Code at the speed of thought – Zed is a high-performance, multiplayer code editor from the creators of Atom and Tree-sitter.
language_model: Make LanguageModel plain data served by its provider (#64750)
Mikayla Maki975845875bb2
diff --git a/Cargo.lock b/Cargo.lockindex f2f375f5ceb..edff55c3997 100644--- a/Cargo.lock+++ b/Cargo.lock@@ -9729,7 +9729,6 @@ dependencies = ["ai_onboarding","anthropic","anyhow",- "async-lock","aws-config","aws-credential-types","aws-sigv4",@@ -9781,7 +9780,6 @@ dependencies = ["settings","sha2","strum 0.28.0",- "tokio","ui","ui_input","util",diff --git a/crates/agent/src/agent.rs b/crates/agent/src/agent.rsindex 89dc13d6bea..ebede3a87d9 100644--- a/crates/agent/src/agent.rs+++ b/crates/agent/src/agent.rs@@ -254,7 +254,7 @@ struct PendingSession {pub struct LanguageModels {/// Access language model by ID- models: HashMap<AgentModelId, Arc<dyn LanguageModel>>,+ models: HashMap<AgentModelId, LanguageModel>,/// Cached list for returning language model informationmodel_list: acp_thread::AgentModelList,refresh_models_rx: watch::Receiver<()>,@@ -332,12 +332,12 @@ impl LanguageModels {self.refresh_models_tx.send(()).ok();}- pub fn model_from_id(&self, model_id: &AgentModelId) -> Option<Arc<dyn LanguageModel>> {+ pub fn model_from_id(&self, model_id: &AgentModelId) -> Option<LanguageModel> {self.models.get(model_id).cloned()}fn map_language_model_to_info(- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,provider: &Arc<dyn LanguageModelProvider>,) -> acp_thread::AgentModelInfo {acp_thread::AgentModelInfo {@@ -354,7 +354,7 @@ impl LanguageModels {}}- fn model_id(model: &Arc<dyn LanguageModel>) -> AgentModelId {+ fn model_id(model: &LanguageModel) -> AgentModelId {AgentModelId::new(format!("{}/{}", model.provider_id().0, model.id().0))}@@ -771,7 +771,7 @@ impl NativeAgent {let default_model = registry.default_model().and_then(|default_model| {self.models- .model_from_id(&LanguageModels::model_id(&default_model.model))+ .model_from_id(&LanguageModels::model_id(&default_model))});let thread = cx.new(|cx| {Thread::new(@@ -824,7 +824,7 @@ impl NativeAgent {});let registry = LanguageModelRegistry::read_global(cx);- let summarization_model = registry.thread_summary_model(cx).map(|c| c.model);+ let summarization_model = registry.thread_summary_model(cx);let weak = cx.weak_entity();let weak_thread = thread_handle.downgrade();@@ -1444,12 +1444,15 @@ impl NativeAgent {self.models.refresh_list(cx);let registry = LanguageModelRegistry::read_global(cx);- let default_model = registry.default_model().map(|m| m.model);- let summarization_model = registry.thread_summary_model(cx).map(|m| m.model);+ let default_model = registry.default_model();+ let summarization_model = registry.thread_summary_model(cx);for session in self.sessions.values_mut() {session.thread.update(cx, |thread, cx| {thread.ensure_model(default_model.as_ref(), cx);+ if let language_model::Event::ProviderStateChanged(provider_id) = event {+ thread.refresh_model(provider_id, cx);+ }if let Some(model) = summarization_model.clone() {if thread.summarization_model().is_none()@@ -1655,9 +1658,8 @@ impl NativeAgent {.projects.get(&project_id).context("project state not found")?;- let summarization_model = LanguageModelRegistry::read_global(cx)- .thread_summary_model(cx)- .map(|c| c.model);+ let summarization_model =+ LanguageModelRegistry::read_global(cx).thread_summary_model(cx);Ok(cx.new(|cx| {let mut thread = Thread::from_db(@@ -2746,7 +2748,7 @@ pub fn available_native_agent(cx: &App) -> AvailableAgent {for model in provider.provided_models(cx) {let id = format!("{}/{}", provider_id.0, model.id().0);let is_default = default.as_ref().is_some_and(|default| {- default.provider.id() == provider_id && default.model.id() == model.id()+ default.provider_id == provider_id && default.id == model.id});models.push(AvailableModel {id,@@ -3922,10 +3924,10 @@ mod internal_tests {use fs::FakeFs;use gpui::TestAppContext;use indoc::formatdoc;- use language_model::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};+ use language_model::fake_provider::FakeLanguageModelProvider;use language_model::{CompletionIntent, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelProviderId, LanguageModelProviderName, Speed,+ LanguageModelName, LanguageModelProviderId, LanguageModelProviderName, Speed,};use serde_json::json;use settings::{LanguageModelProviderSetting, SettingsStore};@@ -3957,19 +3959,14 @@ mod internal_tests {let thread_store = cx.new(|cx| ThreadStore::new(cx));let agent = cx.update(|cx| NativeAgent::new(thread_store, Templates::new(), fs, cx));let connection = Rc::new(NativeAgentConnection(agent.clone()));- let model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "subagent-model",- "Subagent Model",- true,+ let provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![model.clone()]),- );+ let model = provider.update_model("subagent-model", |model| {+ model.name = LanguageModelName::from("Subagent Model".to_string());+ model.supports_thinking = true;+ });cx.update(|cx| {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(provider, cx);@@ -4230,11 +4227,11 @@ mod internal_tests {AcpBetaFeatureFlag, FeatureFlag as _, FeatureFlagAppExt as _, FeatureFlagsSettings,};- init_test(cx);+ let fake = init_test(cx);let (connection, agent, project, acp_thread) = setup_native_agent_session(cx).await;let session_id = cx.update(|cx| acp_thread.read(cx).session_id().clone());let thread = cx.update(|cx| native_thread_for_session(&agent, &session_id, cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let old_message_id = ClientUserMessageId::new();cx.update(|cx| {@@ -4273,7 +4270,7 @@ mod internal_tests {});cx.run_until_parked();- let request = model.pending_completions().pop().unwrap();+ let request = fake.pending_completions().pop().unwrap();assert_eq!(request.intent,Some(CompletionIntent::ThreadContextSummarization)@@ -4299,7 +4296,7 @@ mod internal_tests {compaction.id.clone()});- model.send_completion_stream_text_chunk(&request, "retained ");+ fake.send_text(&model, &request, "retained ");cx.run_until_parked();let summary = acp_thread.read_with(cx, |thread, cx| {let Some(acp_thread::AgentThreadEntry::ContextCompaction(compaction)) =@@ -4318,8 +4315,8 @@ mod internal_tests {assert_eq!(markdown.read(cx).source().as_ref(), "retained ");markdown.clone()});- model.send_completion_stream_text_chunk(&request, "context");- model.end_completion_stream(&request);+ fake.send_text(&model, &request, "context");+ fake.end_stream(&model, &request);cx.run_until_parked();prompt_task.await@@ -4454,12 +4451,11 @@ mod internal_tests {#[gpui::test]async fn test_retried_auto_compaction_terminalizes_each_attempt(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let (connection, agent, _project, acp_thread) = setup_native_agent_session(cx).await;let session_id = cx.update(|cx| acp_thread.read(cx).session_id().clone());let thread = cx.update(|cx| native_thread_for_session(&agent, &session_id, cx));- let model = Arc::new(FakeLanguageModel::default());- model.set_max_token_count(1_000_000);+ let model = fake.model("fake");cx.update(|cx| {let mut settings = agent_settings::AgentSettings::get_global(cx).clone();@@ -4481,19 +4477,20 @@ mod internal_tests {});cx.run_until_parked();- let first_request = model+ let first_request = fake.pending_completions().pop().expect("first user prompt should start a model request");- model.send_completion_stream_event(+ fake.send_event(+ &model,&first_request,LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 750_000,..Default::default()}),);- model.send_completion_stream_text_chunk(&first_request, "old assistant");- model.end_completion_stream(&first_request);+ fake.send_text(&model, &first_request, "old assistant");+ fake.end_stream(&model, &first_request);cx.run_until_parked();first_prompt.await@@ -4509,7 +4506,7 @@ mod internal_tests {});cx.run_until_parked();- let first_compaction_request = model+ let first_compaction_request = fake.pending_completions().pop().expect("token threshold should start automatic compaction");@@ -4517,7 +4514,8 @@ mod internal_tests {first_compaction_request.intent,Some(CompletionIntent::ThreadContextSummarization));- model.send_completion_stream_error(+ fake.send_error(+ &model,&first_compaction_request,LanguageModelCompletionError::from_provider_response(language_model::ANTHROPIC_PROVIDER_NAME,@@ -4528,10 +4526,10 @@ mod internal_tests {language_model::ProviderErrorCategory::RateLimit,),);- model.end_completion_stream(&first_compaction_request);+ fake.end_stream(&model, &first_compaction_request);cx.run_until_parked();- let second_compaction_request = model+ let second_compaction_request = fake.pending_completions().pop().expect("retryable error should start another compaction attempt");@@ -4560,8 +4558,8 @@ mod internal_tests {assert!(thread.is_compacting());});- model.send_completion_stream_text_chunk(&second_compaction_request, "summary");- model.end_completion_stream(&second_compaction_request);+ fake.send_text(&model, &second_compaction_request, "summary");+ fake.end_stream(&model, &second_compaction_request);cx.run_until_parked();acp_thread.read_with(cx, |thread, _cx| {@@ -4588,13 +4586,13 @@ mod internal_tests {);});- let final_request = model+ let final_request = fake.pending_completions().pop().expect("successful compaction should continue the user turn");assert_eq!(final_request.intent, Some(CompletionIntent::UserPrompt));- model.send_completion_stream_text_chunk(&final_request, "new assistant");- model.end_completion_stream(&final_request);+ fake.send_text(&model, &final_request, "new assistant");+ fake.end_stream(&model, &final_request);cx.run_until_parked();second_prompt.await@@ -4603,7 +4601,7 @@ mod internal_tests {#[gpui::test]async fn test_native_compaction_cancellation_is_bridged_before_stop(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);for (scenario, cancel_before_first_poll, partial_summary) in [("before first poll", true, None),@@ -4617,7 +4615,7 @@ mod internal_tests {let (connection, agent, project, acp_thread) = setup_native_agent_session(cx).await;let session_id = acp_thread.read_with(cx, |thread, _| thread.session_id().clone());let thread = cx.update(|cx| native_thread_for_session(&agent, &session_id, cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");cx.update(|cx| {let path_style = project.read(cx).path_style(cx);@@ -4649,12 +4647,12 @@ mod internal_tests {.expect("manual compaction should start");cx.run_until_parked();- let request = model+ let request = fake.pending_completions().pop().expect("manual compaction should reach the model");if let Some(partial_summary) = partial_summary {- model.send_completion_stream_text_chunk(&request, partial_summary);+ fake.send_text(&model, &request, partial_summary);cx.run_until_parked();}@@ -4723,11 +4721,11 @@ mod internal_tests {async fn test_native_tool_names_survive_reused_ids_and_reload(cx: &mut TestAppContext) {use language_model::{LanguageModelToolUse, LanguageModelToolUseInput};- init_test(cx);+ let fake = init_test(cx);let (connection, agent, project, acp_thread) = setup_native_agent_session(cx).await;let session_id = acp_thread.read_with(cx, |thread, _| thread.session_id().clone());let thread = cx.update(|cx| native_thread_for_session(&agent, &session_id, cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("thread");let fs = project.read_with(cx, |project, _| project.fs().clone());fs.write(Path::new("/a/file.txt"), b"retained file contents").await@@ -4750,12 +4748,13 @@ mod internal_tests {)});cx.run_until_parked();- let request = model- .pending_completions()+ let request = fake+ .pending_completions_for(&model).pop().expect("user prompt should reach the model");let input = json!({"path": "a/file.txt"});- model.send_completion_stream_event(+ fake.send_event(+ &model,&request,LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "reused_id".into(),@@ -4766,15 +4765,15 @@ mod internal_tests {thought_signature: None,}),);- model.end_completion_stream(&request);+ fake.end_stream(&model, &request);cx.run_until_parked();- let request = model- .pending_completions()+ let request = fake+ .pending_completions_for(&model).pop().expect("tool result should reach the model");assert_eq!(request.intent, Some(CompletionIntent::ToolResults));- model.send_completion_stream_text_chunk(&request, "done");- model.end_completion_stream(&request);+ fake.send_text(&model, &request, "done");+ fake.end_stream(&model, &request);cx.run_until_parked();prompt_task.await.expect("native tool turn should complete");}@@ -6653,19 +6652,14 @@ mod internal_tests {// Register a thinking model and select it.cx.update(|cx| {- let thinking_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "fake-thinking",- "Fake Thinking",- true,+ let thinking_provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let thinking_provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![thinking_model]),- );+ thinking_provider.update_model("fake-thinking", |model| {+ model.name = LanguageModelName::from("Fake Thinking".to_string());+ model.supports_thinking = true;+ });LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(thinking_provider, cx);});@@ -6717,19 +6711,14 @@ mod internal_tests {// Register a second provider with a thinking model.cx.update(|cx| {- let thinking_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "fake-thinking",- "Fake Thinking",- true,+ let thinking_provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let thinking_provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![thinking_model]),- );+ thinking_provider.update_model("fake-thinking", |model| {+ model.name = LanguageModelName::from("Fake Thinking".to_string());+ model.supports_thinking = true;+ });LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(thinking_provider, cx);});@@ -6847,22 +6836,17 @@ mod internal_tests {let connection = Rc::new(NativeAgentConnection(agent.clone()));// Register a thinking model.- let thinking_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "fake-thinking",- "Fake Thinking",- true,+ let thinking_provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let thinking_provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![thinking_model.clone()]),- );+ let thinking_model = thinking_provider.update_model("fake-thinking", |model| {+ model.name = LanguageModelName::from("Fake Thinking".to_string());+ model.supports_thinking = true;+ });cx.update(|cx| {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {- registry.register_provider(thinking_provider, cx);+ registry.register_provider(thinking_provider.clone(), cx);});});agent.update(cx, |agent, cx| agent.models.refresh_list(cx));@@ -6901,8 +6885,8 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- thinking_model.send_last_completion_stream_text_chunk("Response.");- thinking_model.end_last_completion_stream();+ thinking_provider.send_last_text(&thinking_model, "Response.");+ thinking_provider.end_last(&thinking_model);send.await.unwrap();cx.run_until_parked();@@ -6948,22 +6932,16 @@ mod internal_tests {// Register a model where id() != name(), like real Anthropic models// (e.g. id="claude-sonnet-4-5-thinking-latest", name="Claude Sonnet 4.5 Thinking").- let model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "custom-model-id",- "Custom Model Display Name",- false,+ let provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![model.clone()]),- );+ let model = provider.update_model("custom-model-id", |model| {+ model.name = LanguageModelName::from("Custom Model Display Name".to_string());+ });cx.update(|cx| {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {- registry.register_provider(provider, cx);+ registry.register_provider(provider.clone(), cx);});});agent.update(cx, |agent, cx| agent.models.refresh_list(cx));@@ -7002,8 +6980,8 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("Response.");- model.end_last_completion_stream();+ provider.send_last_text(&model, "Response.");+ provider.end_last(&model);send.await.unwrap();cx.run_until_parked();@@ -7057,19 +7035,13 @@ mod internal_tests {.update(|cx| NativeAgent::new(thread_store.clone(), Templates::new(), fs.clone(), cx));let connection = Rc::new(NativeAgentConnection(agent.clone()));- let model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "custom-model-id",- "Custom Model Display Name",- false,+ let provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));- let provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![model.clone()]),- );+ let model = provider.update_model("custom-model-id", |model| {+ model.name = LanguageModelName::from("Custom Model Display Name".to_string());+ });cx.update(|cx| {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(provider.clone(), cx);@@ -7097,8 +7069,8 @@ mod internal_tests {let send = acp_thread.update(cx, |thread, cx| thread.send(vec!["Hello".into()], cx));let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("Response.");- model.end_last_completion_stream();+ provider.send_last_text(&model, "Response.");+ provider.end_last(&model);send.await.unwrap();cx.run_until_parked();@@ -7196,19 +7168,13 @@ mod internal_tests {});// The user explicitly picks a different, available model.- let other_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "other-corp",- "other-model-id",- "Other Model",- false,+ let other_provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("other-corp".to_string()),+ LanguageModelProviderName::from("Other Corp".to_string()),));- let other_provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("other-corp".to_string()),- LanguageModelProviderName::from("Other Corp".to_string()),- )- .with_models(vec![other_model.clone()]),- );+ other_provider.update_model("other-model-id", |model| {+ model.name = LanguageModelName::from("Other Model".to_string());+ });cx.update(|cx| {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(other_provider, cx);@@ -7246,7 +7212,7 @@ mod internal_tests {#[gpui::test]async fn test_save_load_thread(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree("/",@@ -7277,8 +7243,8 @@ mod internal_tests {});// Ensure empty threads are not saved, even if they get mutated.- let model = Arc::new(FakeLanguageModel::default());- let summary_model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");+ let summary_model = fake.model("summary");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);thread.set_summarization_model(Some(summary_model.clone()), cx);@@ -7306,19 +7272,26 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("Lorem.");- model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "Lorem.");+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 150,output_tokens: 75,..Default::default()- },- ));- model.end_last_completion_stream();+ }),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();- summary_model- .send_last_completion_stream_text_chunk(&format!("Explaining {}", path!("/a/b.md")));- summary_model.end_last_completion_stream();+ let summary_request = fake.pending_completions_for(&summary_model).pop().unwrap();+ fake.send_text(+ &summary_model,+ &summary_request,+ format!("Explaining {}", path!("/a/b.md")),+ );+ fake.end_stream(&summary_model, &summary_request);send.await.unwrap();let uri = MentionUri::File {@@ -7425,7 +7398,7 @@ mod internal_tests {#[gpui::test]async fn test_releasing_session_saves_thread(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree("/",@@ -7455,7 +7428,7 @@ mod internal_tests {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);});@@ -7465,8 +7438,8 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("world");- model.end_last_completion_stream();+ fake.send_last_text(&model, "world");+ fake.end_last(&model);send.await.unwrap();cx.run_until_parked();@@ -7589,7 +7562,7 @@ mod internal_tests {#[gpui::test]async fn test_save_burst_preserves_in_flight_write(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree("/",@@ -7619,7 +7592,7 @@ mod internal_tests {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);});@@ -7628,8 +7601,8 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("world");- model.end_last_completion_stream();+ fake.send_last_text(&model, "world");+ fake.end_last(&model);send.await.unwrap();cx.run_until_parked();@@ -7680,7 +7653,7 @@ mod internal_tests {#[gpui::test]async fn test_thread_summary_releases_loaded_session(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree("/",@@ -7710,8 +7683,8 @@ mod internal_tests {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());- let summary_model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");+ let summary_model = fake.model("summary");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);thread.set_summarization_model(Some(summary_model.clone()), cx);@@ -7721,8 +7694,9 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- model.send_last_completion_stream_text_chunk("world");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "world");+ fake.end_stream(&model, &request);send.await.unwrap();cx.run_until_parked();@@ -7731,8 +7705,9 @@ mod internal_tests {});cx.run_until_parked();- summary_model.send_last_completion_stream_text_chunk("summary");- summary_model.end_last_completion_stream();+ let summary_request = fake.pending_completions_for(&summary_model).pop().unwrap();+ fake.send_text(&summary_model, &summary_request, "summary");+ fake.end_stream(&summary_model, &summary_request);assert_eq!(summary.await.unwrap(), "summary");cx.run_until_parked();@@ -7758,7 +7733,7 @@ mod internal_tests {#[gpui::test]async fn test_loaded_sessions_keep_state_until_last_close(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree("/",@@ -7788,13 +7763,7 @@ mod internal_tests {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = cx.update(|cx| {- LanguageModelRegistry::read_global(cx)- .default_model()- .map(|default_model| default_model.model)- .expect("default test model should be available")- });- let fake_model = model.as_fake();+ let model = fake.model("fake");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);});@@ -7803,8 +7772,8 @@ mod internal_tests {let send = cx.foreground_executor().spawn(send);cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("world");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "world");+ fake.end_last(&model);send.await.unwrap();cx.run_until_parked();@@ -7861,8 +7830,8 @@ mod internal_tests {let follow_up = cx.foreground_executor().spawn(follow_up);cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("yes");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "yes");+ fake.end_last(&model);follow_up.await.unwrap();cx.run_until_parked();@@ -7973,14 +7942,15 @@ mod internal_tests {})}- fn init_test(cx: &mut TestAppContext) {+ /// Returns the fake provider that serves the registry's default model.+ fn init_test(cx: &mut TestAppContext) -> Arc<FakeLanguageModelProvider> {env_logger::try_init().ok();cx.update(|cx| {let settings_store = SettingsStore::test(cx);cx.set_global(settings_store);- LanguageModelRegistry::test(cx);- });+ LanguageModelRegistry::test(cx)+ })}#[test]diff --git a/crates/agent/src/tests/mod.rs b/crates/agent/src/tests/mod.rsindex c93e11d031f..4cf9dc4e2f1 100644--- a/crates/agent/src/tests/mod.rs+++ b/crates/agent/src/tests/mod.rs@@ -25,12 +25,11 @@ use gpui::{};use indoc::indoc;use language_model::{- CompletionIntent, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelId, LanguageModelImageExt, LanguageModelProviderId, LanguageModelProviderName,+ CompletionIntent, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,+ LanguageModelImageExt, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,LanguageModelRegistry, LanguageModelRequest, LanguageModelRequestMessage,LanguageModelToolResult, LanguageModelToolUse, MessageContent, ProviderErrorCategory, Role,- StopReason, TokenUsage,- fake_provider::{FakeLanguageModel, FakeLanguageModelProvider},+ StopReason, TokenUsage, fake_provider::FakeLanguageModelProvider,};use pretty_assertions::assert_eq;use project::{@@ -57,11 +56,13 @@ use util::path;mod test_tools;use test_tools::*;-pub(crate) fn init_test(cx: &mut TestAppContext) {+/// Returns the fake provider that serves the registry's default model.+pub(crate) fn init_test(cx: &mut TestAppContext) -> Arc<FakeLanguageModelProvider> {cx.update(|cx| {let settings_store = SettingsStore::test(cx);cx.set_global(settings_store);- });+ LanguageModelRegistry::test(cx)+ })}pub(crate) fn release_dropped_entities(cx: &mut TestAppContext) {@@ -324,8 +325,12 @@ fn disable_sandboxing(cx: &mut TestAppContext) {#[gpui::test]async fn test_echo(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let events = thread.update(cx, |thread, cx| {@@ -337,10 +342,12 @@ async fn test_echo(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hello");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hello");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events = events.collect().await;thread.update(cx, |thread, _cx| {@@ -479,8 +486,12 @@ async fn test_terminal_tool_without_timeout_does_not_kill_handle(cx: &mut TestAp#[gpui::test]async fn test_thinking(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let events = thread.update(cx, |thread, cx| {@@ -497,14 +508,19 @@ async fn test_thinking(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Thinking {- text: "Think".to_string(),- signature: None,- });- fake_model.send_last_completion_stream_text_chunk("Hello");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Thinking {+ text: "Think".to_string(),+ signature: None,+ },+ );+ fake.send_last_text(&model, "Hello");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events = events.collect().await;thread.update(cx, |thread, _cx| {@@ -528,13 +544,18 @@ async fn test_thinking(cx: &mut TestAppContext) {#[gpui::test]async fn test_thinking_allowed_when_model_cannot_disable_thinking(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();- fake_model.set_supports_thinking(true);+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;+ fake.update_model("fake", |model| model.supports_thinking = true);// With thinking toggled off, a model that can disable thinking honors// the toggle...thread.update(cx, |thread, cx| {+ thread.refresh_model(&model.provider_id, cx);thread.set_thinking_enabled(false, cx);let request = thread.build_completion_request(CompletionIntent::UserPrompt, cx)@@ -543,8 +564,9 @@ async fn test_thinking_allowed_when_model_cannot_disable_thinking(cx: &mut TestA});// ...but a model that always thinks ignores the stale toggle state.- fake_model.set_supports_disabling_thinking(false);+ fake.update_model("fake", |model| model.supports_disabling_thinking = false);thread.update(cx, |thread, cx| {+ thread.refresh_model(&model.provider_id, cx);let request = thread.build_completion_request(CompletionIntent::UserPrompt, cx).unwrap();@@ -555,12 +577,11 @@ async fn test_thinking_allowed_when_model_cannot_disable_thinking(cx: &mut TestA#[gpui::test]async fn test_system_prompt(cx: &mut TestAppContext) {let ThreadTest {- model,+ fake,thread,project_context,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();project_context.update(cx, |project_context, _cx| {project_context.shell = "test-shell".into()@@ -572,7 +593,7 @@ async fn test_system_prompt(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- let mut pending_completions = fake_model.pending_completions();+ let mut pending_completions = fake.pending_completions();assert_eq!(pending_completions.len(),1,@@ -601,8 +622,7 @@ async fn test_system_prompt(cx: &mut TestAppContext) {#[gpui::test]async fn test_system_prompt_without_tools(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest { fake, thread, .. } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, cx| {@@ -610,7 +630,7 @@ async fn test_system_prompt_without_tools(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- let mut pending_completions = fake_model.pending_completions();+ let mut pending_completions = fake.pending_completions();assert_eq!(pending_completions.len(),1,@@ -639,8 +659,12 @@ async fn test_system_prompt_without_tools(cx: &mut TestAppContext) {#[gpui::test]async fn test_prompt_caching(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;// Send initial user message and verify it's cachedthread@@ -650,7 +674,7 @@ async fn test_prompt_caching(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(completion.messages[1..],vec![LanguageModelRequestMessage {@@ -660,10 +684,11 @@ async fn test_prompt_caching(cx: &mut TestAppContext) {reasoning_details: None,}]);- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Text(- "Response to Message 1".into(),- ));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Text("Response to Message 1".into()),+ );+ fake.end_last(&model);cx.run_until_parked();// Send another user message and verify only the latest is cached@@ -674,7 +699,7 @@ async fn test_prompt_caching(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(completion.messages[1..],vec![@@ -698,10 +723,11 @@ async fn test_prompt_caching(cx: &mut TestAppContext) {}]);- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::Text(- "Response to Message 2".into(),- ));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Text("Response to Message 2".into()),+ );+ fake.end_last(&model);cx.run_until_parked();// Simulate a tool call and verify that the latest tool result is cached@@ -721,12 +747,14 @@ async fn test_prompt_caching(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(tool_use.clone()),+ );+ fake.end_last(&model);cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_result = LanguageModelToolResult {tool_use_id: "tool_1".into(),tool_name: EchoTool::NAME.into(),@@ -905,8 +933,12 @@ async fn test_streaming_tool_calls(cx: &mut TestAppContext) {#[gpui::test]async fn test_tool_authorization(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -915,37 +947,40 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_1".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_2".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_interrupted".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let tool_call_auth_1 = next_tool_call_authorization(&mut events).await;let tool_call_auth_2 = next_tool_call_authorization(&mut events).await;let interrupted_tool_call_auth = next_tool_call_authorization(&mut events).await;@@ -977,7 +1012,7 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();assert_eq!(message.content,@@ -1007,17 +1042,18 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {);// Simulate yet another tool call.- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_3".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Respond by always allowing tools - send transformed option_id// (UI transforms "always:tool_requiring_permission" to "always_allow:tool_requiring_permission")@@ -1030,7 +1066,7 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {)).unwrap();cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();assert_eq!(message.content,@@ -1046,19 +1082,20 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {);// Simulate a final tool call, ensuring we don't trigger authorization.- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_4".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();assert_eq!(message.content,@@ -1076,8 +1113,12 @@ async fn test_tool_authorization(cx: &mut TestAppContext) {#[gpui::test]async fn test_tool_hallucination(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -1085,17 +1126,18 @@ async fn test_tool_hallucination(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_1".into(),name: "nonexistent_tool".into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let tool_call = expect_tool_call(&mut events).await;assert_eq!(tool_call.title, "nonexistent_tool");@@ -1111,8 +1153,12 @@ async fn test_tool_hallucination(cx: &mut TestAppContext) {/// one in `AcpThread::upsert_tool_call`.#[gpui::test]async fn test_tool_call_id_scoped_per_completion_request(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| {thread.add_tool(EchoTool);@@ -1125,39 +1171,43 @@ async fn test_tool_call_id_scoped_per_completion_request(cx: &mut TestAppContext.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "call_1".into(),name: EchoTool::NAME.into(),raw_input: json!({"text": "first"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "first"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let first_tool_call = next_tool_call(&mut events).await;cx.run_until_parked();// Same turn, second cycle: the id counter has reset, so "call_1" recurs// for a different tool call.- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "call_1".into(),name: EchoTool::NAME.into(),raw_input: json!({"text": "second"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "second"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let second_tool_call = next_tool_call(&mut events).await;cx.run_until_parked();- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);events.collect::<Vec<_>>().await;assert_ne!(@@ -1175,11 +1225,11 @@ async fn test_tool_call_id_scoped_per_completion_request(cx: &mut TestAppContextasync fn test_replayed_tool_call_ids_scoped_across_messages(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,project_context,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();thread.update(cx, |thread, _cx| {thread.add_tool(EchoTool);@@ -1191,20 +1241,21 @@ async fn test_replayed_tool_call_ids_scoped_across_messages(cx: &mut TestAppCont}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "call_1".into(),name: EchoTool::NAME.into(),raw_input: json!({"text": "first"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "first"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Done with first");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done with first");+ fake.end_last(&model);cx.run_until_parked();// Different turn: the id counter has reset, so "call_1" recurs in a@@ -1215,20 +1266,21 @@ async fn test_replayed_tool_call_ids_scoped_across_messages(cx: &mut TestAppCont}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "call_1".into(),name: EchoTool::NAME.into(),raw_input: json!({"text": "second"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "second"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Done with second");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done with second");+ fake.end_last(&model);cx.run_until_parked();let db_thread = thread.read_with(cx, |thread, cx| thread.to_db(cx)).await;@@ -1637,9 +1689,12 @@ async fn test_concurrent_tool_calls(cx: &mut TestAppContext) {#[gpui::test]async fn test_profiles(cx: &mut TestAppContext) {let ThreadTest {- model, thread, fs, ..+ model,+ fake,+ thread,+ fs,+ ..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();thread.update(cx, |thread, _cx| {thread.add_tool(DelayTool);@@ -1684,7 +1739,7 @@ async fn test_profiles(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let mut pending_completions = fake_model.pending_completions();+ let mut pending_completions = fake.pending_completions();assert_eq!(pending_completions.len(), 1);let completion = pending_completions.pop().unwrap();let tool_names: Vec<String> = completion@@ -1693,7 +1748,7 @@ async fn test_profiles(cx: &mut TestAppContext) {.map(|tool| tool.name.clone()).collect();assert_eq!(tool_names, vec![DelayTool::NAME, EchoTool::NAME]);- fake_model.end_last_completion_stream();+ fake.end_last(&model);// Switch to test-2 profile, and verify that it has only the infinite tool.thread@@ -1703,7 +1758,7 @@ async fn test_profiles(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- let mut pending_completions = fake_model.pending_completions();+ let mut pending_completions = fake.pending_completions();assert_eq!(pending_completions.len(), 1);let completion = pending_completions.pop().unwrap();let tool_names: Vec<String> = completion@@ -1718,12 +1773,12 @@ async fn test_profiles(cx: &mut TestAppContext) {async fn test_mcp_tools(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,context_server_store,fs,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();// Override profiles and wait for settings to be loaded.fs.insert_file(@@ -1773,19 +1828,20 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {cx.run_until_parked();// Simulate the model calling the MCP tool.- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(tool_names_for_completion(&completion), vec!["echo"]);- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "echo".into(),raw_input: json!({"text": "test"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "test"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let (tool_call_params, tool_call_response) = mcp_tool_calls.next().await.unwrap();@@ -1804,8 +1860,8 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {cx.run_until_parked();assert_eq!(tool_names_for_completion(&completion), vec!["echo"]);- fake_model.send_last_completion_stream_text_chunk("Done!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done!");+ fake.end_last(&model);events.collect::<Vec<_>>().await;// Send again after adding the echo tool, ensuring the name collision is resolved.@@ -1814,32 +1870,34 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {thread.send(ClientUserMessageId::new(), ["Go"], cx).unwrap()});cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(tool_names_for_completion(&completion),vec!["echo", "test_server_echo"]);- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_2".into(),name: "test_server_echo".into(),raw_input: json!({"text": "mcp"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "mcp"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_3".into(),name: "echo".into(),raw_input: json!({"text": "native"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "native"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let (tool_call_params, tool_call_response) = mcp_tool_calls.next().await.unwrap();@@ -1856,7 +1914,7 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {cx.run_until_parked();// Ensure the tool results were inserted with the correct names.- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(completion.messages.last().unwrap().content,vec![@@ -1876,7 +1934,7 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {},),]);- fake_model.end_last_completion_stream();+ fake.end_last(&model);events.collect::<Vec<_>>().await;}@@ -1884,12 +1942,12 @@ async fn test_mcp_tools(cx: &mut TestAppContext) {async fn test_mcp_tool_names_are_sanitized_for_providers(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,context_server_store,fs,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();fs.insert_file(paths::settings_file(),@@ -1934,22 +1992,23 @@ async fn test_mcp_tool_names_are_sanitized_for_providers(cx: &mut TestAppContext});cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(tool_names_for_completion(&completion),vec!["snake_case_PascalCase"]);- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "snake_case_PascalCase".into(),raw_input: json!({}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let (tool_call_params, tool_call_response) = mcp_tool_calls.next().await.unwrap();@@ -1966,8 +2025,8 @@ async fn test_mcp_tool_names_are_sanitized_for_providers(cx: &mut TestAppContext.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Done!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done!");+ fake.end_last(&model);events.collect::<Vec<_>>().await;}@@ -1975,13 +2034,16 @@ async fn test_mcp_tool_names_are_sanitized_for_providers(cx: &mut TestAppContextasync fn test_mcp_tool_multi_content_response(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,context_server_store,fs,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();- fake_model.set_supports_images(true);+ fake.update_model("fake", |model| model.supports_images = true);+ thread.update(cx, |thread, cx| {+ thread.refresh_model(&model.provider_id, cx)+ });fs.insert_file(paths::settings_file(),@@ -2027,18 +2089,19 @@ async fn test_mcp_tool_multi_content_response(cx: &mut TestAppContext) {});cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ let completion = fake.pending_completions().pop().unwrap();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "screenshot".into(),raw_input: json!({}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let _ = completion;@@ -2067,7 +2130,7 @@ async fn test_mcp_tool_multi_content_response(cx: &mut TestAppContext) {cx.run_until_parked();// Verify the tool result round-trips back to the model as a multi-part Vec.- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_result = completion.messages.last()@@ -2101,7 +2164,7 @@ async fn test_mcp_tool_multi_content_response(cx: &mut TestAppContext) {tool_result.content[2],language_model::LanguageModelToolResultContent::Text(Arc::from("Some more text")));- fake_model.end_last_completion_stream();+ fake.end_last(&model);events.collect::<Vec<_>>().await;}@@ -2109,12 +2172,12 @@ async fn test_mcp_tool_multi_content_response(cx: &mut TestAppContext) {async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,context_server_store,fs,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();// Setup settings to allow MCP toolsfs.insert_file(@@ -2169,7 +2232,7 @@ async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestApcx.run_until_parked();// Verify the MCP tool is available to the model- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(tool_names_for_completion(&completion),vec!["issue_read"],@@ -2177,8 +2240,9 @@ async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestAp);// Simulate the model calling the MCP tool- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "issue_read".into(),raw_input: json!({"issue_url": "https://github.com/zed-industries/zed/issues/47404"})@@ -2188,9 +2252,9 @@ async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestAp),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// The MCP server receives the tool call and responds with content@@ -2211,11 +2275,13 @@ async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestAp// After tool completes, the model continues with a new completion request// that includes the tool results. We need to respond to this.- let _completion = fake_model.pending_completions().pop().unwrap();- fake_model.send_last_completion_stream_text_chunk("I found the issue!");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ let _completion = fake.pending_completions().pop().unwrap();+ fake.send_last_text(&model, "I found the issue!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);events.collect::<Vec<_>>().await;// Verify the tool result is stored in the thread by checking the markdown output.@@ -2297,13 +2363,12 @@ async fn test_mcp_tool_result_displayed_when_server_disconnected(cx: &mut TestAp#[gpui::test]async fn test_mcp_tool_truncation(cx: &mut TestAppContext) {let ThreadTest {- model,+ fake,thread,context_server_store,fs,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();// Set up a profile with all tools enabledfs.insert_file(@@ -2457,7 +2522,7 @@ async fn test_mcp_tool_truncation(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(tool_names_for_completion(&completion),vec![@@ -2567,10 +2632,14 @@ async fn test_cancellation(cx: &mut TestAppContext) {#[gpui::test]async fn test_terminal_tool_cancellation_captures_output(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);disable_sandboxing(cx);- let fake_model = model.as_fake();let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))@@ -2590,8 +2659,9 @@ async fn test_terminal_tool_cancellation_captures_output(cx: &mut TestAppContextcx.run_until_parked();// Simulate the model calling the terminal tool- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_1".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 1000", "cd": "."}"#.into(),@@ -2600,9 +2670,9 @@ async fn test_terminal_tool_cancellation_captures_output(cx: &mut TestAppContext),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Wait for the terminal tool to start runningwait_for_terminal_tool_started(&mut events, cx).await;@@ -2659,7 +2729,7 @@ async fn test_terminal_tool_cancellation_captures_output(cx: &mut TestAppContext});// Verify we can send a new message after cancellation- verify_thread_recovery(&thread, &fake_model, cx).await;+ verify_thread_recovery(&thread, &fake, cx).await;}#[gpui::test]@@ -2667,9 +2737,13 @@ async fn test_cancellation_aware_tool_responds_to_cancellation(cx: &mut TestAppC// This test verifies that tools which properly handle cancellation via// `event_stream.cancelled_by_user()` (like edit_file_tool) respond promptly// to cancellation and report that they were cancelled.- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);- let fake_model = model.as_fake();let (tool, was_cancelled) = CancellationAwareTool::new();@@ -2687,17 +2761,18 @@ async fn test_cancellation_aware_tool_responds_to_cancellation(cx: &mut TestAppCcx.run_until_parked();// Simulate the model calling the cancellation-aware tool- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "cancellation_aware_1".into(),name: "cancellation_aware".into(),raw_input: r#"{}"#.into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();@@ -2754,13 +2829,13 @@ async fn test_cancellation_aware_tool_responds_to_cancellation(cx: &mut TestAppC);// Verify we can send a new message after cancellation- verify_thread_recovery(&thread, &fake_model, cx).await;+ verify_thread_recovery(&thread, &fake, cx).await;}/// Helper to verify thread can recover after cancellation by sending a simple message.async fn verify_thread_recovery(thread: &Entity<Thread>,- fake_model: &FakeLanguageModel,+ fake: &FakeLanguageModelProvider,cx: &mut TestAppContext,) {let events = thread@@ -2773,10 +2848,13 @@ async fn verify_thread_recovery(}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hello");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ let model = thread.read_with(cx, |thread, _| thread.model().cloned().unwrap());+ fake.send_last_text(&model, "Hello");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events = events.collect::<Vec<_>>().await;thread.update(cx, |thread, _cx| {@@ -2849,10 +2927,14 @@ async fn collect_events_until_stop(#[gpui::test]async fn test_truncate_while_terminal_tool_running(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);disable_sandboxing(cx);- let fake_model = model.as_fake();let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))@@ -2873,8 +2955,9 @@ async fn test_truncate_while_terminal_tool_running(cx: &mut TestAppContext) {cx.run_until_parked();// Simulate the model calling the terminal tool- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_1".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 1000", "cd": "."}"#.into(),@@ -2883,9 +2966,9 @@ async fn test_truncate_while_terminal_tool_running(cx: &mut TestAppContext) {),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Wait for the terminal tool to start runningwait_for_terminal_tool_started(&mut events, cx).await;@@ -2914,16 +2997,20 @@ async fn test_truncate_while_terminal_tool_running(cx: &mut TestAppContext) {});// Verify we can send a new message after truncation- verify_thread_recovery(&thread, &fake_model, cx).await;+ verify_thread_recovery(&thread, &fake, cx).await;}#[gpui::test]async fn test_cancel_multiple_concurrent_terminal_tools(cx: &mut TestAppContext) {// Tests that cancellation properly kills all running terminal tools when multiple are active.- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);disable_sandboxing(cx);- let fake_model = model.as_fake();let environment = Rc::new(MultiTerminalEnvironment::new());@@ -2940,8 +3027,9 @@ async fn test_cancel_multiple_concurrent_terminal_tools(cx: &mut TestAppContext)cx.run_until_parked();// Simulate the model calling two terminal tools- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_1".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 1000", "cd": "."}"#.into(),@@ -2950,10 +3038,11 @@ async fn test_cancel_multiple_concurrent_terminal_tools(cx: &mut TestAppContext)),is_input_complete: true,thought_signature: None,- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_2".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 2000", "cd": "."}"#.into(),@@ -2962,9 +3051,9 @@ async fn test_cancel_multiple_concurrent_terminal_tools(cx: &mut TestAppContext)),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Wait for both terminal tools to start by counting terminal content updateslet mut terminals_started = 0;@@ -3035,10 +3124,14 @@ async fn test_cancel_multiple_concurrent_terminal_tools(cx: &mut TestAppContext)async fn test_terminal_tool_stopped_via_terminal_card_button(cx: &mut TestAppContext) {// Tests that clicking the stop button on the terminal card (as opposed to the main// cancel button) properly reports user stopped via the was_stopped_by_user path.- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);disable_sandboxing(cx);- let fake_model = model.as_fake();let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))@@ -3058,8 +3151,9 @@ async fn test_terminal_tool_stopped_via_terminal_card_button(cx: &mut TestAppConcx.run_until_parked();// Simulate the model calling the terminal tool- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_1".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 1000", "cd": "."}"#.into(),@@ -3068,9 +3162,9 @@ async fn test_terminal_tool_stopped_via_terminal_card_button(cx: &mut TestAppCon),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Wait for the terminal tool to start runningwait_for_terminal_tool_started(&mut events, cx).await;@@ -3085,9 +3179,11 @@ async fn test_terminal_tool_stopped_via_terminal_card_button(cx: &mut TestAppConcx.run_until_parked();// The thread continues after tool completion - simulate the model ending its turn- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);// Collect remaining eventslet remaining_events = collect_events_until_stop(&mut events, cx).await;@@ -3129,10 +3225,14 @@ async fn test_terminal_tool_stopped_via_terminal_card_button(cx: &mut TestAppCon#[gpui::test]async fn test_terminal_tool_timeout_expires(cx: &mut TestAppContext) {// Tests that when a timeout is configured and expires, the tool result indicates timeout.- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;always_allow_tools(cx);disable_sandboxing(cx);- let fake_model = model.as_fake();let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))@@ -3156,8 +3256,9 @@ async fn test_terminal_tool_timeout_expires(cx: &mut TestAppContext) {cx.run_until_parked();// Simulate the model calling the terminal tool with a short timeout- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "terminal_tool_1".into(),name: TerminalTool::NAME.into(),raw_input: r#"{"command": "sleep 1000", "cd": ".", "timeout_ms": 100}"#.into(),@@ -3166,9 +3267,9 @@ async fn test_terminal_tool_timeout_expires(cx: &mut TestAppContext) {),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);// Wait for the terminal tool to start runningwait_for_terminal_tool_started(&mut events, cx).await;@@ -3178,9 +3279,11 @@ async fn test_terminal_tool_timeout_expires(cx: &mut TestAppContext) {cx.run_until_parked();// The thread continues after tool completion - simulate the model ending its turn- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);// Collect remaining eventslet remaining_events = collect_events_until_stop(&mut events, cx).await;@@ -3231,8 +3334,12 @@ async fn test_terminal_tool_timeout_expires(cx: &mut TestAppContext) {#[gpui::test]async fn test_in_progress_send_canceled_by_next_send(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let events_1 = thread.update(cx, |thread, cx| {@@ -3240,7 +3347,7 @@ async fn test_in_progress_send_canceled_by_next_send(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey 1!");+ fake.send_last_text(&model, "Hey 1!");cx.run_until_parked();let events_2 = thread@@ -3249,10 +3356,12 @@ async fn test_in_progress_send_canceled_by_next_send(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey 2!");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey 2!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events_1 = events_1.collect::<Vec<_>>().await;assert_eq!(stop_events(events_1), vec![acp::StopReason::Cancelled]);@@ -3266,8 +3375,12 @@ async fn test_retry_cancelled_promptly_on_new_send(cx: &mut TestAppContext) {// the retry loop waits on a timer. If the user switches models and sends a new message// during that delay, the old turn should exit immediately instead of retrying with the// stale model.- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let model_a = model.as_fake();+ let ThreadTest {+ model: model_a,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;// Start a turn with model_a.let events_1 = thread@@ -3276,26 +3389,29 @@ async fn test_retry_cancelled_promptly_on_new_send(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- assert_eq!(model_a.completion_count(), 1);+ assert_eq!(fake.pending_completions_for(&model_a).len(), 1);// Model returns a retryable upstream 500. The turn enters the retry delay.- model_a.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- language_model::LanguageModelProviderName::new("test"),- http_client::StatusCode::INTERNAL_SERVER_ERROR,- "Internal server error".to_string(),- None,- ));- model_a.end_last_completion_stream();+ fake.send_last_error(+ &model_a,+ LanguageModelCompletionError::from_http_status(+ language_model::LanguageModelProviderName::new("test"),+ http_client::StatusCode::INTERNAL_SERVER_ERROR,+ "Internal server error".to_string(),+ None,+ ),+ );+ fake.end_last(&model_a);cx.run_until_parked();// The old completion was consumed; model_a has no pending requests yet because the// retry timer hasn't fired.- assert_eq!(model_a.completion_count(), 0);+ assert_eq!(fake.pending_completions_for(&model_a).len(), 0);// Switch to model_b and send a new message. This cancels the old turn.- let model_b = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake", "model-b", "Model B", false,- ));+ let model_b = fake.update_model("model-b", |model| {+ model.name = LanguageModelName::from("Model B".to_string())+ });thread.update(cx, |thread, cx| {thread.set_model(model_b.clone(), cx);});@@ -3307,7 +3423,7 @@ async fn test_retry_cancelled_promptly_on_new_send(cx: &mut TestAppContext) {cx.run_until_parked();// model_b should have received its completion request.- assert_eq!(model_b.as_fake().completion_count(), 1);+ assert_eq!(fake.pending_completions_for(&model_b).len(), 1);// Advance the clock well past the retry delay (BASE_RETRY_DELAY = 5s).cx.executor().advance_clock(Duration::from_secs(10));@@ -3316,19 +3432,20 @@ async fn test_retry_cancelled_promptly_on_new_send(cx: &mut TestAppContext) {// model_a must NOT have received another completion request — the cancelled turn// should have exited during the retry delay rather than retrying with the old model.assert_eq!(- model_a.completion_count(),+ fake.pending_completions_for(&model_a).len(),0,"old model should not receive a retry request after cancellation");// Complete model_b's turn.- model_b- .as_fake()- .send_last_completion_stream_text_chunk("Done!");- model_b- .as_fake()- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- model_b.as_fake().end_last_completion_stream();+ let model_b_request = fake.pending_completions_for(&model_b).pop().unwrap();+ fake.send_text(&model_b, &model_b_request, "Done!");+ fake.send_event(+ &model_b,+ &model_b_request,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_stream(&model_b, &model_b_request);let events_1 = events_1.collect::<Vec<_>>().await;assert_eq!(stop_events(events_1), vec![acp::StopReason::Cancelled]);@@ -3339,8 +3456,12 @@ async fn test_retry_cancelled_promptly_on_new_send(cx: &mut TestAppContext) {#[gpui::test]async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let events_1 = thread.update(cx, |thread, cx| {@@ -3348,10 +3469,12 @@ async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey 1!");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey 1!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events_1 = events_1.collect::<Vec<_>>().await;let events_2 = thread@@ -3360,10 +3483,12 @@ async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey 2!");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey 2!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let events_2 = events_2.collect::<Vec<_>>().await;assert_eq!(stop_events(events_1), vec![acp::StopReason::EndTurn]);@@ -3372,8 +3497,12 @@ async fn test_subsequent_successful_sends_dont_cancel(cx: &mut TestAppContext) {#[gpui::test]async fn test_refusal(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let events = thread.update(cx, |thread, cx| {@@ -3392,7 +3521,7 @@ async fn test_refusal(cx: &mut TestAppContext) {);});- fake_model.send_last_completion_stream_text_chunk("Hey!");+ fake.send_last_text(&model, "Hey!");cx.run_until_parked();thread.read_with(cx, |thread, _| {assert_eq!(@@ -3410,8 +3539,10 @@ async fn test_refusal(cx: &mut TestAppContext) {});// If the model refuses to continue, the thread should remove all the messages after the last user message.- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::Refusal));+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::Refusal),+ );let events = events.collect::<Vec<_>>().await;assert_eq!(stop_events(events), vec![acp::StopReason::Refusal]);thread.read_with(cx, |thread, _| {@@ -3421,8 +3552,12 @@ async fn test_refusal(cx: &mut TestAppContext) {#[gpui::test]async fn test_truncate_first_message(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let message_id = ClientUserMessageId::new();thread@@ -3443,15 +3578,16 @@ async fn test_truncate_first_message(cx: &mut TestAppContext) {assert_eq!(thread.latest_token_usage(), None);});- fake_model.send_last_completion_stream_text_chunk("Hey!");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Hey!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 32_000,output_tokens: 16_000,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));+ }),+ );cx.run_until_parked();thread.read_with(cx, |thread, _| {assert_eq!(@@ -3504,15 +3640,16 @@ async fn test_truncate_first_message(cx: &mut TestAppContext) {);});cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Ahoy!");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Ahoy!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 40_000,output_tokens: 20_000,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));+ }),+ );cx.run_until_parked();thread.read_with(cx, |thread, _| {assert_eq!(@@ -3543,8 +3680,12 @@ async fn test_truncate_first_message(cx: &mut TestAppContext) {#[gpui::test]async fn test_latest_token_usage_counts_cached_input_tokens(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let message_1_id = ClientUserMessageId::new();thread@@ -3554,16 +3695,17 @@ async fn test_latest_token_usage_counts_cached_input_tokens(cx: &mut TestAppCont.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Response 1");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 1");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 100,output_tokens: 50,cache_creation_input_tokens: 25,cache_read_input_tokens: 75,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3594,8 +3736,12 @@ async fn test_latest_token_usage_counts_cached_input_tokens(cx: &mut TestAppCont#[gpui::test]async fn test_prompt_too_large_marks_token_usage_exceeded(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, cx| {@@ -3604,15 +3750,16 @@ async fn test_prompt_too_large_marks_token_usage_exceeded(cx: &mut TestAppContex.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Response 1");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 1");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 100,output_tokens: 50,..Default::default()- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3629,13 +3776,16 @@ async fn test_prompt_too_large_marks_token_usage_exceeded(cx: &mut TestAppContex.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- LanguageModelProviderName::new("test"),- http_client::StatusCode::PAYLOAD_TOO_LARGE,- "prompt too large".to_string(),- None,- ));- fake_model.end_last_completion_stream();+ fake.send_last_error(+ &model,+ LanguageModelCompletionError::from_http_status(+ LanguageModelProviderName::new("test"),+ http_client::StatusCode::PAYLOAD_TOO_LARGE,+ "prompt too large".to_string(),+ None,+ ),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3648,8 +3798,12 @@ async fn test_prompt_too_large_marks_token_usage_exceeded(cx: &mut TestAppContex#[gpui::test]async fn test_prompt_too_large_uses_reported_token_count(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, cx| {@@ -3658,13 +3812,16 @@ async fn test_prompt_too_large_uses_reported_token_count(cx: &mut TestAppContext.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- LanguageModelProviderName::new("test"),- http_client::StatusCode::PAYLOAD_TOO_LARGE,- "prompt is too long: 1500000 tokens".to_string(),- None,- ));- fake_model.end_last_completion_stream();+ fake.send_last_error(+ &model,+ LanguageModelCompletionError::from_http_status(+ LanguageModelProviderName::new("test"),+ http_client::StatusCode::PAYLOAD_TOO_LARGE,+ "prompt is too long: 1500000 tokens".to_string(),+ None,+ ),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3678,11 +3835,11 @@ async fn test_prompt_too_large_uses_reported_token_count(cx: &mut TestAppContextasync fn test_cumulative_token_usage(cx: &mut TestAppContext) {let ThreadTest {model,+ fake,thread,project_context,..} = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();thread.update(cx, |thread, cx| {@@ -3694,43 +3851,47 @@ async fn test_cumulative_token_usage(cx: &mut TestAppContext) {// The first request emits two cumulative snapshots; only the final values// must be counted, exactly once.- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 100,output_tokens: 10,..Default::default()- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 100,output_tokens: 50,..Default::default()- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: EchoTool::NAME.into(),raw_input: json!({"text": "hello"}).to_string(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "hello"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// The second request (after the tool call) is counted in addition to the first.- fake_model.send_last_completion_stream_text_chunk("Done");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {+ fake.send_last_text(&model, "Done");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 200,output_tokens: 30,..Default::default()- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let expected = TokenUsage {@@ -3772,8 +3933,12 @@ async fn test_cumulative_token_usage(cx: &mut TestAppContext) {#[gpui::test]async fn test_cumulative_token_usage_keeps_accounted_usage_monotonic(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, cx| {@@ -3782,24 +3947,27 @@ async fn test_cumulative_token_usage_keeps_accounted_usage_monotonic(cx: &mut Te.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 100,output_tokens: 10,..Default::default()- },- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage::default(),- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage::default()),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 100,output_tokens: 50,..Default::default()- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3816,8 +3984,12 @@ async fn test_cumulative_token_usage_keeps_accounted_usage_monotonic(cx: &mut Te#[gpui::test]async fn test_truncate_second_message(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, cx| {@@ -3825,16 +3997,17 @@ async fn test_truncate_second_message(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Message 1 response");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Message 1 response");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 32_000,output_tokens: 16_000,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let assert_first_message_state = |cx: &mut TestAppContext| {@@ -3875,16 +4048,17 @@ async fn test_truncate_second_message(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Message 2 response");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Message 2 response");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 40_000,output_tokens: 20_000,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3931,10 +4105,14 @@ async fn test_truncate_second_message(cx: &mut TestAppContext) {#[gpui::test]async fn test_title_generation(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;- let summary_model = Arc::new(FakeLanguageModel::default());+ let summary_model = fake.model("summary");thread.update(cx, |thread, cx| {thread.set_summarization_model(Some(summary_model.clone()), cx)});@@ -3946,16 +4124,17 @@ async fn test_title_generation(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey!");+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _| assert_eq!(thread.title(), None));// Ensure the summary model has been invoked to generate a title.- summary_model.send_last_completion_stream_text_chunk("Hello ");- summary_model.send_last_completion_stream_text_chunk("world\nG");- summary_model.send_last_completion_stream_text_chunk("oodnight Moon");- summary_model.end_last_completion_stream();+ let summary_request = fake.pending_completions_for(&summary_model).pop().unwrap();+ fake.send_text(&summary_model, &summary_request, "Hello ");+ fake.send_text(&summary_model, &summary_request, "world\nG");+ fake.send_text(&summary_model, &summary_request, "oodnight Moon");+ fake.end_stream(&summary_model, &summary_request);send.collect::<Vec<_>>().await;cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -3969,10 +4148,10 @@ async fn test_title_generation(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey again!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey again!");+ fake.end_last(&model);cx.run_until_parked();- assert_eq!(summary_model.pending_completions(), Vec::new());+ assert_eq!(fake.pending_completions_for(&summary_model), Vec::new());send.collect::<Vec<_>>().await;thread.read_with(cx, |thread, _| {assert_eq!(thread.title(), Some("Hello world".into()))@@ -3981,18 +4160,19 @@ async fn test_title_generation(cx: &mut TestAppContext) {#[gpui::test]async fn test_stream_thread_title_keeps_only_first_line(cx: &mut TestAppContext) {- let model = Arc::new(FakeLanguageModel::default());let request = LanguageModelRequest::default();+ let fake = cx.update(LanguageModelRegistry::test);+ let registered_model = fake.model("fake");let title_task = cx.spawn({- let model = model.clone();- async move |cx| crate::stream_thread_title(model, request, &cx).await+ let registered_model = registered_model.clone();+ async move |cx| crate::stream_thread_title(registered_model, request, &cx).await});cx.run_until_parked();- model.send_last_completion_stream_text_chunk("Hello world\nGoodnight Moon");- model.end_last_completion_stream();+ fake.send_last_text(®istered_model, "Hello world\nGoodnight Moon");+ fake.end_last(®istered_model);let title = title_task.await.unwrap();assert_eq!(title, "Hello world");@@ -4000,19 +4180,20 @@ async fn test_stream_thread_title_keeps_only_first_line(cx: &mut TestAppContext)#[gpui::test]async fn test_stream_thread_title_stops_when_newline_ends_chunk(cx: &mut TestAppContext) {- let model = Arc::new(FakeLanguageModel::default());let request = LanguageModelRequest::default();+ let fake = cx.update(LanguageModelRegistry::test);+ let registered_model = fake.model("fake");let title_task = cx.spawn({- let model = model.clone();- async move |cx| crate::stream_thread_title(model, request, &cx).await+ let registered_model = registered_model.clone();+ async move |cx| crate::stream_thread_title(registered_model, request, &cx).await});cx.run_until_parked();- model.send_last_completion_stream_text_chunk("Hello world\n");- model.send_last_completion_stream_text_chunk("Goodnight Moon");- model.end_last_completion_stream();+ fake.send_last_text(®istered_model, "Hello world\n");+ fake.send_last_text(®istered_model, "Goodnight Moon");+ fake.end_last(®istered_model);let title = title_task.await.unwrap();assert_eq!(title, "Hello world");@@ -4024,8 +4205,12 @@ async fn test_stream_thread_title_stops_when_newline_ends_chunk(cx: &mut TestApp// they share a single rendering path.#[gpui::test]async fn test_db_thread_markdown_matches_live_thread(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let send = thread.update(cx, |thread, cx| {@@ -4033,8 +4218,8 @@ async fn test_db_thread_markdown_matches_live_thread(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey there!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey there!");+ fake.end_last(&model);send.collect::<Vec<_>>().await;cx.run_until_parked();@@ -4047,11 +4232,14 @@ async fn test_db_thread_markdown_matches_live_thread(cx: &mut TestAppContext) {#[gpui::test]async fn test_title_generation_failure_allows_retry(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;- let summary_model = Arc::new(FakeLanguageModel::default());- let fake_summary_model = summary_model.as_fake();+ let summary_model = fake.model("summary");thread.update(cx, |thread, cx| {thread.set_summarization_model(Some(summary_model.clone()), cx)});@@ -4063,11 +4251,14 @@ async fn test_title_generation_failure_allows_retry(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey!");+ fake.end_last(&model);cx.run_until_parked();- fake_summary_model.send_last_completion_stream_error(+ let summary_request = fake.pending_completions_for(&summary_model).pop().unwrap();+ fake.send_error(+ &summary_model,+ &summary_request,LanguageModelCompletionError::from_http_status(language_model::LanguageModelProviderName::new("test"),gpui::http_client::StatusCode::INTERNAL_SERVER_ERROR,@@ -4075,7 +4266,7 @@ async fn test_title_generation_failure_allows_retry(cx: &mut TestAppContext) {None,),);- fake_summary_model.end_last_completion_stream();+ fake.end_stream(&summary_model, &summary_request);send.collect::<Vec<_>>().await;cx.run_until_parked();@@ -4101,8 +4292,9 @@ async fn test_title_generation_failure_allows_retry(cx: &mut TestAppContext) {assert!(thread.is_generating_title());});- fake_summary_model.send_last_completion_stream_text_chunk("Retried title");- fake_summary_model.end_last_completion_stream();+ let summary_request = fake.pending_completions_for(&summary_model).pop().unwrap();+ fake.send_text(&summary_model, &summary_request, "Retried title");+ fake.end_stream(&summary_model, &summary_request);cx.run_until_parked();thread.read_with(cx, |thread, _| {@@ -4115,8 +4307,12 @@ async fn test_title_generation_failure_allows_retry(cx: &mut TestAppContext) {#[gpui::test]async fn test_building_request_with_pending_tools(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let _events = thread.update(cx, |thread, cx| {@@ -4143,14 +4339,16 @@ async fn test_building_request_with_pending_tools(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- fake_model.send_last_completion_stream_text_chunk("Hi!");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- permission_tool_use,- ));- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- echo_tool_use.clone(),- ));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hi!");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(permission_tool_use),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(echo_tool_use.clone()),+ );+ fake.end_last(&model);cx.run_until_parked();// Ensure pending tools are skipped when building a request.@@ -4211,8 +4409,8 @@ async fn test_agent_connection(cx: &mut TestAppContext) {language_model::init(cx);RefreshLlmTokenListener::register(client.clone(), user_store.clone(), cx);language_models::init(user_store, client.clone(), cx);- LanguageModelRegistry::test(cx);});+ let fake = cx.update(LanguageModelRegistry::test);cx.executor().forbid_parking();// Create a project for new_thread@@ -4268,12 +4466,11 @@ async fn test_agent_connection(cx: &mut TestAppContext) {let model = cx.update(|cx| agent.read(cx).models().model_from_id(&model.id)).unwrap();- let model = model.as_fake();assert_eq!(model.id().0, "fake", "should return default model");let request = acp_thread.update(cx, |thread, cx| thread.send(vec!["abc".into()], cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("def");+ fake.send_last_text(&model, "def");cx.run_until_parked();acp_thread.read_with(cx, |thread, cx| {assert_eq!(@@ -4318,9 +4515,13 @@ async fn test_agent_connection(cx: &mut TestAppContext) {#[gpui::test]async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {- let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;+ let ThreadTest {+ model,+ thread,+ fake,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| thread.add_tool(EchoTool));- let fake_model = model.as_fake();let mut events = thread.update(cx, |thread, cx| {@@ -4331,30 +4532,32 @@ async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {// Simulate streaming partial input.let input = json!({});- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "1".into(),name: EchoTool::NAME.into(),raw_input: input.to_string(),input: language_model::LanguageModelToolUseInput::Json(input),is_input_complete: false,thought_signature: None,- },- ));+ }),+ );// Input streaming completedlet input = json!({ "text": "Hello!" });- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "1".into(),name: "echo".into(),raw_input: input.to_string(),input: language_model::LanguageModelToolUseInput::Json(input),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// User message is index 0, so the tool call is scoped to index 1 (see@@ -4401,8 +4604,12 @@ async fn test_tool_updates_to_completion(cx: &mut TestAppContext) {#[gpui::test]async fn test_send_no_retry_on_success(cx: &mut TestAppContext) {- let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ thread,+ fake,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -4411,8 +4618,8 @@ async fn test_send_no_retry_on_success(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey!");+ fake.end_last(&model);let mut retry_events = Vec::new();while let Some(Ok(event)) = events.next().await {@@ -4444,8 +4651,12 @@ async fn test_send_no_retry_on_success(cx: &mut TestAppContext) {#[gpui::test]async fn test_send_retry_on_error(cx: &mut TestAppContext) {- let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ thread,+ fake,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -4454,14 +4665,17 @@ async fn test_send_retry_on_error(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Hey,");- fake_model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- LanguageModelProviderName::new("Anthropic"),- http_client::StatusCode::SERVICE_UNAVAILABLE,- "Anthropic's API servers are overloaded right now".to_string(),- Some(Duration::from_secs(3)),- ));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Hey,");+ fake.send_last_error(+ &model,+ LanguageModelCompletionError::from_http_status(+ LanguageModelProviderName::new("Anthropic"),+ http_client::StatusCode::SERVICE_UNAVAILABLE,+ "Anthropic's API servers are overloaded right now".to_string(),+ Some(Duration::from_secs(3)),+ ),+ );+ fake.end_last(&model);cx.executor().advance_clock(crate::maximum_retry_delay_with_jitter(Duration::from_secs(@@ -4469,8 +4683,8 @@ async fn test_send_retry_on_error(cx: &mut TestAppContext) {)));cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("there!");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "there!");+ fake.end_last(&model);cx.run_until_parked();let mut retry_events = Vec::new();@@ -4513,8 +4727,12 @@ async fn test_send_retry_on_error(cx: &mut TestAppContext) {#[gpui::test]async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {- let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ thread,+ fake,+ ..+ } = setup(cx, TestModel::Fake).await;let events = thread.update(cx, |thread, cx| {@@ -4532,22 +4750,26 @@ async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- tool_use_1.clone(),- ));- fake_model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- LanguageModelProviderName::new("Anthropic"),- http_client::StatusCode::SERVICE_UNAVAILABLE,- "Anthropic's API servers are overloaded right now".to_string(),- Some(Duration::from_secs(3)),- ));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(tool_use_1.clone()),+ );+ fake.send_last_error(+ &model,+ LanguageModelCompletionError::from_http_status(+ LanguageModelProviderName::new("Anthropic"),+ http_client::StatusCode::SERVICE_UNAVAILABLE,+ "Anthropic's API servers are overloaded right now".to_string(),+ Some(Duration::from_secs(3)),+ ),+ );+ fake.end_last(&model);cx.executor().advance_clock(crate::maximum_retry_delay_with_jitter(Duration::from_secs(3,)));- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();assert_eq!(completion.messages[1..],vec![@@ -4580,8 +4802,8 @@ async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {]);- fake_model.send_last_completion_stream_text_chunk("Done");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done");+ fake.end_last(&model);cx.run_until_parked();events.collect::<Vec<_>>().await;thread.read_with(cx, |thread, _cx| {@@ -4598,8 +4820,12 @@ async fn test_send_retry_finishes_tool_calls_on_error(cx: &mut TestAppContext) {#[gpui::test]async fn test_send_max_retries_exceeded(cx: &mut TestAppContext) {- let ThreadTest { thread, model, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ thread,+ fake,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -4609,7 +4835,8 @@ async fn test_send_max_retries_exceeded(cx: &mut TestAppContext) {cx.run_until_parked();for _ in 0..crate::thread::MAX_RETRY_ATTEMPTS + 1 {- fake_model.send_last_completion_stream_error(+ fake.send_last_error(+ &model,LanguageModelCompletionError::from_http_status(LanguageModelProviderName::new("Anthropic"),http_client::StatusCode::SERVICE_UNAVAILABLE,@@ -4617,7 +4844,7 @@ async fn test_send_max_retries_exceeded(cx: &mut TestAppContext) {Some(Duration::from_secs(3)),),);- fake_model.end_last_completion_stream();+ fake.end_last(&model);cx.executor().advance_clock(crate::maximum_retry_delay_with_jitter(Duration::from_secs(3,@@ -4665,8 +4892,12 @@ async fn test_streaming_tool_completes_when_llm_stream_ends_without_final_input(init_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| {thread.add_tool(StreamingEchoTool::new());@@ -4693,21 +4924,26 @@ async fn test_streaming_tool_completes_when_llm_stream_ends_without_final_input(is_input_complete: false,thought_signature: None,};- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(tool_use.clone()),+ );cx.run_until_parked();// Send a stream error WITHOUT ever sending is_input_complete = true.// Before the fix, this would deadlock: the tool waits for more partials// (or cancellation), run_turn_internal waits for the tool, and the sender// keeping the channel open lives inside RunningTurn.- fake_model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- language_model::LanguageModelProviderName::new("test"),- http_client::StatusCode::INTERNAL_SERVER_ERROR,- "Internal server error".to_string(),- None,- ));- fake_model.end_last_completion_stream();+ fake.send_last_error(+ &model,+ LanguageModelCompletionError::from_http_status(+ language_model::LanguageModelProviderName::new("test"),+ http_client::StatusCode::INTERNAL_SERVER_ERROR,+ "Internal server error".to_string(),+ None,+ ),+ );+ fake.end_last(&model);// Advance past the retry delay so run_turn_internal retries.cx.executor()@@ -4718,10 +4954,7 @@ async fn test_streaming_tool_completes_when_llm_stream_ends_without_final_input(// The retry request should contain the streaming tool's error result,// proving the tool terminated and its result was forwarded.- let completion = fake_model- .pending_completions()- .pop()- .expect("No running turn");+ let completion = fake.pending_completions().pop().expect("No running turn");assert_eq!(completion.messages[1..],vec![@@ -4755,8 +4988,8 @@ async fn test_streaming_tool_completes_when_llm_stream_ends_without_final_input();// Finish the retry round so the turn completes cleanly.- fake_model.send_last_completion_stream_text_chunk("Done");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done");+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _cx| {@@ -4774,8 +5007,12 @@ async fn test_streaming_tool_json_parse_error_is_forwarded_to_running_tool(init_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| {thread.add_tool(StreamingJsonErrorContextTool);@@ -4800,10 +5037,11 @@ async fn test_streaming_tool_json_parse_error_is_forwarded_to_running_tool(is_input_complete: false,thought_signature: None,};- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use));+ fake.send_last_event(&model, LanguageModelCompletionEvent::ToolUse(tool_use));cx.run_until_parked();- fake_model.send_last_completion_stream_event(+ fake.send_last_event(+ &model,LanguageModelCompletionEvent::ToolUseJsonParseError {id: "tool_1".into(),tool_name: StreamingJsonErrorContextTool::NAME.into(),@@ -4811,18 +5049,17 @@ async fn test_streaming_tool_json_parse_error_is_forwarded_to_running_tool(json_parse_error: "EOF while parsing a string at line 1 column 17".into(),},);- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::ToolUse));- fake_model.end_last_completion_stream();+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::ToolUse),+ );+ fake.end_last(&model);cx.run_until_parked();cx.executor().advance_clock(Duration::from_secs(5));cx.run_until_parked();- let completion = fake_model- .pending_completions()- .pop()- .expect("No running turn");+ let completion = fake.pending_completions().pop().expect("No running turn");let tool_results: Vec<_> = completion.messages@@ -4863,8 +5100,8 @@ async fn test_streaming_tool_json_parse_error_is_forwarded_to_running_tool("Should not contain orphaned sender error, got: {content_text}");- fake_model.send_last_completion_stream_text_chunk("Done");- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "Done");+ fake.end_last(&model);cx.run_until_parked();thread.read_with(cx, |thread, _cx| {@@ -4887,7 +5124,9 @@ fn stop_events(result_events: Vec<Result<ThreadEvent>>) -> Vec<acp::StopReason>}struct ThreadTest {- model: Arc<dyn LanguageModel>,+ model: LanguageModel,+ /// Serves `model`; unused with `TestModel::Sonnet4`.+ fake: Arc<FakeLanguageModelProvider>,thread: Entity<Thread>,project_context: Entity<ProjectContext>,context_server_store: Entity<ContextServerStore>,@@ -4945,11 +5184,11 @@ async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {).await;- cx.update(|cx| {+ let fake = cx.update(|cx| {settings::init(cx);- match model {- TestModel::Fake => {}+ let fake = match model {+ TestModel::Fake => LanguageModelRegistry::test(cx),TestModel::Sonnet4 => {<dyn fs::Fs>::set_global(fs.clone(), cx);gpui_tokio::init(cx);@@ -4960,10 +5199,12 @@ async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {language_model::init(cx);RefreshLlmTokenListener::register(client.clone(), user_store.clone(), cx);language_models::init(user_store, client.clone(), cx);+ Arc::new(FakeLanguageModelProvider::default())}};watch_settings(fs.clone(), cx);+ fake});let templates = Templates::new();@@ -4974,7 +5215,7 @@ async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {let model = cx.update(|cx| {if let TestModel::Fake = model {- Task::ready(Arc::new(FakeLanguageModel::default()) as Arc<_>)+ Task::ready(fake.model("fake"))} else {let model_id = model.id();let models = LanguageModelRegistry::read_global(cx);@@ -5010,6 +5251,7 @@ async fn setup(cx: &mut TestAppContext, model: TestModel) -> ThreadTest {});ThreadTest {model,+ fake,thread,project_context,context_server_store,@@ -5141,8 +5383,12 @@ fn setup_context_server(#[gpui::test]async fn test_tokens_before_message(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;// First messagelet message_1_id = ClientUserMessageId::new();@@ -5163,16 +5409,17 @@ async fn test_tokens_before_message(cx: &mut TestAppContext) {});// Complete first message with usage- fake_model.send_last_completion_stream_text_chunk("Response 1");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 1");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 100,output_tokens: 50,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// First message still has no tokens before it@@ -5203,16 +5450,17 @@ async fn test_tokens_before_message(cx: &mut TestAppContext) {});// Complete second message- fake_model.send_last_completion_stream_text_chunk("Response 2");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 2");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 250, // Total for this request (includes previous context)output_tokens: 75,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// Third message@@ -5248,8 +5496,12 @@ async fn test_tokens_before_message(cx: &mut TestAppContext) {#[gpui::test]async fn test_tokens_before_message_after_truncate(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;// Set up three messages with responseslet message_1_id = ClientUserMessageId::new();@@ -5259,16 +5511,17 @@ async fn test_tokens_before_message_after_truncate(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Response 1");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 1");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 100,output_tokens: 50,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();let message_2_id = ClientUserMessageId::new();@@ -5278,16 +5531,17 @@ async fn test_tokens_before_message_after_truncate(cx: &mut TestAppContext) {}).unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_text_chunk("Response 2");- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- language_model::TokenUsage {+ fake.send_last_text(&model, "Response 2");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::UsageUpdate(language_model::TokenUsage {input_tokens: 250,output_tokens: 75,cache_creation_input_tokens: 0,cache_read_input_tokens: 0,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);cx.run_until_parked();// Verify initial state@@ -5612,10 +5866,7 @@ async fn test_spawn_agent_tool_rejects_model_when_resuming(cx: &mut TestAppConte#[gpui::test]async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {- init_test(cx);- cx.update(|cx| {- LanguageModelRegistry::test(cx);- });+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);});@@ -5648,7 +5899,7 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {let thread = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("thread");// Ensure empty threads are not saved, even if they get mutated.thread.update(cx, |thread, cx| {@@ -5658,7 +5909,8 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {let send = acp_thread.update(cx, |thread, cx| thread.send_raw("Prompt", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("spawning subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "spawning subagent");let subagent_tool_input = SpawnAgentToolInput {label: "label".to_string(),message: "subagent task prompt".to_string(),@@ -5675,10 +5927,12 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- subagent_tool_use,- ));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(subagent_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -5700,8 +5954,9 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {.expect("subagent thread should be alive")});- model.send_last_completion_stream_text_chunk("subagent task response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "subagent task response");+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -5719,8 +5974,9 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {"});- model.send_last_completion_stream_text_chunk("Response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "Response");+ fake.end_stream(&model, &request);send.await.unwrap();@@ -5750,10 +6006,7 @@ async fn test_subagent_tool_call_end_to_end(cx: &mut TestAppContext) {#[gpui::test]async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppContext) {- init_test(cx);- cx.update(|cx| {- LanguageModelRegistry::test(cx);- });+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);});@@ -5786,7 +6039,7 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppConlet thread = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("thread");// Ensure empty threads are not saved, even if they get mutated.thread.update(cx, |thread, cx| {@@ -5796,7 +6049,8 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppConlet send = acp_thread.update(cx, |thread, cx| thread.send_raw("Prompt", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("spawning subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "spawning subagent");let subagent_tool_input = SpawnAgentToolInput {label: "label".to_string(),message: "subagent task prompt".to_string(),@@ -5813,10 +6067,12 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppConis_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- subagent_tool_use,- ));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(subagent_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -5838,13 +6094,18 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppCon.expect("subagent thread should be alive")});- model.send_last_completion_stream_text_chunk("subagent task response 1");- model.send_last_completion_stream_event(LanguageModelCompletionEvent::Thinking {- text: "thinking more about the subagent task".into(),- signature: None,- });- model.send_last_completion_stream_text_chunk("subagent task response 2");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "subagent task response 1");+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::Thinking {+ text: "thinking more about the subagent task".into(),+ signature: None,+ },+ );+ fake.send_text(&model, &request, "subagent task response 2");+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -5868,8 +6129,9 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppCon"});- model.send_last_completion_stream_text_chunk("Response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "Response");+ fake.end_stream(&model, &request);send.await.unwrap();@@ -5901,10 +6163,7 @@ async fn test_subagent_tool_output_does_not_include_thinking(cx: &mut TestAppCon#[gpui::test]async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestAppContext) {- init_test(cx);- cx.update(|cx| {- LanguageModelRegistry::test(cx);- });+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);});@@ -5937,7 +6196,7 @@ async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestAplet thread = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");// Ensure empty threads are not saved, even if they get mutated.thread.update(cx, |thread, cx| {@@ -5947,7 +6206,8 @@ async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestAplet send = acp_thread.update(cx, |thread, cx| thread.send_raw("Prompt", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("spawning subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "spawning subagent");let subagent_tool_input = SpawnAgentToolInput {label: "label".to_string(),message: "subagent task prompt".to_string(),@@ -5964,10 +6224,12 @@ async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestApis_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- subagent_tool_use,- ));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(subagent_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -5988,8 +6250,8 @@ async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestAp.expect("subagent thread should be alive")});- // model.send_last_completion_stream_text_chunk("subagent task response");- // model.end_last_completion_stream();+ // fake.send_last_text(&model, "subagent task response");+ // fake.end_last(&model);// cx.run_until_parked();@@ -6034,10 +6296,7 @@ async fn test_subagent_tool_call_cancellation_during_task_prompt(cx: &mut TestAp#[gpui::test]async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {- init_test(cx);- cx.update(|cx| {- LanguageModelRegistry::test(cx);- });+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);});@@ -6070,7 +6329,7 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {let thread = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("thread");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);@@ -6080,7 +6339,8 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {// === First turn: create subagent ===let send = acp_thread.update(cx, |thread, cx| thread.send_raw("First prompt", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("spawning subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "spawning subagent");let subagent_tool_input = SpawnAgentToolInput {label: "initial task".to_string(),message: "do the first task".to_string(),@@ -6097,10 +6357,12 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- subagent_tool_use,- ));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(subagent_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -6123,14 +6385,16 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {});// Subagent responds- model.send_last_completion_stream_text_chunk("first task response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "first task response");+ fake.end_stream(&model, &request);cx.run_until_parked();// Parent model responds to complete first turn- model.send_last_completion_stream_text_chunk("First response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "First response");+ fake.end_stream(&model, &request);send.await.unwrap();@@ -6145,7 +6409,8 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {// === Second turn: resume subagent with session_id ===let send2 = acp_thread.update(cx, |thread, cx| thread.send_raw("Follow up", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("resuming subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "resuming subagent");let resume_tool_input = SpawnAgentToolInput {label: "follow-up task".to_string(),message: "do the follow-up task".to_string(),@@ -6162,8 +6427,12 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(resume_tool_use));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(resume_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -6175,14 +6444,16 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {});// Subagent responds to follow-up- model.send_last_completion_stream_text_chunk("follow-up task response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "follow-up task response");+ fake.end_stream(&model, &request);cx.run_until_parked();// Parent model responds to complete second turn- model.send_last_completion_stream_text_chunk("Second response");- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "Second response");+ fake.end_stream(&model, &request);send2.await.unwrap();@@ -6220,7 +6491,7 @@ async fn test_subagent_tool_resume_session(cx: &mut TestAppContext) {#[gpui::test]async fn test_subagent_thread_inherits_parent_thread_properties(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);@@ -6233,7 +6504,7 @@ async fn test_subagent_thread_inherits_parent_thread_properties(cx: &mut TestApplet context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let parent_thread = cx.new(|cx| {Thread::new(@@ -6268,7 +6539,7 @@ async fn test_subagent_thread_inherits_parent_thread_properties(cx: &mut TestApp#[gpui::test]async fn test_subagent_thread_model_selection(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree(path!("/test"), json!({})).await;@@ -6277,30 +6548,20 @@ async fn test_subagent_thread_model_selection(cx: &mut TestAppContext) {let context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let parent_model = Arc::new(FakeLanguageModel::default());- let subagent_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "subagent-model",- "Subagent Model",- true,- ));- let explicit_model = Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake-corp",- "explicit-model",- "Explicit Model",- false,+ let parent_model = fake.model("fake");+ let provider = Arc::new(FakeLanguageModelProvider::new(+ LanguageModelProviderId::from("fake-corp".to_string()),+ LanguageModelProviderName::from("Fake Corp".to_string()),));+ let subagent_model = provider.update_model("subagent-model", |model| {+ model.name = LanguageModelName::from("Subagent Model".to_string());+ model.supports_thinking = true;+ });+ let explicit_model = provider.update_model("explicit-model", |model| {+ model.name = LanguageModelName::from("Explicit Model".to_string());+ });cx.update(|cx| {- LanguageModelRegistry::test(cx);-- let provider = Arc::new(- FakeLanguageModelProvider::new(- LanguageModelProviderId::from("fake-corp".to_string()),- LanguageModelProviderName::from("Fake Corp".to_string()),- )- .with_models(vec![subagent_model.clone(), explicit_model.clone()]),- );LanguageModelRegistry::global(cx).update(cx, |registry, cx| {registry.register_provider(provider, cx);});@@ -6386,7 +6647,7 @@ async fn test_subagent_thread_model_selection(cx: &mut TestAppContext) {#[gpui::test]async fn test_max_subagent_depth_prevents_tool_registration(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);@@ -6399,7 +6660,7 @@ async fn test_max_subagent_depth_prevents_tool_registration(cx: &mut TestAppContlet context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))}));@@ -6436,7 +6697,7 @@ async fn test_max_subagent_depth_prevents_tool_registration(cx: &mut TestAppCont#[gpui::test]async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);let fs = FakeFs::new(cx.executor());fs.insert_tree(path!("/test"), json!({})).await;@@ -6445,7 +6706,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {let context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))}));@@ -6456,7 +6717,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {project_context,context_server_registry,Templates::new(),- Some(model.clone() as Arc<dyn LanguageModel>),+ Some(model.clone()),cx,);thread.add_default_tools(environment, cx);@@ -6497,7 +6758,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_names = tool_names_for_completion(&completion);for name in &lsp_tool_names {assert!(@@ -6516,7 +6777,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {tool_names.iter().any(|t| t == ReadFileTool::NAME),"expected non-LSP tools to still be exposed, got: {tool_names:?}");- model.end_last_completion_stream();+ fake.end_last(&model);cx.run_until_parked();// Enable the `lsp-tool` flag and send another message; the LSP tools@@ -6532,7 +6793,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_names = tool_names_for_completion(&completion);for name in &lsp_tool_names {assert!(@@ -6550,7 +6811,7 @@ async fn test_lsp_tools_gated_by_feature_flag(cx: &mut TestAppContext) {#[gpui::test]async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);// `CreateThreadToolFeatureFlag::enabled_for_staff()` returns true, which// means tests in debug builds resolve it to ON unless we explicitly@@ -6585,7 +6846,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {let context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let environment = Rc::new(cx.update(|cx| {FakeThreadEnvironment::default().with_terminal(FakeTerminalHandle::new_never_exits(cx))}));@@ -6596,7 +6857,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {project_context,context_server_registry,Templates::new(),- Some(model.clone() as Arc<dyn LanguageModel>),+ Some(model.clone()),cx,);thread.add_default_tools(environment, cx);@@ -6619,7 +6880,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_names = tool_names_for_completion(&completion);assert!(!tool_names.iter().any(|tool| tool == CreateThreadTool::NAME));assert!(@@ -6632,7 +6893,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {tool_names.iter().any(|t| t == ReadFileTool::NAME),"expected non-sibling-thread tools to still be exposed, got: {tool_names:?}");- model.end_last_completion_stream();+ fake.end_last(&model);cx.run_until_parked();// Flag explicitly on: the next completion request includes both tools.@@ -6644,7 +6905,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {.unwrap();cx.run_until_parked();- let completion = model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let tool_names = tool_names_for_completion(&completion);assert!(tool_names.iter().any(|tool| tool == CreateThreadTool::NAME));assert!(@@ -6656,7 +6917,7 @@ async fn test_thread_tools_feature_gating(cx: &mut TestAppContext) {#[gpui::test]async fn test_parent_cancel_stops_subagent(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);cx.update(|cx| {cx.update_flags(true, vec!["subagents".to_string()]);@@ -6669,7 +6930,7 @@ async fn test_parent_cancel_stops_subagent(cx: &mut TestAppContext) {let context_server_store = project.read_with(cx, |project, _| project.context_server_store());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(context_server_store.clone(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");let parent = cx.new(|cx| {Thread::new(@@ -6741,7 +7002,7 @@ async fn test_subagent_auto_compaction(cx: &mut TestAppContext) {cx,);test.assert_running(cx);- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());cx.executor().advance_clock(crate::maximum_retry_delay_with_jitter(Duration::from_secs(3,@@ -6751,12 +7012,12 @@ async fn test_subagent_auto_compaction(cx: &mut TestAppContext) {assert_eq!(retry_request.messages, request.messages);let request = retry_request;- test.model- .send_completion_stream_text_chunk(&request, "subagent summary");- test.model.end_completion_stream(&request);+ test.fake+ .send_text(&test.model, &request, "subagent summary");+ test.fake.end_stream(&test.model, &request);cx.run_until_parked();- let request = test.model.pending_completions().pop().unwrap();+ let request = test.fake.pending_completions().pop().unwrap();assert_eq!(request.intent, Some(CompletionIntent::ToolResults));assert_eq!(request.thread_id, Some(test.handle.id().to_string()));assert_eq!(@@ -6771,9 +7032,9 @@ async fn test_subagent_auto_compaction(cx: &mut TestAppContext) {"The previous conversation was compacted. Use this summary as context:\n\nsubagent summary",],);- test.model- .send_completion_stream_text_chunk(&request, "subagent answer");- test.model.end_completion_stream(&request);+ test.fake+ .send_text(&test.model, &request, "subagent answer");+ test.fake.end_stream(&test.model, &request);assert_eq!(send.await.unwrap(), "subagent answer");test.assert_stopped(cx);@@ -6786,15 +7047,15 @@ async fn test_subagent_auto_compaction(cx: &mut TestAppContext) {assert_eq!(handle.id(), test.handle.id());let send = cx.update(|cx| handle.send("follow-up task".to_string(), &cx.to_async()));cx.run_until_parked();- let request = test.model.pending_completions().pop().unwrap();+ let request = test.fake.pending_completions().pop().unwrap();assert_eq!(request.intent, Some(CompletionIntent::Subagent));assert_eq!(request.messages.last().unwrap().string_contents(),"follow-up task");- test.model- .send_completion_stream_text_chunk(&request, "follow-up answer");- test.model.end_completion_stream(&request);+ test.fake+ .send_text(&test.model, &request, "follow-up answer");+ test.fake.end_stream(&test.model, &request);assert_eq!(send.await.unwrap(), "follow-up answer");test.assert_stopped(cx);@@ -6803,11 +7064,11 @@ async fn test_subagent_auto_compaction(cx: &mut TestAppContext) {if use_tool {test.tool_round(0, cx);}- test.model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(- StopReason::Refusal,- ));- test.model.end_last_completion_stream();+ test.fake.send_last_event(+ &test.model,+ LanguageModelCompletionEvent::Stop(StopReason::Refusal),+ );+ test.fake.end_last(&test.model);assert_eq!(send.await.unwrap_err().to_string(),"The agent refused to process that prompt. Try again."@@ -6828,12 +7089,12 @@ async fn test_subagent_compaction_respects_settings_and_context_window(cx: &muttest.configure_compaction(enabled, max_tokens, max_output_tokens, cx);let send = test.send("subagent task prompt", cx);test.tool_round(input_tokens, cx);- let request = test.model.pending_completions().pop().unwrap();+ let request = test.fake.pending_completions().pop().unwrap();assert_eq!(request.intent, Some(CompletionIntent::ToolResults));test.assert_running(cx);- test.model- .send_completion_stream_text_chunk(&request, "subagent answer");- test.model.end_completion_stream(&request);+ test.fake+ .send_text(&test.model, &request, "subagent answer");+ test.fake.end_stream(&test.model, &request);assert_eq!(send.await.unwrap(), "subagent answer");test.assert_stopped(cx);}@@ -6849,12 +7110,11 @@ async fn test_subagent_context_limit_warning_when_compaction_unavailable(cx: &mulet test = SubagentCompactionTest::new(cx).await;test.configure_compaction(enabled, max_tokens, max_output_tokens, cx);let send = test.send("subagent task prompt", cx);- let request = test.model.pending_completions().pop().unwrap();- test.model- .send_completion_stream_text_chunk(&request, "partial work");+ let request = test.fake.pending_completions().pop().unwrap();+ test.fake.send_text(&test.model, &request, "partial work");test.update_usage(warning_tokens - 1, cx);test.assert_running(cx);- assert!(!test.model.is_completion_stream_closed(&request));+ assert!(!test.fake.is_stream_closed(&test.model, &request));test.update_usage(warning_tokens, cx);assert_eq!(send.await.unwrap_err().to_string(),@@ -6863,18 +7123,17 @@ async fn test_subagent_context_limit_warning_when_compaction_unavailable(cx: &mu),);test.assert_stopped(cx);- assert!(test.model.is_completion_stream_closed(&request));- test.model.end_completion_stream(&request);- assert_eq!(test.model.pending_completions(), Vec::new());+ assert!(test.fake.is_stream_closed(&test.model, &request));+ test.fake.end_stream(&test.model, &request);+ assert_eq!(test.fake.pending_completions(), Vec::new());test.assert_no_compaction(cx).await;let send = test.send("wrap up", cx);- let request = test.model.pending_completions().pop().unwrap();- test.model- .send_completion_stream_text_chunk(&request, "wrapped up");+ let request = test.fake.pending_completions().pop().unwrap();+ test.fake.send_text(&test.model, &request, "wrapped up");test.update_usage(warning_tokens + 1, cx);test.assert_running(cx);- test.model.end_completion_stream(&request);+ test.fake.end_stream(&test.model, &request);assert_eq!(send.await.unwrap(), "wrapped up");test.assert_stopped(cx);}@@ -6885,16 +7144,16 @@ async fn test_subagent_context_limit_exceeded_without_partial_output(cx: &mut Telet test = SubagentCompactionTest::new(cx).await;test.configure_compaction(false, 1_000_000, None, cx);let send = test.send("subagent task prompt", cx);- let request = test.model.pending_completions().pop().unwrap();+ let request = test.fake.pending_completions().pop().unwrap();test.update_usage(1_000_000, cx);assert_eq!(send.await.unwrap_err().to_string(),SUBAGENT_CONTEXT_LIMIT_WARNING,);test.assert_stopped(cx);- assert!(test.model.is_completion_stream_closed(&request));- test.model.end_completion_stream(&request);- assert_eq!(test.model.pending_completions(), Vec::new());+ assert!(test.fake.is_stream_closed(&test.model, &request));+ test.fake.end_stream(&test.model, &request);+ assert_eq!(test.fake.pending_completions(), Vec::new());}#[gpui::test]@@ -6918,10 +7177,10 @@ async fn test_subagent_context_limit_preserves_terminal_result_during_checkpointlet checkpoint_job = repository.update(cx, |repository, _| {repository.send_job("hold checkpoint", None, move |_, _| checkpoint_gate)});- test.model- .send_last_completion_stream_text_chunk("partial work");+ test.fake.send_last_text(&test.model, "partial work");if provider_error {- test.model.send_last_completion_stream_error(+ test.fake.send_last_error(+ &test.model,LanguageModelCompletionError::from_http_status(LanguageModelProviderName::new("test"),http_client::StatusCode::BAD_REQUEST,@@ -6930,7 +7189,7 @@ async fn test_subagent_context_limit_preserves_terminal_result_during_checkpoint),);}- test.model.end_last_completion_stream();+ test.fake.end_last(&test.model);cx.run_until_parked();test.thread@@ -6978,7 +7237,7 @@ async fn test_subagent_context_limit_preserves_terminal_result_during_checkpointassert_eq!(send.await.unwrap(), "partial work");}test.assert_stopped(cx);- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());}}@@ -6993,7 +7252,7 @@ async fn test_subagent_compaction_exhausts_transient_retries(cx: &mut TestAppConSome(Duration::from_secs(3)),cx,);- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());if request_index < 4 {test.assert_running(cx);cx.executor()@@ -7018,7 +7277,7 @@ async fn test_subagent_compaction_exhausts_transient_retries(cx: &mut TestAppCon3,)));cx.run_until_parked();- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());}#[gpui::test]@@ -7027,19 +7286,19 @@ async fn test_subagent_compaction_parent_cancellation(cx: &mut TestAppContext) {let send = test.send("subagent task prompt", cx);test.tool_round(950_000, cx);let request = test.compaction_request();- test.model- .send_completion_stream_text_chunk(&request, "partial summary");+ test.fake+ .send_text(&test.model, &request, "partial summary");cx.run_until_parked();- assert!(!test.model.is_completion_stream_closed(&request));+ assert!(!test.fake.is_stream_closed(&test.model, &request));test.parent.update(cx, |thread, cx| thread.cancel(cx)).await;assert_eq!(send.await.unwrap_err().to_string(), "User canceled");cx.run_until_parked();- assert!(test.model.is_completion_stream_closed(&request));+ assert!(test.fake.is_stream_closed(&test.model, &request));test.assert_stopped(cx);test.assert_no_compaction(cx).await;- test.model.end_completion_stream(&request);+ test.fake.end_stream(&test.model, &request);cx.run_until_parked();- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());}#[gpui::test]@@ -7052,7 +7311,8 @@ async fn test_subagent_compaction_error_propagation(cx: &mut TestAppContext) {test.fail_compaction(http_client::StatusCode::BAD_REQUEST, None, cx);"compaction provider error"} else {- test.model.end_completion_stream(&test.compaction_request());+ test.fake+ .end_stream(&test.model, &test.compaction_request());"Compaction produced an empty summary"};assert_eq!(@@ -7065,15 +7325,14 @@ async fn test_subagent_compaction_error_propagation(cx: &mut TestAppContext) {);test.assert_stopped(cx);test.assert_no_compaction(cx).await;- assert_eq!(test.model.pending_completions(), Vec::new());+ assert_eq!(test.fake.pending_completions(), Vec::new());}}#[gpui::test]async fn test_subagent_error_propagation(cx: &mut TestAppContext) {- init_test(cx);+ let fake = init_test(cx);cx.update(|cx| {- LanguageModelRegistry::test(cx);let mut settings = AgentSettings::get_global(cx).clone();settings.auto_compact.enabled = false;AgentSettings::override_global(settings, cx);@@ -7110,7 +7369,7 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {let thread = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("thread");thread.update(cx, |thread, cx| {thread.set_model(model.clone(), cx);@@ -7120,7 +7379,8 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {// Start the parent turnlet send = acp_thread.update(cx, |thread, cx| thread.send_raw("Prompt", cx));cx.run_until_parked();- model.send_last_completion_stream_text_chunk("spawning subagent");+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "spawning subagent");let subagent_tool_input = SpawnAgentToolInput {label: "label".to_string(),message: "subagent task prompt".to_string(),@@ -7137,10 +7397,12 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {is_input_complete: true,thought_signature: None,};- model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- subagent_tool_use,- ));- model.end_last_completion_stream();+ fake.send_event(+ &model,+ &request,+ LanguageModelCompletionEvent::ToolUse(subagent_tool_use),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -7150,14 +7412,19 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {subagent_ids.into_iter().next().unwrap()});- model.send_last_completion_stream_text_chunk("partial work");- model.send_last_completion_stream_error(LanguageModelCompletionError::from_http_status(- LanguageModelProviderName::new("test"),- http_client::StatusCode::PAYLOAD_TOO_LARGE,- "prompt too large".to_string(),- None,- ));- model.end_last_completion_stream();+ let request = fake.pending_completions_for(&model).pop().unwrap();+ fake.send_text(&model, &request, "partial work");+ fake.send_error(+ &model,+ &request,+ LanguageModelCompletionError::from_http_status(+ LanguageModelProviderName::new("test"),+ http_client::StatusCode::PAYLOAD_TOO_LARGE,+ "prompt too large".to_string(),+ None,+ ),+ );+ fake.end_stream(&model, &request);cx.run_until_parked();@@ -7169,7 +7436,7 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {);});- let request = model.pending_completions().pop().unwrap();+ let request = fake.pending_completions_for(&model).pop().unwrap();assert_eq!(request.intent, Some(CompletionIntent::ToolResults));let tool_results = request.messages@@ -7197,8 +7464,8 @@ async fn test_subagent_error_propagation(cx: &mut TestAppContext) {}).to_string(),);- model.send_completion_stream_text_chunk(&request, "Response after error");- model.end_completion_stream(&request);+ fake.send_text(&model, &request, "Response after error");+ fake.end_stream(&model, &request);send.await.unwrap();}@@ -8114,8 +8381,12 @@ async fn test_fetch_tool_follows_same_host_redirect(cx: &mut TestAppContext) {/// sibling pending authorizations for the same tool in the same turn.#[gpui::test]async fn test_always_allow_resolves_pending_authorizations(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -8127,18 +8398,19 @@ async fn test_always_allow_resolves_pending_authorizations(cx: &mut TestAppConte// Two parallel tool calls, both require permission.for id in ["tool_id_1", "tool_id_2"] {- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: id.into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));+ }),+ );}- fake_model.end_last_completion_stream();+ fake.end_last(&model);let tool_call_auth_1 = next_tool_call_authorization(&mut events).await;let tool_call_auth_2 = next_tool_call_authorization(&mut events).await;@@ -8168,7 +8440,7 @@ async fn test_always_allow_resolves_pending_authorizations(cx: &mut TestAppConte"expected tool 2's response receiver to be dropped after auto-resolve");- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();let results: Vec<_> = message.content@@ -8195,8 +8467,12 @@ async fn test_always_allow_resolves_pending_authorizations(cx: &mut TestAppConte/// for tool calls that match the new rule.#[gpui::test]async fn test_external_settings_edit_resolves_pending_authorization(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -8206,17 +8482,18 @@ async fn test_external_settings_edit_resolves_pending_authorization(cx: &mut Tes.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_1".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let tool_call_auth = next_tool_call_authorization(&mut events).await;@@ -8249,7 +8526,7 @@ async fn test_external_settings_edit_resolves_pending_authorization(cx: &mut Tes"response receiver should have been dropped after settings-driven auto-resolve");- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();let result = message.content@@ -8266,8 +8543,12 @@ async fn test_external_settings_edit_resolves_pending_authorization(cx: &mut Tes/// authorization prompt and returns the tool call as denied.#[gpui::test]async fn test_external_deny_rule_resolves_pending_authorization(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -8277,17 +8558,18 @@ async fn test_external_deny_rule_resolves_pending_authorization(cx: &mut TestApp.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_1".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let tool_call_auth = next_tool_call_authorization(&mut events).await;@@ -8319,7 +8601,7 @@ async fn test_external_deny_rule_resolves_pending_authorization(cx: &mut TestApp"response receiver should have been dropped after deny auto-resolve");- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();let result = message.content@@ -8342,8 +8624,12 @@ async fn test_external_deny_rule_resolves_pending_authorization(cx: &mut TestAppasync fn test_unrelated_settings_change_does_not_resolve_pending_authorization(cx: &mut TestAppContext,) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -8353,17 +8639,18 @@ async fn test_unrelated_settings_change_does_not_resolve_pending_authorization(.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_id_1".into(),name: ToolRequiringPermission::NAME.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));- fake_model.end_last_completion_stream();+ }),+ );+ fake.end_last(&model);let tool_call_auth = next_tool_call_authorization(&mut events).await;@@ -8386,7 +8673,7 @@ async fn test_unrelated_settings_change_does_not_resolve_pending_authorization(.expect("response receiver should still be alive");cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();let result = message.content@@ -8405,8 +8692,12 @@ async fn test_unrelated_settings_change_does_not_resolve_pending_authorization(/// and waits for the user.#[gpui::test]async fn test_always_allow_does_not_resolve_unrelated_tool_authorization(cx: &mut TestAppContext) {- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let mut events = thread.update(cx, |thread, cx| {@@ -8423,18 +8714,19 @@ async fn test_always_allow_does_not_resolve_unrelated_tool_authorization(cx: &mu("tool_id_1", ToolRequiringPermission::NAME),("tool_id_2", ToolRequiringPermission2::NAME),] {- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: id.into(),name: name.into(),raw_input: "{}".into(),input: language_model::LanguageModelToolUseInput::Json(json!({})),is_input_complete: true,thought_signature: None,- },- ));+ }),+ );}- fake_model.end_last_completion_stream();+ fake.end_last(&model);let auth_a = next_tool_call_authorization(&mut events).await;let auth_b = next_tool_call_authorization(&mut events).await;@@ -8476,7 +8768,7 @@ async fn test_always_allow_does_not_resolve_unrelated_tool_authorization(cx: &mu.expect("tool 2's response receiver should still be alive");cx.run_until_parked();- let completion = fake_model.pending_completions().pop().unwrap();+ let completion = fake.pending_completions().pop().unwrap();let message = completion.messages.last().unwrap();let results: Vec<_> = message.content@@ -8503,8 +8795,12 @@ async fn test_queued_message_ends_turn_at_boundary(cx: &mut TestAppContext) {init_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;// Add a tool so we can simulate tool callsthread.update(cx, |thread, _cx| {@@ -8520,18 +8816,21 @@ async fn test_queued_message_ends_turn_at_boundary(cx: &mut TestAppContext) {cx.run_until_parked();// Simulate the model making a tool call- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "echo".into(),raw_input: r#"{"text": "hello"}"#.into(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "hello"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::ToolUse));+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::ToolUse),+ );// Request that the turn end at the next boundary (a "steering" queued message)thread.update(cx, |thread, _cx| {@@ -8539,7 +8838,7 @@ async fn test_queued_message_ends_turn_at_boundary(cx: &mut TestAppContext) {});// Now end the stream - tool will run, and the boundary check should see the queue- fake_model.end_last_completion_stream();+ fake.end_last(&model);// Collect all events until the turn stopslet all_events = collect_events_until_stop(&mut events, cx).await;@@ -8590,8 +8889,12 @@ async fn test_queued_message_does_not_end_turn_without_boundary_flag(cx: &mut Teinit_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| {thread.add_tool(EchoTool);@@ -8604,28 +8907,31 @@ async fn test_queued_message_does_not_end_turn_without_boundary_flag(cx: &mut Te.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "echo".into(),raw_input: r#"{"text": "hello"}"#.into(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "hello"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::ToolUse));+ }),+ );+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::ToolUse),+ );// Default behavior: even though a message is conceptually queued, we do NOT// set the boundary flag, so the agent must keep going past the tool boundary// (running to completion) rather than ending the turn early.- fake_model.end_last_completion_stream();+ fake.end_last(&model);cx.run_until_parked();// The agent should have issued a fresh completion request with the tool// results instead of stopping — proof it continued past the boundary.- let continuation = fake_model.pending_completions();+ let continuation = fake.pending_completions();assert_eq!(continuation.len(),1,@@ -8633,10 +8939,12 @@ async fn test_queued_message_does_not_end_turn_without_boundary_flag(cx: &mut Te);// Let the continuation finish the turn naturally.- fake_model.send_last_completion_stream_text_chunk("All done");- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::Stop(StopReason::EndTurn));- fake_model.end_last_completion_stream();+ fake.send_last_text(&model, "All done");+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::Stop(StopReason::EndTurn),+ );+ fake.end_last(&model);let all_events = collect_events_until_stop(&mut events, cx).await;let stop_reasons = stop_events(all_events);@@ -8652,8 +8960,12 @@ async fn test_streaming_tool_error_breaks_stream_loop_immediately(cx: &mut TestAinit_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;thread.update(cx, |thread, _cx| {thread.add_tool(StreamingFailingEchoTool {@@ -8681,12 +8993,14 @@ async fn test_streaming_tool_error_breaks_stream_loop_immediately(cx: &mut TestAthought_signature: None,};- fake_model- .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(tool_use.clone()));+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(tool_use.clone()),+ );cx.run_until_parked();- let completions = fake_model.pending_completions();+ let completions = fake.pending_completions();let last_completion = completions.last().unwrap();assert_eq!(@@ -8727,8 +9041,12 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Teinit_test(cx);always_allow_tools(cx);- let ThreadTest { model, thread, .. } = setup(cx, TestModel::Fake).await;- let fake_model = model.as_fake();+ let ThreadTest {+ model,+ fake,+ thread,+ ..+ } = setup(cx, TestModel::Fake).await;let (complete_streaming_echo_tool_call_tx, complete_streaming_echo_tool_call_rx) =oneshot::channel();@@ -8753,16 +9071,17 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Te.unwrap();cx.run_until_parked();- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "call_1".into(),name: StreamingEchoTool::NAME.into(),raw_input: "hello".into(),input: language_model::LanguageModelToolUseInput::Json(json!({ "text": "hello" })),is_input_complete: false,thought_signature: None,- },- ));+ }),+ );let first_tool_use = LanguageModelToolUse {id: "call_1".into(),name: StreamingEchoTool::NAME.into(),@@ -8771,9 +9090,10 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Teis_input_complete: true,thought_signature: None,};- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- first_tool_use.clone(),- ));+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(first_tool_use.clone()),+ );let second_tool_use = LanguageModelToolUse {name: StreamingFailingEchoTool::NAME.into(),raw_input: "hello".into(),@@ -8782,9 +9102,10 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Tethought_signature: None,id: "call_2".into(),};- fake_model.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- second_tool_use.clone(),- ));+ fake.send_last_event(+ &model,+ LanguageModelCompletionEvent::ToolUse(second_tool_use.clone()),+ );cx.run_until_parked();@@ -8792,7 +9113,7 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Tecx.run_until_parked();- let completions = fake_model.pending_completions();+ let completions = fake.pending_completions();let last_completion = completions.last().unwrap();assert_eq!(@@ -8843,9 +9164,12 @@ async fn test_streaming_tool_error_waits_for_prior_tools_to_complete(cx: &mut Te#[gpui::test]async fn test_mid_turn_model_and_settings_refresh(cx: &mut TestAppContext) {let ThreadTest {- model, thread, fs, ..+ model: model_a,+ fake,+ thread,+ fs,+ ..} = setup(cx, TestModel::Fake).await;- let fake_model_a = model.as_fake();thread.update(cx, |thread, _cx| {thread.add_tool(EchoTool);@@ -8894,36 +9218,35 @@ async fn test_mid_turn_model_and_settings_refresh(cx: &mut TestAppContext) {cx.run_until_parked();// Verify first request has both tools and thinking disabled.- let completions = fake_model_a.pending_completions();+ let completions = fake.pending_completions_for(&model_a);assert_eq!(completions.len(), 1);let first_tools = tool_names_for_completion(&completions[0]);assert_eq!(first_tools, vec![DelayTool::NAME, EchoTool::NAME]);assert!(!completions[0].thinking_allowed);// Model A responds with an echo tool call.- fake_model_a.send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {+ fake.send_last_event(+ &model_a,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {id: "tool_1".into(),name: "echo".into(),raw_input: r#"{"text":"hello"}"#.into(),input: language_model::LanguageModelToolUseInput::Json(json!({"text": "hello"})),is_input_complete: true,thought_signature: None,- },- ));- fake_model_a.end_last_completion_stream();+ }),+ );+ fake.end_last(&model_a);// Before the next iteration runs, switch to profile-b (only DelayTool),// swap in a new model, and enable thinking.- let fake_model_b = Arc::new(FakeLanguageModel::with_id_and_thinking(- "test-provider",- "model-b",- "Model B",- true,- ));+ let model_b = fake.update_model("model-b", |model| {+ model.name = LanguageModelName::from("Model B".to_string());+ model.supports_thinking = true;+ });thread.update(cx, |thread, cx| {thread.set_profile(AgentProfileId("profile-b".into()), cx);- thread.set_model(fake_model_b.clone() as Arc<dyn LanguageModel>, cx);+ thread.set_model(model_b.clone(), cx);thread.set_thinking_enabled(true, cx);});@@ -8932,7 +9255,7 @@ async fn test_mid_turn_model_and_settings_refresh(cx: &mut TestAppContext) {cx.run_until_parked();// The second request should have gone to model B.- let model_b_completions = fake_model_b.pending_completions();+ let model_b_completions = fake.pending_completions_for(&model_b);assert_eq!(model_b_completions.len(),1,@@ -8956,7 +9279,9 @@ struct SubagentCompactionTest {thread: Entity<Thread>,environment: NativeThreadEnvironment,handle: Rc<dyn SubagentHandle>,- model: Arc<FakeLanguageModel>,+ fake: Arc<FakeLanguageModelProvider>,+ /// The model the subagent thread sends to.+ model: LanguageModel,}impl SubagentCompactionTest {@@ -8965,9 +9290,8 @@ impl SubagentCompactionTest {}async fn new_with_files(files: serde_json::Value, cx: &mut TestAppContext) -> Self {- init_test(cx);+ let fake = init_test(cx);cx.update(|cx| {- LanguageModelRegistry::test(cx);let mut settings = AgentSettings::get_global(cx).clone();settings.auto_compact.enabled = true;settings.auto_compact.threshold = AutoCompactThreshold::Percentage(0.9);@@ -8992,7 +9316,7 @@ impl SubagentCompactionTest {let parent = agent.read_with(cx, |agent, _| {agent.sessions.get(&session_id).unwrap().thread.clone()});- let model = Arc::new(FakeLanguageModel::default());+ let model = fake.model("fake");parent.update(cx, |thread, cx| thread.set_model(model.clone(), cx));let environment = NativeThreadEnvironment {agent: agent.downgrade(),@@ -9014,6 +9338,7 @@ impl SubagentCompactionTest {thread,environment,handle,+ fake,model,}}@@ -9030,8 +9355,13 @@ impl SubagentCompactionTest {settings.auto_compact.enabled = enabled;AgentSettings::override_global(settings, cx);});- self.model.set_max_token_count(max_tokens);- self.model.set_max_output_tokens(max_output_tokens);+ let model = self.fake.update_model("fake", |model| {+ model.max_token_count = max_tokens;+ model.max_input_tokens = max_tokens;+ model.max_output_tokens = max_output_tokens;+ });+ self.thread+ .update(cx, |thread, cx| thread.set_model(model, cx));}fn send(&self, message: &str, cx: &mut TestAppContext) -> Task<Result<String>> {@@ -9041,40 +9371,39 @@ impl SubagentCompactionTest {}fn update_usage(&self, input_tokens: u64, cx: &mut TestAppContext) {- self.model- .send_last_completion_stream_event(LanguageModelCompletionEvent::UsageUpdate(- TokenUsage {- input_tokens,- ..TokenUsage::default()- },- ));+ self.fake.send_last_event(+ &self.model,+ LanguageModelCompletionEvent::UsageUpdate(TokenUsage {+ input_tokens,+ ..TokenUsage::default()+ }),+ );cx.run_until_parked();}fn tool_round(&self, input_tokens: u64, cx: &mut TestAppContext) {- self.model- .send_last_completion_stream_text_chunk("partial work");+ self.fake.send_last_text(&self.model, "partial work");self.update_usage(input_tokens, cx);self.assert_running(cx);- self.model- .send_last_completion_stream_event(LanguageModelCompletionEvent::ToolUse(- LanguageModelToolUse {- id: "echo_1".into(),- name: EchoTool::NAME.into(),- raw_input: r#"{"text":"tool output"}"#.to_string(),- input: language_model::LanguageModelToolUseInput::Json(- json!({"text": "tool output"}),- ),- is_input_complete: true,- thought_signature: None,- },- ));- self.model.end_last_completion_stream();+ self.fake.send_last_event(+ &self.model,+ LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {+ id: "echo_1".into(),+ name: EchoTool::NAME.into(),+ raw_input: r#"{"text":"tool output"}"#.to_string(),+ input: language_model::LanguageModelToolUseInput::Json(+ json!({"text": "tool output"}),+ ),+ is_input_complete: true,+ thought_signature: None,+ }),+ );+ self.fake.end_last(&self.model);cx.run_until_parked();}fn compaction_request(&self) -> LanguageModelRequest {- let mut requests = self.model.pending_completions();+ let mut requests = self.fake.pending_completions();assert_eq!(requests.len(), 1);let request = requests.pop().unwrap();assert_eq!(@@ -9091,9 +9420,9 @@ impl SubagentCompactionTest {cx: &mut TestAppContext,) {let request = self.compaction_request();- self.model- .send_completion_stream_text_chunk(&request, "failed summary");- self.model.send_completion_stream_error(+ self.fake.send_text(&self.model, &request, "failed summary");+ self.fake.send_error(+ &self.model,&request,LanguageModelCompletionError::from_http_status(LanguageModelProviderName::new("test"),@@ -9102,7 +9431,7 @@ impl SubagentCompactionTest {retry_after,),);- self.model.end_completion_stream(&request);+ self.fake.end_stream(&self.model, &request);cx.run_until_parked();}diff --git a/crates/agent/src/thread.rs b/crates/agent/src/thread.rsindex f8dbde992b2..2969b1beb1a 100644--- a/crates/agent/src/thread.rs+++ b/crates/agent/src/thread.rs@@ -1239,13 +1239,13 @@ enum CompletionError {}pub enum ThreadModel {- Ready(Arc<dyn LanguageModel>),+ Ready(LanguageModel),Unresolved(SelectedModel),Unset,}impl ThreadModel {- fn as_model(&self) -> Option<&Arc<dyn LanguageModel>> {+ fn as_model(&self) -> Option<&LanguageModel> {match self {Self::Ready(model) => Some(model),Self::Unresolved(_) | Self::Unset => None,@@ -1307,7 +1307,7 @@ pub struct Thread {project_context: Entity<ProjectContext>,pub(crate) templates: Arc<Templates>,model: ThreadModel,- summarization_model: Option<Arc<dyn LanguageModel>>,+ summarization_model: Option<LanguageModel>,thinking_enabled: bool,thinking_effort: Option<String>,speed: Option<Speed>,@@ -1332,7 +1332,7 @@ pub struct Thread {}impl Thread {- fn prompt_capabilities(model: Option<&dyn LanguageModel>) -> acp::PromptCapabilities {+ fn prompt_capabilities(model: Option<&LanguageModel>) -> acp::PromptCapabilities {let image = model.map_or(true, |model| model.supports_images());acp::PromptCapabilities::new().image(image)@@ -1381,7 +1381,7 @@ impl Thread {project_context: Entity<ProjectContext>,context_server_registry: Entity<ContextServerRegistry>,templates: Arc<Templates>,- model: Option<Arc<dyn LanguageModel>>,+ model: Option<LanguageModel>,cx: &mut Context<Self>,) -> Self {Self::new_internal(@@ -1400,7 +1400,7 @@ impl Thread {project_context: Entity<ProjectContext>,context_server_registry: Entity<ContextServerRegistry>,templates: Arc<Templates>,- model: Option<Arc<dyn LanguageModel>>,+ model: Option<LanguageModel>,action_log: Entity<ActionLog>,cx: &mut Context<Self>,) -> Self {@@ -1420,7 +1420,7 @@ impl Thread {.as_ref().and_then(|model| model.speed);let (prompt_capabilities_tx, prompt_capabilities_rx) =- watch::channel(Self::prompt_capabilities(model.as_deref()));+ watch::channel(Self::prompt_capabilities(model.as_ref()));let model = match model {Some(model) => ThreadModel::Ready(model),None => Self::user_configured_model_selection(cx)@@ -1508,7 +1508,7 @@ impl Thread {self.thinking_effort = selection.effort.clone();self.speed = selection.speed.filter(|_| model.supports_fast_mode());self.prompt_capabilities_tx- .send(Self::prompt_capabilities(Some(model.as_ref())))+ .send(Self::prompt_capabilities(Some(&model))).log_err();self.model = ThreadModel::Ready(model);}@@ -1784,7 +1784,6 @@ impl Thread {saved_selection.as_ref().and_then(|selection| registry.select_model(selection, cx))- .map(|configured| configured.model)});let model = match (resolved_saved_model, saved_selection) {@@ -1792,16 +1791,14 @@ impl Thread {(None, Some(selection)) => ThreadModel::Unresolved(selection),(None, None) => Self::resolve_profile_model(&profile_id, cx).or_else(|| {- LanguageModelRegistry::global(cx).update(cx, |registry, _cx| {- registry.default_model().map(|model| model.model)- })+ LanguageModelRegistry::global(cx)+ .update(cx, |registry, _cx| registry.default_model())}).map_or(ThreadModel::Unset, ThreadModel::Ready),};- let (prompt_capabilities_tx, prompt_capabilities_rx) = watch::channel(- Self::prompt_capabilities(model.as_model().map(|model| model.as_ref())),- );+ let (prompt_capabilities_tx, prompt_capabilities_rx) =+ watch::channel(Self::prompt_capabilities(model.as_model()));let action_log = cx.new(|_| ActionLog::new(project.clone()));@@ -2009,7 +2006,7 @@ impl Thread {self.ui_scroll_position = position;}- pub fn model(&self) -> Option<&Arc<dyn LanguageModel>> {+ pub fn model(&self) -> Option<&LanguageModel> {self.model.as_model()}@@ -2019,18 +2016,13 @@ impl Thread {pub(crate) fn ensure_model(&mut self,- default_model: Option<&Arc<dyn LanguageModel>>,+ default_model: Option<&LanguageModel>,cx: &mut Context<Self>,) {let resolved = match &self.model {ThreadModel::Ready(_) => return,- ThreadModel::Unresolved(selection) => {- LanguageModelRegistry::global(cx).update(cx, |registry, cx| {- registry- .select_model(selection, cx)- .map(|configured| configured.model)- })- }+ ThreadModel::Unresolved(selection) => LanguageModelRegistry::global(cx)+ .update(cx, |registry, cx| registry.select_model(selection, cx)),ThreadModel::Unset => default_model.cloned(),};@@ -2039,10 +2031,10 @@ impl Thread {}}- pub fn set_model(&mut self, model: Arc<dyn LanguageModel>, cx: &mut Context<Self>) {+ pub fn set_model(&mut self, model: LanguageModel, cx: &mut Context<Self>) {let old_usage = self.latest_token_usage();self.model = ThreadModel::Ready(model.clone());- let new_caps = Self::prompt_capabilities(self.model.as_model().map(|model| model.as_ref()));+ let new_caps = Self::prompt_capabilities(self.model.as_model());let new_usage = self.latest_token_usage();if old_usage != new_usage {cx.emit(TokenUsageUpdated(new_usage));@@ -2063,13 +2055,37 @@ impl Thread {cx.notify()}- pub fn summarization_model(&self) -> Option<&Arc<dyn LanguageModel>> {+ /// Re-selects this thread's model from its provider when that provider's+ /// state changes, so names and capabilities that update after a model is+ /// created (e.g. a local model finishing loading) reach the thread.+ pub fn refresh_model(&mut self, provider_id: &LanguageModelProviderId, cx: &mut Context<Self>) {+ let Some(current) = self.model() else {+ return;+ };+ if ¤t.provider_id != provider_id {+ return;+ }+ let Some(model) = LanguageModelRegistry::read_global(cx)+ .provider(provider_id)+ .into_iter()+ .flat_map(|provider| provider.provided_models(cx))+ .find(|model| model.id == current.id)+ else {+ return;+ };+ if model == *current {+ return;+ }+ self.set_model(model, cx);+ }++ pub fn summarization_model(&self) -> Option<&LanguageModel> {self.summarization_model.as_ref()}pub fn set_summarization_model(&mut self,- model: Option<Arc<dyn LanguageModel>>,+ model: Option<LanguageModel>,cx: &mut Context<Self>,) {self.summarization_model = model.clone();@@ -2495,7 +2511,7 @@ impl Thread {fn resolve_profile_model(profile_id: &AgentProfileId,cx: &mut Context<Self>,- ) -> Option<Arc<dyn LanguageModel>> {+ ) -> Option<LanguageModel> {let selection = AgentSettings::get_global(cx).profiles.get(profile_id)?@@ -2522,16 +2538,13 @@ impl Thread {fn resolve_model_from_selection(selection: &LanguageModelSelection,cx: &mut Context<Self>,- ) -> Option<Arc<dyn LanguageModel>> {+ ) -> Option<LanguageModel> {let selected = SelectedModel {provider: LanguageModelProviderId::from(selection.provider.0.clone()),model: LanguageModelId::from(selection.model.clone()),};- LanguageModelRegistry::global(cx).update(cx, |registry, cx| {- registry- .select_model(&selected, cx)- .map(|configured| configured.model)- })+ LanguageModelRegistry::global(cx)+ .update(cx, |registry, cx| registry.select_model(&selected, cx))}pub fn resume(@@ -2786,7 +2799,7 @@ impl Thread {let mut attempt = 0;let mut intent = CompletionIntent::UserPrompt;// Set when a refusal fallback occurs so subsequent iterations use the fallback model.- let mut refusal_fallback_model: Option<Arc<dyn LanguageModel>> = None;+ let mut refusal_fallback_model: Option<LanguageModel> = None;loop {match Self::perform_compaction_if_needed(this,@@ -2869,15 +2882,16 @@ impl Thread {// mid-turn changes (e.g. the user switches model, toggles tools,// or changes profile) take effect between tool-call rounds.// If a refusal fallback is active, use that model instead.- let (model, request) = this.update(cx, |this, cx| {+ let (model, provider, request) = this.update(cx, |this, cx| {let model = refusal_fallback_model.clone().or_else(|| this.model().cloned()).ok_or_else(|| anyhow!(NoModelConfiguredError))?;+ let provider = LanguageModelRegistry::read_global(cx).provider_for_model(&model);this.refresh_turn_tools(cx);let request = this.build_completion_request(intent, cx)?;this.current_request_token_usage = TokenUsage::default();- anyhow::Ok((model, request))+ anyhow::Ok((model, provider, request))})??;telemetry::event!(@@ -2894,7 +2908,11 @@ impl Thread {log::debug!("Calling model.stream_completion, attempt {}", attempt);- let (mut events, mut error) = match model.stream_completion(request, cx).await {+ let events = match provider {+ Ok(provider) => provider.stream_completion(&model, request, cx).await,+ Err(error) => Err(error),+ };+ let (mut events, mut error) = match events {Ok(events) => (events.fuse(), None),Err(err) => (stream::empty().boxed().fuse(), Some(err)),};@@ -3019,8 +3037,10 @@ impl Thread {})?;if had_refusal {- let maybe_fallback = this.update(cx, |this, cx| -> Option<Arc<dyn LanguageModel>> {- let current_model = refusal_fallback_model.as_ref().or(this.model())?;+ let maybe_fallback = this.update(cx, |this, cx| -> Option<LanguageModel> {+ let current_model = refusal_fallback_model+ .as_ref()+ .or(this.model())?;let fallback_id = match current_model.refusal_fallback_model_id() {Some(id) => id,None => {@@ -3199,7 +3219,7 @@ impl Thread {this: &WeakEntity<Self>,event_stream: &ThreadEventStream,mut cancellation_rx: watch::Receiver<bool>,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,request: LanguageModelRequest,insertion: CompactionInsertion,cx: &mut AsyncApp,@@ -3215,8 +3235,10 @@ impl Thread {acp_thread::ContextCompactionStatus::InProgress,);let result: Result<ControlFlow<()>> = async {+ let provider =+ cx.update(|cx| LanguageModelRegistry::read_global(cx).provider_for_model(&model))?;let stream = futures::select! {- result = model.stream_completion(request, cx).fuse() => result,+ result = provider.stream_completion(&model, request, cx).fuse() => result,_ = cancellation_rx.changed().fuse() => {if *cancellation_rx.borrow() {log::debug!("Compaction cancelled before request started");@@ -3911,7 +3933,13 @@ impl Thread {let task = cx.spawn(async move |this, cx| {let mut summary = String::new();- let mut messages = model.stream_completion(request, cx).await.log_err()?;+ let provider = cx+ .update(|cx| LanguageModelRegistry::read_global(cx).provider_for_model(&model))+ .log_err()?;+ let mut messages = provider+ .stream_completion(&model, request, cx)+ .await+ .log_err()?;while let Some(event) = messages.next().await {let event = event.log_err()?;let text = match event {@@ -3974,7 +4002,7 @@ impl Thread {fn spawn_title_generation(&mut self,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,on_generated_title: Option<Box<dyn FnOnce(SharedString, &mut Context<Self>)>>,cx: &mut Context<Self>,) {@@ -4423,7 +4451,7 @@ impl Thread {fn build_compaction_telemetry(&self,trigger: &'static str,- compaction_model: &Arc<dyn LanguageModel>,+ compaction_model: &LanguageModel,cx: &App,) -> Option<CompactionTelemetry> {let model = self.model()?;@@ -4438,8 +4466,8 @@ impl Thread {thread_id: self.id.to_string(),parent_thread_id: self.parent_thread_id().map(|id| id.to_string()),prompt_id: self.prompt_id.to_string(),- model: model.telemetry_id(),- compaction_model: compaction_model.telemetry_id(),+ model: model.telemetry_id.to_string(),+ compaction_model: compaction_model.telemetry_id.to_string(),thinking_effort: self.thinking_effort.clone(),max_tokens,tokens_before,@@ -4524,24 +4552,23 @@ impl Thread {Some(self.messages.len())}- fn compaction_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn compaction_model(&self, cx: &App) -> Option<LanguageModel> {LanguageModelRegistry::read_global(cx).compaction_model()- .map(|m| m.model).or_else(|| self.model().cloned())}fn build_compaction_request(&self,insertion_ix: usize,- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,cx: &App,) -> LanguageModelRequest {let mut request = LanguageModelRequest {thread_id: Some(self.id.to_string()),prompt_id: Some(self.prompt_id.to_string()),intent: Some(CompletionIntent::ThreadContextSummarization),- temperature: AgentSettings::temperature_for_model(model, cx),+ temperature: AgentSettings::temperature_for_model(&model, cx),messages: self.build_request_messages_until(Vec::new(), insertion_ix, cx),..Default::default()};@@ -4612,6 +4639,8 @@ impl Thread {// Retrying won't help until the user consents to data retention// or switches models.DataRetentionConsentRequired { .. } => None,+ // Retrying won't help until the thread picks another model.+ ModelUnavailable { .. } => None,// `Other` includes mid-stream mapping failures that can be caused by// a transient malformed or interrupted provider event.Other(..) => Some(RetryStrategy::FixedDelay {@@ -4984,12 +5013,14 @@ pub fn build_thread_title_request(}pub async fn stream_thread_title(- model: Arc<dyn LanguageModel>,+ model: LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,) -> Result<String> {let mut title = String::new();- let mut events = model.stream_completion(request, cx).await?;+ let provider =+ cx.update(|cx| LanguageModelRegistry::read_global(cx).provider_for_model(&model))?;+ let mut events = provider.stream_completion(&model, request, cx).await?;while let Some(event) = events.next().await {let LanguageModelCompletionEvent::Text(text) = event? else {continue;@@ -6962,7 +6993,7 @@ mod tests {use super::*;use gpui::TestAppContext;use language_model::LanguageModelToolUseId;- use language_model::fake_provider::FakeLanguageModel;+ use language_model::fake_provider::FakeLanguageModelProvider;use serde_json::json;use settings::LanguageModelProviderSetting;use std::sync::Arc;@@ -6991,12 +7022,18 @@ mod tests {);}- async fn setup_thread_for_test(cx: &mut TestAppContext) -> (Entity<Thread>, ThreadEventStream) {- cx.update(|cx| {+ async fn setup_thread_for_test(+ cx: &mut TestAppContext,+ ) -> (+ Entity<Thread>,+ ThreadEventStream,+ Arc<FakeLanguageModelProvider>,+ ) {+ let fake = cx.update(|cx| {let settings_store = settings::SettingsStore::test(cx);cx.set_global(settings_store);- LanguageModelRegistry::test(cx);+ LanguageModelRegistry::test(cx)});let fs = fs::FakeFs::new(cx.background_executor.clone());@@ -7023,7 +7060,7 @@ mod tests {let (event_tx, _event_rx) = mpsc::unbounded();let event_stream = ThreadEventStream::new(event_tx);- (thread, event_stream)+ (thread, event_stream, fake)})}@@ -7033,16 +7070,9 @@ mod tests {AgentSettings::override_global(settings, cx);}- fn set_registry_compaction_model(cx: &mut App, model: Option<Arc<dyn LanguageModel>>) {- use language_model::fake_provider::FakeLanguageModelProvider;- use language_model::{ConfiguredModel, LanguageModelProvider};+ fn set_registry_compaction_model(cx: &mut App, model: Option<LanguageModel>) {LanguageModelRegistry::global(cx).update(cx, |registry, cx| {- let configured = model.map(|m| ConfiguredModel {- provider: Arc::new(FakeLanguageModelProvider::default())- as Arc<dyn LanguageModelProvider>,- model: m,- });- registry.set_compaction_model(configured, cx);+ registry.set_compaction_model(model, cx);});}@@ -7233,9 +7263,9 @@ mod tests {#[gpui::test]async fn test_thread_summary_request_uses_compacted_history(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;let thread_id = thread.read_with(cx, |thread, _| thread.id().to_string());- let summary_model = Arc::new(FakeLanguageModel::default());+ let summary_model = fake.model("summary");let summary_task = cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -7263,7 +7293,7 @@ mod tests {});cx.run_until_parked();- let summary_request = summary_model.pending_completions().pop().unwrap();+ let summary_request = fake.pending_completions().pop().unwrap();assert_eq!(summary_request.thread_id.as_deref(),Some(thread_id.as_str())@@ -7284,8 +7314,8 @@ mod tests {]);- summary_model.send_completion_stream_text_chunk(&summary_request, "thread summary");- summary_model.end_completion_stream(&summary_request);+ fake.send_text(&summary_model, &summary_request, "thread summary");+ fake.end_stream(&summary_model, &summary_request);assert_eq!(summary_task.await.as_deref(), Some("thread summary"));}@@ -7323,13 +7353,13 @@ mod tests {#[gpui::test]async fn test_compaction_threshold_uses_percentage_setting(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");let user_message_id = ClientUserMessageId::new();cx.update(|cx| {thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(), "below limit"));@@ -7358,14 +7388,13 @@ mod tests {#[gpui::test]async fn test_compaction_threshold_accounts_for_max_output_tokens(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());- model.set_max_output_tokens(Some(32_000));+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.update_model("fake", |model| model.max_output_tokens = Some(32_000));let user_message_id = ClientUserMessageId::new();cx.update(|cx| {thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(),"near input limit",@@ -7422,12 +7451,12 @@ mod tests {#[gpui::test]async fn test_compaction_threshold_respects_independent_input_limit(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let mut model = FakeLanguageModel::default();- model.set_max_token_count(200_000);- model.set_max_input_tokens(90_000);- model.set_max_output_tokens(Some(16_384));- let model = Arc::new(model);+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.update_model("fake", |model| {+ model.max_token_count = 200_000;+ model.max_input_tokens = 90_000;+ model.max_output_tokens = Some(16_384);+ });let user_message_id = ClientUserMessageId::new();cx.update(|cx| {@@ -7439,7 +7468,7 @@ mod tests {},);thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(),"near the independent input limit",@@ -7463,8 +7492,8 @@ mod tests {#[gpui::test]async fn test_compaction_threshold_respects_enabled_setting(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");let user_message_id = ClientUserMessageId::new();cx.update(|cx| {@@ -7476,7 +7505,7 @@ mod tests {},);thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(), "near limit"));@@ -7495,8 +7524,8 @@ mod tests {#[gpui::test]async fn test_compaction_threshold_respects_token_settings(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");let user_message_id = ClientUserMessageId::new();cx.update(|cx| {@@ -7508,7 +7537,7 @@ mod tests {},);thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(),"fixed token limit",@@ -7565,15 +7594,17 @@ mod tests {#[gpui::test]async fn test_compaction_unavailable_for_small_context_window(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;// A context window below the minimum disables auto-compaction.- model.set_max_token_count(MIN_COMPACTION_CONTEXT_WINDOW - 1);+ let model = fake.update_model("fake", |model| {+ model.max_token_count = MIN_COMPACTION_CONTEXT_WINDOW - 1;+ model.max_input_tokens = MIN_COMPACTION_CONTEXT_WINDOW - 1;+ });let user_message_id = ClientUserMessageId::new();cx.update(|cx| {thread.update(cx, |thread, cx| {- thread.set_model(model, cx);+ thread.set_model(model.clone(), cx);thread.messages.push(user_text_message(user_message_id.clone(), "near limit"));@@ -7594,8 +7625,8 @@ mod tests {async fn test_compaction_inserts_before_new_user_and_requests_compacted_window(cx: &mut TestAppContext,) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");let old_user_message_id = ClientUserMessageId::new();let new_user_message_id = ClientUserMessageId::new();@@ -7625,7 +7656,7 @@ mod tests {.unwrap();cx.run_until_parked();- let compaction_request = model.pending_completions().pop().unwrap();+ let compaction_request = fake.pending_completions().pop().unwrap();assert_eq!(compaction_request.intent,Some(CompletionIntent::ThreadContextSummarization)@@ -7636,11 +7667,11 @@ mod tests {assert_eq!(compaction_texts[1], "old assistant");assert_eq!(compaction_texts[2], COMPACTION_PROMPT);- model.send_completion_stream_text_chunk(&compaction_request, "compacted old context");- model.end_completion_stream(&compaction_request);+ fake.send_text(&model, &compaction_request, "compacted old context");+ fake.end_stream(&model, &compaction_request);cx.run_until_parked();- let final_request = model.pending_completions().pop().unwrap();+ let final_request = fake.pending_completions().pop().unwrap();assert_eq!(final_request.intent, Some(CompletionIntent::UserPrompt));assert_eq!(request_texts_after_system(&final_request.messages),@@ -7651,8 +7682,8 @@ mod tests {]);- model.send_completion_stream_text_chunk(&final_request, "answer");- model.end_completion_stream(&final_request);+ fake.send_text(&model, &final_request, "answer");+ fake.end_stream(&model, &final_request);cx.run_until_parked();cx.update(|cx| {@@ -7670,11 +7701,13 @@ mod tests {#[gpui::test]async fn test_manual_compact_forces_summary(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;// A context window below the minimum and no recorded token usage would// both disable *automatic* compaction. Manual compaction forces it anyway.- model.set_max_token_count(MIN_COMPACTION_CONTEXT_WINDOW - 1);+ let model = fake.update_model("fake", |model| {+ model.max_token_count = MIN_COMPACTION_CONTEXT_WINDOW - 1;+ model.max_input_tokens = MIN_COMPACTION_CONTEXT_WINDOW - 1;+ });let user_message_id = ClientUserMessageId::new();let compact_message_id = ClientUserMessageId::new();@@ -7699,7 +7732,7 @@ mod tests {.unwrap();cx.run_until_parked();- let compaction_request = model.pending_completions().pop().unwrap();+ let compaction_request = fake.pending_completions().pop().unwrap();assert_eq!(compaction_request.intent,Some(CompletionIntent::ThreadContextSummarization)@@ -7710,14 +7743,14 @@ mod tests {assert_eq!(compaction_texts[1], "old assistant");assert_eq!(compaction_texts[2], COMPACTION_PROMPT);- model.send_completion_stream_text_chunk(&compaction_request, "summary of old context");- model.end_completion_stream(&compaction_request);+ fake.send_text(&model, &compaction_request, "summary of old context");+ fake.end_stream(&model, &compaction_request);cx.run_until_parked();// The compaction summary is appended after a zero-content user message// marker, and no follow-up model turn is requested — `/compact` only// compacts.- assert!(model.pending_completions().is_empty());+ assert!(fake.pending_completions().is_empty());cx.update(|cx| {thread.read_with(cx, |thread, _cx| {assert!(matches!(&*thread.messages[0], Message::User(_)));@@ -7751,8 +7784,8 @@ mod tests {/// rewind marker (or a partial summary) dangling at the end of the thread.#[gpui::test]async fn test_manual_compact_cancelled_leaves_no_marker(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -7773,7 +7806,7 @@ mod tests {.unwrap();cx.run_until_parked();// The compaction request is in flight but hasn't streamed a summary.- assert_eq!(model.pending_completions().len(), 1);+ assert_eq!(fake.pending_completions().len(), 1);cx.update(|cx| thread.update(cx, |thread, cx| thread.cancel(cx))).await;@@ -7790,8 +7823,8 @@ mod tests {/// the thread untouched — no marker, no compaction.#[gpui::test]async fn test_manual_compact_empty_summary_leaves_no_marker(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -7812,9 +7845,9 @@ mod tests {.unwrap();cx.run_until_parked();- let request = model.pending_completions().pop().unwrap();+ let request = fake.pending_completions().pop().unwrap();// End the stream without emitting any summary text.- model.end_completion_stream(&request);+ fake.end_stream(&model, &request);cx.run_until_parked();// An error is surfaced, and the thread is left exactly as it was. The@@ -7838,8 +7871,8 @@ mod tests {/// issues no model request and adds no marker.#[gpui::test]async fn test_manual_compact_noop_on_empty_thread(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");cx.update(|cx| thread.update(cx, |thread, cx| thread.set_model(model.clone(), cx)));let _events = cx@@ -7851,7 +7884,7 @@ mod tests {.unwrap();cx.run_until_parked();- assert!(model.pending_completions().is_empty());+ assert!(fake.pending_completions().is_empty());thread.read_with(cx, |thread, _cx| {assert!(thread.messages.is_empty());});@@ -7862,7 +7895,7 @@ mod tests {/// a compacted thread doesn't surface an empty `/compact` bubble.#[gpui::test]async fn test_manual_compact_marker_replays_as_empty_user_message(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, _) = setup_thread_for_test(cx).await;let marker_id = ClientUserMessageId::new();let mut replay_events = cx.update(|cx| {@@ -7907,9 +7940,9 @@ mod tests {/// to the configured model rather than the thread's primary model.#[gpui::test]async fn test_compaction_uses_configured_compaction_model(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let thread_model = Arc::new(FakeLanguageModel::default());- let compaction_model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let thread_model = fake.model("fake");+ let compaction_model = fake.model("compaction");cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -7919,7 +7952,8 @@ mod tests {.push(user_text_message(ClientUserMessageId::new(), "old user"));thread.messages.push(agent_text_message("old assistant"));});- set_registry_compaction_model(cx, Some(compaction_model.clone()));+ let compaction_model = compaction_model.clone();+ set_registry_compaction_model(cx, Some(compaction_model));});let _events = cx@@ -7932,12 +7966,12 @@ mod tests {cx.run_until_parked();assert_eq!(- thread_model.pending_completions().len(),+ fake.pending_completions_for(&thread_model).len(),0,"thread's primary model should not have been used for compaction");- let request = compaction_model- .pending_completions()+ let request = fake+ .pending_completions_for(&compaction_model).pop().expect("compaction model should have received the request");assert_eq!(@@ -7953,8 +7987,8 @@ mod tests {assert_eq!(telemetry.model, compaction_model.telemetry_id());});- compaction_model.send_completion_stream_text_chunk(&request, "summary");- compaction_model.end_completion_stream(&request);+ fake.send_text(&compaction_model, &request, "summary");+ fake.end_stream(&compaction_model, &request);cx.run_until_parked();}@@ -7964,8 +7998,8 @@ mod tests {/// model — not the one the user tried to configure.#[gpui::test]async fn test_compaction_falls_back_when_compaction_model_unavailable(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let thread_model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let thread_model = fake.model("fake");cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -7997,8 +8031,8 @@ mod tests {.unwrap();cx.run_until_parked();- let request = thread_model- .pending_completions()+ let request = fake+ .pending_completions_for(&thread_model).pop().expect("thread model should have received the fallback request");assert_eq!(@@ -8014,8 +8048,8 @@ mod tests {assert_eq!(telemetry.model, thread_model.telemetry_id());});- thread_model.send_completion_stream_text_chunk(&request, "summary");- thread_model.end_completion_stream(&request);+ fake.send_text(&thread_model, &request, "summary");+ fake.end_stream(&thread_model, &request);cx.run_until_parked();}@@ -8023,9 +8057,9 @@ mod tests {/// `agent.compaction_model`.#[gpui::test]async fn test_auto_compaction_uses_compaction_model(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let thread_model = Arc::new(FakeLanguageModel::default());- let compaction_model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let thread_model = fake.model("fake");+ let compaction_model = fake.model("compaction");let old_user_message_id = ClientUserMessageId::new();cx.update(|cx| {@@ -8050,7 +8084,8 @@ mod tests {threshold: agent_settings::AutoCompactThreshold::Percentage(0.5),},);- set_registry_compaction_model(cx, Some(compaction_model.clone()));+ let compaction_model = compaction_model.clone();+ set_registry_compaction_model(cx, Some(compaction_model));});// The auto-compact gate fires inside `run_turn` when we kick off a@@ -8065,9 +8100,9 @@ mod tests {.unwrap();cx.run_until_parked();- assert_eq!(thread_model.pending_completions().len(), 0);- let request = compaction_model- .pending_completions()+ assert_eq!(fake.pending_completions_for(&thread_model).len(), 0);+ let request = fake+ .pending_completions_for(&compaction_model).pop().expect("compaction model should have received the auto-compaction request");assert_eq!(@@ -8075,15 +8110,15 @@ mod tests {Some(CompletionIntent::ThreadContextSummarization));- compaction_model.send_completion_stream_text_chunk(&request, "summary");- compaction_model.end_completion_stream(&request);+ fake.send_text(&compaction_model, &request, "summary");+ fake.end_stream(&compaction_model, &request);cx.run_until_parked();}#[gpui::test]async fn test_compaction_usage_counts_toward_cumulative_usage(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;- let model = Arc::new(FakeLanguageModel::default());+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");let old_user_message_id = ClientUserMessageId::new();let new_user_message_id = ClientUserMessageId::new();let prior_usage = TokenUsage {@@ -8127,13 +8162,14 @@ mod tests {.unwrap();cx.run_until_parked();- let compaction_request = model.pending_completions().pop().unwrap();+ let compaction_request = fake.pending_completions().pop().unwrap();assert_eq!(compaction_request.intent,Some(CompletionIntent::ThreadContextSummarization));- model.send_completion_stream_event(+ fake.send_event(+ &model,&compaction_request,LanguageModelCompletionEvent::UsageUpdate(TokenUsage {input_tokens: 40,@@ -8141,12 +8177,13 @@ mod tests {..Default::default()}),);- model.send_completion_stream_event(+ fake.send_event(+ &model,&compaction_request,LanguageModelCompletionEvent::UsageUpdate(compaction_usage),);- model.send_completion_stream_text_chunk(&compaction_request, "compacted old context");- model.end_completion_stream(&compaction_request);+ fake.send_text(&model, &compaction_request, "compacted old context");+ fake.end_stream(&model, &compaction_request);cx.run_until_parked();let expected_after_compaction = prior_usage + compaction_usage;@@ -8159,14 +8196,15 @@ mod tests {);});- let final_request = model.pending_completions().pop().unwrap();+ let final_request = fake.pending_completions().pop().unwrap();assert_eq!(final_request.intent, Some(CompletionIntent::UserPrompt));- model.send_completion_stream_event(+ fake.send_event(+ &model,&final_request,LanguageModelCompletionEvent::UsageUpdate(final_usage),);- model.end_completion_stream(&final_request);+ fake.end_stream(&model, &final_request);cx.run_until_parked();thread.read_with(cx, |thread, _cx| {@@ -8183,7 +8221,7 @@ mod tests {#[gpui::test]async fn test_replay_emits_context_compaction(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, _) = setup_thread_for_test(cx).await;let user_message_id = ClientUserMessageId::new();let mut replay_events = cx.update(|cx| {@@ -8232,7 +8270,7 @@ mod tests {#[gpui::test]async fn test_native_compaction_boundary(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, _) = setup_thread_for_test(cx).await;let request_messages = cx.update(|cx| {thread.update(cx, |thread, cx| {@@ -8263,7 +8301,7 @@ mod tests {#[gpui::test]async fn test_retained_users_truncate_oldest(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, _) = setup_thread_for_test(cx).await;let mut long_text = "START".to_string();long_text.push_str(&"x".repeat(COMPACTION_RETAINED_USER_MESSAGES_BYTE_BUDGET));long_text.push_str("END");@@ -8860,7 +8898,7 @@ mod tests {#[gpui::test]async fn test_replay_tool_call_replays_image_content(cx: &mut TestAppContext) {- let (thread, _event_stream) = setup_thread_for_test(cx).await;+ let (thread, _event_stream, _) = setup_thread_for_test(cx).await;let registered_tool_use_id = LanguageModelToolUseId::from("registered_tool_id");let missing_tool_use_id = LanguageModelToolUseId::from("missing_tool_id");@@ -8976,21 +9014,44 @@ mod tests {);}+ #[gpui::test]+ async fn test_refresh_model_picks_up_provider_capability_changes(cx: &mut TestAppContext) {+ let (thread, _event_stream, fake) = setup_thread_for_test(cx).await;+ let model = fake.model("fake");+ let provider_id = model.provider_id.clone();+ let model_changes = Rc::new(std::cell::Cell::new(0));+ let _subscription = cx.update(|cx| {+ thread.update(cx, |thread, cx| thread.set_model(model, cx));+ let model_changes = model_changes.clone();+ cx.subscribe(&thread, move |_, _: &ModelChanged, _| {+ model_changes.set(model_changes.get() + 1)+ })+ });++ thread.update(cx, |thread, cx| thread.refresh_model(&provider_id, cx));+ assert_eq!(model_changes.get(), 0, "an unchanged model is not re-set");++ fake.update_model("fake", |model| model.supports_images = true);+ thread.update(cx, |thread, cx| {+ thread.refresh_model(&LanguageModelProviderId::from("other".to_string()), cx);+ assert!(!thread.model().unwrap().supports_images());++ thread.refresh_model(&provider_id, cx);+ assert!(thread.model().unwrap().supports_images());+ });+ assert_eq!(model_changes.get(), 1);+ }+#[gpui::test]async fn test_set_model_propagates_to_subagents(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, fake) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 2);- let new_model: Arc<dyn LanguageModel> = Arc::new(FakeLanguageModel::with_id_and_thinking(- "test-provider",- "new-model",- "New Model",- false,- ));+ let new_model = fake.model("new-model");cx.update(|cx| {parent.update(cx, |thread, cx| {- thread.set_model(new_model, cx);+ thread.set_model(new_model.clone(), cx);});for subagent in &subagents {@@ -9006,20 +9067,14 @@ mod tests {#[gpui::test]async fn test_set_summarization_model_propagates_to_subagents(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, fake) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 2);- let summary_model: Arc<dyn LanguageModel> =- Arc::new(FakeLanguageModel::with_id_and_thinking(- "test-provider",- "summary-model",- "Summary Model",- false,- ));+ let summary_model = fake.model("summary-model");cx.update(|cx| {parent.update(cx, |thread, cx| {- thread.set_summarization_model(Some(summary_model), cx);+ thread.set_summarization_model(Some(summary_model.clone()), cx);});for subagent in &subagents {@@ -9035,7 +9090,7 @@ mod tests {#[gpui::test]async fn test_set_thinking_enabled_propagates_to_subagents(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, _) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 2);cx.update(|cx| {@@ -9065,7 +9120,7 @@ mod tests {#[gpui::test]async fn test_set_thinking_effort_propagates_to_subagents(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, _) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 2);cx.update(|cx| {@@ -9097,7 +9152,7 @@ mod tests {#[gpui::test]async fn test_subagent_inherits_settings_at_creation(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, _) = setup_thread_for_test(cx).await;cx.update(|cx| {parent.update(cx, |thread, cx| {@@ -9121,7 +9176,7 @@ mod tests {#[gpui::test]async fn test_set_speed_propagates_to_subagents(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, _) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 2);cx.update(|cx| {@@ -9141,7 +9196,7 @@ mod tests {#[gpui::test]async fn test_dropped_subagent_does_not_panic(cx: &mut TestAppContext) {- let (parent, _event_stream) = setup_thread_for_test(cx).await;+ let (parent, _event_stream, _) = setup_thread_for_test(cx).await;let subagents = setup_parent_with_subagents(cx, &parent, 1);// Drop the subagent so the WeakEntity can no longer be upgraded@@ -9161,7 +9216,7 @@ mod tests {async fn test_handle_tool_use_json_parse_error_adds_tool_use_to_content(cx: &mut TestAppContext,) {- let (thread, event_stream) = setup_thread_for_test(cx).await;+ let (thread, event_stream, _) = setup_thread_for_test(cx).await;let tool_use_id = LanguageModelToolUseId::from("test_tool_id");let tool_name: Arc<str> = Arc::from("test_tool");diff --git a/crates/agent/src/tools/edit_file_tool.rs b/crates/agent/src/tools/edit_file_tool.rsindex 64f6af0a49c..e645273d799 100644--- a/crates/agent/src/tools/edit_file_tool.rs+++ b/crates/agent/src/tools/edit_file_tool.rs@@ -289,7 +289,7 @@ mod tests {use crate::{ContextServerRegistry, Templates, ToolInputSender};use fs::Fs as _;use gpui::{AppContext as _, TestAppContext, UpdateGlobal};- use language_model::fake_provider::FakeLanguageModel;+ use language_model::LanguageModelRegistry;use project::ProjectPath;use prompt_store::ProjectContext;use serde_json::json;@@ -3089,7 +3089,7 @@ mod tests {let language_registry = project.read_with(cx, |project, _cx| project.languages().clone());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(project.read(cx).context_server_store(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = cx.update(|cx| LanguageModelRegistry::test(cx).model("fake"));let thread = cx.new(|cx| {crate::Thread::new(project.clone(),diff --git a/crates/agent/src/tools/evals/edit_file.rs b/crates/agent/src/tools/evals/edit_file.rsindex 124664fd39d..9ec852b3f9f 100644--- a/crates/agent/src/tools/evals/edit_file.rs+++ b/crates/agent/src/tools/evals/edit_file.rs@@ -75,7 +75,7 @@ trait AssertionFn: 'static + Send + Sync {fn assert<'a>(&'a self,sample: &'a EvalSample,- judge_model: Arc<dyn LanguageModel>,+ judge_model: LanguageModel,cx: &'a mut TestAppContext,) -> LocalBoxFuture<'a, Result<EvalAssertionOutcome>>;}@@ -85,16 +85,12 @@ whereF: 'static+ Send+ Sync- + AsyncFn(- &EvalSample,- Arc<dyn LanguageModel>,- &mut TestAppContext,- ) -> Result<EvalAssertionOutcome>,+ + AsyncFn(&EvalSample, LanguageModel, &mut TestAppContext) -> Result<EvalAssertionOutcome>,{fn assert<'a>(&'a self,sample: &'a EvalSample,- judge_model: Arc<dyn LanguageModel>,+ judge_model: LanguageModel,cx: &'a mut TestAppContext,) -> LocalBoxFuture<'a, Result<EvalAssertionOutcome>> {(self)(sample, judge_model, cx).boxed_local()@@ -110,11 +106,7 @@ impl EvalAssertion {F: 'static+ Send+ Sync- + AsyncFn(- &EvalSample,- Arc<dyn LanguageModel>,- &mut TestAppContext,- ) -> Result<EvalAssertionOutcome>,+ + AsyncFn(&EvalSample, LanguageModel, &mut TestAppContext) -> Result<EvalAssertionOutcome>,{EvalAssertion(Arc::new(f))}@@ -159,9 +151,11 @@ impl EvalAssertion {.map(|effort_level| effort_level.value.to_string()),..Default::default()};+ let provider =+ cx.update(|cx| LanguageModelRegistry::read_global(cx).provider_for_model(&judge))?;let mut response = retry_on_rate_limit(async || {- Ok(judge- .stream_completion_text(request.clone(), &cx.to_async())+ Ok(provider+ .stream_completion_text(&judge, request.clone(), &cx.to_async()).await?)}).await?;@@ -190,7 +184,7 @@ impl EvalAssertion {async fn run(&self,input: &EvalSample,- judge_model: Arc<dyn LanguageModel>,+ judge_model: LanguageModel,cx: &mut TestAppContext,) -> Result<EvalAssertionOutcome> {self.0.assert(input, judge_model, cx).await@@ -224,8 +218,8 @@ struct EvalAssertionOutcome {struct EditToolTest {fs: Arc<FakeFs>,project: Entity<Project>,- model: Arc<dyn LanguageModel>,- judge_model: Arc<dyn LanguageModel>,+ model: LanguageModel,+ judge_model: LanguageModel,model_thinking_effort: Option<String>,}@@ -309,7 +303,7 @@ impl EditToolTest {async fn load_model(selected_model: &SelectedModel,cx: &mut AsyncApp,- ) -> Result<Arc<dyn LanguageModel>> {+ ) -> Result<LanguageModel> {cx.update(|cx| {let registry = LanguageModelRegistry::read_global(cx);let provider = registry@@ -319,13 +313,14 @@ impl EditToolTest {}).await?;Ok(cx.update(|cx| {- let models = LanguageModelRegistry::read_global(cx);- models- .available_models(cx)- .find(|model| {- model.provider_id() == selected_model.provider- && model.id() == selected_model.model- })+ let registry = LanguageModelRegistry::read_global(cx);+ let provider = registry+ .provider(&selected_model.provider)+ .expect("Provider not found");+ provider+ .provided_models(cx)+ .into_iter()+ .find(|model| model.id() == selected_model.model).unwrap_or_else(|| panic!("Model {} not found", selected_model.model.0))}))}@@ -482,9 +477,13 @@ impl EditToolTest {let model = self.model.clone();let events = cx.update(|cx| {+ let provider = LanguageModelRegistry::read_global(cx).provider_for_model(&model);let async_cx = cx.to_async();- cx.foreground_executor()- .spawn(async move { model.stream_completion(request, &async_cx).await })+ cx.foreground_executor().spawn(async move {+ provider?+ .stream_completion(&model, request, &async_cx)+ .await+ })}).await.map_err(|err| anyhow::anyhow!("completion error: {}", err))?;diff --git a/crates/agent/src/tools/evals/terminal_tool.rs b/crates/agent/src/tools/evals/terminal_tool.rsindex 03b6c3a07c2..8e5bcbc49a3 100644--- a/crates/agent/src/tools/evals/terminal_tool.rs+++ b/crates/agent/src/tools/evals/terminal_tool.rs@@ -145,7 +145,7 @@ impl Display for EvalOutput {}struct TerminalToolTest {- model: Arc<dyn LanguageModel>,+ model: LanguageModel,model_thinking_effort: Option<String>,}@@ -274,10 +274,7 @@ impl TerminalToolTest {}}-async fn load_model(- selected_model: &SelectedModel,- cx: &mut AsyncApp,-) -> Result<Arc<dyn LanguageModel>> {+async fn load_model(selected_model: &SelectedModel, cx: &mut AsyncApp) -> Result<LanguageModel> {cx.update(|cx| {let registry = LanguageModelRegistry::read_global(cx);let provider = registry@@ -287,12 +284,14 @@ async fn load_model(}).await?;Ok(cx.update(|cx| {- let models = LanguageModelRegistry::read_global(cx);- models- .available_models(cx)- .find(|model| {- model.provider_id() == selected_model.provider && model.id() == selected_model.model- })+ let registry = LanguageModelRegistry::read_global(cx);+ let provider = registry+ .provider(&selected_model.provider)+ .expect("Provider not found");+ provider+ .provided_models(cx)+ .into_iter()+ .find(|model| model.id() == selected_model.model).unwrap_or_else(|| panic!("Model {} not found", selected_model.model.0))}))}@@ -300,16 +299,20 @@ async fn load_model(/// Stream the model completion and extract the first complete tool use whose/// name matches `TerminalTool::NAME`, parsed as `TerminalToolInput`.async fn extract_tool_use(- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,request: LanguageModelRequest,cx: &mut TestAppContext,) -> Result<TerminalToolInput> {let model = model.clone();let events = cx.update(|cx| {+ let provider = LanguageModelRegistry::read_global(cx).provider_for_model(&model);let async_cx = cx.to_async();- cx.foreground_executor()- .spawn(async move { model.stream_completion(request, &async_cx).await })+ cx.foreground_executor().spawn(async move {+ provider?+ .stream_completion(&model, request, &async_cx)+ .await+ })}).await.map_err(|err| anyhow::anyhow!("completion error: {}", err))?;diff --git a/crates/agent/src/tools/evals/write_file.rs b/crates/agent/src/tools/evals/write_file.rsindex 4c471ab61af..6e93c4a8879 100644--- a/crates/agent/src/tools/evals/write_file.rs+++ b/crates/agent/src/tools/evals/write_file.rs@@ -67,7 +67,7 @@ impl Display for WriteEvalOutput {struct WriteToolTest {fs: Arc<FakeFs>,project: Entity<Project>,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,model_thinking_effort: Option<String>,}@@ -143,7 +143,7 @@ impl WriteToolTest {async fn load_model(selected_model: &SelectedModel,cx: &mut AsyncApp,- ) -> Result<Arc<dyn LanguageModel>> {+ ) -> Result<LanguageModel> {cx.update(|cx| {let registry = LanguageModelRegistry::read_global(cx);let provider = registry@@ -153,13 +153,14 @@ impl WriteToolTest {}).await?;Ok(cx.update(|cx| {- let models = LanguageModelRegistry::read_global(cx);- models- .available_models(cx)- .find(|model| {- model.provider_id() == selected_model.provider- && model.id() == selected_model.model- })+ let registry = LanguageModelRegistry::read_global(cx);+ let provider = registry+ .provider(&selected_model.provider)+ .expect("Provider not found");+ provider+ .provided_models(cx)+ .into_iter()+ .find(|model| model.id() == selected_model.model).unwrap_or_else(|| panic!("Model {} not found", selected_model.model.0))}))}@@ -302,9 +303,13 @@ impl WriteToolTest {let model = self.model.clone();let events = cx.update(|cx| {+ let provider = LanguageModelRegistry::read_global(cx).provider_for_model(&model);let async_cx = cx.to_async();- cx.foreground_executor()- .spawn(async move { model.stream_completion(request, &async_cx).await })+ cx.foreground_executor().spawn(async move {+ provider?+ .stream_completion(&model, request, &async_cx)+ .await+ })}).await.map_err(|err| anyhow::anyhow!("completion error: {}", err))?;diff --git a/crates/agent/src/tools/write_file_tool.rs b/crates/agent/src/tools/write_file_tool.rsindex 59561556947..451a0416aa7 100644--- a/crates/agent/src/tools/write_file_tool.rs+++ b/crates/agent/src/tools/write_file_tool.rs@@ -274,7 +274,7 @@ mod tests {use futures::StreamExt as _;use gpui::{AppContext as _, Entity, TestAppContext, UpdateGlobal};use language::language_settings::FormatOnSave;- use language_model::fake_provider::FakeLanguageModel;+ use language_model::LanguageModelRegistry;use project::{Project, ProjectPath};use prompt_store::ProjectContext;use serde_json::json;@@ -1382,7 +1382,7 @@ mod tests {let language_registry = project.read_with(cx, |project, _cx| project.languages().clone());let context_server_registry =cx.new(|cx| ContextServerRegistry::new(project.read(cx).context_server_store(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = cx.update(|cx| LanguageModelRegistry::test(cx).model("fake"));let thread = cx.new(|cx| {crate::Thread::new(project.clone(),diff --git a/crates/agent_settings/src/agent_settings.rs b/crates/agent_settings/src/agent_settings.rsindex 193cc5a8c0f..56f97f643eb 100644--- a/crates/agent_settings/src/agent_settings.rs+++ b/crates/agent_settings/src/agent_settings.rs@@ -263,7 +263,7 @@ impl AgentSettings {self.enabled && !DisableAiSettings::get_global(cx).disable_ai}- pub fn temperature_for_model(model: &Arc<dyn LanguageModel>, cx: &App) -> Option<f32> {+ pub fn temperature_for_model(model: &LanguageModel, cx: &App) -> Option<f32> {let settings = Self::get_global(cx);for setting in settings.model_parameters.iter().rev() {if let Some(provider) = &setting.provider@@ -301,7 +301,7 @@ impl AgentSettings {}pub fn language_model_to_selection(- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,override_selection: Option<&LanguageModelSelection>,) -> LanguageModelSelection {let provider = model.provider_id().0.to_string().into();diff --git a/crates/agent_ui/src/agent_configuration/manage_profiles_modal.rs b/crates/agent_ui/src/agent_configuration/manage_profiles_modal.rsindex d38db4a0855..314eb7afae2 100644--- a/crates/agent_ui/src/agent_configuration/manage_profiles_modal.rs+++ b/crates/agent_ui/src/agent_configuration/manage_profiles_modal.rs@@ -108,7 +108,7 @@ pub struct NewProfileMode {pub struct ManageProfilesModal {fs: Arc<dyn Fs>,context_server_registry: Entity<ContextServerRegistry>,- active_model: Option<Arc<dyn LanguageModel>>,+ active_model: Option<LanguageModel>,focus_handle: FocusHandle,mode: Mode,_settings_subscription: Subscription,@@ -144,7 +144,7 @@ impl ManageProfilesModal {pub fn new(fs: Arc<dyn Fs>,- active_model: Option<Arc<dyn LanguageModel>>,+ active_model: Option<LanguageModel>,context_server_registry: Entity<ContextServerRegistry>,window: &mut Window,cx: &mut Context<Self>,@@ -244,13 +244,11 @@ impl ManageProfilesModal {let provider_id = language_model::LanguageModelProviderId(gpui::SharedString::from(selection.provider.0.clone()),);- let provider = registry.provider(&provider_id)?;- let model = provider+ registry+ .provider(&provider_id)?.provided_models(cx)- .iter()- .find(|m| m.id().0 == selection.model.as_str())?- .clone();- Some(language_model::ConfiguredModel { provider, model })+ .into_iter()+ .find(|m| m.id().0 == selection.model.as_str())})}},diff --git a/crates/agent_ui/src/agent_model_selector.rs b/crates/agent_ui/src/agent_model_selector.rsindex cbffced96df..89eeccc1e3c 100644--- a/crates/agent_ui/src/agent_model_selector.rs+++ b/crates/agent_ui/src/agent_model_selector.rs@@ -5,7 +5,7 @@ use crate::{};use fs::Fs;use gpui::{Entity, FocusHandle, SharedString};-use language_model::IconOrSvg;+use language_model::{IconOrSvg, LanguageModelRegistry};use picker::popover_menu::PickerPopoverMenu;use settings::update_settings_file;use std::sync::Arc;@@ -30,7 +30,7 @@ impl AgentModelSelector {language_model_selector({let model_context = model_usage_context.clone();- move |cx| model_context.configured_model(cx)+ move |cx| model_context.model(cx)},{let fs = fs.clone();@@ -74,7 +74,7 @@ impl AgentModelSelector {self.menu_handle.toggle(window, cx);}- pub fn active_model(&self, cx: &App) -> Option<language_model::ConfiguredModel> {+ pub fn active_model(&self, cx: &App) -> Option<language_model::LanguageModel> {self.selector.read(cx).delegate.active_model(cx)}@@ -90,10 +90,14 @@ impl Render for AgentModelSelector {let model = self.selector.read(cx).delegate.active_model(cx);let model_name = model.as_ref()- .map(|model| model.model.name().0)+ .map(|model| model.name().0).unwrap_or_else(|| SharedString::from("Select a Model"));- let provider_icon = model.as_ref().map(|model| model.provider.icon());+ let provider_icon = model.as_ref().and_then(|model| {+ LanguageModelRegistry::read_global(cx)+ .provider(&model.provider_id)+ .map(|provider| provider.icon())+ });let color = if self.menu_handle.is_deployed() {Color::Accent} else {diff --git a/crates/agent_ui/src/agent_panel.rs b/crates/agent_ui/src/agent_panel.rsindex 8a146247b87..4942a140a7d 100644--- a/crates/agent_ui/src/agent_panel.rs+++ b/crates/agent_ui/src/agent_panel.rs@@ -4710,16 +4710,16 @@ pub(crate) fn apply_native_model_override();return;};- let configured = LanguageModelRegistry::global(cx)+ let model = LanguageModelRegistry::global(cx).update(cx, |registry, cx| registry.select_model(&selected, cx));- let Some(configured) = configured else {+ let Some(model) = model else {log::warn!("create_thread: no model registered for {model_id:?}; using thread's default model");return;};thread.update(cx, |thread, cx| {- thread.set_model(configured.model, cx);+ thread.set_model(model, cx);});}@@ -6178,9 +6178,7 @@ impl AgentPanel {if LanguageModelRegistry::global(cx).read(cx).default_model()- .is_some_and(|model| {- model.provider.id() != language_model::ZED_CLOUD_PROVIDER_ID- })+ .is_some_and(|model| model.provider_id != language_model::ZED_CLOUD_PROVIDER_ID){return false;}diff --git a/crates/agent_ui/src/agent_ui.rs b/crates/agent_ui/src/agent_ui.rsindex c9e823bd26e..b3a4d9fd7d3 100644--- a/crates/agent_ui/src/agent_ui.rs+++ b/crates/agent_ui/src/agent_ui.rs@@ -56,7 +56,7 @@ use language::{language_settings::{AllLanguageSettings, EditPredictionProvider},};use language_model::{- ConfiguredModel, LanguageModelId, LanguageModelProviderId, LanguageModelRegistry,+ LanguageModel, LanguageModelId, LanguageModelProviderId, LanguageModelRegistry,};use project::{AgentId, DisableAiSettings};use prompt_store::{self, PromptBuilder, rules_to_skills_migration};@@ -538,7 +538,7 @@ pub(crate) enum ModelUsageContext {}impl ModelUsageContext {- pub fn configured_model(&self, cx: &App) -> Option<ConfiguredModel> {+ pub fn model(&self, cx: &App) -> Option<LanguageModel> {match self {Self::InlineAssistant => {LanguageModelRegistry::read_global(cx).inline_assistant_model()diff --git a/crates/agent_ui/src/buffer_codegen.rs b/crates/agent_ui/src/buffer_codegen.rsindex 38798cadd99..636758c7abb 100644--- a/crates/agent_ui/src/buffer_codegen.rs+++ b/crates/agent_ui/src/buffer_codegen.rs@@ -9,7 +9,6 @@ use futures::{channel::mpsc,future::{LocalBoxFuture, Shared},join,- stream::BoxStream,};use gpui::{App, AppContext as _, AsyncApp, Context, Entity, EventEmitter, Subscription, Task};use language::{@@ -17,9 +16,10 @@ use language::{};use language_model::{CompletionIntent, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelRegistry, LanguageModelRequest, LanguageModelRequestMessage,- LanguageModelRequestTool, LanguageModelTextStream, LanguageModelToolChoice,- LanguageModelToolUse, LanguageModelToolUseId, Role, StopReason, TokenUsage,+ LanguageModelCompletionStream, LanguageModelRegistry, LanguageModelRequest,+ LanguageModelRequestMessage, LanguageModelRequestTool, LanguageModelTextStream,+ LanguageModelToolChoice, LanguageModelToolUse, LanguageModelToolUseId, Role, StopReason,+ TokenUsage,};use language_models::provider::anthropic::telemetry::{AnthropicCompletionType, AnthropicEventData, AnthropicEventReporter, AnthropicEventType,@@ -174,7 +174,7 @@ impl BufferCodegen {pub fn start(&mut self,- primary_model: Arc<dyn LanguageModel>,+ primary_model: LanguageModel,user_prompt: String,context_task: Shared<Task<Option<LoadedContext>>>,cx: &mut Context<Self>,@@ -401,7 +401,7 @@ impl CodegenAlternative {&self.last_equal_ranges}- pub fn use_streaming_tools(model: &dyn LanguageModel, cx: &App) -> bool {+ pub fn use_streaming_tools(model: &LanguageModel, cx: &App) -> bool {model.supports_streaming_tools()&& AgentSettings::get_global(cx).inline_assistant_use_streaming_tools}@@ -410,7 +410,7 @@ impl CodegenAlternative {&mut self,user_prompt: String,context_task: Shared<Task<Option<LoadedContext>>>,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,cx: &mut Context<Self>,) -> Result<()> {// Clear the model explanation since the user has started a new generation.@@ -425,11 +425,12 @@ impl CodegenAlternative {self.edit_position = Some(self.range.start.bias_right(&self.snapshot));- if Self::use_streaming_tools(model.as_ref(), cx) {+ let provider = LanguageModelRegistry::read_global(cx).provider_for_model(&model)?;+ if Self::use_streaming_tools(&model, cx) {let request = self.build_request(&model, user_prompt, context_task, cx)?;let completion_events = cx.spawn({let model = model.clone();- async move |_, cx| model.stream_completion(request.await, cx).await+ async move |_, cx| provider.stream_completion(&model, request.await, cx).await});self.generation = self.handle_completion(model, completion_events, cx);} else {@@ -441,7 +442,9 @@ impl CodegenAlternative {cx.spawn({let model = model.clone();async move |_, cx| {- Ok(model.stream_completion_text(request.await, cx).await?)+ Ok(provider+ .stream_completion_text(&model, request.await, cx)+ .await?)}}).boxed_local()@@ -455,7 +458,7 @@ impl CodegenAlternative {fn build_request_tools(&self,- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,user_prompt: String,context_task: Shared<Task<Option<LoadedContext>>>,cx: &mut App,@@ -500,7 +503,8 @@ impl CodegenAlternative {let session_id = self.session_id.to_string();let tool_choice = model- .supports_tool_choice(LanguageModelToolChoice::Any)+ .tool_choice_support+ .supports(LanguageModelToolChoice::Any).then_some(LanguageModelToolChoice::Any);Ok(cx.spawn(async move |_cx| {@@ -565,12 +569,12 @@ impl CodegenAlternative {fn build_request(&self,- model: &Arc<dyn LanguageModel>,+ model: &LanguageModel,user_prompt: String,context_task: Shared<Task<Option<LoadedContext>>>,cx: &mut App,) -> Result<Task<LanguageModelRequest>> {- if Self::use_streaming_tools(model.as_ref(), cx) {+ if Self::use_streaming_tools(model, cx) {return self.build_request_tools(model, user_prompt, context_task, cx);}@@ -649,7 +653,7 @@ impl CodegenAlternative {pub fn handle_stream(&mut self,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,strip_invalid_spans: bool,stream: impl 'static + Future<Output = Result<LanguageModelTextStream>>,cx: &mut Context<Self>,@@ -1135,15 +1139,9 @@ impl CodegenAlternative {fn handle_completion(&mut self,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,completion_stream: Task<- Result<- BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,+ Result<LanguageModelCompletionStream, LanguageModelCompletionError>,>,cx: &mut Context<Self>,) -> Task<()> {@@ -1646,7 +1644,7 @@ mod tests {use gpui::TestAppContext;use indoc::indoc;use language::{Buffer, Point};- use language_model::fake_provider::FakeLanguageModel;+ use language_model::fake_provider::FakeLanguageModelProvider;use language_model::{LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelRegistry,LanguageModelToolUse, StopReason, TokenUsage,@@ -2507,7 +2505,9 @@ mod tests {.start("Rewrite the selection".into(),Task::ready(None).shared(),- Arc::new(FakeLanguageModel::default()),+ LanguageModelRegistry::read_global(cx)+ .default_model()+ .unwrap(),cx,).unwrap();@@ -2593,7 +2593,7 @@ mod tests {cx: &mut TestAppContext,) -> mpsc::UnboundedSender<String> {let (chunks_tx, chunks_rx) = mpsc::unbounded();- let model = Arc::new(FakeLanguageModel::default());+ let model = FakeLanguageModelProvider::default().model("fake");codegen.update(cx, |codegen, cx| {codegen.generation = codegen.handle_stream(model,@@ -2615,7 +2615,7 @@ mod tests {) -> mpsc::UnboundedSender<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>{let (events_tx, events_rx) = mpsc::unbounded();- let model = Arc::new(FakeLanguageModel::default());+ let model = FakeLanguageModelProvider::default().model("fake");codegen.update(cx, |codegen, cx| {codegen.generation =codegen.handle_completion(model, Task::ready(Ok(events_rx.boxed())), cx);@@ -2628,13 +2628,11 @@ mod tests {cx: &mut TestAppContext,) -> mpsc::UnboundedSender<LanguageModelCompletionEvent> {let (events_tx, events_rx) = mpsc::unbounded();- let model = Arc::new(FakeLanguageModel::default());+ let model = FakeLanguageModelProvider::default().model("fake");codegen.update(cx, |codegen, cx| {- let completion_stream = Task::ready(Ok(events_rx.map(Ok).boxed()- as BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >));+ let completion_stream = Task::ready(Ok(+ events_rx.map(Ok).boxed() as LanguageModelCompletionStream+ ));codegen.generation = codegen.handle_completion(model, completion_stream, cx);});events_txdiff --git a/crates/agent_ui/src/conversation_view/thread_view.rs b/crates/agent_ui/src/conversation_view/thread_view.rsindex f31a1227ae7..97fc7ea325a 100644--- a/crates/agent_ui/src/conversation_view/thread_view.rs+++ b/crates/agent_ui/src/conversation_view/thread_view.rs@@ -5144,9 +5144,9 @@ impl ThreadView {let enable_thinking = !thread.thinking_enabled();thread.set_thinking_enabled(enable_thinking, cx);- let favorite_key = thread.model().map(|model| {- (model.provider_id().0.to_string(), model.id().0.to_string())- });+ let favorite_key = thread+ .model()+ .map(|model| (model.provider_id.0.to_string(), model.id.0.to_string()));let fs = thread.project().read(cx).fs().clone();update_settings_file(fs, cx, move |settings, _| {if let Some(agent) = settings.agent.as_mut() {@@ -11960,7 +11960,7 @@ impl ThreadView {/// Returns the model to offer as a downgrade target when the current model/// requires data retention consent (e.g. Opus 4.8 for Fable).- fn data_retention_fallback_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn data_retention_fallback_model(&self, cx: &App) -> Option<LanguageModel> {let thread = self.as_native_thread(cx)?;let model = thread.read(cx).model()?.clone();let fallback_id = model.refusal_fallback_model_id()?;diff --git a/crates/agent_ui/src/favorite_models.rs b/crates/agent_ui/src/favorite_models.rsindex c655f9b6a55..563da9dd098 100644--- a/crates/agent_ui/src/favorite_models.rs+++ b/crates/agent_ui/src/favorite_models.rs@@ -7,7 +7,7 @@ use settings::{Settings as _, update_settings_file};use ui::App;pub fn toggle_in_settings(- model: Arc<dyn LanguageModel>,+ model: LanguageModel,should_be_favorite: bool,fs: Arc<dyn Fs>,cx: &mut App,diff --git a/crates/agent_ui/src/inline_assistant.rs b/crates/agent_ui/src/inline_assistant.rsindex 8f96989ac89..e65a1f2acd7 100644--- a/crates/agent_ui/src/inline_assistant.rs+++ b/crates/agent_ui/src/inline_assistant.rs@@ -39,7 +39,7 @@ use gpui::{UpdateGlobal, WeakEntity, Window, point,};use language::{Buffer, Point, Selection, TransactionId};-use language_model::{ConfigurationError, ConfiguredModel, LanguageModelRegistry};+use language_model::{ConfigurationError, LanguageModelRegistry};use multi_buffer::MultiBufferRow;use parking_lot::Mutex;use project::{DisableAiSettings, Project};@@ -407,13 +407,13 @@ impl InlineAssistant {"Assistant Invoked",kind = "inline",phase = "invoked",- model = model.model.telemetry_id(),- model_provider = model.provider.id().to_string(),+ model = model.telemetry_id(),+ model_provider = model.provider_id().to_string(),language_name = buffer.language().map(|language| language.name().to_proto()));report_anthropic_event(- &model.model,+ &model,AnthropicEventData {completion_type: AnthropicCompletionType::Editor,event: AnthropicEventType::Invoked,@@ -990,8 +990,8 @@ impl InlineAssistant {let codegen = assist.codegen.read(cx);let session_id = codegen.session_id();let message_id = active_alternative.read(cx).message_id.clone();- let model_telemetry_id = model.model.telemetry_id();- let model_provider_id = model.model.provider_id().to_string();+ let model_telemetry_id = model.telemetry_id();+ let model_provider_id = model.provider_id().to_string();let (phase, event_type, anthropic_event_type) = if undo {(@@ -1019,7 +1019,7 @@ impl InlineAssistant {);report_anthropic_event(- &model.model,+ &model,AnthropicEventData {completion_type: AnthropicCompletionType::Editor,event: anthropic_event_type,@@ -1245,9 +1245,7 @@ impl InlineAssistant {self.prompt_history.pop_front();}- let Some(ConfiguredModel { model, .. }) =- LanguageModelRegistry::read_global(cx).inline_assistant_model()- else {+ let Some(model) = LanguageModelRegistry::read_global(cx).inline_assistant_model() else {return;};diff --git a/crates/agent_ui/src/inline_prompt_editor.rs b/crates/agent_ui/src/inline_prompt_editor.rsindex 9d506aa5e4d..3d5baa652fd 100644--- a/crates/agent_ui/src/inline_prompt_editor.rs+++ b/crates/agent_ui/src/inline_prompt_editor.rs@@ -568,8 +568,8 @@ impl<T: 'static> PromptEditor<T> {return;};- let model_telemetry_id = model.model.telemetry_id();- let model_provider_id = model.provider.id().to_string();+ let model_telemetry_id = model.telemetry_id();+ let model_provider_id = model.provider_id().to_string();let (kind, language_name) = match &self.mode {PromptEditorMode::Buffer { codegen, .. } => {@@ -610,16 +610,13 @@ impl<T: 'static> PromptEditor<T> {CompletionState::Generated { completion_text } => {let model_info = self.model_selector.read(cx).active_model(cx);let (model_id, use_streaming_tools) = {- let Some(configured_model) = model_info else {+ let Some(model) = model_info else {self.toast("No configured model", None, cx);return;};(- configured_model.model.telemetry_id(),- CodegenAlternative::use_streaming_tools(- configured_model.model.as_ref(),- cx,- ),+ model.telemetry_id(),+ CodegenAlternative::use_streaming_tools(&model, cx),)};@@ -673,16 +670,13 @@ impl<T: 'static> PromptEditor<T> {CompletionState::Generated { completion_text } => {let model_info = self.model_selector.read(cx).active_model(cx);let (model_telemetry_id, use_streaming_tools) = {- let Some(configured_model) = model_info else {+ let Some(model) = model_info else {self.toast("No configured model", None, cx);return;};(- configured_model.model.telemetry_id(),- CodegenAlternative::use_streaming_tools(- configured_model.model.as_ref(),- cx,- ),+ model.telemetry_id(),+ CodegenAlternative::use_streaming_tools(&model, cx),)};@@ -1022,11 +1016,11 @@ impl<T: 'static> PromptEditor<T> {let disabled = matches!(codegen.status(cx), CodegenStatus::Idle);let model_registry = LanguageModelRegistry::read_global(cx);- let default_model = model_registry.default_model().map(|default| default.model);+ let default_model = model_registry.default_model();let alternative_models = model_registry.inline_alternative_models();let get_model_name = |index: usize| -> String {- let name = |model: &Arc<dyn LanguageModel>| model.name().0.to_string();+ let name = |model: &LanguageModel| model.name.0.to_string();match index {0 => default_model.as_ref().map_or_else(String::new, name),@@ -1201,7 +1195,7 @@ struct PromptEditorCompletionProviderDelegate;fn inline_assistant_model_supports_images(cx: &App) -> bool {LanguageModelRegistry::read_global(cx).inline_assistant_model()- .map_or(false, |m| m.model.supports_images())+ .map_or(false, |m| m.supports_images())}impl PromptCompletionProviderDelegate for PromptEditorCompletionProviderDelegate {diff --git a/crates/agent_ui/src/language_model_selector.rs b/crates/agent_ui/src/language_model_selector.rsindex 2c19e111571..ba60f9c2db0 100644--- a/crates/agent_ui/src/language_model_selector.rs+++ b/crates/agent_ui/src/language_model_selector.rs@@ -8,8 +8,8 @@ use gpui::{Subscription, Task,};use language_model::{- ConfiguredModel, IconOrSvg, LanguageModel, LanguageModelId, LanguageModelProvider,- LanguageModelProviderId, LanguageModelRegistry,+ IconOrSvg, LanguageModel, LanguageModelId, LanguageModelProvider, LanguageModelProviderId,+ LanguageModelRegistry,};use ordered_float::OrderedFloat;use picker::{Picker, PickerDelegate};@@ -19,16 +19,16 @@ use zed_actions::agent::OpenSettings;use crate::ui::{ModelSelectorFooter, ModelSelectorHeader, ModelSelectorListItem};-type OnModelChanged = Arc<dyn Fn(Arc<dyn LanguageModel>, &mut App) + 'static>;-type GetActiveModel = Arc<dyn Fn(&App) -> Option<ConfiguredModel> + 'static>;-type OnToggleFavorite = Arc<dyn Fn(Arc<dyn LanguageModel>, bool, &mut App) + 'static>;+type OnModelChanged = Arc<dyn Fn(LanguageModel, &mut App) + 'static>;+type GetActiveModel = Arc<dyn Fn(&App) -> Option<LanguageModel> + 'static>;+type OnToggleFavorite = Arc<dyn Fn(LanguageModel, bool, &mut App) + 'static>;pub type LanguageModelSelector = Picker<LanguageModelPickerDelegate>;pub fn language_model_selector(- get_active_model: impl Fn(&App) -> Option<ConfiguredModel> + 'static,- on_model_changed: impl Fn(Arc<dyn LanguageModel>, &mut App) + 'static,- on_toggle_favorite: impl Fn(Arc<dyn LanguageModel>, bool, &mut App) + 'static,+ get_active_model: impl Fn(&App) -> Option<LanguageModel> + 'static,+ on_model_changed: impl Fn(LanguageModel, &mut App) + 'static,+ on_toggle_favorite: impl Fn(LanguageModel, bool, &mut App) + 'static,popover_styles: bool,focus_handle: FocusHandle,window: &mut Window,@@ -96,7 +96,7 @@ type FavoritesIndex = HashMap<LanguageModelProviderId, HashSet<LanguageModelId>>#[derive(Clone)]struct ModelInfo {- model: Arc<dyn LanguageModel>,+ model: LanguageModel,icon: IconOrSvg,is_favorite: bool,}@@ -104,7 +104,7 @@ struct ModelInfo {impl ModelInfo {fn new(provider: &dyn LanguageModelProvider,- model: Arc<dyn LanguageModel>,+ model: LanguageModel,favorites_index: &FavoritesIndex,) -> Self {let is_favorite = favorites_index@@ -133,9 +133,9 @@ pub struct LanguageModelPickerDelegate {impl LanguageModelPickerDelegate {fn new(- get_active_model: impl Fn(&App) -> Option<ConfiguredModel> + 'static,- on_model_changed: impl Fn(Arc<dyn LanguageModel>, &mut App) + 'static,- on_toggle_favorite: impl Fn(Arc<dyn LanguageModel>, bool, &mut App) + 'static,+ get_active_model: impl Fn(&App) -> Option<LanguageModel> + 'static,+ on_model_changed: impl Fn(LanguageModel, &mut App) + 'static,+ on_toggle_favorite: impl Fn(LanguageModel, bool, &mut App) + 'static,popover_styles: bool,focus_handle: FocusHandle,window: &mut Window,@@ -177,7 +177,7 @@ impl LanguageModelPickerDelegate {fn get_active_model_index(entries: &[LanguageModelPickerEntry],- active_model: Option<ConfiguredModel>,+ active_model: Option<LanguageModel>,) -> usize {entries.iter()@@ -185,10 +185,7 @@ impl LanguageModelPickerDelegate {if let LanguageModelPickerEntry::Model(model) = entry {active_model.as_ref()- .map(|active_model| {- active_model.model.id() == model.model.id()- && active_model.provider.id() == model.model.provider_id()- })+ .map(|active_model| active_model.is_same_as(&model.model)).unwrap_or_default()} else {false@@ -197,7 +194,7 @@ impl LanguageModelPickerDelegate {.unwrap_or(0)}- pub fn active_model(&self, cx: &App) -> Option<ConfiguredModel> {+ pub fn active_model(&self, cx: &App) -> Option<LanguageModel> {(self.get_active_model)(cx)}@@ -211,8 +208,8 @@ impl LanguageModelPickerDelegate {}let active_model = (self.get_active_model)(cx);- let active_provider_id = active_model.as_ref().map(|m| m.provider.id());- let active_model_id = active_model.as_ref().map(|m| m.model.id());+ let active_provider_id = active_model.as_ref().map(|m| m.provider_id());+ let active_model_id = active_model.as_ref().map(|m| m.id());let current_index = self.all_models@@ -484,7 +481,7 @@ impl PickerDelegate for LanguageModelPickerDelegate {self.filtered_entries.get(self.selected_index){let model = model_info.model.clone();- (self.on_model_changed)(model.clone(), cx);+ (self.on_model_changed)(model, cx);let current_index = self.selected_index;self.set_selected_index(current_index, window, cx);@@ -510,15 +507,16 @@ impl PickerDelegate for LanguageModelPickerDelegate {}LanguageModelPickerEntry::Model(model_info) => {let active_model = (self.get_active_model)(cx);- let active_provider_id = active_model.as_ref().map(|m| m.provider.id());- let active_model_id = active_model.map(|m| m.model.id());+ let active_provider_id = active_model.as_ref().map(|m| m.provider_id());+ let active_model_id = active_model.map(|m| m.id());let is_selected = Some(model_info.model.provider_id()) == active_provider_id&& Some(model_info.model.id()) == active_model_id;let model_cost = model_info.model- .model_cost_info()+ .cost_info+ .as_ref().map(|cost| cost.to_shared_string());let is_favorite = model_info.is_favorite;@@ -567,87 +565,21 @@ impl PickerDelegate for LanguageModelPickerDelegate {#[cfg(test)]mod tests {use super::*;- use futures::{future::BoxFuture, stream::BoxStream};- use gpui::{AsyncApp, TestAppContext};+ use gpui::TestAppContext;use language_model::{- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,- LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelRequest, LanguageModelToolChoice,+ LanguageModelId, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,};use ui::IconName;- #[derive(Clone)]- struct TestLanguageModel {- name: LanguageModelName,- id: LanguageModelId,- provider_id: LanguageModelProviderId,- provider_name: LanguageModelProviderName,- }-- impl TestLanguageModel {- fn new(name: &str, provider: &str) -> Self {- Self {- name: LanguageModelName::from(name.to_string()),- id: LanguageModelId::from(name.to_string()),- provider_id: LanguageModelProviderId::from(provider.to_string()),- provider_name: LanguageModelProviderName::from(provider.to_string()),- }- }- }-- impl LanguageModel for TestLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- self.name.clone()- }-- fn provider_id(&self) -> LanguageModelProviderId {- self.provider_id.clone()- }-- fn provider_name(&self) -> LanguageModelProviderName {- self.provider_name.clone()- }-- fn supports_tools(&self) -> bool {- false- }-- fn supports_tool_choice(&self, _choice: LanguageModelToolChoice) -> bool {- false- }-- fn supports_images(&self) -> bool {- false- }-- fn telemetry_id(&self) -> String {- format!("{}/{}", self.provider_id.0, self.name.0)- }-- fn max_token_count(&self) -> u64 {- 1000- }-- fn stream_completion(- &self,- _: LanguageModelRequest,- _: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- unimplemented!()- }+ fn test_language_model(name: &str, provider: &str) -> LanguageModel {+ LanguageModel::new(+ LanguageModelId::from(name.to_string()),+ LanguageModelName::from(name.to_string()),+ LanguageModelProviderId::from(provider.to_string()),+ LanguageModelProviderName::from(provider.to_string()),+ format!("{provider}/{name}"),+ 1000,+ )}fn create_models(model_specs: Vec<(&str, &str)>) -> Vec<ModelInfo> {@@ -665,7 +597,7 @@ mod tests {.iter().any(|(fav_provider, fav_name)| *fav_provider == provider && *fav_name == name);ModelInfo {- model: Arc::new(TestLanguageModel::new(name, provider)),+ model: test_language_model(name, provider),icon: IconOrSvg::Icon(IconName::ZedAgent),is_favorite,}diff --git a/crates/agent_ui/src/terminal_codegen.rs b/crates/agent_ui/src/terminal_codegen.rsindex 19adf2d8806..ec83b7710f6 100644--- a/crates/agent_ui/src/terminal_codegen.rs+++ b/crates/agent_ui/src/terminal_codegen.rs@@ -1,7 +1,7 @@use crate::inline_prompt_editor::CodegenStatus;use futures::{SinkExt, StreamExt, channel::mpsc};use gpui::{App, AppContext as _, Context, Entity, EventEmitter, Task};-use language_model::{ConfiguredModel, LanguageModelRegistry, LanguageModelRequest};+use language_model::{LanguageModelRegistry, LanguageModelRequest};use language_models::provider::anthropic::telemetry::{AnthropicCompletionType, AnthropicEventData, AnthropicEventReporter, AnthropicEventType,};@@ -37,11 +37,11 @@ impl TerminalCodegen {}pub fn start(&mut self, prompt_task: Task<LanguageModelRequest>, cx: &mut Context<Self>) {- let Some(ConfiguredModel { model, .. }) =- LanguageModelRegistry::read_global(cx).inline_assistant_model()- else {+ let registry = LanguageModelRegistry::read_global(cx);+ let Some(model) = registry.inline_assistant_model() else {return;};+ let provider = registry.provider_for_model(&model);let anthropic_reporter = AnthropicEventReporter::new(&model, cx);let session_id = self.session_id;@@ -52,7 +52,10 @@ impl TerminalCodegen {self.transaction = Some(TerminalTransaction::start(self.terminal.clone()));self.generation = cx.spawn(async move |this, cx| {let prompt = prompt_task.await;- let response = model.stream_completion_text(prompt, cx).await;+ let response = match provider {+ Ok(provider) => provider.stream_completion_text(&model, prompt, cx).await,+ Err(error) => Err(error),+ };let generate = async {let message_id = response.as_ref()diff --git a/crates/agent_ui/src/terminal_inline_assistant.rs b/crates/agent_ui/src/terminal_inline_assistant.rsindex c7c9e7fb8e2..a3eca1d845b 100644--- a/crates/agent_ui/src/terminal_inline_assistant.rs+++ b/crates/agent_ui/src/terminal_inline_assistant.rs@@ -15,8 +15,8 @@ use fs::Fs;use gpui::{App, Entity, Focusable, Global, Subscription, Task, UpdateGlobal, WeakEntity};use language::Buffer;use language_model::{- CompletionIntent, ConfiguredModel, LanguageModelRegistry, LanguageModelRequest,- LanguageModelRequestMessage, Role,+ CompletionIntent, LanguageModelRegistry, LanguageModelRequest, LanguageModelRequestMessage,+ Role,};use language_models::provider::anthropic::telemetry::{AnthropicCompletionType, AnthropicEventData, AnthropicEventType, report_anthropic_event,@@ -214,7 +214,7 @@ impl TerminalInlineAssistant {assist_id: TerminalInlineAssistId,cx: &mut App,) -> Result<Task<LanguageModelRequest>> {- let ConfiguredModel { model, .. } = LanguageModelRegistry::read_global(cx)+ let model = LanguageModelRegistry::read_global(cx).inline_assistant_model().context("No inline assistant model")?;@@ -301,9 +301,7 @@ impl TerminalInlineAssistant {}).log_err();- if let Some(ConfiguredModel { model, .. }) =- LanguageModelRegistry::read_global(cx).inline_assistant_model()- {+ if let Some(model) = LanguageModelRegistry::read_global(cx).inline_assistant_model() {let codegen = assist.codegen.read(cx);let session_id = codegen.session_id();let message_id = codegen.message_id.clone();diff --git a/crates/benchmarks/benches/edit_file_tool.rs b/crates/benchmarks/benches/edit_file_tool.rsindex 2adf7cd8ad6..204628cfe80 100644--- a/crates/benchmarks/benches/edit_file_tool.rs+++ b/crates/benchmarks/benches/edit_file_tool.rs@@ -26,7 +26,7 @@ use gpui::{UpdateGlobal as _,};use language::{FakeLspAdapter, rust_lang};-use language_model::fake_provider::FakeLanguageModel;+use language_model::LanguageModelRegistry;use project::{FakeFs, Project};use prompt_store::ProjectContext;use rand::{Rng as _, SeedableRng as _, rngs::StdRng};@@ -206,7 +206,7 @@ async fn setup_editor_and_tool(cx: &mut TestAppContext, file_text: String) -> Halet context_server_registry =cx.new(|cx| ContextServerRegistry::new(project.read(cx).context_server_store(), cx));- let model = Arc::new(FakeLanguageModel::default());+ let model = cx.update(|cx| LanguageModelRegistry::test(cx).model("fake"));let thread = cx.new(|cx| {Thread::new(project.clone(),diff --git a/crates/copilot_chat/src/copilot_chat.rs b/crates/copilot_chat/src/copilot_chat.rsindex 9b9b208a1b7..c56bee14dbe 100644--- a/crates/copilot_chat/src/copilot_chat.rs+++ b/crates/copilot_chat/src/copilot_chat.rs@@ -2,7 +2,7 @@ pub mod copilot_oauth;mod model;pub mod responses;-pub use model::{PROVIDER_ID, PROVIDER_NAME, create_language_model};+pub use model::{PROVIDER_ID, PROVIDER_NAME, language_model, stream_completion};use std::sync::Arc;@@ -768,7 +768,7 @@ impl CopilotChat {Self::get_auth_details(&copilot_chat, &mut cx).await?;let api_url = configuration.chat_completions_url(&api_endpoint);- stream_completion(+ stream_chat_completion(client.clone(),oauth_token,api_url.into(),@@ -1083,7 +1083,7 @@ async fn request_models(Ok(models)}-async fn stream_completion(+async fn stream_chat_completion(client: Arc<dyn HttpClient>,oauth_token: String,completion_url: Arc<str>,diff --git a/crates/copilot_chat/src/model.rs b/crates/copilot_chat/src/model.rsindex e778a182583..ee7e2ddd29b 100644--- a/crates/copilot_chat/src/model.rs+++ b/crates/copilot_chat/src/model.rs@@ -1,4 +1,4 @@-use std::{pin::Pin, str::FromStr as _, sync::Arc};+use std::{pin::Pin, str::FromStr as _};use crate::responses as copilot_responses;use crate::{@@ -13,17 +13,16 @@ use anthropic::{use anyhow::{Result, anyhow};use collections::HashMap;use futures::future::BoxFuture;-use futures::stream::BoxStream;use futures::{FutureExt, Stream, StreamExt};use gpui::{AsyncApp, Entity};use http_client::StatusCode;use language_model::{CompletionIntent, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelCostInfo, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,- LanguageModelProviderId, LanguageModelProviderName, LanguageModelRequest,- LanguageModelRequestMessage, LanguageModelToolChoice, LanguageModelToolResultContent,- LanguageModelToolUse, MessageContent, ProviderErrorCategory, RateLimiter, Role, StopReason,- TokenUsage,+ LanguageModelCompletionStream, LanguageModelCostInfo, LanguageModelEffortLevel,+ LanguageModelId, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,+ LanguageModelRequest, LanguageModelRequestMessage, LanguageModelToolChoice,+ LanguageModelToolChoiceSupport, LanguageModelToolResultContent, LanguageModelToolUse,+ MessageContent, ProviderErrorCategory, RateLimiter, Role, StopReason, TokenUsage,};use util::debug_panic;@@ -33,61 +32,13 @@ pub const PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("cpub const PROVIDER_NAME: LanguageModelProviderName =LanguageModelProviderName::new("GitHub Copilot Chat");-pub fn create_language_model(- model: CopilotChatModel,- copilot_chat: Entity<CopilotChat>,-) -> Arc<dyn LanguageModel> {- Arc::new(CopilotChatLanguageModel {- model,- copilot_chat,- request_limiter: RateLimiter::new(4),- })-}--struct CopilotChatLanguageModel {- model: CopilotChatModel,- copilot_chat: Entity<CopilotChat>,- request_limiter: RateLimiter,-}--impl LanguageModel for CopilotChatLanguageModel {- fn id(&self) -> LanguageModelId {- LanguageModelId::from(self.model.id().to_string())- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.supports_tools()- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_images(&self) -> bool {- self.model.supports_vision()- }-- fn supports_thinking(&self) -> bool {- self.model.can_think()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- let levels = self.model.reasoning_effort_levels();- if levels.is_empty() {- return vec![];- }+/// Describes a Copilot Chat model as a [`LanguageModel`].+pub fn language_model(model: &CopilotChatModel) -> LanguageModel {+ let supports_tools = model.supports_tools();+ let levels = model.reasoning_effort_levels();+ let supported_effort_levels = if levels.is_empty() {+ vec![]+ } else {levels.iter().map(|level| {@@ -105,228 +56,216 @@ impl LanguageModel for CopilotChatLanguageModel {}}).collect()- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto- | LanguageModelToolChoice::Any- | LanguageModelToolChoice::None => self.supports_tools(),- }- }-- fn model_cost_info(&self) -> Option<LanguageModelCostInfo> {- LanguageModelCostInfo::RequestCost {- cost_per_request: self.model.multiplier(),+ };+ LanguageModel {+ supports_tools,+ supports_streaming_tools: true,+ supports_images: model.supports_vision(),+ supports_thinking: model.can_think(),+ supported_effort_levels: supported_effort_levels.into(),+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: supports_tools,+ any: supports_tools,+ none: supports_tools,+ },+ cost_info: LanguageModelCostInfo::RequestCost {+ cost_per_request: model.multiplier(),}- .into()- }-- fn telemetry_id(&self) -> String {- format!("copilot_chat/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_input_tokens(&self) -> u64 {- self.model- .max_prompt_tokens()- .unwrap_or_else(|| self.model.max_token_count())- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()+ .into(),+ max_input_tokens: model.max_prompt_tokens().unwrap_or(model.max_token_count()),+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("copilot_chat/{}", model.id()),+ model.max_token_count(),+ )}+}- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let is_user_initiated = request.intent.is_none_or(|intent| match intent {- CompletionIntent::UserPrompt- | CompletionIntent::ThreadContextSummarization- | CompletionIntent::InlineAssist- | CompletionIntent::TerminalInlineAssist- | CompletionIntent::GenerateGitCommitMessage => true,-- CompletionIntent::Subagent- | CompletionIntent::ToolResults- | CompletionIntent::ThreadSummarization- | CompletionIntent::CreateFile- | CompletionIntent::EditFile => false,- });+/// Streams a completion of `request` from the Copilot Chat model `model`.+pub fn stream_completion(+ model: &CopilotChatModel,+ copilot_chat: &Entity<CopilotChat>,+ request_limiter: &RateLimiter,+ request: LanguageModelRequest,+ cx: &AsyncApp,+) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>> {+ let is_user_initiated = request.intent.is_none_or(|intent| match intent {+ CompletionIntent::UserPrompt+ | CompletionIntent::ThreadContextSummarization+ | CompletionIntent::InlineAssist+ | CompletionIntent::TerminalInlineAssist+ | CompletionIntent::GenerateGitCommitMessage => true,++ CompletionIntent::Subagent+ | CompletionIntent::ToolResults+ | CompletionIntent::ThreadSummarization+ | CompletionIntent::CreateFile+ | CompletionIntent::EditFile => false,+ });- if self.model.supports_messages() {- let location = intent_to_chat_location(request.intent);- let model = self.model.clone();- let copilot_chat = self.copilot_chat.clone();- let request_limiter = self.request_limiter.clone();- let future = cx.spawn(async move |cx| {- let max_output_tokens = model- .max_output_tokens()- .or(request.max_output_tokens)- .ok_or_else(|| {- anyhow!("Copilot did not provide an output limit for this model")- })?;- let effort = request- .thinking_effort- .as_ref()- .and_then(|e| anthropic::Effort::from_str(e).ok());-- let mut anthropic_request = into_anthropic(- request,- model.id().to_string(),- 0.0,- max_output_tokens,- if model.supports_adaptive_thinking() {- AnthropicModelMode::Thinking {- budget_tokens: None,- }- } else if model.supports_thinking() {- AnthropicModelMode::Thinking {- budget_tokens: compute_thinking_budget(- model.min_thinking_budget(),- model.max_thinking_budget(),- max_output_tokens.min(u32::MAX as u64) as u32,- ),- }- } else {- AnthropicModelMode::Default- },- AnthropicPromptCacheMode::Legacy,- &PROVIDER_ID,- )?;+ if model.supports_messages() {+ let location = intent_to_chat_location(request.intent);+ let model = model.clone();+ let copilot_chat = copilot_chat.clone();+ let request_limiter = request_limiter.clone();+ let future = cx.spawn(async move |cx| {+ let max_output_tokens = model+ .max_output_tokens()+ .or(request.max_output_tokens)+ .ok_or_else(|| anyhow!("Copilot did not provide an output limit for this model"))?;+ let effort = request+ .thinking_effort+ .as_ref()+ .and_then(|e| anthropic::Effort::from_str(e).ok());++ let mut anthropic_request = into_anthropic(+ request,+ model.id().to_string(),+ 0.0,+ max_output_tokens,+ if model.supports_adaptive_thinking() {+ AnthropicModelMode::Thinking {+ budget_tokens: None,+ }+ } else if model.supports_thinking() {+ AnthropicModelMode::Thinking {+ budget_tokens: compute_thinking_budget(+ model.min_thinking_budget(),+ model.max_thinking_budget(),+ max_output_tokens.min(u32::MAX as u64) as u32,+ ),+ }+ } else {+ AnthropicModelMode::Default+ },+ AnthropicPromptCacheMode::Legacy,+ &PROVIDER_ID,+ )?;- anthropic_request.temperature = None;+ anthropic_request.temperature = None;- // The Copilot proxy doesn't support eager_input_streaming on tools.- for tool in &mut anthropic_request.tools {- tool.eager_input_streaming = false;- }+ // The Copilot proxy doesn't support eager_input_streaming on tools.+ for tool in &mut anthropic_request.tools {+ tool.eager_input_streaming = false;+ }- if model.supports_adaptive_thinking() {- if anthropic_request.thinking.is_some() {- anthropic_request.thinking = Some(anthropic::Thinking::Adaptive {- display: Some(anthropic::AdaptiveThinkingDisplay::Summarized),- // Thinking block binding needs a beta header on- // the upstream Anthropic request, which the- // Copilot proxy controls, so opting in belongs- // server-side.- block_binding: None,+ if model.supports_adaptive_thinking() {+ if anthropic_request.thinking.is_some() {+ anthropic_request.thinking = Some(anthropic::Thinking::Adaptive {+ display: Some(anthropic::AdaptiveThinkingDisplay::Summarized),+ // Thinking block binding needs a beta header on+ // the upstream Anthropic request, which the+ // Copilot proxy controls, so opting in belongs+ // server-side.+ block_binding: None,+ });+ anthropic_request.output_config =+ effort.map(|effort| anthropic::OutputConfig {+ effort: Some(effort),});- anthropic_request.output_config =- effort.map(|effort| anthropic::OutputConfig {- effort: Some(effort),- });- }}+ }- let anthropic_beta =- if !model.supports_adaptive_thinking() && model.supports_thinking() {- Some("interleaved-thinking-2025-05-14".to_string())- } else {- None- };+ let anthropic_beta = if !model.supports_adaptive_thinking() && model.supports_thinking()+ {+ Some("interleaved-thinking-2025-05-14".to_string())+ } else {+ None+ };- let body = serde_json::to_string(&anthropic::StreamingRequest {- base: anthropic_request,- stream: true,- })- .map_err(|e| anyhow::anyhow!(e))?;-- let stream = CopilotChat::stream_messages(- copilot_chat,- body,- location,- is_user_initiated,- anthropic_beta,- cx.clone(),- );- let executor = cx.background_executor().clone();-- request_limiter- .stream(async move {- let events = stream.await?;- let mapper = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID);- Ok(language_model::stream_in_background(- mapper.map_stream(events).boxed(),- executor,- ))- })- .await- });- return async move { Ok(future.await?.boxed()) }.boxed();- }+ let body = serde_json::to_string(&anthropic::StreamingRequest {+ base: anthropic_request,+ stream: true,+ })+ .map_err(|e| anyhow::anyhow!(e))?;- if self.model.supports_response() {- let location = intent_to_chat_location(request.intent);- let responses_request = match into_copilot_responses(&self.model, request) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let copilot_chat = self.copilot_chat.clone();- let request_limiter = self.request_limiter.clone();- let future = cx.spawn(async move |cx| {- let request = CopilotChat::stream_response(- copilot_chat,- responses_request,- location,- is_user_initiated,- cx.clone(),- );- let executor = cx.background_executor().clone();- request_limiter- .stream(async move {- let stream = request.await?;- let mapper = CopilotResponsesEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(stream).boxed(),- executor,- ))- })- .await- });- return async move { Ok(future.await?.boxed()) }.boxed();- }+ let stream = CopilotChat::stream_messages(+ copilot_chat,+ body,+ location,+ is_user_initiated,+ anthropic_beta,+ cx.clone(),+ );+ let executor = cx.background_executor().clone();++ request_limiter+ .stream(async move {+ let events = stream.await?;+ let mapper = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID);+ Ok(language_model::stream_in_background(+ mapper.map_stream(events).boxed(),+ executor,+ ))+ })+ .await+ });+ return async move { Ok(future.await?.boxed()) }.boxed();+ }+ if model.supports_response() {let location = intent_to_chat_location(request.intent);- let copilot_request = match into_copilot_chat(&self.model, request) {+ let responses_request = match into_copilot_responses(model, request) {Ok(request) => request,- Err(err) => return futures::future::ready(Err(err.into())).boxed(),+ Err(error) => return async move { Err(error.into()) }.boxed(),};- let is_streaming = copilot_request.stream;-- let copilot_chat = self.copilot_chat.clone();- let request_limiter = self.request_limiter.clone();+ let copilot_chat = copilot_chat.clone();+ let request_limiter = request_limiter.clone();let future = cx.spawn(async move |cx| {- let request = CopilotChat::stream_completion(+ let request = CopilotChat::stream_response(copilot_chat,- copilot_request,+ responses_request,location,is_user_initiated,cx.clone(),);+ let executor = cx.background_executor().clone();request_limiter.stream(async move {- let response = request.await?;- Ok(map_to_language_model_completion_events(- response,- is_streaming,+ let stream = request.await?;+ let mapper = CopilotResponsesEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(stream).boxed(),+ executor,))}).await});- async move { Ok(future.await?.boxed()) }.boxed()+ return async move { Ok(future.await?.boxed()) }.boxed();}++ let location = intent_to_chat_location(request.intent);+ let copilot_request = match into_copilot_chat(model, request) {+ Ok(request) => request,+ Err(err) => return futures::future::ready(Err(err.into())).boxed(),+ };+ let is_streaming = copilot_request.stream;++ let copilot_chat = copilot_chat.clone();+ let request_limiter = request_limiter.clone();+ let future = cx.spawn(async move |cx| {+ let request = CopilotChat::stream_completion(+ copilot_chat,+ copilot_request,+ location,+ is_user_initiated,+ cx.clone(),+ );+ request_limiter+ .stream(async move {+ let response = request.await?;+ Ok(map_to_language_model_completion_events(+ response,+ is_streaming,+ ))+ })+ .await+ });+ async move { Ok(future.await?.boxed()) }.boxed()}pub fn map_to_language_model_completion_events(@@ -1311,19 +1250,10 @@ mod tests {use futures::StreamExt;use language_model::ProviderErrorCategory;use serde_json::json;+ use std::sync::Arc;- #[gpui::test]- fn language_model_exposes_token_limits(cx: &mut gpui::TestAppContext) {- use gpui::AppContext as _;-- let copilot_chat = cx.new(|cx| {- CopilotChat::new(- Arc::new(http_client::BlockedHttpClient::new()),- Arc::new(EmptyCredentialsProvider),- crate::CopilotChatConfiguration::default(),- cx,- )- });+ #[test]+ fn language_model_exposes_token_limits() {for (limits, input, context, output) in [(json!({"max_context_window_tokens": 200_000, "max_prompt_tokens": 90_000, "max_output_tokens": 16_384}),@@ -1353,41 +1283,10 @@ mod tests {let mut value = serde_json::to_value(test_responses_model()).unwrap();value["capabilities"]["limits"] = limits;let descriptor = serde_json::from_value(value).unwrap();- let model = create_language_model(descriptor, copilot_chat.clone());- assert_eq!(model.max_token_count(), context);- assert_eq!(model.max_input_tokens(), input);- assert_eq!(model.max_total_tokens(), Some(context));- assert_eq!(model.max_output_tokens(), output);- }- }-- struct EmptyCredentialsProvider;-- impl credentials_provider::CredentialsProvider for EmptyCredentialsProvider {- fn read_credentials<'a>(- &'a self,- _url: &'a str,- _cx: &'a AsyncApp,- ) -> Pin<Box<dyn Future<Output = Result<Option<(String, Vec<u8>)>>> + 'a>> {- Box::pin(async { Ok(None) })- }-- fn write_credentials<'a>(- &'a self,- _url: &'a str,- _username: &'a str,- _password: &'a [u8],- _cx: &'a AsyncApp,- ) -> Pin<Box<dyn Future<Output = Result<()>> + 'a>> {- Box::pin(async { Ok(()) })- }-- fn delete_credentials<'a>(- &'a self,- _url: &'a str,- _cx: &'a AsyncApp,- ) -> Pin<Box<dyn Future<Output = Result<()>> + 'a>> {- Box::pin(async { Ok(()) })+ let model = language_model(&descriptor);+ assert_eq!(model.max_token_count, context);+ assert_eq!(model.max_input_tokens, input);+ assert_eq!(model.max_output_tokens, output);}}diff --git a/crates/eval_cli/src/main.rs b/crates/eval_cli/src/main.rsindex e25730dc5e0..b223fb2f2ed 100644--- a/crates/eval_cli/src/main.rs+++ b/crates/eval_cli/src/main.rs@@ -463,10 +463,7 @@ fn ensure_provider_authenticated(selected: &SelectedModel, cx: &gpui::App) -> ReOk(())}-fn find_available_model(- selected: &SelectedModel,- cx: &gpui::App,-) -> Option<Arc<dyn LanguageModel>> {+fn find_available_model(selected: &SelectedModel, cx: &gpui::App) -> Option<LanguageModel> {let registry = LanguageModelRegistry::global(cx);let models = registry.read(cx).available_models(cx).collect::<Vec<_>>();@@ -619,19 +616,11 @@ async fn run_agent(let registry = LanguageModelRegistry::global(cx);let model = find_available_model(&selected, cx).ok_or_else(|| model_not_found_error(model_name, cx))?;- let provider = registry- .read(cx)- .provider(&model.provider_id())- .context("Provider not found")?;-let supports_thinking = model.supports_thinking();let model_id = model.id().0.to_string();registry.update(cx, |registry, cx| {- registry.set_default_model(- Some(language_model::ConfiguredModel { provider, model }),- cx,- );+ registry.set_default_model(Some(model), cx);});let enable_thinking = thinking_override.unwrap_or(supports_thinking);diff --git a/crates/git_ui/src/git_panel.rs b/crates/git_ui/src/git_panel.rsindex 31c7e55cb4e..1e3e49da653 100644--- a/crates/git_ui/src/git_panel.rs+++ b/crates/git_ui/src/git_panel.rs@@ -57,8 +57,8 @@ use gpui::{use itertools::Itertools;use language::{Buffer, BufferEvent, File};use language_model::{- CompletionIntent, ConfiguredModel, Event as LanguageModelEvent, LanguageModelRegistry,- LanguageModelRequest, LanguageModelRequestMessage, Role,+ CompletionIntent, Event as LanguageModelEvent, LanguageModelRegistry, LanguageModelRequest,+ LanguageModelRequestMessage, Role,};use menu;use multi_buffer::ExcerptBoundaryInfo;@@ -4117,11 +4117,22 @@ impl GitPanel {return;}- let Some(ConfiguredModel { provider, model }) =- LanguageModelRegistry::read_global(cx).commit_message_model(cx)- else {+ let registry = LanguageModelRegistry::read_global(cx);+ let Some(model) = registry.commit_message_model(cx) else {return;};+ let provider = match registry.provider_for_model(&model) {+ Ok(provider) => provider,+ Err(error) => {+ if let Some(workspace) = self.workspace.upgrade() {+ workspace.update(cx, |workspace, cx| {+ workspace+ .show_error(format!("Failed to generate commit message: {error}"), cx);+ });+ }+ return;+ }+ };let Some(repo) = self.active_repository.as_ref() else {return;@@ -4155,11 +4166,7 @@ impl GitPanel {});if let Some(task) = cx.update(|cx| {- if !provider.is_authenticated(cx) {- Some(provider.authenticate(cx))- } else {- None- }+ (!provider.is_authenticated(cx)).then(|| provider.authenticate(cx))}) {task.await.log_err();}@@ -4240,7 +4247,7 @@ impl GitPanel {max_output_tokens: None,};- let stream = model.stream_completion_text(request, cx);+ let stream = provider.stream_completion_text(&model, request, cx);match stream.await {Ok(mut messages) => {if !text_empty {diff --git a/crates/language_model/src/fake_provider.rs b/crates/language_model/src/fake_provider.rsindex 4aa9d34cfeb..d8dee6dcba0 100644--- a/crates/language_model/src/fake_provider.rs+++ b/crates/language_model/src/fake_provider.rs@@ -1,34 +1,48 @@use crate::{AuthenticateError, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice,+ LanguageModelCompletionStream, LanguageModelId, LanguageModelName, LanguageModelProvider,+ LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,+ LanguageModelRequest,};use anyhow::anyhow;-use futures::{FutureExt, channel::mpsc, future::BoxFuture, stream::BoxStream, stream::StreamExt};+use futures::{FutureExt, channel::mpsc, future::BoxFuture, stream::StreamExt};use gpui::{App, AsyncApp, Entity, Task};use http_client::Result;use parking_lot::Mutex;use std::collections::VecDeque;-use std::sync::{- Arc,- atomic::{AtomicBool, AtomicU64, Ordering::SeqCst},-};-#[derive(Clone)]+/// A provider for tests that serves any model id and records every request.+///+/// Tests get models with [`Self::model`] and drive responses to the requests+/// the provider has received, like `send_last_text`.pub struct FakeLanguageModelProvider {id: LanguageModelProviderId,name: LanguageModelProviderName,- models: Vec<Arc<dyn LanguageModel>>,+ state: Mutex<FakeProviderState>,+}++#[derive(Default)]+struct FakeProviderState {+ models: Vec<LanguageModel>,+ pending_completions: Vec<PendingCompletion>,+ forbid_requests: bool,+ input_token_counts: VecDeque<u64>,+ input_token_count_requests: Vec<LanguageModelRequest>,+}++struct PendingCompletion {+ model_id: LanguageModelId,+ request: LanguageModelRequest,+ events:+ mpsc::UnboundedSender<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,}impl Default for FakeLanguageModelProvider {fn default() -> Self {- Self {- id: LanguageModelProviderId::from("fake".to_string()),- name: LanguageModelProviderName::from("Fake".to_string()),- models: vec![Arc::new(FakeLanguageModel::default())],- }+ Self::new(+ LanguageModelProviderId::from("fake".to_string()),+ LanguageModelProviderName::from("Fake".to_string()),+ )}}@@ -49,16 +63,16 @@ impl LanguageModelProvider for FakeLanguageModelProvider {self.name.clone()}- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- self.models.first().cloned()+ fn default_model(&self, _cx: &App) -> Option<LanguageModel> {+ self.state.lock().models.first().cloned()}- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- self.models.first().cloned()+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ self.state.lock().models.first().cloned()}- fn provided_models(&self, _: &App) -> Vec<Arc<dyn LanguageModel>> {- self.models.clone()+ fn provided_models(&self, _: &App) -> Vec<LanguageModel> {+ self.state.lock().models.clone()}fn is_authenticated(&self, _: &App) -> bool {@@ -72,348 +86,256 @@ impl LanguageModelProvider for FakeLanguageModelProvider {fn settings_view(&self, _: &mut App) -> Option<crate::ProviderSettingsView> {None}-}-impl FakeLanguageModelProvider {- pub fn new(id: LanguageModelProviderId, name: LanguageModelProviderName) -> Self {- Self {- id,- name,- models: vec![Arc::new(FakeLanguageModel::default())],+ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ _: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let mut state = self.state.lock();+ if state.forbid_requests {+ return async { Err(anyhow!("requests are forbidden").into()) }.boxed();}+ let (events, rx) = mpsc::unbounded();+ state.pending_completions.push(PendingCompletion {+ model_id: model.id.clone(),+ request,+ events,+ });+ async move { Ok(rx.boxed()) }.boxed()}- pub fn with_models(mut self, models: Vec<Arc<dyn LanguageModel>>) -> Self {- self.models = models;- self- }-- pub fn test_model(&self) -> FakeLanguageModel {- FakeLanguageModel::default()- }-}--#[derive(Debug, PartialEq)]-pub struct ToolUseRequest {- pub request: LanguageModelRequest,- pub name: String,- pub description: String,- pub schema: serde_json::Value,-}--pub struct FakeLanguageModel {- id: LanguageModelId,- name: LanguageModelName,- provider_id: LanguageModelProviderId,- provider_name: LanguageModelProviderName,- current_completion_txs: Mutex<- Vec<(- LanguageModelRequest,- mpsc::UnboundedSender<- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- )>,- >,- forbid_requests: AtomicBool,- supports_thinking: AtomicBool,- supports_disabling_thinking: AtomicBool,- supports_streaming_tools: AtomicBool,- supports_images: AtomicBool,- supports_server_side_compaction: AtomicBool,- max_token_count: AtomicU64,- max_input_tokens: Option<u64>,- max_output_tokens: AtomicU64,- input_token_counts: Mutex<VecDeque<u64>>,- input_token_count_requests: Mutex<Vec<LanguageModelRequest>>,-}--impl Default for FakeLanguageModel {- fn default() -> Self {- Self {- id: LanguageModelId::from("fake".to_string()),- name: LanguageModelName::from("Fake".to_string()),- provider_id: LanguageModelProviderId::from("fake".to_string()),- provider_name: LanguageModelProviderName::from("Fake".to_string()),- current_completion_txs: Mutex::new(Vec::new()),- forbid_requests: AtomicBool::new(false),- supports_thinking: AtomicBool::new(false),- supports_disabling_thinking: AtomicBool::new(true),- supports_streaming_tools: AtomicBool::new(false),- supports_images: AtomicBool::new(false),- supports_server_side_compaction: AtomicBool::new(false),- max_token_count: AtomicU64::new(1_000_000),- max_input_tokens: None,- max_output_tokens: AtomicU64::new(0),- input_token_counts: Mutex::new(VecDeque::new()),- input_token_count_requests: Mutex::new(Vec::new()),+ fn count_input_tokens(+ &self,+ _model: &LanguageModel,+ request: LanguageModelRequest,+ _: &AsyncApp,+ ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {+ let mut state = self.state.lock();+ if state.forbid_requests {+ return async { Err(anyhow!("requests are forbidden").into()) }.boxed();}+ state.input_token_count_requests.push(request);+ let count = state.input_token_counts.pop_front();+ async move { Ok(count) }.boxed()}}-impl FakeLanguageModel {- pub fn with_id_and_thinking(- provider_id: &str,- id: &str,- name: &str,- supports_thinking: bool,- ) -> Self {- Self {- id: LanguageModelId::from(id.to_string()),- name: LanguageModelName::from(name.to_string()),- provider_id: LanguageModelProviderId::from(provider_id.to_string()),- supports_thinking: AtomicBool::new(supports_thinking),- ..Default::default()- }+impl FakeLanguageModelProvider {+ /// A provider offering one model, with id `fake`.+ pub fn new(id: LanguageModelProviderId, name: LanguageModelProviderName) -> Self {+ let provider = Self {+ id,+ name,+ state: Mutex::default(),+ };+ provider.update_model("fake", |model| {+ model.name = LanguageModelName::from("Fake".to_string())+ });+ provider+ }++ /// The model this provider offers as `id`, adding it with default+ /// capabilities if it isn't offered yet.+ pub fn model(&self, id: &str) -> LanguageModel {+ self.update_model(id, |_| {})+ }++ /// Changes the capabilities of the model offered as `id`, adding it first+ /// if needed, and returns the updated model.+ pub fn update_model(&self, id: &str, update: impl FnOnce(&mut LanguageModel)) -> LanguageModel {+ let mut state = self.state.lock();+ let index = match state+ .models+ .iter()+ .position(|model| model.id.0.as_ref() == id)+ {+ Some(index) => index,+ None => {+ state.models.push(LanguageModel {+ supports_disabling_thinking: true,+ ..LanguageModel::new(+ LanguageModelId::from(id.to_string()),+ LanguageModelName::from(id.to_string()),+ self.id.clone(),+ self.name.clone(),+ "fake",+ 1_000_000,+ )+ });+ state.models.len() - 1+ }+ };+ update(&mut state.models[index]);+ state.models[index].clone()}pub fn allow_requests(&self) {- self.forbid_requests.store(false, SeqCst);+ self.state.lock().forbid_requests = false;}pub fn forbid_requests(&self) {- self.forbid_requests.store(true, SeqCst);- }-- pub fn set_supports_thinking(&self, supports: bool) {- self.supports_thinking.store(supports, SeqCst);- }-- pub fn set_supports_disabling_thinking(&self, supports: bool) {- self.supports_disabling_thinking.store(supports, SeqCst);- }-- pub fn set_supports_streaming_tools(&self, supports: bool) {- self.supports_streaming_tools.store(supports, SeqCst);- }-- pub fn set_supports_images(&self, supports: bool) {- self.supports_images.store(supports, SeqCst);- }-- pub fn set_supports_server_side_compaction(&self, supports: bool) {- self.supports_server_side_compaction.store(supports, SeqCst);- }-- pub fn set_max_token_count(&self, count: u64) {- self.max_token_count.store(count, SeqCst);- }-- pub fn set_max_input_tokens(&mut self, count: u64) {- self.max_input_tokens = Some(count);- }-- pub fn set_max_output_tokens(&self, count: Option<u64>) {- self.max_output_tokens- .store(count.unwrap_or_default(), SeqCst);+ self.state.lock().forbid_requests = true;}pub fn queue_input_token_count(&self, count: u64) {- self.input_token_counts.lock().push_back(count);+ self.state.lock().input_token_counts.push_back(count);}pub fn input_token_count_requests(&self) -> Vec<LanguageModelRequest> {- self.input_token_count_requests.lock().clone()+ self.state.lock().input_token_count_requests.clone()}+ /// Requests still streaming, to any model, oldest first.pub fn pending_completions(&self) -> Vec<LanguageModelRequest> {- self.current_completion_txs+ self.state.lock()+ .pending_completions.iter()- .map(|(request, _)| request.clone())+ .map(|pending| pending.request.clone())+ .collect()+ }++ /// Requests still streaming to `model`, oldest first.+ pub fn pending_completions_for(&self, model: &LanguageModel) -> Vec<LanguageModelRequest> {+ self.state+ .lock()+ .pending_completions+ .iter()+ .filter(|pending| pending.model_id == model.id)+ .map(|pending| pending.request.clone()).collect()}pub fn completion_count(&self) -> usize {- self.current_completion_txs.lock().len()+ self.state.lock().pending_completions.len()}- pub fn send_completion_stream_text_chunk(+ /// Streams `chunk` as text to the oldest pending `request` to `model`.+ pub fn send_text(&self,+ model: &LanguageModel,request: &LanguageModelRequest,chunk: impl Into<String>,) {- self.send_completion_stream_event(+ self.send_event(+ model,request,LanguageModelCompletionEvent::Text(chunk.into()),);}- pub fn send_completion_stream_event(+ /// Streams `event` to the oldest pending `request` to `model`.+ pub fn send_event(&self,+ model: &LanguageModel,request: &LanguageModelRequest,event: impl Into<LanguageModelCompletionEvent>,) {- let current_completion_txs = self.current_completion_txs.lock();- let tx = current_completion_txs- .iter()- .find(|(req, _)| req == request)- .map(|(_, tx)| tx)- .unwrap();- tx.unbounded_send(Ok(event.into())).unwrap();+ let state = self.state.lock();+ state.send(state.pending_index(model, request), Ok(event.into()));}- pub fn send_completion_stream_error(+ /// Streams `error` to the oldest pending `request` to `model`.+ pub fn send_error(&self,+ model: &LanguageModel,request: &LanguageModelRequest,error: impl Into<LanguageModelCompletionError>,) {- let current_completion_txs = self.current_completion_txs.lock();- let tx = current_completion_txs- .iter()- .find(|(req, _)| req == request)- .map(|(_, tx)| tx)- .unwrap();- tx.unbounded_send(Err(error.into())).unwrap();+ let state = self.state.lock();+ state.send(state.pending_index(model, request), Err(error.into()));}- pub fn end_completion_stream(&self, request: &LanguageModelRequest) {- self.current_completion_txs- .lock()- .retain(|(req, _)| req != request);+ /// Ends the stream of the oldest pending `request` to `model`.+ pub fn end_stream(&self, model: &LanguageModel, request: &LanguageModelRequest) {+ let mut state = self.state.lock();+ let index = state.pending_index(model, request);+ state.pending_completions.remove(index);}- pub fn is_completion_stream_closed(&self, request: &LanguageModelRequest) -> bool {- self.current_completion_txs+ /// Whether the oldest pending `request` to `model` has ended or had its+ /// receiver dropped.+ pub fn is_stream_closed(&self, model: &LanguageModel, request: &LanguageModelRequest) -> bool {+ self.state.lock()+ .pending_completions.iter()- .find(|(pending_request, _)| pending_request == request)- .is_none_or(|(_, sender)| sender.is_closed())+ .find(|pending| pending.model_id == model.id && &pending.request == request)+ .is_none_or(|pending| pending.events.is_closed())}- pub fn send_last_completion_stream_text_chunk(&self, chunk: impl Into<String>) {- self.send_completion_stream_text_chunk(self.pending_completions().last().unwrap(), chunk);+ /// Streams `chunk` as text to the most recent pending request to `model`.+ pub fn send_last_text(&self, model: &LanguageModel, chunk: impl Into<String>) {+ self.send_last_event(model, LanguageModelCompletionEvent::Text(chunk.into()));}- pub fn send_last_completion_stream_event(+ /// Streams `event` to the most recent pending request to `model`.+ pub fn send_last_event(&self,+ model: &LanguageModel,event: impl Into<LanguageModelCompletionEvent>,) {- self.send_completion_stream_event(self.pending_completions().last().unwrap(), event);+ let state = self.state.lock();+ state.send(state.last_pending_index(model), Ok(event.into()));}- pub fn send_last_completion_stream_error(+ /// Streams `error` to the most recent pending request to `model`.+ pub fn send_last_error(&self,+ model: &LanguageModel,error: impl Into<LanguageModelCompletionError>,) {- self.send_completion_stream_error(self.pending_completions().last().unwrap(), error);+ let state = self.state.lock();+ state.send(state.last_pending_index(model), Err(error.into()));}- pub fn end_last_completion_stream(&self) {- self.end_completion_stream(self.pending_completions().last().unwrap());+ /// Ends the stream of the most recent pending request to `model`.+ pub fn end_last(&self, model: &LanguageModel) {+ let mut state = self.state.lock();+ let index = state.last_pending_index(model);+ state.pending_completions.remove(index);}}-impl LanguageModel for FakeLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- self.name.clone()- }-- fn provider_id(&self) -> LanguageModelProviderId {- self.provider_id.clone()- }-- fn provider_name(&self) -> LanguageModelProviderName {- self.provider_name.clone()- }-- fn supports_tools(&self) -> bool {- false- }-- fn supports_tool_choice(&self, _choice: LanguageModelToolChoice) -> bool {- false- }-- fn supports_images(&self) -> bool {- self.supports_images.load(SeqCst)- }-- fn supports_server_side_compaction(&self) -> bool {- self.supports_server_side_compaction.load(SeqCst)- }-- fn supports_thinking(&self) -> bool {- self.supports_thinking.load(SeqCst)- }-- fn supports_disabling_thinking(&self) -> bool {- self.supports_disabling_thinking.load(SeqCst)- }-- fn supports_streaming_tools(&self) -> bool {- self.supports_streaming_tools.load(SeqCst)- }-- fn telemetry_id(&self) -> String {- "fake".to_string()- }-- fn max_token_count(&self) -> u64 {- self.max_token_count.load(SeqCst)- }-- fn max_input_tokens(&self) -> u64 {- self.max_input_tokens- .unwrap_or_else(|| self.max_token_count())+impl FakeProviderState {+ fn pending_index(&self, model: &LanguageModel, request: &LanguageModelRequest) -> usize {+ self.pending_completions+ .iter()+ .position(|pending| pending.model_id == model.id && &pending.request == request)+ .unwrap_or_else(|| {+ panic!(+ "no pending completion to model `{}` matches the request",+ model.id.0+ )+ })}- fn max_output_tokens(&self) -> Option<u64> {- let max_output_tokens = self.max_output_tokens.load(SeqCst);- if max_output_tokens == 0 {- None- } else {- Some(max_output_tokens)- }+ fn last_pending_index(&self, model: &LanguageModel) -> usize {+ self.pending_completions+ .iter()+ .rposition(|pending| pending.model_id == model.id)+ .unwrap_or_else(|| panic!("no pending completion to model `{}`", model.id.0))}- fn count_input_tokens(+ fn send(&self,- request: LanguageModelRequest,- _: &AsyncApp,- ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {- if self.forbid_requests.load(SeqCst) {- return async {- Err(LanguageModelCompletionError::Other(anyhow!(- "requests are forbidden"- )))- }- .boxed();- }- self.input_token_count_requests.lock().push(request);- let count = self.input_token_counts.lock().pop_front();- async move { Ok(count) }.boxed()- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- _: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- if self.forbid_requests.load(SeqCst) {- async move {- Err(LanguageModelCompletionError::Other(anyhow!(- "requests are forbidden"- )))- }- .boxed()- } else {- let (tx, rx) = mpsc::unbounded();- self.current_completion_txs.lock().push((request, tx));- async move { Ok(rx.boxed()) }.boxed()- }+ index: usize,+ event: Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,+ ) {+ self.pending_completions[index]+ .events+ .unbounded_send(event)+ .unwrap();}+}- fn as_fake(&self) -> &Self {- self- }+#[derive(Debug, PartialEq)]+pub struct ToolUseRequest {+ pub request: LanguageModelRequest,+ pub name: String,+ pub description: String,+ pub schema: serde_json::Value,}diff --git a/crates/language_model/src/language_model.rs b/crates/language_model/src/language_model.rsindex dc12a4559bf..5cd16b10063 100644--- a/crates/language_model/src/language_model.rs+++ b/crates/language_model/src/language_model.rs@@ -88,187 +88,324 @@ impl Default for LanguageModelTextStream {}}-pub trait LanguageModel: Send + Sync {- fn id(&self) -> LanguageModelId;- fn name(&self) -> LanguageModelName;- fn provider_id(&self) -> LanguageModelProviderId;- fn provider_name(&self) -> LanguageModelProviderName;- fn upstream_provider_id(&self) -> LanguageModelProviderId {- self.provider_id()+/// A language model offered by a [`LanguageModelProvider`].+///+/// This is plain data: identity, capabilities, and limits. Everything that+/// talks to a backend (streaming completions, counting tokens, compaction)+/// lives on the provider, which is handed the model it should serve and+/// resolves its own configuration for it by `id`. Find that provider with+/// [`LanguageModelRegistry::provider_for_model`].+#[derive(Clone, Debug, PartialEq)]+pub struct LanguageModel {+ /// Identifies the model within its provider. Stable across provider+ /// refreshes; the name and capabilities may change.+ pub id: LanguageModelId,+ pub name: LanguageModelName,+ pub provider_id: LanguageModelProviderId,+ pub provider_name: LanguageModelProviderName,+ /// The provider that ultimately serves requests, when it differs from+ /// `provider_id` (for example, a model offered through a gateway).+ pub upstream_provider_id: Option<LanguageModelProviderId>,+ pub upstream_provider_name: Option<LanguageModelProviderName>,+ pub telemetry_id: SharedString,+ /// Whether this model is the "latest", so we can highlight it in the UI.+ pub is_latest: bool,+ /// Why the model is currently disabled, if it is.+ pub disabled_reason: Option<DisabledReason>,+ /// Whether requests to this model require the user to consent to the+ /// upstream provider retaining inference logs (i.e. the model cannot be+ /// offered with Zero Data Retention).+ pub requires_data_retention: bool,+ /// When this model refuses a request, the model ID to fall back to (same provider).+ pub refusal_fallback_model_id: Option<&'static str>,+ /// Information about the cost of using this model, if available.+ pub cost_info: Option<LanguageModelCostInfo>,+ pub supports_thinking: bool,+ /// Whether thinking can be turned off entirely for this model. Some+ /// models (e.g. Claude Fable 5) always think and cannot honor an "off"+ /// request. Only meaningful when `supports_thinking` is `true`.+ pub supports_disabling_thinking: bool,+ pub supports_fast_mode: bool,+ /// The effort levels that can be used when thinking.+ pub supported_effort_levels: Arc<[LanguageModelEffortLevel]>,+ /// Whether this model supports provider-side automatic context+ /// compaction (requested via `LanguageModelRequest::compact_at_tokens`).+ pub supports_server_side_compaction: bool,+ pub supports_explicit_compaction: bool,+ /// Whether native compaction honors `LanguageModelRequest::max_output_tokens`.+ pub supports_explicit_compaction_output_limit: bool,+ /// The provider-enforced input size required for explicit compaction.+ pub minimum_explicit_compaction_input_tokens: Option<u64>,+ pub supports_images: bool,+ pub supports_tools: bool,+ pub tool_choice_support: LanguageModelToolChoiceSupport,+ /// Whether this model or provider supports streaming tool calls.+ pub supports_streaming_tools: bool,+ /// Whether this model/provider reports accurate split input/output token+ /// counts. When true, the UI may show separate input/output token indicators.+ pub supports_split_token_display: bool,+ /// The model's context-window capacity.+ pub max_token_count: u64,+ /// The input ceiling before reserving output from any shared window.+ /// Equals `max_token_count` unless the model has a separate prompt limit.+ pub max_input_tokens: u64,+ pub max_output_tokens: Option<u64>,+}++/// Which [`LanguageModelToolChoice`] values a model accepts.+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]+pub struct LanguageModelToolChoiceSupport {+ pub auto: bool,+ pub any: bool,+ pub none: bool,+}++impl LanguageModelToolChoiceSupport {+ pub const ALL: Self = Self {+ auto: true,+ any: true,+ none: true,+ };++ pub fn supports(&self, choice: LanguageModelToolChoice) -> bool {+ match choice {+ LanguageModelToolChoice::Auto => self.auto,+ LanguageModelToolChoice::Any => self.any,+ LanguageModelToolChoice::None => self.none,+ }}- fn upstream_provider_name(&self) -> LanguageModelProviderName {- self.provider_name()+}++impl LanguageModel {+ /// Creates a model with the given identity and conservative defaults for+ /// every capability. Providers override the capabilities they support+ /// with struct update syntax.+ pub fn new(+ id: LanguageModelId,+ name: LanguageModelName,+ provider_id: LanguageModelProviderId,+ provider_name: LanguageModelProviderName,+ telemetry_id: impl Into<SharedString>,+ max_token_count: u64,+ ) -> Self {+ Self {+ id,+ name,+ provider_id,+ provider_name,+ upstream_provider_id: None,+ upstream_provider_name: None,+ telemetry_id: telemetry_id.into(),+ is_latest: false,+ disabled_reason: None,+ requires_data_retention: false,+ refusal_fallback_model_id: None,+ cost_info: None,+ supports_thinking: false,+ supports_disabling_thinking: true,+ supports_fast_mode: false,+ supported_effort_levels: Arc::default(),+ supports_server_side_compaction: false,+ supports_explicit_compaction: false,+ supports_explicit_compaction_output_limit: false,+ minimum_explicit_compaction_input_tokens: None,+ supports_images: false,+ supports_tools: false,+ tool_choice_support: LanguageModelToolChoiceSupport::default(),+ supports_streaming_tools: false,+ supports_split_token_display: false,+ max_token_count,+ max_input_tokens: max_token_count,+ max_output_tokens: None,+ }}- /// Returns whether this model is the "latest", so we can highlight it in the UI.- fn is_latest(&self) -> bool {- false+ pub fn id(&self) -> LanguageModelId {+ self.id.clone()}- /// Whether the model is currently disabled and, if so, why this is the case.- fn is_disabled(&self) -> Option<DisabledReason> {- None+ pub fn name(&self) -> LanguageModelName {+ self.name.clone()}- /// Whether requests to this model require the user to consent to the- /// upstream provider retaining inference logs (i.e. the model cannot be- /// offered with Zero Data Retention).- fn requires_data_retention(&self) -> bool {- false+ pub fn provider_id(&self) -> LanguageModelProviderId {+ self.provider_id.clone()}- /// When this model refuses a request, the model ID to fall back to (same provider).- fn refusal_fallback_model_id(&self) -> Option<&'static str> {- None+ pub fn provider_name(&self) -> LanguageModelProviderName {+ self.provider_name.clone()}- fn telemetry_id(&self) -> String;+ pub fn upstream_provider_id(&self) -> LanguageModelProviderId {+ self.upstream_provider_id+ .clone()+ .unwrap_or_else(|| self.provider_id.clone())+ }- fn api_key(&self, _cx: &App) -> Option<String> {- None+ pub fn upstream_provider_name(&self) -> LanguageModelProviderName {+ self.upstream_provider_name+ .clone()+ .unwrap_or_else(|| self.provider_name.clone())}- /// Information about the cost of using this model, if available.- fn model_cost_info(&self) -> Option<LanguageModelCostInfo> {- None+ pub fn is_latest(&self) -> bool {+ self.is_latest}- /// Whether this model supports thinking.- fn supports_thinking(&self) -> bool {- false+ pub fn is_disabled(&self) -> Option<DisabledReason> {+ self.disabled_reason.clone()}- /// Whether thinking can be turned off entirely for this model. Some- /// models (e.g. Claude Fable 5) always think and cannot honor an "off"- /// request. Only meaningful when `supports_thinking` returns `true`.- fn supports_disabling_thinking(&self) -> bool {- true+ pub fn requires_data_retention(&self) -> bool {+ self.requires_data_retention}- fn supports_fast_mode(&self) -> bool {- false+ pub fn refusal_fallback_model_id(&self) -> Option<&'static str> {+ self.refusal_fallback_model_id}- /// Returns the list of supported effort levels that can be used when thinking.- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- Vec::new()+ pub fn telemetry_id(&self) -> String {+ self.telemetry_id.to_string()+ }++ pub fn model_cost_info(&self) -> Option<LanguageModelCostInfo> {+ self.cost_info.clone()+ }++ pub fn supports_thinking(&self) -> bool {+ self.supports_thinking+ }++ pub fn supports_disabling_thinking(&self) -> bool {+ self.supports_disabling_thinking+ }++ pub fn supports_fast_mode(&self) -> bool {+ self.supports_fast_mode+ }++ pub fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {+ self.supported_effort_levels.to_vec()}/// Returns the default effort level to use when thinking.- fn default_effort_level(&self) -> Option<LanguageModelEffortLevel> {- self.supported_effort_levels()- .into_iter()+ pub fn default_effort_level(&self) -> Option<LanguageModelEffortLevel> {+ self.supported_effort_levels+ .iter().find(|effort_level| effort_level.is_default)+ .cloned()}- /// Whether this model supports provider-side automatic context- /// compaction (requested via `LanguageModelRequest::compact_at_tokens`).- fn supports_server_side_compaction(&self) -> bool {- false+ pub fn supports_server_side_compaction(&self) -> bool {+ self.supports_server_side_compaction}- fn supports_explicit_compaction(&self) -> bool {- false+ pub fn supports_explicit_compaction(&self) -> bool {+ self.supports_explicit_compaction}- /// Whether native compaction honors `LanguageModelRequest::max_output_tokens`.- fn supports_explicit_compaction_output_limit(&self) -> bool {- false+ pub fn supports_explicit_compaction_output_limit(&self) -> bool {+ self.supports_explicit_compaction_output_limit}- /// The provider-enforced input size required for explicit compaction.- fn minimum_explicit_compaction_input_tokens(&self) -> Option<u64> {- None+ pub fn minimum_explicit_compaction_input_tokens(&self) -> Option<u64> {+ self.minimum_explicit_compaction_input_tokens}- fn compact(- &self,- _request: LanguageModelRequest,- _cx: &AsyncApp,- ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- let provider = self.provider_name();- async move {- Err(LanguageModelCompletionError::Other(anyhow::anyhow!(- "{provider} does not support explicit compaction"- )))- }- .boxed()+ pub fn supports_images(&self) -> bool {+ self.supports_images}- /// Whether this model supports images- fn supports_images(&self) -> bool;-- /// Whether this model supports tools.- fn supports_tools(&self) -> bool;+ pub fn supports_tools(&self) -> bool {+ self.supports_tools+ }- /// Whether this model supports choosing which tool to use.- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool;+ pub fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {+ self.tool_choice_support.supports(choice)+ }- /// Returns whether this model or provider supports streaming tool calls;- fn supports_streaming_tools(&self) -> bool {- false+ pub fn supports_streaming_tools(&self) -> bool {+ self.supports_streaming_tools}- /// Returns whether this model/provider reports accurate split input/output token counts.- /// When true, the UI may show separate input/output token indicators.- fn supports_split_token_display(&self) -> bool {- false+ pub fn supports_split_token_display(&self) -> bool {+ self.supports_split_token_display}- /// Returns the model's context-window capacity.- fn max_token_count(&self) -> u64;+ pub fn max_token_count(&self) -> u64 {+ self.max_token_count+ }/// Returns the input ceiling before reserving output from any shared window.- ///- /// Models with a separate prompt limit override the context-window default.- fn max_input_tokens(&self) -> u64 {- self.max_token_count()+ pub fn max_input_tokens(&self) -> u64 {+ self.max_input_tokens}- /// Counts request input without generating output, when supported by the provider.+ /// Returns the combined input and output ceiling, if one applies.///- /// Counts may be estimates and differ from subsequent measured usage. Callers- /// choose the content to count; this does not infer which input is already- /// covered by a previous usage report. Unsupported providers return `None`.- fn count_input_tokens(- &self,- _request: LanguageModelRequest,- _cx: &AsyncApp,- ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {- async { Ok(None) }.boxed()+ /// This shares the context window with output.+ pub fn max_total_tokens(&self) -> Option<u64> {+ Some(self.max_token_count)}- /// Returns the combined input and output ceiling, if one applies.- ///- /// The conservative default shares the context window with output. `None`- /// means generation does not consume that window, not merely that the API- /// validates input separately or stops generation at the window boundary.- fn max_total_tokens(&self) -> Option<u64> {- Some(self.max_token_count())+ pub fn max_output_tokens(&self) -> Option<u64> {+ self.max_output_tokens}+}- fn max_output_tokens(&self) -> Option<u64> {- None+/// The error for a request to `model` when its provider doesn't offer it.+pub fn unavailable_error(model: &LanguageModel) -> LanguageModelCompletionError {+ LanguageModelCompletionError::ModelUnavailable {+ provider: model.provider_name.clone(),+ model: model.id.clone(),+ }+}++/// Either a built-in icon name or a path to an external SVG.+#[derive(Debug, Clone, PartialEq, Eq)]+pub enum IconOrSvg {+ /// A built-in icon from Zed's icon set.+ Icon(IconName),+ /// Path to a custom SVG icon file.+ Svg(SharedString),+}++impl Default for IconOrSvg {+ fn default() -> Self {+ Self::Icon(IconName::ZedAssistant)}+}+pub trait LanguageModelProvider: 'static {+ fn id(&self) -> LanguageModelProviderId;+ fn name(&self) -> LanguageModelProviderName;+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::default()+ }+ fn default_model(&self, cx: &App) -> Option<LanguageModel>;+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel>;+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel>;+ fn recommended_models(&self, _cx: &App) -> Vec<LanguageModel> {+ Vec::new()+ }++ /// Streams a completion of `request` from `model`, which must be one of+ /// this provider's models.fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- >;+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>;+ /// Streams the text of a completion of `request` from `model`.fn stream_completion_text(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<LanguageModelTextStream, LanguageModelCompletionError>> {- let future = self.stream_completion(request, cx);+ let future = self.stream_completion(model, request, cx);async move {let events = future.await?;@@ -329,18 +466,19 @@ pub trait LanguageModel: Send + Sync {.boxed()}+ /// Completes `request` from `model`, resolving to its first complete tool use.fn stream_completion_tool(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<LanguageModelToolUse, LanguageModelCompletionError>> {- let future = self.stream_completion(request, cx);+ let future = self.stream_completion(model, request, cx);async move {let events = future.await?;let mut events = events.fuse();- // Iterate through events until we find a complete ToolUsewhile let Some(event) = events.next().await {match event {Ok(LanguageModelCompletionEvent::ToolUse(tool_use))@@ -355,7 +493,6 @@ pub trait LanguageModel: Send + Sync {}}- // Stream ended without a complete tool useErr(LanguageModelCompletionError::Other(anyhow::anyhow!("Stream ended without receiving a complete tool use")))@@ -363,54 +500,42 @@ pub trait LanguageModel: Send + Sync {.boxed()}- #[cfg(any(test, feature = "test-support"))]- fn as_fake(&self) -> &fake_provider::FakeLanguageModel {- unimplemented!()+ /// Counts request input without generating output, when supported by the provider.+ ///+ /// Counts may be estimates and differ from subsequent measured usage. Callers+ /// choose the content to count; this does not infer which input is already+ /// covered by a previous usage report. Unsupported providers return `None`.+ fn count_input_tokens(+ &self,+ _model: &LanguageModel,+ _request: LanguageModelRequest,+ _cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {+ async { Ok(None) }.boxed()}-}-impl std::fmt::Debug for dyn LanguageModel {- fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {- f.debug_struct("<dyn LanguageModel>")- .field("id", &self.id())- .field("name", &self.name())- .field("provider_id", &self.provider_id())- .field("provider_name", &self.provider_name())- .field("upstream_provider_name", &self.upstream_provider_name())- .field("upstream_provider_id", &self.upstream_provider_id())- .field("upstream_provider_id", &self.upstream_provider_id())- .field("supports_streaming_tools", &self.supports_streaming_tools())- .finish()+ /// Compacts `request` into replacement context, for models that report+ /// [`LanguageModel::supports_explicit_compaction`].+ fn compact(+ &self,+ model: &LanguageModel,+ _request: LanguageModelRequest,+ _cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let provider = model.provider_name.clone();+ async move {+ Err(LanguageModelCompletionError::Other(anyhow::anyhow!(+ "{provider} does not support explicit compaction"+ )))+ }+ .boxed()}-}--/// Either a built-in icon name or a path to an external SVG.-#[derive(Debug, Clone, PartialEq, Eq)]-pub enum IconOrSvg {- /// A built-in icon from Zed's icon set.- Icon(IconName),- /// Path to a custom SVG icon file.- Svg(SharedString),-}-impl Default for IconOrSvg {- fn default() -> Self {- Self::Icon(IconName::ZedAssistant)+ /// The API key used to serve `model`, if this provider uses one.+ fn api_key(&self, _model: &LanguageModel, _cx: &App) -> Option<String> {+ None}-}-pub trait LanguageModelProvider: 'static {- fn id(&self) -> LanguageModelProviderId;- fn name(&self) -> LanguageModelProviderName;- fn icon(&self) -> IconOrSvg {- IconOrSvg::default()- }- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>>;- fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {- Vec::new()- }fn is_authenticated(&self, cx: &App) -> bool;fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>>;fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView>;diff --git a/crates/language_model/src/registry.rs b/crates/language_model/src/registry.rsindex 43e243c7100..011a1d4efb3 100644--- a/crates/language_model/src/registry.rs+++ b/crates/language_model/src/registry.rs@@ -1,6 +1,6 @@use crate::{- LanguageModel, LanguageModelId, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderState, ZED_CLOUD_PROVIDER_ID,+ LanguageModel, LanguageModelCompletionError, LanguageModelId, LanguageModelProvider,+ LanguageModelProviderId, LanguageModelProviderState, ZED_CLOUD_PROVIDER_ID, unavailable_error,};use collections::{BTreeMap, HashSet};use gpui::{App, Context, Entity, EventEmitter, Global, prelude::*};@@ -46,16 +46,16 @@ impl std::fmt::Debug for ConfigurationError {pub struct LanguageModelRegistry {/// True if the user has *NO* default model configured in settingsshould_use_fallback: bool,- default_model: Option<ConfiguredModel>,+ default_model: Option<LanguageModel>,/// This model is automatically configured by a user's environment after/// authenticating all providers. It's only used when `default_model` is not set.- available_fallback_model: Option<ConfiguredModel>,- inline_assistant_model: Option<ConfiguredModel>,- commit_message_model: Option<ConfiguredModel>,- thread_summary_model: Option<ConfiguredModel>,- compaction_model: Option<ConfiguredModel>,+ available_fallback_model: Option<LanguageModel>,+ inline_assistant_model: Option<LanguageModel>,+ commit_message_model: Option<LanguageModel>,+ thread_summary_model: Option<LanguageModel>,+ compaction_model: Option<LanguageModel>,providers: BTreeMap<LanguageModelProviderId, Arc<dyn LanguageModelProvider>>,- inline_alternatives: Vec<Arc<dyn LanguageModel>>,+ inline_alternatives: Vec<LanguageModel>,/// Set of installed extension IDs that provide language models./// Used to determine which built-in providers should be hidden.installed_llm_extension_ids: HashSet<Arc<str>>,@@ -92,19 +92,15 @@ impl FromStr for SelectedModel {}}-#[derive(Clone)]-pub struct ConfiguredModel {- pub provider: Arc<dyn LanguageModelProvider>,- pub model: Arc<dyn LanguageModel>,-}--impl ConfiguredModel {- pub fn is_same_as(&self, other: &ConfiguredModel) -> bool {- self.model.id() == other.model.id() && self.provider.id() == other.provider.id()+impl LanguageModel {+ /// Whether `other` is the same model from the same provider, regardless+ /// of changes to its capabilities.+ pub fn is_same_as(&self, other: &LanguageModel) -> bool {+ self.id == other.id && self.provider_id == other.provider_id}pub fn is_provided_by_zed(&self) -> bool {- self.provider.id() == ZED_CLOUD_PROVIDER_ID+ self.provider_id == ZED_CLOUD_PROVIDER_ID}}@@ -133,28 +129,20 @@ impl LanguageModelRegistry {}#[cfg(any(test, feature = "test-support"))]+ /// Installs a registry whose only provider is a fake, with its model as+ /// the default, and returns the fake.pub fn test(cx: &mut App) -> Arc<crate::fake_provider::FakeLanguageModelProvider> {let fake_provider = Arc::new(crate::fake_provider::FakeLanguageModelProvider::default());let registry = cx.new(|cx| {let mut registry = Self::default();registry.register_provider(fake_provider.clone(), cx);- let model = fake_provider.provided_models(cx)[0].clone();- let configured_model = ConfiguredModel {- provider: fake_provider.clone(),- model,- };- registry.set_default_model(Some(configured_model), cx);+ registry.set_default_model(Some(fake_provider.model("fake")), cx);registry});cx.set_global(GlobalLanguageModelRegistry(registry));fake_provider}- #[cfg(any(test, feature = "test-support"))]- pub fn fake_model(&self) -> Arc<dyn LanguageModel> {- self.default_model.as_ref().unwrap().model.clone()- }-pub fn set_should_use_fallback(&mut self, value: bool) {self.should_use_fallback = value;}@@ -258,7 +246,7 @@ impl LanguageModelRegistry {pub fn configuration_error(&self,- model: Option<ConfiguredModel>,+ model: Option<LanguageModel>,cx: &App,) -> Option<ConfigurationError> {let Some(model) = model else {@@ -268,8 +256,11 @@ impl LanguageModelRegistry {return Some(ConfigurationError::ModelNotFound);};- if !model.provider.is_authenticated(cx) {- return Some(ConfigurationError::ProviderNotAuthenticated(model.provider));+ let Some(provider) = self.provider(&model.provider_id) else {+ return Some(ConfigurationError::ModelNotFound);+ };+ if !provider.is_authenticated(cx) {+ return Some(ConfigurationError::ProviderNotAuthenticated(provider));}None@@ -280,10 +271,7 @@ impl LanguageModelRegistry {self.providers.values().any(|p| p.is_authenticated(cx))}- pub fn available_models<'a>(- &'a self,- cx: &'a App,- ) -> impl Iterator<Item = Arc<dyn LanguageModel>> + 'a {+ pub fn available_models<'a>(&'a self, cx: &'a App) -> impl Iterator<Item = LanguageModel> + 'a {self.providers.values().filter(|provider| provider.is_authenticated(cx))@@ -294,6 +282,20 @@ impl LanguageModelRegistry {self.providers.get(id).cloned()}+ /// The registered provider that serves `model`.+ ///+ /// # Errors+ ///+ /// Returns [`LanguageModelCompletionError::ModelUnavailable`] when no+ /// provider is registered under the model's provider id.+ pub fn provider_for_model(+ &self,+ model: &LanguageModel,+ ) -> Result<Arc<dyn LanguageModelProvider>, LanguageModelCompletionError> {+ self.provider(&model.provider_id)+ .ok_or_else(|| unavailable_error(model))+ }+pub fn select_default_model(&mut self, model: Option<&SelectedModel>, cx: &mut Context<Self>) {let configured_model = model.and_then(|model| self.select_model(model, cx));self.set_default_model(configured_model, cx);@@ -344,10 +346,7 @@ impl LanguageModelRegistry {) {self.inline_alternatives = alternatives.into_iter()- .flat_map(|alternative| {- self.select_model(&alternative, cx)- .map(|configured_model| configured_model.model)- })+ .flat_map(|alternative| self.select_model(&alternative, cx)).collect::<Vec<_>>();}@@ -355,17 +354,14 @@ impl LanguageModelRegistry {&mut self,selected_model: &SelectedModel,cx: &mut Context<Self>,- ) -> Option<ConfiguredModel> {- let provider = self.provider(&selected_model.provider)?;- let model = provider+ ) -> Option<LanguageModel> {+ self.provider(&selected_model.provider)?.provided_models(cx)- .iter()- .find(|model| model.id() == selected_model.model)?- .clone();- Some(ConfiguredModel { provider, model })+ .into_iter()+ .find(|model| model.id == selected_model.model)}- pub fn set_default_model(&mut self, model: Option<ConfiguredModel>, cx: &mut Context<Self>) {+ pub fn set_default_model(&mut self, model: Option<LanguageModel>, cx: &mut Context<Self>) {match (self.default_model(), model.as_ref()) {(Some(old), Some(new)) if old.is_same_as(new) => {}(None, None) => {}@@ -385,19 +381,15 @@ impl LanguageModelRegistry {.iter().filter(|provider| provider.is_authenticated(cx)).find_map(|provider| {- let model = provider+ provider.default_model(cx)- .or_else(|| provider.recommended_models(cx).first().cloned())?;- Some(ConfiguredModel {- provider: provider.clone(),- model,- })+ .or_else(|| provider.recommended_models(cx).first().cloned())});self.set_fallback_model(fallback_model, cx);}- fn set_fallback_model(&mut self, model: Option<ConfiguredModel>, cx: &mut Context<Self>) {+ fn set_fallback_model(&mut self, model: Option<LanguageModel>, cx: &mut Context<Self>) {if self.default_model.is_none() {match (self.available_fallback_model.as_ref(), model.as_ref()) {(Some(old), Some(new)) if old.is_same_as(new) => {}@@ -410,7 +402,7 @@ impl LanguageModelRegistry {pub fn set_inline_assistant_model(&mut self,- model: Option<ConfiguredModel>,+ model: Option<LanguageModel>,cx: &mut Context<Self>,) {match (self.inline_assistant_model.as_ref(), model.as_ref()) {@@ -423,7 +415,7 @@ impl LanguageModelRegistry {pub fn set_commit_message_model(&mut self,- model: Option<ConfiguredModel>,+ model: Option<LanguageModel>,cx: &mut Context<Self>,) {match (self.commit_message_model.as_ref(), model.as_ref()) {@@ -436,7 +428,7 @@ impl LanguageModelRegistry {pub fn set_thread_summary_model(&mut self,- model: Option<ConfiguredModel>,+ model: Option<LanguageModel>,cx: &mut Context<Self>,) {match (self.thread_summary_model.as_ref(), model.as_ref()) {@@ -447,7 +439,7 @@ impl LanguageModelRegistry {self.thread_summary_model = model;}- pub fn set_compaction_model(&mut self, model: Option<ConfiguredModel>, cx: &mut Context<Self>) {+ pub fn set_compaction_model(&mut self, model: Option<LanguageModel>, cx: &mut Context<Self>) {match (self.compaction_model.as_ref(), model.as_ref()) {(Some(old), Some(new)) if old.is_same_as(new) => {}(None, None) => {}@@ -456,7 +448,7 @@ impl LanguageModelRegistry {self.compaction_model = model;}- pub fn default_model(&self) -> Option<ConfiguredModel> {+ pub fn default_model(&self) -> Option<LanguageModel> {#[cfg(debug_assertions)]if std::env::var("ZED_SIMULATE_NO_LLM_PROVIDER").is_ok() {return None;@@ -471,16 +463,13 @@ impl LanguageModelRegistry {})}- pub fn default_fast_model(&self, cx: &App) -> Option<ConfiguredModel> {- let configured = self.default_model()?;- let fast_model = configured.provider.default_fast_model(cx)?;- Some(ConfiguredModel {- provider: configured.provider,- model: fast_model,- })+ pub fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ let default_model = self.default_model()?;+ self.provider(&default_model.provider_id)?+ .default_fast_model(cx)}- pub fn inline_assistant_model(&self) -> Option<ConfiguredModel> {+ pub fn inline_assistant_model(&self) -> Option<LanguageModel> {#[cfg(debug_assertions)]if std::env::var("ZED_SIMULATE_NO_LLM_PROVIDER").is_ok() {return None;@@ -491,7 +480,7 @@ impl LanguageModelRegistry {.or_else(|| self.default_model())}- pub fn commit_message_model(&self, cx: &App) -> Option<ConfiguredModel> {+ pub fn commit_message_model(&self, cx: &App) -> Option<LanguageModel> {#[cfg(debug_assertions)]if std::env::var("ZED_SIMULATE_NO_LLM_PROVIDER").is_ok() {return None;@@ -503,7 +492,7 @@ impl LanguageModelRegistry {.or_else(|| self.default_model())}- pub fn thread_summary_model(&self, cx: &App) -> Option<ConfiguredModel> {+ pub fn thread_summary_model(&self, cx: &App) -> Option<LanguageModel> {#[cfg(debug_assertions)]if std::env::var("ZED_SIMULATE_NO_LLM_PROVIDER").is_ok() {return None;@@ -518,7 +507,7 @@ impl LanguageModelRegistry {/// Returns the configured compaction model without falling back through/// `default_fast_model`/`default_model`. Callers that want a fallback to/// the thread's primary model should handle `None` themselves.- pub fn compaction_model(&self) -> Option<ConfiguredModel> {+ pub fn compaction_model(&self) -> Option<LanguageModel> {#[cfg(debug_assertions)]if std::env::var("ZED_SIMULATE_NO_LLM_PROVIDER").is_ok() {return None;@@ -530,7 +519,7 @@ impl LanguageModelRegistry {/// The models to use for inline assists. Returns the union of the active/// model and all inline alternatives. When there are multiple models, the/// user will be able to cycle through results.- pub fn inline_alternative_models(&self) -> &[Arc<dyn LanguageModel>] {+ pub fn inline_alternative_models(&self) -> &[LanguageModel] {&self.inline_alternatives}}@@ -538,7 +527,7 @@ impl LanguageModelRegistry {#[cfg(test)]mod tests {use super::*;- use crate::fake_provider::{FakeLanguageModel, FakeLanguageModelProvider};+ use crate::fake_provider::FakeLanguageModelProvider;#[test]fn selected_model_allows_slashes_in_model_id() {@@ -692,13 +681,7 @@ mod tests {let provider = registry.provider(&provider.id()).unwrap();let model = provider.default_model(cx).unwrap();- registry.set_fallback_model(- Some(ConfiguredModel {- provider: provider.clone(),- model: model.clone(),- }),- cx,- );+ registry.set_fallback_model(Some(model.clone()), cx);assert!(registry.default_model().is_none());assert!(registry.inline_assistant_model().is_none());@@ -706,8 +689,8 @@ mod tests {registry.set_should_use_fallback(true);let default_model = registry.default_model().unwrap();- assert_eq!(default_model.model.id(), model.id());- assert_eq!(default_model.provider.id(), provider.id());+ assert_eq!(default_model.id, model.id);+ assert_eq!(default_model.provider_id, provider.id());assert!(registry.inline_assistant_model()@@ -719,14 +702,9 @@ mod tests {#[gpui::test]fn test_inline_assistant_model_precedence(cx: &mut App) {let registry = cx.new(|_| LanguageModelRegistry::default());- let provider = Arc::new(FakeLanguageModelProvider::default());+ let provider = FakeLanguageModelProvider::default();let [inline_model, default_model, fallback_model] =- ["inline", "default", "fallback"].map(|model_id| ConfiguredModel {- provider: provider.clone(),- model: Arc::new(FakeLanguageModel::with_id_and_thinking(- "fake", model_id, model_id, false,- )),- });+ ["inline", "default", "fallback"].map(|model_id| provider.model(model_id));registry.update(cx, |registry, cx| {registry.set_should_use_fallback(true);diff --git a/crates/language_model_core/src/language_model_core.rs b/crates/language_model_core/src/language_model_core.rsindex a91ba901219..25f911d4ced 100644--- a/crates/language_model_core/src/language_model_core.rs+++ b/crates/language_model_core/src/language_model_core.rs@@ -31,6 +31,12 @@ pub use crate::util::{};pub use gpui_shared_string::SharedString;+/// The events of a streaming completion, as produced by a provider.+pub type LanguageModelCompletionStream = futures::stream::BoxStream<+ 'static,+ Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,+>;+/// A completion event from a language model.#[derive(Debug, PartialEq, Clone, Serialize, Deserialize)]pub enum LanguageModelCompletionEvent {@@ -243,6 +249,13 @@ pub enum LanguageModelCompletionError {},#[error("stream from {provider} ended unexpectedly")]StreamEndedUnexpectedly { provider: LanguageModelProviderName },+ /// The provider no longer offers a model with this id, for example after+ /// its settings or fetched model list changed.+ #[error("{provider} no longer offers the model {}", model.0)]+ ModelUnavailable {+ provider: LanguageModelProviderName,+ model: LanguageModelId,+ },#[error(transparent)]Other(#[from] anyhow::Error),}@@ -397,6 +410,7 @@ impl LanguageModelCompletionError {| Self::BuildRequestBody { .. }| Self::DeserializeResponse { .. }| Self::StreamEndedUnexpectedly { .. }+ | Self::ModelUnavailable { .. }| Self::Other(_) => false,}}@@ -426,6 +440,7 @@ impl LanguageModelCompletionError {| Self::BuildRequestBody { .. }| Self::DeserializeResponse { .. }| Self::StreamEndedUnexpectedly { .. }+ | Self::ModelUnavailable { .. }| Self::Other(_) => None,}}@@ -674,7 +689,7 @@ impl LanguageModelToolUseInput {}}-#[derive(Debug, Clone)]+#[derive(Debug, Clone, PartialEq, Eq)]pub struct LanguageModelEffortLevel {pub name: SharedString,pub value: SharedString,diff --git a/crates/language_model_core/src/rate_limiter.rs b/crates/language_model_core/src/rate_limiter.rsindex 28e92056f9d..a2f6a7df9f8 100644--- a/crates/language_model_core/src/rate_limiter.rs+++ b/crates/language_model_core/src/rate_limiter.rs@@ -7,7 +7,53 @@ use std::{task::{Context, Poll},};-use crate::LanguageModelCompletionError;+use crate::{LanguageModelCompletionError, LanguageModelId};+use collections::HashMap;+use std::sync::Mutex;++/// Default number of concurrent requests a provider allows per model.+///+/// One limit is shared by every request to a model across the app: threads,+/// subagents, summaries, and inline assists. It is set well above normal+/// parallel use so it only stops runaway fan-out.+pub const DEFAULT_MODEL_CONCURRENCY: usize = 16;++/// A provider's local concurrency caps, one [`RateLimiter`] per model id.+///+/// Upstream rate limits are per credential and model family, and exceeding+/// them surfaces as retryable errors. This only bounds how many requests Zed+/// has in flight to one model at once.+pub struct ModelRateLimiters {+ limit: usize,+ limiters: Mutex<HashMap<LanguageModelId, RateLimiter>>,+}++impl ModelRateLimiters {+ pub fn new(limit: usize) -> Self {+ Self {+ limit,+ limiters: Mutex::default(),+ }+ }++ /// Returns the limiter shared by every request to `model_id`.+ pub fn for_model(&self, model_id: &LanguageModelId) -> RateLimiter {+ let mut limiters = self+ .limiters+ .lock()+ .unwrap_or_else(|poisoned| poisoned.into_inner());+ limiters+ .entry(model_id.clone())+ .or_insert_with(|| RateLimiter::new(self.limit))+ .clone()+ }+}++impl Default for ModelRateLimiters {+ fn default() -> Self {+ Self::new(DEFAULT_MODEL_CONCURRENCY)+ }+}#[derive(Clone)]pub struct RateLimiter {diff --git a/crates/language_models/Cargo.toml b/crates/language_models/Cargo.tomlindex 552e9e296fc..b4ccec6d853 100644--- a/crates/language_models/Cargo.toml+++ b/crates/language_models/Cargo.toml@@ -17,7 +17,6 @@ test-support = [][dependencies]ai_onboarding.workspace = true-async-lock.workspace = trueanthropic = { workspace = true, features = ["schemars"] }anyhow.workspace = trueaws-config = { workspace = true, features = ["behavior-version-latest"] }@@ -66,7 +65,6 @@ serde_json.workspace = truesettings.workspace = truesha2.workspace = truestrum.workspace = true-tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }ui.workspace = trueui_input.workspace = trueutil.workspace = truediff --git a/crates/language_models/src/provider/anthropic.rs b/crates/language_models/src/provider/anthropic.rsindex e8d380b00fc..91c039c41ca 100644--- a/crates/language_models/src/provider/anthropic.rs+++ b/crates/language_models/src/provider/anthropic.rs@@ -10,10 +10,10 @@ use http_client::{CustomHeaders, HttpClient};use language_model::{ANTHROPIC_PROVIDER_ID, ANTHROPIC_PROVIDER_NAME, ApiKeyConfiguration, ApiKeyState,AuthenticateError, CompactionResult, EnvVar, FastModeConfirmation, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,- LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,- ProviderSettingsView, RateLimiter, env_var,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelId,+ LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,+ LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoiceSupport,+ ModelRateLimiters, ProviderSettingsView, env_var, unavailable_error,};use settings::{Settings, SettingsStore};use std::sync::{Arc, LazyLock};@@ -38,6 +38,7 @@ pub struct AnthropicSettings {pub struct AnthropicLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}const API_KEY_ENV_VAR_NAME: &str = "ANTHROPIC_API_KEY";@@ -173,17 +174,93 @@ impl AnthropicLanguageModelProvider {}});- Self { http_client, state }+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ }}- fn create_language_model(&self, model: anthropic::Model) -> Arc<dyn LanguageModel> {- Arc::new(AnthropicModel {- id: LanguageModelId::from(model.id.to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })+ /// Every model this provider offers, keyed by id: the fetched list, with+ /// settings entries added or overriding fetched ones.+ fn anthropic_models(&self, cx: &App) -> BTreeMap<String, anthropic::Model> {+ let mut models: BTreeMap<String, anthropic::Model> = BTreeMap::default();++ // Models reported by Anthropic's `/v1/models` endpoint are the+ // primary source. The list will be empty until authentication has+ // succeeded and the first fetch completes.+ for model in &self.state.read(cx).fetched_models {+ models.insert(model.id.to_string(), model.clone());+ }++ // User-defined `available_models` from settings can either add+ // entirely new entries or override fields on a fetched model with+ // the same id (e.g. enable Fast mode or set a tool override).+ for available in &AnthropicLanguageModelProvider::settings(cx).available_models {+ let model = available_model_to_anthropic_model(available);+ models.insert(model.id.to_string(), model);+ }++ models+ }++ /// Describes the offered model with `id`.+ fn offered_model(&self, id: &str, cx: &App) -> Option<LanguageModel> {+ self.anthropic_models(cx).get(id).map(language_model)+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<anthropic::Model, LanguageModelCompletionError> {+ self.anthropic_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))+ }++ fn stream_anthropic_request(+ &self,+ config: &anthropic::Model,+ request: anthropic::Request,+ cx: &AsyncApp,+ ) -> BoxFuture<+ 'static,+ Result<+ BoxStream<'static, Result<anthropic::Event, AnthropicError>>,+ LanguageModelCompletionError,+ >,+ > {+ let http_client = self.http_client.clone();++ let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {+ let api_url = AnthropicLanguageModelProvider::api_url(cx);+ let extra_headers = AnthropicLanguageModelProvider::settings(cx)+ .custom_headers+ .clone();+ (state.api_key_state.key(&api_url), api_url, extra_headers)+ });++ let beta_headers = config.beta_headers();++ async move {+ let Some(api_key) = api_key else {+ return Err(LanguageModelCompletionError::NoApiKey {+ provider: PROVIDER_NAME,+ });+ };+ let request = anthropic::stream_completion(+ http_client.as_ref(),+ &api_url,+ &api_key,+ request,+ beta_headers,+ &extra_headers,+ );+ request.await.map_err(Into::into)+ }+ .boxed()}fn settings(cx: &App) -> &AnthropicSettings {@@ -221,49 +298,34 @@ impl LanguageModelProvider for AnthropicLanguageModelProvider {IconOrSvg::Icon(IconName::AiAnthropic)}- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {let fetched = self.state.read(cx).fetched_models.clone();// Pick the highest-version Sonnet we know about; otherwise the first// Claude model returned. Returning `None` until the fetch completes// matches the Ollama provider's behavior.- pick_preferred_model(&fetched, &["claude-sonnet-", "claude-opus-", "claude-"])- .map(|model| self.create_language_model(model))+ let preferred =+ pick_preferred_model(&fetched, &["claude-sonnet-", "claude-opus-", "claude-"])?;+ self.offered_model(&preferred.id, cx)}- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {let fetched = self.state.read(cx).fetched_models.clone();- pick_preferred_model(&fetched, &["claude-haiku-", "claude-"])- .map(|model| self.create_language_model(model))+ let preferred = pick_preferred_model(&fetched, &["claude-haiku-", "claude-"])?;+ self.offered_model(&preferred.id, cx)}- fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn recommended_models(&self, cx: &App) -> Vec<LanguageModel> {let fetched = self.state.read(cx).fetched_models.clone();pick_preferred_model(&fetched, &["claude-sonnet-"])- .map(|model| vec![self.create_language_model(model)])- .unwrap_or_default()+ .and_then(|preferred| self.offered_model(&preferred.id, cx))+ .into_iter()+ .collect()}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let mut models: BTreeMap<String, anthropic::Model> = BTreeMap::default();-- // Models reported by Anthropic's `/v1/models` endpoint are the- // primary source. The list will be empty until authentication has- // succeeded and the first fetch completes.- for model in &self.state.read(cx).fetched_models {- models.insert(model.id.to_string(), model.clone());- }-- // User-defined `available_models` from settings can either add- // entirely new entries or override fields on a fetched model with- // the same id (e.g. enable Fast mode or set a tool override).- for available in &AnthropicLanguageModelProvider::settings(cx).available_models {- let model = available_model_to_anthropic_model(available);- models.insert(model.id.to_string(), model);- }-- models+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.anthropic_models(cx).into_values()- .map(|model| self.create_language_model(model))+ .map(|model| language_model(&model)).collect()}@@ -290,6 +352,153 @@ impl LanguageModelProvider for AnthropicLanguageModelProvider {.update(cx, |state, cx| state.set_api_key(api_key, cx))}+ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let has_tools = !request.tools.is_empty();+ let request_id = config.request_id(has_tools).to_string();+ let mut request = match into_anthropic(+ request,+ request_id,+ config.default_temperature,+ config.max_output_tokens,+ config.mode.clone(),+ AnthropicPromptCacheMode::Automatic,+ &PROVIDER_ID,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ if !config.supports_speed {+ request.speed = None;+ }+ let request = self.stream_anthropic_request(&config, request, cx);+ let executor = cx.background_executor().clone();+ let future = request_limiter.stream(async move {+ let response = request.await?;+ let events = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID).map_stream(response);+ Ok(language_model::stream_in_background(+ events.boxed(),+ executor,+ ))+ });+ async move { Ok(future.await?.boxed()) }.boxed()+ }++ fn count_input_tokens(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request_id = config.request_id(!request.tools.is_empty()).to_string();+ let request = into_anthropic(+ request,+ request_id,+ config.default_temperature,+ config.max_output_tokens,+ config.mode.clone(),+ AnthropicPromptCacheMode::Automatic,+ &PROVIDER_ID,+ );+ let http_client = self.http_client.clone();+ let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {+ let api_url = AnthropicLanguageModelProvider::api_url(cx);+ let extra_headers = AnthropicLanguageModelProvider::settings(cx)+ .custom_headers+ .clone();+ (state.api_key_state.key(&api_url), api_url, extra_headers)+ });+ let beta_headers = config.beta_headers();+ request_limiter+ .run(async move {+ let request = request?.into_count_tokens_request();+ let api_key = api_key.ok_or(LanguageModelCompletionError::NoApiKey {+ provider: PROVIDER_NAME,+ })?;+ anthropic::count_input_tokens(+ http_client.as_ref(),+ &api_url,+ &api_key,+ request,+ beta_headers,+ &extra_headers,+ )+ .await+ .map(Some)+ .map_err(Into::into)+ })+ .boxed()+ }++ fn compact(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if !config.supports_compaction {+ return async {+ Err(LanguageModelCompletionError::Other(anyhow::anyhow!(+ "this Anthropic model does not support explicit compaction"+ )))+ }+ .boxed();+ }++ let mut request = match into_anthropic(+ request,+ config.request_id(false).to_string(),+ config.default_temperature,+ config.max_output_tokens,+ config.mode.clone(),+ AnthropicPromptCacheMode::Automatic,+ &PROVIDER_ID,+ ) {+ Ok(request) => request.into_compact_request(),+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ if !config.supports_speed {+ request.speed = None;+ }+ let request = self.stream_anthropic_request(&config, request, cx);+ let executor = cx.background_executor().clone();+ let future = request_limiter.run(async move {+ let response = request.await?;+ let stream = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID).map_stream(response);+ let stream = language_model::stream_in_background(stream.boxed(), executor);+ let (context, usage) = collect_compaction_result(stream.boxed(), PROVIDER_NAME).await?;+ Ok(CompactionResult { context, usage })+ });+ future.boxed()+ }++ fn api_key(&self, _model: &LanguageModel, cx: &App) -> Option<String> {+ self.state.read_with(cx, |state, cx| {+ let api_url = AnthropicLanguageModelProvider::api_url(cx);+ state.api_key_state.key(&api_url).map(|key| key.to_string())+ })+ }+fn fast_mode_confirmation(&self, _cx: &App) -> Option<FastModeConfirmation> {Some(FastModeConfirmation {title: "Enable Fast Mode for Anthropic?".into(),@@ -464,10 +673,9 @@ mod tests {cx: &mut gpui::TestAppContext,) {let provider = direct_anthropic_test_provider(FakeHttpClient::with_404_response(), cx);- let model = direct_anthropic_test_model(&provider);+ let model = direct_anthropic_test_model(&provider, cx);- assert!(model.supports_explicit_compaction());- assert_eq!(model.max_total_tokens(), Some(model.max_token_count()));+ assert!(model.supports_explicit_compaction);assert_eq!(model.minimum_explicit_compaction_input_tokens(),Some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS)@@ -569,7 +777,7 @@ mod tests {let provider = direct_anthropic_test_provider(http_client, cx);let store_key = cx.update(|cx| provider.set_api_key(Some("test-key".to_string()), cx));store_key.await.unwrap();- let model = direct_anthropic_test_model(&provider);+ let model = direct_anthropic_test_model(&provider, cx);let request = LanguageModelRequest {max_output_tokens: Some(8192),messages: vec![LanguageModelRequestMessage {@@ -581,7 +789,10 @@ mod tests {..Default::default()};- let result = model.compact(request, &cx.to_async()).await.unwrap();+ let result = provider+ .compact(&model, request, &cx.to_async())+ .await+ .unwrap();assert_eq!(result.usage,@@ -644,16 +855,23 @@ mod tests {})}+ /// Adds a listed model to the provider's fetched list, as a successful+ /// `/v1/models` fetch would, and returns it.fn direct_anthropic_test_model(provider: &AnthropicLanguageModelProvider,- ) -> Arc<dyn LanguageModel> {- provider.create_language_model(anthropic::Model::from_listed(anthropic::ListModelEntry {+ cx: &mut gpui::TestAppContext,+ ) -> LanguageModel {+ let model = anthropic::Model::from_listed(anthropic::ListModelEntry {id: "claude-opus-4-6".to_string(),display_name: "Claude Opus 4.6".to_string(),max_input_tokens: 1_000_000,max_tokens: 128_000,capabilities: None,- }))+ });+ provider+ .state+ .update(cx, |state, _| state.fetched_models.push(model.clone()));+ language_model(&model)}struct TestCredentialsProvider;@@ -689,170 +907,29 @@ mod tests {}}-pub struct AnthropicModel {- id: LanguageModelId,- model: anthropic::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--impl AnthropicModel {- fn stream_completion(- &self,- request: anthropic::Request,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<anthropic::Event, AnthropicError>>,- LanguageModelCompletionError,- >,- > {- let http_client = self.http_client.clone();-- let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {- let api_url = AnthropicLanguageModelProvider::api_url(cx);- let extra_headers = AnthropicLanguageModelProvider::settings(cx)- .custom_headers- .clone();- (state.api_key_state.key(&api_url), api_url, extra_headers)- });-- let beta_headers = self.model.beta_headers();-- async move {- let Some(api_key) = api_key else {- return Err(LanguageModelCompletionError::NoApiKey {- provider: PROVIDER_NAME,- });- };- let request = anthropic::stream_completion(- http_client.as_ref(),- &api_url,- &api_key,- request,- beta_headers,- &extra_headers,- );- request.await.map_err(Into::into)- }- .boxed()- }-}--impl LanguageModel for AnthropicModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name.clone())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- true- }-- fn supports_images(&self) -> bool {- self.model.supports_images- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::None => true,- LanguageModelToolChoice::Any => anthropic::supports_forced_tool_use(&self.model.id),- }- }-- fn supports_thinking(&self) -> bool {- self.model.supports_thinking- }-- fn supports_fast_mode(&self) -> bool {- self.model.supports_speed- }-- fn refusal_fallback_model_id(&self) -> Option<&'static str> {- if self.model.id.starts_with(anthropic::FABLE_MODEL_ID_PREFIX) {- Some(anthropic::FABLE_FALLBACK_MODEL_ID)- } else {- None- }- }-- fn supports_server_side_compaction(&self) -> bool {- self.model.supports_compaction- }-- fn supports_explicit_compaction(&self) -> bool {- self.model.supports_compaction- }-- fn supports_explicit_compaction_output_limit(&self) -> bool {- self.supports_explicit_compaction()- }-- fn minimum_explicit_compaction_input_tokens(&self) -> Option<u64> {- self.supports_explicit_compaction()- .then_some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS)- }-- fn compact(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- if !self.supports_explicit_compaction() {- return async {- Err(LanguageModelCompletionError::Other(anyhow::anyhow!(- "this Anthropic model does not support explicit compaction"- )))- }- .boxed();- }-- let mut request = match into_anthropic(- request,- self.model.request_id(false).to_string(),- self.model.default_temperature,- self.model.max_output_tokens,- self.model.mode.clone(),- AnthropicPromptCacheMode::Automatic,- &PROVIDER_ID,- ) {- Ok(request) => request.into_compact_request(),- Err(error) => return async move { Err(error.into()) }.boxed(),- };- if !self.model.supports_speed {- request.speed = None;- }- let request = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();- let future = self.request_limiter.run(async move {- let response = request.await?;- let stream = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID).map_stream(response);- let stream = language_model::stream_in_background(stream.boxed(), executor);- let (context, usage) = collect_compaction_result(stream.boxed(), PROVIDER_NAME).await?;- Ok(CompactionResult { context, usage })- });- future.boxed()- }-- fn supported_effort_levels(&self) -> Vec<language_model::LanguageModelEffortLevel> {- self.model+fn language_model(model: &anthropic::Model) -> LanguageModel {+ let supports_explicit_compaction = model.supports_compaction;+ LanguageModel {+ supports_tools: true,+ supports_images: model.supports_images,+ supports_streaming_tools: true,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: true,+ none: true,+ any: anthropic::supports_forced_tool_use(&model.id),+ },+ supports_thinking: model.supports_thinking,+ supports_fast_mode: model.supports_speed,+ refusal_fallback_model_id: model+ .id+ .starts_with(anthropic::FABLE_MODEL_ID_PREFIX)+ .then_some(anthropic::FABLE_FALLBACK_MODEL_ID),+ supports_server_side_compaction: model.supports_compaction,+ supports_explicit_compaction,+ supports_explicit_compaction_output_limit: supports_explicit_compaction,+ minimum_explicit_compaction_input_tokens: supports_explicit_compaction+ .then_some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS),+ supported_effort_levels: model.supported_effort_levels.iter().map(|e| {@@ -870,111 +947,15 @@ impl LanguageModel for AnthropicModel {is_default,}})- .collect::<Vec<_>>()- }-- fn telemetry_id(&self) -> String {- format!("anthropic/{}", self.model.id)- }-- fn api_key(&self, cx: &App) -> Option<String> {- self.state.read_with(cx, |state, cx| {- let api_url = AnthropicLanguageModelProvider::api_url(cx);- state.api_key_state.key(&api_url).map(|key| key.to_string())- })- }-- fn max_token_count(&self) -> u64 {- self.model.max_input_tokens- }-- fn max_output_tokens(&self) -> Option<u64> {- Some(self.model.max_output_tokens)- }-- fn count_input_tokens(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {- let request_id = self.model.request_id(!request.tools.is_empty()).to_string();- let request = into_anthropic(- request,- request_id,- self.model.default_temperature,- self.model.max_output_tokens,- self.model.mode.clone(),- AnthropicPromptCacheMode::Automatic,- &PROVIDER_ID,- );- let http_client = self.http_client.clone();- let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {- let api_url = AnthropicLanguageModelProvider::api_url(cx);- let extra_headers = AnthropicLanguageModelProvider::settings(cx)- .custom_headers- .clone();- (state.api_key_state.key(&api_url), api_url, extra_headers)- });- let beta_headers = self.model.beta_headers();- self.request_limiter- .run(async move {- let request = request?.into_count_tokens_request();- let api_key = api_key.ok_or(LanguageModelCompletionError::NoApiKey {- provider: PROVIDER_NAME,- })?;- anthropic::count_input_tokens(- http_client.as_ref(),- &api_url,- &api_key,- request,- beta_headers,- &extra_headers,- )- .await- .map(Some)- .map_err(Into::into)- })- .boxed()- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let has_tools = !request.tools.is_empty();- let request_id = self.model.request_id(has_tools).to_string();- let mut request = match into_anthropic(- request,- request_id,- self.model.default_temperature,- self.model.max_output_tokens,- self.model.mode.clone(),- AnthropicPromptCacheMode::Automatic,- &PROVIDER_ID,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- if !self.model.supports_speed {- request.speed = None;- }- let request = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {- let response = request.await?;- let events = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID).map_stream(response);- Ok(language_model::stream_in_background(- events.boxed(),- executor,- ))- });- async move { Ok(future.await?.boxed()) }.boxed()+ .collect(),+ max_output_tokens: Some(model.max_output_tokens),+ ..LanguageModel::new(+ LanguageModelId::from(model.id.to_string()),+ LanguageModelName::from(model.display_name.clone()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("anthropic/{}", model.id),+ model.max_input_tokens,+ )}}diff --git a/crates/language_models/src/provider/anthropic/telemetry.rs b/crates/language_models/src/provider/anthropic/telemetry.rsindex 75fb11a81b4..ffb9223377b 100644--- a/crates/language_models/src/provider/anthropic/telemetry.rs+++ b/crates/language_models/src/provider/anthropic/telemetry.rs@@ -2,7 +2,7 @@ use anthropic::ANTHROPIC_API_URL;use anyhow::{Context as _, anyhow};use gpui::BackgroundExecutor;use http_client::{AsyncBody, HttpClient, Method, Request as HttpRequest};-use language_model::{ANTHROPIC_PROVIDER_ID, LanguageModel};+use language_model::{ANTHROPIC_PROVIDER_ID, LanguageModel, LanguageModelRegistry};use std::env;use std::sync::Arc;use util::ResultExt;@@ -51,11 +51,7 @@ impl AnthropicEventType {}}-pub fn report_anthropic_event(- model: &Arc<dyn LanguageModel>,- event: AnthropicEventData,- cx: &gpui::App,-) {+pub fn report_anthropic_event(model: &LanguageModel, event: AnthropicEventData, cx: &gpui::App) {let reporter = AnthropicEventReporter::new(model, cx);reporter.report(event);}@@ -69,12 +65,15 @@ pub struct AnthropicEventReporter {}impl AnthropicEventReporter {- pub fn new(model: &Arc<dyn LanguageModel>, cx: &gpui::App) -> Self {+ pub fn new(model: &LanguageModel, cx: &gpui::App) -> Self {Self {http_client: cx.http_client(),executor: cx.background_executor().clone(),- api_key: model.api_key(cx),- is_anthropic: model.provider_id() == ANTHROPIC_PROVIDER_ID,+ api_key: LanguageModelRegistry::read_global(cx)+ .provider_for_model(model)+ .ok()+ .and_then(|provider| provider.api_key(model, cx)),+ is_anthropic: model.provider_id == ANTHROPIC_PROVIDER_ID,}}diff --git a/crates/language_models/src/provider/anthropic_compatible.rs b/crates/language_models/src/provider/anthropic_compatible.rsindex 0ad754a913b..8542c08fdcf 100644--- a/crates/language_models/src/provider/anthropic_compatible.rs+++ b/crates/language_models/src/provider/anthropic_compatible.rs@@ -7,10 +7,10 @@ use gpui::{App, AppContext, AsyncApp, Entity, Task};use http_client::{CustomHeaders, HttpClient};use language_model::{AuthenticateError, IconOrSvg, LanguageModel, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelId, LanguageModelName, LanguageModelProvider,+ LanguageModelCompletionStream, LanguageModelId, LanguageModelName, LanguageModelProvider,LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,- LanguageModelRequest, LanguageModelToolChoice, ProviderSettingsView, RateLimiter,- SubPageProviderSettings,+ LanguageModelRequest, LanguageModelToolChoiceSupport, ModelRateLimiters, ProviderSettingsView,+ SubPageProviderSettings, unavailable_error,};use settings::Settings;use std::sync::Arc;@@ -38,6 +38,7 @@ pub struct AnthropicCompatibleLanguageModelProvider {name: LanguageModelProviderName,http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}impl ApiCompatibleProviderSettings for AnthropicCompatibleSettings {@@ -185,126 +186,85 @@ impl AnthropicCompatibleLanguageModelProvider {name: id.into(),http_client,state,+ request_limiters: ModelRateLimiters::default(),}}- fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {- let capabilities = model.capabilities.clone();+ /// Every model this provider offers: the settings entries, in order.+ fn available_models<'a>(&self, cx: &'a App) -> &'a [AvailableModel] {+ &self.state.read(cx).settings.available_models+ }++ fn create_language_model(&self, available: &AvailableModel) -> LanguageModel {+ let capabilities = &available.capabilities;+ let model = available_model_to_anthropic_model(available);+ LanguageModel {+ supports_tools: capabilities.tools,+ supports_images: capabilities.images,+ supports_streaming_tools: capabilities.tools,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: capabilities.tools,+ any: capabilities.tools,+ none: true,+ },+ supports_thinking: model.supports_thinking,+ supported_effort_levels: model+ .supported_effort_levels+ .iter()+ .map(|effort| {+ let is_default = matches!(effort, anthropic::Effort::High);+ let (name, value) = match effort {+ anthropic::Effort::Low => ("Low".into(), "low".into()),+ anthropic::Effort::Medium => ("Medium".into(), "medium".into()),+ anthropic::Effort::High => ("High".into(), "high".into()),+ anthropic::Effort::XHigh => ("XHigh".into(), "xhigh".into()),+ anthropic::Effort::Max => ("Max".into(), "max".into()),+ };+ language_model::LanguageModelEffortLevel {+ name,+ value,+ is_default,+ }+ })+ .collect(),+ max_output_tokens: Some(model.max_output_tokens),+ ..LanguageModel::new(+ LanguageModelId::from(model.id.clone()),+ LanguageModelName::from(model.display_name.clone()),+ self.id.clone(),+ self.name.clone(),+ format!("anthropic/{}", model.id),+ model.max_input_tokens,+ )+ }+ }++ /// The current configuration of `model` and the prompt cache mode to+ /// request it with, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<(anthropic::Model, AnthropicPromptCacheMode), LanguageModelCompletionError> {+ let available = self+ .available_models(cx)+ .iter()+ .find(|available| available.name == model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))?;// Compatible providers may not support Anthropic's automatic prompt// caching; only request explicit (legacy) cache breakpoints when the// user has opted in via the `prompt_caching` capability.- let cache_mode = if capabilities.prompt_caching {+ let cache_mode = if available.capabilities.prompt_caching {AnthropicPromptCacheMode::Legacy} else {AnthropicPromptCacheMode::Disabled};- let model = available_model_to_anthropic_model(&model);-- Arc::new(AnthropicCompatibleLanguageModel {- id: LanguageModelId::from(model.id.clone()),- provider_id: self.id.clone(),- provider_name: self.name.clone(),- model,- capabilities,- cache_mode,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-}--impl LanguageModelProviderState for AnthropicCompatibleLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for AnthropicCompatibleLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- self.id.clone()- }-- fn name(&self) -> LanguageModelProviderName {- self.name.clone()- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiAnthropicCompat)- }-- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {- self.state- .read(cx)- .settings- .available_models- .first()- .map(|model| self.create_language_model(model.clone()))- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- None- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- self.state- .read(cx)- .settings- .available_models- .iter()- .map(|model| self.create_language_model(model.clone()))- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.clone();- Some(ProviderSettingsView::SubPage(SubPageProviderSettings::new(- move |window, cx| {- cx.new(|cx| {- ApiCompatibleProviderConfigurationView::new(- state.clone(),- "Anthropic",- API_KEY_PLACEHOLDER,- window,- cx,- )- })- .into()- },- )))+ Ok((available_model_to_anthropic_model(available), cache_mode))}- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))- }-}--pub struct AnthropicCompatibleLanguageModel {- id: LanguageModelId,- provider_id: LanguageModelProviderId,- provider_name: LanguageModelProviderName,- model: anthropic::Model,- capabilities: ModelCapabilities,- cache_mode: AnthropicPromptCacheMode,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--impl AnthropicCompatibleLanguageModel {- fn stream_completion(+ fn stream_anthropic_request(&self,+ config: &anthropic::Model,request: anthropic::Request,cx: &AsyncApp,) -> BoxFuture<@@ -315,7 +275,7 @@ impl AnthropicCompatibleLanguageModel {>,> {let http_client = self.http_client.clone();- let provider_name = self.provider_name.clone();+ let provider_name = self.name.clone();let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, _cx| {let api_url = state.settings.api_url.clone();@@ -326,7 +286,7 @@ impl AnthropicCompatibleLanguageModel {)});- let beta_headers = self.model.beta_headers();+ let beta_headers = config.beta_headers();async move {let Some(api_key) = api_key else {@@ -352,113 +312,109 @@ impl AnthropicCompatibleLanguageModel {}}-impl LanguageModel for AnthropicCompatibleLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name.clone())- }-- fn provider_id(&self) -> LanguageModelProviderId {- self.provider_id.clone()- }+impl LanguageModelProviderState for AnthropicCompatibleLanguageModelProvider {+ type ObservableEntity = State;- fn provider_name(&self) -> LanguageModelProviderName {- self.provider_name.clone()+ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn supports_tools(&self) -> bool {- self.capabilities.tools+impl LanguageModelProvider for AnthropicCompatibleLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {+ self.id.clone()}- fn supports_images(&self) -> bool {- self.capabilities.images+ fn name(&self) -> LanguageModelProviderName {+ self.name.clone()}- fn supports_streaming_tools(&self) -> bool {- self.capabilities.tools+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiAnthropicCompat)}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => self.capabilities.tools,- LanguageModelToolChoice::None => true,- }+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.available_models(cx)+ .first()+ .map(|model| self.create_language_model(model))}- fn supports_thinking(&self) -> bool {- self.model.supports_thinking+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ None}- fn supported_effort_levels(&self) -> Vec<language_model::LanguageModelEffortLevel> {- self.model- .supported_effort_levels+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.available_models(cx).iter()- .map(|effort| {- let is_default = matches!(effort, anthropic::Effort::High);- let (name, value) = match effort {- anthropic::Effort::Low => ("Low".into(), "low".into()),- anthropic::Effort::Medium => ("Medium".into(), "medium".into()),- anthropic::Effort::High => ("High".into(), "high".into()),- anthropic::Effort::XHigh => ("XHigh".into(), "xhigh".into()),- anthropic::Effort::Max => ("Max".into(), "max".into()),- };- language_model::LanguageModelEffortLevel {- name,- value,- is_default,- }- })+ .map(|model| self.create_language_model(model)).collect()}- fn telemetry_id(&self) -> String {- format!("anthropic/{}", self.model.id)+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn max_token_count(&self) -> u64 {- self.model.max_input_tokens+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))+ }++ fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.clone();+ Some(ProviderSettingsView::SubPage(SubPageProviderSettings::new(+ move |window, cx| {+ cx.new(|cx| {+ ApiCompatibleProviderConfigurationView::new(+ state.clone(),+ "Anthropic",+ API_KEY_PLACEHOLDER,+ window,+ cx,+ )+ })+ .into()+ },+ )))}- fn max_output_tokens(&self) -> Option<u64> {- Some(self.model.max_output_tokens)+ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let (config, cache_mode) = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);let has_tools = !request.tools.is_empty();- let request_id = self.model.request_id(has_tools).to_string();+ let request_id = config.request_id(has_tools).to_string();let mut request = match into_anthropic(request,request_id,- self.model.default_temperature,- self.model.max_output_tokens,- self.model.mode.clone(),- self.cache_mode,- &self.provider_id,+ config.default_temperature,+ config.max_output_tokens,+ config.mode.clone(),+ cache_mode,+ &self.id,) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- if !self.model.supports_speed {+ if !config.supports_speed {request.speed = None;}- let completion_request = self.stream_completion(request, cx);- let provider_name = self.provider_name.clone();- let provider_id = self.provider_id.clone();+ let completion_request = self.stream_anthropic_request(&config, request, cx);+ let provider_name = self.name.clone();+ let provider_id = self.id.clone();let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let response = completion_request.await?;let events = AnthropicEventMapper::new(provider_name, provider_id).map_stream(response);Ok(language_model::stream_in_background(diff --git a/crates/language_models/src/provider/bedrock.rs b/crates/language_models/src/provider/bedrock.rsindex 092dcd1ae10..239d7fd0911 100644--- a/crates/language_models/src/provider/bedrock.rs+++ b/crates/language_models/src/provider/bedrock.rs@@ -2,10 +2,9 @@ use std::pin::Pin;use std::sync::Arc;use anyhow::{Context as _, Result, anyhow};-use async_lock::OnceCell;use aws_config::stalled_stream_protection::StalledStreamProtectionConfig;use aws_config::{BehaviorVersion, Region};-use aws_credential_types::provider::{ProvideCredentials, SharedCredentialsProvider};+use aws_credential_types::provider::ProvideCredentials;use aws_credential_types::{Credentials, Token};use aws_http_client::AwsHttpClient;use aws_sigv4::http_request::{SignableBody, SignableRequest, SigningSettings, sign};@@ -25,7 +24,7 @@ use bedrock::{BedrockToolResultContentBlock, BedrockToolResultStatus, BedrockToolSpec, BedrockToolUseBlock,ConverseModel, MantleModel, MantleProtocol, value_to_aws_document,};-use collections::{BTreeMap, HashMap};+use collections::{BTreeMap, HashMap, IndexMap};use credentials_provider::CredentialsProvider;use futures::{AsyncBufReadExt, AsyncReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, io::BufReader,@@ -41,12 +40,13 @@ use http_client::{};use language_model::{AuthenticateError, EnvVar, IconOrSvg, InlineDescription, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,- LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice, LanguageModelToolResultContent, LanguageModelToolUse, MessageContent,+ LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelCompletionStream,+ LanguageModelEffortLevel, LanguageModelId, LanguageModelName, LanguageModelProvider,+ LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,+ LanguageModelRequest, LanguageModelToolChoice, LanguageModelToolChoiceSupport,+ LanguageModelToolResultContent, LanguageModelToolUse, MessageContent, ModelRateLimiters,ProviderErrorCategory, ProviderSettingsView, RateLimiter, Role, SubPageProviderSettings,- TokenUsage, env_var,+ TokenUsage, env_var, unavailable_error,};use open_ai::responses::Request as OpenAiResponseRequest;use open_ai::responses::{ResponseOutputItem, StreamEvent as OpenAiResponseStreamEvent};@@ -58,7 +58,7 @@ use settings::{Settings, SettingsStore,};use std::sync::LazyLock;-use std::time::SystemTime;+use std::time::{Duration, SystemTime};use strum::{EnumIter, IntoEnumIterator, IntoStaticStr};use ui::{ButtonLink, ConfiguredApiCard, Divider, List, ListBulletItem, prelude::*};use ui_input::InputField;@@ -254,20 +254,22 @@ fn mantle_endpoint_url(region: &str) -> String {format!("https://bedrock-mantle.{region}.api.aws/openai/v1")}-enum MantleAuth {+/// Auth resolved for one request, shared by the Converse and Mantle APIs.+#[derive(Clone)]+enum BedrockRequestAuth {ApiKey { api_key: String },SigV4 { credentials: Credentials },}-impl MantleAuth {+impl BedrockRequestAuth {fn apply(&self, request: &mut HttpRequest<AsyncBody>, body: &[u8], region: &str) -> Result<()> {match self {- MantleAuth::ApiKey { api_key } => {+ BedrockRequestAuth::ApiKey { api_key } => {let value = HeaderValue::from_str(&format!("Bearer {}", api_key.trim())).context("building Mantle bearer token authorization header")?;request.headers_mut().insert(AUTHORIZATION, value);}- MantleAuth::SigV4 { credentials } => {+ BedrockRequestAuth::SigV4 { credentials } => {sign_mantle_request_sigv4(request, body, credentials, region)?;}}@@ -348,10 +350,54 @@ pub struct State {/// Whether credentials came from environment variables (only relevant for static credentials)credentials_from_env: bool,credentials_provider: Arc<dyn CredentialsProvider>,+ /// AWS credentials that expire, such as those from SSO, STS or a credential+ /// process, reused until shortly before they expire. Credentials without an+ /// expiry are resolved for every request instead, because the AWS SDK never+ /// rereads credential files and external tools rewrite them.+ expiring_credentials: Option<ExpiringCredentials>,+ /// Incremented on every auth change, so credentials resolved for earlier+ /// auth are not cached once they arrive.+ auth_generation: u64,_subscription: Subscription,}+struct ExpiringCredentials {+ auth: Option<BedrockAuth>,+ region: String,+ credentials: Credentials,+}++/// How long before expiry cached AWS credentials are replaced, so a request+/// that streams for a while does not outlive them.+const CREDENTIALS_EXPIRY_MARGIN: Duration = Duration::from_secs(5 * 60);++/// How long resolving AWS credentials may take before the request fails,+/// matching the AWS SDK's default identity cache load timeout. Without it a+/// hung credential process or metadata endpoint would stall the request.+const CREDENTIALS_LOAD_TIMEOUT: Duration = Duration::from_secs(5);++/// The Bedrock bearer token the AWS SDK would find in the environment, looked+/// up with `environment_variable`.+///+/// Unless a token is configured explicitly, the SDK prefers this token over any+/// AWS credentials when it builds a client, so it is honored before resolving+/// credentials. Like the SDK, this checks the Bedrock-specific variable, then+/// the generic one, and uses a set value even when it is empty.+fn environment_bearer_token(+ environment_variable: impl Fn(&str) -> Option<String>,+) -> Option<String> {+ environment_variable("AWS_BEARER_TOKEN_BEDROCK")+ .or_else(|| environment_variable("AWS_BEARER_TOKEN"))+}+impl State {+ /// Sets the auth method and drops credentials resolved for the old one.+ fn set_auth(&mut self, auth: Option<BedrockAuth>) {+ self.auth = auth;+ self.expiring_credentials = None;+ self.auth_generation += 1;+ }+fn reset_auth(&self, cx: &mut Context<Self>) -> Task<Result<()>> {let credentials_provider = self.credentials_provider.clone();cx.spawn(async move |this, cx| {@@ -360,7 +406,7 @@ impl State {.await.log_err();this.update(cx, |this, cx| {- this.auth = None;+ this.set_auth(None);this.credentials_from_env = false;cx.notify();})@@ -384,7 +430,7 @@ impl State {).await?;this.update(cx, |this, cx| {- this.auth = auth;+ this.set_auth(auth);this.credentials_from_env = false;cx.notify();})@@ -422,7 +468,7 @@ impl State {return cx.spawn(async move |this, cx| {this.update(cx, |this, cx| {- this.auth = Some(auth);+ this.set_auth(Some(auth));this.credentials_from_env = false;cx.notify();})?;@@ -483,7 +529,7 @@ impl State {// If we got auth from env vars, use itif let Some(auth) = auth {this.update(cx, |this, cx| {- this.auth = Some(auth);+ this.set_auth(Some(auth));this.credentials_from_env = from_env;cx.notify();})?;@@ -507,7 +553,7 @@ impl State {.ok_or(AuthenticateError::CredentialsNotFound)?;this.update(cx, |this, cx| {- this.auth = Some(auth);+ this.set_auth(Some(auth));this.credentials_from_env = false;cx.notify();})?;@@ -548,8 +594,14 @@ impl State {pub struct BedrockLanguageModelProvider {http_client: AwsHttpClient,plain_http_client: Arc<dyn HttpClient>,- handle: tokio::runtime::Handle,state: Entity<State>,+ request_limiters: ModelRateLimiters,+}++/// A model this provider offers, served by either the Converse API or Mantle.+enum BedrockModelConfig {+ Converse(ConverseModel),+ Mantle(MantleModel),}impl BedrockLanguageModelProvider {@@ -563,6 +615,8 @@ impl BedrockLanguageModelProvider {settings: Some(AllLanguageModelSettings::get_global(cx).bedrock.clone()),credentials_from_env: false,credentials_provider,+ expiring_credentials: None,+ auth_generation: 0,_subscription: cx.observe_global::<SettingsStore>(|_, cx| {cx.notify();}),@@ -571,58 +625,16 @@ impl BedrockLanguageModelProvider {Self {http_client: AwsHttpClient::new(http_client.clone()),plain_http_client: http_client,- handle: Tokio::handle(cx),state,+ request_limiters: ModelRateLimiters::default(),}}- fn create_language_model(&self, model: bedrock::ConverseModel) -> Arc<dyn LanguageModel> {- Arc::new(BedrockModel {- id: LanguageModelId::from(model.id().to_string()),- model,- http_client: self.http_client.clone(),- handle: self.handle.clone(),- state: self.state.clone(),- client: OnceCell::new(),- request_limiter: RateLimiter::new(4),- })- }-- fn create_mantle_language_model(&self, model: bedrock::MantleModel) -> Arc<dyn LanguageModel> {- Arc::new(BedrockMantleModel {- id: LanguageModelId::from(model.id().to_string()),- model,- http_client: self.plain_http_client.clone(),- state: self.state.clone(),- credentials_provider: Arc::new(OnceCell::new()),- request_limiter: RateLimiter::new(4),- })- }-}--impl LanguageModelProvider for BedrockLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiBedrock)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(bedrock::ConverseModel::default()))- }-- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {- let region = self.state.read(cx).get_region();- Some(self.create_language_model(bedrock::ConverseModel::default_fast(region.as_str())))- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: the built-in Converse+ /// models with settings entries added or overriding them, then the same+ /// for Mantle. A Mantle model replaces a Converse model with the same id,+ /// so each id is offered and served exactly once.+ fn bedrock_models(&self, cx: &App) -> IndexMap<String, BedrockModelConfig> {let bedrock_settings = &AllLanguageModelSettings::get_global(cx).bedrock;let mut models = BTreeMap::default();@@ -632,7 +644,6 @@ impl LanguageModelProvider for BedrockLanguageModelProvider {}}- // Override with available models from settingsfor model in bedrock_settings.available_models.iter() {models.insert(model.name.clone(),@@ -652,11 +663,6 @@ impl LanguageModelProvider for BedrockLanguageModelProvider {);}- let mut models: Vec<Arc<dyn LanguageModel>> = models- .into_values()- .map(|model| self.create_language_model(model))- .collect();-let mut mantle_models = BTreeMap::default();for model in bedrock::MantleModel::iter() {@@ -665,7 +671,6 @@ impl LanguageModelProvider for BedrockLanguageModelProvider {}}- // Override with available Mantle models from settingsfor model in bedrock_settings.mantle_available_models.iter() {mantle_models.insert(model.name.clone(),@@ -682,13 +687,92 @@ impl LanguageModelProvider for BedrockLanguageModelProvider {);}- models.extend(- mantle_models- .into_values()- .map(|model| self.create_mantle_language_model(model)),- );+ let mut all_models = IndexMap::default();+ for (id, model) in models {+ all_models.insert(id, BedrockModelConfig::Converse(model));+ }+ for (id, model) in mantle_models {+ all_models.insert(id, BedrockModelConfig::Mantle(model));+ }+ all_models+ }++ /// Describes the offered model with `id`.+ fn offered_model(&self, id: &str, cx: &App) -> Option<LanguageModel> {+ self.bedrock_models(cx).get(id).map(language_model)+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<BedrockModelConfig, LanguageModelCompletionError> {+ self.bedrock_models(cx)+ .swap_remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))+ }+}++fn language_model(config: &BedrockModelConfig) -> LanguageModel {+ match config {+ BedrockModelConfig::Converse(model) => converse_language_model(model),+ BedrockModelConfig::Mantle(model) => mantle_language_model(model),+ }+}++impl LanguageModelProvider for BedrockLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {+ PROVIDER_ID+ }++ fn name(&self) -> LanguageModelProviderName {+ PROVIDER_NAME+ }++ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiBedrock)+ }++ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.offered_model(bedrock::ConverseModel::default().id(), cx)+ }- models+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ let region = self.state.read(cx).get_region();+ self.offered_model(+ bedrock::ConverseModel::default_fast(region.as_str()).id(),+ cx,+ )+ }++ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.bedrock_models(cx)+ .values()+ .map(language_model)+ .collect()+ }++ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ match config {+ BedrockModelConfig::Converse(config) => {+ self.stream_converse_completion(&config, &request_limiter, request, cx)+ }+ BedrockModelConfig::Mantle(config) => {+ self.stream_mantle_completion(&config, &request_limiter, request, cx)+ }+ }}fn is_authenticated(&self, cx: &App) -> bool {@@ -721,79 +805,85 @@ impl LanguageModelProviderState for BedrockLanguageModelProvider {}}-struct BedrockModel {- id: LanguageModelId,- model: ConverseModel,- http_client: AwsHttpClient,- handle: tokio::runtime::Handle,- client: OnceCell<BedrockClient>,- state: Entity<State>,- request_limiter: RateLimiter,-}--impl BedrockModel {- fn get_or_init_client(&self, cx: &AsyncApp) -> anyhow::Result<&BedrockClient> {- self.client- .get_or_try_init_blocking(|| {- let (auth, endpoint, region) = cx.read_entity(&self.state, |state, _cx| {- let endpoint = state.settings.as_ref().and_then(|s| s.endpoint.clone());- let region = state.get_region();- (state.auth.clone(), endpoint, region)- });-- let mut config_builder = aws_config::defaults(BehaviorVersion::latest())- .stalled_stream_protection(StalledStreamProtectionConfig::disabled())- .http_client(self.http_client.clone())- .region(Region::new(region))- .timeout_config(TimeoutConfig::disabled());+impl BedrockLanguageModelProvider {+ /// Resolves auth for one request, reusing unexpired cached credentials.+ fn resolve_request_auth(&self, cx: &AsyncApp) -> Task<Result<BedrockRequestAuth>> {+ let http_client = self.http_client.clone();+ self.resolve_request_auth_with(+ cx,+ |name| std::env::var(name).ok(),+ move |auth, region| resolve_request_auth(http_client, auth, region),+ )+ }- if let Some(endpoint_url) = endpoint- && !endpoint_url.is_empty()- {- config_builder = config_builder.endpoint_url(endpoint_url);- }+ /// [`Self::resolve_request_auth`], reading environment variables with+ /// `environment_variable` and resolving uncached auth with `resolve`.+ fn resolve_request_auth_with<ResolveFuture>(+ &self,+ cx: &AsyncApp,+ environment_variable: impl Fn(&str) -> Option<String>,+ resolve: impl FnOnce(Option<BedrockAuth>, String) -> ResolveFuture,+ ) -> Task<Result<BedrockRequestAuth>>+ where+ ResolveFuture: Future<Output = Result<BedrockRequestAuth>> + Send + 'static,+ {+ let (auth, region, auth_generation, cached_credentials) =+ cx.read_entity(&self.state, |state, _cx| {+ let (auth, region) = (state.auth.clone(), state.get_region());+ let fresh_until = SystemTime::now() + CREDENTIALS_EXPIRY_MARGIN;+ let cached_credentials = state+ .expiring_credentials+ .as_ref()+ .filter(|cached| {+ cached.auth == auth+ && cached.region == region+ && cached+ .credentials+ .expiry()+ .is_some_and(|expiry| expiry > fresh_until)+ })+ .map(|cached| cached.credentials.clone());+ (auth, region, state.auth_generation, cached_credentials)+ });+ if !matches!(auth, Some(BedrockAuth::ApiKey { .. }))+ && let Some(api_key) = environment_bearer_token(environment_variable)+ {+ return Task::ready(Ok(BedrockRequestAuth::ApiKey { api_key }));+ }+ if let Some(credentials) = cached_credentials {+ return Task::ready(Ok(BedrockRequestAuth::SigV4 { credentials }));+ }- match auth {- Some(BedrockAuth::Automatic) | None => {- // Use default AWS credential provider chain- }- Some(BedrockAuth::NamedProfile { profile_name })- | Some(BedrockAuth::SingleSignOn { profile_name }) => {- if !profile_name.is_empty() {- config_builder = config_builder.profile_name(profile_name);- }- }- Some(BedrockAuth::IamCredentials {- access_key_id,- secret_access_key,- session_token,- }) => {- let aws_creds = Credentials::new(- access_key_id,- secret_access_key,- session_token,- None,- "zed-bedrock-provider",- );- config_builder = config_builder.credentials_provider(aws_creds);- }- Some(BedrockAuth::ApiKey { api_key }) => {- config_builder = config_builder- .auth_scheme_preference(["httpBearerAuth".into()]) // https://github.com/smithy-lang/smithy-rs/pull/4241- .token_provider(Token::new(api_key, None));+ let resolve_task = Tokio::spawn_result(cx, resolve(auth.clone(), region.clone()));+ let timeout = cx.background_executor().timer(CREDENTIALS_LOAD_TIMEOUT);+ let state = self.state.clone();+ cx.spawn(async move |cx| {+ // Dropping `resolve_task` on timeout cancels the Tokio task.+ let request_auth = futures::select_biased! {+ request_auth = resolve_task.fuse() => request_auth?,+ _ = timeout.fuse() => anyhow::bail!(+ "timed out after {CREDENTIALS_LOAD_TIMEOUT:?} resolving AWS credentials"+ ),+ };+ if let BedrockRequestAuth::SigV4 { credentials } = &request_auth+ && credentials.expiry().is_some()+ {+ let credentials = credentials.clone();+ state.update(cx, |state, _| {+ if state.auth_generation == auth_generation {+ state.expiring_credentials = Some(ExpiringCredentials {+ auth,+ region,+ credentials,+ });}- }-- let config = self.handle.block_on(config_builder.load());-- anyhow::Ok(BedrockClient::new(&config))- })- .context("initializing Bedrock client")?;-- self.client.get().context("Bedrock client not initialized")+ });+ }+ Ok(request_auth)+ })}- fn stream_completion(+ fn stream_bedrock_request(&self,request: bedrock::Request,cx: &AsyncApp,@@ -801,72 +891,78 @@ impl BedrockModel {'static,Result<BoxStream<'static, Result<BedrockStreamingResponse, anyhow::Error>>, BedrockError>,> {- let Ok(runtime_client) = self- .get_or_init_client(cx)- .cloned()- .context("Bedrock client not initialized")- else {- return futures::future::ready(Err(BedrockError::Other(anyhow!("App state dropped"))))- .boxed();- };- let extra_headers = self.state.read_with(cx, |_, cx| {- AllLanguageModelSettings::get_global(cx)- .bedrock- .custom_headers- .clone()- });+ let http_client = self.http_client.clone();+ let auth_task = self.resolve_request_auth(cx);+ let (profile_name, endpoint, region, extra_headers) =+ self.state.read_with(cx, |state, cx| {+ (+ aws_profile_name(&state.auth),+ state.settings.as_ref().and_then(|s| s.endpoint.clone()),+ state.get_region(),+ AllLanguageModelSettings::get_global(cx)+ .bedrock+ .custom_headers+ .clone(),+ )+ });- let task = Tokio::spawn(- cx,- bedrock::stream_completion(runtime_client, request, extra_headers),- );+ let task = Tokio::spawn(cx, async move {+ let request_auth = auth_task.await.map_err(BedrockError::Other)?;+ let client =+ build_converse_client(http_client, request_auth, profile_name, endpoint, region)+ .await;+ bedrock::stream_completion(client, request, extra_headers).await+ });async move { task.await.map_err(|e| BedrockError::Other(e.into()))? }.boxed()}}-impl LanguageModel for BedrockModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME+/// Builds a Converse client for one request from already-resolved auth. The+/// profile still applies, since it can configure more than credentials.+async fn build_converse_client(+ http_client: AwsHttpClient,+ request_auth: BedrockRequestAuth,+ profile_name: Option<String>,+ endpoint: Option<String>,+ region: String,+) -> BedrockClient {+ let mut config_builder = aws_config::defaults(BehaviorVersion::latest())+ .stalled_stream_protection(StalledStreamProtectionConfig::disabled())+ .http_client(http_client)+ .region(Region::new(region))+ .timeout_config(TimeoutConfig::disabled());++ if let Some(profile_name) = profile_name {+ config_builder = config_builder.profile_name(profile_name);}-- fn supports_tools(&self) -> bool {- self.model.supports_tool_use()+ if let Some(endpoint_url) = endpoint+ && !endpoint_url.is_empty()+ {+ config_builder = config_builder.endpoint_url(endpoint_url);}- fn supports_images(&self) -> bool {- self.model.supports_images()- }+ config_builder = match request_auth {+ BedrockRequestAuth::SigV4 { credentials } => {+ config_builder.credentials_provider(credentials)+ }+ BedrockRequestAuth::ApiKey { api_key } => config_builder+ .auth_scheme_preference(["httpBearerAuth".into()]) // https://github.com/smithy-lang/smithy-rs/pull/4241+ .token_provider(Token::new(api_key, None)),+ };- fn supports_thinking(&self) -> bool {- self.model.supports_thinking()- }+ BedrockClient::new(&config_builder.load().await)+}- fn refusal_fallback_model_id(&self) -> Option<&'static str> {- if self- .model+fn converse_language_model(model: &ConverseModel) -> LanguageModel {+ LanguageModel {+ supports_tools: model.supports_tool_use(),+ supports_images: model.supports_images(),+ supports_thinking: model.supports_thinking(),+ refusal_fallback_model_id: model.id().starts_with(anthropic::FABLE_MODEL_ID_PREFIX)- {- Some(anthropic::FABLE_FALLBACK_MODEL_ID)- } else {- None- }- }-- fn supported_effort_levels(&self) -> Vec<language_model::LanguageModelEffortLevel> {- if self.model.supports_adaptive_thinking() {+ .then_some(anthropic::FABLE_FALLBACK_MODEL_ID),+ supported_effort_levels: if model.supports_adaptive_thinking() {vec![language_model::LanguageModelEffortLevel {name: "Low".into(),@@ -896,51 +992,40 @@ impl LanguageModel for BedrockModel {].into_iter().filter(|effort_level| {- effort_level.value != "xhigh" || self.model.supports_xhigh_adaptive_thinking()+ effort_level.value != "xhigh" || model.supports_xhigh_adaptive_thinking()}).collect()} else {- Vec::new()- }- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => {- self.model.supports_tool_use()- }- // Add support for None - we'll filter tool calls at response- LanguageModelToolChoice::None => self.model.supports_tool_use(),- }- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn telemetry_id(&self) -> String {- format!("bedrock/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- Some(self.model.max_output_tokens())+ Arc::default()+ },+ // Add support for None - we'll filter tool calls at response+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: model.supports_tool_use(),+ any: model.supports_tool_use(),+ none: model.supports_tool_use(),+ },+ supports_streaming_tools: true,+ max_output_tokens: Some(model.max_output_tokens()),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("bedrock/{}", model.id()),+ model.max_token_count(),+ )}+}- fn stream_completion(+impl BedrockLanguageModelProvider {+ fn stream_converse_completion(&self,+ config: &ConverseModel,+ request_limiter: &RateLimiter,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {if request.contains_custom_tool_input() {return async move {Err(anyhow::anyhow!("Bedrock does not support custom tools").into())@@ -954,7 +1039,7 @@ impl LanguageModel for BedrockModel {(state.get_region(), state.get_allow_global(), gid, gv)});- let model_id = match self.model.cross_region_inference_id(®ion, allow_global) {+ let model_id = match config.cross_region_inference_id(®ion, allow_global) {Ok(s) => s,Err(e) => {return async move { Err(e.into()) }.boxed();@@ -966,11 +1051,11 @@ impl LanguageModel for BedrockModel {let request = match into_bedrock(request,model_id,- self.model.default_temperature(),- self.model.max_output_tokens(),- self.model.thinking_mode(),- self.model.supports_caching(),- self.model.supports_tool_use(),+ config.default_temperature(),+ config.max_output_tokens(),+ config.thinking_mode(),+ config.supports_caching(),+ config.supports_tool_use(),guardrail_identifier,guardrail_version,) {@@ -978,10 +1063,10 @@ impl LanguageModel for BedrockModel {Err(err) => return futures::future::ready(Err(err.into())).boxed(),};- let request = self.stream_completion(request, cx);- let display_name = self.model.display_name().to_string();+ let request = self.stream_bedrock_request(request, cx);+ let display_name = config.display_name().to_string();let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let response = request.await.map_err(|err| match err {BedrockError::Validation(ref msg) => {if msg.contains("model identifier is invalid") {@@ -1135,82 +1220,61 @@ fn map_mantle_error(model: &MantleModel, error: RequestError) -> LanguageModelCoerror.into()}-/// Resolves an AWS credentials provider for profile/SSO/automatic auth.-/// Cached in `cell` since building it may read config files from disk;-/// credentials themselves are still re-resolved on every call. Async so this-/// never blocks the foreground thread (unlike `BedrockModel::get_or_init_client`).-async fn resolve_mantle_credentials_provider(- cell: &OnceCell<SharedCredentialsProvider>,- profile_name: Option<String>,- region: String,-) -> Result<SharedCredentialsProvider> {- let provider = cell- .get_or_try_init(move || async move {- let mut config_builder =- aws_config::defaults(BehaviorVersion::latest()).region(Region::new(region));-- if let Some(profile_name) = profile_name.filter(|name| !name.is_empty()) {- config_builder = config_builder.profile_name(profile_name);- }-- let config = config_builder.load().await;- config- .credentials_provider()- .context("no AWS credentials provider is configured")- })- .await- .context("resolving AWS credentials for Bedrock Mantle")?;- Ok(provider.clone())+/// The AWS profile named by profile or SSO auth, if any.+fn aws_profile_name(auth: &Option<BedrockAuth>) -> Option<String> {+ match auth {+ Some(BedrockAuth::NamedProfile { profile_name })+ | Some(BedrockAuth::SingleSignOn { profile_name }) => {+ Some(profile_name.clone()).filter(|name| !name.is_empty())+ }+ _ => None,+ }}-/// Resolves provider settings into concrete Mantle request auth. A configured-/// Bedrock API key is sent as bearer auth; every AWS-credential-based method-/// signs the Mantle HTTP request directly with SigV4.-async fn resolve_mantle_auth(- credentials_provider: Arc<OnceCell<SharedCredentialsProvider>>,+/// Resolves provider settings into concrete request auth. A configured Bedrock+/// API key is sent as bearer auth; every other method resolves AWS+/// credentials to sign the request with SigV4.+async fn resolve_request_auth(+ http_client: AwsHttpClient,auth: Option<BedrockAuth>,region: String,-) -> Result<MantleAuth> {- match auth {- Some(BedrockAuth::ApiKey { api_key }) => Ok(MantleAuth::ApiKey { api_key }),+) -> Result<BedrockRequestAuth> {+ let profile_name = match auth {+ Some(BedrockAuth::ApiKey { api_key }) => return Ok(BedrockRequestAuth::ApiKey { api_key }),Some(BedrockAuth::IamCredentials {access_key_id,secret_access_key,session_token,- }) => Ok(MantleAuth::SigV4 {- credentials: Credentials::new(- access_key_id,- secret_access_key,- session_token,- None,- "zed-bedrock-provider",- ),- }),- Some(BedrockAuth::NamedProfile { profile_name })- | Some(BedrockAuth::SingleSignOn { profile_name }) => {- let provider = resolve_mantle_credentials_provider(- &credentials_provider,- Some(profile_name),- region.clone(),- )- .await?;- let credentials = provider- .provide_credentials()- .await- .context("failed to resolve AWS credentials")?;- Ok(MantleAuth::SigV4 { credentials })- }- Some(BedrockAuth::Automatic) | None => {- let provider =- resolve_mantle_credentials_provider(&credentials_provider, None, region.clone())- .await?;- let credentials = provider- .provide_credentials()- .await- .context("failed to resolve AWS credentials")?;- Ok(MantleAuth::SigV4 { credentials })+ }) => {+ return Ok(BedrockRequestAuth::SigV4 {+ credentials: Credentials::new(+ access_key_id,+ secret_access_key,+ session_token,+ None,+ "zed-bedrock-provider",+ ),+ });}+ _ => aws_profile_name(&auth),+ };++ let mut config_builder = aws_config::defaults(BehaviorVersion::latest())+ .http_client(http_client)+ .region(Region::new(region));+ if let Some(profile_name) = profile_name {+ config_builder = config_builder.profile_name(profile_name);}+ let credentials = config_builder+ .load()+ .await+ .credentials_provider()+ .context("no AWS credentials provider is configured")+ .context("resolving AWS credentials for Bedrock")?+ .provide_credentials()+ .await+ .context("failed to resolve AWS credentials")?;+ Ok(BedrockRequestAuth::SigV4 { credentials })}fn parse_mantle_chat_stream_line(line: &str) -> Result<ResponseStreamEvent> {@@ -1244,7 +1308,7 @@ async fn stream_mantle_sse<Request, Event>(provider_name: &str,url: &str,region: &str,- auth: &MantleAuth,+ auth: &BedrockRequestAuth,request: Request,extra_headers: &CustomHeaders,parse_stream_line: fn(&str) -> Result<Event>,@@ -1363,8 +1427,7 @@ impl MantleResponseEventMapper {fn map_stream(mut self,events: BoxStream<'static, Result<OpenAiResponseStreamEvent>>,- ) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>- {+ ) -> LanguageModelCompletionStream {events.flat_map(move |event| {futures::stream::iter(match event {@@ -1738,18 +1801,11 @@ impl MantleResponseEventMapper {}}-struct BedrockMantleModel {- id: LanguageModelId,- model: MantleModel,- http_client: Arc<dyn HttpClient>,- state: Entity<State>,- credentials_provider: Arc<OnceCell<SharedCredentialsProvider>>,- request_limiter: RateLimiter,-}--impl BedrockMantleModel {+impl BedrockLanguageModelProvider {fn stream_mantle_request<Request, Event>(&self,+ config: &MantleModel,+ request_limiter: &RateLimiter,request: Request,cx: &AsyncApp,endpoint: &'static str,@@ -1759,12 +1815,9 @@ impl BedrockMantleModel {Request: Serialize + Send + 'static,Event: Send + 'static,{- let http_client = self.http_client.clone();- let model = self.model.clone();- let credentials_provider = self.credentials_provider.clone();- let (auth, region) = cx.read_entity(&self.state, |state, _cx| {- (state.auth.clone(), state.get_region())- });+ let http_client = self.plain_http_client.clone();+ let model = config.clone();+ let region = cx.read_entity(&self.state, |state, _cx| state.get_region());let url = format!("{}/{}", mantle_endpoint_url(®ion), endpoint);let extra_headers = cx.read_entity(&self.state, |_, cx| {AllLanguageModelSettings::get_global(cx)@@ -1773,12 +1826,9 @@ impl BedrockMantleModel {.clone()});let provider_name = PROVIDER_NAME.0.to_string();- let auth_task = Tokio::spawn_result(- cx,- resolve_mantle_auth(credentials_provider, auth, region.clone()),- );+ let auth_task = self.resolve_request_auth(cx);- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let auth = auth_task.await.map_err(LanguageModelCompletionError::Other)?;@@ -1799,8 +1849,10 @@ impl BedrockMantleModel {async move { Ok(future.await?.boxed()) }.boxed()}- fn stream_completion(+ fn stream_chat_completion(&self,+ config: &MantleModel,+ request_limiter: &RateLimiter,request: open_ai::Request,cx: &AsyncApp,) -> BoxFuture<@@ -1808,6 +1860,8 @@ impl BedrockMantleModel {Result<BoxStream<'static, Result<ResponseStreamEvent>>, LanguageModelCompletionError>,> {self.stream_mantle_request(+ config,+ request_limiter,request,cx,"chat/completions",@@ -1817,6 +1871,8 @@ impl BedrockMantleModel {fn stream_response(&self,+ config: &MantleModel,+ request_limiter: &RateLimiter,request: OpenAiResponseRequest,cx: &AsyncApp,) -> BoxFuture<@@ -1828,87 +1884,55 @@ impl BedrockMantleModel {> {let mut request = request;strip_unsupported_mantle_response_fields(&mut request);- self.stream_mantle_request(request, cx, "responses", parse_mantle_response_stream_line)+ self.stream_mantle_request(+ config,+ request_limiter,+ request,+ cx,+ "responses",+ parse_mantle_response_stream_line,+ )}}-impl LanguageModel for BedrockMantleModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.supports_tools()- }-- fn supports_images(&self) -> bool {- self.model.supports_images()- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => {- self.model.supports_tools()- }- LanguageModelToolChoice::None => true,- }- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_thinking(&self) -> bool {- self.model.supports_thinking()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- mantle_supported_effort_levels(&self.model)- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn telemetry_id(&self) -> String {- format!("bedrock-mantle/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- Some(self.model.max_output_tokens())+fn mantle_language_model(model: &MantleModel) -> LanguageModel {+ LanguageModel {+ supports_tools: model.supports_tools(),+ supports_images: model.supports_images(),+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: model.supports_tools(),+ any: model.supports_tools(),+ none: true,+ },+ supports_streaming_tools: true,+ supports_thinking: model.supports_thinking(),+ supported_effort_levels: mantle_supported_effort_levels(model).into(),+ supports_split_token_display: true,+ max_output_tokens: Some(model.max_output_tokens()),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("bedrock-mantle/{}", model.id()),+ model.max_token_count(),+ )}+}- fn stream_completion(+impl BedrockLanguageModelProvider {+ fn stream_mantle_completion(&self,+ config: &MantleModel,+ request_limiter: &RateLimiter,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {let region = cx.read_entity(&self.state, |state, _cx| state.get_region());if !MANTLE_SUPPORTED_REGIONS.contains(®ion.as_str()) {- let display_name = self.model.display_name().to_string();+ let display_name = config.display_name().to_string();let supported = MANTLE_SUPPORTED_REGIONS.join(", ");return futures::future::ready(Err(LanguageModelCompletionError::Other(anyhow!("{display_name} is not available in {region} because Bedrock Mantle isn't offered \@@ -1917,25 +1941,25 @@ impl LanguageModel for BedrockMantleModel {.boxed();}- let model_id = self.model.request_id().to_string();- let max_output_tokens = Some(self.model.max_output_tokens());+ let model_id = config.request_id().to_string();+ let max_output_tokens = Some(config.max_output_tokens());- match self.model.protocol() {+ match config.protocol() {MantleProtocol::Responses => {let request = match into_open_ai_response(request,&model_id,- self.model.supports_tools(),+ config.supports_tools(),false,max_output_tokens,- mantle_default_reasoning_effort(&self.model),- self.model.supports_thinking(),+ mantle_default_reasoning_effort(config),+ config.supports_thinking(),&PROVIDER_ID,) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let completions = self.stream_response(request, cx);+ let completions = self.stream_response(config, request_limiter, request, cx);let executor = cx.background_executor().clone();async move {let mapper = MantleResponseEventMapper::new();@@ -1947,11 +1971,11 @@ impl LanguageModel for BedrockMantleModel {.boxed()}MantleProtocol::ChatCompletions => {- let reasoning_effort = mantle_selected_reasoning_effort(&request, &self.model);+ let reasoning_effort = mantle_selected_reasoning_effort(&request, config);let request = match into_open_ai(request,&model_id,- self.model.supports_tools(),+ config.supports_tools(),false,max_output_tokens,ChatCompletionMaxTokensParameter::MaxCompletionTokens,@@ -1961,7 +1985,7 @@ impl LanguageModel for BedrockMantleModel {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let completions = self.stream_completion(request, cx);+ let completions = self.stream_chat_completion(config, request_limiter, request, cx);let executor = cx.background_executor().clone();async move {let mapper = ChatCompletionEventMapper::new();@@ -2946,6 +2970,246 @@ mod tests {.unwrap()}+ #[gpui::test]+ fn mantle_model_replaces_converse_model_with_the_same_id(cx: &mut gpui::TestAppContext) {+ cx.update(|cx| {+ let settings_store = SettingsStore::test(cx);+ cx.set_global(settings_store);+ gpui_tokio::init(cx);+ let content = serde_json::json!({+ "language_models": {+ "bedrock": {+ "available_models": [{+ "name": "shared-model",+ "display_name": "Converse Shared",+ "max_tokens": 1000,+ }],+ "mantle_available_models": [{+ "name": "shared-model",+ "display_name": "Mantle Shared",+ "max_tokens": 2000,+ "protocol": "chat_completions",+ }],+ }+ }+ })+ .to_string();+ cx.update_global::<SettingsStore, _>(|store, cx| {+ store+ .set_user_settings(&content, cx)+ .expect("test settings should parse");+ });+ });+ let provider = cx.update(|cx| {+ BedrockLanguageModelProvider::new(+ http_client::FakeHttpClient::with_404_response(),+ Arc::new(NoCredentialsProvider),+ cx,+ )+ });++ let shared = cx+ .update(|cx| provider.provided_models(cx))+ .into_iter()+ .filter(|model| model.id.0.as_ref() == "shared-model")+ .collect::<Vec<_>>();+ assert_eq!(shared.len(), 1, "each id should be offered once");+ assert_eq!(shared[0].name.0.as_ref(), "Mantle Shared");++ let config = cx+ .update(|cx| provider.config(&shared[0], cx))+ .expect("the offered model should resolve");+ assert!(matches!(+ config,+ BedrockModelConfig::Mantle(model) if model.display_name() == "Mantle Shared"+ ));+ }++ #[test]+ fn environment_bearer_token_prefers_the_bedrock_variable_like_the_aws_sdk() {+ let environment = |variables: &[(&str, &str)]| {+ let variables: HashMap<String, String> = variables+ .iter()+ .map(|(name, value)| (name.to_string(), value.to_string()))+ .collect();+ move |name: &str| variables.get(name).cloned()+ };++ assert_eq!(environment_bearer_token(environment(&[])), None);+ assert_eq!(+ environment_bearer_token(environment(&[("AWS_BEARER_TOKEN", "generic")])),+ Some("generic".to_string())+ );+ assert_eq!(+ environment_bearer_token(environment(&[+ ("AWS_BEARER_TOKEN", "generic"),+ ("AWS_BEARER_TOKEN_BEDROCK", "bedrock"),+ ])),+ Some("bedrock".to_string())+ );+ assert_eq!(+ environment_bearer_token(environment(&[+ ("AWS_BEARER_TOKEN", "generic"),+ ("AWS_BEARER_TOKEN_BEDROCK", ""),+ ])),+ Some(String::new()),+ "a set but empty Bedrock variable still wins, as in the SDK"+ );+ }++ #[gpui::test]+ async fn expiring_credentials_are_cached_per_auth_until_near_expiry(+ cx: &mut gpui::TestAppContext,+ ) {+ let provider = test_provider(cx);+ let auth = Some(BedrockAuth::NamedProfile {+ profile_name: "work".into(),+ });+ let other_auth = Some(BedrockAuth::NamedProfile {+ profile_name: "other".into(),+ });+ provider+ .state+ .update(cx, |state, _| state.set_auth(auth.clone()));++ // Resolved credentials that expire are cached and reused.+ let resolved = provider+ .resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| {+ resolved_credentials("first", Some(Duration::from_secs(60 * 60)))+ })+ .await;+ assert_eq!(access_key_id(resolved), "first");+ let reused = provider+ .resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| {+ resolved_credentials("second", None)+ })+ .await;+ assert_eq!(+ access_key_id(reused),+ "first",+ "unexpired credentials are reused"+ );++ // Credentials close to expiry, or cached for other auth, are replaced.+ for (cached_auth, expires_in, case) in [+ (&auth, Duration::from_secs(60), "near expiry"),+ (&other_auth, Duration::from_secs(60 * 60), "different auth"),+ ] {+ provider.state.update(cx, |state, _| {+ state.expiring_credentials = Some(ExpiringCredentials {+ auth: cached_auth.clone(),+ region: state.get_region(),+ credentials: Credentials::new(+ "cached",+ "secret",+ None,+ Some(SystemTime::now() + expires_in),+ "test",+ ),+ });+ });+ let resolved = provider+ .resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| {+ resolved_credentials("fresh", None)+ })+ .await;+ assert_eq!(access_key_id(resolved), "fresh", "{case}");+ }++ // Credentials without an expiry are never cached.+ provider+ .state+ .update(cx, |state, _| state.expiring_credentials = None);+ provider+ .resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| {+ resolved_credentials("static", None)+ })+ .await+ .expect("auth should resolve");+ assert!(+ provider+ .state+ .read_with(cx, |state, _| state.expiring_credentials.is_none()),+ "credentials without an expiry should not be cached"+ );++ // A resolution that completes after an auth change, even back to the+ // same auth, is not cached.+ let (send_credentials, receive_credentials) = futures::channel::oneshot::channel();+ let pending =+ provider.resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| async move {+ receive_credentials+ .await+ .expect("the test sends credentials")+ });+ provider+ .state+ .update(cx, |state, _| state.set_auth(auth.clone()));+ assert!(+ send_credentials+ .send(resolved_credentials("stale", Some(Duration::from_secs(60 * 60))).await)+ .is_ok(),+ "the pending resolution should still be waiting"+ );+ pending.await.expect("auth should resolve");+ assert!(+ provider+ .state+ .read_with(cx, |state, _| state.expiring_credentials.is_none()),+ "credentials resolved for earlier auth should not be cached"+ );+ }++ #[gpui::test]+ async fn resolving_request_auth_prefers_the_environment_token_and_times_out(+ cx: &mut gpui::TestAppContext,+ ) {+ let provider = test_provider(cx);+ provider.state.update(cx, |state, _| {+ state.set_auth(Some(BedrockAuth::Automatic));+ });++ let request_auth = provider+ .resolve_request_auth_with(+ &cx.to_async(),+ |name| (name == "AWS_BEARER_TOKEN").then(|| "token".to_string()),+ |_, _| async { panic!("credentials should not be resolved") },+ )+ .await+ .expect("auth should resolve");+ assert!(matches!(+ request_auth,+ BedrockRequestAuth::ApiKey { api_key } if api_key == "token"+ ));++ // An API key configured in Zed takes precedence over the environment.+ provider.state.update(cx, |state, _| {+ state.set_auth(Some(BedrockAuth::ApiKey {+ api_key: "configured".into(),+ }));+ });+ let http_client = provider.http_client.clone();+ let request_auth = provider+ .resolve_request_auth_with(+ &cx.to_async(),+ |name| (name == "AWS_BEARER_TOKEN").then(|| "token".to_string()),+ move |auth, region| resolve_request_auth(http_client, auth, region),+ )+ .await+ .expect("auth should resolve");+ assert!(matches!(+ request_auth,+ BedrockRequestAuth::ApiKey { api_key } if api_key == "configured"+ ));++ let pending = provider.resolve_request_auth_with(&cx.to_async(), no_environment, |_, _| {+ futures::future::pending()+ });+ cx.executor().advance_clock(CREDENTIALS_LOAD_TIMEOUT);+ let error = pending.await.err().expect("resolution should time out");+ assert!(error.to_string().contains("timed out"), "{error}");+ }+#[test]fn test_thinking_disallowed_sends_explicit_opt_out_only_on_opus_5() {// Claude Opus 5 runs adaptive thinking by default when the `thinking`@@ -3822,4 +4086,74 @@ mod tests {let request = serde_json::to_value(&request).unwrap();assert!(request.get("context_management").is_none());}++ struct NoCredentialsProvider;++ impl CredentialsProvider for NoCredentialsProvider {+ fn read_credentials<'a>(+ &'a self,+ _url: &'a str,+ _cx: &'a AsyncApp,+ ) -> Pin<Box<dyn Future<Output = Result<Option<(String, Vec<u8>)>>> + 'a>> {+ Box::pin(async { Ok(None) })+ }++ fn write_credentials<'a>(+ &'a self,+ _url: &'a str,+ _username: &'a str,+ _password: &'a [u8],+ _cx: &'a AsyncApp,+ ) -> Pin<Box<dyn Future<Output = Result<()>> + 'a>> {+ Box::pin(async { Ok(()) })+ }++ fn delete_credentials<'a>(+ &'a self,+ _url: &'a str,+ _cx: &'a AsyncApp,+ ) -> Pin<Box<dyn Future<Output = Result<()>> + 'a>> {+ Box::pin(async { Ok(()) })+ }+ }++ fn test_provider(cx: &mut gpui::TestAppContext) -> BedrockLanguageModelProvider {+ cx.executor().allow_parking();+ cx.update(|cx| {+ let settings_store = SettingsStore::test(cx);+ cx.set_global(settings_store);+ gpui_tokio::init(cx);+ BedrockLanguageModelProvider::new(+ http_client::FakeHttpClient::with_404_response(),+ Arc::new(NoCredentialsProvider),+ cx,+ )+ })+ }++ fn no_environment(_: &str) -> Option<String> {+ None+ }++ async fn resolved_credentials(+ access_key_id: &str,+ expires_in: Option<Duration>,+ ) -> Result<BedrockRequestAuth> {+ Ok(BedrockRequestAuth::SigV4 {+ credentials: Credentials::new(+ access_key_id,+ "secret",+ None,+ expires_in.map(|expires_in| SystemTime::now() + expires_in),+ "test",+ ),+ })+ }++ fn access_key_id(request_auth: Result<BedrockRequestAuth>) -> String {+ match request_auth.expect("auth should resolve") {+ BedrockRequestAuth::SigV4 { credentials } => credentials.access_key_id().to_string(),+ BedrockRequestAuth::ApiKey { .. } => panic!("expected SigV4 credentials"),+ }+ }}diff --git a/crates/language_models/src/provider/cloud.rs b/crates/language_models/src/provider/cloud.rsindex a577d8b43a8..f24c6a5ffee 100644--- a/crates/language_models/src/provider/cloud.rs+++ b/crates/language_models/src/provider/cloud.rs@@ -9,14 +9,16 @@ use cloud_api_types::Plan;use futures::FutureExt;use futures::StreamExt;use futures::future::BoxFuture;-use gpui::{AnyElement, App, AppContext, Context, Entity, Subscription, Task, TaskExt};++use gpui::{AnyElement, App, AppContext, AsyncApp, Context, Entity, Subscription, Task, TaskExt};use language_model::{- AuthenticateError, FastModeConfirmation, IconOrSvg, InlineDescription, LanguageModel,+ AuthenticateError, CompactionResult, FastModeConfirmation, IconOrSvg, InlineDescription,+ LanguageModel, LanguageModelCompletionError, LanguageModelCompletionStream,LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, ProviderSettingsView, ZED_CLOUD_PROVIDER_ID,+ LanguageModelProviderState, LanguageModelRequest, ProviderSettingsView, ZED_CLOUD_PROVIDER_ID,ZED_CLOUD_PROVIDER_NAME,};-use language_models_cloud::{CloudLlmTokenProvider, CloudModelProvider};+use language_models_cloud::{CloudLlmTokenProvider, CloudModelProvider, language_model};use rand::{Rng as _, SeedableRng as _, rngs::StdRng};use release_channel::AppVersion;@@ -287,40 +289,84 @@ impl LanguageModelProvider for CloudLanguageModelProvider {IconOrSvg::Icon(IconName::AiZed)}- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {let state = self.state.read(cx);let provider = state.provider.read(cx);- let model = provider.default_model()?;- Some(provider.create_model(model))+ Some(language_model(provider.default_model()?))}- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {let state = self.state.read(cx);let provider = state.provider.read(cx);- let model = provider.default_fast_model()?;- Some(provider.create_model(model))+ Some(language_model(provider.default_fast_model()?))}- fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn recommended_models(&self, cx: &App) -> Vec<LanguageModel> {let state = self.state.read(cx);let provider = state.provider.read(cx);provider.recommended_models().iter()- .map(|model| provider.create_model(model))+ .map(|model| language_model(model)).collect()}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {let state = self.state.read(cx);let provider = state.provider.read(cx);provider.models().iter()- .map(|model| provider.create_model(model))+ .map(|model| language_model(model)).collect()}+ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ cx.update(|cx| {+ self.state+ .read(cx)+ .provider+ .read(cx)+ .stream_completion(model, request, cx)+ })+ }++ fn count_input_tokens(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {+ cx.update(|cx| {+ self.state+ .read(cx)+ .provider+ .read(cx)+ .count_input_tokens(model, request, cx)+ })+ }++ fn compact(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ cx.update(|cx| {+ self.state+ .read(cx)+ .provider+ .read(cx)+ .compact(model, request, cx)+ })+ }+fn is_authenticated(&self, cx: &App) -> bool {let state = self.state.read(cx);!state.is_signed_out(cx)diff --git a/crates/language_models/src/provider/copilot_chat.rs b/crates/language_models/src/provider/copilot_chat.rsindex db39f17c68f..218ffddc446 100644--- a/crates/language_models/src/provider/copilot_chat.rs+++ b/crates/language_models/src/provider/copilot_chat.rs@@ -2,20 +2,23 @@ use std::sync::Arc;use anyhow::Result;use copilot_chat::{- CopilotChat, CopilotChatConfiguration, Model as CopilotChatModel, PROVIDER_ID, PROVIDER_NAME,- create_language_model,+ CopilotChat, CopilotChatConfiguration, PROVIDER_ID, PROVIDER_NAME, language_model,};-use gpui::{App, Entity, Subscription, Task};+use futures::{FutureExt as _, future::BoxFuture};+use gpui::{App, AsyncApp, Entity, Subscription, Task};use language::language_settings::all_language_settings;use language_model::{- AuthenticateError, IconOrSvg, LanguageModel, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelProviderState, ProviderSettingsView,+ AuthenticateError, IconOrSvg, LanguageModel, LanguageModelCompletionError,+ LanguageModelCompletionStream, LanguageModelProvider, LanguageModelProviderId,+ LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest, ModelRateLimiters,+ ProviderSettingsView, unavailable_error,};use settings::SettingsStore;use ui::prelude::*;pub struct CopilotChatLanguageModelProvider {state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -57,15 +60,32 @@ impl CopilotChatLanguageModelProvider {}});- Self { state }+ Self {+ state,+ request_limiters: ModelRateLimiters::default(),+ }}- fn create_language_model(+ /// The current configuration of `model`, and the Copilot Chat client that+ /// serves it, if Copilot Chat still offers it.+ fn config(&self,- model: CopilotChatModel,- copilot_chat: Entity<CopilotChat>,- ) -> Arc<dyn LanguageModel> {- create_language_model(model, copilot_chat)+ model: &LanguageModel,+ cx: &App,+ ) -> Result<(copilot_chat::Model, Entity<CopilotChat>), LanguageModelCompletionError> {+ let unavailable = || unavailable_error(model);+ let copilot_chat = CopilotChat::global(cx).ok_or_else(unavailable)?;+ let config = copilot_chat+ .read(cx)+ .models()+ .and_then(|models| {+ models+ .iter()+ .find(|config| config.id() == model.id.0.as_ref())+ })+ .cloned()+ .ok_or_else(unavailable)?;+ Ok((config, copilot_chat))}}@@ -90,30 +110,40 @@ impl LanguageModelProvider for CopilotChatLanguageModelProvider {IconOrSvg::Icon(IconName::Copilot)}- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {let copilot_chat = CopilotChat::global(cx)?;- let model = copilot_chat.read(cx).models()?.first()?.clone();- Some(self.create_language_model(model, copilot_chat))+ Some(language_model(copilot_chat.read(cx).models()?.first()?))}- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {// The default model should be Copilot Chat's 'base model', which is likely a relatively fast// model (e.g. 4o) and a sensible choice when considering premium requestsself.default_model(cx)}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {let Some(copilot_chat) = CopilotChat::global(cx) else {return Vec::new();};let Some(models) = copilot_chat.read(cx).models() else {return Vec::new();};- models- .iter()- .cloned()- .map(|model| self.create_language_model(model, copilot_chat.clone()))- .collect()+ models.iter().map(language_model).collect()+ }++ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let (config, copilot_chat) = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ copilot_chat::stream_completion(&config, &copilot_chat, &request_limiter, request, cx)}fn is_authenticated(&self, cx: &App) -> bool {diff --git a/crates/language_models/src/provider/deepseek.rs b/crates/language_models/src/provider/deepseek.rsindex 64fe5fbd690..7185f0b8fde 100644--- a/crates/language_models/src/provider/deepseek.rs+++ b/crates/language_models/src/provider/deepseek.rs@@ -9,11 +9,12 @@ use gpui::{App, AppContext, AsyncApp, Context, Entity, SharedString, Task};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,- LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice, LanguageModelToolResultContent, LanguageModelToolUse, MessageContent,- ProviderSettingsView, RateLimiter, Role, StopReason, TokenUsage, env_var,+ LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelCompletionStream,+ LanguageModelEffortLevel, LanguageModelId, LanguageModelName, LanguageModelProvider,+ LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,+ LanguageModelRequest, LanguageModelToolChoice, LanguageModelToolChoiceSupport,+ LanguageModelToolResultContent, LanguageModelToolUse, MessageContent, ModelRateLimiters,+ ProviderSettingsView, RateLimiter, Role, StopReason, TokenUsage, env_var, unavailable_error,};pub use settings::DeepseekAvailableModel as AvailableModel;use settings::{Settings, SettingsStore};@@ -46,6 +47,7 @@ pub struct DeepSeekSettings {pub struct DeepSeekLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -107,17 +109,46 @@ impl DeepSeekLanguageModelProvider {}});- Self { http_client, state }+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ }}- fn create_language_model(&self, model: deepseek::Model) -> Arc<dyn LanguageModel> {- Arc::new(DeepSeekLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })+ /// Every model this provider offers, keyed by id: the built-in models,+ /// with settings entries added or overriding built-in ones.+ fn deepseek_models(&self, cx: &App) -> IndexMap<String, deepseek::Model> {+ let mut models = IndexMap::default();++ for model in [deepseek::Model::V4_1Flash, deepseek::Model::V4Pro] {+ models.insert(model.id().to_string(), model);+ }++ for available_model in &Self::settings(cx).available_models {+ models.insert(+ available_model.name.clone(),+ deepseek::Model::Custom {+ name: available_model.name.clone(),+ display_name: available_model.display_name.clone(),+ max_tokens: available_model.max_tokens,+ max_output_tokens: available_model.max_output_tokens,+ },+ );+ }++ models+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<deepseek::Model, LanguageModelCompletionError> {+ self.deepseek_models(cx)+ .swap_remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}fn settings(cx: &App) -> &DeepSeekSettings {@@ -155,35 +186,22 @@ impl LanguageModelProvider for DeepSeekLanguageModelProvider {IconOrSvg::Icon(IconName::AiDeepSeek)}- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(deepseek::Model::default()))+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.deepseek_models(cx)+ .get(deepseek::Model::default().id())+ .map(language_model)}- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(deepseek::Model::default_fast()))+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ self.deepseek_models(cx)+ .get(deepseek::Model::default_fast().id())+ .map(language_model)}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let mut models = IndexMap::default();-- models.insert("deepseek-flash", deepseek::Model::V4_1Flash);- models.insert("deepseek-v4-pro", deepseek::Model::V4Pro);-- for available_model in &Self::settings(cx).available_models {- models.insert(- &available_model.name,- deepseek::Model::Custom {- name: available_model.name.clone(),- display_name: available_model.display_name.clone(),- max_tokens: available_model.max_tokens,- max_output_tokens: available_model.max_output_tokens,- },- );- }-- models- .into_values()- .map(|model| self.create_language_model(model))+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.deepseek_models(cx)+ .values()+ .map(language_model).collect()}@@ -209,20 +227,42 @@ impl LanguageModelProvider for DeepSeekLanguageModelProvider {self.state.update(cx, |state, cx| state.set_api_key(api_key, cx))}-}-pub struct DeepSeekLanguageModel {- id: LanguageModelId,- model: deepseek::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,+ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request = match into_deepseek(request, &config, config.max_output_tokens()) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let stream = self.stream_deepseek_request(request, &request_limiter, cx);+ let executor = cx.background_executor().clone();++ async move {+ let mapper = DeepSeekEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(stream.await?).boxed(),+ executor,+ ))+ }+ .boxed()+ }}-impl DeepSeekLanguageModel {- fn stream_completion(+impl DeepSeekLanguageModelProvider {+ fn stream_deepseek_request(&self,request: deepseek::Request,+ request_limiter: &RateLimiter,cx: &AsyncApp,) -> BoxFuture<'static, Result<BoxStream<'static, Result<deepseek::StreamResponse>>>> {let http_client = self.http_client.clone();@@ -235,7 +275,7 @@ impl DeepSeekLanguageModel {(state.api_key_state.key(&api_url), api_url, extra_headers)});- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -256,109 +296,45 @@ impl DeepSeekLanguageModel {}}-impl LanguageModel for DeepSeekLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- true- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_thinking(&self) -> bool {- matches!(- self.model,- deepseek::Model::V4_1Flash | deepseek::Model::V4Pro+fn language_model(model: &deepseek::Model) -> LanguageModel {+ let supports_thinking = matches!(model, deepseek::Model::V4_1Flash | deepseek::Model::V4Pro);+ LanguageModel {+ supports_tools: true,+ supports_streaming_tools: true,+ supports_thinking,+ supported_effort_levels: if supports_thinking {+ Arc::new([+ LanguageModelEffortLevel {+ name: "Low".into(),+ value: "low".into(),+ is_default: false,+ },+ LanguageModelEffortLevel {+ name: "High".into(),+ value: "high".into(),+ is_default: true,+ },+ LanguageModelEffortLevel {+ name: "Max".into(),+ value: "max".into(),+ is_default: false,+ },+ ])+ } else {+ Arc::default()+ },+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_images: model.supports_images(),+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("deepseek/{}", model.id()),+ model.max_token_count(),)}-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- if !self.supports_thinking() {- return Vec::new();- }-- vec![- LanguageModelEffortLevel {- name: "Low".into(),- value: "low".into(),- is_default: false,- },- LanguageModelEffortLevel {- name: "High".into(),- value: "high".into(),- is_default: true,- },- LanguageModelEffortLevel {- name: "Max".into(),- value: "max".into(),- is_default: false,- },- ]- }-- fn supports_tool_choice(&self, _choice: LanguageModelToolChoice) -> bool {- true- }-- fn supports_images(&self) -> bool {- self.model.supports_images()- }-- fn telemetry_id(&self) -> String {- format!("deepseek/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let request = match into_deepseek(request, &self.model, self.max_output_tokens()) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let stream = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();-- async move {- let mapper = DeepSeekEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(stream.await?).boxed(),- executor,- ))- }- .boxed()- }}pub fn into_deepseek(diff --git a/crates/language_models/src/provider/google.rs b/crates/language_models/src/provider/google.rsindex 57e00f8d021..e7e69940421 100644--- a/crates/language_models/src/provider/google.rs+++ b/crates/language_models/src/provider/google.rs@@ -8,13 +8,13 @@ use gpui::{App, AppContext, AsyncApp, Context, Entity, SharedString, Task};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyConfiguration, AuthenticateError, EnvVar, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelToolChoice,+ LanguageModelCompletionStream, LanguageModelToolChoiceSupport, unavailable_error,};use language_model::{GOOGLE_PROVIDER_ID, GOOGLE_PROVIDER_NAME, IconOrSvg, LanguageModel, LanguageModelEffortLevel,LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- ProviderSettingsView, RateLimiter,+ LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest, ModelRateLimiters,+ ProviderSettingsView,};use schemars::JsonSchema;use serde::{Deserialize, Serialize};@@ -50,6 +50,7 @@ pub enum ModelMode {pub struct GoogleLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -119,73 +120,24 @@ impl GoogleLanguageModelProvider {}});- Self { http_client, state }- }-- fn create_language_model(&self, model: google_ai::Model) -> Arc<dyn LanguageModel> {- Arc::new(GoogleLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-- fn settings(cx: &App) -> &GoogleSettings {- &crate::AllLanguageModelSettings::get_global(cx).google- }-- fn api_url(cx: &App) -> SharedString {- let api_url = &Self::settings(cx).api_url;- if api_url.is_empty() {- google_ai::API_URL.into()- } else {- SharedString::new(api_url.as_str())+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),}}-}--impl LanguageModelProviderState for GoogleLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for GoogleLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiGoogle)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(google_ai::Model::default()))- }- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(google_ai::Model::default_fast()))- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: the built-in models,+ /// with settings entries added or overriding built-in ones.+ fn google_models(&self, cx: &App) -> BTreeMap<String, google_ai::Model> {let mut models = BTreeMap::default();- // Add base models from google_ai::Model::iter()for model in google_ai::Model::iter() {if !matches!(model, google_ai::Model::Custom { .. }) {models.insert(model.id().to_string(), model);}}- // Override with available models from settingsfor model in &GoogleLanguageModelProvider::settings(cx).available_models {models.insert(model.name.clone(),@@ -199,53 +151,20 @@ impl LanguageModelProvider for GoogleLanguageModelProvider {}models- .into_values()- .map(|model| {- Arc::new(GoogleLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- }) as Arc<dyn LanguageModel>- })- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))}- fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.read(cx);- Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(- state.api_key_state.has_key(),- state.api_key_state.is_from_env_var(),- state.api_key_state.env_var_name().clone(),- "https://aistudio.google.com/app/apikey".into(),- )))- }-- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<google_ai::Model, LanguageModelCompletionError> {+ self.google_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}-}--pub struct GoogleLanguageModel {- id: LanguageModelId,- model: google_ai::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}-impl GoogleLanguageModel {- fn stream_completion(+ fn stream_google_request(&self,request: google_ai::GenerateContentRequest,cx: &AsyncApp,@@ -276,102 +195,138 @@ impl GoogleLanguageModel {}.boxed()}-}-impl LanguageModel for GoogleLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()+ fn settings(cx: &App) -> &GoogleSettings {+ &crate::AllLanguageModelSettings::get_global(cx).google}- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())+ fn api_url(cx: &App) -> SharedString {+ let api_url = &Self::settings(cx).api_url;+ if api_url.is_empty() {+ google_ai::API_URL.into()+ } else {+ SharedString::new(api_url.as_str())+ }}+}- fn provider_id(&self) -> LanguageModelProviderId {+impl LanguageModelProviderState for GoogleLanguageModelProvider {+ type ObservableEntity = State;++ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())+ }+}++impl LanguageModelProvider for GoogleLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {PROVIDER_ID}- fn provider_name(&self) -> LanguageModelProviderName {+ fn name(&self) -> LanguageModelProviderName {PROVIDER_NAME}- fn supports_tools(&self) -> bool {- self.model.supports_tools()+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiGoogle)}- fn supports_images(&self) -> bool {- self.model.supports_images()+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.google_models(cx)+ .get(google_ai::Model::default().id())+ .map(language_model)}- fn supports_thinking(&self) -> bool {- self.model.supports_thinking()+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ self.google_models(cx)+ .get(google_ai::Model::default_fast().id())+ .map(language_model)}- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- let default_level = self.model.default_thinking_level();- self.model- .supported_thinking_levels()- .iter()- .map(|level| LanguageModelEffortLevel {- name: level.name().into(),- value: level.value().into(),- is_default: Some(*level) == default_level,- })+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.google_models(cx)+ .values()+ .map(language_model).collect()}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto- | LanguageModelToolChoice::Any- | LanguageModelToolChoice::None => true,- }+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn telemetry_id(&self) -> String {- format!("google/{}", self.model.request_id())+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn max_token_count(&self) -> u64 {- self.model.max_token_count()+ fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.read(cx);+ Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(+ state.api_key_state.has_key(),+ state.api_key_state.is_from_env_var(),+ state.api_key_state.env_var_name().clone(),+ "https://aistudio.google.com/app/apikey".into(),+ )))}- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()+ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);let mut request = request;if request.max_output_tokens.is_some() {request.max_output_tokens =- request.effective_max_output_tokens(self.max_output_tokens());+ request.effective_max_output_tokens(config.max_output_tokens());}- let request = match into_google(- request,- self.model.request_id().to_string(),- self.model.mode(),- ) {+ let request = match into_google(request, config.request_id().to_string(), config.mode()) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let request = self.stream_completion(request, cx);- let future = self.request_limiter.stream(async move {+ let request = self.stream_google_request(request, cx);+ let future = request_limiter.stream(async move {let response = request.await.map_err(LanguageModelCompletionError::from)?;Ok(GoogleEventMapper::new().map_stream(response))});async move { Ok(future.await?.boxed()) }.boxed()}}++fn language_model(model: &google_ai::Model) -> LanguageModel {+ let default_level = model.default_thinking_level();+ LanguageModel {+ supports_tools: model.supports_tools(),+ supports_images: model.supports_images(),+ supports_thinking: model.supports_thinking(),+ supported_effort_levels: model+ .supported_thinking_levels()+ .iter()+ .map(|level| LanguageModelEffortLevel {+ name: level.name().into(),+ value: level.value().into(),+ is_default: Some(*level) == default_level,+ })+ .collect(),+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("google/{}", model.request_id()),+ model.max_token_count(),+ )+ }+}diff --git a/crates/language_models/src/provider/llama_cpp.rs b/crates/language_models/src/provider/llama_cpp.rsindex c21494ffc4f..e33fe2d7d5d 100644--- a/crates/language_models/src/provider/llama_cpp.rs+++ b/crates/language_models/src/provider/llama_cpp.rs@@ -2,16 +2,17 @@ use anyhow::Result;use collections::{HashMap, HashSet};use credentials_provider::CredentialsProvider;use fs::Fs;-use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};+use futures::{FutureExt, StreamExt, future::BoxFuture};use gpui::{App, AsyncApp, Context, Entity, Task, TaskExt};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, InlineDescription, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,- LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelId,+ LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,- LanguageModelToolResultContent, MessageContent, ProviderSettingsView, RateLimiter, Role,- SubPageProviderSettings, env_var,+ LanguageModelToolChoiceSupport, LanguageModelToolResultContent, MessageContent,+ ModelRateLimiters, ProviderSettingsView, RateLimiter, Role, SubPageProviderSettings, env_var,+ unavailable_error,};use llama_cpp::{LLAMA_CPP_API_URL, ModelEntry, Props, get_models, get_props, stream_chat_completion,@@ -60,10 +61,11 @@ pub struct LlamaCppSettings {pub struct LlamaCppLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,- /// Live capabilities shared with the agent's models (see [`LiveCapabilities`]).+ /// Live capabilities shared with [`State`] (see [`LiveCapabilities`]).capability_cells: CapabilityCells,- /// Live model-load progress shared with the models (see [`LoadingProgress`]).+ /// Live model-load progress shared with [`State`] (see [`LoadingProgress`]).loading_progress: LoadingProgress,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -337,11 +339,10 @@ impl State {}}-/// Capabilities that only become known once a router model loads. `LanguageModel`'s-/// methods take no `cx`, yet the agent reads them live each turn, so we share-/// them through a map that re-discovery updates — an already-selected model-/// picks up real values without re-selection. Image support is available from-/// `/v1/models` hints, so it stays a field.+/// Capabilities that only become known once a router model loads. Re-discovery+/// updates them in a shared map, which each request reads when it is bound, so+/// a request uses the model's real values as soon as it has loaded. Image+/// support is available from `/v1/models` hints, so it is not tracked here.#[derive(Clone, Copy, Debug, PartialEq)]struct LiveCapabilities {max_tokens: u64,@@ -359,7 +360,7 @@ impl LiveCapabilities {}}-/// Live capabilities keyed by model name, shared by the provider and its models.+/// Live capabilities keyed by model name, shared by the provider and its state.type CapabilityCells = Arc<RwLock<HashMap<String, LiveCapabilities>>>;/// Model name → load-status label (e.g. `"Loading weights 42%"`) while a router@@ -402,8 +403,8 @@ fn compute_effective_models(models}-/// Updates the shared capability map from the effective models, so a model held-/// by an open conversation observes the new values (it reads the map by name).+/// Updates the shared capability map from the effective models, so requests+/// for a model held by an open conversation observe the new values.fn sync_capability_cells(cells: &CapabilityCells, effective: &HashMap<String, llama_cpp::Model>) {let mut cells = write_recover(cells);for model in effective.values() {@@ -461,6 +462,7 @@ impl LlamaCppLanguageModelProvider {http_client: http_client.clone(),capability_cells: capability_cells.clone(),loading_progress: loading_progress.clone(),+ request_limiters: ModelRateLimiters::default(),state: cx.new(|cx| {cx.observe_global::<SettingsStore>({let mut last_settings = LlamaCppLanguageModelProvider::settings(cx).clone();@@ -526,6 +528,103 @@ impl LlamaCppLanguageModelProvider {fn has_custom_url(cx: &App) -> bool {Self::settings(cx).api_url != LLAMA_CPP_API_URL}++ /// Every model this provider offers, keyed by id; see [`compute_effective_models`].+ fn llama_cpp_models(&self, cx: &App) -> HashMap<String, llama_cpp::Model> {+ compute_effective_models(+ &self.state.read(cx).fetched_models,+ LlamaCppLanguageModelProvider::settings(cx),+ )+ }++ /// The model's live capabilities, or `model`'s own if the map lacks it.+ fn capabilities(&self, model: &llama_cpp::Model) -> LiveCapabilities {+ read_recover(&self.capability_cells)+ .get(&model.name)+ .copied()+ .unwrap_or_else(|| LiveCapabilities::of(model))+ }++ fn create_language_model(&self, model: &llama_cpp::Model) -> LanguageModel {+ let capabilities = self.capabilities(model);+ let display_name = model.display_name();+ let loading_label = read_recover(&self.loading_progress)+ .get(&model.name)+ .cloned();+ let name = match loading_label {+ // Surface load progress in the display name so it shows wherever the+ // model is named, without provider-agnostic UI changes. The agent+ // rebuilds the name on `ProviderStateChanged`, which our ticks emit.+ Some(label) => LanguageModelName::from(format!("{display_name} · {label}")),+ None => LanguageModelName::from(display_name.to_string()),+ };+ LanguageModel {+ supports_tools: capabilities.supports_tools,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: capabilities.supports_tools,+ any: capabilities.supports_tools,+ none: capabilities.supports_tools,+ },+ supports_images: model.supports_images,+ supports_thinking: capabilities.supports_thinking,+ ..LanguageModel::new(+ LanguageModelId::from(model.name.clone()),+ name,+ PROVIDER_ID,+ PROVIDER_NAME,+ telemetry_id_for(&model.name),+ capabilities.max_tokens,+ )+ }+ }++ /// The current configuration of `model` and its live capabilities, if+ /// this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<(llama_cpp::Model, LiveCapabilities), LanguageModelCompletionError> {+ let config = self+ .llama_cpp_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))?;+ let capabilities = self.capabilities(&config);+ Ok((config, capabilities))+ }++ fn stream_llama_cpp_request(+ &self,+ request: llama_cpp::ChatCompletionRequest,+ request_limiter: &RateLimiter,+ cx: &AsyncApp,+ ) -> BoxFuture<+ 'static,+ Result<futures::stream::BoxStream<'static, Result<llama_cpp::ResponseStreamEvent>>>,+ > {+ let http_client = self.http_client.clone();+ let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {+ let api_url = LlamaCppLanguageModelProvider::api_url(cx);+ let extra_headers = LlamaCppLanguageModelProvider::settings(cx)+ .custom_headers+ .clone();+ (state.api_key_state.key(&api_url), api_url, extra_headers)+ });++ let future = request_limiter.stream(async move {+ let stream = stream_chat_completion(+ http_client.as_ref(),+ &api_url,+ api_key.as_deref(),+ request,+ &extra_headers,+ )+ .await?;+ Ok(stream)+ });++ async move { Ok(future.await?.boxed()) }.boxed()+ }}impl LanguageModelProviderState for LlamaCppLanguageModelProvider {@@ -549,39 +648,25 @@ impl LanguageModelProvider for LlamaCppLanguageModelProvider {IconOrSvg::Icon(IconName::AiLlamaCpp)}- fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, _: &App) -> Option<LanguageModel> {// No default model: in router mode it could trigger an expensive load of// an unloaded model on a constrained machine.None}- fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, _: &App) -> Option<LanguageModel> {// See explanation for default_model.None}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let settings = LlamaCppLanguageModelProvider::settings(cx);- let effective = compute_effective_models(&self.state.read(cx).fetched_models, settings);+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ let effective = self.llama_cpp_models(cx);// Refresh the shared capability map so open conversations pick up settings changes.sync_capability_cells(&self.capability_cells, &effective);let mut models = effective- .into_values()- .map(|model| {- Arc::new(LlamaCppLanguageModel {- id: LanguageModelId::from(model.name.clone()),- name: model.name.clone(),- display_name: model.display_name().to_string(),- fallback_capabilities: LiveCapabilities::of(&model),- supports_images: model.supports_images,- capability_cells: self.capability_cells.clone(),- loading_progress: self.loading_progress.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- state: self.state.clone(),- }) as Arc<dyn LanguageModel>- })+ .values()+ .map(|model| self.create_language_model(model)).collect::<Vec<_>>();models.sort_by_key(|model| model.name());models@@ -607,86 +692,34 @@ impl LanguageModelProvider for LlamaCppLanguageModelProvider {)),))}-}--pub struct LlamaCppLanguageModel {- id: LanguageModelId,- /// The model id sent to the server.- name: String,- display_name: String,- /// Live capabilities shared with the provider, read fresh on each access so an- /// open conversation reflects the model's real values once it has loaded.- capability_cells: CapabilityCells,- /// Used when `capability_cells` has no entry (e.g. model removed mid-conversation).- fallback_capabilities: LiveCapabilities,- /// Available from `/v1/models` hints, so captured at build time.- supports_images: bool,- /// Shared with the provider; this model's load progress, read by `name` so the- /// selector can show a loading indicator.- loading_progress: LoadingProgress,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,- state: Entity<State>,-}--impl LlamaCppLanguageModel {- /// The model's live capabilities, or the build-time fallback if the map lacks it.- fn capabilities(&self) -> LiveCapabilities {- read_recover(&self.capability_cells)- .get(&self.name)- .copied()- .unwrap_or(self.fallback_capabilities)- }-- /// This model's load-status label while loading, read live from the shared map.- fn loading_label(&self) -> Option<SharedString> {- read_recover(&self.loading_progress)- .get(&self.name)- .cloned()- }-- fn to_llama_cpp_request(- &self,- request: LanguageModelRequest,- ) -> Result<llama_cpp::ChatCompletionRequest> {- build_llama_cpp_request(- &self.name,- self.supports_images,- self.capabilities(),- request,- )- }fn stream_completion(&self,- request: llama_cpp::ChatCompletionRequest,+ model: &LanguageModel,+ request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<futures::stream::BoxStream<'static, Result<llama_cpp::ResponseStreamEvent>>>,- > {- let http_client = self.http_client.clone();- let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {- let api_url = LlamaCppLanguageModelProvider::api_url(cx);- let extra_headers = LlamaCppLanguageModelProvider::settings(cx)- .custom_headers- .clone();- (state.api_key_state.key(&api_url), api_url, extra_headers)- });-- let future = self.request_limiter.stream(async move {- let stream = stream_chat_completion(- http_client.as_ref(),- &api_url,- api_key.as_deref(),- request,- &extra_headers,- )- .await?;- Ok(stream)- });-- async move { Ok(future.await?.boxed()) }.boxed()+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let (config, capabilities) = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request = match build_llama_cpp_request(+ &config.name,+ config.supports_images,+ capabilities,+ request,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_llama_cpp_request(request, &request_limiter, cx);+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(mapper.map_stream(completions.await?).boxed())+ }+ .boxed()}}@@ -871,82 +904,6 @@ fn build_llama_cpp_request(})}-impl LanguageModel for LlamaCppLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- match self.loading_label() {- // Surface load progress in the display name so it shows wherever the- // model is named, without provider-agnostic UI changes. The agent- // rebuilds the name on `ProviderStateChanged`, which our ticks emit.- Some(label) => LanguageModelName::from(format!("{} · {}", self.display_name, label)),- None => LanguageModelName::from(self.display_name.clone()),- }- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.capabilities().supports_tools- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- self.supports_tools()- && match choice {- LanguageModelToolChoice::Auto => true,- LanguageModelToolChoice::Any => true,- LanguageModelToolChoice::None => true,- }- }-- fn supports_images(&self) -> bool {- self.supports_images- }-- fn supports_thinking(&self) -> bool {- self.capabilities().supports_thinking- }-- fn telemetry_id(&self) -> String {- telemetry_id_for(&self.name)- }-- fn max_token_count(&self) -> u64 {- self.capabilities().max_tokens- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let request = match self.to_llama_cpp_request(request) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_completion(request, cx);- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(mapper.map_stream(completions.await?).boxed())- }- .boxed()- }-}-fn add_message_content_part(new_part: llama_cpp::MessagePart,role: Role,@@ -1837,6 +1794,80 @@ mod tests {}}+ #[gpui::test]+ fn config_uses_live_capabilities_for_a_stable_model_id(cx: &mut TestAppContext) {+ cx.update(|cx| {+ let settings_store = SettingsStore::test(cx);+ cx.set_global(settings_store);+ });+ let provider = cx.update(|cx| {+ LlamaCppLanguageModelProvider::new(+ FakeHttpClient::with_404_response(),+ Arc::new(FakeCredentialsProvider {+ api_key: Vec::new(),+ }),+ cx,+ )+ });+ cx.run_until_parked();++ let cold = llama_cpp::Model::new(+ "m",+ Some("m"),+ Some(ASSUMED_UNLOADED_CONTEXT),+ true,+ false,+ false,+ );+ provider+ .state+ .update(cx, |state, _| state.fetched_models = vec![cold]);+ write_recover(&provider.loading_progress)+ .insert("m".to_string(), SharedString::from("Loading 42%"));+ let selected = cx.update(|cx| provider.provided_models(cx)).remove(0);+ assert_eq!(selected.id.0.as_ref(), "m");+ assert_eq!(selected.name.0.as_ref(), "m · Loading 42%");+ assert!(!selected.supports_thinking);++ // The model finishes loading and re-discovery reports its real values.+ let loaded = llama_cpp::Model::new("m", Some("m"), Some(262_144), false, false, true);+ provider.state.update(cx, |state, cx| {+ state.fetched_models = vec![loaded];+ sync_capability_cells(+ &state.capability_cells,+ &compute_effective_models(+ &state.fetched_models,+ LlamaCppLanguageModelProvider::settings(cx),+ ),+ );+ });+ write_recover(&provider.loading_progress).remove("m");++ let (config, capabilities) = cx+ .update(|cx| provider.config(&selected, cx))+ .expect("model is still offered");+ assert_eq!(config.name, "m");+ assert_eq!(+ capabilities,+ LiveCapabilities {+ max_tokens: 262_144,+ supports_tools: false,+ supports_thinking: true,+ }+ );+ let refreshed = cx.update(|cx| provider.provided_models(cx)).remove(0);+ assert_eq!(refreshed.id, selected.id);+ assert_eq!(refreshed.name.0.as_ref(), "m");++ provider+ .state+ .update(cx, |state, _| state.fetched_models.clear());+ assert!(matches!(+ cx.update(|cx| provider.config(&selected, cx)),+ Err(LanguageModelCompletionError::ModelUnavailable { .. })+ ));+ }+#[gpui::test]async fn authenticate_fetches_models_after_loading_api_key(cx: &mut TestAppContext) {cx.update(|cx| {diff --git a/crates/language_models/src/provider/lmstudio.rs b/crates/language_models/src/provider/lmstudio.rsindex 310932cfc45..f7ee4ab6ffd 100644--- a/crates/language_models/src/provider/lmstudio.rs+++ b/crates/language_models/src/provider/lmstudio.rs@@ -1,18 +1,19 @@use anyhow::{Result, anyhow};use credentials_provider::CredentialsProvider;use fs::Fs;-use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};+use futures::{FutureExt, StreamExt, future::BoxFuture};use gpui::{App, AsyncApp, Context, Entity, Subscription, Task, TaskExt};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelToolChoice, LanguageModelToolResultContent,- MessageContent, env_var,+ LanguageModelCompletionStream, LanguageModelToolChoice, LanguageModelToolChoiceSupport,+ LanguageModelToolResultContent, MessageContent, env_var, unavailable_error,};use language_model::{InlineDescription, LanguageModelId, LanguageModelName, LanguageModelProvider,LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,- LanguageModelRequest, ProviderSettingsView, RateLimiter, Role, SubPageProviderSettings,+ LanguageModelRequest, ModelRateLimiters, ProviderSettingsView, RateLimiter, Role,+ SubPageProviderSettings,};use lmstudio::{LMSTUDIO_API_URL, ModelType, get_models};@@ -46,6 +47,7 @@ pub struct LmStudioSettings {pub struct LmStudioLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -208,6 +210,7 @@ impl LmStudioLanguageModelProvider {_subscription: subscription,}}),+ request_limiters: ModelRateLimiters::default(),};this.state.update(cx, |state, cx| state.restart_fetch_models_task(cx));@@ -224,6 +227,46 @@ impl LmStudioLanguageModelProvider {fn has_custom_url(cx: &App) -> bool {Self::api_url(cx) != LMSTUDIO_API_URL}++ /// Every model this provider offers, keyed by id: the models LM Studio+ /// reports, with settings entries added or overriding reported ones.+ fn lmstudio_models(&self, cx: &App) -> BTreeMap<String, lmstudio::Model> {+ let mut models: BTreeMap<String, lmstudio::Model> = BTreeMap::default();++ for model in self.state.read(cx).available_models.iter() {+ models.insert(model.name.clone(), model.clone());+ }++ for model in AllLanguageModelSettings::get_global(cx)+ .lmstudio+ .available_models+ .iter()+ {+ models.insert(+ model.name.clone(),+ lmstudio::Model {+ name: model.name.clone(),+ display_name: model.display_name.clone(),+ max_tokens: model.max_tokens,+ supports_tool_calls: model.supports_tool_calls,+ supports_images: model.supports_images,+ },+ );+ }++ models+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<lmstudio::Model, LanguageModelCompletionError> {+ self.lmstudio_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))+ }}impl LanguageModelProviderState for LmStudioLanguageModelProvider {@@ -247,55 +290,22 @@ impl LanguageModelProvider for LmStudioLanguageModelProvider {IconOrSvg::Icon(IconName::AiLmStudio)}- fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, _: &App) -> Option<LanguageModel> {// We shouldn't try to select default model, because it might lead to a load call for an unloaded model.// In a constrained environment where user might not have enough resources it'll be a bad UX to select something// to load by default.None}- fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, _: &App) -> Option<LanguageModel> {// See explanation for default_model.None}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let mut models: BTreeMap<String, lmstudio::Model> = BTreeMap::default();-- // Add models from the LM Studio API- for model in self.state.read(cx).available_models.iter() {- models.insert(model.name.clone(), model.clone());- }-- // Override with available models from settings- for model in AllLanguageModelSettings::get_global(cx)- .lmstudio- .available_models- .iter()- {- models.insert(- model.name.clone(),- lmstudio::Model {- name: model.name.clone(),- display_name: model.display_name.clone(),- max_tokens: model.max_tokens,- supports_tool_calls: model.supports_tool_calls,- supports_images: model.supports_images,- },- );- }-- models- .into_values()- .map(|model| {- Arc::new(LmStudioLanguageModel {- id: LanguageModelId::from(model.name.clone()),- model,- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- state: self.state.clone(),- }) as Arc<dyn LanguageModel>- })+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.lmstudio_models(cx)+ .values()+ .map(language_model).collect()}@@ -319,157 +329,175 @@ impl LanguageModelProvider for LmStudioLanguageModelProvider {)),))}-}--pub struct LmStudioLanguageModel {- id: LanguageModelId,- model: lmstudio::Model,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,- state: Entity<State>,-}-impl LmStudioLanguageModel {- fn to_lmstudio_request(+ fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,- ) -> Result<lmstudio::ChatCompletionRequest> {- if request.contains_custom_tool_input() {- anyhow::bail!("LM Studio does not support custom tools");+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request = match to_lmstudio_request(&config, request) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_lmstudio_request(request, &request_limiter, cx);+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(mapper.map_stream(completions.await?).boxed())}+ .boxed()+ }+}- let mut messages = Vec::new();-- for message in request.messages {- for content in message.content {- match content {- MessageContent::Text(text) => add_message_content_part(- lmstudio::MessagePart::Text { text },+fn to_lmstudio_request(+ config: &lmstudio::Model,+ request: LanguageModelRequest,+) -> Result<lmstudio::ChatCompletionRequest> {+ if request.contains_custom_tool_input() {+ anyhow::bail!("LM Studio does not support custom tools");+ }++ let mut messages = Vec::new();++ for message in request.messages {+ for content in message.content {+ match content {+ MessageContent::Text(text) => add_message_content_part(+ lmstudio::MessagePart::Text { text },+ message.role,+ &mut messages,+ ),+ MessageContent::Thinking { .. } => {}+ MessageContent::RedactedThinking(_) => {}+ MessageContent::Compaction(_) => {}+ MessageContent::Image(image) => {+ add_message_content_part(+ lmstudio::MessagePart::Image {+ image_url: lmstudio::ImageUrl {+ url: image.to_base64_url(),+ detail: None,+ },+ },message.role,&mut messages,- ),- MessageContent::Thinking { .. } => {}- MessageContent::RedactedThinking(_) => {}- MessageContent::Compaction(_) => {}- MessageContent::Image(image) => {- add_message_content_part(- lmstudio::MessagePart::Image {- image_url: lmstudio::ImageUrl {- url: image.to_base64_url(),- detail: None,- },- },- message.role,- &mut messages,- );- }- MessageContent::ToolUse(tool_use) => {- let input = tool_use.input.as_json().ok_or_else(|| {- anyhow!("LM Studio does not support custom tool calls")- })?;- let tool_call = lmstudio::ToolCall {- id: tool_use.id.to_string(),- content: lmstudio::ToolCallContent::Function {- function: lmstudio::FunctionContent {- name: tool_use.name.to_string(),- arguments: serde_json::to_string(input).unwrap_or_default(),- },+ );+ }+ MessageContent::ToolUse(tool_use) => {+ let input = tool_use+ .input+ .as_json()+ .ok_or_else(|| anyhow!("LM Studio does not support custom tool calls"))?;+ let tool_call = lmstudio::ToolCall {+ id: tool_use.id.to_string(),+ content: lmstudio::ToolCallContent::Function {+ function: lmstudio::FunctionContent {+ name: tool_use.name.to_string(),+ arguments: serde_json::to_string(input).unwrap_or_default(),},- };+ },+ };- if let Some(lmstudio::ChatMessage::Assistant { tool_calls, .. }) =- messages.last_mut()- {- tool_calls.push(tool_call);- } else {- messages.push(lmstudio::ChatMessage::Assistant {- content: None,- tool_calls: vec![tool_call],- });- }+ if let Some(lmstudio::ChatMessage::Assistant { tool_calls, .. }) =+ messages.last_mut()+ {+ tool_calls.push(tool_call);+ } else {+ messages.push(lmstudio::ChatMessage::Assistant {+ content: None,+ tool_calls: vec![tool_call],+ });}- MessageContent::ToolResult(tool_result) => {- let content: Vec<lmstudio::MessagePart> = tool_result- .content- .iter()- .map(|part| match part {- LanguageModelToolResultContent::Text(text) => {- lmstudio::MessagePart::Text {- text: text.to_string(),- }+ }+ MessageContent::ToolResult(tool_result) => {+ let content: Vec<lmstudio::MessagePart> = tool_result+ .content+ .iter()+ .map(|part| match part {+ LanguageModelToolResultContent::Text(text) => {+ lmstudio::MessagePart::Text {+ text: text.to_string(),}- LanguageModelToolResultContent::Image(image) => {- lmstudio::MessagePart::Image {- image_url: lmstudio::ImageUrl {- url: image.to_base64_url(),- detail: None,- },- }+ }+ LanguageModelToolResultContent::Image(image) => {+ lmstudio::MessagePart::Image {+ image_url: lmstudio::ImageUrl {+ url: image.to_base64_url(),+ detail: None,+ },}- })- .collect();+ }+ })+ .collect();- messages.push(lmstudio::ChatMessage::Tool {- content: content.into(),- tool_call_id: tool_result.tool_use_id.to_string(),- });- }+ messages.push(lmstudio::ChatMessage::Tool {+ content: content.into(),+ tool_call_id: tool_result.tool_use_id.to_string(),+ });}}}+ }- Ok(lmstudio::ChatCompletionRequest {- model: self.model.name.clone(),- messages,- stream: true,- stream_options: Some(lmstudio::StreamOptions {- include_usage: true,- }),- max_tokens: Some(- request- .max_output_tokens- .map(i32::try_from)- .transpose()?- .unwrap_or(-1),- ),- stop: Some(request.stop),- // In LM Studio you can configure specific settings you'd like to use for your model.- // For example Qwen3 is recommended to be used with 0.7 temperature.- // It would be a bad UX to silently override these settings from Zed, so we pass no temperature as a default.- temperature: request.temperature.or(None),- tools: request- .tools- .into_iter()- .map(|tool| {- let input_schema = match tool.input {- language_model::LanguageModelRequestToolInput::Function {- input_schema,- ..- } => input_schema,- language_model::LanguageModelRequestToolInput::Custom { .. } => {- return Err(anyhow::anyhow!("LM Studio does not support custom tools"));- }- };- Ok(lmstudio::ToolDefinition::Function {- function: lmstudio::FunctionDefinition {- name: tool.name,- description: Some(tool.description),- parameters: Some(input_schema),- },- })+ Ok(lmstudio::ChatCompletionRequest {+ model: config.name.clone(),+ messages,+ stream: true,+ stream_options: Some(lmstudio::StreamOptions {+ include_usage: true,+ }),+ max_tokens: Some(+ request+ .max_output_tokens+ .map(i32::try_from)+ .transpose()?+ .unwrap_or(-1),+ ),+ stop: Some(request.stop),+ // In LM Studio you can configure specific settings you'd like to use for your model.+ // For example Qwen3 is recommended to be used with 0.7 temperature.+ // It would be a bad UX to silently override these settings from Zed, so we pass no temperature as a default.+ temperature: request.temperature.or(None),+ tools: request+ .tools+ .into_iter()+ .map(|tool| {+ let input_schema = match tool.input {+ language_model::LanguageModelRequestToolInput::Function {+ input_schema,+ ..+ } => input_schema,+ language_model::LanguageModelRequestToolInput::Custom { .. } => {+ return Err(anyhow::anyhow!("LM Studio does not support custom tools"));+ }+ };+ Ok(lmstudio::ToolDefinition::Function {+ function: lmstudio::FunctionDefinition {+ name: tool.name,+ description: Some(tool.description),+ parameters: Some(input_schema),+ },})- .collect::<Result<_>>()?,- tool_choice: request.tool_choice.map(|choice| match choice {- LanguageModelToolChoice::Auto => lmstudio::ToolChoice::Auto,- LanguageModelToolChoice::Any => lmstudio::ToolChoice::Required,- LanguageModelToolChoice::None => lmstudio::ToolChoice::None,- }),- })- }+ })+ .collect::<Result<_>>()?,+ tool_choice: request.tool_choice.map(|choice| match choice {+ LanguageModelToolChoice::Auto => lmstudio::ToolChoice::Auto,+ LanguageModelToolChoice::Any => lmstudio::ToolChoice::Required,+ LanguageModelToolChoice::None => lmstudio::ToolChoice::None,+ }),+ })+}- fn stream_completion(+impl LmStudioLanguageModelProvider {+ fn stream_lmstudio_request(&self,request: lmstudio::ChatCompletionRequest,+ request_limiter: &RateLimiter,cx: &AsyncApp,) -> BoxFuture<'static,@@ -485,7 +513,7 @@ impl LmStudioLanguageModel {(state.api_key_state.key(&api_url), api_url, extra_headers)});- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let stream = lmstudio::stream_chat_completion(http_client.as_ref(),&api_url,@@ -501,69 +529,24 @@ impl LmStudioLanguageModel {}}-impl LanguageModel for LmStudioLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.supports_tool_calls()- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- self.supports_tools()- && match choice {- LanguageModelToolChoice::Auto => true,- LanguageModelToolChoice::Any => true,- LanguageModelToolChoice::None => true,- }- }-- fn supports_images(&self) -> bool {- self.model.supports_images- }-- fn telemetry_id(&self) -> String {- format!("lmstudio/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let request = match self.to_lmstudio_request(request) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_completion(request, cx);- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(mapper.map_stream(completions.await?).boxed())- }- .boxed()+fn language_model(model: &lmstudio::Model) -> LanguageModel {+ let supports_tools = model.supports_tool_calls();+ LanguageModel {+ supports_tools,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: supports_tools,+ any: supports_tools,+ none: supports_tools,+ },+ supports_images: model.supports_images,+ ..LanguageModel::new(+ LanguageModelId::from(model.name.clone()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("lmstudio/{}", model.id()),+ model.max_token_count(),+ )}}diff --git a/crates/language_models/src/provider/mistral.rs b/crates/language_models/src/provider/mistral.rsindex 3884896cb8b..bf105035b2b 100644--- a/crates/language_models/src/provider/mistral.rs+++ b/crates/language_models/src/provider/mistral.rs@@ -2,16 +2,17 @@ use anyhow::{Result, anyhow};use collections::{BTreeMap, HashMap};use credentials_provider::CredentialsProvider;-use futures::{FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream};+use futures::{FutureExt, Stream, StreamExt, future::BoxFuture};use gpui::{App, AppContext, AsyncApp, Context, Entity, Global, SharedString, Task};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,- LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,- LanguageModelToolResultContent, LanguageModelToolUse, MessageContent, ProviderSettingsView,- RateLimiter, Role, StopReason, TokenUsage, env_var,+ LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelCompletionStream,+ LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,+ LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,+ LanguageModelToolChoice, LanguageModelToolChoiceSupport, LanguageModelToolResultContent,+ LanguageModelToolUse, MessageContent, ModelRateLimiters, ProviderSettingsView, RateLimiter,+ Role, StopReason, TokenUsage, env_var, unavailable_error,};pub use mistral::{MISTRAL_API_URL, StreamResponse};pub use settings::MistralAvailableModel as AvailableModel;@@ -40,6 +41,7 @@ pub struct MistralSettings {pub struct MistralLanguageModelProvider {http_client: Arc<dyn HttpClient>,pub state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -113,75 +115,26 @@ impl MistralLanguageModelProvider {}});- let this = Arc::new(Self { http_client, state });+ let this = Arc::new(Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ });cx.set_global(GlobalMistralLanguageModelProvider(this));cx.global::<GlobalMistralLanguageModelProvider>().0.clone()}- fn create_language_model(&self, model: mistral::Model) -> Arc<dyn LanguageModel> {- Arc::new(MistralLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-- fn settings(cx: &App) -> &MistralSettings {- &crate::AllLanguageModelSettings::get_global(cx).mistral- }-- pub fn api_url(cx: &App) -> SharedString {- let api_url = &Self::settings(cx).api_url;- if api_url.is_empty() {- mistral::MISTRAL_API_URL.into()- } else {- SharedString::new(api_url.as_str())- }- }-}--impl LanguageModelProviderState for MistralLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for MistralLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiMistral)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(mistral::Model::default()))- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(mistral::Model::default_fast()))- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: the built-in models,+ /// with settings entries added or overriding built-in ones.+ fn mistral_models(&self, cx: &App) -> BTreeMap<String, mistral::Model> {let mut models = BTreeMap::default();- // Add base models from mistral::Model::iter()for model in mistral::Model::iter() {if !matches!(model, mistral::Model::Custom { .. }) {models.insert(model.id().to_string(), model);}}- // Override with available models from settingsfor model in &Self::settings(cx).available_models {models.insert(model.name.clone(),@@ -199,54 +152,22 @@ impl LanguageModelProvider for MistralLanguageModelProvider {}models- .into_values()- .map(|model| {- Arc::new(MistralLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- }) as Arc<dyn LanguageModel>- })- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.read(cx);- Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(- state.api_key_state.has_key(),- state.api_key_state.is_from_env_var(),- state.api_key_state.env_var_name().clone(),- "https://console.mistral.ai/api-keys".into(),- )))}- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<mistral::Model, LanguageModelCompletionError> {+ self.mistral_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}-}--pub struct MistralLanguageModel {- id: LanguageModelId,- model: mistral::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}-impl MistralLanguageModel {- fn stream_completion(+ fn stream_mistral_request(&self,+ request_limiter: &RateLimiter,request: mistral::Request,affinity: Option<String>,cx: &AsyncApp,@@ -264,7 +185,7 @@ impl MistralLanguageModel {(state.api_key_state.key(&api_url), api_url, extra_headers)});- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -284,78 +205,102 @@ impl MistralLanguageModel {async move { Ok(future.await?.boxed()) }.boxed()}-}-impl LanguageModel for MistralLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()+ fn settings(cx: &App) -> &MistralSettings {+ &crate::AllLanguageModelSettings::get_global(cx).mistral}- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())+ pub fn api_url(cx: &App) -> SharedString {+ let api_url = &Self::settings(cx).api_url;+ if api_url.is_empty() {+ mistral::MISTRAL_API_URL.into()+ } else {+ SharedString::new(api_url.as_str())+ }}+}- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID+impl LanguageModelProviderState for MistralLanguageModelProvider {+ type ObservableEntity = State;++ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME+impl LanguageModelProvider for MistralLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {+ PROVIDER_ID}- fn supports_tools(&self) -> bool {- self.model.supports_tools()+ fn name(&self) -> LanguageModelProviderName {+ PROVIDER_NAME}- fn supports_streaming_tools(&self) -> bool {- true+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiMistral)}- fn supports_tool_choice(&self, _choice: LanguageModelToolChoice) -> bool {- self.model.supports_tools()+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.mistral_models(cx)+ .get(mistral::Model::default().id())+ .map(language_model)}- fn supports_images(&self) -> bool {- self.model.supports_images()+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ self.mistral_models(cx)+ .get(mistral::Model::default_fast().id())+ .map(language_model)}- fn supports_thinking(&self) -> bool {- self.model.supports_thinking()+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.mistral_models(cx)+ .values()+ .map(language_model)+ .collect()}- fn supports_disabling_thinking(&self) -> bool {- self.model.supports_disabling_thinking()+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn telemetry_id(&self) -> String {- format!("mistral/{}", self.model.id())+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn max_token_count(&self) -> u64 {- self.model.max_token_count()+ fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.read(cx);+ Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(+ state.api_key_state.has_key(),+ state.api_key_state.is_from_env_var(),+ state.api_key_state.env_var_name().clone(),+ "https://console.mistral.ai/api-keys".into(),+ )))}- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()+ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);let (request, affinity) =- match into_mistral(request, self.model.clone(), self.max_output_tokens()) {+ match into_mistral(request, config.clone(), config.max_output_tokens()) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let stream = self.stream_completion(request, affinity, cx);+ let stream = self.stream_mistral_request(&request_limiter, request, affinity, cx);let executor = cx.background_executor().clone();async move {@@ -370,6 +315,31 @@ impl LanguageModel for MistralLanguageModel {}}+fn language_model(model: &mistral::Model) -> LanguageModel {+ let supports_tools = model.supports_tools();+ LanguageModel {+ supports_tools,+ supports_streaming_tools: true,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: supports_tools,+ any: supports_tools,+ none: supports_tools,+ },+ supports_images: model.supports_images(),+ supports_thinking: model.supports_thinking(),+ supports_disabling_thinking: model.supports_disabling_thinking(),+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("mistral/{}", model.id()),+ model.max_token_count(),+ )+ }+}+pub fn into_mistral(request: LanguageModelRequest,model: mistral::Model,diff --git a/crates/language_models/src/provider/ollama.rs b/crates/language_models/src/provider/ollama.rsindex 486111cadca..d17edc278de 100644--- a/crates/language_models/src/provider/ollama.rs+++ b/crates/language_models/src/provider/ollama.rs@@ -2,18 +2,18 @@ use anyhow::{Result, anyhow};use collections::HashMap;use credentials_provider::CredentialsProvider;use fs::Fs;-use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};+use futures::{FutureExt, StreamExt, future::BoxFuture};use futures::{Stream, TryFutureExt, stream};use gpui::{App, AsyncApp, Context, Entity, Task, TaskExt};use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyState, AuthenticateError, DisabledReason, EnvVar, IconOrSvg, InlineDescription,- LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId,- LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelRequestTool,- LanguageModelToolChoice, LanguageModelToolUse, LanguageModelToolUseId, MessageContent,- ProviderSettingsView, RateLimiter, Role, StopReason, SubPageProviderSettings, TokenUsage,- env_var,+ LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,+ LanguageModelCompletionStream, LanguageModelId, LanguageModelName, LanguageModelProvider,+ LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,+ LanguageModelRequest, LanguageModelRequestTool, LanguageModelToolUse, LanguageModelToolUseId,+ MessageContent, ModelRateLimiters, ProviderSettingsView, Role, StopReason,+ SubPageProviderSettings, TokenUsage, env_var, unavailable_error,};use menu;use ollama::{@@ -54,6 +54,7 @@ pub struct OllamaSettings {pub struct OllamaLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -234,67 +235,18 @@ impl OllamaLanguageModelProvider {credentials_provider,}}),+ request_limiters: ModelRateLimiters::default(),};this}- fn settings(cx: &App) -> &OllamaSettings {- &AllLanguageModelSettings::get_global(cx).ollama- }-- fn api_url(cx: &App) -> SharedString {- let api_url = &Self::settings(cx).api_url;- if api_url.is_empty() {- OLLAMA_API_URL.into()- } else {- SharedString::new(api_url.as_str())- }- }-- fn has_custom_url(cx: &App) -> bool {- Self::settings(cx).api_url != OLLAMA_API_URL- }-}--impl LanguageModelProviderState for OllamaLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for OllamaLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiOllama)- }-- fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {- // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.- // In a constrained environment where user might not have enough resources it'll be a bad UX to select something- // to load by default.- None- }-- fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {- // See explanation for default_model.- None- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: the fetched list when+ /// auto-discovery is on, with settings entries merged over it.+ fn ollama_models(&self, cx: &App) -> HashMap<String, ollama::Model> {let mut models: HashMap<String, ollama::Model> = HashMap::default();let settings = OllamaLanguageModelProvider::settings(cx);if settings.auto_discover {- // Add models from the Ollama APIfor model in self.state.read(cx).fetched_models.iter() {let mut model = model.clone();if let Some(context_window) = settings.context_window {@@ -304,68 +256,35 @@ impl LanguageModelProvider for OllamaLanguageModelProvider {}}- // Override with available models from settingsmerge_settings_into_models(&mut models,&settings.available_models,settings.context_window,);- let mut models = models- .into_values()- .map(|model| {- Arc::new(OllamaLanguageModel {- id: LanguageModelId::from(model.name.clone()),- disabled: model.disabled.as_ref().map(|d| DisabledReason::new(d)),- model,- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- state: self.state.clone(),- }) as Arc<dyn LanguageModel>- })- .collect::<Vec<_>>();- models.sort_by_key(|model| model.name());models}- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.clone();- Some(ProviderSettingsView::SubPage(- SubPageProviderSettings::new(move |window, cx| {- cx.new(|cx| ConfigurationView::new(state.clone(), window, cx))- .into()- })- .description(InlineDescription::Text(- "Run local models on your machine with Ollama.".into(),- )),- ))+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<ollama::Model, LanguageModelCompletionError> {+ self.ollama_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}-}--pub struct OllamaLanguageModel {- id: LanguageModelId,- model: ollama::Model,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,- state: Entity<State>,- disabled: Option<DisabledReason>,-}-impl OllamaLanguageModel {- fn to_ollama_request(&self, request: LanguageModelRequest) -> Result<ChatRequest> {+ fn to_ollama_request(+ config: &ollama::Model,+ request: LanguageModelRequest,+ ) -> Result<ChatRequest> {if request.contains_custom_tool_input() {anyhow::bail!("Ollama does not support custom tools");}- let supports_vision = self.model.supports_vision.unwrap_or(false);+ let supports_vision = config.supports_vision.unwrap_or(false);let mut messages = Vec::with_capacity(request.messages.len());@@ -468,12 +387,12 @@ impl OllamaLanguageModel {}}Ok(ChatRequest {- model: self.model.name.clone(),+ model: config.name.clone(),messages,- keep_alive: self.model.keep_alive.clone().unwrap_or_default(),+ keep_alive: config.keep_alive.clone().unwrap_or_default(),stream: true,options: Some(ChatOptions {- num_ctx: Some(self.model.max_tokens),+ num_ctx: Some(config.max_tokens),num_predict: request.max_output_tokens.map(isize::try_from).transpose()?,// Only send stop tokens if explicitly provided. When empty/None,// Ollama will use the model's default stop tokens from its Modelfile.@@ -486,11 +405,10 @@ impl OllamaLanguageModel {temperature: request.temperature.or(Some(1.0)),..Default::default()}),- think: self- .model+ think: config.supports_thinking.map(|supports_thinking| supports_thinking && request.thinking_allowed),- tools: if self.model.supports_tools.unwrap_or(false) {+ tools: if config.supports_tools.unwrap_or(false) {request.tools.into_iter()@@ -501,69 +419,102 @@ impl OllamaLanguageModel {},})}-}-impl LanguageModel for OllamaLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()+ fn settings(cx: &App) -> &OllamaSettings {+ &AllLanguageModelSettings::get_global(cx).ollama+ }++ fn api_url(cx: &App) -> SharedString {+ let api_url = &Self::settings(cx).api_url;+ if api_url.is_empty() {+ OLLAMA_API_URL.into()+ } else {+ SharedString::new(api_url.as_str())+ }+ }++ fn has_custom_url(cx: &App) -> bool {+ Self::settings(cx).api_url != OLLAMA_API_URL}+}++impl LanguageModelProviderState for OllamaLanguageModelProvider {+ type ObservableEntity = State;- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())+ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn provider_id(&self) -> LanguageModelProviderId {+impl LanguageModelProvider for OllamaLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {PROVIDER_ID}- fn provider_name(&self) -> LanguageModelProviderName {+ fn name(&self) -> LanguageModelProviderName {PROVIDER_NAME}- fn supports_tools(&self) -> bool {- self.model.supports_tools.unwrap_or(false)+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiOllama)}- fn supports_images(&self) -> bool {- self.model.supports_vision.unwrap_or(false)+ fn default_model(&self, _: &App) -> Option<LanguageModel> {+ // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.+ // In a constrained environment where user might not have enough resources it'll be a bad UX to select something+ // to load by default.+ None}- fn supports_thinking(&self) -> bool {- self.model.supports_thinking.unwrap_or(false)+ fn default_fast_model(&self, _: &App) -> Option<LanguageModel> {+ // See explanation for default_model.+ None}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto => false,- LanguageModelToolChoice::Any => false,- LanguageModelToolChoice::None => false,- }+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ let mut models = self+ .ollama_models(cx)+ .values()+ .map(language_model)+ .collect::<Vec<_>>();+ models.sort_by_key(|model| model.name());+ models}- fn telemetry_id(&self) -> String {- format!("ollama/{}", self.model.id())+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn is_disabled(&self) -> Option<DisabledReason> {- self.disabled.clone()+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn max_token_count(&self) -> u64 {- self.model.max_token_count()+ fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.clone();+ Some(ProviderSettingsView::SubPage(+ SubPageProviderSettings::new(move |window, cx| {+ cx.new(|cx| ConfigurationView::new(state.clone(), window, cx))+ .into()+ })+ .description(InlineDescription::Text(+ "Run local models on your machine with Ollama.".into(),+ )),+ ))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- let request = match self.to_ollama_request(request) {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request = match Self::to_ollama_request(&config, request) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};@@ -577,7 +528,7 @@ impl LanguageModel for OllamaLanguageModel {(state.api_key_state.key(&api_url), api_url, extra_headers)});- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let stream = stream_chat_completion(http_client.as_ref(),&api_url,@@ -594,6 +545,23 @@ impl LanguageModel for OllamaLanguageModel {}}+fn language_model(model: &ollama::Model) -> LanguageModel {+ LanguageModel {+ supports_tools: model.supports_tools.unwrap_or(false),+ supports_images: model.supports_vision.unwrap_or(false),+ supports_thinking: model.supports_thinking.unwrap_or(false),+ disabled_reason: model.disabled.as_ref().map(|d| DisabledReason::new(d)),+ ..LanguageModel::new(+ LanguageModelId::from(model.name.clone()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("ollama/{}", model.id()),+ model.max_token_count(),+ )+ }+}+fn map_to_language_model_completion_events(stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {diff --git a/crates/language_models/src/provider/open_ai.rs b/crates/language_models/src/provider/open_ai.rsindex 496c7b867ff..6af07c57edd 100644--- a/crates/language_models/src/provider/open_ai.rs+++ b/crates/language_models/src/provider/open_ai.rs@@ -7,10 +7,11 @@ use http_client::{CustomHeaders, HttpClient};use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, CompactionResult, EnvVar,FastModeConfirmation, IconOrSvg, LanguageModel, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,+ LanguageModelCompletionStream, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice, OPEN_AI_PROVIDER_ID,- OPEN_AI_PROVIDER_NAME, ProviderSettingsView, RateLimiter, env_var, stream_in_background,+ LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoiceSupport,+ ModelRateLimiters, OPEN_AI_PROVIDER_ID, OPEN_AI_PROVIDER_NAME, ProviderSettingsView,+ RateLimiter, env_var, stream_in_background, unavailable_error,};use open_ai::{ResponseStreamEvent,@@ -48,6 +49,7 @@ pub struct OpenAiSettings {pub struct OpenAiLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -109,78 +111,25 @@ impl OpenAiLanguageModelProvider {}});- Self { http_client, state }- }-- fn create_language_model(&self, model: open_ai::Model) -> Arc<dyn LanguageModel> {- Arc::new(OpenAiLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-- fn settings(cx: &App) -> &OpenAiSettings {- &crate::AllLanguageModelSettings::get_global(cx).openai- }-- fn api_url(cx: &App) -> SharedString {- let api_url = &Self::settings(cx).api_url;- if api_url.is_empty() {- open_ai::OPEN_AI_API_URL.into()- } else {- SharedString::new(api_url.as_str())+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),}}-}--impl LanguageModelProviderState for OpenAiLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for OpenAiLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiOpenAi)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(open_ai::Model::default()))- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(open_ai::Model::default_fast()))- }-- fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {- vec![self.create_language_model(open_ai::Model::FivePointSixSol)]- }- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: the built-in models,+ /// with settings entries added or overriding built-in ones.+ fn open_ai_models(&self, cx: &App) -> BTreeMap<String, open_ai::Model> {let mut models = BTreeMap::default();- // Add base models from open_ai::Model::iter()for model in open_ai::Model::iter() {if !matches!(model, open_ai::Model::Custom { .. }) {models.insert(model.id().to_string(), model);}}- // Override with available models from settings- for model in &OpenAiLanguageModelProvider::settings(cx).available_models {+ for model in &Self::settings(cx).available_models {models.insert(model.name.clone(),open_ai::Model::Custom {@@ -197,174 +146,22 @@ impl LanguageModelProvider for OpenAiLanguageModelProvider {}models- .into_values()- .map(|model| self.create_language_model(model))- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.read(cx);- Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(- state.api_key_state.has_key(),- state.api_key_state.is_from_env_var(),- state.api_key_state.env_var_name().clone(),- "https://platform.openai.com/api-keys".into(),- )))- }-- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))- }-- fn fast_mode_confirmation(&self, _cx: &App) -> Option<FastModeConfirmation> {- Some(FastModeConfirmation {- title: "Enable Fast Mode for OpenAI?".into(),- message: "Fast mode sends requests using OpenAI's Priority processing tier, which \- targets significantly lower latency than the standard tier and is billed at a \- premium per-token rate."- .into(),- })- }-}--fn default_thinking_reasoning_effort(model: &open_ai::Model) -> Option<open_ai::ReasoningEffort> {- use open_ai::ReasoningEffort;-- model- .reasoning_effort()- .filter(|effort| open_ai_reasoning_effort_is_supported(*effort))- .or_else(|| {- let supported_efforts = model.supported_reasoning_efforts();- if supported_efforts.contains(&ReasoningEffort::Medium) {- Some(ReasoningEffort::Medium)- } else {- supported_efforts- .iter()- .copied()- .find(|effort| open_ai_reasoning_effort_is_supported(*effort))- }- })-}--fn open_ai_reasoning_effort_is_supported(effort: open_ai::ReasoningEffort) -> bool {- effort != open_ai::ReasoningEffort::None-}--fn normalize_open_ai_response_thinking_effort(- request: &mut LanguageModelRequest,- model: &open_ai::Model,-) {- let selected_effort_is_supported = request- .thinking_effort- .as_deref()- .and_then(|effort| effort.parse::<open_ai::ReasoningEffort>().ok())- .is_some_and(|effort| {- open_ai_reasoning_effort_is_supported(effort)- && model.supported_reasoning_efforts().contains(&effort)- });-- if !selected_effort_is_supported {- request.thinking_effort = None;- }-}--fn supports_selectable_thinking_effort(model: &open_ai::Model) -> bool {- model.uses_responses_api()- && model- .supported_reasoning_efforts()- .iter()- .any(|effort| open_ai_reasoning_effort_is_supported(*effort))-}--fn supported_thinking_effort_levels(model: &open_ai::Model) -> Vec<LanguageModelEffortLevel> {- if !supports_selectable_thinking_effort(model) {- return Vec::new();- }-- let default_effort = default_thinking_reasoning_effort(model);- model- .supported_reasoning_efforts()- .iter()- .copied()- .filter_map(|effort| {- if !open_ai_reasoning_effort_is_supported(effort) {- return None;- }-- Some(LanguageModelEffortLevel {- name: effort.label().into(),- value: effort.value().into(),- is_default: Some(effort) == default_effort,- })- })- .collect()-}--#[cfg(test)]-mod tests {- use super::*;-- #[test]- fn supported_thinking_effort_levels_hide_none() {- let effort_levels = supported_thinking_effort_levels(&open_ai::Model::FivePointTwo);- let values = effort_levels- .iter()- .map(|level| level.value.as_ref())- .collect::<Vec<_>>();-- assert_eq!(values, ["low", "medium", "high", "xhigh"]);- assert_eq!(- effort_levels- .iter()- .find(|level| level.is_default)- .map(|level| level.value.as_ref()),- Some("medium")- );}- #[test]- fn models_supporting_only_none_have_no_selectable_thinking_effort() {- let model = open_ai::Model::Custom {- name: "custom-model".to_string(),- display_name: None,- max_tokens: 128_000,- max_output_tokens: None,- max_completion_tokens: None,- reasoning_effort: Some(open_ai::ReasoningEffort::None),- supports_chat_completions: false,- supports_images: true,- };-- assert!(!supports_selectable_thinking_effort(&model));- assert!(supported_thinking_effort_levels(&model).is_empty());- assert!(- model- .supported_reasoning_efforts()- .contains(&open_ai::ReasoningEffort::None)- );+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<open_ai::Model, LanguageModelCompletionError> {+ self.open_ai_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}-}--pub struct OpenAiLanguageModel {- id: LanguageModelId,- model: open_ai::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}-impl OpenAiLanguageModel {- fn stream_completion(+ fn stream_chat_completion(&self,+ request_limiter: &RateLimiter,request: open_ai::Request,cx: &AsyncApp,) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>>>@@ -379,7 +176,7 @@ impl OpenAiLanguageModel {(state.api_key_state.key(&api_url), api_url, extra_headers)});- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let provider = PROVIDER_NAME;let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey { provider });@@ -401,6 +198,7 @@ impl OpenAiLanguageModel {fn stream_response(&self,+ request_limiter: &RateLimiter,request: ResponseRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponsesStreamEvent>>>>@@ -416,7 +214,7 @@ impl OpenAiLanguageModel {});let provider = PROVIDER_NAME;- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey { provider });};@@ -437,6 +235,7 @@ impl OpenAiLanguageModel {fn compact_response(&self,+ request_limiter: &RateLimiter,request: CompactRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<CompactedResponse, LanguageModelCompletionError>> {@@ -451,7 +250,7 @@ impl OpenAiLanguageModel {});let provider = PROVIDER_NAME;- let future = self.request_limiter.run(async move {+ let future = request_limiter.run(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey { provider });};@@ -468,167 +267,184 @@ impl OpenAiLanguageModel {future.boxed()}-}-impl LanguageModel for OpenAiLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()+ fn settings(cx: &App) -> &OpenAiSettings {+ &crate::AllLanguageModelSettings::get_global(cx).openai+ }++ fn api_url(cx: &App) -> SharedString {+ let api_url = &Self::settings(cx).api_url;+ if api_url.is_empty() {+ open_ai::OPEN_AI_API_URL.into()+ } else {+ SharedString::new(api_url.as_str())+ }}+}++impl LanguageModelProviderState for OpenAiLanguageModelProvider {+ type ObservableEntity = State;- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())+ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn provider_id(&self) -> LanguageModelProviderId {+impl LanguageModelProvider for OpenAiLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {PROVIDER_ID}- fn provider_name(&self) -> LanguageModelProviderName {+ fn name(&self) -> LanguageModelProviderName {PROVIDER_NAME}- fn supports_tools(&self) -> bool {- true- }-- fn supports_images(&self) -> bool {- use open_ai::Model;- match &self.model {- Model::FourOmniMini- | Model::Five- | Model::FiveMini- | Model::FiveNano- | Model::FivePointOne- | Model::FivePointTwo- | Model::FivePointThreeCodex- | Model::FivePointFour- | Model::FivePointFourMini- | Model::FivePointFourNano- | Model::FivePointFourPro- | Model::FivePointFive- | Model::FivePointFivePro- | Model::FivePointSixSol- | Model::FivePointSixTerra- | Model::FivePointSixLuna- | Model::SixAstra- | Model::SixSol- | Model::SixLuna- | Model::O3 => true,- Model::Four => false,- Model::Custom {- supports_images, ..- } => *supports_images,- }+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiOpenAi)}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto => true,- LanguageModelToolChoice::Any => true,- LanguageModelToolChoice::None => true,- }+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.open_ai_models(cx)+ .get(open_ai::Model::default().id())+ .map(language_model)}- fn supports_streaming_tools(&self) -> bool {- true+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ self.open_ai_models(cx)+ .get(open_ai::Model::default_fast().id())+ .map(language_model)}- fn supports_thinking(&self) -> bool {- supports_selectable_thinking_effort(&self.model)+ fn recommended_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.open_ai_models(cx)+ .get(open_ai::Model::FivePointSixSol.id())+ .map(language_model)+ .into_iter()+ .collect()}- fn supports_fast_mode(&self) -> bool {- self.model.supports_priority()+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.open_ai_models(cx)+ .values()+ .map(language_model)+ .collect()}- fn supports_server_side_compaction(&self) -> bool {- self.model.supports_compaction()+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn supports_explicit_compaction(&self) -> bool {- self.model.supports_compaction()+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn compact(+ fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.read(cx);+ Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(+ state.api_key_state.has_key(),+ state.api_key_state.is_from_env_var(),+ state.api_key_state.env_var_name().clone(),+ "https://platform.openai.com/api-keys".into(),+ )))+ }++ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))+ }++ fn stream_completion(&self,+ model: &LanguageModel,mut request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- if !self.supports_explicit_compaction() {- return async {- Err(LanguageModelCompletionError::Other(anyhow::anyhow!(- "this OpenAI model does not support explicit compaction"- )))- }- .boxed();- }-- normalize_open_ai_response_thinking_effort(&mut request, &self.model);- let request = match into_open_ai_response(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- self.max_output_tokens(),- default_thinking_reasoning_effort(&self.model),- self.model- .supported_reasoning_efforts()- .contains(&open_ai::ReasoningEffort::None),- &OPEN_AI_PROVIDER_ID,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),};- let request = request.into_compact_request();- let response = self.compact_response(request, cx);- async move {- let response = response.await?;- let usage = token_usage_from_response_usage(&response.usage);- let context = response- .into_compacted_context(OPEN_AI_PROVIDER_ID)- .map_err(LanguageModelCompletionError::Other)?;- Ok(CompactionResult { context, usage })+ let request_limiter = self.request_limiters.for_model(&model.id);+ if !config.supports_priority() {+ request.speed = None;+ }+ if config.uses_responses_api() {+ normalize_open_ai_response_thinking_effort(&mut request, &config);+ let request = match into_open_ai_response(+ request,+ config.id(),+ config.supports_parallel_tool_calls(),+ config.supports_prompt_cache_key(),+ config.max_output_tokens(),+ default_thinking_reasoning_effort(&config),+ config+ .supported_reasoning_efforts()+ .contains(&open_ai::ReasoningEffort::None),+ &OPEN_AI_PROVIDER_ID,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_response(&request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = OpenAiResponseEventMapper::new(OPEN_AI_PROVIDER_ID);+ Ok(stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()+ } else {+ let request = match into_open_ai(+ request,+ config.id(),+ config.supports_parallel_tool_calls(),+ config.supports_prompt_cache_key(),+ config.max_output_tokens(),+ ChatCompletionMaxTokensParameter::MaxCompletionTokens,+ None,+ false,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_chat_completion(&request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()}- .boxed()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- supported_thinking_effort_levels(&self.model)- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn telemetry_id(&self) -> String {- format!("openai/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()}fn count_input_tokens(&self,+ model: &LanguageModel,mut request: LanguageModelRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {- if !self.model.uses_responses_api() {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if !config.uses_responses_api() {return async { Ok(None) }.boxed();}- normalize_open_ai_response_thinking_effort(&mut request, &self.model);+ normalize_open_ai_response_thinking_effort(&mut request, &config);let request = into_open_ai_response(request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- self.max_output_tokens(),- default_thinking_reasoning_effort(&self.model),- self.model+ config.id(),+ config.supports_parallel_tool_calls(),+ config.supports_prompt_cache_key(),+ config.max_output_tokens(),+ default_thinking_reasoning_effort(&config),+ config.supported_reasoning_efforts().contains(&open_ai::ReasoningEffort::None),&OPEN_AI_PROVIDER_ID,@@ -641,7 +457,7 @@ impl LanguageModel for OpenAiLanguageModel {.clone();(state.api_key_state.key(&api_url), api_url, extra_headers)});- self.request_limiter+ request_limiter.run(async move {let request = request?.into_count_tokens_request();let api_key = api_key.ok_or(LanguageModelCompletionError::NoApiKey {@@ -662,74 +478,231 @@ impl LanguageModel for OpenAiLanguageModel {.boxed()}- fn stream_completion(+ fn compact(&self,+ model: &LanguageModel,mut request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- if !self.model.supports_priority() {- request.speed = None;+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if !config.supports_compaction() {+ return async {+ Err(LanguageModelCompletionError::Other(anyhow::anyhow!(+ "this OpenAI model does not support explicit compaction"+ )))+ }+ .boxed();}- if self.model.uses_responses_api() {- normalize_open_ai_response_thinking_effort(&mut request, &self.model);- let request = match into_open_ai_response(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- self.max_output_tokens(),- default_thinking_reasoning_effort(&self.model),- self.model- .supported_reasoning_efforts()- .contains(&open_ai::ReasoningEffort::None),- &OPEN_AI_PROVIDER_ID,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_response(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = OpenAiResponseEventMapper::new(OPEN_AI_PROVIDER_ID);- Ok(stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))++ normalize_open_ai_response_thinking_effort(&mut request, &config);+ let request = match into_open_ai_response(+ request,+ config.id(),+ config.supports_parallel_tool_calls(),+ config.supports_prompt_cache_key(),+ config.max_output_tokens(),+ default_thinking_reasoning_effort(&config),+ config+ .supported_reasoning_efforts()+ .contains(&open_ai::ReasoningEffort::None),+ &OPEN_AI_PROVIDER_ID,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let request = request.into_compact_request();+ let response = self.compact_response(&request_limiter, request, cx);+ async move {+ let response = response.await?;+ let usage = token_usage_from_response_usage(&response.usage);+ let context = response+ .into_compacted_context(OPEN_AI_PROVIDER_ID)+ .map_err(LanguageModelCompletionError::Other)?;+ Ok(CompactionResult { context, usage })+ }+ .boxed()+ }++ fn fast_mode_confirmation(&self, _cx: &App) -> Option<FastModeConfirmation> {+ Some(FastModeConfirmation {+ title: "Enable Fast Mode for OpenAI?".into(),+ message: "Fast mode sends requests using OpenAI's Priority processing tier, which \+ targets significantly lower latency than the standard tier and is billed at a \+ premium per-token rate."+ .into(),+ })+ }+}++fn default_thinking_reasoning_effort(model: &open_ai::Model) -> Option<open_ai::ReasoningEffort> {+ use open_ai::ReasoningEffort;++ model+ .reasoning_effort()+ .filter(|effort| open_ai_reasoning_effort_is_supported(*effort))+ .or_else(|| {+ let supported_efforts = model.supported_reasoning_efforts();+ if supported_efforts.contains(&ReasoningEffort::Medium) {+ Some(ReasoningEffort::Medium)+ } else {+ supported_efforts+ .iter()+ .copied()+ .find(|effort| open_ai_reasoning_effort_is_supported(*effort))}- .boxed()- } else {- let request = match into_open_ai(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- self.max_output_tokens(),- ChatCompletionMaxTokensParameter::MaxCompletionTokens,- None,- false,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))+ })+}++fn open_ai_reasoning_effort_is_supported(effort: open_ai::ReasoningEffort) -> bool {+ effort != open_ai::ReasoningEffort::None+}++fn normalize_open_ai_response_thinking_effort(+ request: &mut LanguageModelRequest,+ model: &open_ai::Model,+) {+ let selected_effort_is_supported = request+ .thinking_effort+ .as_deref()+ .and_then(|effort| effort.parse::<open_ai::ReasoningEffort>().ok())+ .is_some_and(|effort| {+ open_ai_reasoning_effort_is_supported(effort)+ && model.supported_reasoning_efforts().contains(&effort)+ });++ if !selected_effort_is_supported {+ request.thinking_effort = None;+ }+}++fn supports_selectable_thinking_effort(model: &open_ai::Model) -> bool {+ model.uses_responses_api()+ && model+ .supported_reasoning_efforts()+ .iter()+ .any(|effort| open_ai_reasoning_effort_is_supported(*effort))+}++fn supported_thinking_effort_levels(model: &open_ai::Model) -> Vec<LanguageModelEffortLevel> {+ if !supports_selectable_thinking_effort(model) {+ return Vec::new();+ }++ let default_effort = default_thinking_reasoning_effort(model);+ model+ .supported_reasoning_efforts()+ .iter()+ .copied()+ .filter_map(|effort| {+ if !open_ai_reasoning_effort_is_supported(effort) {+ return None;}- .boxed()- }++ Some(LanguageModelEffortLevel {+ name: effort.label().into(),+ value: effort.value().into(),+ is_default: Some(effort) == default_effort,+ })+ })+ .collect()+}++#[cfg(test)]+mod tests {+ use super::*;++ #[test]+ fn supported_thinking_effort_levels_hide_none() {+ let effort_levels = supported_thinking_effort_levels(&open_ai::Model::FivePointTwo);+ let values = effort_levels+ .iter()+ .map(|level| level.value.as_ref())+ .collect::<Vec<_>>();++ assert_eq!(values, ["low", "medium", "high", "xhigh"]);+ assert_eq!(+ effort_levels+ .iter()+ .find(|level| level.is_default)+ .map(|level| level.value.as_ref()),+ Some("medium")+ );+ }++ #[test]+ fn models_supporting_only_none_have_no_selectable_thinking_effort() {+ let model = open_ai::Model::Custom {+ name: "custom-model".to_string(),+ display_name: None,+ max_tokens: 128_000,+ max_output_tokens: None,+ max_completion_tokens: None,+ reasoning_effort: Some(open_ai::ReasoningEffort::None),+ supports_chat_completions: false,+ supports_images: true,+ };++ assert!(!supports_selectable_thinking_effort(&model));+ assert!(supported_thinking_effort_levels(&model).is_empty());+ assert!(+ model+ .supported_reasoning_efforts()+ .contains(&open_ai::ReasoningEffort::None)+ );+ }+}++fn language_model(model: &open_ai::Model) -> LanguageModel {+ use open_ai::Model;+ let supports_images = match model {+ Model::FourOmniMini+ | Model::Five+ | Model::FiveMini+ | Model::FiveNano+ | Model::FivePointOne+ | Model::FivePointTwo+ | Model::FivePointThreeCodex+ | Model::FivePointFour+ | Model::FivePointFourMini+ | Model::FivePointFourNano+ | Model::FivePointFourPro+ | Model::FivePointFive+ | Model::FivePointFivePro+ | Model::FivePointSixSol+ | Model::FivePointSixTerra+ | Model::FivePointSixLuna+ | Model::SixAstra+ | Model::SixSol+ | Model::SixLuna+ | Model::O3 => true,+ Model::Four => false,+ Model::Custom {+ supports_images, ..+ } => *supports_images,+ };+ LanguageModel {+ supports_tools: true,+ supports_images,+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_streaming_tools: true,+ supports_thinking: supports_selectable_thinking_effort(model),+ supports_fast_mode: model.supports_priority(),+ supports_server_side_compaction: model.supports_compaction(),+ supports_explicit_compaction: model.supports_compaction(),+ supported_effort_levels: supported_thinking_effort_levels(model).into(),+ supports_split_token_display: true,+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("openai/{}", model.id()),+ model.max_token_count(),+ )}}diff --git a/crates/language_models/src/provider/open_ai_compatible.rs b/crates/language_models/src/provider/open_ai_compatible.rsindex 090c348f167..e4bcf241162 100644--- a/crates/language_models/src/provider/open_ai_compatible.rs+++ b/crates/language_models/src/provider/open_ai_compatible.rs@@ -6,10 +6,11 @@ use http_client::{CustomHeaders, HttpClient};use language_model::chat_completion::ChatCompletionEventMapper;use language_model::{AuthenticateError, IconOrSvg, LanguageModel, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,+ LanguageModelCompletionStream, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,- ProviderSettingsView, RateLimiter, SubPageProviderSettings,+ LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoiceSupport,+ ModelRateLimiters, ProviderSettingsView, RateLimiter, SubPageProviderSettings,+ unavailable_error,};use open_ai::{ResponseStreamEvent,@@ -50,6 +51,7 @@ pub struct OpenAiCompatibleLanguageModelProvider {name: LanguageModelProviderName,http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}impl OpenAiCompatibleLanguageModelProvider {@@ -75,111 +77,61 @@ impl OpenAiCompatibleLanguageModelProvider {name: id.into(),http_client,state,+ request_limiters: ModelRateLimiters::default(),}}- fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {- Arc::new(OpenAiCompatibleLanguageModel {- id: LanguageModelId::from(model.name.clone()),- provider_id: self.id.clone(),- provider_name: self.name.clone(),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })+ /// Every model this provider offers, in settings order.+ fn available_models<'a>(&self, cx: &'a App) -> &'a [AvailableModel] {+ &self.state.read(cx).settings.available_models}-}--impl LanguageModelProviderState for OpenAiCompatibleLanguageModelProvider {- type ObservableEntity = State;- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for OpenAiCompatibleLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- self.id.clone()- }-- fn name(&self) -> LanguageModelProviderName {- self.name.clone()- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiOpenAiCompat)- }-- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {- self.state- .read(cx)- .settings- .available_models- .first()- .map(|model| self.create_language_model(model.clone()))- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- None- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- self.state- .read(cx)- .settings- .available_models- .iter()- .map(|model| self.create_language_model(model.clone()))- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()- }-- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.clone();- Some(ProviderSettingsView::SubPage(SubPageProviderSettings::new(- move |window, cx| {- cx.new(|cx| {- ApiCompatibleProviderConfigurationView::new(- state.clone(),- "OpenAI",- API_KEY_PLACEHOLDER,- window,- cx,- )- })- .into()+ fn language_model(&self, model: &AvailableModel) -> LanguageModel {+ LanguageModel {+ supports_tools: model.capabilities.tools,+ supports_images: model.capabilities.images,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: model.capabilities.tools,+ any: model.capabilities.tools,+ none: true,},- )))+ supports_streaming_tools: true,+ supports_thinking: default_thinking_reasoning_effort(model).is_some(),+ supported_effort_levels: supported_thinking_effort_levels(model).into(),+ supports_split_token_display: true,+ max_output_tokens: model.max_output_tokens,+ ..LanguageModel::new(+ LanguageModelId::from(model.name.clone()),+ LanguageModelName::from(+ model+ .display_name+ .clone()+ .unwrap_or_else(|| model.name.clone()),+ ),+ self.id.clone(),+ self.name.clone(),+ format!("openai/{}", model.name),+ model.max_tokens,+ )+ }}- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<AvailableModel, LanguageModelCompletionError> {+ self.available_models(cx)+ .iter()+ .find(|available| available.name == model.id.0.as_ref())+ .cloned()+ .ok_or_else(|| unavailable_error(model))}-}--pub struct OpenAiCompatibleLanguageModel {- id: LanguageModelId,- provider_id: LanguageModelProviderId,- provider_name: LanguageModelProviderName,- model: AvailableModel,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}-impl OpenAiCompatibleLanguageModel {- fn stream_completion(+ fn stream_chat_completion(&self,+ request_limiter: &RateLimiter,request: open_ai::Request,cx: &AsyncApp,) -> BoxFuture<@@ -200,8 +152,8 @@ impl OpenAiCompatibleLanguageModel {)});- let provider = self.provider_name.clone();- let future = self.request_limiter.stream(async move {+ let provider = self.name.clone();+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey { provider });};@@ -222,6 +174,7 @@ impl OpenAiCompatibleLanguageModel {fn stream_response(&self,+ request_limiter: &RateLimiter,request: ResponseRequest,cx: &AsyncApp,) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponsesStreamEvent>>>>@@ -237,8 +190,8 @@ impl OpenAiCompatibleLanguageModel {)});- let provider = self.provider_name.clone();- let future = self.request_limiter.stream(async move {+ let provider = self.name.clone();+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey { provider });};@@ -258,6 +211,146 @@ impl OpenAiCompatibleLanguageModel {}}+impl LanguageModelProviderState for OpenAiCompatibleLanguageModelProvider {+ type ObservableEntity = State;++ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())+ }+}++impl LanguageModelProvider for OpenAiCompatibleLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {+ self.id.clone()+ }++ fn name(&self) -> LanguageModelProviderName {+ self.name.clone()+ }++ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiOpenAiCompat)+ }++ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.available_models(cx)+ .first()+ .map(|model| self.language_model(model))+ }++ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ None+ }++ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.available_models(cx)+ .iter()+ .map(|model| self.language_model(model))+ .collect()+ }++ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()+ }++ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))+ }++ fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.clone();+ Some(ProviderSettingsView::SubPage(SubPageProviderSettings::new(+ move |window, cx| {+ cx.new(|cx| {+ ApiCompatibleProviderConfigurationView::new(+ state.clone(),+ "OpenAI",+ API_KEY_PLACEHOLDER,+ window,+ cx,+ )+ })+ .into()+ },+ )))+ }++ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))+ }++ fn stream_completion(+ &self,+ model: &LanguageModel,+ mut request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ // `speed` can leak in from a parent thread's model; this provider never+ // supports fast mode, and arbitrary compatible endpoints reject `service_tier`.+ request.speed = None;++ if config.capabilities.chat_completions {+ let reasoning_effort = chat_completion_reasoning_effort(&request, &config);+ let request = match into_open_ai(+ request,+ &config.name,+ config.capabilities.parallel_tool_calls,+ config.capabilities.prompt_cache_key,+ config.max_output_tokens,+ chat_completion_max_tokens_parameter(&config),+ reasoning_effort,+ config.capabilities.interleaved_reasoning,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_chat_completion(&request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()+ } else {+ disable_response_thinking_for_none_effort(&mut request, &config);+ let request = match into_open_ai_response(+ request,+ &config.name,+ config.capabilities.parallel_tool_calls,+ config.capabilities.prompt_cache_key,+ config.max_output_tokens,+ default_thinking_reasoning_effort(&config),+ supports_none_reasoning_effort(&config),+ &self.id,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_response(&request_limiter, request, cx);+ let compaction_state_owner = self.id.clone();+ let executor = cx.background_executor().clone();+ async move {+ let mapper = OpenAiResponseEventMapper::new(compaction_state_owner);+ Ok(language_model::stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()+ }+ }+}+fn default_thinking_reasoning_effort(model: &AvailableModel) -> Option<open_ai::ReasoningEffort> {model.reasoning_effort@@ -331,147 +424,6 @@ fn disable_response_thinking_for_none_effort(}}-impl LanguageModel for OpenAiCompatibleLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(- self.model- .display_name- .clone()- .unwrap_or_else(|| self.model.name.clone()),- )- }-- fn provider_id(&self) -> LanguageModelProviderId {- self.provider_id.clone()- }-- fn provider_name(&self) -> LanguageModelProviderName {- self.provider_name.clone()- }-- fn supports_tools(&self) -> bool {- self.model.capabilities.tools- }-- fn supports_images(&self) -> bool {- self.model.capabilities.images- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto => self.model.capabilities.tools,- LanguageModelToolChoice::Any => self.model.capabilities.tools,- LanguageModelToolChoice::None => true,- }- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_thinking(&self) -> bool {- default_thinking_reasoning_effort(&self.model).is_some()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- supported_thinking_effort_levels(&self.model)- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn telemetry_id(&self) -> String {- format!("openai/{}", self.model.name)- }-- fn max_token_count(&self) -> u64 {- self.model.max_tokens- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens- }-- fn stream_completion(- &self,- mut request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- // `speed` can leak in from a parent thread's model; this provider never- // supports fast mode, and arbitrary compatible endpoints reject `service_tier`.- if !self.supports_fast_mode() {- request.speed = None;- }-- if self.model.capabilities.chat_completions {- let reasoning_effort = chat_completion_reasoning_effort(&request, &self.model);- let request = match into_open_ai(- request,- &self.model.name,- self.model.capabilities.parallel_tool_calls,- self.model.capabilities.prompt_cache_key,- self.max_output_tokens(),- chat_completion_max_tokens_parameter(&self.model),- reasoning_effort,- self.model.capabilities.interleaved_reasoning,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))- }- .boxed()- } else {- disable_response_thinking_for_none_effort(&mut request, &self.model);- let request = match into_open_ai_response(- request,- &self.model.name,- self.model.capabilities.parallel_tool_calls,- self.model.capabilities.prompt_cache_key,- self.max_output_tokens(),- default_thinking_reasoning_effort(&self.model),- supports_none_reasoning_effort(&self.model),- &self.provider_id,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_response(request, cx);- let compaction_state_owner = self.provider_id.clone();- let executor = cx.background_executor().clone();- async move {- let mapper = OpenAiResponseEventMapper::new(compaction_state_owner);- Ok(language_model::stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))- }- .boxed()- }- }-}-#[cfg(test)]mod tests {use super::*;diff --git a/crates/language_models/src/provider/open_router.rs b/crates/language_models/src/provider/open_router.rsindex 9a11aedce59..c7a23b4e959 100644--- a/crates/language_models/src/provider/open_router.rs+++ b/crates/language_models/src/provider/open_router.rs@@ -7,11 +7,11 @@ use http_client::{CustomHeaders, HttpClient};use language_model::chat_completion::ChatCompletionEventMapper;use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelEffortLevel,LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice, LanguageModelToolResultContent, MessageContent, ProviderSettingsView,- RateLimiter, Role, env_var,+ LanguageModelToolChoice, LanguageModelToolChoiceSupport, LanguageModelToolResultContent,+ MessageContent, ModelRateLimiters, ProviderSettingsView, Role, env_var, unavailable_error,};use open_router::{Model, ModelMode as OpenRouterModelMode, OPEN_ROUTER_API_URL, ReasoningEffort,@@ -39,6 +39,7 @@ pub struct OpenRouterSettings {pub struct OpenRouterLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -159,7 +160,11 @@ impl OpenRouterLanguageModelProvider {}});- Self { http_client, state }+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ }}fn settings(cx: &App) -> &OpenRouterSettings {@@ -175,47 +180,9 @@ impl OpenRouterLanguageModelProvider {}}- fn create_language_model(&self, model: open_router::Model) -> Arc<dyn LanguageModel> {- Arc::new(OpenRouterLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-}--impl LanguageModelProviderState for OpenRouterLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for OpenRouterLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiOpenRouter)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(open_router::Model::default()))- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- None- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider lists, in order: the fetched list, with+ /// settings entries replacing fetched ones of the same name or appended.+ fn open_router_models(&self, cx: &App) -> Vec<open_router::Model> {let mut models_from_api = self.state.read(cx).available_models.clone();let mut settings_models = Vec::new();@@ -257,45 +224,31 @@ impl LanguageModelProvider for OpenRouterLanguageModelProvider {}models_from_api- .into_iter()- .map(|model| self.create_language_model(model))- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()}- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.read(cx);- Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(- state.api_key_state.has_key(),- state.api_key_state.is_from_env_var(),- state.api_key_state.env_var_name().clone(),- "https://openrouter.ai/keys".into(),- )))+ /// The listed model with `id`, or the built-in default model when `id`+ /// names it, even if the fetched list doesn't include it.+ fn open_router_model(&self, id: &str, cx: &App) -> Option<open_router::Model> {+ self.open_router_models(cx)+ .into_iter()+ .find(|listed| listed.id() == id)+ .or_else(|| {+ let default = open_router::Model::default();+ (default.id() == id).then_some(default)+ })}- fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {- self.state- .update(cx, |state, cx| state.set_api_key(api_key, cx))+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<open_router::Model, LanguageModelCompletionError> {+ self.open_router_model(model.id.0.as_ref(), cx)+ .ok_or_else(|| unavailable_error(model))}-}-pub struct OpenRouterLanguageModel {- id: LanguageModelId,- model: open_router::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--impl OpenRouterLanguageModel {- fn stream_completion(+ fn stream_open_router_request(&self,request: open_router::Request,cx: &AsyncApp,@@ -337,111 +290,87 @@ impl OpenRouterLanguageModel {}}-impl LanguageModel for OpenRouterLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }+impl LanguageModelProviderState for OpenRouterLanguageModelProvider {+ type ObservableEntity = State;- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())+ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn provider_id(&self) -> LanguageModelProviderId {+impl LanguageModelProvider for OpenRouterLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {PROVIDER_ID}- fn provider_name(&self) -> LanguageModelProviderName {+ fn name(&self) -> LanguageModelProviderName {PROVIDER_NAME}- fn supports_tools(&self) -> bool {- self.model.supports_tool_calls()- }-- fn supports_streaming_tools(&self) -> bool {- true+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiOpenRouter)}- fn supports_thinking(&self) -> bool {- matches!(- self.model.mode,- OpenRouterModelMode::Thinking { .. } | OpenRouterModelMode::Adaptive- )+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ let default = open_router::Model::default();+ self.open_router_model(default.id(), cx)+ .map(|model| language_model(&model))}- fn supports_disabling_thinking(&self) -> bool {- !self.model.mandatory_reasoning+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ None}- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- let efforts: &[ReasoningEffort] = if !self.model.supported_efforts.is_empty() {- &self.model.supported_efforts- } else if self.model.supports_max_tokens {- &ReasoningEffort::OPENAI_COMPATIBLE_SELECTABLE- } else {- return Vec::new();- };- let default_effort = self.model.default_effort.or_else(|| {- self.model- .supports_max_tokens- .then_some(ReasoningEffort::Medium)- });- efforts+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.open_router_models(cx).iter()- .map(|&effort| LanguageModelEffortLevel {- name: effort.label().into(),- value: effort.value().into(),- is_default: Some(effort) == default_effort,- })+ .map(language_model).collect()}- fn telemetry_id(&self) -> String {- format!("openrouter/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto => true,- LanguageModelToolChoice::Any => true,- LanguageModelToolChoice::None => true,- }+ fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.read(cx);+ Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(+ state.api_key_state.has_key(),+ state.api_key_state.is_from_env_var(),+ state.api_key_state.env_var_name().clone(),+ "https://openrouter.ai/keys".into(),+ )))}- fn supports_images(&self) -> bool {- self.model.supports_images.unwrap_or(false)+ fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {+ self.state+ .update(cx, |state, cx| state.set_api_key(api_key, cx))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);let openrouter_request =- match into_open_router(request, &self.model, self.max_output_tokens()) {+ match into_open_router(request, &config, config.max_output_tokens()) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let request = self.stream_completion(openrouter_request, cx);+ let request = self.stream_open_router_request(openrouter_request, cx);let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let response = request.await?;let events = ChatCompletionEventMapper::new().map_stream(response);Ok(language_model::stream_in_background(@@ -453,6 +382,48 @@ impl LanguageModel for OpenRouterLanguageModel {}}+fn language_model(model: &open_router::Model) -> LanguageModel {+ let efforts: &[ReasoningEffort] = if !model.supported_efforts.is_empty() {+ &model.supported_efforts+ } else if model.supports_max_tokens {+ &ReasoningEffort::OPENAI_COMPATIBLE_SELECTABLE+ } else {+ &[]+ };+ let default_effort = model+ .default_effort+ .or_else(|| model.supports_max_tokens.then_some(ReasoningEffort::Medium));+ let supported_effort_levels = efforts+ .iter()+ .map(|&effort| LanguageModelEffortLevel {+ name: effort.label().into(),+ value: effort.value().into(),+ is_default: Some(effort) == default_effort,+ })+ .collect();+ LanguageModel {+ supports_tools: model.supports_tool_calls(),+ supports_streaming_tools: true,+ supports_thinking: matches!(+ model.mode,+ OpenRouterModelMode::Thinking { .. } | OpenRouterModelMode::Adaptive+ ),+ supports_disabling_thinking: !model.mandatory_reasoning,+ supported_effort_levels,+ max_output_tokens: model.max_output_tokens(),+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_images: model.supports_images.unwrap_or(false),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("openrouter/{}", model.id()),+ model.max_token_count(),+ )+ }+}+pub fn into_open_router(request: LanguageModelRequest,model: &Model,diff --git a/crates/language_models/src/provider/openai_subscribed.rs b/crates/language_models/src/provider/openai_subscribed.rsindex bac7831a6d3..3ab8e6bb868 100644--- a/crates/language_models/src/provider/openai_subscribed.rs+++ b/crates/language_models/src/provider/openai_subscribed.rs@@ -1,14 +1,18 @@use anyhow::{Result, anyhow};use credentials_provider::CredentialsProvider;-use futures::future::Shared;-use gpui::{App, Context, Entity, SharedString, Task, Window};+use futures::FutureExt as _;+use futures::future::{BoxFuture, Shared};++use gpui::{App, AsyncApp, Context, Entity, SharedString, Task, Window};use http_client::HttpClient;use language_model::{- AuthenticateError, FastModeConfirmation, IconOrSvg, InlineDescription, LanguageModel,+ AuthenticateError, CompactionResult, FastModeConfirmation, IconOrSvg, InlineDescription,+ LanguageModel, LanguageModelCompletionError, LanguageModelCompletionStream,LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, ProviderSettingsView,+ LanguageModelProviderState, LanguageModelRequest, ModelRateLimiters, ProviderSettingsView,+ unavailable_error,};-use openai_subscribed::{PROVIDER_ID, PROVIDER_NAME, State, create_language_model};+use openai_subscribed::{ChatGptModel, PROVIDER_ID, PROVIDER_NAME, State, language_model};use std::sync::Arc;use ui::{ConfiguredApiCard, prelude::*};@@ -17,6 +21,7 @@ const SUBSCRIPTION_DESCRIPTION: &str =pub struct OpenAiSubscribedProvider {state: Entity<State>,+ request_limiters: ModelRateLimiters,}impl OpenAiSubscribedProvider {@@ -26,7 +31,25 @@ impl OpenAiSubscribedProvider {cx: &mut App,) -> Self {let state = cx.new(|cx| State::new(http_client, credentials_provider, cx));- Self { state }+ Self {+ state,+ request_limiters: ModelRateLimiters::default(),+ }+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<ChatGptModel, LanguageModelCompletionError> {+ self.state+ .read(cx)+ .available_models()+ .iter()+ .find(|config| config.id() == model.id.0.as_ref())+ .cloned()+ .ok_or_else(|| unavailable_error(model))}}@@ -51,31 +74,30 @@ impl LanguageModelProvider for OpenAiSubscribedProvider {IconOrSvg::Icon(IconName::AiOpenAiGptSub)}- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {self.state.read(cx).default_model()- .map(|model| create_language_model(model, &self.state, cx))+ .map(|model| language_model(&model))}- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {self.state.read(cx).default_fast_model()- .map(|model| create_language_model(model, &self.state, cx))+ .map(|model| language_model(&model))}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {self.state.read(cx).available_models().iter()- .cloned()- .map(|model| create_language_model(model, &self.state, cx))+ .map(language_model).collect()}- fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn recommended_models(&self, cx: &App) -> Vec<LanguageModel> {self.default_model(cx).into_iter().collect()}@@ -151,6 +173,35 @@ impl LanguageModelProvider for OpenAiSubscribedProvider {.into()}+ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ openai_subscribed::stream_completion(&config, &self.state, &request_limiter, request, cx)+ }++ fn compact(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ openai_subscribed::compact(&config, &self.state, &request_limiter, request, cx)+ }+fn fast_mode_confirmation(&self, _cx: &App) -> Option<FastModeConfirmation> {Some(FastModeConfirmation {title: "Enable Fast Mode for OpenAI?".into(),diff --git a/crates/language_models/src/provider/opencode.rs b/crates/language_models/src/provider/opencode.rsindex 17c490e8242..f446610a98b 100644--- a/crates/language_models/src/provider/opencode.rs+++ b/crates/language_models/src/provider/opencode.rs@@ -9,11 +9,11 @@ use gpui::{use http_client::{AsyncBody, CustomHeaders, HttpClient, Method, Request, http};use language_model::{ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, InlineDescription, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelEffortLevel,LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice, ProviderSettingsView, RateLimiter, ReasoningEffort,- SubPageProviderSettings, env_var,+ LanguageModelToolChoiceSupport, ModelRateLimiters, ProviderSettingsView, RateLimiter,+ ReasoningEffort, SubPageProviderSettings, env_var, unavailable_error,};use opencode::{ApiProtocol, OPENCODE_API_URL, OpenCodeSubscription};use serde::Deserialize;@@ -96,6 +96,7 @@ pub struct OpenCodeSettings {pub struct OpenCodeLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -370,81 +371,19 @@ impl OpenCodeLanguageModelProvider {}});- Self { http_client, state }- }-- fn create_language_model_with_capabilities(- &self,- model: opencode::Model,- subscription: OpenCodeSubscription,- supports_images: bool,- supports_thinking: bool,- ) -> Arc<dyn LanguageModel> {- let id_str = format!("{}/{}", subscription.id_prefix(), model.id());- Arc::new(OpenCodeLanguageModel {- id: LanguageModelId::from(id_str),- model,- subscription,- supports_images,- supports_thinking,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })- }-- pub fn settings(cx: &App) -> &OpenCodeSettings {- &crate::AllLanguageModelSettings::get_global(cx).opencode- }-- fn subscription_enabled(subscription: OpenCodeSubscription, cx: &App) -> bool {- let settings = Self::settings(cx);- match subscription {- OpenCodeSubscription::Zen => settings.show_zen_models,- OpenCodeSubscription::Go => settings.show_go_models,- }- }-- fn api_url(cx: &App) -> SharedString {- let api_url = &Self::settings(cx).api_url;- if api_url.is_empty() {- OPENCODE_API_URL.into()- } else {- SharedString::new(api_url.as_str())+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),}}-}--impl LanguageModelProviderState for OpenCodeLanguageModelProvider {- type ObservableEntity = State;-- fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {- Some(self.state.clone())- }-}--impl LanguageModelProvider for OpenCodeLanguageModelProvider {- fn id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }- fn icon(&self) -> IconOrSvg {- IconOrSvg::Icon(IconName::AiOpenCode)- }-- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- None- }-- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- None- }-- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ /// Every model this provider offers, keyed by id: discovered models of+ /// enabled subscriptions, with settings entries added or overriding them.+ fn opencode_models(+ &self,+ cx: &App,+ ) -> BTreeMap<String, (DiscoveredModel, OpenCodeSubscription)> {let mut models: BTreeMap<String, (DiscoveredModel, OpenCodeSubscription)> =BTreeMap::default();let settings = Self::settings(cx);@@ -454,7 +393,7 @@ impl LanguageModelProvider for OpenCodeLanguageModelProvider {if Self::subscription_enabled(subscription, cx) {if let Some(discovered) = discovered_models.get(&subscription) {for model in discovered {- let key = format!("{}/{}", subscription.id_prefix(), model.model.id());+ let key = model_id(&model.model, subscription);models.insert(key, (model.clone(), subscription));}}@@ -486,7 +425,7 @@ impl LanguageModelProvider for OpenCodeLanguageModelProvider {model.custom_model_api_url.clone(),model.interleaved_reasoning,);- let key = format!("{}/{}", subscription.id_prefix(), model.name);+ let key = model_id(&custom_model, subscription);models.insert(key,(@@ -501,88 +440,27 @@ impl LanguageModelProvider for OpenCodeLanguageModelProvider {}models- .into_values()- .map(|(model, subscription)| {- self.create_language_model_with_capabilities(- model.model,- subscription,- model.supports_images,- model.supports_thinking,- )- })- .collect()- }-- fn is_authenticated(&self, cx: &App) -> bool {- self.state.read(cx).is_authenticated()}- fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {- self.state.update(cx, |state, cx| state.authenticate(cx))- }-- fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {- let state = self.state.clone();- Some(ProviderSettingsView::SubPage(- SubPageProviderSettings::new(move |window, cx| {- cx.new(|cx| ConfigurationView::new(state.clone(), window, cx))- .into()- })- .description(InlineDescription::Text(- "To use OpenCode models in Zed, you need an API key.".into(),- )),- ))- }-}--pub struct OpenCodeLanguageModel {- id: LanguageModelId,- model: opencode::Model,- subscription: OpenCodeSubscription,- supports_images: bool,- supports_thinking: bool,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--struct InjectHeaderClient {- inner: Arc<dyn HttpClient>,- name: http::HeaderName,- value: http::HeaderValue,-}--impl HttpClient for InjectHeaderClient {- fn user_agent(&self) -> Option<&http::HeaderValue> {- self.inner.user_agent()- }-- fn proxy(&self) -> Option<&http_client::Url> {- self.inner.proxy()- }-- fn send(+ /// The current configuration of `model` and the subscription serving it,+ /// if this provider still offers it.+ fn config(&self,- mut req: http::Request<AsyncBody>,- ) -> futures::future::BoxFuture<'static, anyhow::Result<http::Response<AsyncBody>>> {- req.headers_mut()- .insert(self.name.clone(), self.value.clone());- self.inner.send(req)+ model: &LanguageModel,+ cx: &App,+ ) -> Result<(DiscoveredModel, OpenCodeSubscription), LanguageModelCompletionError> {+ self.opencode_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}-}-// Standalone requests do not have a conversation ID, but OpenCode requires a-// non-empty session header.-fn opencode_session_header_value(thread_id: Option<&str>) -> http::HeaderValue {- thread_id- .filter(|thread_id| !thread_id.is_empty())- .and_then(|thread_id| http::HeaderValue::from_str(thread_id).ok())- .unwrap_or_else(|| http::HeaderValue::from(rand::random::<u64>()))-}--impl OpenCodeLanguageModel {- fn base_api_url(&self, cx: &AsyncApp) -> SharedString {- if let Some(url) = self.model.custom_model_api_url() {+ fn base_api_url(+ &self,+ config: &DiscoveredModel,+ subscription: OpenCodeSubscription,+ cx: &AsyncApp,+ ) -> SharedString {+ if let Some(url) = config.model.custom_model_api_url() {if !url.is_empty() {return url.to_string().into();}@@ -593,12 +471,12 @@ impl OpenCodeLanguageModel {.state.read_with(cx, |_, cx| OpenCodeLanguageModelProvider::api_url(cx));- let suffix = self.subscription.api_path_suffix();+ let suffix = subscription.api_path_suffix();let base_str = base.as_ref().trim_end_matches('/');format!("{}{}", base_str, suffix).into()}- fn api_key(&self, cx: &AsyncApp) -> Option<Arc<str>> {+ fn current_api_key(&self, cx: &AsyncApp) -> Option<Arc<str>> {self.state.read_with(cx, |state, cx| {let api_url = OpenCodeLanguageModelProvider::api_url(cx);state.api_key_state.key(&api_url)@@ -615,6 +493,9 @@ impl OpenCodeLanguageModel {fn stream_anthropic(&self,+ config: &DiscoveredModel,+ subscription: OpenCodeSubscription,+ request_limiter: &RateLimiter,request: anthropic::Request,http_client: Arc<dyn HttpClient>,extra_headers: CustomHeaders,@@ -630,10 +511,10 @@ impl OpenCodeLanguageModel {>,> {// Anthropic crate appends /v1/messages to api_url- let api_url = self.base_api_url(cx);- let api_key = self.api_key(cx);+ let api_url = self.base_api_url(config, subscription, cx);+ let api_key = self.current_api_key(cx);- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -656,6 +537,9 @@ impl OpenCodeLanguageModel {fn stream_openai_chat(&self,+ config: &DiscoveredModel,+ subscription: OpenCodeSubscription,+ request_limiter: &RateLimiter,request: open_ai::Request,http_client: Arc<dyn HttpClient>,extra_headers: CustomHeaders,@@ -663,12 +547,12 @@ impl OpenCodeLanguageModel {) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>>>{// OpenAI crate appends /chat/completions to api_url, so we pass base + "/v1"- let base_url = self.base_api_url(cx);+ let base_url = self.base_api_url(config, subscription, cx);let api_url: SharedString = format!("{base_url}/v1").into();- let api_key = self.api_key(cx);+ let api_key = self.current_api_key(cx);let provider_name = PROVIDER_NAME.0.to_string();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -691,6 +575,9 @@ impl OpenCodeLanguageModel {fn stream_openai_response(&self,+ config: &DiscoveredModel,+ subscription: OpenCodeSubscription,+ request_limiter: &RateLimiter,request: open_ai::responses::Request,http_client: Arc<dyn HttpClient>,extra_headers: CustomHeaders,@@ -700,12 +587,12 @@ impl OpenCodeLanguageModel {Result<futures::stream::BoxStream<'static, Result<open_ai::responses::StreamEvent>>>,> {// Responses crate appends /responses to api_url, so we pass base + "/v1"- let base_url = self.base_api_url(cx);+ let base_url = self.base_api_url(config, subscription, cx);let api_url: SharedString = format!("{base_url}/v1").into();- let api_key = self.api_key(cx);+ let api_key = self.current_api_key(cx);let provider_name = PROVIDER_NAME.0.to_string();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -728,6 +615,9 @@ impl OpenCodeLanguageModel {fn stream_google(&self,+ config: &DiscoveredModel,+ subscription: OpenCodeSubscription,+ request_limiter: &RateLimiter,request: google_ai::GenerateContentRequest,http_client: Arc<dyn HttpClient>,extra_headers: CustomHeaders,@@ -736,10 +626,10 @@ impl OpenCodeLanguageModel {'static,Result<futures::stream::BoxStream<'static, Result<google_ai::GenerateContentResponse>>>,> {- let api_url = self.base_api_url(cx);- let api_key = self.api_key(cx);+ let api_url = self.base_api_url(config, subscription, cx);+ let api_key = self.current_api_key(cx);- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let Some(api_key) = api_key else {return Err(LanguageModelCompletionError::NoApiKey {provider: PROVIDER_NAME,@@ -758,128 +648,98 @@ impl OpenCodeLanguageModel {async move { Ok(future.await?.boxed()) }.boxed()}-}-impl LanguageModel for OpenCodeLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()+ pub fn settings(cx: &App) -> &OpenCodeSettings {+ &crate::AllLanguageModelSettings::get_global(cx).opencode}- fn name(&self) -> LanguageModelName {- LanguageModelName::from(format!(- "{}: {}",- self.subscription.display_name(),- self.model.display_name()- ))+ fn subscription_enabled(subscription: OpenCodeSubscription, cx: &App) -> bool {+ let settings = Self::settings(cx);+ match subscription {+ OpenCodeSubscription::Zen => settings.show_zen_models,+ OpenCodeSubscription::Go => settings.show_go_models,+ }}- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID+ fn api_url(cx: &App) -> SharedString {+ let api_url = &Self::settings(cx).api_url;+ if api_url.is_empty() {+ OPENCODE_API_URL.into()+ } else {+ SharedString::new(api_url.as_str())+ }}+}- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME+impl LanguageModelProviderState for OpenCodeLanguageModelProvider {+ type ObservableEntity = State;++ fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {+ Some(self.state.clone())}+}- fn supports_tools(&self) -> bool {- self.model.supports_tools()+impl LanguageModelProvider for OpenCodeLanguageModelProvider {+ fn id(&self) -> LanguageModelProviderId {+ PROVIDER_ID}- fn supports_images(&self) -> bool {- self.supports_images+ fn name(&self) -> LanguageModelProviderName {+ PROVIDER_NAME}- fn supports_thinking(&self) -> bool {- self.supports_thinking+ fn icon(&self) -> IconOrSvg {+ IconOrSvg::Icon(IconName::AiOpenCode)}- fn supports_disabling_thinking(&self) -> bool {- self.model- .supported_reasoning_effort_levels()- .is_some_and(|levels| levels.contains(&ReasoningEffort::None))+ fn default_model(&self, _cx: &App) -> Option<LanguageModel> {+ None}- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- self.model- .supported_reasoning_effort_levels()- .map(|levels| {- let levels = levels- .iter()- .copied()- .filter(|effort| *effort != ReasoningEffort::None)- .collect::<Vec<_>>();- if levels.is_empty() {- return Vec::new();- }- let default_index = if levels.contains(&ReasoningEffort::Max) {- [- ReasoningEffort::High,- ReasoningEffort::Medium,- ReasoningEffort::Low,- ReasoningEffort::Minimal,- ]- .into_iter()- .find_map(|effort| levels.iter().position(|level| *level == effort))- .unwrap_or(levels.len() - 1)- } else {- levels.len() - 1- };- levels- .into_iter()- .enumerate()- .map(|(i, effort)| {- let (name, value) = reasoning_effort_display(effort);- LanguageModelEffortLevel {- name: name.into(),- value: value.into(),- is_default: i == default_index,- }- })- .collect()- })- .unwrap_or_default()+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ None}- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => true,- LanguageModelToolChoice::None => {- // Google models don't support None tool choice- self.model.protocol() != ApiProtocol::Google- }- }+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.opencode_models(cx)+ .values()+ .map(|(model, subscription)| language_model(model, *subscription))+ .collect()}- fn telemetry_id(&self) -> String {- format!(- "opencode/{}/{}",- self.subscription.id_prefix(),- self.model.id()- )+ fn is_authenticated(&self, cx: &App) -> bool {+ self.state.read(cx).is_authenticated()}- fn max_token_count(&self) -> u64 {- self.model.max_token_count()+ fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {+ self.state.update(cx, |state, cx| state.authenticate(cx))}- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()+ fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {+ let state = self.state.clone();+ Some(ProviderSettingsView::SubPage(+ SubPageProviderSettings::new(move |window, cx| {+ cx.new(|cx| ConfigurationView::new(state.clone(), window, cx))+ .into()+ })+ .description(InlineDescription::Text(+ "To use OpenCode models in Zed, you need an API key.".into(),+ )),+ ))}fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let (config, subscription) = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);let http_client: Arc<dyn HttpClient> = Arc::new(InjectHeaderClient {inner: self.http_client.clone(),name: http::HeaderName::from_static(OPENCODE_SESSION_HEADER_NAME),@@ -887,18 +747,18 @@ impl LanguageModel for OpenCodeLanguageModel {});let extra_headers = self.custom_headers(cx);- match self.model.protocol() {+ match config.model.protocol() {ApiProtocol::Anthropic => {- let mode = if self.supports_thinking() && request.thinking_allowed {+ let mode = if config.supports_thinking && request.thinking_allowed {anthropic::AnthropicModelMode::AdaptiveThinking} else {anthropic::AnthropicModelMode::Default};let anthropic_request = match into_anthropic(request,- self.model.id().to_string(),+ config.model.id().to_string(),1.0,- self.model.max_output_tokens().unwrap_or(8192),+ config.model.max_output_tokens().unwrap_or(8192),mode,anthropic::completion::AnthropicPromptCacheMode::Automatic,&PROVIDER_ID,@@ -906,8 +766,15 @@ impl LanguageModel for OpenCodeLanguageModel {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let stream =- self.stream_anthropic(anthropic_request, http_client, extra_headers, cx);+ let stream = self.stream_anthropic(+ &config,+ subscription,+ &request_limiter,+ anthropic_request,+ http_client,+ extra_headers,+ cx,+ );let executor = cx.background_executor().clone();async move {let mapper = AnthropicEventMapper::new(PROVIDER_NAME, PROVIDER_ID);@@ -929,19 +796,26 @@ impl LanguageModel for OpenCodeLanguageModel {};let openai_request = match into_open_ai(request,- self.model.id(),+ config.model.id(),true,false,- self.model.max_output_tokens(),+ config.model.max_output_tokens(),ChatCompletionMaxTokensParameter::MaxCompletionTokens,reasoning_effort,- self.model.interleaved_reasoning(),+ config.model.interleaved_reasoning(),) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let stream =- self.stream_openai_chat(openai_request, http_client, extra_headers, cx);+ let stream = self.stream_openai_chat(+ &config,+ subscription,+ &request_limiter,+ openai_request,+ http_client,+ extra_headers,+ cx,+ );let executor = cx.background_executor().clone();async move {let mapper = ChatCompletionEventMapper::new();@@ -953,16 +827,16 @@ impl LanguageModel for OpenCodeLanguageModel {.boxed()}ApiProtocol::OpenAiResponses => {- let supports_none_reasoning_effort = self+ let supports_none_reasoning_effort = config.model.supported_reasoning_effort_levels().is_some_and(|levels| levels.contains(&ReasoningEffort::None));let response_request = match into_open_ai_response(request,- self.model.id(),+ config.model.id(),true,false,- self.model.max_output_tokens(),+ config.model.max_output_tokens(),None,supports_none_reasoning_effort,&PROVIDER_ID,@@ -970,8 +844,15 @@ impl LanguageModel for OpenCodeLanguageModel {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let stream =- self.stream_openai_response(response_request, http_client, extra_headers, cx);+ let stream = self.stream_openai_response(+ &config,+ subscription,+ &request_limiter,+ response_request,+ http_client,+ extra_headers,+ cx,+ );let executor = cx.background_executor().clone();async move {let mapper = OpenAiResponseEventMapper::new(PROVIDER_ID);@@ -986,20 +867,29 @@ impl LanguageModel for OpenCodeLanguageModel {let mut request = request;if request.max_output_tokens.is_some() {request.max_output_tokens =- request.effective_max_output_tokens(self.max_output_tokens());+ request.effective_max_output_tokens(config.model.max_output_tokens());}- let mode = if self.supports_thinking() && request.thinking_allowed {+ let mode = if config.supports_thinking && request.thinking_allowed {google_ai::GoogleModelMode::Thinking {budget_tokens: None,}} else {google_ai::GoogleModelMode::Default};- let google_request = match into_google(request, self.model.id().to_string(), mode) {+ let google_request = match into_google(request, config.model.id().to_string(), mode)+ {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};- let stream = self.stream_google(google_request, http_client, extra_headers, cx);+ let stream = self.stream_google(+ &config,+ subscription,+ &request_limiter,+ google_request,+ http_client,+ extra_headers,+ cx,+ );async move {let mapper = GoogleEventMapper::new();Ok(mapper.map_stream(stream.await?.boxed()).boxed())@@ -1010,6 +900,118 @@ impl LanguageModel for OpenCodeLanguageModel {}}+struct InjectHeaderClient {+ inner: Arc<dyn HttpClient>,+ name: http::HeaderName,+ value: http::HeaderValue,+}++impl HttpClient for InjectHeaderClient {+ fn user_agent(&self) -> Option<&http::HeaderValue> {+ self.inner.user_agent()+ }++ fn proxy(&self) -> Option<&http_client::Url> {+ self.inner.proxy()+ }++ fn send(+ &self,+ mut req: http::Request<AsyncBody>,+ ) -> futures::future::BoxFuture<'static, anyhow::Result<http::Response<AsyncBody>>> {+ req.headers_mut()+ .insert(self.name.clone(), self.value.clone());+ self.inner.send(req)+ }+}++// Standalone requests do not have a conversation ID, but OpenCode requires a+// non-empty session header.+fn opencode_session_header_value(thread_id: Option<&str>) -> http::HeaderValue {+ thread_id+ .filter(|thread_id| !thread_id.is_empty())+ .and_then(|thread_id| http::HeaderValue::from_str(thread_id).ok())+ .unwrap_or_else(|| http::HeaderValue::from(rand::random::<u64>()))+}++fn model_id(model: &opencode::Model, subscription: OpenCodeSubscription) -> String {+ format!("{}/{}", subscription.id_prefix(), model.id())+}++fn language_model(model: &DiscoveredModel, subscription: OpenCodeSubscription) -> LanguageModel {+ let DiscoveredModel {+ model,+ supports_images,+ supports_thinking,+ } = model;+ let supported_effort_levels = model+ .supported_reasoning_effort_levels()+ .map(|levels| {+ let levels = levels+ .iter()+ .copied()+ .filter(|effort| *effort != ReasoningEffort::None)+ .collect::<Vec<_>>();+ if levels.is_empty() {+ return Vec::new();+ }+ let default_index = if levels.contains(&ReasoningEffort::Max) {+ [+ ReasoningEffort::High,+ ReasoningEffort::Medium,+ ReasoningEffort::Low,+ ReasoningEffort::Minimal,+ ]+ .into_iter()+ .find_map(|effort| levels.iter().position(|level| *level == effort))+ .unwrap_or(levels.len() - 1)+ } else {+ levels.len() - 1+ };+ levels+ .into_iter()+ .enumerate()+ .map(|(i, effort)| {+ let (name, value) = reasoning_effort_display(effort);+ LanguageModelEffortLevel {+ name: name.into(),+ value: value.into(),+ is_default: i == default_index,+ }+ })+ .collect()+ })+ .unwrap_or_default();+ LanguageModel {+ supports_tools: model.supports_tools(),+ supports_images: *supports_images,+ supports_thinking: *supports_thinking,+ supports_disabling_thinking: model+ .supported_reasoning_effort_levels()+ .is_some_and(|levels| levels.contains(&ReasoningEffort::None)),+ supported_effort_levels: supported_effort_levels.into(),+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: true,+ any: true,+ // Google models don't support None tool choice+ none: model.protocol() != ApiProtocol::Google,+ },+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model_id(model, subscription)),+ LanguageModelName::from(format!(+ "{}: {}",+ subscription.display_name(),+ model.display_name()+ )),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("opencode/{}/{}", subscription.id_prefix(), model.id()),+ model.max_token_count(),+ )+ }+}+#[derive(Deserialize)]struct RichCatalogProvider {npm: Option<String>,@@ -1692,34 +1694,25 @@ mod tests {assert!(!errors.contains_key(&OpenCodeSubscription::Go));}- #[gpui::test]- fn test_default_reasoning_effort(cx: &mut gpui::TestAppContext) {+ #[test]+ fn test_default_reasoning_effort() {use ReasoningEffort::{High, Low, Max, Medium, Minimal, XHigh};- cx.update(|cx| {- let settings_store = SettingsStore::test(cx);- cx.set_global(settings_store);- let provider = OpenCodeLanguageModelProvider::new_with_cache(- FakeHttpClient::with_404_response(),- Arc::new(TestCredentialsProvider),- FakeFs::new(cx.background_executor().clone()),- PathBuf::from("/cache/models.json"),- cx,- );- for (levels, expected) in [- (vec![Low, Medium, High, XHigh, Max], Some("high")),- (vec![Max, High, Medium], Some("high")),- (vec![Low, Medium, Max], Some("medium")),- (vec![Minimal, Low, Max], Some("low")),- (vec![Minimal, Max], Some("minimal")),- (vec![Low, High, XHigh], Some("xhigh")),- (vec![High, Low], Some("low")),- (vec![Max], Some("max")),- (vec![ReasoningEffort::None], None),- (vec![], None),- ] {- let model = provider.create_language_model_with_capabilities(- opencode::Model::new(+ for (levels, expected) in [+ (vec![Low, Medium, High, XHigh, Max], Some("high")),+ (vec![Max, High, Medium], Some("high")),+ (vec![Low, Medium, Max], Some("medium")),+ (vec![Minimal, Low, Max], Some("low")),+ (vec![Minimal, Max], Some("minimal")),+ (vec![Low, High, XHigh], Some("xhigh")),+ (vec![High, Low], Some("low")),+ (vec![Max], Some("max")),+ (vec![ReasoningEffort::None], None),+ (vec![], None),+ ] {+ let model = language_model(+ &DiscoveredModel {+ model: opencode::Model::new("test-model".to_string(),None,1000,@@ -1729,19 +1722,19 @@ mod tests {None,false,),- OpenCodeSubscription::Zen,- false,- true,- );- assert_eq!(- model- .default_effort_level()- .map(|effort| effort.value.to_string())- .as_deref(),- expected,- );- }- });+ supports_images: false,+ supports_thinking: true,+ },+ OpenCodeSubscription::Zen,+ );+ assert_eq!(+ model+ .default_effort_level()+ .map(|effort| effort.value.to_string())+ .as_deref(),+ expected,+ );+ }}#[test]@@ -1959,11 +1952,15 @@ mod tests {state.set_api_key(Some("test-key".to_string()), cx)});store_key.await.unwrap();- let model = provider.create_language_model_with_capabilities(- test_model("test-model", ApiProtocol::Anthropic),+ let model = add_discovered_model(+ &provider,+ DiscoveredModel {+ model: test_model("test-model", ApiProtocol::Anthropic),+ supports_images: true,+ supports_thinking: false,+ },OpenCodeSubscription::Go,- true,- false,+ cx,);let request = LanguageModelRequest {thread_id: None,@@ -1976,8 +1973,8 @@ mod tests {..Default::default()};- let stream = model- .stream_completion(request, &cx.to_async())+ let stream = provider+ .stream_completion(&model, request, &cx.to_async()).await.unwrap();drop(stream);@@ -1998,6 +1995,25 @@ mod tests {.expect("generated session id should be a u64");}+ /// Adds a model to the provider's discovered list, as a successful catalog+ /// fetch would, and returns it.+ fn add_discovered_model(+ provider: &OpenCodeLanguageModelProvider,+ model: DiscoveredModel,+ subscription: OpenCodeSubscription,+ cx: &mut gpui::TestAppContext,+ ) -> LanguageModel {+ let language_model = language_model(&model, subscription);+ provider.state.update(cx, |state, _| {+ state+ .discovered_models+ .entry(subscription)+ .or_default()+ .push(model)+ });+ language_model+ }+fn test_model(name: &str, protocol: ApiProtocol) -> opencode::Model {opencode::Model::new(name.to_string(),diff --git a/crates/language_models/src/provider/vercel_ai_gateway.rs b/crates/language_models/src/provider/vercel_ai_gateway.rsindex b22c7d3b4a4..cc643a04f9b 100644--- a/crates/language_models/src/provider/vercel_ai_gateway.rs+++ b/crates/language_models/src/provider/vercel_ai_gateway.rs@@ -9,10 +9,10 @@ use http_client::{use language_model::chat_completion::{ChatCompletionEventMapper, ResponseStreamEvent};use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,- LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,- LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,- ProviderSettingsView, RateLimiter, env_var,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelId,+ LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,+ LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoiceSupport,+ ModelRateLimiters, ProviderSettingsView, RateLimiter, env_var, unavailable_error,};use serde::Deserialize;pub use settings::OpenAiCompatibleModelCapabilities as ModelCapabilities;@@ -39,6 +39,7 @@ pub struct VercelAiGatewaySettings {pub struct VercelAiGatewayLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -149,7 +150,11 @@ impl VercelAiGatewayLanguageModelProvider {}});- Self { http_client, state }+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ }}fn settings(cx: &App) -> &VercelAiGatewaySettings {@@ -176,14 +181,75 @@ impl VercelAiGatewayLanguageModelProvider {}}- fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {- Arc::new(VercelAiGatewayLanguageModel {- id: LanguageModelId::from(model.name.clone()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })+ /// Every model this provider offers, keyed by id: the default model, then+ /// the fetched list, then settings entries, each overriding the last.+ fn vercel_ai_gateway_models(&self, cx: &App) -> BTreeMap<String, AvailableModel> {+ let mut models = BTreeMap::default();++ let default_model = Self::default_available_model();+ models.insert(default_model.name.clone(), default_model);++ for model in self.state.read(cx).available_models.clone() {+ models.insert(model.name.clone(), model);+ }++ for model in &Self::settings(cx).available_models {+ models.insert(model.name.clone(), model.clone());+ }++ models+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<AvailableModel, LanguageModelCompletionError> {+ self.vercel_ai_gateway_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))+ }++ fn stream_open_ai(+ &self,+ request_limiter: &RateLimiter,+ request: open_ai::Request,+ cx: &AsyncApp,+ ) -> BoxFuture<+ 'static,+ Result<+ futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,+ LanguageModelCompletionError,+ >,+ > {+ let http_client = self.http_client.clone();+ let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {+ let api_url = VercelAiGatewayLanguageModelProvider::api_url(cx);+ let extra_headers = VercelAiGatewayLanguageModelProvider::settings(cx)+ .custom_headers+ .clone();+ (state.api_key_state.key(&api_url), api_url, extra_headers)+ });++ let future = request_limiter.stream(async move {+ let provider = PROVIDER_NAME;+ let Some(api_key) = api_key else {+ return Err(LanguageModelCompletionError::NoApiKey { provider });+ };+ let request = open_ai::stream_completion(+ http_client.as_ref(),+ provider.0.as_str(),+ &api_url,+ &api_key,+ request,+ &extra_headers,+ );+ let response = request.await.map_err(map_open_ai_error)?;+ Ok(response)+ });++ async move { Ok(future.await?.boxed()) }.boxed()}}@@ -208,31 +274,20 @@ impl LanguageModelProvider for VercelAiGatewayLanguageModelProvider {IconOrSvg::Icon(IconName::AiVercel)}- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(Self::default_available_model()))+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.vercel_ai_gateway_models(cx)+ .get(&Self::default_available_model().name)+ .map(language_model)}- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {None}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let mut models = BTreeMap::default();-- let default_model = Self::default_available_model();- models.insert(default_model.name.clone(), default_model);-- for model in self.state.read(cx).available_models.clone() {- models.insert(model.name.clone(), model);- }-- for model in &Self::settings(cx).available_models {- models.insert(model.name.clone(), model.clone());- }-- models- .into_values()- .map(|model| self.create_language_model(model))+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.vercel_ai_gateway_models(cx)+ .values()+ .map(language_model).collect()}@@ -259,55 +314,42 @@ impl LanguageModelProvider for VercelAiGatewayLanguageModelProvider {self.state.update(cx, |state, cx| state.set_api_key(api_key, cx))}-}--pub struct VercelAiGatewayLanguageModel {- id: LanguageModelId,- model: AvailableModel,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}-impl VercelAiGatewayLanguageModel {- fn stream_open_ai(+ fn stream_completion(&self,- request: open_ai::Request,+ model: &LanguageModel,+ request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,- LanguageModelCompletionError,- >,- > {- let http_client = self.http_client.clone();- let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {- let api_url = VercelAiGatewayLanguageModelProvider::api_url(cx);- let extra_headers = VercelAiGatewayLanguageModelProvider::settings(cx)- .custom_headers- .clone();- (state.api_key_state.key(&api_url), api_url, extra_headers)- });-- let future = self.request_limiter.stream(async move {- let provider = PROVIDER_NAME;- let Some(api_key) = api_key else {- return Err(LanguageModelCompletionError::NoApiKey { provider });- };- let request = open_ai::stream_completion(- http_client.as_ref(),- provider.0.as_str(),- &api_url,- &api_key,- request,- &extra_headers,- );- let response = request.await.map_err(map_open_ai_error)?;- Ok(response)- });-- async move { Ok(future.await?.boxed()) }.boxed()+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let request = match crate::provider::open_ai::into_open_ai(+ request,+ &config.name,+ config.capabilities.parallel_tool_calls,+ config.capabilities.prompt_cache_key,+ config.max_output_tokens,+ crate::provider::open_ai::ChatCompletionMaxTokensParameter::MaxCompletionTokens,+ None,+ false,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_open_ai(&request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()}}@@ -375,102 +417,32 @@ fn has_tag(tags: &[String], expected: &str) -> bool {.any(|tag| tag.trim().eq_ignore_ascii_case(expected))}-impl LanguageModel for VercelAiGatewayLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(- self.model- .display_name- .clone()- .unwrap_or_else(|| self.model.name.clone()),+fn language_model(model: &AvailableModel) -> LanguageModel {+ LanguageModel {+ supports_tools: model.capabilities.tools,+ supports_images: model.capabilities.images,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: model.capabilities.tools,+ any: model.capabilities.tools,+ none: true,+ },+ supports_streaming_tools: true,+ supports_split_token_display: true,+ max_output_tokens: model.max_output_tokens,+ ..LanguageModel::new(+ LanguageModelId::from(model.name.clone()),+ LanguageModelName::from(+ model+ .display_name+ .clone()+ .unwrap_or_else(|| model.name.clone()),+ ),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("vercel_ai_gateway/{}", model.name),+ model.max_tokens,)}-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.capabilities.tools- }-- fn supports_images(&self) -> bool {- self.model.capabilities.images- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto => self.model.capabilities.tools,- LanguageModelToolChoice::Any => self.model.capabilities.tools,- LanguageModelToolChoice::None => true,- }- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn telemetry_id(&self) -> String {- format!("vercel_ai_gateway/{}", self.model.name)- }-- fn max_token_count(&self) -> u64 {- self.model.max_tokens- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- let request = match crate::provider::open_ai::into_open_ai(- request,- &self.model.name,- self.model.capabilities.parallel_tool_calls,- self.model.capabilities.prompt_cache_key,- self.max_output_tokens(),- crate::provider::open_ai::ChatCompletionMaxTokensParameter::MaxCompletionTokens,- None,- false,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_open_ai(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))- }- .boxed()- }}#[derive(Deserialize)]diff --git a/crates/language_models/src/provider/x_ai.rs b/crates/language_models/src/provider/x_ai.rsindex 708d05f3830..13c77f1f2d4 100644--- a/crates/language_models/src/provider/x_ai.rs+++ b/crates/language_models/src/provider/x_ai.rs@@ -7,10 +7,11 @@ use http_client::{CustomHeaders, HttpClient};use language_model::chat_completion::{ChatCompletionEventMapper, ResponseStreamEvent};use language_model::{ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,- LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,+ LanguageModelCompletionError, LanguageModelCompletionStream, LanguageModelEffortLevel,LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,- LanguageModelToolChoice, ProviderSettingsView, RateLimiter, env_var,+ LanguageModelToolChoiceSupport, ModelRateLimiters, ProviderSettingsView, RateLimiter, env_var,+ unavailable_error,};pub use settings::XaiAvailableModel as AvailableModel;use settings::{Settings, SettingsStore};@@ -35,6 +36,7 @@ pub struct XAiSettings {pub struct XAiLanguageModelProvider {http_client: Arc<dyn HttpClient>,state: Entity<State>,+ request_limiters: ModelRateLimiters,}pub struct State {@@ -96,17 +98,94 @@ impl XAiLanguageModelProvider {}});- Self { http_client, state }+ Self {+ http_client,+ state,+ request_limiters: ModelRateLimiters::default(),+ }}- fn create_language_model(&self, model: x_ai::Model) -> Arc<dyn LanguageModel> {- Arc::new(XAiLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- model,- state: self.state.clone(),- http_client: self.http_client.clone(),- request_limiter: RateLimiter::new(4),- })+ /// Every model this provider offers, keyed by id: the built-in models,+ /// with settings entries added or overriding built-in ones.+ fn x_ai_models(&self, cx: &App) -> BTreeMap<String, x_ai::Model> {+ let mut models = BTreeMap::default();++ for model in x_ai::Model::iter() {+ if !matches!(model, x_ai::Model::Custom { .. }) {+ models.insert(model.id().to_string(), model);+ }+ }++ for model in &Self::settings(cx).available_models {+ models.insert(+ model.name.clone(),+ x_ai::Model::Custom {+ name: model.name.clone(),+ display_name: model.display_name.clone(),+ max_tokens: model.max_tokens,+ max_output_tokens: model.max_output_tokens,+ max_completion_tokens: model.max_completion_tokens,+ supports_images: model.supports_images,+ supports_tools: model.supports_tools,+ parallel_tool_calls: model.parallel_tool_calls,+ },+ );+ }++ models+ }++ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ cx: &App,+ ) -> Result<x_ai::Model, LanguageModelCompletionError> {+ self.x_ai_models(cx)+ .remove(model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))+ }++ fn stream_chat_completion(+ &self,+ request_limiter: &RateLimiter,+ request: open_ai::Request,+ cx: &AsyncApp,+ ) -> BoxFuture<+ 'static,+ Result<+ futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,+ LanguageModelCompletionError,+ >,+ > {+ let http_client = self.http_client.clone();++ let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {+ let api_url = XAiLanguageModelProvider::api_url(cx);+ let extra_headers = XAiLanguageModelProvider::settings(cx)+ .custom_headers+ .clone();+ (state.api_key_state.key(&api_url), api_url, extra_headers)+ });++ let future = request_limiter.stream(async move {+ let provider = PROVIDER_NAME;+ let Some(api_key) = api_key else {+ return Err(LanguageModelCompletionError::NoApiKey { provider });+ };+ let request = open_ai::stream_completion(+ http_client.as_ref(),+ provider.0.as_str(),+ &api_url,+ &api_key,+ request,+ &extra_headers,+ );+ let response = request.await?;+ Ok(response)+ });++ async move { Ok(future.await?.boxed()) }.boxed()}fn settings(cx: &App) -> &XAiSettings {@@ -144,43 +223,20 @@ impl LanguageModelProvider for XAiLanguageModelProvider {IconOrSvg::Icon(IconName::AiXAi)}- fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(x_ai::Model::default()))+ fn default_model(&self, cx: &App) -> Option<LanguageModel> {+ self.x_ai_models(cx)+ .get(x_ai::Model::default().id())+ .map(language_model)}- fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_language_model(x_ai::Model::default_fast()))+ fn default_fast_model(&self, cx: &App) -> Option<LanguageModel> {+ self.x_ai_models(cx)+ .get(x_ai::Model::default_fast().id())+ .map(language_model)}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {- let mut models = BTreeMap::default();-- for model in x_ai::Model::iter() {- if !matches!(model, x_ai::Model::Custom { .. }) {- models.insert(model.id().to_string(), model);- }- }-- for model in &Self::settings(cx).available_models {- models.insert(- model.name.clone(),- x_ai::Model::Custom {- name: model.name.clone(),- display_name: model.display_name.clone(),- max_tokens: model.max_tokens,- max_output_tokens: model.max_output_tokens,- max_completion_tokens: model.max_completion_tokens,- supports_images: model.supports_images,- supports_tools: model.supports_tools,- parallel_tool_calls: model.parallel_tool_calls,- },- );- }-- models- .into_values()- .map(|model| self.create_language_model(model))- .collect()+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {+ self.x_ai_models(cx).values().map(language_model).collect()}fn is_authenticated(&self, cx: &App) -> bool {@@ -205,56 +261,43 @@ impl LanguageModelProvider for XAiLanguageModelProvider {self.state.update(cx, |state, cx| state.set_api_key(api_key, cx))}-}-pub struct XAiLanguageModel {- id: LanguageModelId,- model: x_ai::Model,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--impl XAiLanguageModel {fn stream_completion(&self,- request: open_ai::Request,+ model: &LanguageModel,+ request: LanguageModelRequest,cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,- LanguageModelCompletionError,- >,- > {- let http_client = self.http_client.clone();-- let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {- let api_url = XAiLanguageModelProvider::api_url(cx);- let extra_headers = XAiLanguageModelProvider::settings(cx)- .custom_headers- .clone();- (state.api_key_state.key(&api_url), api_url, extra_headers)- });-- let future = self.request_limiter.stream(async move {- let provider = PROVIDER_NAME;- let Some(api_key) = api_key else {- return Err(LanguageModelCompletionError::NoApiKey { provider });- };- let request = open_ai::stream_completion(- http_client.as_ref(),- provider.0.as_str(),- &api_url,- &api_key,- request,- &extra_headers,- );- let response = request.await?;- Ok(response)- });-- async move { Ok(future.await?.boxed()) }.boxed()+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match cx.update(|cx| self.config(model, cx)) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ let reasoning_effort = reasoning_effort_for_request(&request, &config);+ let request = match crate::provider::open_ai::into_open_ai(+ request,+ config.id(),+ config.supports_parallel_tool_calls(),+ config.supports_prompt_cache_key(),+ config.max_output_tokens(),+ crate::provider::open_ai::ChatCompletionMaxTokensParameter::MaxCompletionTokens,+ reasoning_effort,+ false,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = self.stream_chat_completion(&request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()}}@@ -340,105 +383,24 @@ fn supported_thinking_effort_levels(model: &x_ai::Model) -> Vec<LanguageModelEff.collect()}-impl LanguageModel for XAiLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.supports_tool()- }-- fn supports_images(&self) -> bool {- self.model.supports_images()- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto- | LanguageModelToolChoice::Any- | LanguageModelToolChoice::None => true,- }- }-- fn supports_thinking(&self) -> bool {- self.model.supports_reasoning_effort()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- supported_thinking_effort_levels(&self.model)- }-- fn telemetry_id(&self) -> String {- format!("x_ai/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- let reasoning_effort = reasoning_effort_for_request(&request, &self.model);- let request = match crate::provider::open_ai::into_open_ai(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- self.max_output_tokens(),- crate::provider::open_ai::ChatCompletionMaxTokensParameter::MaxCompletionTokens,- reasoning_effort,- false,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_completion(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))- }- .boxed()+fn language_model(model: &x_ai::Model) -> LanguageModel {+ LanguageModel {+ supports_tools: model.supports_tool(),+ supports_images: model.supports_images(),+ supports_streaming_tools: true,+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_thinking: model.supports_reasoning_effort(),+ supported_effort_levels: supported_thinking_effort_levels(model).into(),+ max_output_tokens: model.max_output_tokens(),+ supports_split_token_display: true,+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("x_ai/{}", model.id()),+ model.max_token_count(),+ )}}diff --git a/crates/language_models/src/provider/x_ai_subscribed.rs b/crates/language_models/src/provider/x_ai_subscribed.rsindex a9362c3d304..8b99d7c42ef 100644--- a/crates/language_models/src/provider/x_ai_subscribed.rs+++ b/crates/language_models/src/provider/x_ai_subscribed.rs@@ -1,22 +1,26 @@use anyhow::{Result, anyhow};use credentials_provider::CredentialsProvider;-use futures::future::Shared;-use gpui::{App, Context, Entity, SharedString, Task, Window};+use futures::FutureExt as _;+use futures::future::{BoxFuture, Shared};++use gpui::{App, AsyncApp, Context, Entity, SharedString, Task, Window};use http_client::HttpClient;use language_model::{- AuthenticateError, IconOrSvg, InlineDescription, LanguageModel, LanguageModelProvider,- LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,- ProviderSettingsView,+ AuthenticateError, IconOrSvg, InlineDescription, LanguageModel, LanguageModelCompletionError,+ LanguageModelCompletionStream, LanguageModelProvider, LanguageModelProviderId,+ LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest, ModelRateLimiters,+ ProviderSettingsView, unavailable_error,};use std::sync::Arc;use ui::{ConfiguredApiCard, prelude::*};-use x_ai_subscribed::{PROVIDER_ID, PROVIDER_NAME, State, SuperGrokModel, create_language_model};+use x_ai_subscribed::{PROVIDER_ID, PROVIDER_NAME, State, SuperGrokModel, language_model};const SUBSCRIPTION_DESCRIPTION: &str ="Sign in with your SuperGrok subscription to use Grok models in Zed's agent.";pub struct XAiSubscribedProvider {state: Entity<State>,+ request_limiters: ModelRateLimiters,}impl XAiSubscribedProvider {@@ -26,11 +30,21 @@ impl XAiSubscribedProvider {cx: &mut App,) -> Self {let state = cx.new(|cx| State::new(http_client, credentials_provider, cx));- Self { state }+ Self {+ state,+ request_limiters: ModelRateLimiters::default(),+ }}- fn create_model(&self, model: SuperGrokModel, cx: &App) -> Arc<dyn LanguageModel> {- create_language_model(model, &self.state, cx)+ /// The current configuration of `model`, if this provider still offers it.+ fn config(+ &self,+ model: &LanguageModel,+ ) -> Result<SuperGrokModel, LanguageModelCompletionError> {+ SuperGrokModel::all()+ .into_iter()+ .find(|config| config.id() == model.id.0.as_ref())+ .ok_or_else(|| unavailable_error(model))}}@@ -55,25 +69,22 @@ impl LanguageModelProvider for XAiSubscribedProvider {IconOrSvg::Icon(IconName::AiXAi)}- fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_model(SuperGrokModel::Grok47, cx))+ fn default_model(&self, _cx: &App) -> Option<LanguageModel> {+ Some(language_model(&SuperGrokModel::Grok47))}- fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {- Some(self.create_model(SuperGrokModel::GrokBuild01, cx))+ fn default_fast_model(&self, _cx: &App) -> Option<LanguageModel> {+ Some(language_model(&SuperGrokModel::GrokBuild01))}- fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn provided_models(&self, cx: &App) -> Vec<LanguageModel> {if !self.is_authenticated(cx) {return Vec::new();}- SuperGrokModel::all()- .into_iter()- .map(|model| self.create_model(model, cx))- .collect()+ SuperGrokModel::all().iter().map(language_model).collect()}- fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {+ fn recommended_models(&self, cx: &App) -> Vec<LanguageModel> {self.default_model(cx).into_iter().collect()}@@ -148,6 +159,21 @@ impl LanguageModelProvider for XAiSubscribedProvider {Sign in via Settings > AI > LLM Providers to continue.".into()}++ fn stream_completion(+ &self,+ model: &LanguageModel,+ request: LanguageModelRequest,+ cx: &AsyncApp,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match self.config(model) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ x_ai_subscribed::stream_completion(&config, &self.state, &request_limiter, request, cx)+ }}struct ConfigurationView {diff --git a/crates/language_models_cloud/src/language_models_cloud.rs b/crates/language_models_cloud/src/language_models_cloud.rsindex 6aa0fc2c164..0a0dc1ead51 100644--- a/crates/language_models_cloud/src/language_models_cloud.rs+++ b/crates/language_models_cloud/src/language_models_cloud.rs@@ -7,13 +7,11 @@ use cloud_llm_client::{SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, ZED_VERSION_HEADER_NAME,};use futures::{- AsyncBufReadExt, AsyncReadExt as _, FutureExt, Stream, StreamExt,- future::BoxFuture,- io::BufReader,- stream::{self, BoxStream},+ AsyncBufReadExt, AsyncReadExt as _, FutureExt, Stream, StreamExt, future::BoxFuture,+ io::BufReader, stream,};use google_ai::GoogleModelMode;-use gpui::{AppContext, AsyncApp, Context, Task};+use gpui::{App, AppContext, Context, Task};use http_client::http::{HeaderMap, HeaderValue};use http_client::{AsyncBody, HttpClient, HttpClientWithUrl, HttpRequestExt, Method, Response, StatusCode,@@ -21,11 +19,11 @@ use http_client::{use language_model::{ANTHROPIC_PROVIDER_ID, ANTHROPIC_PROVIDER_NAME, CompactionResult, DisabledReason,GOOGLE_PROVIDER_ID, GOOGLE_PROVIDER_NAME, LanguageModel, LanguageModelCompletionError,- LanguageModelCompletionEvent, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,- LanguageModelProviderId, LanguageModelProviderName, LanguageModelRequest,- LanguageModelToolChoice, OPEN_AI_PROVIDER_ID, OPEN_AI_PROVIDER_NAME, ProviderErrorCategory,- RateLimiter, X_AI_PROVIDER_ID, X_AI_PROVIDER_NAME, ZED_CLOUD_PROVIDER_ID,- ZED_CLOUD_PROVIDER_NAME,+ LanguageModelCompletionEvent, LanguageModelCompletionStream, LanguageModelEffortLevel,+ LanguageModelId, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,+ LanguageModelRequest, LanguageModelToolChoiceSupport, ModelRateLimiters, OPEN_AI_PROVIDER_ID,+ OPEN_AI_PROVIDER_NAME, ProviderErrorCategory, RateLimiter, X_AI_PROVIDER_ID,+ X_AI_PROVIDER_NAME, ZED_CLOUD_PROVIDER_ID, ZED_CLOUD_PROVIDER_NAME, unavailable_error,};use schemars::JsonSchema;@@ -106,21 +104,12 @@ impl From<ModelMode> for AnthropicModelMode {}}-pub struct CloudLanguageModel<TP: CloudLlmTokenProvider> {- pub id: LanguageModelId,- pub model: Arc<cloud_llm_client::LanguageModel>,- pub token_provider: Arc<TP>,- pub http_client: Arc<HttpClientWithUrl>,- pub app_version: Option<Version>,- pub request_limiter: RateLimiter,-}-pub struct PerformLlmCompletionResponse {pub response: Response<AsyncBody>,pub includes_status_messages: bool,}-impl<TP: CloudLlmTokenProvider> CloudLanguageModel<TP> {+impl<TP: CloudLlmTokenProvider> CloudModelProvider<TP> {pub async fn perform_llm_completion(http_client: &HttpClientWithUrl,token_provider: &TP,@@ -247,148 +236,69 @@ impl<TP: CloudLlmTokenProvider> CloudLanguageModel<TP> {}}-impl<TP: CloudLlmTokenProvider + 'static> CloudLanguageModel<TP> {- fn check_data_retention_consent(- &self,- cx: &AsyncApp,- ) -> Result<(), LanguageModelCompletionError> {- if self.requires_data_retention() && !self.token_provider.has_data_retention_consent(cx) {- return Err(LanguageModelCompletionError::DataRetentionConsentRequired {- model_name: self.model.display_name.clone(),- });- }- Ok(())- }-- fn anthropic_request(&self, request: LanguageModelRequest) -> Result<anthropic::Request> {- let enable_thinking = request.thinking_allowed && self.model.supports_thinking;- let effort = request- .thinking_effort- .as_ref()- .and_then(|effort| anthropic::Effort::from_str(effort).ok());- let mut request = into_anthropic(- request,- self.model.id.to_string(),- 1.0,- self.model.max_output_tokens as u64,- if enable_thinking && effort.is_some() {- AnthropicModelMode::AdaptiveThinking- } else if enable_thinking {- AnthropicModelMode::Thinking {- budget_tokens: Some(4_096),- }- } else {- AnthropicModelMode::Default- },- AnthropicPromptCacheMode::Automatic,- &ANTHROPIC_PROVIDER_ID,- )?;- if !self.model.supports_fast_mode {- request.speed = None;- }- Ok(request)- }-- fn open_ai_request(- &self,- request: LanguageModelRequest,- ) -> Result<open_ai::responses::Request> {- let enable_thinking = request.thinking_allowed && self.model.supports_thinking;- let effort = request- .thinking_effort- .as_ref()- .and_then(|effort| open_ai::ReasoningEffort::from_str(effort).ok())- .filter(|effort| *effort != open_ai::ReasoningEffort::None);- let supports_none_reasoning_effort =- self.model.supported_effort_levels.iter().any(|effort| {- open_ai::ReasoningEffort::from_str(&effort.value)- .is_ok_and(|effort| effort == open_ai::ReasoningEffort::None)- });- let mut request = into_open_ai_response(- request,- &self.model.id.0,- self.model.supports_parallel_tool_calls,- true,- None,- None,- supports_none_reasoning_effort,- &OPEN_AI_PROVIDER_ID,- )?;- if enable_thinking && let Some(effort) = effort {- request.reasoning = Some(open_ai::responses::ReasoningConfig {- effort,- summary: Some(open_ai::responses::ReasoningSummaryMode::Auto),- });- }- Ok(request)+fn anthropic_request(+ config: &cloud_llm_client::LanguageModel,+ request: LanguageModelRequest,+) -> Result<anthropic::Request> {+ let enable_thinking = request.thinking_allowed && config.supports_thinking;+ let effort = request+ .thinking_effort+ .as_ref()+ .and_then(|effort| anthropic::Effort::from_str(effort).ok());+ let mut request = into_anthropic(+ request,+ config.id.to_string(),+ 1.0,+ config.max_output_tokens as u64,+ if enable_thinking && effort.is_some() {+ AnthropicModelMode::AdaptiveThinking+ } else if enable_thinking {+ AnthropicModelMode::Thinking {+ budget_tokens: Some(4_096),+ }+ } else {+ AnthropicModelMode::Default+ },+ AnthropicPromptCacheMode::Automatic,+ &ANTHROPIC_PROVIDER_ID,+ )?;+ if !config.supports_fast_mode {+ request.speed = None;}+ Ok(request)+}- fn compact_anthropic(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- let thread_id = request.thread_id.clone();- let prompt_id = request.prompt_id.clone();- let app_version = self.app_version.clone();- let mut request = match into_anthropic(- request,- self.model.id.to_string(),- 1.0,- self.model.max_output_tokens as u64,- AnthropicModelMode::Default,- AnthropicPromptCacheMode::Automatic,- &ANTHROPIC_PROVIDER_ID,- ) {- Ok(request) => request.into_compact_request(),- Err(error) => return async move { Err(error.into()) }.boxed(),- };- if !self.model.supports_fast_mode {- request.speed = None;- }-- let http_client = self.http_client.clone();- let token_provider = self.token_provider.clone();- let auth_context = token_provider.auth_context(cx);- let executor = cx.background_executor().clone();- let future = self.request_limiter.run(async move {- let PerformLlmCompletionResponse {- response,- includes_status_messages,- } = Self::perform_llm_completion(- &http_client,- &*token_provider,- auth_context,- app_version,- CompletionBody {- thread_id,- prompt_id,- provider: cloud_llm_client::LanguageModelProvider::Anthropic,- model: request.model.clone(),- provider_request: serde_json::to_value(&request).map_err(|error| {- LanguageModelCompletionError::SerializeRequest {- provider: ANTHROPIC_PROVIDER_NAME,- error,- }- })?,- },- )- .await?;-- let mut mapper =- AnthropicEventMapper::new(ANTHROPIC_PROVIDER_NAME, ANTHROPIC_PROVIDER_ID);- let stream = map_cloud_completion_events(- Box::pin(response_lines(response, includes_status_messages)),- &ANTHROPIC_PROVIDER_NAME,- move |event| mapper.map_event(event),- );- let stream = language_model::stream_in_background(stream.boxed(), executor);- let (context, usage) =- collect_compaction_result(stream, ANTHROPIC_PROVIDER_NAME).await?;- Ok(CompactionResult { context, usage })+fn open_ai_request(+ config: &cloud_llm_client::LanguageModel,+ request: LanguageModelRequest,+) -> Result<open_ai::responses::Request> {+ let enable_thinking = request.thinking_allowed && config.supports_thinking;+ let effort = request+ .thinking_effort+ .as_ref()+ .and_then(|effort| open_ai::ReasoningEffort::from_str(effort).ok())+ .filter(|effort| *effort != open_ai::ReasoningEffort::None);+ let supports_none_reasoning_effort = config.supported_effort_levels.iter().any(|effort| {+ open_ai::ReasoningEffort::from_str(&effort.value)+ .is_ok_and(|effort| effort == open_ai::ReasoningEffort::None)+ });+ let mut request = into_open_ai_response(+ request,+ &config.id.0,+ config.supports_parallel_tool_calls,+ true,+ None,+ None,+ supports_none_reasoning_effort,+ &OPEN_AI_PROVIDER_ID,+ )?;+ if enable_thinking && let Some(effort) = effort {+ request.reasoning = Some(open_ai::responses::ReasoningConfig {+ effort,+ summary: Some(open_ai::responses::ReasoningSummaryMode::Auto),});- future.boxed()}+ Ok(request)}fn needs_llm_token_refresh(response: &Response<AsyncBody>) -> bool {@@ -492,49 +402,260 @@ impl From<ApiError> for LanguageModelCompletionError {}}-impl<TP: CloudLlmTokenProvider + 'static> CloudLanguageModel<TP> {- /// Explicit compaction via OpenAI's dedicated compact operation, proxied- /// through the cloud's `/completions/compact` endpoint.- fn compact_open_ai(+/// Describes a model from the Zed cloud model list.+pub fn language_model(model: &cloud_llm_client::LanguageModel) -> LanguageModel {+ use cloud_llm_client::LanguageModelProvider::*;+ let supports_explicit_compaction = supports_explicit_compaction(model);+ LanguageModel {+ upstream_provider_id: Some(match model.provider {+ Anthropic => ANTHROPIC_PROVIDER_ID,+ OpenAi => OPEN_AI_PROVIDER_ID,+ Google => GOOGLE_PROVIDER_ID,+ XAi => X_AI_PROVIDER_ID,+ }),+ upstream_provider_name: Some(match model.provider {+ Anthropic => ANTHROPIC_PROVIDER_NAME,+ OpenAi => OPEN_AI_PROVIDER_NAME,+ Google => GOOGLE_PROVIDER_NAME,+ XAi => X_AI_PROVIDER_NAME,+ }),+ is_latest: model.is_latest,+ disabled_reason: if model.is_disabled {+ model.disabled_reason.clone().map(DisabledReason::new)+ } else {+ None+ },+ requires_data_retention: requires_data_retention(model),+ refusal_fallback_model_id: if model.id.0.starts_with(anthropic::FABLE_MODEL_ID_PREFIX) {+ Some(anthropic::FABLE_FALLBACK_MODEL_ID)+ } else {+ None+ },+ supports_tools: model.supports_tools,+ supports_images: model.supports_images,+ supports_thinking: model.supports_thinking,+ supports_disabling_thinking: model.supports_disabling_thinking,+ supports_fast_mode: model.supports_fast_mode,+ supports_server_side_compaction: model.supports_server_side_compaction,+ supports_explicit_compaction,+ supports_explicit_compaction_output_limit: model.provider == Anthropic+ && supports_explicit_compaction,+ minimum_explicit_compaction_input_tokens: (model.provider == Anthropic+ && supports_explicit_compaction)+ .then_some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS),+ supported_effort_levels: model+ .supported_effort_levels+ .iter()+ .map(|effort_level| LanguageModelEffortLevel {+ name: effort_level.name.clone().into(),+ value: effort_level.value.clone().into(),+ is_default: effort_level.is_default.unwrap_or(false),+ })+ .collect(),+ supports_streaming_tools: model.supports_streaming_tools,+ tool_choice_support: LanguageModelToolChoiceSupport {+ auto: true,+ none: true,+ any: model.provider != Anthropic || anthropic::supports_forced_tool_use(&model.id.0),+ },+ supports_split_token_display: matches!(model.provider, OpenAi | XAi),+ max_output_tokens: Some(model.max_output_tokens as u64),+ ..LanguageModel::new(+ LanguageModelId::from(model.id.0.to_string()),+ LanguageModelName::from(model.display_name.clone()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("zed.dev/{}", model.id),+ model.max_token_count as u64,+ )+ }+}++fn requires_data_retention(model: &cloud_llm_client::LanguageModel) -> bool {+ // Anthropic cannot offer Fable models with Zero Data Retention+ model.id.0.starts_with(anthropic::FABLE_MODEL_ID_PREFIX)+}++fn supports_explicit_compaction(model: &cloud_llm_client::LanguageModel) -> bool {+ matches!(+ model.provider,+ cloud_llm_client::LanguageModelProvider::OpenAi+ | cloud_llm_client::LanguageModelProvider::Anthropic+ ) && model.supports_server_side_compaction+}++pub struct CloudModelProvider<TP: CloudLlmTokenProvider> {+ token_provider: Arc<TP>,+ http_client: Arc<HttpClientWithUrl>,+ app_version: Option<Version>,+ models: Vec<Arc<cloud_llm_client::LanguageModel>>,+ default_model: Option<Arc<cloud_llm_client::LanguageModel>>,+ default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,+ recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,+ request_limiters: ModelRateLimiters,+}++impl<TP: CloudLlmTokenProvider + 'static> CloudModelProvider<TP> {+ pub fn new(+ token_provider: Arc<TP>,+ http_client: Arc<HttpClientWithUrl>,+ app_version: Option<Version>,+ ) -> Self {+ Self {+ token_provider,+ http_client,+ app_version,+ models: Vec::new(),+ default_model: None,+ default_fast_model: None,+ recommended_models: Vec::new(),+ request_limiters: ModelRateLimiters::default(),+ }+ }++ pub fn refresh_models(&self, cx: &mut Context<Self>) -> Task<Result<()>> {+ let http_client = self.http_client.clone();+ let token_provider = self.token_provider.clone();+ cx.spawn(async move |this, cx| {+ let auth_context = token_provider.auth_context(cx);+ let response =+ Self::fetch_models_request(&http_client, &*token_provider, auth_context).await?;+ this.update(cx, |this, cx| {+ this.update_models(response);+ cx.notify();+ })+ })+ }++ async fn fetch_models_request(+ http_client: &HttpClientWithUrl,+ token_provider: &TP,+ auth_context: TP::AuthContext,+ ) -> Result<ListModelsResponse> {+ let url = http_client.build_zed_llm_url("/models", &[])?;+ let mut response =+ authenticated_llm_request(http_client, token_provider, auth_context, |token| {+ Ok(http_client::Request::builder()+ .method(Method::GET)+ .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")+ .uri(url.as_ref())+ .header("Authorization", format!("Bearer {token}"))+ .body(AsyncBody::empty())?)+ })+ .await+ .context("failed to send list models request")?;++ if response.status().is_success() {+ let mut body = String::new();+ response.body_mut().read_to_string(&mut body).await?;+ Ok(serde_json::from_str(&body)?)+ } else {+ let mut body = String::new();+ response.body_mut().read_to_string(&mut body).await?;+ anyhow::bail!(+ "error listing models.\nStatus: {:?}\nBody: {body}",+ response.status(),+ );+ }+ }++ pub fn update_models(&mut self, response: ListModelsResponse) {+ let models: Vec<_> = response.models.into_iter().map(Arc::new).collect();++ self.default_model = models+ .iter()+ .find(|model| {+ response+ .default_model+ .as_ref()+ .is_some_and(|default_model_id| &model.id == default_model_id)+ })+ .cloned();+ self.default_fast_model = models+ .iter()+ .find(|model| {+ response+ .default_fast_model+ .as_ref()+ .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)+ })+ .cloned();+ self.recommended_models = response+ .recommended_models+ .iter()+ .filter_map(|id| models.iter().find(|model| &model.id == id))+ .cloned()+ .collect();+ self.models = models;+ }++ pub fn clear_models(&mut self) {+ self.models.clear();+ self.default_model = None;+ self.default_fast_model = None;+ self.recommended_models.clear();+ }++ /// The current configuration of `model`, if the model list still contains it.+ fn config(+ &self,+ model: &LanguageModel,+ ) -> Result<Arc<cloud_llm_client::LanguageModel>, LanguageModelCompletionError> {+ self.models+ .iter()+ .find(|config| config.id.0.as_ref() == model.id.0.as_ref())+ .cloned()+ .ok_or_else(|| unavailable_error(model))+ }++ fn check_data_retention_consent(+ &self,+ config: &cloud_llm_client::LanguageModel,+ cx: &App,+ ) -> Result<(), LanguageModelCompletionError> {+ if requires_data_retention(config) && !self.token_provider.has_data_retention_consent(cx) {+ return Err(LanguageModelCompletionError::DataRetentionConsentRequired {+ model_name: config.display_name.clone(),+ });+ }+ Ok(())+ }++ fn compact_anthropic(&self,+ config: &cloud_llm_client::LanguageModel,+ request_limiter: &RateLimiter,request: LanguageModelRequest,- cx: &AsyncApp,+ cx: &App,) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {let thread_id = request.thread_id.clone();let prompt_id = request.prompt_id.clone();let app_version = self.app_version.clone();- let model_provider = self.model.provider;- let provider_name = provider_name(&self.model.provider);- let supports_none_reasoning_effort =- self.model.supported_effort_levels.iter().any(|effort| {- open_ai::ReasoningEffort::from_str(&effort.value)- .is_ok_and(|effort| effort == open_ai::ReasoningEffort::None)- });- // Cloud proxies to OpenAI's own infrastructure, so the resulting- // compaction state is owned by (and interchangeable with) OpenAI- // proper, not by the cloud transport.- let request = match into_open_ai_response(+ let mut request = match into_anthropic(request,- &self.model.id.0,- self.model.supports_parallel_tool_calls,- true,- None,- None,- supports_none_reasoning_effort,- &OPEN_AI_PROVIDER_ID,+ config.id.to_string(),+ 1.0,+ config.max_output_tokens as u64,+ AnthropicModelMode::Default,+ AnthropicPromptCacheMode::Automatic,+ &ANTHROPIC_PROVIDER_ID,) {- Ok(request) => request,+ Ok(request) => request.into_compact_request(),Err(error) => return async move { Err(error.into()) }.boxed(),};- let compact_request = request.into_compact_request();+ if !config.supports_fast_mode {+ request.speed = None;+ }+let http_client = self.http_client.clone();let token_provider = self.token_provider.clone();let auth_context = token_provider.auth_context(cx);- let future = self.request_limiter.run(async move {+ let executor = cx.background_executor().clone();+ let future = request_limiter.run(async move {let PerformLlmCompletionResponse {response,includes_status_messages,- } = Self::perform_llm_compaction(+ } = Self::perform_llm_completion(&http_client,&*token_provider,auth_context,@@ -542,11 +663,11 @@ impl<TP: CloudLlmTokenProvider + 'static> CloudLanguageModel<TP> {CompletionBody {thread_id,prompt_id,- provider: model_provider,- model: compact_request.model.clone(),- provider_request: serde_json::to_value(compact_request).map_err(|error| {+ provider: cloud_llm_client::LanguageModelProvider::Anthropic,+ model: request.model.clone(),+ provider_request: serde_json::to_value(&request).map_err(|error| {LanguageModelCompletionError::SerializeRequest {- provider: provider_name.clone(),+ provider: ANTHROPIC_PROVIDER_NAME,error,}})?,@@ -554,154 +675,122 @@ impl<TP: CloudLlmTokenProvider + 'static> CloudLanguageModel<TP> {).await?;- let events = response_lines::<open_ai::responses::CompactedResponse>(- response,- includes_status_messages,+ let mut mapper =+ AnthropicEventMapper::new(ANTHROPIC_PROVIDER_NAME, ANTHROPIC_PROVIDER_ID);+ let stream = map_cloud_completion_events(+ Box::pin(response_lines(response, includes_status_messages)),+ &ANTHROPIC_PROVIDER_NAME,+ move |event| mapper.map_event(event),);- futures::pin_mut!(events);- while let Some(event) = events.next().await {- match event.map_err(|error| error.into_completion_error(provider_name.clone()))? {- CompletionEvent::Event(response) => {- let usage = token_usage_from_response_usage(&response.usage);- let context = response- .into_compacted_context(OPEN_AI_PROVIDER_ID)- .map_err(LanguageModelCompletionError::Other)?;- return Ok(CompactionResult { context, usage });- }- CompletionEvent::Status(_) => {}- }- }-- Err(LanguageModelCompletionError::StreamEndedUnexpectedly {- provider: provider_name,- })+ let stream = language_model::stream_in_background(stream.boxed(), executor);+ let (context, usage) =+ collect_compaction_result(stream, ANTHROPIC_PROVIDER_NAME).await?;+ Ok(CompactionResult { context, usage })});future.boxed()}-}--impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<TP> {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name.clone())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn upstream_provider_id(&self) -> LanguageModelProviderId {- use cloud_llm_client::LanguageModelProvider::*;- match self.model.provider {- Anthropic => ANTHROPIC_PROVIDER_ID,- OpenAi => OPEN_AI_PROVIDER_ID,- Google => GOOGLE_PROVIDER_ID,- XAi => X_AI_PROVIDER_ID,- }- }-- fn upstream_provider_name(&self) -> LanguageModelProviderName {- use cloud_llm_client::LanguageModelProvider::*;- match self.model.provider {- Anthropic => ANTHROPIC_PROVIDER_NAME,- OpenAi => OPEN_AI_PROVIDER_NAME,- Google => GOOGLE_PROVIDER_NAME,- XAi => X_AI_PROVIDER_NAME,- }- }-- fn is_latest(&self) -> bool {- self.model.is_latest- }-- fn is_disabled(&self) -> Option<DisabledReason> {- if self.model.is_disabled {- self.model.disabled_reason.clone().map(DisabledReason::new)- } else {- None- }- }-- fn requires_data_retention(&self) -> bool {- // Anthropic cannot offer Fable models with Zero Data Retention- self.id- .0- .as_ref()- .starts_with(anthropic::FABLE_MODEL_ID_PREFIX)- }-- fn refusal_fallback_model_id(&self) -> Option<&'static str> {- if self- .id- .0- .as_ref()- .starts_with(anthropic::FABLE_MODEL_ID_PREFIX)- {- Some(anthropic::FABLE_FALLBACK_MODEL_ID)- } else {- None- }- }-- fn supports_tools(&self) -> bool {- self.model.supports_tools- }-- fn supports_images(&self) -> bool {- self.model.supports_images- }-- fn supports_thinking(&self) -> bool {- self.model.supports_thinking- }-- fn supports_disabling_thinking(&self) -> bool {- self.model.supports_disabling_thinking- }-- fn supports_fast_mode(&self) -> bool {- self.model.supports_fast_mode- }-- fn supports_server_side_compaction(&self) -> bool {- self.model.supports_server_side_compaction- }-- fn supports_explicit_compaction(&self) -> bool {- matches!(- self.model.provider,- cloud_llm_client::LanguageModelProvider::OpenAi- | cloud_llm_client::LanguageModelProvider::Anthropic- ) && self.model.supports_server_side_compaction- }+ /// Explicit compaction via OpenAI's dedicated compact operation, proxied+ /// through the cloud's `/completions/compact` endpoint.+ fn compact_open_ai(+ &self,+ config: &cloud_llm_client::LanguageModel,+ request_limiter: &RateLimiter,+ request: LanguageModelRequest,+ cx: &App,+ ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let thread_id = request.thread_id.clone();+ let prompt_id = request.prompt_id.clone();+ let app_version = self.app_version.clone();+ let model_provider = config.provider;+ let provider_name = provider_name(&config.provider);+ let supports_none_reasoning_effort = config.supported_effort_levels.iter().any(|effort| {+ open_ai::ReasoningEffort::from_str(&effort.value)+ .is_ok_and(|effort| effort == open_ai::ReasoningEffort::None)+ });+ // Cloud proxies to OpenAI's own infrastructure, so the resulting+ // compaction state is owned by (and interchangeable with) OpenAI+ // proper, not by the cloud transport.+ let request = match into_open_ai_response(+ request,+ &config.id.0,+ config.supports_parallel_tool_calls,+ true,+ None,+ None,+ supports_none_reasoning_effort,+ &OPEN_AI_PROVIDER_ID,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let compact_request = request.into_compact_request();+ let http_client = self.http_client.clone();+ let token_provider = self.token_provider.clone();+ let auth_context = token_provider.auth_context(cx);+ let future = request_limiter.run(async move {+ let PerformLlmCompletionResponse {+ response,+ includes_status_messages,+ } = Self::perform_llm_compaction(+ &http_client,+ &*token_provider,+ auth_context,+ app_version,+ CompletionBody {+ thread_id,+ prompt_id,+ provider: model_provider,+ model: compact_request.model.clone(),+ provider_request: serde_json::to_value(compact_request).map_err(|error| {+ LanguageModelCompletionError::SerializeRequest {+ provider: provider_name.clone(),+ error,+ }+ })?,+ },+ )+ .await?;- fn supports_explicit_compaction_output_limit(&self) -> bool {- self.model.provider == cloud_llm_client::LanguageModelProvider::Anthropic- && self.supports_explicit_compaction()- }+ let events = response_lines::<open_ai::responses::CompactedResponse>(+ response,+ includes_status_messages,+ );+ futures::pin_mut!(events);+ while let Some(event) = events.next().await {+ match event.map_err(|error| error.into_completion_error(provider_name.clone()))? {+ CompletionEvent::Event(response) => {+ let usage = token_usage_from_response_usage(&response.usage);+ let context = response+ .into_compacted_context(OPEN_AI_PROVIDER_ID)+ .map_err(LanguageModelCompletionError::Other)?;+ return Ok(CompactionResult { context, usage });+ }+ CompletionEvent::Status(_) => {}+ }+ }- fn minimum_explicit_compaction_input_tokens(&self) -> Option<u64> {- (self.model.provider == cloud_llm_client::LanguageModelProvider::Anthropic- && self.supports_explicit_compaction())- .then_some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS)+ Err(LanguageModelCompletionError::StreamEndedUnexpectedly {+ provider: provider_name,+ })+ });+ future.boxed()}-- fn compact(+ /// Compacts `request`'s conversation with the listed model `model`.+ pub fn compact(&self,+ model: &LanguageModel,request: LanguageModelRequest,- cx: &AsyncApp,+ cx: &App,) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- if let Err(error) = self.check_data_retention_consent(cx) {+ let config = match self.config(model) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if let Err(error) = self.check_data_retention_consent(&config, cx) {return async move { Err(error) }.boxed();}- if !self.supports_explicit_compaction() {+ if !supports_explicit_compaction(&config) {return async {Err(LanguageModelCompletionError::Other(anyhow::anyhow!("this cloud model does not support explicit compaction"@@ -710,10 +799,12 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<T.boxed();}- match self.model.provider {- cloud_llm_client::LanguageModelProvider::OpenAi => self.compact_open_ai(request, cx),+ match config.provider {+ cloud_llm_client::LanguageModelProvider::OpenAi => {+ self.compact_open_ai(&config, &request_limiter, request, cx)+ }cloud_llm_client::LanguageModelProvider::Anthropic => {- self.compact_anthropic(request, cx)+ self.compact_anthropic(&config, &request_limiter, request, cx)}// Unreachable while the `supports_explicit_compaction` guard// above holds, but a provider mismatch should degrade to the@@ -728,64 +819,27 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<T}}- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- self.model- .supported_effort_levels- .iter()- .map(|effort_level| LanguageModelEffortLevel {- name: effort_level.name.clone().into(),- value: effort_level.value.clone().into(),- is_default: effort_level.is_default.unwrap_or(false),- })- .collect()- }-- fn supports_streaming_tools(&self) -> bool {- self.model.supports_streaming_tools- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto | LanguageModelToolChoice::None => true,- LanguageModelToolChoice::Any => {- self.model.provider != cloud_llm_client::LanguageModelProvider::Anthropic- || anthropic::supports_forced_tool_use(self.id.0.as_ref())- }- }- }-- fn supports_split_token_display(&self) -> bool {- use cloud_llm_client::LanguageModelProvider::*;- matches!(self.model.provider, OpenAi | XAi)- }-- fn telemetry_id(&self) -> String {- format!("zed.dev/{}", self.model.id)- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count as u64- }-- fn max_output_tokens(&self) -> Option<u64> {- Some(self.model.max_output_tokens as u64)- }-- fn count_input_tokens(+ /// Counts `request`'s input tokens for the listed model `model`, or+ /// returns `None` when its upstream provider cannot count them.+ pub fn count_input_tokens(&self,+ model: &LanguageModel,request: LanguageModelRequest,- cx: &AsyncApp,+ cx: &App,) -> BoxFuture<'static, Result<Option<u64>, LanguageModelCompletionError>> {- if let Err(error) = self.check_data_retention_consent(cx) {+ let config = match self.config(model) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if let Err(error) = self.check_data_retention_consent(&config, cx) {return async move { Err(error) }.boxed();}use cloud_llm_client::LanguageModelProvider;- let provider_request = match self.model.provider {- LanguageModelProvider::Anthropic => self- .anthropic_request(request)+ let provider_request = match config.provider {+ LanguageModelProvider::Anthropic => anthropic_request(&config, request).and_then(|request| Ok(serde_json::to_value(request)?)),- LanguageModelProvider::OpenAi => self- .open_ai_request(request)+ LanguageModelProvider::OpenAi => open_ai_request(&config, request).and_then(|request| Ok(serde_json::to_value(request)?)),_ => return async { Ok(None) }.boxed(),};@@ -793,8 +847,8 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<TOk(provider_request) => CompletionBody {thread_id: None,prompt_id: None,- provider: self.model.provider,- model: self.model.id.to_string(),+ provider: config.provider,+ model: config.id.to_string(),provider_request,},Err(error) => return async move { Err(error.into()) }.boxed(),@@ -803,7 +857,7 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<Tlet token_provider = self.token_provider.clone();let auth_context = token_provider.auth_context(cx);let app_version = self.app_version.clone();- self.request_limiter+ request_limiter.run(async move {let mut response = Self::perform_llm_request("/count_tokens",@@ -840,33 +894,35 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<T.boxed()}- fn stream_completion(+ /// Streams a completion of `request` from the listed model `model`.+ pub fn stream_completion(&self,+ model: &LanguageModel,request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,- LanguageModelCompletionError,- >,- > {- if let Err(error) = self.check_data_retention_consent(cx) {+ cx: &App,+ ) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>>+ {+ let config = match self.config(model) {+ Ok(config) => config,+ Err(error) => return async move { Err(error) }.boxed(),+ };+ let request_limiter = self.request_limiters.for_model(&model.id);+ if let Err(error) = self.check_data_retention_consent(&config, cx) {return async move { Err(error) }.boxed();}let mut request = request;if request.max_output_tokens.is_some() {request.max_output_tokens =- request.effective_max_output_tokens(self.max_output_tokens());+ request.effective_max_output_tokens(Some(config.max_output_tokens as u64));}let thread_id = request.thread_id.clone();let prompt_id = request.prompt_id.clone();let app_version = self.app_version.clone();- let provider_name = provider_name(&self.model.provider);- match self.model.provider {+ let provider_name = provider_name(&config.provider);+ match config.provider {cloud_llm_client::LanguageModelProvider::Anthropic => {- let request = match self.anthropic_request(request) {+ let request = match anthropic_request(&config, request) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};@@ -875,7 +931,7 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<Tlet token_provider = self.token_provider.clone();let auth_context = token_provider.auth_context(cx);let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let PerformLlmCompletionResponse {response,includes_status_messages,@@ -916,14 +972,14 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<Tcloud_llm_client::LanguageModelProvider::OpenAi => {let http_client = self.http_client.clone();let token_provider = self.token_provider.clone();- let request = match self.open_ai_request(request) {+ let request = match open_ai_request(&config, request) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};let auth_context = token_provider.auth_context(cx);let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let PerformLlmCompletionResponse {response,includes_status_messages,@@ -965,8 +1021,8 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<Tlet token_provider = self.token_provider.clone();let request = match into_open_ai(request,- &self.model.id.0,- self.model.supports_parallel_tool_calls,+ &config.id.0,+ config.supports_parallel_tool_calls,false,None,ChatCompletionMaxTokensParameter::MaxCompletionTokens,@@ -978,7 +1034,7 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<T};let auth_context = token_provider.auth_context(cx);let executor = cx.background_executor().clone();- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let PerformLlmCompletionResponse {response,includes_status_messages,@@ -1019,13 +1075,12 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<Tlet http_client = self.http_client.clone();let token_provider = self.token_provider.clone();let request =- match into_google(request, self.model.id.to_string(), GoogleModelMode::Default)- {+ match into_google(request, config.id.to_string(), GoogleModelMode::Default) {Ok(request) => request,Err(error) => return async move { Err(error.into()) }.boxed(),};let auth_context = token_provider.auth_context(cx);- let future = self.request_limiter.stream(async move {+ let future = request_limiter.stream(async move {let PerformLlmCompletionResponse {response,includes_status_messages,@@ -1060,131 +1115,6 @@ impl<TP: CloudLlmTokenProvider + 'static> LanguageModel for CloudLanguageModel<T}}}-}--pub struct CloudModelProvider<TP: CloudLlmTokenProvider> {- token_provider: Arc<TP>,- http_client: Arc<HttpClientWithUrl>,- app_version: Option<Version>,- models: Vec<Arc<cloud_llm_client::LanguageModel>>,- default_model: Option<Arc<cloud_llm_client::LanguageModel>>,- default_fast_model: Option<Arc<cloud_llm_client::LanguageModel>>,- recommended_models: Vec<Arc<cloud_llm_client::LanguageModel>>,-}--impl<TP: CloudLlmTokenProvider + 'static> CloudModelProvider<TP> {- pub fn new(- token_provider: Arc<TP>,- http_client: Arc<HttpClientWithUrl>,- app_version: Option<Version>,- ) -> Self {- Self {- token_provider,- http_client,- app_version,- models: Vec::new(),- default_model: None,- default_fast_model: None,- recommended_models: Vec::new(),- }- }-- pub fn refresh_models(&self, cx: &mut Context<Self>) -> Task<Result<()>> {- let http_client = self.http_client.clone();- let token_provider = self.token_provider.clone();- cx.spawn(async move |this, cx| {- let auth_context = token_provider.auth_context(cx);- let response =- Self::fetch_models_request(&http_client, &*token_provider, auth_context).await?;- this.update(cx, |this, cx| {- this.update_models(response);- cx.notify();- })- })- }-- async fn fetch_models_request(- http_client: &HttpClientWithUrl,- token_provider: &TP,- auth_context: TP::AuthContext,- ) -> Result<ListModelsResponse> {- let url = http_client.build_zed_llm_url("/models", &[])?;- let mut response =- authenticated_llm_request(http_client, token_provider, auth_context, |token| {- Ok(http_client::Request::builder()- .method(Method::GET)- .header(CLIENT_SUPPORTS_X_AI_HEADER_NAME, "true")- .uri(url.as_ref())- .header("Authorization", format!("Bearer {token}"))- .body(AsyncBody::empty())?)- })- .await- .context("failed to send list models request")?;-- if response.status().is_success() {- let mut body = String::new();- response.body_mut().read_to_string(&mut body).await?;- Ok(serde_json::from_str(&body)?)- } else {- let mut body = String::new();- response.body_mut().read_to_string(&mut body).await?;- anyhow::bail!(- "error listing models.\nStatus: {:?}\nBody: {body}",- response.status(),- );- }- }-- pub fn update_models(&mut self, response: ListModelsResponse) {- let models: Vec<_> = response.models.into_iter().map(Arc::new).collect();-- self.default_model = models- .iter()- .find(|model| {- response- .default_model- .as_ref()- .is_some_and(|default_model_id| &model.id == default_model_id)- })- .cloned();- self.default_fast_model = models- .iter()- .find(|model| {- response- .default_fast_model- .as_ref()- .is_some_and(|default_fast_model_id| &model.id == default_fast_model_id)- })- .cloned();- self.recommended_models = response- .recommended_models- .iter()- .filter_map(|id| models.iter().find(|model| &model.id == id))- .cloned()- .collect();- self.models = models;- }-- pub fn clear_models(&mut self) {- self.models.clear();- self.default_model = None;- self.default_fast_model = None;- self.recommended_models.clear();- }-- pub fn create_model(- &self,- model: &Arc<cloud_llm_client::LanguageModel>,- ) -> Arc<dyn LanguageModel> {- Arc::new(CloudLanguageModel::<TP> {- id: LanguageModelId::from(model.id.0.to_string()),- model: model.clone(),- token_provider: self.token_provider.clone(),- http_client: self.http_client.clone(),- app_version: self.app_version.clone(),- request_limiter: RateLimiter::new(4),- })- }pub fn models(&self) -> &[Arc<cloud_llm_client::LanguageModel>] {&self.models@@ -1207,7 +1137,7 @@ pub fn map_cloud_completion_events<T, F>(stream: Pin<Box<dyn Stream<Item = Result<CompletionEvent<T>, ResponseStreamError>> + Send>>,provider: &LanguageModelProviderName,mut map_callback: F,-) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>+) -> LanguageModelCompletionStreamwhereT: DeserializeOwned + 'static,F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>@@ -1350,6 +1280,7 @@ mod tests {};use serde_json::json;use std::sync::Mutex;+ use std::sync::atomic::{AtomicBool, Ordering};#[gpui::test]async fn cloud_google_completion_emits_stop(cx: &mut gpui::TestAppContext) {@@ -1358,7 +1289,7 @@ mod tests {("MAX_TOKENS", StopReason::MaxTokens),("SAFETY", StopReason::Refusal),] {- let model = cloud_google_test_model(vec![+ let (provider, model) = cloud_google_test_model(vec![json!({"event": {"candidates": [{"content": {"role": "model", "parts": [{"text": "Hello"}]}}]}}),@@ -1371,8 +1302,10 @@ mod tests {}}),json!({"status": "stream_ended"}),]);- let mut stream = model- .stream_completion(LanguageModelRequest::default(), &cx.to_async())+ let mut stream = cx+ .update(|cx| {+ provider.stream_completion(&model, LanguageModelRequest::default(), cx)+ }).await.unwrap();@@ -1411,9 +1344,11 @@ mod tests {if stream_ended {events.push(json!({"status": "stream_ended"}));}- let model = cloud_google_test_model(events);- let mut stream = model- .stream_completion(LanguageModelRequest::default(), &cx.to_async())+ let (provider, model) = cloud_google_test_model(events);+ let mut stream = cx+ .update(|cx| {+ provider.stream_completion(&model, LanguageModelRequest::default(), cx)+ }).await.unwrap();@@ -1447,9 +1382,11 @@ mod tests {if stream_ended {events.push(json!({"status": "stream_ended"}));}- let model = cloud_google_test_model(events);- let mut stream = model- .stream_completion(LanguageModelRequest::default(), &cx.to_async())+ let (provider, model) = cloud_google_test_model(events);+ let mut stream = cx+ .update(|cx| {+ provider.stream_completion(&model, LanguageModelRequest::default(), cx)+ }).await.unwrap();@@ -1527,12 +1464,16 @@ mod tests {)))?)}});- let model = cloud_test_model(http_client);- assert!(model.supports_explicit_compaction());- assert!(!model.supports_explicit_compaction_output_limit());+ let language_model = language_model(&cloud_test_config());+ assert!(language_model.supports_explicit_compaction());+ assert!(!language_model.supports_explicit_compaction_output_limit());+ let (provider, model) = cloud_test_model(http_client);let request = compact_test_request();- let result = model.compact(request, &cx.to_async()).await.unwrap();+ let result = cx+ .update(|cx| provider.compact(&model, request, cx))+ .await+ .unwrap();assert_eq!(result.usage,@@ -1675,11 +1616,14 @@ mod tests {.body(AsyncBody::from(format!("{response_lines}\n")))?)}});- let model = cloud_anthropic_test_model(http_client);+ let (provider, model) = cloud_anthropic_test_model(http_client);let mut request = compact_test_request();request.max_output_tokens = Some(8192);- let result = model.compact(request, &cx.to_async()).await.unwrap();+ let result = cx+ .update(|cx| provider.compact(&model, request, cx))+ .await+ .unwrap();assert_eq!(result.usage,@@ -1756,10 +1700,10 @@ mod tests {}))))?)});- let model = cloud_test_model(http_client);+ let (provider, model) = cloud_test_model(http_client);- let error = model- .compact(compact_test_request(), &cx.to_async())+ let error = cx+ .update(|cx| provider.compact(&model, compact_test_request(), cx)).await.unwrap_err();@@ -1957,10 +1901,10 @@ mod tests {async fn cloud_transport_errors_include_hostname(cx: &mut gpui::TestAppContext) {let http_client =FakeHttpClient::create(|_| async move { Err(anyhow::anyhow!("DNS lookup failed")) });- let model = cloud_test_model(http_client);+ let (provider, model) = cloud_test_model(http_client);- let error = model- .compact(compact_test_request(), &cx.to_async())+ let error = cx+ .update(|cx| provider.compact(&model, compact_test_request(), cx)).await.expect_err("request should fail");@@ -2004,17 +1948,43 @@ mod tests {#[test]fn cloud_anthropic_supports_explicit_compaction_after_minimum_input() {- let model = cloud_anthropic_test_model(FakeHttpClient::with_404_response());+ let model = language_model(&cloud_anthropic_test_config());- assert!(model.supports_explicit_compaction());- assert!(model.supports_explicit_compaction_output_limit());- assert_eq!(model.max_total_tokens(), Some(model.max_token_count()));+ assert!(model.supports_explicit_compaction);+ assert!(model.supports_explicit_compaction_output_limit);assert_eq!(model.minimum_explicit_compaction_input_tokens(),Some(anthropic::MIN_COMPACTION_TRIGGER_TOKENS));}+ #[test]+ fn config_resolves_current_model_list_and_rejects_unlisted_models() {+ let listed = language_model(&cloud_test_config());+ let mut provider = test_provider(+ FakeHttpClient::with_404_response(),+ vec![cloud_test_config(), cloud_anthropic_test_config()],+ );+ assert_eq!(provider.config(&listed).unwrap().id.0.as_ref(), "gpt-5.4");++ let mut refreshed = cloud_test_config();+ refreshed.max_output_tokens = 64_000;+ provider.update_models(ListModelsResponse {+ models: vec![refreshed],+ default_model: None,+ default_fast_model: None,+ recommended_models: Vec::new(),+ });+ assert_eq!(provider.config(&listed).unwrap().max_output_tokens, 64_000);++ let unlisted = language_model(&cloud_anthropic_test_config());+ assert!(matches!(+ provider.config(&unlisted),+ Err(LanguageModelCompletionError::ModelUnavailable { provider, model })+ if provider == PROVIDER_NAME && model == unlisted.id+ ));+ }+#[gpui::test]async fn hosted_input_counts_use_authenticated_count_endpoint(cx: &mut gpui::TestAppContext) {for anthropic in [true, false] {@@ -2044,7 +2014,7 @@ mod tests {}}});- let model = if anthropic {+ let (provider, model) = if anthropic {cloud_anthropic_test_model(http_client)} else {cloud_test_model(http_client)@@ -2056,8 +2026,7 @@ mod tests {},));assert_eq!(- model- .count_input_tokens(request, &cx.to_async())+ cx.update(|cx| provider.count_input_tokens(&model, request, cx)).await.unwrap(),Some(731)@@ -2071,13 +2040,13 @@ mod tests {cx: &mut gpui::TestAppContext,) {for (status, body) in [(429, "rate limit"), (200, r#"{"tokens":-1}"#), (200, "{}")] {- let model = cloud_test_model(FakeHttpClient::create(move |_| async move {+ let (provider, model) = cloud_test_model(FakeHttpClient::create(move |_| async move {Ok(Response::builder().status(status).body(AsyncBody::from(body))?)}));- let error = model- .count_input_tokens(compact_test_request(), &cx.to_async())+ let error = cx+ .update(|cx| provider.count_input_tokens(&model, compact_test_request(), cx)).await.unwrap_err();if status == 200 {@@ -2095,13 +2064,16 @@ mod tests {));}}- let mut model = cloud_test_model(FakeHttpClient::create(|_| async {- panic!("unsupported provider must not send a counting request");- }));- Arc::make_mut(&mut model.model).provider = cloud_llm_client::LanguageModelProvider::XAi;+ let mut config = cloud_test_config();+ config.provider = cloud_llm_client::LanguageModelProvider::XAi;+ let (provider, model) = bind_test_model(+ FakeHttpClient::create(|_| async {+ panic!("unsupported provider must not send a counting request");+ }),+ config,+ );assert_eq!(- model- .count_input_tokens(compact_test_request(), &cx.to_async())+ cx.update(|cx| provider.count_input_tokens(&model, compact_test_request(), cx)).await.unwrap(),None@@ -2110,16 +2082,18 @@ mod tests {#[test]fn opus_55_hosted_requests_use_adaptive_thinking_with_prefix_binding() -> Result<()> {- let mut model = cloud_anthropic_test_model(FakeHttpClient::with_404_response());- Arc::make_mut(&mut model.model).id =- cloud_llm_client::LanguageModelId("claude-opus-5-5".into());+ let mut config = cloud_anthropic_test_config();+ config.id = cloud_llm_client::LanguageModelId("claude-opus-5-5".into());for effort in [None, Some("high")] {- let request = model.anthropic_request(LanguageModelRequest {- thinking_allowed: true,- thinking_effort: effort.map(str::to_string),- temperature: Some(0.25),- ..Default::default()- })?;+ let request = anthropic_request(+ &config,+ LanguageModelRequest {+ thinking_allowed: true,+ thinking_effort: effort.map(str::to_string),+ temperature: Some(0.25),+ ..Default::default()+ },+ )?;let body = serde_json::to_value(request)?;assert_eq!(body["thinking"]["type"], "adaptive");assert_eq!(@@ -2137,22 +2111,24 @@ mod tests {cx: &mut gpui::TestAppContext,) {let calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));- let mut model = cloud_anthropic_test_model(FakeHttpClient::create({- let calls = calls.clone();- move |request| {- calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);- async move {- assert_eq!(request.uri().path(), "/count_tokens");- Ok(Response::builder()- .status(200)- .body(AsyncBody::from(r#"{"tokens":731}"#))?)+ let mut config = cloud_anthropic_test_config();+ config.id = cloud_llm_client::LanguageModelId("claude-fable-5-1".into());+ assert!(language_model(&config).requires_data_retention());+ let (provider, model) = bind_test_model(+ FakeHttpClient::create({+ let calls = calls.clone();+ move |request| {+ calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);+ async move {+ assert_eq!(request.uri().path(), "/count_tokens");+ Ok(Response::builder()+ .status(200)+ .body(AsyncBody::from(r#"{"tokens":731}"#))?)+ }}- }- }));- model.id = LanguageModelId::from("claude-fable-5-1".to_string());- Arc::make_mut(&mut model.model).id =- cloud_llm_client::LanguageModelId("claude-fable-5-1".into());- assert!(model.requires_data_retention());+ }),+ config,+ );let mut request = compact_test_request();request.messages[0].content.push(MessageContent::Image(language_model::LanguageModelImage {@@ -2160,29 +2136,28 @@ mod tests {},));assert!(matches!(- model- .count_input_tokens(request.clone(), &cx.to_async())+ cx.update(|cx| provider.count_input_tokens(&model, request.clone(), cx)).await,Err(LanguageModelCompletionError::DataRetentionConsentRequired { .. })));assert!(matches!(- model.compact(request.clone(), &cx.to_async()).await,+ cx.update(|cx| provider.compact(&model, request.clone(), cx))+ .await,Err(LanguageModelCompletionError::DataRetentionConsentRequired { .. })));assert!(matches!(- model- .stream_completion(request.clone(), &cx.to_async())+ cx.update(|cx| provider.stream_completion(&model, request.clone(), cx)).await,Err(LanguageModelCompletionError::DataRetentionConsentRequired { .. })));assert_eq!(calls.load(std::sync::atomic::Ordering::SeqCst), 0);- Arc::get_mut(&mut model.token_provider)- .unwrap()- .data_retention_consent = true;+ provider+ .token_provider+ .data_retention_consent+ .store(true, Ordering::SeqCst);assert_eq!(- model- .count_input_tokens(request, &cx.to_async())+ cx.update(|cx| provider.count_input_tokens(&model, request, cx)).await.unwrap(),Some(731)@@ -2199,38 +2174,39 @@ mod tests {(Some(8192), Some(8192)),(Some(u64::MAX), Some(128_000)),] {- let model = cloud_test_model(FakeHttpClient::create(move |mut request| async move {- assert_eq!(request.uri().path(), "/completions");- let mut body = String::new();- request.body_mut().read_to_string(&mut body).await?;- let body: serde_json::Value = serde_json::from_str(&body)?;- assert_eq!(- body["provider_request"].get("max_output_tokens").cloned(),- expected.map(|value| json!(value))- );- let completed = CompletionEvent::Event(json!({- "type": "response.completed",- "response": {- "id": "response-1", "status": "completed", "output": [],- "usage": {"input_tokens": 1, "output_tokens": 0, "total_tokens": 1}- }+ let (provider, model) =+ cloud_test_model(FakeHttpClient::create(move |mut request| async move {+ assert_eq!(request.uri().path(), "/completions");+ let mut body = String::new();+ request.body_mut().read_to_string(&mut body).await?;+ let body: serde_json::Value = serde_json::from_str(&body)?;+ assert_eq!(+ body["provider_request"].get("max_output_tokens").cloned(),+ expected.map(|value| json!(value))+ );+ let completed = CompletionEvent::Event(json!({+ "type": "response.completed",+ "response": {+ "id": "response-1", "status": "completed", "output": [],+ "usage": {"input_tokens": 1, "output_tokens": 0, "total_tokens": 1}+ }+ }));+ let ended = CompletionEvent::<serde_json::Value>::Status(+ CompletionRequestStatus::StreamEnded,+ );+ Ok(Response::builder()+ .status(200)+ .header(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")+ .body(AsyncBody::from(format!(+ "{}\n{}\n",+ serde_json::to_string(&completed)?,+ serde_json::to_string(&ended)?,+ )))?)}));- let ended = CompletionEvent::<serde_json::Value>::Status(- CompletionRequestStatus::StreamEnded,- );- Ok(Response::builder()- .status(200)- .header(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")- .body(AsyncBody::from(format!(- "{}\n{}\n",- serde_json::to_string(&completed)?,- serde_json::to_string(&ended)?,- )))?)- }));let mut request = compact_test_request();request.max_output_tokens = limit;- let mut stream = model- .stream_completion(request, &cx.to_async())+ let mut stream = cx+ .update(|cx| provider.stream_completion(&model, request, cx)).await.unwrap();while let Some(event) = stream.next().await {@@ -2241,7 +2217,7 @@ mod tests {fn cloud_google_test_model(events: Vec<serde_json::Value>,- ) -> CloudLanguageModel<TestTokenProvider> {+ ) -> (CloudModelProvider<TestTokenProvider>, LanguageModel) {let body = events.into_iter().map(|event| format!("{event}\n"))@@ -2255,12 +2231,10 @@ mod tests {.body(AsyncBody::from(body))?)}});- let mut model = cloud_test_model(http_client);- model.id = LanguageModelId::from("gemini-3.1-pro-preview".to_string());- let metadata = Arc::make_mut(&mut model.model);- metadata.provider = cloud_llm_client::LanguageModelProvider::Google;- metadata.id = cloud_llm_client::LanguageModelId(Arc::from("gemini-3.1-pro-preview"));- model+ let mut config = cloud_test_config();+ config.provider = cloud_llm_client::LanguageModelProvider::Google;+ config.id = cloud_llm_client::LanguageModelId(Arc::from("gemini-3.1-pro-preview"));+ bind_test_model(http_client, config)}fn compact_test_request() -> LanguageModelRequest {@@ -2279,71 +2253,91 @@ mod tests {fn cloud_anthropic_test_model(http_client: Arc<HttpClientWithUrl>,- ) -> CloudLanguageModel<TestTokenProvider> {- CloudLanguageModel {- id: LanguageModelId::from("claude-opus-4-6".to_string()),- model: Arc::new(cloud_llm_client::LanguageModel {- provider: cloud_llm_client::LanguageModelProvider::Anthropic,- id: cloud_llm_client::LanguageModelId(Arc::from("claude-opus-4-6")),- display_name: "Claude Opus 4.6".to_string(),- is_latest: true,- max_token_count: 1_000_000,- max_token_count_in_max_mode: None,- max_output_tokens: 128_000,- supports_tools: true,- supports_images: true,- supports_thinking: true,- supports_disabling_thinking: true,- supports_fast_mode: true,- supports_server_side_compaction: true,- supported_effort_levels: Vec::new(),- supports_streaming_tools: true,- supports_parallel_tool_calls: false,- is_disabled: false,- disabled_reason: None,- }),- token_provider: Arc::new(TestTokenProvider::default()),- http_client,- app_version: None,- request_limiter: RateLimiter::new(4),- }+ ) -> (CloudModelProvider<TestTokenProvider>, LanguageModel) {+ bind_test_model(http_client, cloud_anthropic_test_config())}fn cloud_test_model(http_client: Arc<HttpClientWithUrl>,- ) -> CloudLanguageModel<TestTokenProvider> {- CloudLanguageModel {- id: LanguageModelId::from("gpt-5.4".to_string()),- model: Arc::new(cloud_llm_client::LanguageModel {- provider: cloud_llm_client::LanguageModelProvider::OpenAi,- id: cloud_llm_client::LanguageModelId(Arc::from("gpt-5.4")),- display_name: "GPT-5.4".to_string(),- is_latest: true,- max_token_count: 1_000_000,- max_token_count_in_max_mode: None,- max_output_tokens: 128_000,- supports_tools: true,- supports_images: true,- supports_thinking: true,- supports_disabling_thinking: true,- supports_fast_mode: true,- supports_server_side_compaction: true,- supported_effort_levels: Vec::new(),- supports_streaming_tools: true,- supports_parallel_tool_calls: true,- is_disabled: false,- disabled_reason: None,- }),- token_provider: Arc::new(TestTokenProvider::default()),- http_client,- app_version: None,- request_limiter: RateLimiter::new(4),+ ) -> (CloudModelProvider<TestTokenProvider>, LanguageModel) {+ bind_test_model(http_client, cloud_test_config())+ }++ /// Lists `configs` in a provider's model list, as a successful `/models`+ /// fetch would.+ fn test_provider(+ http_client: Arc<HttpClientWithUrl>,+ configs: Vec<cloud_llm_client::LanguageModel>,+ ) -> CloudModelProvider<TestTokenProvider> {+ let mut provider =+ CloudModelProvider::new(Arc::new(TestTokenProvider::default()), http_client, None);+ provider.update_models(ListModelsResponse {+ models: configs,+ default_model: None,+ default_fast_model: None,+ recommended_models: Vec::new(),+ });+ provider+ }++ /// Lists `config` in a provider, and describes the model it lists.+ fn bind_test_model(+ http_client: Arc<HttpClientWithUrl>,+ config: cloud_llm_client::LanguageModel,+ ) -> (CloudModelProvider<TestTokenProvider>, LanguageModel) {+ let model = language_model(&config);+ (test_provider(http_client, vec![config]), model)+ }++ fn cloud_anthropic_test_config() -> cloud_llm_client::LanguageModel {+ cloud_llm_client::LanguageModel {+ provider: cloud_llm_client::LanguageModelProvider::Anthropic,+ id: cloud_llm_client::LanguageModelId(Arc::from("claude-opus-4-6")),+ display_name: "Claude Opus 4.6".to_string(),+ is_latest: true,+ max_token_count: 1_000_000,+ max_token_count_in_max_mode: None,+ max_output_tokens: 128_000,+ supports_tools: true,+ supports_images: true,+ supports_thinking: true,+ supports_disabling_thinking: true,+ supports_fast_mode: true,+ supports_server_side_compaction: true,+ supported_effort_levels: Vec::new(),+ supports_streaming_tools: true,+ supports_parallel_tool_calls: false,+ is_disabled: false,+ disabled_reason: None,+ }+ }++ fn cloud_test_config() -> cloud_llm_client::LanguageModel {+ cloud_llm_client::LanguageModel {+ provider: cloud_llm_client::LanguageModelProvider::OpenAi,+ id: cloud_llm_client::LanguageModelId(Arc::from("gpt-5.4")),+ display_name: "GPT-5.4".to_string(),+ is_latest: true,+ max_token_count: 1_000_000,+ max_token_count_in_max_mode: None,+ max_output_tokens: 128_000,+ supports_tools: true,+ supports_images: true,+ supports_thinking: true,+ supports_disabling_thinking: true,+ supports_fast_mode: true,+ supports_server_side_compaction: true,+ supported_effort_levels: Vec::new(),+ supports_streaming_tools: true,+ supports_parallel_tool_calls: true,+ is_disabled: false,+ disabled_reason: None,}}#[derive(Default)]struct TestTokenProvider {- data_retention_consent: bool,+ data_retention_consent: AtomicBool,}impl CloudLlmTokenProvider for TestTokenProvider {@@ -2366,7 +2360,7 @@ mod tests {}fn has_data_retention_consent(&self, _cx: &impl AppContext) -> bool {- self.data_retention_consent+ self.data_retention_consent.load(Ordering::SeqCst)}}}diff --git a/crates/openai_subscribed/src/openai_subscribed.rs b/crates/openai_subscribed/src/openai_subscribed.rsindex ca77c61c49d..86ee184018d 100644--- a/crates/openai_subscribed/src/openai_subscribed.rs+++ b/crates/openai_subscribed/src/openai_subscribed.rs@@ -3,15 +3,16 @@ use base64::Engine as _;use base64::engine::general_purpose::URL_SAFE_NO_PAD;use credentials_provider::CredentialsProvider;use futures::{FutureExt, StreamExt, future::BoxFuture, future::Shared};-use gpui::{App, AsyncApp, Context, Entity, SharedString, Task, WeakEntity};+use gpui::{AsyncApp, Context, Entity, SharedString, Task, WeakEntity};use http_client::{AsyncBody, CustomHeaders, HttpClient, Method, Request as HttpRequest, RequestBuilderExt as _,http::{HeaderName, HeaderValue},};use language_model::{CompactionResult, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,- LanguageModelEffortLevel, LanguageModelId, LanguageModelName, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelRequest, LanguageModelToolChoice, RateLimiter,+ LanguageModelCompletionStream, LanguageModelEffortLevel, LanguageModelId, LanguageModelName,+ LanguageModelProviderId, LanguageModelProviderName, LanguageModelRequest,+ LanguageModelToolChoiceSupport, RateLimiter,};use open_ai::{ReasoningEffort,@@ -518,54 +519,28 @@ impl ChatGptModel {}}-/// Creates a [`LanguageModel`] for `model` that authenticates through-/// `state`'s credentials, refreshing them as needed.-pub fn create_language_model(- model: ChatGptModel,- state: &Entity<State>,- cx: &App,-) -> Arc<dyn LanguageModel> {- Arc::new(OpenAiSubscribedLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- http_client: state.read(cx).http_client.clone(),- model,- state: state.clone(),- request_limiter: RateLimiter::new(4),- })-}--struct OpenAiSubscribedLanguageModel {- id: LanguageModelId,- model: ChatGptModel,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- request_limiter: RateLimiter,-}--impl OpenAiSubscribedLanguageModel {- fn codex_responses_request(- &self,- mut request: LanguageModelRequest,- ) -> Result<open_ai::responses::Request> {- if !self.model.supports_priority() {- request.speed = None;- }- let mut responses_request = into_open_ai_response(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- self.model.supports_prompt_cache_key(),- None,- self.model.default_reasoning_effort(),- self.model- .supported_reasoning_efforts()- .contains(&ReasoningEffort::None),- &PROVIDER_ID,- )?;- responses_request.store = Some(false);- responses_request.instructions.get_or_insert_default();- Ok(responses_request)- }+fn codex_responses_request(+ model: &ChatGptModel,+ mut request: LanguageModelRequest,+) -> Result<open_ai::responses::Request> {+ if !model.supports_priority() {+ request.speed = None;+ }+ let mut responses_request = into_open_ai_response(+ request,+ model.id(),+ model.supports_parallel_tool_calls(),+ model.supports_prompt_cache_key(),+ None,+ model.default_reasoning_effort(),+ model+ .supported_reasoning_efforts()+ .contains(&ReasoningEffort::None),+ &PROVIDER_ID,+ )?;+ responses_request.store = Some(false);+ responses_request.instructions.get_or_insert_default();+ Ok(responses_request)}fn codex_extra_headers(@@ -741,223 +716,180 @@ async fn list_models(Ok(models.into_iter().map(ChatGptModel::from).collect())}-impl LanguageModel for OpenAiSubscribedLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- true- }-- fn supports_images(&self) -> bool {- self.model.supports_images()- }-- fn supports_tool_choice(&self, _choice: LanguageModelToolChoice) -> bool {- true- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_thinking(&self) -> bool {- true- }-- fn supports_fast_mode(&self) -> bool {- self.model.supports_priority()- }-- fn supports_server_side_compaction(&self) -> bool {- true- }-- fn supports_explicit_compaction(&self) -> bool {- true+/// Describes a ChatGPT subscription model as a [`LanguageModel`].+pub fn language_model(model: &ChatGptModel) -> LanguageModel {+ let default_effort = model.default_reasoning_effort();+ let supported_effort_levels = model+ .supported_reasoning_efforts()+ .iter()+ .copied()+ .filter_map(|effort| {+ let (name, value) = match effort {+ ReasoningEffort::None => return None,+ ReasoningEffort::Minimal => ("Minimal", "minimal"),+ ReasoningEffort::Low => ("Low", "low"),+ ReasoningEffort::Medium => ("Medium", "medium"),+ ReasoningEffort::High => ("High", "high"),+ ReasoningEffort::XHigh => ("Extra High", "xhigh"),+ ReasoningEffort::Max => ("Max", "max"),+ };++ Some(LanguageModelEffortLevel {+ name: name.into(),+ value: value.into(),+ is_default: Some(effort) == default_effort,+ })+ })+ .collect();+ LanguageModel {+ supports_tools: true,+ supports_images: model.supports_images(),+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_streaming_tools: true,+ supports_thinking: true,+ supports_fast_mode: model.supports_priority(),+ supports_server_side_compaction: true,+ supports_explicit_compaction: true,+ supported_effort_levels,+ max_output_tokens: model.max_output_tokens(),+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("openai-subscribed/{}", model.id()),+ model.max_token_count(),+ )}+}- fn compact(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {- let mut responses_request = match self.codex_responses_request(request) {- Ok(responses_request) => responses_request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- responses_request.context_management = None;- responses_request- .input- .push(ResponseInputItem::CompactionTrigger);-- let state = self.state.downgrade();- let http_client = self.http_client.clone();- let request_limiter = self.request_limiter.clone();-- cx.spawn(async move |cx| {- let creds = get_fresh_credentials(&state, &http_client, cx).await?;- let extra_headers =- codex_extra_headers(&creds, responses_request.prompt_cache_key.as_deref());- let access_token = creds.access_token.clone();- let response_stream = request_limiter- .stream(async move {- stream_response(- http_client.as_ref(),- PROVIDER_NAME.0.as_str(),- CODEX_BASE_URL,- &access_token,- responses_request,- &extra_headers,- )- .await- .map_err(LanguageModelCompletionError::from)- })- .await?;- let mapper = OpenAiResponseEventMapper::new(PROVIDER_ID);- let mut event_stream = language_model::stream_in_background(- mapper.map_stream(response_stream.boxed()).boxed(),- cx.background_executor().clone(),- );- let mut compacted_context = None;- let mut usage = language_model::TokenUsage::default();-- while let Some(event) = event_stream.next().await {- match event? {- LanguageModelCompletionEvent::Compaction(- language_model::CompactionUpdate::Finished(context),- ) => {- if compacted_context.replace(context).is_some() {- return Err(LanguageModelCompletionError::Other(anyhow!(- "ChatGPT subscription compaction returned multiple replacement contexts"- )));- }- }- LanguageModelCompletionEvent::UsageUpdate(updated_usage) => {- usage = updated_usage;+/// Compacts `request`'s conversation with the ChatGPT model `model`, using+/// `state`'s credentials, which are refreshed as needed.+pub fn compact(+ model: &ChatGptModel,+ state: &Entity<State>,+ request_limiter: &RateLimiter,+ request: LanguageModelRequest,+ cx: &AsyncApp,+) -> BoxFuture<'static, Result<CompactionResult, LanguageModelCompletionError>> {+ let mut responses_request = match codex_responses_request(model, request) {+ Ok(responses_request) => responses_request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ responses_request.context_management = None;+ responses_request+ .input+ .push(ResponseInputItem::CompactionTrigger);++ let http_client = state.read_with(cx, |state, _| state.http_client.clone());+ let state = state.downgrade();+ let request_limiter = request_limiter.clone();++ cx.spawn(async move |cx| {+ let creds = get_fresh_credentials(&state, &http_client, cx).await?;+ let extra_headers =+ codex_extra_headers(&creds, responses_request.prompt_cache_key.as_deref());+ let access_token = creds.access_token.clone();+ let response_stream = request_limiter+ .stream(async move {+ stream_response(+ http_client.as_ref(),+ PROVIDER_NAME.0.as_str(),+ CODEX_BASE_URL,+ &access_token,+ responses_request,+ &extra_headers,+ )+ .await+ .map_err(LanguageModelCompletionError::from)+ })+ .await?;+ let mapper = OpenAiResponseEventMapper::new(PROVIDER_ID);+ let mut event_stream = language_model::stream_in_background(+ mapper.map_stream(response_stream.boxed()).boxed(),+ cx.background_executor().clone(),+ );+ let mut compacted_context = None;+ let mut usage = language_model::TokenUsage::default();++ while let Some(event) = event_stream.next().await {+ match event? {+ LanguageModelCompletionEvent::Compaction(+ language_model::CompactionUpdate::Finished(context),+ ) => {+ if compacted_context.replace(context).is_some() {+ return Err(LanguageModelCompletionError::Other(anyhow!(+ "ChatGPT subscription compaction returned multiple replacement contexts"+ )));}- _ => {}}+ LanguageModelCompletionEvent::UsageUpdate(updated_usage) => {+ usage = updated_usage;+ }+ _ => {}}+ }- let context = compacted_context.ok_or_else(|| {- LanguageModelCompletionError::Other(anyhow!(- "ChatGPT subscription compaction returned no replacement context"- ))- })?;- Ok(CompactionResult { context, usage })- })- .boxed()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- let default_effort = self.model.default_reasoning_effort();- self.model- .supported_reasoning_efforts()- .iter()- .copied()- .filter_map(|effort| {- let (name, value) = match effort {- ReasoningEffort::None => return None,- ReasoningEffort::Minimal => ("Minimal", "minimal"),- ReasoningEffort::Low => ("Low", "low"),- ReasoningEffort::Medium => ("Medium", "medium"),- ReasoningEffort::High => ("High", "high"),- ReasoningEffort::XHigh => ("Extra High", "xhigh"),- ReasoningEffort::Max => ("Max", "max"),- };-- Some(LanguageModelEffortLevel {- name: name.into(),- value: value.into(),- is_default: Some(effort) == default_effort,- })- })- .collect()- }-- fn telemetry_id(&self) -> String {- format!("openai-subscribed/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }+ let context = compacted_context.ok_or_else(|| {+ LanguageModelCompletionError::Other(anyhow!(+ "ChatGPT subscription compaction returned no replacement context"+ ))+ })?;+ Ok(CompactionResult { context, usage })+ })+ .boxed()+}- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- let responses_request = match self.codex_responses_request(request) {- Ok(responses_request) => responses_request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };+/// Streams a completion of `request` from the ChatGPT model `model`, using+/// `state`'s credentials, which are refreshed as needed.+pub fn stream_completion(+ model: &ChatGptModel,+ state: &Entity<State>,+ request_limiter: &RateLimiter,+ request: LanguageModelRequest,+ cx: &AsyncApp,+) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>> {+ let responses_request = match codex_responses_request(model, request) {+ Ok(responses_request) => responses_request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };- let state = self.state.downgrade();- let http_client = self.http_client.clone();- let request_limiter = self.request_limiter.clone();- let executor = cx.background_executor().clone();-- let future = cx.spawn(async move |cx| {- let creds = get_fresh_credentials(&state, &http_client, cx).await?;- let extra_headers =- codex_extra_headers(&creds, responses_request.prompt_cache_key.as_deref());-- let access_token = creds.access_token.clone();- request_limiter- .stream(async move {- stream_response(- http_client.as_ref(),- PROVIDER_NAME.0.as_str(),- CODEX_BASE_URL,- &access_token,- responses_request,- &extra_headers,- )- .await- .map_err(LanguageModelCompletionError::from)- })+ let http_client = state.read_with(cx, |state, _| state.http_client.clone());+ let state = state.downgrade();+ let request_limiter = request_limiter.clone();+ let executor = cx.background_executor().clone();++ let future = cx.spawn(async move |cx| {+ let creds = get_fresh_credentials(&state, &http_client, cx).await?;+ let extra_headers =+ codex_extra_headers(&creds, responses_request.prompt_cache_key.as_deref());++ let access_token = creds.access_token.clone();+ request_limiter+ .stream(async move {+ stream_response(+ http_client.as_ref(),+ PROVIDER_NAME.0.as_str(),+ CODEX_BASE_URL,+ &access_token,+ responses_request,+ &extra_headers,+ ).await- });+ .map_err(LanguageModelCompletionError::from)+ })+ .await+ });- async move {- let mapper = OpenAiResponseEventMapper::new(PROVIDER_ID);- Ok(language_model::stream_in_background(- mapper.map_stream(future.await?.boxed()).boxed(),- executor,- ))- }- .boxed()+ async move {+ let mapper = OpenAiResponseEventMapper::new(PROVIDER_ID);+ Ok(language_model::stream_in_background(+ mapper.map_stream(future.await?.boxed()).boxed(),+ executor,+ ))}+ .boxed()}-async fn get_fresh_credentials(state: &WeakEntity<State>,http_client: &Arc<dyn HttpClient>,@@ -1357,6 +1289,7 @@ mod tests {use super::*;use gpui::{AppContext as _, TestAppContext};use http_client::FakeHttpClient;+ use language_model::DEFAULT_MODEL_CONCURRENCY;use parking_lot::Mutex;use std::future::Future;use std::pin::Pin;@@ -2107,9 +2040,10 @@ mod tests {let mut credentials = make_fresh_credentials();credentials.account_id = Some("account-123".to_string());let state = make_state(http, Some(credentials), cx);- let model = cx.read(|cx| create_language_model(ChatGptModel::Gpt55, &state, cx));- assert!(model.supports_server_side_compaction());- assert!(model.supports_explicit_compaction());+ let language_model = language_model(&ChatGptModel::Gpt55);+ assert!(language_model.supports_server_side_compaction());+ assert!(language_model.supports_explicit_compaction());+ let request_limiter = RateLimiter::new(DEFAULT_MODEL_CONCURRENCY);let request = LanguageModelRequest {messages: vec![language_model::LanguageModelRequestMessage {@@ -2123,12 +2057,17 @@ mod tests {..Default::default()};let async_cx = cx.to_async();- let events = model- .stream_completion(request, &async_cx)- .await- .expect("the response stream should start")- .collect::<Vec<_>>()- .await;+ let events = stream_completion(+ &ChatGptModel::Gpt55,+ &state,+ &request_limiter,+ request,+ &async_cx,+ )+ .await+ .expect("the response stream should start")+ .collect::<Vec<_>>()+ .await;assert_eq!(compaction_request_count.load(Ordering::SeqCst), 1);assert!(matches!(@@ -2200,7 +2139,7 @@ mod tests {let http: Arc<dyn HttpClient> = http_client;let state = make_state(http, Some(make_fresh_credentials()), cx);- let model = cx.read(|cx| create_language_model(ChatGptModel::Gpt55, &state, cx));+ let request_limiter = RateLimiter::new(DEFAULT_MODEL_CONCURRENCY);let request = LanguageModelRequest {messages: vec![language_model::LanguageModelRequestMessage {role: language_model::Role::User,@@ -2213,10 +2152,15 @@ mod tests {..Default::default()};- let result = model- .compact(request, &cx.to_async())- .await- .expect("manual compaction should succeed");+ let result = compact(+ &ChatGptModel::Gpt55,+ &state,+ &request_limiter,+ request,+ &cx.to_async(),+ )+ .await+ .expect("manual compaction should succeed");let language_model::CompactedContext::ProviderState(compaction_state) = result.contextelse {panic!("expected provider compaction state");diff --git a/crates/sidebar/src/sidebar.rs b/crates/sidebar/src/sidebar.rsindex 37491eb69e3..8e934d82162 100644--- a/crates/sidebar/src/sidebar.rs+++ b/crates/sidebar/src/sidebar.rs@@ -3841,9 +3841,7 @@ impl Sidebar {}}- let Some(configured_model) =- LanguageModelRegistry::read_global(cx).thread_summary_model(cx)- else {+ let Some(model) = LanguageModelRegistry::read_global(cx).thread_summary_model(cx) else {if let Some(workspace) = self.active_workspace(cx) {Self::show_no_thread_summary_model_toast(workspace, cx);}@@ -3854,7 +3852,6 @@ impl Sidebar {return;}- let model = configured_model.model;let temperature = AgentSettings::temperature_for_model(&model, cx);let thread_store = ThreadStore::global(cx);diff --git a/crates/x_ai_subscribed/src/x_ai_subscribed.rs b/crates/x_ai_subscribed/src/x_ai_subscribed.rsindex a830b130e70..09a3d267f6c 100644--- a/crates/x_ai_subscribed/src/x_ai_subscribed.rs+++ b/crates/x_ai_subscribed/src/x_ai_subscribed.rs@@ -3,13 +3,13 @@ use base64::Engine as _;use base64::engine::general_purpose::URL_SAFE_NO_PAD;use credentials_provider::CredentialsProvider;use futures::{FutureExt, StreamExt, future::BoxFuture, future::Shared};-use gpui::{App, AsyncApp, Context, Entity, SharedString, Task, WeakEntity};+use gpui::{AsyncApp, Context, Entity, SharedString, Task, WeakEntity};use http_client::{AsyncBody, CustomHeaders, HttpClient, Method, Request as HttpRequest};use language_model::chat_completion::ChatCompletionEventMapper;use language_model::{- LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,+ LanguageModel, LanguageModelCompletionError, LanguageModelCompletionStream,LanguageModelEffortLevel, LanguageModelId, LanguageModelName, LanguageModelProviderId,- LanguageModelProviderName, LanguageModelRequest, LanguageModelToolChoice,+ LanguageModelProviderName, LanguageModelRequest, LanguageModelToolChoiceSupport,ProviderErrorCategory, RateLimiter,};use open_ai::{ReasoningEffort, ResponseStreamEvent};@@ -317,32 +317,6 @@ impl SuperGrokModel {}}-pub fn create_language_model(- model: SuperGrokModel,- state: &Entity<State>,- cx: &App,-) -> Arc<dyn LanguageModel> {- Arc::new(SuperGrokLanguageModel {- id: LanguageModelId::from(model.id().to_string()),- http_client: state.read(cx).http_client(),- model,- state: state.clone(),- api_url: XAI_API_URL.into(),- extra_headers: CustomHeaders::default(),- request_limiter: RateLimiter::new(4),- })-}--struct SuperGrokLanguageModel {- id: LanguageModelId,- model: SuperGrokModel,- state: Entity<State>,- http_client: Arc<dyn HttpClient>,- api_url: SharedString,- extra_headers: CustomHeaders,- request_limiter: RateLimiter,-}-fn advertised_reasoning_efforts(model: &SuperGrokModel) -> &'static [ReasoningEffort] {// xAI rejects `reasoning_effort: "none"` on grok-4.5 and newer. Compact and// title requests disable thinking, so we omit the field instead of sending none.@@ -439,151 +413,101 @@ fn map_completion_error(error: LanguageModelCompletionError) -> LanguageModelCom}}-impl SuperGrokLanguageModel {- fn stream_open_ai_completion(- &self,- request: open_ai::Request,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,- LanguageModelCompletionError,- >,- > {- let http_client = self.http_client.clone();- let api_url = self.api_url.clone();- let extra_headers = self.extra_headers.clone();- let state = self.state.downgrade();- let request_limiter = self.request_limiter.clone();-- let future = cx.spawn(async move |cx| {- let credentials = get_fresh_credentials(&state, &http_client, cx).await?;- let access_token = credentials.access_token.clone();- request_limiter- .stream(async move {- open_ai::stream_completion(- http_client.as_ref(),- PROVIDER_NAME.0.as_str(),- api_url.as_ref(),- &access_token,- request,- &extra_headers,- )- .await- .map_err(|error| {- map_completion_error(LanguageModelCompletionError::from(error))- })- })+fn stream_open_ai_completion(+ state: &Entity<State>,+ request_limiter: &RateLimiter,+ request: open_ai::Request,+ cx: &AsyncApp,+) -> BoxFuture<+ 'static,+ Result<+ futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,+ LanguageModelCompletionError,+ >,+> {+ let http_client = state.read_with(cx, |state, _| state.http_client());+ let extra_headers = CustomHeaders::default();+ let state = state.downgrade();+ let request_limiter = request_limiter.clone();++ let future = cx.spawn(async move |cx| {+ let credentials = get_fresh_credentials(&state, &http_client, cx).await?;+ let access_token = credentials.access_token.clone();+ request_limiter+ .stream(async move {+ open_ai::stream_completion(+ http_client.as_ref(),+ PROVIDER_NAME.0.as_str(),+ XAI_API_URL,+ &access_token,+ request,+ &extra_headers,+ ).await- });+ .map_err(|error| map_completion_error(LanguageModelCompletionError::from(error)))+ })+ .await+ });- async move { Ok(future.await?.boxed()) }.boxed()- }+ async move { Ok(future.await?.boxed()) }.boxed()}-impl LanguageModel for SuperGrokLanguageModel {- fn id(&self) -> LanguageModelId {- self.id.clone()- }-- fn name(&self) -> LanguageModelName {- LanguageModelName::from(self.model.display_name().to_string())- }-- fn provider_id(&self) -> LanguageModelProviderId {- PROVIDER_ID- }-- fn provider_name(&self) -> LanguageModelProviderName {- PROVIDER_NAME- }-- fn supports_tools(&self) -> bool {- self.model.supports_tools()- }-- fn supports_images(&self) -> bool {- self.model.supports_images()- }-- fn supports_streaming_tools(&self) -> bool {- true- }-- fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {- match choice {- LanguageModelToolChoice::Auto- | LanguageModelToolChoice::Any- | LanguageModelToolChoice::None => true,- }- }-- fn supports_thinking(&self) -> bool {- self.model.supports_reasoning_effort()- }-- fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {- supported_thinking_effort_levels(&self.model)- }-- fn telemetry_id(&self) -> String {- format!("x_ai_subscribed/{}", self.model.id())- }-- fn max_token_count(&self) -> u64 {- self.model.max_token_count()- }-- fn max_output_tokens(&self) -> Option<u64> {- self.model.max_output_tokens()- }-- fn supports_split_token_display(&self) -> bool {- true- }-- fn stream_completion(- &self,- request: LanguageModelRequest,- cx: &AsyncApp,- ) -> BoxFuture<- 'static,- Result<- futures::stream::BoxStream<- 'static,- Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,- >,- LanguageModelCompletionError,- >,- > {- let reasoning_effort = reasoning_effort_for_request(&request, &self.model);- let request = match open_ai::completion::into_open_ai(- request,- self.model.id(),- self.model.supports_parallel_tool_calls(),- false,- self.max_output_tokens(),- open_ai::completion::ChatCompletionMaxTokensParameter::MaxCompletionTokens,- reasoning_effort,- false,- ) {- Ok(request) => request,- Err(error) => return async move { Err(error.into()) }.boxed(),- };- let completions = self.stream_open_ai_completion(request, cx);- let executor = cx.background_executor().clone();- async move {- let mapper = ChatCompletionEventMapper::new();- Ok(language_model::stream_in_background(- mapper.map_stream(completions.await?).boxed(),- executor,- ))- }- .boxed()+/// Describes a SuperGrok model as a [`LanguageModel`].+pub fn language_model(model: &SuperGrokModel) -> LanguageModel {+ LanguageModel {+ supports_tools: model.supports_tools(),+ supports_images: model.supports_images(),+ supports_streaming_tools: true,+ tool_choice_support: LanguageModelToolChoiceSupport::ALL,+ supports_thinking: model.supports_reasoning_effort(),+ supported_effort_levels: supported_thinking_effort_levels(model).into(),+ max_output_tokens: model.max_output_tokens(),+ supports_split_token_display: true,+ ..LanguageModel::new(+ LanguageModelId::from(model.id().to_string()),+ LanguageModelName::from(model.display_name().to_string()),+ PROVIDER_ID,+ PROVIDER_NAME,+ format!("x_ai_subscribed/{}", model.id()),+ model.max_token_count(),+ )}}+/// Streams a completion of `request` from the SuperGrok model `model`, using+/// `state`'s credentials, which are refreshed as needed.+pub fn stream_completion(+ model: &SuperGrokModel,+ state: &Entity<State>,+ request_limiter: &RateLimiter,+ request: LanguageModelRequest,+ cx: &AsyncApp,+) -> BoxFuture<'static, Result<LanguageModelCompletionStream, LanguageModelCompletionError>> {+ let reasoning_effort = reasoning_effort_for_request(&request, model);+ let request = match open_ai::completion::into_open_ai(+ request,+ model.id(),+ model.supports_parallel_tool_calls(),+ false,+ model.max_output_tokens(),+ open_ai::completion::ChatCompletionMaxTokensParameter::MaxCompletionTokens,+ reasoning_effort,+ false,+ ) {+ Ok(request) => request,+ Err(error) => return async move { Err(error.into()) }.boxed(),+ };+ let completions = stream_open_ai_completion(state, request_limiter, request, cx);+ let executor = cx.background_executor().clone();+ async move {+ let mapper = ChatCompletionEventMapper::new();+ Ok(language_model::stream_in_background(+ mapper.map_stream(completions.await?).boxed(),+ executor,+ ))+ }+ .boxed()+}async fn get_fresh_credentials(state: &WeakEntity<State>,http_client: &Arc<dyn HttpClient>,