agent: Allow customizing temperature by provider/model (#30033)

Adds a new `agent.model_parameters` setting that allows the user to
specify a custom temperature for a provider AND/OR model:

```json5
    "model_parameters": [
      // To set parameters for all requests to OpenAI models:
      {
        "provider": "openai",
        "temperature": 0.5
      },
      // To set parameters for all requests in general:
      {
        "temperature": 0
      },
      // To set parameters for a specific provider and model:
      {
        "provider": "zed.dev",
        "model": "claude-3-7-sonnet-latest",
        "temperature": 1.0
      }
    ],
```

Release Notes:

- agent: Allow customizing temperature by provider/model

---------

Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Marshall Bowers <git@maxdeviant.com>

Cargo.lock

@@ -595,6 +595,7 @@ version = "0.1.0"
 dependencies = [
  "anthropic",
  "anyhow",
+ "collections",
  "deepseek",
  "feature_flags",
  "fs",
@@ -3006,6 +3007,7 @@ dependencies = [
  "anyhow",
  "assistant",
  "assistant_context_editor",
+ "assistant_settings",
  "assistant_slash_command",
  "assistant_tool",
  "async-stripe",

assets/settings/default.json

@@ -605,13 +605,11 @@
     //
     // Default: main
     "fallback_branch_name": "main",
-
     // Whether to sort entries in the panel by path
     // or by status (the default).
     //
     // Default: false
     "sort_by_path": false,
-
     "scrollbar": {
       // When to show the scrollbar in the git panel.
       //
@@ -661,6 +659,28 @@
       // The model to use.
       "model": "claude-3-7-sonnet-latest"
     },
+    // Additional parameters for language model requests. When making a request to a model, parameters will be taken
+    // from the last entry in this list that matches the model's provider and name. In each entry, both provider
+    // and model are optional, so that you can specify parameters for either one.
+    "model_parameters": [
+      // To set parameters for all requests to OpenAI models:
+      // {
+      //   "provider": "openai",
+      //   "temperature": 0.5
+      // }
+      //
+      // To set parameters for all requests in general:
+      // {
+      //   "temperature": 0
+      // }
+      //
+      // To set parameters for a specific provider and model:
+      // {
+      //   "provider": "zed.dev",
+      //   "model": "claude-3-7-sonnet-latest",
+      //   "temperature": 1.0
+      // }
+    ],
     // When enabled, the agent can run potentially destructive actions without asking for your confirmation.
     "always_allow_tool_actions": false,
     // When enabled, the agent will stream edits.

crates/agent/src/active_thread.rs

@@ -1417,7 +1417,10 @@ impl ActiveThread {
                         messages: vec![request_message],
                         tools: vec![],
                         stop: vec![],
-                        temperature: None,
+                        temperature: AssistantSettings::temperature_for_model(
+                            &configured_model.model,
+                            cx,
+                        ),
                     };
 
                     Some(configured_model.model.count_tokens(request, cx))

crates/agent/src/buffer_codegen.rs

@@ -2,6 +2,7 @@ use crate::context::ContextLoadResult;
 use crate::inline_prompt_editor::CodegenStatus;
 use crate::{context::load_context, context_store::ContextStore};
 use anyhow::Result;
+use assistant_settings::AssistantSettings;
 use client::telemetry::Telemetry;
 use collections::HashSet;
 use editor::{Anchor, AnchorRangeExt, MultiBuffer, MultiBufferSnapshot, ToOffset as _, ToPoint};
@@ -383,7 +384,7 @@ impl CodegenAlternative {
             if user_prompt.trim().to_lowercase() == "delete" {
                 async { Ok(LanguageModelTextStream::default()) }.boxed_local()
             } else {
-                let request = self.build_request(user_prompt, cx)?;
+                let request = self.build_request(&model, user_prompt, cx)?;
                 cx.spawn(async move |_, cx| model.stream_completion_text(request.await, &cx).await)
                     .boxed_local()
             };
@@ -393,6 +394,7 @@ impl CodegenAlternative {
 
     fn build_request(
         &self,
+        model: &Arc<dyn LanguageModel>,
         user_prompt: String,
         cx: &mut App,
     ) -> Result<Task<LanguageModelRequest>> {
@@ -441,6 +443,8 @@ impl CodegenAlternative {
             }
         });
 
+        let temperature = AssistantSettings::temperature_for_model(&model, cx);
+
         Ok(cx.spawn(async move |_cx| {
             let mut request_message = LanguageModelRequestMessage {
                 role: Role::User,
@@ -463,7 +467,7 @@ impl CodegenAlternative {
                 mode: None,
                 tools: Vec::new(),
                 stop: Vec::new(),
-                temperature: None,
+                temperature,
                 messages: vec![request_message],
             }
         }))

crates/agent/src/message_editor.rs

@@ -8,7 +8,7 @@ use crate::ui::{
     AnimatedLabel, MaxModeTooltip,
     preview::{AgentPreview, UsageCallout},
 };
-use assistant_settings::CompletionMode;
+use assistant_settings::{AssistantSettings, CompletionMode};
 use buffer_diff::BufferDiff;
 use client::UserStore;
 use collections::{HashMap, HashSet};
@@ -1273,7 +1273,7 @@ impl MessageEditor {
                         messages: vec![request_message],
                         tools: vec![],
                         stop: vec![],
-                        temperature: None,
+                        temperature: AssistantSettings::temperature_for_model(&model.model, cx),
                     };
 
                     Some(model.model.count_tokens(request, cx))

crates/agent/src/terminal_inline_assistant.rs

@@ -6,6 +6,7 @@ use crate::inline_prompt_editor::{
 use crate::terminal_codegen::{CLEAR_INPUT, CodegenEvent, TerminalCodegen};
 use crate::thread_store::{TextThreadStore, ThreadStore};
 use anyhow::{Context as _, Result};
+use assistant_settings::AssistantSettings;
 use client::telemetry::Telemetry;
 use collections::{HashMap, VecDeque};
 use editor::{MultiBuffer, actions::SelectAll};
@@ -266,6 +267,12 @@ impl TerminalInlineAssistant {
             load_context(contexts, project, &assist.prompt_store, cx)
         })?;
 
+        let ConfiguredModel { model, .. } = LanguageModelRegistry::read_global(cx)
+            .inline_assistant_model()
+            .context("No inline assistant model")?;
+
+        let temperature = AssistantSettings::temperature_for_model(&model, cx);
+
         Ok(cx.background_spawn(async move {
             let mut request_message = LanguageModelRequestMessage {
                 role: Role::User,
@@ -287,7 +294,7 @@ impl TerminalInlineAssistant {
                 messages: vec![request_message],
                 tools: Vec::new(),
                 stop: Vec::new(),
-                temperature: None,
+                temperature,
             }
         }))
     }

crates/agent/src/thread.rs

@@ -1145,7 +1145,7 @@ impl Thread {
             messages: vec![],
             tools: Vec::new(),
             stop: Vec::new(),
-            temperature: None,
+            temperature: AssistantSettings::temperature_for_model(&model, cx),
         };
 
         let available_tools = self.available_tools(cx, model.clone());
@@ -1251,7 +1251,12 @@ impl Thread {
         request
     }
 
-    fn to_summarize_request(&self, added_user_message: String) -> LanguageModelRequest {
+    fn to_summarize_request(
+        &self,
+        model: &Arc<dyn LanguageModel>,
+        added_user_message: String,
+        cx: &App,
+    ) -> LanguageModelRequest {
         let mut request = LanguageModelRequest {
             thread_id: None,
             prompt_id: None,
@@ -1259,7 +1264,7 @@ impl Thread {
             messages: vec![],
             tools: Vec::new(),
             stop: Vec::new(),
-            temperature: None,
+            temperature: AssistantSettings::temperature_for_model(model, cx),
         };
 
         for message in &self.messages {
@@ -1696,7 +1701,7 @@ impl Thread {
             If the conversation is about a specific subject, include it in the title. \
             Be descriptive. DO NOT speak in the first person.";
 
-        let request = self.to_summarize_request(added_user_message.into());
+        let request = self.to_summarize_request(&model.model, added_user_message.into(), cx);
 
         self.pending_summary = cx.spawn(async move |this, cx| {
             async move {
@@ -1782,7 +1787,7 @@ impl Thread {
              4. Any action items or next steps if any\n\
              Format it in Markdown with headings and bullet points.";
 
-        let request = self.to_summarize_request(added_user_message.into());
+        let request = self.to_summarize_request(&model, added_user_message.into(), cx);
 
         *self.detailed_summary_tx.borrow_mut() = DetailedSummaryState::Generating {
             message_id: last_message_id,
@@ -2655,7 +2660,7 @@ struct PendingCompletion {
 mod tests {
     use super::*;
     use crate::{ThreadStore, context::load_context, context_store::ContextStore, thread_store};
-    use assistant_settings::AssistantSettings;
+    use assistant_settings::{AssistantSettings, LanguageModelParameters};
     use assistant_tool::ToolRegistry;
     use editor::EditorSettings;
     use gpui::TestAppContext;
@@ -3066,6 +3071,100 @@ fn main() {{
         );
     }
 
+    #[gpui::test]
+    async fn test_temperature_setting(cx: &mut TestAppContext) {
+        init_test_settings(cx);
+
+        let project = create_test_project(
+            cx,
+            json!({"code.rs": "fn main() {\n    println!(\"Hello, world!\");\n}"}),
+        )
+        .await;
+
+        let (_workspace, _thread_store, thread, _context_store, model) =
+            setup_test_environment(cx, project.clone()).await;
+
+        // Both model and provider
+        cx.update(|cx| {
+            AssistantSettings::override_global(
+                AssistantSettings {
+                    model_parameters: vec![LanguageModelParameters {
+                        provider: Some(model.provider_id().0.to_string().into()),
+                        model: Some(model.id().0.clone()),
+                        temperature: Some(0.66),
+                    }],
+                    ..AssistantSettings::get_global(cx).clone()
+                },
+                cx,
+            );
+        });
+
+        let request = thread.update(cx, |thread, cx| {
+            thread.to_completion_request(model.clone(), cx)
+        });
+        assert_eq!(request.temperature, Some(0.66));
+
+        // Only model
+        cx.update(|cx| {
+            AssistantSettings::override_global(
+                AssistantSettings {
+                    model_parameters: vec![LanguageModelParameters {
+                        provider: None,
+                        model: Some(model.id().0.clone()),
+                        temperature: Some(0.66),
+                    }],
+                    ..AssistantSettings::get_global(cx).clone()
+                },
+                cx,
+            );
+        });
+
+        let request = thread.update(cx, |thread, cx| {
+            thread.to_completion_request(model.clone(), cx)
+        });
+        assert_eq!(request.temperature, Some(0.66));
+
+        // Only provider
+        cx.update(|cx| {
+            AssistantSettings::override_global(
+                AssistantSettings {
+                    model_parameters: vec![LanguageModelParameters {
+                        provider: Some(model.provider_id().0.to_string().into()),
+                        model: None,
+                        temperature: Some(0.66),
+                    }],
+                    ..AssistantSettings::get_global(cx).clone()
+                },
+                cx,
+            );
+        });
+
+        let request = thread.update(cx, |thread, cx| {
+            thread.to_completion_request(model.clone(), cx)
+        });
+        assert_eq!(request.temperature, Some(0.66));
+
+        // Same model name, different provider
+        cx.update(|cx| {
+            AssistantSettings::override_global(
+                AssistantSettings {
+                    model_parameters: vec![LanguageModelParameters {
+                        provider: Some("anthropic".into()),
+                        model: Some(model.id().0.clone()),
+                        temperature: Some(0.66),
+                    }],
+                    ..AssistantSettings::get_global(cx).clone()
+                },
+                cx,
+            );
+        });
+
+        let request = thread.update(cx, |thread, cx| {
+            thread.to_completion_request(model.clone(), cx)
+        });
+        assert_eq!(request.temperature, None);
+    }
+
     fn init_test_settings(cx: &mut TestAppContext) {
         cx.update(|cx| {
             let settings_store = SettingsStore::test(cx);

crates/assistant/src/assistant.rs

@@ -163,7 +163,7 @@ fn update_active_language_model_from_settings(cx: &mut App) {
 
     fn to_selected_model(selection: &LanguageModelSelection) -> language_model::SelectedModel {
         language_model::SelectedModel {
-            provider: LanguageModelProviderId::from(selection.provider.clone()),
+            provider: LanguageModelProviderId::from(selection.provider.0.clone()),
             model: LanguageModelId::from(selection.model.clone()),
         }
     }

crates/assistant/src/inline_assistant.rs

@@ -2484,7 +2484,7 @@ impl InlineAssist {
                     .read(cx)
                     .active_context(cx)?
                     .read(cx)
-                    .to_completion_request(RequestType::Chat, cx),
+                    .to_completion_request(None, RequestType::Chat, cx),
             )
         } else {
             None
@@ -2870,7 +2870,8 @@ impl CodegenAlternative {
         if let Some(ConfiguredModel { model, .. }) =
             LanguageModelRegistry::read_global(cx).inline_assistant_model()
         {
-            let request = self.build_request(user_prompt, assistant_panel_context.clone(), cx);
+            let request =
+                self.build_request(&model, user_prompt, assistant_panel_context.clone(), cx);
             match request {
                 Ok(request) => {
                     let total_count = model.count_tokens(request.clone(), cx);
@@ -2915,7 +2916,8 @@ impl CodegenAlternative {
             if user_prompt.trim().to_lowercase() == "delete" {
                 async { Ok(LanguageModelTextStream::default()) }.boxed_local()
             } else {
-                let request = self.build_request(user_prompt, assistant_panel_context, cx)?;
+                let request =
+                    self.build_request(&model, user_prompt, assistant_panel_context, cx)?;
                 self.request = Some(request.clone());
 
                 cx.spawn(async move |_, cx| model.stream_completion_text(request, &cx).await)
@@ -2927,6 +2929,7 @@ impl CodegenAlternative {
 
     fn build_request(
         &self,
+        model: &Arc<dyn LanguageModel>,
         user_prompt: String,
         assistant_panel_context: Option<LanguageModelRequest>,
         cx: &App,
@@ -2981,7 +2984,7 @@ impl CodegenAlternative {
             messages,
             tools: Vec::new(),
             stop: Vec::new(),
-            temperature: None,
+            temperature: AssistantSettings::temperature_for_model(&model, cx),
         })
     }
 

crates/assistant/src/terminal_inline_assistant.rs

@@ -261,7 +261,7 @@ impl TerminalInlineAssistant {
                         .read(cx)
                         .active_context(cx)?
                         .read(cx)
-                        .to_completion_request(RequestType::Chat, cx),
+                        .to_completion_request(None, RequestType::Chat, cx),
                 )
             })
         } else {

crates/assistant_context_editor/src/context.rs

@@ -3,6 +3,7 @@ mod context_tests;
 
 use crate::patch::{AssistantEdit, AssistantPatch, AssistantPatchStatus};
 use anyhow::{Context as _, Result, anyhow};
+use assistant_settings::AssistantSettings;
 use assistant_slash_command::{
     SlashCommandContent, SlashCommandEvent, SlashCommandLine, SlashCommandOutputSection,
     SlashCommandResult, SlashCommandWorkingSet,
@@ -1273,10 +1274,10 @@ impl AssistantContext {
     pub(crate) fn count_remaining_tokens(&mut self, cx: &mut Context<Self>) {
         // Assume it will be a Chat request, even though that takes fewer tokens (and risks going over the limit),
         // because otherwise you see in the UI that your empty message has a bunch of tokens already used.
-        let request = self.to_completion_request(RequestType::Chat, cx);
         let Some(model) = LanguageModelRegistry::read_global(cx).default_model() else {
             return;
         };
+        let request = self.to_completion_request(Some(&model.model), RequestType::Chat, cx);
         let debounce = self.token_count.is_some();
         self.pending_token_count = cx.spawn(async move |this, cx| {
             async move {
@@ -1422,7 +1423,7 @@ impl AssistantContext {
         }
 
         let request = {
-            let mut req = self.to_completion_request(RequestType::Chat, cx);
+            let mut req = self.to_completion_request(Some(&model), RequestType::Chat, cx);
             // Skip the last message because it's likely to change and
             // therefore would be a waste to cache.
             req.messages.pop();
@@ -2321,7 +2322,7 @@ impl AssistantContext {
         // Compute which messages to cache, including the last one.
         self.mark_cache_anchors(&model.cache_configuration(), false, cx);
 
-        let request = self.to_completion_request(request_type, cx);
+        let request = self.to_completion_request(Some(&model), request_type, cx);
 
         let assistant_message = self
             .insert_message_after(last_message_id, Role::Assistant, MessageStatus::Pending, cx)
@@ -2561,6 +2562,7 @@ impl AssistantContext {
 
     pub fn to_completion_request(
         &self,
+        model: Option<&Arc<dyn LanguageModel>>,
         request_type: RequestType,
         cx: &App,
     ) -> LanguageModelRequest {
@@ -2584,7 +2586,8 @@ impl AssistantContext {
             messages: Vec::new(),
             tools: Vec::new(),
             stop: Vec::new(),
-            temperature: None,
+            temperature: model
+                .and_then(|model| AssistantSettings::temperature_for_model(model, cx)),
         };
         for message in self.messages(cx) {
             if message.status != MessageStatus::Done {
@@ -2981,7 +2984,7 @@ impl AssistantContext {
                 return;
             }
 
-            let mut request = self.to_completion_request(RequestType::Chat, cx);
+            let mut request = self.to_completion_request(Some(&model.model), RequestType::Chat, cx);
             request.messages.push(LanguageModelRequestMessage {
                 role: Role::User,
                 content: vec![

crates/assistant_context_editor/src/context/context_tests.rs

@@ -43,9 +43,8 @@ use workspace::Workspace;
 
 #[gpui::test]
 fn test_inserting_and_removing_messages(cx: &mut App) {
-    let settings_store = SettingsStore::test(cx);
+    init_test(cx);
-    LanguageModelRegistry::test(cx);
+
-    cx.set_global(settings_store);
     let registry = Arc::new(LanguageRegistry::test(cx.background_executor().clone()));
     let prompt_builder = Arc::new(PromptBuilder::new(None).unwrap());
     let context = cx.new(|cx| {
@@ -182,9 +181,8 @@ fn test_inserting_and_removing_messages(cx: &mut App) {
 
 #[gpui::test]
 fn test_message_splitting(cx: &mut App) {
-    let settings_store = SettingsStore::test(cx);
+    init_test(cx);
-    cx.set_global(settings_store);
+
-    LanguageModelRegistry::test(cx);
     let registry = Arc::new(LanguageRegistry::test(cx.background_executor().clone()));
 
     let prompt_builder = Arc::new(PromptBuilder::new(None).unwrap());
@@ -285,9 +283,8 @@ fn test_message_splitting(cx: &mut App) {
 
 #[gpui::test]
 fn test_messages_for_offsets(cx: &mut App) {
-    let settings_store = SettingsStore::test(cx);
+    init_test(cx);
-    LanguageModelRegistry::test(cx);
+
-    cx.set_global(settings_store);
     let registry = Arc::new(LanguageRegistry::test(cx.background_executor().clone()));
     let prompt_builder = Arc::new(PromptBuilder::new(None).unwrap());
     let context = cx.new(|cx| {
@@ -378,10 +375,8 @@ fn test_messages_for_offsets(cx: &mut App) {
 
 #[gpui::test]
 async fn test_slash_commands(cx: &mut TestAppContext) {
-    let settings_store = cx.update(SettingsStore::test);
+    cx.update(init_test);
-    cx.set_global(settings_store);
+
-    cx.update(LanguageModelRegistry::test);
-    cx.update(Project::init_settings);
     let fs = FakeFs::new(cx.background_executor.clone());
 
     fs.insert_tree(
@@ -671,22 +666,19 @@ async fn test_slash_commands(cx: &mut TestAppContext) {
 
 #[gpui::test]
 async fn test_workflow_step_parsing(cx: &mut TestAppContext) {
-    cx.update(prompt_store::init);
-    let mut settings_store = cx.update(SettingsStore::test);
     cx.update(|cx| {
-        settings_store
+        init_test(cx);
-            .set_user_settings(
+        cx.update_global(|settings_store: &mut SettingsStore, cx| {
-                r#"{ "assistant": { "enable_experimental_live_diffs": true } }"#,
+            settings_store
-                cx,
+                .set_user_settings(
-            )
+                    r#"{ "assistant": { "enable_experimental_live_diffs": true } }"#,
-            .unwrap()
+                    cx,
+                )
+                .unwrap()
+        })
     });
-    cx.set_global(settings_store);
-    cx.update(language::init);
-    cx.update(Project::init_settings);
     let fs = FakeFs::new(cx.executor());
     let project = Project::test(fs, [Path::new("/root")], cx).await;
-    cx.update(LanguageModelRegistry::test);
 
     let registry = Arc::new(LanguageRegistry::test(cx.executor()));
 
@@ -1069,9 +1061,8 @@ async fn test_workflow_step_parsing(cx: &mut TestAppContext) {
 
 #[gpui::test]
 async fn test_serialization(cx: &mut TestAppContext) {
-    let settings_store = cx.update(SettingsStore::test);
+    cx.update(init_test);
-    cx.set_global(settings_store);
+
-    cx.update(LanguageModelRegistry::test);
     let registry = Arc::new(LanguageRegistry::test(cx.executor()));
     let prompt_builder = Arc::new(PromptBuilder::new(None).unwrap());
     let context = cx.new(|cx| {
@@ -1147,6 +1138,8 @@ async fn test_serialization(cx: &mut TestAppContext) {
 
 #[gpui::test(iterations = 100)]
 async fn test_random_context_collaboration(cx: &mut TestAppContext, mut rng: StdRng) {
+    cx.update(init_test);
+
     let min_peers = env::var("MIN_PEERS")
         .map(|i| i.parse().expect("invalid `MIN_PEERS` variable"))
         .unwrap_or(2);
@@ -1157,10 +1150,6 @@ async fn test_random_context_collaboration(cx: &mut TestAppContext, mut rng: Std
         .map(|i| i.parse().expect("invalid `OPERATIONS` variable"))
         .unwrap_or(50);
 
-    let settings_store = cx.update(SettingsStore::test);
-    cx.set_global(settings_store);
-    cx.update(LanguageModelRegistry::test);
-
     let slash_commands = cx.update(SlashCommandRegistry::default_global);
     slash_commands.register_command(FakeSlashCommand("cmd-1".into()), false);
     slash_commands.register_command(FakeSlashCommand("cmd-2".into()), false);
@@ -1429,9 +1418,8 @@ async fn test_random_context_collaboration(cx: &mut TestAppContext, mut rng: Std
 
 #[gpui::test]
 fn test_mark_cache_anchors(cx: &mut App) {
-    let settings_store = SettingsStore::test(cx);
+    init_test(cx);
-    LanguageModelRegistry::test(cx);
+
-    cx.set_global(settings_store);
     let registry = Arc::new(LanguageRegistry::test(cx.background_executor().clone()));
     let prompt_builder = Arc::new(PromptBuilder::new(None).unwrap());
     let context = cx.new(|cx| {
@@ -1606,6 +1594,16 @@ fn messages_cache(
         .collect()
 }
 
+fn init_test(cx: &mut App) {
+    let settings_store = SettingsStore::test(cx);
+    prompt_store::init(cx);
+    LanguageModelRegistry::test(cx);
+    cx.set_global(settings_store);
+    language::init(cx);
+    assistant_settings::init(cx);
+    Project::init_settings(cx);
+}
+
 #[derive(Clone)]
 struct FakeSlashCommand(String);
 

crates/assistant_settings/Cargo.toml

@@ -14,6 +14,7 @@ path = "src/assistant_settings.rs"
 [dependencies]
 anthropic = { workspace = true, features = ["schemars"] }
 anyhow.workspace = true
+collections.workspace = true
 feature_flags.workspace = true
 gpui.workspace = true
 indexmap.workspace = true

crates/assistant_settings/src/agent_profile.rs

@@ -1,7 +1,7 @@
 use std::sync::Arc;
 
+use collections::IndexMap;
 use gpui::SharedString;
-use indexmap::IndexMap;
 use schemars::JsonSchema;
 use serde::{Deserialize, Serialize};
 

crates/assistant_settings/src/assistant_settings.rs

@@ -5,10 +5,10 @@ use std::sync::Arc;
 use ::open_ai::Model as OpenAiModel;
 use anthropic::Model as AnthropicModel;
 use anyhow::{Result, bail};
+use collections::IndexMap;
 use deepseek::Model as DeepseekModel;
 use feature_flags::{AgentStreamEditsFeatureFlag, Assistant2FeatureFlag, FeatureFlagAppExt};
-use gpui::{App, Pixels};
+use gpui::{App, Pixels, SharedString};
-use indexmap::IndexMap;
 use language_model::{CloudModel, LanguageModel};
 use lmstudio::Model as LmStudioModel;
 use ollama::Model as OllamaModel;
@@ -18,6 +18,10 @@ use settings::{Settings, SettingsSources};
 
 pub use crate::agent_profile::*;
 
+pub fn init(cx: &mut App) {
+    AssistantSettings::register(cx);
+}
+
 #[derive(Copy, Clone, Default, Debug, Serialize, Deserialize, JsonSchema)]
 #[serde(rename_all = "snake_case")]
 pub enum AssistantDockPosition {
@@ -89,10 +93,20 @@ pub struct AssistantSettings {
     pub notify_when_agent_waiting: NotifyWhenAgentWaiting,
     pub stream_edits: bool,
     pub single_file_review: bool,
+    pub model_parameters: Vec<LanguageModelParameters>,
     pub preferred_completion_mode: CompletionMode,
 }
 
 impl AssistantSettings {
+    pub fn temperature_for_model(model: &Arc<dyn LanguageModel>, cx: &App) -> Option<f32> {
+        let settings = Self::get_global(cx);
+        settings
+            .model_parameters
+            .iter()
+            .rfind(|setting| setting.matches(model))
+            .and_then(|m| m.temperature)
+    }
+
     pub fn stream_edits(&self, cx: &App) -> bool {
         cx.has_flag::<AgentStreamEditsFeatureFlag>() || self.stream_edits
     }
@@ -106,15 +120,47 @@ impl AssistantSettings {
     }
 
     pub fn set_inline_assistant_model(&mut self, provider: String, model: String) {
-        self.inline_assistant_model = Some(LanguageModelSelection { provider, model });
+        self.inline_assistant_model = Some(LanguageModelSelection {
+            provider: provider.into(),
+            model,
+        });
     }
 
     pub fn set_commit_message_model(&mut self, provider: String, model: String) {
-        self.commit_message_model = Some(LanguageModelSelection { provider, model });
+        self.commit_message_model = Some(LanguageModelSelection {
+            provider: provider.into(),
+            model,
+        });
     }
 
     pub fn set_thread_summary_model(&mut self, provider: String, model: String) {
-        self.thread_summary_model = Some(LanguageModelSelection { provider, model });
+        self.thread_summary_model = Some(LanguageModelSelection {
+            provider: provider.into(),
+            model,
+        });
+    }
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize, JsonSchema, PartialEq)]
+pub struct LanguageModelParameters {
+    pub provider: Option<LanguageModelProviderSetting>,
+    pub model: Option<SharedString>,
+    pub temperature: Option<f32>,
+}
+
+impl LanguageModelParameters {
+    pub fn matches(&self, model: &Arc<dyn LanguageModel>) -> bool {
+        if let Some(provider) = &self.provider {
+            if provider.0 != model.provider_id().0 {
+                return false;
+            }
+        }
+        if let Some(setting_model) = &self.model {
+            if *setting_model != model.id().0 {
+                return false;
+            }
+        }
+        true
     }
 }
 
@@ -181,37 +227,37 @@ impl AssistantSettingsContent {
                         .and_then(|provider| match provider {
                             AssistantProviderContentV1::ZedDotDev { default_model } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "zed.dev".to_string(),
+                                    provider: "zed.dev".into(),
                                     model: model.id().to_string(),
                                 })
                             }
                             AssistantProviderContentV1::OpenAi { default_model, .. } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "openai".to_string(),
+                                    provider: "openai".into(),
                                     model: model.id().to_string(),
                                 })
                             }
                             AssistantProviderContentV1::Anthropic { default_model, .. } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "anthropic".to_string(),
+                                    provider: "anthropic".into(),
                                     model: model.id().to_string(),
                                 })
                             }
                             AssistantProviderContentV1::Ollama { default_model, .. } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "ollama".to_string(),
+                                    provider: "ollama".into(),
                                     model: model.id().to_string(),
                                 })
                             }
                             AssistantProviderContentV1::LmStudio { default_model, .. } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "lmstudio".to_string(),
+                                    provider: "lmstudio".into(),
                                     model: model.id().to_string(),
                                 })
                             }
                             AssistantProviderContentV1::DeepSeek { default_model, .. } => {
                                 default_model.map(|model| LanguageModelSelection {
-                                    provider: "deepseek".to_string(),
+                                    provider: "deepseek".into(),
                                     model: model.id().to_string(),
                                 })
                             }
@@ -227,6 +273,7 @@ impl AssistantSettingsContent {
                     notify_when_agent_waiting: None,
                     stream_edits: None,
                     single_file_review: None,
+                    model_parameters: Vec::new(),
                     preferred_completion_mode: None,
                 },
                 VersionedAssistantSettingsContent::V2(ref settings) => settings.clone(),
@@ -238,7 +285,7 @@ impl AssistantSettingsContent {
                 default_width: settings.default_width,
                 default_height: settings.default_height,
                 default_model: Some(LanguageModelSelection {
-                    provider: "openai".to_string(),
+                    provider: "openai".into(),
                     model: settings
                         .default_open_ai_model
                         .clone()
@@ -257,6 +304,7 @@ impl AssistantSettingsContent {
                 notify_when_agent_waiting: None,
                 stream_edits: None,
                 single_file_review: None,
+                model_parameters: Vec::new(),
                 preferred_completion_mode: None,
             },
             None => AssistantSettingsContentV2::default(),
@@ -370,7 +418,10 @@ impl AssistantSettingsContent {
                     }
                 }
                 VersionedAssistantSettingsContent::V2(ref mut settings) => {
-                    settings.default_model = Some(LanguageModelSelection { provider, model });
+                    settings.default_model = Some(LanguageModelSelection {
+                        provider: provider.into(),
+                        model,
+                    });
                 }
             },
             Some(AssistantSettingsContentInner::Legacy(settings)) => {
@@ -381,7 +432,10 @@ impl AssistantSettingsContent {
             None => {
                 self.inner = Some(AssistantSettingsContentInner::for_v2(
                     AssistantSettingsContentV2 {
-                        default_model: Some(LanguageModelSelection { provider, model }),
+                        default_model: Some(LanguageModelSelection {
+                            provider: provider.into(),
+                            model,
+                        }),
                         ..Default::default()
                     },
                 ));
@@ -391,7 +445,10 @@ impl AssistantSettingsContent {
 
     pub fn set_inline_assistant_model(&mut self, provider: String, model: String) {
         self.v2_setting(|setting| {
-            setting.inline_assistant_model = Some(LanguageModelSelection { provider, model });
+            setting.inline_assistant_model = Some(LanguageModelSelection {
+                provider: provider.into(),
+                model,
+            });
             Ok(())
         })
         .ok();
@@ -399,7 +456,10 @@ impl AssistantSettingsContent {
 
     pub fn set_commit_message_model(&mut self, provider: String, model: String) {
         self.v2_setting(|setting| {
-            setting.commit_message_model = Some(LanguageModelSelection { provider, model });
+            setting.commit_message_model = Some(LanguageModelSelection {
+                provider: provider.into(),
+                model,
+            });
             Ok(())
         })
         .ok();
@@ -427,7 +487,10 @@ impl AssistantSettingsContent {
 
     pub fn set_thread_summary_model(&mut self, provider: String, model: String) {
         self.v2_setting(|setting| {
-            setting.thread_summary_model = Some(LanguageModelSelection { provider, model });
+            setting.thread_summary_model = Some(LanguageModelSelection {
+                provider: provider.into(),
+                model,
+            });
             Ok(())
         })
         .ok();
@@ -523,6 +586,7 @@ impl Default for VersionedAssistantSettingsContent {
             notify_when_agent_waiting: None,
             stream_edits: None,
             single_file_review: None,
+            model_parameters: Vec::new(),
             preferred_completion_mode: None,
         })
     }
@@ -587,6 +651,15 @@ pub struct AssistantSettingsContentV2 {
     ///
     /// Default: true
     single_file_review: Option<bool>,
+    /// Additional parameters for language model requests. When making a request
+    /// to a model, parameters will be taken from the last entry in this list
+    /// that matches the model's provider and name. In each entry, both provider
+    /// and model are optional, so that you can specify parameters for either
+    /// one.
+    ///
+    /// Default: []
+    #[serde(default)]
+    model_parameters: Vec<LanguageModelParameters>,
 
     /// What completion mode to enable for new threads
     ///
@@ -613,33 +686,53 @@ impl From<CompletionMode> for zed_llm_client::CompletionMode {
 
 #[derive(Clone, Debug, Serialize, Deserialize, JsonSchema, PartialEq)]
 pub struct LanguageModelSelection {
-    #[schemars(schema_with = "providers_schema")]
+    pub provider: LanguageModelProviderSetting,
-    pub provider: String,
     pub model: String,
 }
 
-fn providers_schema(_: &mut schemars::r#gen::SchemaGenerator) -> schemars::schema::Schema {
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
-    schemars::schema::SchemaObject {
+pub struct LanguageModelProviderSetting(pub String);
-        enum_values: Some(vec![
+
-            "anthropic".into(),
+impl JsonSchema for LanguageModelProviderSetting {
-            "bedrock".into(),
+    fn schema_name() -> String {
-            "google".into(),
+        "LanguageModelProviderSetting".into()
-            "lmstudio".into(),
+    }
-            "ollama".into(),
+
-            "openai".into(),
+    fn json_schema(_: &mut schemars::r#gen::SchemaGenerator) -> Schema {
-            "zed.dev".into(),
+        schemars::schema::SchemaObject {
-            "copilot_chat".into(),
+            enum_values: Some(vec![
-            "deepseek".into(),
+                "anthropic".into(),
-        ]),
+                "bedrock".into(),
-        ..Default::default()
+                "google".into(),
+                "lmstudio".into(),
+                "ollama".into(),
+                "openai".into(),
+                "zed.dev".into(),
+                "copilot_chat".into(),
+                "deepseek".into(),
+            ]),
+            ..Default::default()
+        }
+        .into()
+    }
+}
+
+impl From<String> for LanguageModelProviderSetting {
+    fn from(provider: String) -> Self {
+        Self(provider)
+    }
+}
+
+impl From<&str> for LanguageModelProviderSetting {
+    fn from(provider: &str) -> Self {
+        Self(provider.to_string())
     }
-    .into()
 }
 
 impl Default for LanguageModelSelection {
     fn default() -> Self {
         Self {
-            provider: "openai".to_string(),
+            provider: LanguageModelProviderSetting("openai".to_string()),
             model: "gpt-4".to_string(),
         }
     }
@@ -781,6 +874,10 @@ impl Settings for AssistantSettings {
                 value.preferred_completion_mode,
             );
 
+            settings
+                .model_parameters
+                .extend_from_slice(&value.model_parameters);
+
             if let Some(profiles) = value.profiles {
                 settings
                     .profiles
@@ -913,6 +1010,7 @@ mod tests {
                                 notify_when_agent_waiting: None,
                                 stream_edits: None,
                                 single_file_review: None,
+                                model_parameters: Vec::new(),
                                 preferred_completion_mode: None,
                             },
                         )),
@@ -976,7 +1074,7 @@ mod tests {
                         AssistantSettingsContentV2 {
                             enabled: Some(false),
                             default_model: Some(LanguageModelSelection {
-                                provider: "xai".to_owned(),
+                                provider: "xai".to_owned().into(),
                                 model: "grok".to_owned(),
                             }),
                             ..Default::default()

crates/collab/Cargo.toml

@@ -78,6 +78,7 @@ zed_llm_client.workspace = true
 [dev-dependencies]
 assistant = { workspace = true, features = ["test-support"] }
 assistant_context_editor.workspace = true
+assistant_settings.workspace = true
 assistant_slash_command.workspace = true
 assistant_tool.workspace = true
 async-trait.workspace = true

crates/collab/src/tests/test_server.rs

@@ -307,6 +307,7 @@ impl TestServer {
             );
             language_model::LanguageModelRegistry::test(cx);
             assistant_context_editor::init(client.clone(), cx);
+            assistant_settings::init(cx);
         });
 
         client

crates/git_ui/src/git_panel.rs

@@ -1735,6 +1735,8 @@ impl GitPanel {
             }
         });
 
+        let temperature = AssistantSettings::temperature_for_model(&model, cx);
+
         self.generate_commit_message_task = Some(cx.spawn(async move |this, cx| {
              async move {
                 let _defer = cx.on_drop(&this, |this, _cx| {
@@ -1773,7 +1775,7 @@ impl GitPanel {
                     }],
                     tools: Vec::new(),
                     stop: Vec::new(),
-                    temperature: None,
+                    temperature,
                 };
 
                 let stream = model.stream_completion_text(request, &cx);

crates/language_models/src/settings.rs

@@ -87,8 +87,8 @@ pub struct AllLanguageModelSettingsContent {
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq, JsonSchema)]
 #[serde(untagged)]
 pub enum AnthropicSettingsContent {
-    Legacy(LegacyAnthropicSettingsContent),
     Versioned(VersionedAnthropicSettingsContent),
+    Legacy(LegacyAnthropicSettingsContent),
 }
 
 impl AnthropicSettingsContent {
@@ -197,8 +197,8 @@ pub struct MistralSettingsContent {
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq, JsonSchema)]
 #[serde(untagged)]
 pub enum OpenAiSettingsContent {
-    Legacy(LegacyOpenAiSettingsContent),
     Versioned(VersionedOpenAiSettingsContent),
+    Legacy(LegacyOpenAiSettingsContent),
 }
 
 impl OpenAiSettingsContent {

crates/project/src/lsp_store.rs

@@ -3522,7 +3522,7 @@ impl LspStore {
             )
             .detach();
         } else {
-            log::info!("No extension events global found. Skipping JSON schema auto-reload setup");
+            log::debug!("No extension events global found. Skipping JSON schema auto-reload setup");
         }
         cx.observe_global::<SettingsStore>(Self::on_settings_changed)
             .detach();

crates/worktree/src/worktree.rs

@@ -3871,7 +3871,7 @@ impl BackgroundScanner {
             Some(ancestor_dot_git)
         });
 
-        log::info!("containing git repository: {containing_git_repository:?}");
+        log::trace!("containing git repository: {containing_git_repository:?}");
 
         let (scan_job_tx, scan_job_rx) = channel::unbounded();
         {

docs/src/configuring-zed.md

@@ -3058,14 +3058,14 @@ Run the `theme selector: toggle` action in the command palette to see a current
 }
 ```
 
-## Assistant Panel
+## Agent
 
-- Description: Customize assistant panel
+- Description: Customize agent behavior
-- Setting: `assistant`
+- Setting: `agent`
 - Default:
 
 ```json
-"assistant": {
+"agent": {
   "version": "2",
   "enabled": true,
   "button": true,