feat(rust): add the openai_like chat config foundation (#43379)

* feat(rust): add the openai_like chat config foundation

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(rust): let max_completion_tokens outrank max_tokens and decline refusal responses

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: Yujong Lee <yujong@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm-rust/crates/core/src/chat_completions/common_utils.rs

@@ -3,6 +3,7 @@ use litellm_llms::{
     anthropic::chat::transformation::ANTHROPIC_CHAT_COMPLETIONS_CONFIG,
     base_llm::chat::transformation::BaseConfig,
     bedrock::chat::converse_transformation::BEDROCK_CHAT_COMPLETIONS_CONFIG,
+    openai_like::chat::transformation::OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG,
 };
 use serde_json::{Map, Value};
 
@@ -14,6 +15,7 @@ pub(super) fn chat_completions_provider_config(provider: &str) -> Option<&'stati
     match provider {
         "anthropic" => Some(&ANTHROPIC_CHAT_COMPLETIONS_CONFIG),
         "bedrock" => Some(&BEDROCK_CHAT_COMPLETIONS_CONFIG),
+        "openai_like" => Some(&OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG),
         _ => None,
     }
 }

litellm-rust/crates/llms/src/lib.rs

@@ -7,6 +7,7 @@ pub mod cohere;
 mod error;
 pub mod mistral;
 pub mod openai;
+pub mod openai_like;
 pub mod reducto;
 pub mod vertex_ai;
 

litellm-rust/crates/llms/src/openai_like/chat/mod.rs

@@ -0,0 +1 @@
+pub mod transformation;

litellm-rust/crates/llms/src/openai_like/chat/transformation.rs

@@ -0,0 +1,270 @@
+//! `litellm/llms/openai_like/chat/transformation.py`: the chat config every
+//! OpenAI-compatible endpoint shares. The body is already OpenAI-shaped, so
+//! parameters pass through verbatim; the port keeps Python's two deviations,
+//! the `max_completion_tokens` -> `max_tokens` rename and the usage
+//! `*_tokens` null-to-zero sanitize.
+
+use litellm_auth::{CredentialPlacement, SecretValue};
+use litellm_core_utils::core_helpers::unix_now;
+use litellm_types::{
+    llms::openai::ChatMessage,
+    utils::{ChatCompletionsChoice, ChatCompletionsChoiceMessage, ChatCompletionsResponse},
+};
+use serde_json::{Map, Value, json};
+
+use crate::{
+    Error,
+    base_llm::{
+        auth::AuthScheme,
+        chat::transformation::{
+            BaseConfig, Headers, ProviderChatRequestData, ProviderChatResponseData,
+            ValidatedEnvironment,
+        },
+    },
+    openai_like::common_utils::{complete_openai_like_url, openai_compatible_provider_info},
+};
+
+/// OpenAI parameter names the Rust path can place verbatim in the request body.
+/// Tool parameters are absent on purpose: the message gate already declines
+/// tool-call content, and a `tools` request that did get through would produce
+/// a tool-call response this port cannot normalize yet, so it declines before
+/// the call instead of after it.
+const SUPPORTED_PARAMS: &[(&str, &str)] = &[
+    ("frequency_penalty", "frequency_penalty"),
+    ("logit_bias", "logit_bias"),
+    ("logprobs", "logprobs"),
+    ("top_logprobs", "top_logprobs"),
+    ("max_tokens", "max_tokens"),
+    ("max_completion_tokens", "max_completion_tokens"),
+    ("modalities", "modalities"),
+    ("prediction", "prediction"),
+    ("n", "n"),
+    ("presence_penalty", "presence_penalty"),
+    ("seed", "seed"),
+    ("stop", "stop"),
+    ("stream_options", "stream_options"),
+    ("temperature", "temperature"),
+    ("top_p", "top_p"),
+    ("audio", "audio"),
+    ("web_search_options", "web_search_options"),
+    ("service_tier", "service_tier"),
+    ("safety_identifier", "safety_identifier"),
+    ("prompt_cache_key", "prompt_cache_key"),
+    ("prompt_cache_retention", "prompt_cache_retention"),
+    ("store", "store"),
+    ("response_format", "response_format"),
+];
+
+/// Call configuration the caller may pass that never enters the request body.
+const CONFIG_PARAMS: &[&str] = &["custom_endpoint", "extra_headers", "max_retries"];
+
+pub struct OpenAILikeChatConfig;
+
+pub const OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG: OpenAILikeChatConfig = OpenAILikeChatConfig;
+
+impl BaseConfig for OpenAILikeChatConfig {
+    fn supported_openai_param_mappings(&self) -> &'static [(&'static str, &'static str)] {
+        SUPPORTED_PARAMS
+    }
+
+    fn get_complete_url(
+        &self,
+        api_base: Option<&str>,
+        _model: &str,
+        optional_params: &Map<String, Value>,
+        env_lookup: &dyn Fn(&str) -> Option<String>,
+    ) -> Result<String, Error> {
+        let custom_endpoint = optional_params
+            .get("custom_endpoint")
+            .and_then(Value::as_bool)
+            .unwrap_or(false);
+        complete_openai_like_url(api_base, custom_endpoint, env_lookup)
+    }
+
+    fn transform_request(
+        &self,
+        model: &str,
+        messages: Vec<ChatMessage>,
+        optional_params: Map<String, Value>,
+    ) -> Result<ProviderChatRequestData, Error> {
+        let mut params = Map::from_iter(
+            optional_params
+                .into_iter()
+                .filter(|(key, _)| !CONFIG_PARAMS.contains(&key.as_str())),
+        );
+        // Most OpenAI-compatible endpoints take `max_tokens`, not
+        // `max_completion_tokens`, so Python's `map_openai_params` renames it
+        // and lets it overwrite a `max_tokens` the caller also sent.
+        if let Some(limit) = params.remove("max_completion_tokens") {
+            params.insert("max_tokens".to_string(), limit);
+        }
+        let body = Map::from_iter(
+            [
+                ("model".to_string(), json!(model)),
+                ("messages".to_string(), json!(messages)),
+            ]
+            .into_iter()
+            .chain(params),
+        );
+        Ok(ProviderChatRequestData {
+            body: Value::Object(body),
+            stream_shape: Default::default(),
+        })
+    }
+
+    fn transform_response(
+        &self,
+        model: &str,
+        response: ProviderChatResponseData,
+    ) -> Result<ChatCompletionsResponse, Error> {
+        let mut body = response.body;
+        sanitize_usage(&mut body);
+        let body = body
+            .as_object()
+            .ok_or_else(|| Error::InvalidResponse("chat response is not an object".into()))?;
+
+        let choices = body
+            .get("choices")
+            .and_then(Value::as_array)
+            .ok_or(Error::MissingField("choices"))?
+            .iter()
+            .enumerate()
+            .map(|(position, choice)| normalize_choice(position, choice))
+            .collect::<Result<Vec<_>, _>>()?;
+
+        let usage = body.get("usage").and_then(Value::as_object);
+        let field = |name: &str| {
+            usage
+                .and_then(|usage| usage.get(name))
+                .and_then(Value::as_u64)
+                .unwrap_or(0)
+        };
+        let details = usage.and_then(|usage| usage.get("prompt_tokens_details"));
+
+        Ok(ChatCompletionsResponse {
+            created: body
+                .get("created")
+                .and_then(Value::as_u64)
+                .unwrap_or_else(unix_now),
+            model: body
+                .get("model")
+                .and_then(Value::as_str)
+                .unwrap_or(model)
+                .to_string(),
+            choices,
+            usage: litellm_types::utils::ChatCompletionsUsage {
+                prompt_tokens: field("prompt_tokens"),
+                completion_tokens: field("completion_tokens"),
+                total_tokens: field("total_tokens"),
+                prompt_tokens_details: litellm_types::utils::PromptTokensDetails {
+                    cached_tokens: details
+                        .and_then(|d| d.get("cached_tokens"))
+                        .and_then(Value::as_u64)
+                        .unwrap_or(0),
+                    cache_creation_tokens: details
+                        .and_then(|d| d.get("cache_creation_tokens"))
+                        .and_then(Value::as_u64)
+                        .unwrap_or(0),
+                    text_tokens: details
+                        .and_then(|d| d.get("text_tokens"))
+                        .and_then(Value::as_u64)
+                        .unwrap_or(0),
+                },
+            },
+        })
+    }
+
+    /// `OpenAILikeBase._validate_environment`: a forwarded `authorization` is
+    /// the whole credential, and any other call authenticates with the
+    /// resolved key as a bearer. The key resolves to `""` when neither the
+    /// deployment nor `OPENAI_LIKE_API_KEY` sets one, because vllm-compatible
+    /// endpoints take no key; Python still sends `Bearer ` in that case.
+    fn validate_environment(
+        &self,
+        headers: Headers,
+        api_key: Option<&str>,
+        _model: &str,
+        _optional_params: &Map<String, Value>,
+        env_lookup: &dyn Fn(&str) -> Option<String>,
+    ) -> Result<ValidatedEnvironment, Error> {
+        if headers
+            .iter()
+            .any(|(name, _)| name.eq_ignore_ascii_case("authorization"))
+        {
+            return Ok(ValidatedEnvironment {
+                headers,
+                auth: AuthScheme::Forwarded,
+            });
+        }
+        let (_, key) = openai_compatible_provider_info(None, api_key, env_lookup);
+        Ok(ValidatedEnvironment {
+            headers,
+            auth: AuthScheme::Credential {
+                placement: CredentialPlacement::Bearer,
+                secret: SecretValue::new(key.unwrap_or_default()),
+            },
+        })
+    }
+
+    fn config_params(&self) -> &'static [&'static str] {
+        CONFIG_PARAMS
+    }
+}
+
+/// `OpenAILikeChatConfig._sanitize_usage_obj`: a provider that reports a null
+/// `*_tokens` entry breaks OpenAI clients, so nulls become 0. Python scrubs
+/// every top-level usage key ending in `_tokens`.
+fn sanitize_usage(body: &mut Value) {
+    if let Some(usage) = body.get_mut("usage").and_then(Value::as_object_mut) {
+        for (key, value) in usage.iter_mut() {
+            if key.ends_with("_tokens") && value.is_null() {
+                *value = json!(0);
+            }
+        }
+    }
+}
+
+fn normalize_choice(position: usize, choice: &Value) -> Result<ChatCompletionsChoice, Error> {
+    let message = choice
+        .get("message")
+        .and_then(Value::as_object)
+        .ok_or(Error::MissingField("message"))?;
+    if message
+        .get("tool_calls")
+        .and_then(Value::as_array)
+        .is_some_and(|calls| !calls.is_empty())
+    {
+        // Python rewrites the lone tool call into content only under
+        // `json_mode`, a request flag `transform_response` cannot see, and the
+        // normalized type cannot carry tool calls at all. Declining is
+        // terminal at this point, but passing back an empty assistant turn
+        // would fabricate the reply.
+        return Err(Error::Unsupported("tool call response"));
+    }
+    if message.get("refusal").is_some_and(|value| !value.is_null()) {
+        return Err(Error::Unsupported("refusal response"));
+    }
+    let content = message.get("content");
+    if content.is_some_and(|value| !value.is_null() && !value.is_string()) {
+        return Err(Error::Unsupported("non-text response content"));
+    }
+    Ok(ChatCompletionsChoice {
+        index: choice
+            .get("index")
+            .and_then(Value::as_u64)
+            .unwrap_or(position as u64),
+        message: ChatCompletionsChoiceMessage {
+            role: message
+                .get("role")
+                .and_then(Value::as_str)
+                .unwrap_or("assistant")
+                .to_string(),
+            content: content.and_then(Value::as_str).map(str::to_string),
+        },
+        finish_reason: choice
+            .get("finish_reason")
+            .and_then(Value::as_str)
+            .unwrap_or("")
+            .to_string(),
+    })
+}

litellm-rust/crates/llms/src/openai_like/common_utils.rs

@@ -0,0 +1,58 @@
+//! Shared OpenAI-like credential and endpoint resolution, mirroring
+//! `litellm/llms/openai_like/common_utils.py`.
+
+use crate::Error;
+
+/// `OpenAILikeChatConfig._get_openai_compatible_provider_info`: the deployment's
+/// `api_base` wins over `OPENAI_LIKE_API_BASE`, and the deployment key over
+/// `OPENAI_LIKE_API_KEY`, with an empty key allowed because vllm-compatible
+/// endpoints do not require one.
+pub fn openai_compatible_provider_info(
+    api_base: Option<&str>,
+    api_key: Option<&str>,
+    env_lookup: &dyn Fn(&str) -> Option<String>,
+) -> (Option<String>, Option<String>) {
+    let api_base = api_base
+        .map(str::to_string)
+        .or_else(|| env_lookup("OPENAI_LIKE_API_BASE"));
+    let api_key = api_key
+        .map(str::to_string)
+        .or_else(|| env_lookup("OPENAI_LIKE_API_KEY"))
+        .or(Some(String::new()));
+    (api_base, api_key)
+}
+
+/// `OpenAILikeBase._validate_environment` requires an api base and, when the
+/// caller gave no `custom_endpoint`, appends the route suffix. A caller-supplied
+/// `custom_endpoint` base is used as is.
+pub fn complete_openai_like_url(
+    api_base: Option<&str>,
+    custom_endpoint: bool,
+    env_lookup: &dyn Fn(&str) -> Option<String>,
+) -> Result<String, Error> {
+    let (api_base, _) = openai_compatible_provider_info(api_base, None, env_lookup);
+    let api_base = api_base.ok_or_else(|| {
+        Error::InvalidRequest(
+            "Missing API Base - A call is being made to LLM Provider but no api base is set either in the environment variables ({LLM_PROVIDER}_API_KEY) or via params"
+                .to_string(),
+        )
+    })?;
+    if custom_endpoint {
+        return Ok(api_base);
+    }
+    Ok(format!(
+        "{}/chat/completions",
+        api_base.trim_end_matches('/')
+    ))
+}
+
+/// The api key the call resolves to. `None` means neither the deployment nor the
+/// environment supplied one, which is valid for endpoints that take no key.
+pub fn resolve_openai_like_api_key(
+    api_key: Option<&str>,
+    env_lookup: &dyn Fn(&str) -> Option<String>,
+) -> Option<String> {
+    openai_compatible_provider_info(None, api_key, env_lookup)
+        .1
+        .filter(|key| !key.is_empty())
+}

litellm-rust/crates/llms/src/openai_like/mod.rs

@@ -0,0 +1,2 @@
+pub mod chat;
+pub mod common_utils;

litellm-rust/crates/llms/tests/openai_like_chat_transformation.rs

@@ -0,0 +1,343 @@
+use litellm_llms::{
+    Error,
+    base_llm::{
+        auth::AuthScheme,
+        chat::transformation::{BaseConfig, ProviderChatResponseData, Unsupported},
+    },
+    openai_like::chat::transformation::OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG,
+};
+use litellm_types::{llms::openai::ChatMessage, utils::ChatCompletionsResponse};
+use rstest::rstest;
+use serde_json::{Map, Value, json};
+
+fn messages(value: Value) -> Vec<ChatMessage> {
+    serde_json::from_value(value).expect("valid messages")
+}
+
+fn params(value: Value) -> Map<String, Value> {
+    match value {
+        Value::Object(map) => map,
+        other => panic!("params must be an object, got {other}"),
+    }
+}
+
+fn no_env(_: &str) -> Option<String> {
+    None
+}
+
+fn env_with<'a>(name: &'a str, value: &'a str) -> impl Fn(&str) -> Option<String> + 'a {
+    move |key| (key == name).then(|| value.to_string())
+}
+
+fn transform(model: &str, msgs: Value, opts: Value) -> Value {
+    OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .transform_request(model, messages(msgs), params(opts))
+        .expect("request transforms")
+        .body
+}
+
+fn transform_response(body: Value) -> Result<ChatCompletionsResponse, Error> {
+    OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .transform_response("some-model", ProviderChatResponseData { body })
+}
+
+fn reason(msgs: Value, opts: Value) -> Option<Unsupported> {
+    OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG.unsupported_reason(&messages(msgs), &params(opts))
+}
+
+#[rstest]
+fn builds_the_openai_shaped_body() {
+    let body = transform(
+        "my-model",
+        json!([
+            {"role": "system", "content": "be terse"},
+            {"role": "user", "content": "hi"},
+        ]),
+        json!({"temperature": 0.5, "max_tokens": 8}),
+    );
+    assert_eq!(body["model"], json!("my-model"));
+    assert_eq!(
+        body["messages"],
+        json!([
+            {"role": "system", "content": "be terse"},
+            {"role": "user", "content": "hi"},
+        ])
+    );
+    assert_eq!(body["temperature"], json!(0.5));
+    assert_eq!(body["max_tokens"], json!(8));
+}
+
+#[rstest]
+fn renames_max_completion_tokens_to_max_tokens() {
+    // `OpenAILikeChatConfig.map_openai_params`: most OpenAI-compatible providers
+    // support `max_tokens`, not `max_completion_tokens`.
+    let body = transform(
+        "my-model",
+        json!([{"role": "user", "content": "hi"}]),
+        json!({"max_completion_tokens": 12}),
+    );
+    assert_eq!(body["max_tokens"], json!(12));
+    assert!(body.get("max_completion_tokens").is_none());
+}
+
+#[rstest]
+fn max_completion_tokens_wins_when_both_limits_are_sent() {
+    // Python assigns `max_tokens = max_completion_tokens` after copying the
+    // params, so the renamed value outranks a caller-supplied `max_tokens`.
+    let body = transform(
+        "my-model",
+        json!([{"role": "user", "content": "hi"}]),
+        json!({"max_tokens": 8, "max_completion_tokens": 12}),
+    );
+    assert_eq!(body["max_tokens"], json!(12));
+    assert!(body.get("max_completion_tokens").is_none());
+}
+
+#[rstest]
+fn call_configuration_never_enters_the_body() {
+    let body = transform(
+        "my-model",
+        json!([{"role": "user", "content": "hi"}]),
+        json!({"custom_endpoint": true, "extra_headers": {"x": "y"}, "max_retries": 2}),
+    );
+    assert_eq!(
+        body.as_object().unwrap().keys().collect::<Vec<_>>(),
+        vec!["model", "messages"]
+    );
+}
+
+#[rstest]
+#[case::appends_the_chat_completions_suffix("https://vllm.example.com/v1", json!({}), "https://vllm.example.com/v1/chat/completions")]
+#[case::trims_a_trailing_slash("https://vllm.example.com/v1/", json!({}), "https://vllm.example.com/v1/chat/completions")]
+#[case::a_custom_endpoint_is_used_as_is("https://vllm.example.com/v1/chat/completions", json!({"custom_endpoint": true}), "https://vllm.example.com/v1/chat/completions")]
+fn complete_url(#[case] api_base: &str, #[case] opts: Value, #[case] expected: &str) {
+    assert_eq!(
+        OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+            .get_complete_url(Some(api_base), "my-model", &params(opts), &no_env)
+            .expect("url resolves"),
+        expected
+    );
+}
+
+#[rstest]
+fn api_base_falls_back_to_the_environment() {
+    assert_eq!(
+        OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+            .get_complete_url(
+                None,
+                "my-model",
+                &params(json!({})),
+                &env_with("OPENAI_LIKE_API_BASE", "https://env.example.com/v1"),
+            )
+            .expect("url resolves"),
+        "https://env.example.com/v1/chat/completions"
+    );
+}
+
+#[rstest]
+fn a_missing_api_base_is_an_error() {
+    assert!(matches!(
+        OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+            .get_complete_url(None, "my-model", &params(json!({})), &no_env),
+        Err(Error::InvalidRequest(message)) if message.starts_with("Missing API Base")
+    ));
+}
+
+#[rstest]
+fn the_resolved_key_authenticates_as_a_bearer() {
+    let validated = OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .validate_environment(
+            vec![],
+            Some("sk-test"),
+            "my-model",
+            &params(json!({})),
+            &no_env,
+        )
+        .expect("validates");
+    assert!(matches!(
+        validated.auth,
+        AuthScheme::Credential {
+            placement: litellm_auth::CredentialPlacement::Bearer,
+            ..
+        }
+    ));
+}
+
+#[rstest]
+fn the_key_falls_back_to_the_environment() {
+    let validated = OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .validate_environment(
+            vec![],
+            None,
+            "my-model",
+            &params(json!({})),
+            &env_with("OPENAI_LIKE_API_KEY", "sk-env"),
+        )
+        .expect("validates");
+    let AuthScheme::Credential { secret, .. } = validated.auth else {
+        panic!("expected a bearer credential");
+    };
+    assert_eq!(secret.expose(), "sk-env");
+}
+
+#[rstest]
+fn a_forwarded_authorization_is_the_whole_credential() {
+    // Python adds `Bearer <key>` only when the caller did not already send
+    // `Authorization`, so the forwarded header wins over the deployment key.
+    let headers = vec![("Authorization".to_string(), "Bearer caller".to_string())];
+    let validated = OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .validate_environment(
+            headers,
+            Some("sk-test"),
+            "my-model",
+            &params(json!({})),
+            &no_env,
+        )
+        .expect("validates");
+    assert!(matches!(validated.auth, AuthScheme::Forwarded));
+}
+
+#[rstest]
+fn keyless_calls_still_validate_for_endpoints_that_take_no_key() {
+    // vllm-compatible endpoints require no api key; Python resolves `""` and
+    // sends `Bearer `, so validation must not fail on the missing key.
+    let validated = OPENAI_LIKE_CHAT_COMPLETIONS_CONFIG
+        .validate_environment(vec![], None, "my-model", &params(json!({})), &no_env)
+        .expect("validates");
+    let AuthScheme::Credential { secret, .. } = validated.auth else {
+        panic!("expected a bearer credential");
+    };
+    assert_eq!(secret.expose(), "");
+}
+
+#[rstest]
+fn normalizes_an_openai_response() {
+    let response = transform_response(json!({
+        "created": 1_700_000_000,
+        "model": "served-model-name",
+        "choices": [{
+            "index": 0,
+            "message": {"role": "assistant", "content": "hello"},
+            "finish_reason": "stop",
+        }],
+        "usage": {"prompt_tokens": 3, "completion_tokens": 5, "total_tokens": 8},
+    }))
+    .expect("response normalizes");
+    assert_eq!(response.created, 1_700_000_000);
+    assert_eq!(response.model, "served-model-name");
+    assert_eq!(
+        response.choices[0].message.content.as_deref(),
+        Some("hello")
+    );
+    assert_eq!(response.choices[0].finish_reason, "stop");
+    assert_eq!(response.usage.prompt_tokens, 3);
+    assert_eq!(response.usage.completion_tokens, 5);
+    assert_eq!(response.usage.total_tokens, 8);
+}
+
+#[rstest]
+fn null_token_fields_in_usage_become_zero() {
+    // `_sanitize_usage_obj`: providers that return null token values break
+    // OpenAI clients, so the response is scrubbed at the source.
+    let response = transform_response(json!({
+        "model": "m",
+        "choices": [{"message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}],
+        "usage": {"prompt_tokens": 3, "completion_tokens": null, "total_tokens": null},
+    }))
+    .expect("response normalizes");
+    assert_eq!(response.usage.completion_tokens, 0);
+    assert_eq!(response.usage.total_tokens, 0);
+    assert_eq!(response.usage.prompt_tokens, 3);
+}
+
+#[rstest]
+fn a_tool_call_response_declines_instead_of_dropping_the_calls() {
+    // The `json_mode` rewrite needs a request flag the route does not carry, so
+    // a tool-call answer falls back to Python rather than losing the calls.
+    assert_eq!(
+        transform_response(json!({
+            "model": "m",
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": null,
+                    "tool_calls": [{
+                        "id": "call_1",
+                        "type": "function",
+                        "function": {"name": "f", "arguments": "{}"},
+                    }],
+                },
+                "finish_reason": "tool_calls",
+            }],
+        })),
+        Err(Error::Unsupported("tool call response"))
+    );
+}
+
+#[rstest]
+fn a_refusal_declines_instead_of_returning_an_empty_reply() {
+    assert_eq!(
+        transform_response(json!({
+            "model": "m",
+            "choices": [{
+                "message": {"role": "assistant", "content": null, "refusal": "cannot help"},
+                "finish_reason": "stop",
+            }],
+        })),
+        Err(Error::Unsupported("refusal response"))
+    );
+}
+
+#[rstest]
+fn a_non_text_response_content_declines() {
+    assert_eq!(
+        transform_response(json!({
+            "model": "m",
+            "choices": [{
+                "message": {"role": "assistant", "content": [{"type": "text", "text": "hi"}]},
+                "finish_reason": "stop",
+            }],
+        })),
+        Err(Error::Unsupported("non-text response content"))
+    );
+}
+
+#[rstest]
+#[case::streaming(json!({"stream": true}), "streaming")]
+#[case::unrecognized_param(json!({"some_provider_knob": 1}), "unrecognized request parameter")]
+fn declines(#[case] opts: Value, #[case] expected: &'static str) {
+    assert_eq!(
+        reason(json!([{"role": "user", "content": "hi"}]), opts),
+        Some(Unsupported(expected))
+    );
+}
+
+#[rstest]
+fn accepts_standard_openai_params() {
+    assert_eq!(
+        reason(
+            json!([{"role": "user", "content": "hi"}]),
+            json!({
+                "temperature": 0.2,
+                "top_p": 0.9,
+                "max_tokens": 16,
+                "response_format": {"type": "json_object"},
+                "custom_endpoint": true,
+            }),
+        ),
+        None
+    );
+}
+
+#[rstest]
+fn tool_parameters_decline_before_the_call() {
+    // A `tools` request would come back with tool calls this port cannot
+    // normalize, so it declines at the gate instead of after the call.
+    assert_eq!(
+        reason(
+            json!([{"role": "user", "content": "hi"}]),
+            json!({"tools": [{"type": "function", "function": {"name": "f"}}]}),
+        ),
+        Some(Unsupported("unrecognized request parameter"))
+    );
+}