fix(provider): clamp reasoning effort and filter chat extensions

crates/aether-ai-formats/src/formats/conversion/request.rs

@@ -165,6 +165,7 @@ mod tests {
         convert_openai_chat_request_to_claude_request,
         convert_openai_chat_request_to_openai_responses_request,
         normalize_claude_request_to_openai_chat_request,
+        normalize_gemini_request_to_openai_chat_request,
         normalize_openai_responses_request_to_openai_chat_request,
     };
 
@@ -219,6 +220,43 @@ mod tests {
         assert_eq!(converted["messages"][0]["content"], "hello");
     }
 
+    #[test]
+    fn claude_request_to_chat_clamps_max_reasoning_effort_to_high() {
+        let body = json!({
+            "model": "claude-sonnet",
+            "messages": [{"role": "user", "content": "hello"}],
+            "thinking": {"type": "enabled", "budget_tokens": 1024},
+            "output_config": {"effort": "max"},
+            "max_tokens": 128,
+        });
+
+        let converted =
+            normalize_claude_request_to_openai_chat_request(&body).expect("openai chat request");
+
+        assert_eq!(converted["reasoning_effort"], "high");
+    }
+
+    #[test]
+    fn gemini_request_to_chat_clamps_xhigh_reasoning_effort_to_high() {
+        let body = json!({
+            "contents": [{
+                "role": "user",
+                "parts": [{"text": "hello"}]
+            }],
+            "generationConfig": {
+                "thinkingConfig": {"thinkingBudget": 8192}
+            }
+        });
+
+        let converted = normalize_gemini_request_to_openai_chat_request(
+            &body,
+            "/v1beta/models/gemini-2.5-pro:generateContent",
+        )
+        .expect("openai chat request");
+
+        assert_eq!(converted["reasoning_effort"], "high");
+    }
+
     #[test]
     fn responses_request_normalizer_keeps_tool_history_chat_safe() {
         let call_id_one = "call_weather_123";
@@ -333,6 +371,34 @@ mod tests {
         assert_eq!(messages[0]["content"], "");
     }
 
+    #[test]
+    fn responses_request_normalizer_clamps_chat_reasoning_effort_and_filters_extensions() {
+        let body = json!({
+            "model": "gpt-5.1",
+            "input": "hello",
+            "reasoning": {"effort": "xhigh"},
+            "text": {"verbosity": "high"},
+            "include": ["reasoning.encrypted_content"],
+            "store": false,
+            "service_tier": "priority",
+            "prompt_cache_key": "cache_123",
+            "safety_identifier": "user_123"
+        });
+
+        let converted = normalize_openai_responses_request_to_openai_chat_request(&body)
+            .expect("openai chat request");
+
+        assert_eq!(converted["reasoning_effort"], "high");
+        assert_eq!(converted["verbosity"], "high");
+        assert_eq!(converted["service_tier"], "priority");
+        assert_eq!(converted["prompt_cache_key"], "cache_123");
+        assert_eq!(converted["safety_identifier"], "user_123");
+        assert!(converted.get("include").is_none());
+        assert!(converted.get("store").is_none());
+        assert!(converted.get("text").is_none());
+        assert!(converted.get("reasoning").is_none());
+    }
+
     #[test]
     fn request_normalizer_preserves_multiple_claude_tool_results() {
         let body = json!({

crates/aether-ai-formats/src/formats/openai/chat/request.rs

@@ -1,4 +1,4 @@
-use serde_json::{json, Value};
+use serde_json::{json, Map, Value};
 
 use crate::{
     formats::context::FormatContext,
@@ -194,6 +194,7 @@ pub fn to_raw(canonical: &CanonicalRequest) -> Value {
                     .and_then(|value| value.get("effort"))
                     .and_then(Value::as_str)
             })
+            .and_then(openai_chat_reasoning_effort)
         {
             output.insert(
                 "reasoning_effort".to_string(),
@@ -206,12 +207,12 @@ pub fn to_raw(canonical: &CanonicalRequest) -> Value {
         "openai",
         &output,
     ));
-    output.extend(namespace_extension_object(
+    output.extend(chat_compatible_openai_responses_extension_object(
         &canonical.extensions,
         OPENAI_RESPONSES_EXTENSION_NAMESPACE,
         &output,
     ));
-    output.extend(namespace_extension_object(
+    output.extend(chat_compatible_openai_responses_extension_object(
         &canonical.extensions,
         OPENAI_RESPONSES_LEGACY_EXTENSION_NAMESPACE,
         &output,
@@ -219,6 +220,31 @@ pub fn to_raw(canonical: &CanonicalRequest) -> Value {
     Value::Object(output)
 }
 
+fn openai_chat_reasoning_effort(value: &str) -> Option<&'static str> {
+    match value.trim().to_ascii_lowercase().as_str() {
+        "low" => Some("low"),
+        "medium" => Some("medium"),
+        "high" | "xhigh" | "max" => Some("high"),
+        _ => None,
+    }
+}
+
+fn chat_compatible_openai_responses_extension_object(
+    extensions: &std::collections::BTreeMap<String, Value>,
+    namespace: &str,
+    existing: &Map<String, Value>,
+) -> Map<String, Value> {
+    namespace_extension_object(extensions, namespace, existing)
+        .into_iter()
+        .filter(|(key, _)| {
+            matches!(
+                key.as_str(),
+                "verbosity" | "service_tier" | "prompt_cache_key" | "safety_identifier" | "user"
+            )
+        })
+        .collect()
+}
+
 fn force_stream_options(body: &mut Value, upstream_is_stream: bool) {
     if !upstream_is_stream {
         return;

crates/aether-ai-formats/src/formats/shared/model_directives.rs

@@ -38,8 +38,7 @@ impl ReasoningEffort {
             Self::Low => "low",
             Self::Medium => "medium",
             Self::High => "high",
-            Self::XHigh => "xhigh",
-            Self::Max => "xhigh",
+            Self::XHigh | Self::Max => "high",
         }
     }
 
@@ -524,7 +523,7 @@ mod tests {
             "gpt-5.4-xhigh",
         )
         .expect("directive should apply");
-        assert_eq!(openai_chat["reasoning_effort"], "xhigh");
+        assert_eq!(openai_chat["reasoning_effort"], "high");
 
         let mut responses = json!({
             "model": "gpt-5-upstream",
@@ -597,7 +596,7 @@ mod tests {
             "gpt-5.4-fast-xhigh",
         )
         .expect("directive should apply");
-        assert_eq!(openai_chat["reasoning_effort"], "xhigh");
+        assert_eq!(openai_chat["reasoning_effort"], "high");
         assert_eq!(openai_chat["service_tier"], "priority");
 
         let mut reversed = json!({"model": "gpt-5-upstream", "reasoning_effort": "low"});

crates/aether-ai-formats/src/formats/shared/standard_normalize.rs

@@ -738,7 +738,7 @@ mod tests {
         .expect("openai chat body should build");
 
         assert_eq!(provider_request_body["model"], "gpt-5-upstream");
-        assert_eq!(provider_request_body["reasoning_effort"], "xhigh");
+        assert_eq!(provider_request_body["reasoning_effort"], "high");
     }
 
     #[test]