tenant.openagents/omega

1use anthropic::completion::{AnthropicEventMapper, AnthropicPromptCacheMode, into_anthropic};
2use anthropic::{AnthropicError, AnthropicModelMode};
3use anyhow::Result;
4use credentials_provider::CredentialsProvider;
5use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};
6use gpui::{App, AppContext, AsyncApp, Entity, Task};
7use http_client::{CustomHeaders, HttpClient};
8use language_model::{
9    AuthenticateError, IconOrSvg, LanguageModel, LanguageModelCompletionError,
10    LanguageModelCompletionEvent, LanguageModelId, LanguageModelName, LanguageModelProvider,
11    LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
12    LanguageModelRequest, LanguageModelToolChoice, ProviderSettingsView, RateLimiter,
13    SubPageProviderSettings,
14};
15use settings::Settings;
16use std::sync::Arc;
17use ui::IconName;
18
19use crate::provider::api_compatible::{
20    ApiCompatibleProviderConfigurationView, ApiCompatibleProviderSettings,
21    ApiCompatibleProviderState,
22};
23
24pub use settings::AnthropicCompatibleAvailableModel as AvailableModel;
25pub use settings::AnthropicCompatibleModelCapabilities as ModelCapabilities;
26
27const API_KEY_PLACEHOLDER: &str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
28
29#[derive(Default, Clone, Debug, PartialEq)]
30pub struct AnthropicCompatibleSettings {
31    pub api_url: String,
32    pub available_models: Vec<AvailableModel>,
33    pub custom_headers: CustomHeaders,
34}
35
36pub struct AnthropicCompatibleLanguageModelProvider {
37    id: LanguageModelProviderId,
38    name: LanguageModelProviderName,
39    http_client: Arc<dyn HttpClient>,
40    state: Entity<State>,
41}
42
43impl ApiCompatibleProviderSettings for AnthropicCompatibleSettings {
44    fn api_url(&self) -> &str {
45        &self.api_url
46    }
47}
48
49pub type State = ApiCompatibleProviderState<AnthropicCompatibleSettings>;
50
51fn available_model_to_anthropic_model(available: &AvailableModel) -> anthropic::Model {
52    let mode = match available.mode.unwrap_or_default() {
53        settings::ModelMode::Default => AnthropicModelMode::Default,
54        settings::ModelMode::Thinking { budget_tokens } => {
55            AnthropicModelMode::Thinking { budget_tokens }
56        }
57        settings::ModelMode::Adaptive => AnthropicModelMode::AdaptiveThinking,
58    };
59    let supports_thinking = matches!(
60        mode,
61        AnthropicModelMode::Thinking { .. } | AnthropicModelMode::AdaptiveThinking
62    );
63    let supports_adaptive_thinking = matches!(mode, AnthropicModelMode::AdaptiveThinking { .. });
64
65    anthropic::Model {
66        display_name: available
67            .display_name
68            .clone()
69            .unwrap_or_else(|| available.name.clone()),
70        id: available.name.clone(),
71        max_input_tokens: available.max_tokens,
72        max_output_tokens: available.max_output_tokens.unwrap_or(4_096),
73        default_temperature: available.default_temperature.unwrap_or(1.0),
74        mode,
75        supports_thinking,
76        supports_adaptive_thinking,
77        supports_images: available.capabilities.images,
78        supports_speed: false,
79        supports_compaction: false,
80        supported_effort_levels: if supports_adaptive_thinking {
81            vec![
82                anthropic::Effort::Low,
83                anthropic::Effort::Medium,
84                anthropic::Effort::High,
85                anthropic::Effort::XHigh,
86                anthropic::Effort::Max,
87            ]
88        } else {
89            Vec::new()
90        },
91        tool_override: available.tool_override.clone(),
92        extra_beta_headers: available.extra_beta_headers.clone(),
93    }
94}
95
96#[cfg(test)]
97mod tests {
98    use super::*;
99
100    fn parse_available_model(json: &str) -> AvailableModel {
101        serde_json::from_str(json).expect("test fixture should parse")
102    }
103
104    #[test]
105    fn adaptive_mode_maps_to_adaptive_thinking_with_all_effort_levels() {
106        let available = parse_available_model(
107            r#"{
108                "name": "claude-opus-4-7",
109                "max_tokens": 1000000,
110                "max_output_tokens": 128000,
111                "mode": { "type": "adaptive" }
112            }"#,
113        );
114        let model = available_model_to_anthropic_model(&available);
115
116        assert_eq!(model.mode, AnthropicModelMode::AdaptiveThinking);
117        assert!(model.supports_thinking);
118        assert!(model.supports_adaptive_thinking);
119        assert_eq!(
120            model.supported_effort_levels,
121            vec![
122                anthropic::Effort::Low,
123                anthropic::Effort::Medium,
124                anthropic::Effort::High,
125                anthropic::Effort::XHigh,
126                anthropic::Effort::Max,
127            ]
128        );
129    }
130
131    #[test]
132    fn thinking_mode_does_not_enable_adaptive() {
133        let available = parse_available_model(
134            r#"{
135                "name": "claude-sonnet-4-5",
136                "max_tokens": 200000,
137                "mode": { "type": "thinking", "budget_tokens": 4096 }
138            }"#,
139        );
140        let model = available_model_to_anthropic_model(&available);
141
142        assert!(matches!(model.mode, AnthropicModelMode::Thinking { .. }));
143        assert!(model.supports_thinking);
144        assert!(!model.supports_adaptive_thinking);
145        assert!(model.supported_effort_levels.is_empty());
146    }
147
148    #[test]
149    fn default_mode_disables_thinking() {
150        let available = parse_available_model(
151            r#"{
152                "name": "claude-3-5-haiku",
153                "max_tokens": 200000
154            }"#,
155        );
156        let model = available_model_to_anthropic_model(&available);
157
158        assert_eq!(model.mode, AnthropicModelMode::Default);
159        assert!(!model.supports_thinking);
160        assert!(!model.supports_adaptive_thinking);
161        assert!(model.supported_effort_levels.is_empty());
162    }
163}
164
165impl AnthropicCompatibleLanguageModelProvider {
166    pub fn new(
167        id: Arc<str>,
168        http_client: Arc<dyn HttpClient>,
169        credentials_provider: Arc<dyn CredentialsProvider>,
170        cx: &mut App,
171    ) -> Self {
172        let state = State::new(
173            id.clone(),
174            credentials_provider,
175            |id, cx| {
176                crate::AllLanguageModelSettings::get_global(cx)
177                    .anthropic_compatible
178                    .get(id)
179            },
180            cx,
181        );
182
183        Self {
184            id: id.clone().into(),
185            name: id.into(),
186            http_client,
187            state,
188        }
189    }
190
191    fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {
192        let capabilities = model.capabilities.clone();
193        // Compatible providers may not support Anthropic's automatic prompt
194        // caching; only request explicit (legacy) cache breakpoints when the
195        // user has opted in via the `prompt_caching` capability.
196        let cache_mode = if capabilities.prompt_caching {
197            AnthropicPromptCacheMode::Legacy
198        } else {
199            AnthropicPromptCacheMode::Disabled
200        };
201        let model = available_model_to_anthropic_model(&model);
202
203        Arc::new(AnthropicCompatibleLanguageModel {
204            id: LanguageModelId::from(model.id.clone()),
205            provider_id: self.id.clone(),
206            provider_name: self.name.clone(),
207            model,
208            capabilities,
209            cache_mode,
210            state: self.state.clone(),
211            http_client: self.http_client.clone(),
212            request_limiter: RateLimiter::new(4),
213        })
214    }
215}
216
217impl LanguageModelProviderState for AnthropicCompatibleLanguageModelProvider {
218    type ObservableEntity = State;
219
220    fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
221        Some(self.state.clone())
222    }
223}
224
225impl LanguageModelProvider for AnthropicCompatibleLanguageModelProvider {
226    fn id(&self) -> LanguageModelProviderId {
227        self.id.clone()
228    }
229
230    fn name(&self) -> LanguageModelProviderName {
231        self.name.clone()
232    }
233
234    fn icon(&self) -> IconOrSvg {
235        IconOrSvg::Icon(IconName::AiAnthropicCompat)
236    }
237
238    fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
239        self.state
240            .read(cx)
241            .settings
242            .available_models
243            .first()
244            .map(|model| self.create_language_model(model.clone()))
245    }
246
247    fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
248        None
249    }
250
251    fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
252        self.state
253            .read(cx)
254            .settings
255            .available_models
256            .iter()
257            .map(|model| self.create_language_model(model.clone()))
258            .collect()
259    }
260
261    fn is_authenticated(&self, cx: &App) -> bool {
262        self.state.read(cx).is_authenticated()
263    }
264
265    fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
266        self.state.update(cx, |state, cx| state.authenticate(cx))
267    }
268
269    fn settings_view(&self, _cx: &mut App) -> Option<ProviderSettingsView> {
270        let state = self.state.clone();
271        Some(ProviderSettingsView::SubPage(SubPageProviderSettings::new(
272            move |window, cx| {
273                cx.new(|cx| {
274                    ApiCompatibleProviderConfigurationView::new(
275                        state.clone(),
276                        "Anthropic",
277                        API_KEY_PLACEHOLDER,
278                        window,
279                        cx,
280                    )
281                })
282                .into()
283            },
284        )))
285    }
286
287    fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {
288        self.state
289            .update(cx, |state, cx| state.set_api_key(api_key, cx))
290    }
291}
292
293pub struct AnthropicCompatibleLanguageModel {
294    id: LanguageModelId,
295    provider_id: LanguageModelProviderId,
296    provider_name: LanguageModelProviderName,
297    model: anthropic::Model,
298    capabilities: ModelCapabilities,
299    cache_mode: AnthropicPromptCacheMode,
300    state: Entity<State>,
301    http_client: Arc<dyn HttpClient>,
302    request_limiter: RateLimiter,
303}
304
305impl AnthropicCompatibleLanguageModel {
306    fn stream_completion(
307        &self,
308        request: anthropic::Request,
309        cx: &AsyncApp,
310    ) -> BoxFuture<
311        'static,
312        Result<
313            BoxStream<'static, Result<anthropic::Event, AnthropicError>>,
314            LanguageModelCompletionError,
315        >,
316    > {
317        let http_client = self.http_client.clone();
318        let provider_name = self.provider_name.clone();
319
320        let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, _cx| {
321            let api_url = state.settings.api_url.clone();
322            (
323                state.api_key_state.key(&api_url),
324                api_url,
325                state.settings.custom_headers.clone(),
326            )
327        });
328
329        let beta_headers = self.model.beta_headers();
330
331        async move {
332            let Some(api_key) = api_key else {
333                return Err(LanguageModelCompletionError::NoApiKey {
334                    provider: provider_name,
335                });
336            };
337
338            let request = anthropic::stream_completion(
339                http_client.as_ref(),
340                &api_url,
341                &api_key,
342                request,
343                beta_headers,
344                &extra_headers,
345            );
346
347            request
348                .await
349                .map_err(|error| anthropic::completion_error_from_anthropic(error, provider_name))
350        }
351        .boxed()
352    }
353}
354
355impl LanguageModel for AnthropicCompatibleLanguageModel {
356    fn id(&self) -> LanguageModelId {
357        self.id.clone()
358    }
359
360    fn name(&self) -> LanguageModelName {
361        LanguageModelName::from(self.model.display_name.clone())
362    }
363
364    fn provider_id(&self) -> LanguageModelProviderId {
365        self.provider_id.clone()
366    }
367
368    fn provider_name(&self) -> LanguageModelProviderName {
369        self.provider_name.clone()
370    }
371
372    fn supports_tools(&self) -> bool {
373        self.capabilities.tools
374    }
375
376    fn supports_images(&self) -> bool {
377        self.capabilities.images
378    }
379
380    fn supports_streaming_tools(&self) -> bool {
381        self.capabilities.tools
382    }
383
384    fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
385        match choice {
386            LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => self.capabilities.tools,
387            LanguageModelToolChoice::None => true,
388        }
389    }
390
391    fn supports_thinking(&self) -> bool {
392        self.model.supports_thinking
393    }
394
395    fn supported_effort_levels(&self) -> Vec<language_model::LanguageModelEffortLevel> {
396        self.model
397            .supported_effort_levels
398            .iter()
399            .map(|effort| {
400                let is_default = matches!(effort, anthropic::Effort::High);
401                let (name, value) = match effort {
402                    anthropic::Effort::Low => ("Low".into(), "low".into()),
403                    anthropic::Effort::Medium => ("Medium".into(), "medium".into()),
404                    anthropic::Effort::High => ("High".into(), "high".into()),
405                    anthropic::Effort::XHigh => ("XHigh".into(), "xhigh".into()),
406                    anthropic::Effort::Max => ("Max".into(), "max".into()),
407                };
408                language_model::LanguageModelEffortLevel {
409                    name,
410                    value,
411                    is_default,
412                }
413            })
414            .collect()
415    }
416
417    fn telemetry_id(&self) -> String {
418        format!("anthropic/{}", self.model.id)
419    }
420
421    fn max_token_count(&self) -> u64 {
422        self.model.max_input_tokens
423    }
424
425    fn max_output_tokens(&self) -> Option<u64> {
426        Some(self.model.max_output_tokens)
427    }
428
429    fn stream_completion(
430        &self,
431        request: LanguageModelRequest,
432        cx: &AsyncApp,
433    ) -> BoxFuture<
434        'static,
435        Result<
436            BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
437            LanguageModelCompletionError,
438        >,
439    > {
440        let has_tools = !request.tools.is_empty();
441        let request_id = self.model.request_id(has_tools).to_string();
442        let mut request = match into_anthropic(
443            request,
444            request_id,
445            self.model.default_temperature,
446            self.model.max_output_tokens,
447            self.model.mode.clone(),
448            self.cache_mode,
449            &self.provider_id,
450        ) {
451            Ok(request) => request,
452            Err(error) => return async move { Err(error.into()) }.boxed(),
453        };
454        if !self.model.supports_speed {
455            request.speed = None;
456        }
457        let completion_request = self.stream_completion(request, cx);
458        let provider_name = self.provider_name.clone();
459        let provider_id = self.provider_id.clone();
460        let future = self.request_limiter.stream(async move {
461            let response = completion_request.await?;
462            Ok(AnthropicEventMapper::new(provider_name, provider_id).map_stream(response))
463        });
464        async move { Ok(future.await?.boxed()) }.boxed()
465    }
466}
467