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
// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
//! Router guidance uses the shared request edits so preserved provider content stays intact.
use switchyard_protocol::Request;
/// This helper appends temporary guidance and updates preserved provider bodies.
pub fn append_note(request: &mut Request, note: &str) {
switchyard_translation::append_note(&mut request.llm_request, note);
}
/// This helper clears preserved bodies and the note count after other conversation edits.
/// Subsequent encoding uses the edited request instead of replaying stale provider bodies.
pub(crate) fn drop_exact_replay(request: &mut Request) {
request.llm_request.preservation.requests.clear();
request.llm_request.temporary_note_count = 0;
}
#[cfg(test)]
mod tests {
use super::*;
use serde_json::json;
use switchyard_protocol::WireFormat;
use switchyard_translation::{TranslationEngine, TranslationPolicy};
// Native bodies retain provider fields and tool argument text that decoding may not preserve.
#[test]
fn edits_preserve_native_prefixes_and_provider_fields() -> Result<(), Box<dyn std::error::Error>>
{
let engine = TranslationEngine::default();
let policy = TranslationPolicy::default();
let fixtures = [
(
WireFormat::OpenAiChat,
json!({
"model":"auto", "messages":[
{"role":"developer","content":[{"type":"text","text":"Keep this instruction.","vendor":"keep"}]},
{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"run","arguments":"{ \"cmd\": \"test\" }"}}]},
{"role":"tool","tool_call_id":"call_1","content":"exit 1","vendor":"keep"}
], "metadata":{"caller":"keep"}, "temperature":0.2
}),
),
(
WireFormat::OpenAiResponses,
json!({
"model":"auto", "instructions":"Keep this instruction.", "input":[
{"type":"message","role":"developer","content":[{"type":"input_text","text":"Keep this too."}]},
{"type":"function_call","call_id":"call_1","name":"run","arguments":"{ \"cmd\": \"test\" }"},
{"type":"function_call_output","call_id":"call_1","output":"exit 1"}
], "metadata":{"caller":"keep"}, "store":false
}),
),
(
WireFormat::AnthropicMessages,
json!({
"model":"auto", "max_tokens":32,
"system":[{"type":"text","text":"Keep this instruction.","cache_control":{"type":"ephemeral","ttl":"1h"}}],
"messages":[
{"role":"assistant","content":[{"type":"tool_use","id":"call_1","name":"run","input":{"cmd":"test"}}]},
{"role":"user","content":[{"type":"tool_result","tool_use_id":"call_1","content":"exit 1","is_error":true,"cache_control":{"type":"ephemeral"}}]}
], "metadata":{"user_id":"keep"}
}),
),
];
for (format, original) in fixtures {
let mut request = Request {
llm_request: engine.decode_request(format, &original, &policy)?.request,
..Default::default()
};
append_note(&mut request, "Reassess the failure.");
let encoded = engine
.encode_request(format, &request.llm_request, &policy)?
.body;
let mut expected = original.clone();
match format {
WireFormat::OpenAiChat => expected["messages"].as_array_mut().ok_or("messages")?.push(json!({"role":"user","content":[{"type":"text","text":"Reassess the failure."}]})),
WireFormat::OpenAiResponses => expected["input"].as_array_mut().ok_or("input")?.push(json!({"role":"user","content":[{"type":"input_text","text":"Reassess the failure."}]})),
WireFormat::AnthropicMessages => expected["messages"][1]["content"].as_array_mut().ok_or("content")?.push(json!({"type":"text","text":"Reassess the failure."})),
}
assert_eq!(encoded, expected, "{format}: only the note may change");
switchyard_translation::prepend_system_prompt(
&mut request.llm_request,
"Use the capable tier.",
);
let edited = engine
.encode_request(format, &request.llm_request, &policy)?
.body;
// The system prompt edit must retain the earlier note and all conversation and tool items.
let key = if format == WireFormat::OpenAiResponses {
"input"
} else {
"messages"
};
if format == WireFormat::OpenAiChat {
assert_eq!(
edited[key].as_array().ok_or("messages")?[1..],
encoded[key].as_array().ok_or("messages")?[..]
);
} else {
assert_eq!(edited[key], encoded[key]);
}
assert_eq!(edited["metadata"], original["metadata"]);
}
Ok(())
}
#[test]
fn string_inputs_and_assistant_endings_keep_the_original_message()
-> Result<(), Box<dyn std::error::Error>> {
let engine = TranslationEngine::default();
let policy = TranslationPolicy::default();
for (format, body, key) in [
(
WireFormat::OpenAiChat,
json!({"model":"m","messages":[{"role":"assistant","content":"done"}]}),
"messages",
),
(
WireFormat::AnthropicMessages,
json!({"model":"m","max_tokens":8,"messages":[{"role":"user","content":"question"},{"role":"assistant","content":"done"}]}),
"messages",
),
(
WireFormat::OpenAiResponses,
json!({"model":"m","input":"question"}),
"input",
),
] {
let mut request = Request {
llm_request: engine.decode_request(format, &body, &policy)?.request,
..Default::default()
};
append_note(&mut request, "note one");
append_note(&mut request, "note two");
let encoded = engine
.encode_request(format, &request.llm_request, &policy)?
.body;
let items = encoded[key].as_array().ok_or("items")?;
if body[key].is_array() {
let original = body[key].as_array().ok_or("original")?;
assert_eq!(&items[..original.len()], original);
} else {
assert_eq!(items[0], json!({"role":"user","content":"question"}));
}
}
Ok(())
}
// Unsupported preserved bodies must not cause encoding to omit the note.
#[test]
fn invalid_preservation_falls_back_without_losing_the_note()
-> Result<(), Box<dyn std::error::Error>> {
let engine = TranslationEngine::default();
let policy = TranslationPolicy::default();
let source = json!({"model":"m","messages":[{"role":"user","content":"original task"}]});
let mut request = Request {
llm_request: engine
.decode_request(WireFormat::OpenAiChat, &source, &policy)?
.request,
..Default::default()
};
request
.llm_request
.preservation
.requests
.insert("custom".into(), json!({"messages":"unsupported"}));
request.llm_request.preservation.requests.insert(
WireFormat::OpenAiChat.into(),
json!({"model":"m","messages":"unsupported"}),
);
append_note(&mut request, "reassess the failure");
assert!(request.llm_request.preservation.requests.is_empty());
for target in [
WireFormat::OpenAiChat,
WireFormat::OpenAiResponses,
WireFormat::AnthropicMessages,
] {
let body = engine
.encode_request(target, &request.llm_request, &policy)?
.body;
let decoded = engine.decode_request(target, &body, &policy)?.request;
let text: Vec<_> = decoded
.messages
.iter()
.filter_map(|message| message.text_content(""))
.collect();
assert_eq!(
text,
vec!["original task", "reassess the failure"],
"{target}"
);
}
Ok(())
}
}