tenant.openagents/omega

1use anyhow::Result;
2use collections::BTreeMap;
3use credentials_provider::CredentialsProvider;
4use futures::{FutureExt, StreamExt, future::BoxFuture};
5use gpui::{App, AppContext, AsyncApp, Context, Entity, SharedString, Task};
6use http_client::{CustomHeaders, HttpClient};
7use language_model::{
8    ApiKeyConfiguration, ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel,
9    LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelEffortLevel,
10    LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
11    LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
12    LanguageModelToolChoice, LanguageModelToolSchemaFormat, ProviderSettingsView, RateLimiter,
13    env_var,
14};
15use open_ai::ResponseStreamEvent;
16pub use settings::XaiAvailableModel as AvailableModel;
17use settings::{Settings, SettingsStore};
18use std::sync::{Arc, LazyLock};
19use strum::IntoEnumIterator;
20use ui::IconName;
21use x_ai::XAI_API_URL;
22
23const PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("x_ai");
24const PROVIDER_NAME: LanguageModelProviderName = LanguageModelProviderName::new("xAI");
25
26const API_KEY_ENV_VAR_NAME: &str = "XAI_API_KEY";
27static API_KEY_ENV_VAR: LazyLock<EnvVar> = env_var!(API_KEY_ENV_VAR_NAME);
28
29#[derive(Default, Clone, Debug, PartialEq)]
30pub struct XAiSettings {
31    pub api_url: String,
32    pub available_models: Vec<AvailableModel>,
33    pub custom_headers: CustomHeaders,
34}
35
36pub struct XAiLanguageModelProvider {
37    http_client: Arc<dyn HttpClient>,
38    state: Entity<State>,
39}
40
41pub struct State {
42    api_key_state: ApiKeyState,
43    credentials_provider: Arc<dyn CredentialsProvider>,
44}
45
46impl State {
47    fn is_authenticated(&self) -> bool {
48        self.api_key_state.has_key()
49    }
50
51    fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
52        let credentials_provider = self.credentials_provider.clone();
53        let api_url = XAiLanguageModelProvider::api_url(cx);
54        self.api_key_state.store(
55            api_url,
56            api_key,
57            |this| &mut this.api_key_state,
58            credentials_provider,
59            cx,
60        )
61    }
62
63    fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
64        let credentials_provider = self.credentials_provider.clone();
65        let api_url = XAiLanguageModelProvider::api_url(cx);
66        self.api_key_state.load_if_needed(
67            api_url,
68            |this| &mut this.api_key_state,
69            credentials_provider,
70            cx,
71        )
72    }
73}
74
75impl XAiLanguageModelProvider {
76    pub fn new(
77        http_client: Arc<dyn HttpClient>,
78        credentials_provider: Arc<dyn CredentialsProvider>,
79        cx: &mut App,
80    ) -> Self {
81        let state = cx.new(|cx| {
82            cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
83                let credentials_provider = this.credentials_provider.clone();
84                let api_url = Self::api_url(cx);
85                this.api_key_state.handle_url_change(
86                    api_url,
87                    |this| &mut this.api_key_state,
88                    credentials_provider,
89                    cx,
90                );
91                cx.notify();
92            })
93            .detach();
94            State {
95                api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
96                credentials_provider,
97            }
98        });
99
100        Self { http_client, state }
101    }
102
103    fn create_language_model(&self, model: x_ai::Model) -> Arc<dyn LanguageModel> {
104        Arc::new(XAiLanguageModel {
105            id: LanguageModelId::from(model.id().to_string()),
106            model,
107            state: self.state.clone(),
108            http_client: self.http_client.clone(),
109            request_limiter: RateLimiter::new(4),
110        })
111    }
112
113    fn settings(cx: &App) -> &XAiSettings {
114        &crate::AllLanguageModelSettings::get_global(cx).x_ai
115    }
116
117    fn api_url(cx: &App) -> SharedString {
118        let api_url = &Self::settings(cx).api_url;
119        if api_url.is_empty() {
120            XAI_API_URL.into()
121        } else {
122            SharedString::new(api_url.as_str())
123        }
124    }
125}
126
127impl LanguageModelProviderState for XAiLanguageModelProvider {
128    type ObservableEntity = State;
129
130    fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
131        Some(self.state.clone())
132    }
133}
134
135impl LanguageModelProvider for XAiLanguageModelProvider {
136    fn id(&self) -> LanguageModelProviderId {
137        PROVIDER_ID
138    }
139
140    fn name(&self) -> LanguageModelProviderName {
141        PROVIDER_NAME
142    }
143
144    fn icon(&self) -> IconOrSvg {
145        IconOrSvg::Icon(IconName::AiXAi)
146    }
147
148    fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
149        Some(self.create_language_model(x_ai::Model::default()))
150    }
151
152    fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
153        Some(self.create_language_model(x_ai::Model::default_fast()))
154    }
155
156    fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
157        let mut models = BTreeMap::default();
158
159        for model in x_ai::Model::iter() {
160            if !matches!(model, x_ai::Model::Custom { .. }) {
161                models.insert(model.id().to_string(), model);
162            }
163        }
164
165        for model in &Self::settings(cx).available_models {
166            models.insert(
167                model.name.clone(),
168                x_ai::Model::Custom {
169                    name: model.name.clone(),
170                    display_name: model.display_name.clone(),
171                    max_tokens: model.max_tokens,
172                    max_output_tokens: model.max_output_tokens,
173                    max_completion_tokens: model.max_completion_tokens,
174                    supports_images: model.supports_images,
175                    supports_tools: model.supports_tools,
176                    parallel_tool_calls: model.parallel_tool_calls,
177                },
178            );
179        }
180
181        models
182            .into_values()
183            .map(|model| self.create_language_model(model))
184            .collect()
185    }
186
187    fn is_authenticated(&self, cx: &App) -> bool {
188        self.state.read(cx).is_authenticated()
189    }
190
191    fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
192        self.state.update(cx, |state, cx| state.authenticate(cx))
193    }
194
195    fn settings_view(&self, cx: &mut App) -> Option<ProviderSettingsView> {
196        let state = self.state.read(cx);
197        Some(ProviderSettingsView::ApiKey(ApiKeyConfiguration::new(
198            state.api_key_state.has_key(),
199            state.api_key_state.is_from_env_var(),
200            state.api_key_state.env_var_name().clone(),
201            "https://console.x.ai/team/default/api-keys".into(),
202        )))
203    }
204
205    fn set_api_key(&self, api_key: Option<String>, cx: &mut App) -> Task<Result<()>> {
206        self.state
207            .update(cx, |state, cx| state.set_api_key(api_key, cx))
208    }
209}
210
211pub struct XAiLanguageModel {
212    id: LanguageModelId,
213    model: x_ai::Model,
214    state: Entity<State>,
215    http_client: Arc<dyn HttpClient>,
216    request_limiter: RateLimiter,
217}
218
219impl XAiLanguageModel {
220    fn stream_completion(
221        &self,
222        request: open_ai::Request,
223        cx: &AsyncApp,
224    ) -> BoxFuture<
225        'static,
226        Result<
227            futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,
228            LanguageModelCompletionError,
229        >,
230    > {
231        let http_client = self.http_client.clone();
232
233        let (api_key, api_url, extra_headers) = self.state.read_with(cx, |state, cx| {
234            let api_url = XAiLanguageModelProvider::api_url(cx);
235            let extra_headers = XAiLanguageModelProvider::settings(cx)
236                .custom_headers
237                .clone();
238            (state.api_key_state.key(&api_url), api_url, extra_headers)
239        });
240
241        let future = self.request_limiter.stream(async move {
242            let provider = PROVIDER_NAME;
243            let Some(api_key) = api_key else {
244                return Err(LanguageModelCompletionError::NoApiKey { provider });
245            };
246            let request = open_ai::stream_completion(
247                http_client.as_ref(),
248                provider.0.as_str(),
249                &api_url,
250                &api_key,
251                request,
252                &extra_headers,
253            );
254            let response = request.await?;
255            Ok(response)
256        });
257
258        async move { Ok(future.await?.boxed()) }.boxed()
259    }
260}
261
262fn x_ai_reasoning_efforts(model: &x_ai::Model) -> &'static [open_ai::ReasoningEffort] {
263    if model.supports_reasoning_effort() {
264        &[
265            open_ai::ReasoningEffort::None,
266            open_ai::ReasoningEffort::Low,
267            open_ai::ReasoningEffort::Medium,
268            open_ai::ReasoningEffort::High,
269        ]
270    } else {
271        &[]
272    }
273}
274
275fn default_thinking_reasoning_effort(model: &x_ai::Model) -> Option<open_ai::ReasoningEffort> {
276    if model.supports_reasoning_effort() {
277        Some(open_ai::ReasoningEffort::Low)
278    } else {
279        None
280    }
281}
282
283fn reasoning_effort_for_request(
284    request: &LanguageModelRequest,
285    model: &x_ai::Model,
286) -> Option<open_ai::ReasoningEffort> {
287    let supported_efforts = x_ai_reasoning_efforts(model);
288    if supported_efforts.is_empty() {
289        return None;
290    }
291
292    if request.thinking_allowed {
293        request
294            .thinking_effort
295            .as_deref()
296            .and_then(|effort| effort.parse::<open_ai::ReasoningEffort>().ok())
297            .filter(|effort| supported_efforts.contains(effort))
298            .filter(|effort| *effort != open_ai::ReasoningEffort::None)
299            .or_else(|| default_thinking_reasoning_effort(model))
300    } else if supported_efforts.contains(&open_ai::ReasoningEffort::None) {
301        Some(open_ai::ReasoningEffort::None)
302    } else {
303        None
304    }
305}
306
307fn supported_thinking_effort_levels(model: &x_ai::Model) -> Vec<LanguageModelEffortLevel> {
308    let default_effort = default_thinking_reasoning_effort(model);
309    x_ai_reasoning_efforts(model)
310        .iter()
311        .copied()
312        .filter_map(|effort| {
313            let (name, value) = match effort {
314                open_ai::ReasoningEffort::None => return None,
315                open_ai::ReasoningEffort::Minimal => ("Minimal", "minimal"),
316                open_ai::ReasoningEffort::Low => ("Low", "low"),
317                open_ai::ReasoningEffort::Medium => ("Medium", "medium"),
318                open_ai::ReasoningEffort::High => ("High", "high"),
319                open_ai::ReasoningEffort::XHigh => ("Extra High", "xhigh"),
320                open_ai::ReasoningEffort::Max => return None, // Not supported by any xAI models
321            };
322
323            Some(LanguageModelEffortLevel {
324                name: name.into(),
325                value: value.into(),
326                is_default: Some(effort) == default_effort,
327            })
328        })
329        .collect()
330}
331
332impl LanguageModel for XAiLanguageModel {
333    fn id(&self) -> LanguageModelId {
334        self.id.clone()
335    }
336
337    fn name(&self) -> LanguageModelName {
338        LanguageModelName::from(self.model.display_name().to_string())
339    }
340
341    fn provider_id(&self) -> LanguageModelProviderId {
342        PROVIDER_ID
343    }
344
345    fn provider_name(&self) -> LanguageModelProviderName {
346        PROVIDER_NAME
347    }
348
349    fn supports_tools(&self) -> bool {
350        self.model.supports_tool()
351    }
352
353    fn supports_images(&self) -> bool {
354        self.model.supports_images()
355    }
356
357    fn supports_streaming_tools(&self) -> bool {
358        true
359    }
360
361    fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
362        match choice {
363            LanguageModelToolChoice::Auto
364            | LanguageModelToolChoice::Any
365            | LanguageModelToolChoice::None => true,
366        }
367    }
368
369    fn supports_thinking(&self) -> bool {
370        self.model.supports_reasoning_effort()
371    }
372
373    fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {
374        supported_thinking_effort_levels(&self.model)
375    }
376
377    fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
378        if self.model.requires_json_schema_subset() {
379            LanguageModelToolSchemaFormat::JsonSchemaSubset
380        } else {
381            LanguageModelToolSchemaFormat::JsonSchema
382        }
383    }
384
385    fn telemetry_id(&self) -> String {
386        format!("x_ai/{}", self.model.id())
387    }
388
389    fn max_token_count(&self) -> u64 {
390        self.model.max_token_count()
391    }
392
393    fn max_output_tokens(&self) -> Option<u64> {
394        self.model.max_output_tokens()
395    }
396
397    fn supports_split_token_display(&self) -> bool {
398        true
399    }
400
401    fn stream_completion(
402        &self,
403        request: LanguageModelRequest,
404        cx: &AsyncApp,
405    ) -> BoxFuture<
406        'static,
407        Result<
408            futures::stream::BoxStream<
409                'static,
410                Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
411            >,
412            LanguageModelCompletionError,
413        >,
414    > {
415        let reasoning_effort = reasoning_effort_for_request(&request, &self.model);
416        let request = match crate::provider::open_ai::into_open_ai(
417            request,
418            self.model.id(),
419            self.model.supports_parallel_tool_calls(),
420            self.model.supports_prompt_cache_key(),
421            self.max_output_tokens(),
422            crate::provider::open_ai::ChatCompletionMaxTokensParameter::MaxCompletionTokens,
423            reasoning_effort,
424            false,
425        ) {
426            Ok(request) => request,
427            Err(error) => return async move { Err(error.into()) }.boxed(),
428        };
429        let completions = self.stream_completion(request, cx);
430        async move {
431            let mapper = crate::provider::open_ai::OpenAiEventMapper::new();
432            Ok(mapper.map_stream(completions.await?).boxed())
433        }
434        .boxed()
435    }
436}
437
438#[cfg(test)]
439mod tests {
440    use super::*;
441
442    #[test]
443    fn grok_43_supports_selectable_thinking_effort_levels() {
444        let effort_levels = supported_thinking_effort_levels(&x_ai::Model::Grok43);
445        let values = effort_levels
446            .iter()
447            .map(|level| level.value.as_ref())
448            .collect::<Vec<_>>();
449
450        assert_eq!(values, ["low", "medium", "high"]);
451        assert_eq!(
452            effort_levels
453                .iter()
454                .find(|level| level.is_default)
455                .map(|level| level.value.as_ref()),
456            Some("low")
457        );
458    }
459
460    #[test]
461    fn grok_43_request_uses_selected_reasoning_effort() {
462        let request = LanguageModelRequest {
463            thinking_allowed: true,
464            thinking_effort: Some("high".to_string()),
465            ..Default::default()
466        };
467
468        assert_eq!(
469            reasoning_effort_for_request(&request, &x_ai::Model::Grok43),
470            Some(open_ai::ReasoningEffort::High)
471        );
472    }
473
474    #[test]
475    fn grok_43_request_uses_none_when_thinking_is_disabled() {
476        let request = LanguageModelRequest {
477            thinking_allowed: false,
478            ..Default::default()
479        };
480
481        assert_eq!(
482            reasoning_effort_for_request(&request, &x_ai::Model::Grok43),
483            Some(open_ai::ReasoningEffort::None)
484        );
485    }
486}
487