs
index 3e61f0151..e963af555 100644
--- a/crates/switchyard-translation/tests/response_translation.rs
+++ b/crates/switchyard-translation/tests/response_translation.rs
@@ -353,7 +353,13 @@ fn openai_chat_response_with_tool_call_translates_to_responses_output_item() ->
assert_eq!(output["output"][0]["arguments"], "{\"q\": \"rust\"}");
assert_eq!(
output["usage"],
- json!({"input_tokens": 4, "output_tokens": 3, "total_tokens": 7})
+ json!({
+ "input_tokens": 4,
+ "output_tokens": 3,
+ "total_tokens": 7,
+ "input_tokens_details": {"cached_tokens": 0},
+ "output_tokens_details": {"reasoning_tokens": 0}
+ })
);
Ok(())
}
@@ -396,3 +402,83 @@ fn openai_chat_response_with_text_and_tool_call_translates_both_to_responses() -
assert_eq!(output["output"][1]["call_id"], "call_1");
Ok(())
}
+
+// Verifies both Responses usage detail objects are emitted even when the upstream reports no
+// cache or reasoning breakdown. The Responses schema types them as required, so omitting them
+// makes the payload unparseable by OpenAI-SDK clients.
+#[test]
+fn openai_chat_usage_without_breakdowns_still_emits_responses_usage_details() -> TestResult {
+ let engine = TranslationEngine::default();
+ let body = json!({
+ "id": "chatcmpl-test",
+ "model": "plain-model",
+ "choices": [{
+ "index": 0,
+ "message": {"role": "assistant", "content": "Hi"},
+ "finish_reason": "stop"
+ }],
+ "usage": {"prompt_tokens": 41, "completion_tokens": 3, "total_tokens": 44}
+ });
+
+ let output = engine
+ .translate_response(
+ WireFormat::OpenAiChat,
+ WireFormat::OpenAiResponses,
+ &body,
+ &TranslationPolicy::default(),
+ )?
+ .body;
+
+ assert_eq!(
+ output["usage"],
+ json!({
+ "input_tokens": 41,
+ "output_tokens": 3,
+ "total_tokens": 44,
+ "input_tokens_details": {"cached_tokens": 0},
+ "output_tokens_details": {"reasoning_tokens": 0}
+ })
+ );
+ Ok(())
+}
+
+// Verifies a partial breakdown does not suppress the other detail object: an upstream that
+// reports cached tokens but no reasoning tokens must still carry both.
+#[test]
+fn openai_chat_cache_only_usage_still_emits_reasoning_details() -> TestResult {
+ let engine = TranslationEngine::default();
+ let body = json!({
+ "id": "chatcmpl-test",
+ "model": "cache-only-model",
+ "choices": [{
+ "index": 0,
+ "message": {"role": "assistant", "content": "Hi"},
+ "finish_reason": "stop"
+ }],
+ "usage": {
+ "prompt_tokens": 41,
+ "completion_tokens": 3,
+ "total_tokens": 44,
+ "prompt_tokens_details": {"cached_tokens": 32}
+ }
+ });
+
+ let output = engine
+ .translate_response(
+ WireFormat::OpenAiChat,
+ WireFormat::OpenAiResponses,
+ &body,
+ &TranslationPolicy::default(),
+ )?
+ .body;
+
+ assert_eq!(
+ output["usage"]["input_tokens_details"],
+ json!({"cached_tokens": 32})
+ );
+ assert_eq!(
+ output["usage"]["output_tokens_details"],
+ json!({"reasoning_tokens": 0})
+ );
+ Ok(())
+}
diff --git a/crates/switchyard-translation/tests/stream_translation.rs b/crates/switchyard-translation/tests/stream_translation.rs
index 4cf05b69f..27bd9c411 100644
--- a/crates/switchyard-translation/tests/stream_translation.rs
+++ b/crates/switchyard-translation/tests/stream_translation.rs
@@ -597,3 +597,47 @@ fn openai_chat_error_frame_decodes_to_stream_error() -> TestResult {
}
Ok(())
}
+
+// Verifies the streaming encoder matches the buffered one: both Responses usage detail objects
+// are present even when the upstream reports no cache or reasoning breakdown.
+#[test]
+fn openai_chat_stream_usage_without_breakdowns_still_emits_responses_usage_details() -> TestResult {
+ let engine = TranslationEngine::default();
+ let mut state =
+ StreamTranslationState::new(WireFormat::OpenAiChat, WireFormat::OpenAiResponses);
+ let usage = json!({
+ "id": "chatcmpl-test",
+ "object": "chat.completion.chunk",
+ "model": "plain-model",
+ "choices": [{
+ "index": 0,
+ "delta": {},
+ "finish_reason": "stop"
+ }],
+ "usage": {"prompt_tokens": 41, "completion_tokens": 3, "total_tokens": 44}
+ });
+
+ let mut events = engine.translate_event(
+ &mut state,
+ WireFormat::OpenAiChat,
+ WireFormat::OpenAiResponses,
+ &usage,
+ )?;
+ events.extend(engine.finish_stream(&mut state, WireFormat::OpenAiResponses)?);
+
+ let Some(completed) = events
+ .iter()
+ .find(|event| event["type"] == "response.completed")
+ else {
+ return Err("expected final Responses completion event".into());
+ };
+ assert_eq!(
+ completed["response"]["usage"]["input_tokens_details"],
+ json!({"cached_tokens": 0})
+ );
+ assert_eq!(
+ completed["response"]["usage"]["output_tokens_details"],
+ json!({"reasoning_tokens": 0})
+ );
+ Ok(())
+}