open_ai: Support specifying reasoning effort (#56411) (cherry-pick to stable) (#56808)

Cherry-pick of #56411 to stable

----
Closes #54875

Release Notes:

- Added support for specifying effort level when using OpenAI models

Co-authored-by: Bennet Bo Fenner <bennet@zed.dev>

crates/language_models/src/provider/open_ai.rs

@@ -6,10 +6,10 @@ use gpui::{AnyView, App, AsyncApp, Context, Entity, SharedString, Task, TaskExt,
 use http_client::HttpClient;
 use language_model::{
     ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel, LanguageModelCompletionError,
-    LanguageModelCompletionEvent, LanguageModelId, LanguageModelName, LanguageModelProvider,
-    LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
-    LanguageModelRequest, LanguageModelToolChoice, OPEN_AI_PROVIDER_ID, OPEN_AI_PROVIDER_NAME,
-    RateLimiter, env_var,
+    LanguageModelCompletionEvent, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,
+    LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
+    LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice, OPEN_AI_PROVIDER_ID,
+    OPEN_AI_PROVIDER_NAME, RateLimiter, env_var,
 };
 use menu;
 use open_ai::{
@@ -351,7 +351,34 @@ impl LanguageModel for OpenAiLanguageModel {
     }
 
     fn supports_thinking(&self) -> bool {
-        self.model.reasoning_effort().is_some()
+        self.model.uses_responses_api() && self.model.reasoning_effort().is_some()
+    }
+
+    fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {
+        if !self.supports_thinking() {
+            return Vec::new();
+        }
+
+        let default_effort = self.model.reasoning_effort();
+        self.model
+            .supported_reasoning_efforts()
+            .iter()
+            .map(|effort| {
+                let (name, value) = match effort {
+                    open_ai::ReasoningEffort::Minimal => ("Minimal", "minimal"),
+                    open_ai::ReasoningEffort::Low => ("Low", "low"),
+                    open_ai::ReasoningEffort::Medium => ("Medium", "medium"),
+                    open_ai::ReasoningEffort::High => ("High", "high"),
+                    open_ai::ReasoningEffort::XHigh => ("Extra High", "xhigh"),
+                };
+
+                LanguageModelEffortLevel {
+                    name: name.into(),
+                    value: value.into(),
+                    is_default: Some(*effort) == default_effort,
+                }
+            })
+            .collect()
     }
 
     fn supports_split_token_display(&self) -> bool {
@@ -406,7 +433,7 @@ impl LanguageModel for OpenAiLanguageModel {
                 self.model.supports_parallel_tool_calls(),
                 self.model.supports_prompt_cache_key(),
                 self.max_output_tokens(),
-                self.model.reasoning_effort(),
+                None,
                 false,
             );
             let completions = self.stream_completion(request, cx);

crates/open_ai/src/completion.rs

@@ -190,8 +190,8 @@ pub fn into_open_ai_response(
         tool_choice,
         stop: _,
         temperature,
-        thinking_allowed: _,
-        thinking_effort: _,
+        thinking_allowed,
+        thinking_effort,
         speed: _,
     } = request;
 
@@ -233,10 +233,18 @@ pub fn into_open_ai_response(
         } else {
             None
         },
-        reasoning: reasoning_effort.map(|effort| crate::responses::ReasoningConfig {
-            effort,
-            summary: Some(crate::responses::ReasoningSummaryMode::Auto),
-        }),
+        reasoning: if thinking_allowed {
+            thinking_effort
+                .as_deref()
+                .and_then(|effort| effort.parse::<ReasoningEffort>().ok())
+                .or(reasoning_effort)
+                .map(|effort| crate::responses::ReasoningConfig {
+                    effort,
+                    summary: Some(crate::responses::ReasoningSummaryMode::Auto),
+                })
+        } else {
+            None
+        },
     }
 }
 
@@ -1000,8 +1008,8 @@ mod tests {
             tool_choice: Some(LanguageModelToolChoice::Any),
             stop: vec!["<STOP>".into()],
             temperature: None,
-            thinking_allowed: false,
-            thinking_effort: None,
+            thinking_allowed: true,
+            thinking_effort: Some("high".into()),
             speed: None,
         };
 
@@ -1065,12 +1073,46 @@ mod tests {
                 }
             ],
             "prompt_cache_key": "thread-123",
-            "reasoning": { "effort": "low", "summary": "auto" }
+            "reasoning": { "effort": "high", "summary": "auto" }
         });
 
         assert_eq!(serialized, expected);
     }
 
+    #[test]
+    fn into_open_ai_response_omits_reasoning_when_thinking_is_disabled() {
+        let request = LanguageModelRequest {
+            thread_id: None,
+            prompt_id: None,
+            intent: None,
+            messages: vec![LanguageModelRequestMessage {
+                role: Role::User,
+                content: vec![MessageContent::Text("Hello".into())],
+                cache: false,
+                reasoning_details: None,
+            }],
+            tools: Vec::new(),
+            tool_choice: None,
+            stop: Vec::new(),
+            temperature: None,
+            thinking_allowed: false,
+            thinking_effort: Some("high".into()),
+            speed: None,
+        };
+
+        let response = into_open_ai_response(
+            request,
+            "gpt-5",
+            true,
+            true,
+            None,
+            Some(ReasoningEffort::Medium),
+        );
+
+        let serialized = serde_json::to_value(&response).unwrap();
+        assert_eq!(serialized.get("reasoning"), None);
+    }
+
     #[test]
     fn responses_stream_maps_tool_calls() {
         let events = vec![

crates/open_ai/src/open_ai.rs

@@ -265,13 +265,69 @@ impl Model {
             Self::Custom {
                 reasoning_effort, ..
             } => reasoning_effort.to_owned(),
-            Self::FivePointThreeCodex | Self::FivePointFourPro | Self::FivePointFivePro => {
-                Some(ReasoningEffort::Medium)
-            }
+            Self::O1
+            | Self::O3
+            | Self::O3Mini
+            | Self::Five
+            | Self::FiveCodex
+            | Self::FiveMini
+            | Self::FiveNano
+            | Self::FivePointOne
+            | Self::FivePointTwo
+            | Self::FivePointTwoCodex
+            | Self::FivePointThreeCodex
+            | Self::FivePointFour
+            | Self::FivePointFourPro
+            | Self::FivePointFive
+            | Self::FivePointFivePro => Some(ReasoningEffort::Medium),
             _ => None,
         }
     }
 
+    pub fn supported_reasoning_efforts(&self) -> &'static [ReasoningEffort] {
+        match self {
+            Self::Custom {
+                reasoning_effort: Some(effort),
+                ..
+            } => match effort {
+                ReasoningEffort::Minimal => &[ReasoningEffort::Minimal],
+                ReasoningEffort::Low => &[ReasoningEffort::Low],
+                ReasoningEffort::Medium => &[ReasoningEffort::Medium],
+                ReasoningEffort::High => &[ReasoningEffort::High],
+                ReasoningEffort::XHigh => &[ReasoningEffort::XHigh],
+            },
+            Self::O1 | Self::O3 | Self::O3Mini | Self::FivePointOne => &[
+                ReasoningEffort::Low,
+                ReasoningEffort::Medium,
+                ReasoningEffort::High,
+            ],
+            Self::Five | Self::FiveMini | Self::FiveNano => &[
+                ReasoningEffort::Minimal,
+                ReasoningEffort::Low,
+                ReasoningEffort::Medium,
+                ReasoningEffort::High,
+            ],
+            Self::FiveCodex
+            | Self::FivePointTwoCodex
+            | Self::FivePointThreeCodex
+            | Self::FivePointFourPro => &[
+                ReasoningEffort::Medium,
+                ReasoningEffort::High,
+                ReasoningEffort::XHigh,
+            ],
+            Self::FivePointTwo
+            | Self::FivePointFour
+            | Self::FivePointFive
+            | Self::FivePointFivePro => &[
+                ReasoningEffort::Low,
+                ReasoningEffort::Medium,
+                ReasoningEffort::High,
+                ReasoningEffort::XHigh,
+            ],
+            _ => &[],
+        }
+    }
+
     pub fn uses_responses_api(&self) -> bool {
         match self {
             Self::Custom {