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 OpenAI target prompt requests (#733)

nachiketb-nvidiaeb50a5a363a6
diff --git a/crates/switchyard-translation/src/util.rs b/crates/switchyard-translation/src/util.rs
index 2469157c6..578bba103 100644
--- a/crates/switchyard-translation/src/util.rs
+++ b/crates/switchyard-translation/src/util.rs
@@ -317,6 +317,12 @@ fn prepare_preserved_requests(
body.insert("model".to_string(), Value::String(target.to_string()));
match prompt {
None => true,
+ Some(prompt) if format.as_str() == WireFormat::OpenAiChat.as_str() => {
+ prepend_openai_chat_system(body, prompt)
+ }
+ Some(prompt) if format.as_str() == WireFormat::OpenAiResponses.as_str() => {
+ prepend_openai_responses_instructions(body, prompt)
+ }
Some(prompt) if format.as_str() == WireFormat::AnthropicMessages.as_str() => {
body.remove(crate::codecs::common::ANTHROPIC_REQUEST_KEY);
prepend_anthropic_system(body, prompt)
@@ -326,6 +332,39 @@ fn prepare_preserved_requests(
});
}
+// Prepends a target prompt without rebuilding or otherwise changing a Chat request.
+fn prepend_openai_chat_system(body: &mut Map<String, Value>, prompt: &str) -> bool {
+ let message = json!({"role": "system", "content": prompt});
+ match body.get_mut("messages") {
+ None | Some(Value::Null) => {
+ body.insert("messages".to_string(), Value::Array(vec![message]));
+ }
+ Some(Value::Array(messages)) => messages.insert(0, message),
+ Some(_) => return false,
+ }
+ true
+}
+
+// Prepends a target prompt without rebuilding or otherwise changing a Responses request.
+fn prepend_openai_responses_instructions(body: &mut Map<String, Value>, prompt: &str) -> bool {
+ match body.get_mut("instructions") {
+ None | Some(Value::Null) => {
+ body.insert(
+ "instructions".to_string(),
+ Value::String(prompt.to_string()),
+ );
+ }
+ Some(Value::String(instructions)) if instructions.is_empty() => {
+ *instructions = prompt.to_string();
+ }
+ Some(Value::String(instructions)) => {
+ instructions.insert_str(0, &format!("{prompt}\n\n"));
+ }
+ Some(_) => return false,
+ }
+ true
+}
+
// Prepends a target prompt without rebuilding or otherwise changing an Anthropic request.
fn prepend_anthropic_system(body: &mut Map<String, Value>, prompt: &str) -> bool {
match body.get_mut("system") {
diff --git a/crates/switchyard-translation/tests/request_translation.rs b/crates/switchyard-translation/tests/request_translation.rs
index a6bcccf1b..7e70f82a0 100644
--- a/crates/switchyard-translation/tests/request_translation.rs
+++ b/crates/switchyard-translation/tests/request_translation.rs
@@ -16,36 +16,88 @@ use common::{REASONING_MODEL, normalized_policy, shell_tool_call};
type TestResult<T = ()> = std::result::Result<T, Box<dyn std::error::Error + Send + Sync>>;
-// A target prompt makes every preserved provider body stale.
+// Verifies target preparation restamps the model, prepends native instructions, and preserves all
+// other provider fields.
#[test]
-fn preparing_a_target_prompt_invalidates_exact_replay() -> TestResult {
+fn openai_target_prompt_preserves_native_request_fields() -> TestResult {
let engine = TranslationEngine::default();
let policy = TranslationPolicy::default();
- let body = json!({
- "model": "route",
- "messages": [
- {"role": "system", "name": "caller", "content": "client prompt"},
- {"role": "user", "content": "hi"}
- ]
- });
- let mut request = engine
- .decode_request(WireFormat::OpenAiChat, &body, &policy)?
- .request;
+ for (format, body) in [
+ (
+ WireFormat::OpenAiChat,
+ json!({
+ "model": "route",
+ "messages": [
+ {"role": "system", "name": "caller", "content": "client prompt"},
+ {"role": "user", "content": "hi"}
+ ],
+ "moderation": {"policy": "strict"},
+ "stop": ["<END>"],
+ "prompt_cache_options": {"retention": "24h"},
+ "logprobs": true,
+ "top_logprobs": 2,
+ "modalities": ["text", "audio"],
+ "audio": {"voice": "alloy", "format": "wav"},
+ "web_search_options": {"search_context_size": "high"},
+ "n": 2,
+ "logit_bias": {"42": -1},
+ "frequency_penalty": 0.2,
+ "presence_penalty": 0.3,
+ "seed": 7,
+ "prediction": {"type": "content", "content": "prefix"},
+ "verbosity": "low",
+ "functions": [{"name": "legacy", "parameters": {"type": "object"}}],
+ "function_call": {"name": "legacy"}
+ }),
+ ),
+ (
+ WireFormat::OpenAiResponses,
+ json!({
+ "model": "route",
+ "instructions": "caller prompt",
+ "input": "hi",
+ "moderation": {"policy": "strict"},
+ "prompt_cache_options": {"retention": "24h"},
+ "reasoning": {
+ "effort": "high",
+ "context": "large",
+ "mode": "detailed",
+ "summary": "auto"
+ },
+ "include": ["reasoning.encrypted_content"],
+ "previous_response_id": "resp_previous",
+ "context_management": [{"type": "compaction", "compact_threshold": 1000}],
+ "tools": [
+ {"type": "web_search_preview"},
+ {"type": "file_search", "vector_store_ids": ["vs_1"]},
+ {"type": "mcp", "server_label": "inventory", "server_url": "https://fixture.invalid/mcp"}
+ ]
+ }),
+ ),
+ ] {
+ let mut expected = body.clone();
+ expected["model"] = json!("selected/model");
+ match format {
+ WireFormat::OpenAiChat => expected["messages"]
+ .as_array_mut()
+ .ok_or("expected messages")?
+ .insert(0, json!({"role": "system", "content": "target prompt"})),
+ WireFormat::OpenAiResponses => {
+ expected["instructions"] = json!("target prompt\n\ncaller prompt");
+ }
+ WireFormat::AnthropicMessages => unreachable!(),
+ }
- prepare_request_for_target(
- &mut request,
- &"selected/model".into(),
- Some("target prompt"),
- );
+ let mut request = engine.decode_request(format, &body, &policy)?.request;
+ prepare_request_for_target(
+ &mut request,
+ &"selected/model".into(),
+ Some("target prompt"),
+ );
+ let encoded = engine.encode_request(format, &request, &policy)?.body;
- assert!(request.preservation.requests.is_empty());
- let encoded = engine
- .encode_request(WireFormat::OpenAiChat, &request, &policy)?
- .body;
- assert_eq!(encoded["model"], "selected/model");
- assert_eq!(encoded["messages"][0]["content"], "target prompt");
- assert_eq!(encoded["messages"][1]["content"], "client prompt");
- assert!(encoded["messages"][1].get("name").is_none());
+ assert_eq!(encoded, expected);
+ }
Ok(())
}