1mod api_key;
2mod model;
3mod rate_limiter;
4mod registry;
5mod request;
6mod role;
7mod telemetry;
8pub mod tool_schema;
9
10#[cfg(any(test, feature = "test-support"))]
11pub mod fake_provider;
12
13use anthropic::{AnthropicError, parse_prompt_too_long};
14use anyhow::{Result, anyhow};
15use client::Client;
16use cloud_llm_client::CompletionRequestStatus;
17use futures::FutureExt;
18use futures::{StreamExt, future::BoxFuture, stream::BoxStream};
19use gpui::{AnyView, App, AsyncApp, SharedString, Task, Window};
20use http_client::{StatusCode, http};
21use icons::IconName;
22use open_router::OpenRouterError;
23use parking_lot::Mutex;
24use serde::{Deserialize, Serialize};
25pub use settings::LanguageModelCacheConfiguration;
26use std::ops::{Add, Sub};
27use std::str::FromStr;
28use std::sync::Arc;
29use std::time::Duration;
30use std::{fmt, io};
31use thiserror::Error;
32use util::serde::is_default;
33
34pub use crate::api_key::{ApiKey, ApiKeyState};
35pub use crate::model::*;
36pub use crate::rate_limiter::*;
37pub use crate::registry::*;
38pub use crate::request::*;
39pub use crate::role::*;
40pub use crate::telemetry::*;
41pub use crate::tool_schema::LanguageModelToolSchemaFormat;
42pub use zed_env_vars::{EnvVar, env_var};
43
44pub const ANTHROPIC_PROVIDER_ID: LanguageModelProviderId =
45 LanguageModelProviderId::new("anthropic");
46pub const ANTHROPIC_PROVIDER_NAME: LanguageModelProviderName =
47 LanguageModelProviderName::new("Anthropic");
48
49pub const GOOGLE_PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("google");
50pub const GOOGLE_PROVIDER_NAME: LanguageModelProviderName =
51 LanguageModelProviderName::new("Google AI");
52
53pub const OPEN_AI_PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("openai");
54pub const OPEN_AI_PROVIDER_NAME: LanguageModelProviderName =
55 LanguageModelProviderName::new("OpenAI");
56
57pub const X_AI_PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("x_ai");
58pub const X_AI_PROVIDER_NAME: LanguageModelProviderName = LanguageModelProviderName::new("xAI");
59
60pub const ZED_CLOUD_PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("zed.dev");
61pub const ZED_CLOUD_PROVIDER_NAME: LanguageModelProviderName =
62 LanguageModelProviderName::new("Zed");
63
64pub fn init(client: Arc<Client>, cx: &mut App) {
65 init_settings(cx);
66 RefreshLlmTokenListener::register(client, cx);
67}
68
69pub fn init_settings(cx: &mut App) {
70 registry::init(cx);
71}
72
73/// A completion event from a language model.
74#[derive(Debug, PartialEq, Clone, Serialize, Deserialize)]
75pub enum LanguageModelCompletionEvent {
76 Queued {
77 position: usize,
78 },
79 Started,
80 Stop(StopReason),
81 Text(String),
82 Thinking {
83 text: String,
84 signature: Option<String>,
85 },
86 RedactedThinking {
87 data: String,
88 },
89 ToolUse(LanguageModelToolUse),
90 ToolUseJsonParseError {
91 id: LanguageModelToolUseId,
92 tool_name: Arc<str>,
93 raw_input: Arc<str>,
94 json_parse_error: String,
95 },
96 StartMessage {
97 message_id: String,
98 },
99 ReasoningDetails(serde_json::Value),
100 UsageUpdate(TokenUsage),
101}
102
103impl LanguageModelCompletionEvent {
104 pub fn from_completion_request_status(
105 status: CompletionRequestStatus,
106 upstream_provider: LanguageModelProviderName,
107 ) -> Result<Option<Self>, LanguageModelCompletionError> {
108 match status {
109 CompletionRequestStatus::Queued { position } => {
110 Ok(Some(LanguageModelCompletionEvent::Queued { position }))
111 }
112 CompletionRequestStatus::Started => Ok(Some(LanguageModelCompletionEvent::Started)),
113 CompletionRequestStatus::Unknown | CompletionRequestStatus::StreamEnded => Ok(None),
114 CompletionRequestStatus::UsageUpdated { .. }
115 | CompletionRequestStatus::ToolUseLimitReached => Err(
116 LanguageModelCompletionError::Other(anyhow!("Unexpected status: {status:?}")),
117 ),
118 CompletionRequestStatus::Failed {
119 code,
120 message,
121 request_id: _,
122 retry_after,
123 } => Err(LanguageModelCompletionError::from_cloud_failure(
124 upstream_provider,
125 code,
126 message,
127 retry_after.map(Duration::from_secs_f64),
128 )),
129 }
130 }
131}
132
133#[derive(Error, Debug)]
134pub enum LanguageModelCompletionError {
135 #[error("prompt too large for context window")]
136 PromptTooLarge { tokens: Option<u64> },
137 #[error("missing {provider} API key")]
138 NoApiKey { provider: LanguageModelProviderName },
139 #[error("{provider}'s API rate limit exceeded")]
140 RateLimitExceeded {
141 provider: LanguageModelProviderName,
142 retry_after: Option<Duration>,
143 },
144 #[error("{provider}'s API servers are overloaded right now")]
145 ServerOverloaded {
146 provider: LanguageModelProviderName,
147 retry_after: Option<Duration>,
148 },
149 #[error("{provider}'s API server reported an internal server error: {message}")]
150 ApiInternalServerError {
151 provider: LanguageModelProviderName,
152 message: String,
153 },
154 #[error("{message}")]
155 UpstreamProviderError {
156 message: String,
157 status: StatusCode,
158 retry_after: Option<Duration>,
159 },
160 #[error("HTTP response error from {provider}'s API: status {status_code} - {message:?}")]
161 HttpResponseError {
162 provider: LanguageModelProviderName,
163 status_code: StatusCode,
164 message: String,
165 },
166
167 // Client errors
168 #[error("invalid request format to {provider}'s API: {message}")]
169 BadRequestFormat {
170 provider: LanguageModelProviderName,
171 message: String,
172 },
173 #[error("authentication error with {provider}'s API: {message}")]
174 AuthenticationError {
175 provider: LanguageModelProviderName,
176 message: String,
177 },
178 #[error("Permission error with {provider}'s API: {message}")]
179 PermissionError {
180 provider: LanguageModelProviderName,
181 message: String,
182 },
183 #[error("language model provider API endpoint not found")]
184 ApiEndpointNotFound { provider: LanguageModelProviderName },
185 #[error("I/O error reading response from {provider}'s API")]
186 ApiReadResponseError {
187 provider: LanguageModelProviderName,
188 #[source]
189 error: io::Error,
190 },
191 #[error("error serializing request to {provider} API")]
192 SerializeRequest {
193 provider: LanguageModelProviderName,
194 #[source]
195 error: serde_json::Error,
196 },
197 #[error("error building request body to {provider} API")]
198 BuildRequestBody {
199 provider: LanguageModelProviderName,
200 #[source]
201 error: http::Error,
202 },
203 #[error("error sending HTTP request to {provider} API")]
204 HttpSend {
205 provider: LanguageModelProviderName,
206 #[source]
207 error: anyhow::Error,
208 },
209 #[error("error deserializing {provider} API response")]
210 DeserializeResponse {
211 provider: LanguageModelProviderName,
212 #[source]
213 error: serde_json::Error,
214 },
215
216 #[error("stream from {provider} ended unexpectedly")]
217 StreamEndedUnexpectedly { provider: LanguageModelProviderName },
218
219 // TODO: Ideally this would be removed in favor of having a comprehensive list of errors.
220 #[error(transparent)]
221 Other(#[from] anyhow::Error),
222}
223
224impl LanguageModelCompletionError {
225 fn parse_upstream_error_json(message: &str) -> Option<(StatusCode, String)> {
226 let error_json = serde_json::from_str::<serde_json::Value>(message).ok()?;
227 let upstream_status = error_json
228 .get("upstream_status")
229 .and_then(|v| v.as_u64())
230 .and_then(|status| u16::try_from(status).ok())
231 .and_then(|status| StatusCode::from_u16(status).ok())?;
232 let inner_message = error_json
233 .get("message")
234 .and_then(|v| v.as_str())
235 .unwrap_or(message)
236 .to_string();
237 Some((upstream_status, inner_message))
238 }
239
240 pub fn from_cloud_failure(
241 upstream_provider: LanguageModelProviderName,
242 code: String,
243 message: String,
244 retry_after: Option<Duration>,
245 ) -> Self {
246 if let Some(tokens) = parse_prompt_too_long(&message) {
247 // TODO: currently Anthropic PAYLOAD_TOO_LARGE response may cause INTERNAL_SERVER_ERROR
248 // to be reported. This is a temporary workaround to handle this in the case where the
249 // token limit has been exceeded.
250 Self::PromptTooLarge {
251 tokens: Some(tokens),
252 }
253 } else if code == "upstream_http_error" {
254 if let Some((upstream_status, inner_message)) =
255 Self::parse_upstream_error_json(&message)
256 {
257 return Self::from_http_status(
258 upstream_provider,
259 upstream_status,
260 inner_message,
261 retry_after,
262 );
263 }
264 anyhow!("completion request failed, code: {code}, message: {message}").into()
265 } else if let Some(status_code) = code
266 .strip_prefix("upstream_http_")
267 .and_then(|code| StatusCode::from_str(code).ok())
268 {
269 Self::from_http_status(upstream_provider, status_code, message, retry_after)
270 } else if let Some(status_code) = code
271 .strip_prefix("http_")
272 .and_then(|code| StatusCode::from_str(code).ok())
273 {
274 Self::from_http_status(ZED_CLOUD_PROVIDER_NAME, status_code, message, retry_after)
275 } else {
276 anyhow!("completion request failed, code: {code}, message: {message}").into()
277 }
278 }
279
280 pub fn from_http_status(
281 provider: LanguageModelProviderName,
282 status_code: StatusCode,
283 message: String,
284 retry_after: Option<Duration>,
285 ) -> Self {
286 match status_code {
287 StatusCode::BAD_REQUEST => Self::BadRequestFormat { provider, message },
288 StatusCode::UNAUTHORIZED => Self::AuthenticationError { provider, message },
289 StatusCode::FORBIDDEN => Self::PermissionError { provider, message },
290 StatusCode::NOT_FOUND => Self::ApiEndpointNotFound { provider },
291 StatusCode::PAYLOAD_TOO_LARGE => Self::PromptTooLarge {
292 tokens: parse_prompt_too_long(&message),
293 },
294 StatusCode::TOO_MANY_REQUESTS => Self::RateLimitExceeded {
295 provider,
296 retry_after,
297 },
298 StatusCode::INTERNAL_SERVER_ERROR => Self::ApiInternalServerError { provider, message },
299 StatusCode::SERVICE_UNAVAILABLE => Self::ServerOverloaded {
300 provider,
301 retry_after,
302 },
303 _ if status_code.as_u16() == 529 => Self::ServerOverloaded {
304 provider,
305 retry_after,
306 },
307 _ => Self::HttpResponseError {
308 provider,
309 status_code,
310 message,
311 },
312 }
313 }
314}
315
316impl From<AnthropicError> for LanguageModelCompletionError {
317 fn from(error: AnthropicError) -> Self {
318 let provider = ANTHROPIC_PROVIDER_NAME;
319 match error {
320 AnthropicError::SerializeRequest(error) => Self::SerializeRequest { provider, error },
321 AnthropicError::BuildRequestBody(error) => Self::BuildRequestBody { provider, error },
322 AnthropicError::HttpSend(error) => Self::HttpSend { provider, error },
323 AnthropicError::DeserializeResponse(error) => {
324 Self::DeserializeResponse { provider, error }
325 }
326 AnthropicError::ReadResponse(error) => Self::ApiReadResponseError { provider, error },
327 AnthropicError::HttpResponseError {
328 status_code,
329 message,
330 } => Self::HttpResponseError {
331 provider,
332 status_code,
333 message,
334 },
335 AnthropicError::RateLimit { retry_after } => Self::RateLimitExceeded {
336 provider,
337 retry_after: Some(retry_after),
338 },
339 AnthropicError::ServerOverloaded { retry_after } => Self::ServerOverloaded {
340 provider,
341 retry_after,
342 },
343 AnthropicError::ApiError(api_error) => api_error.into(),
344 }
345 }
346}
347
348impl From<anthropic::ApiError> for LanguageModelCompletionError {
349 fn from(error: anthropic::ApiError) -> Self {
350 use anthropic::ApiErrorCode::*;
351 let provider = ANTHROPIC_PROVIDER_NAME;
352 match error.code() {
353 Some(code) => match code {
354 InvalidRequestError => Self::BadRequestFormat {
355 provider,
356 message: error.message,
357 },
358 AuthenticationError => Self::AuthenticationError {
359 provider,
360 message: error.message,
361 },
362 PermissionError => Self::PermissionError {
363 provider,
364 message: error.message,
365 },
366 NotFoundError => Self::ApiEndpointNotFound { provider },
367 RequestTooLarge => Self::PromptTooLarge {
368 tokens: parse_prompt_too_long(&error.message),
369 },
370 RateLimitError => Self::RateLimitExceeded {
371 provider,
372 retry_after: None,
373 },
374 ApiError => Self::ApiInternalServerError {
375 provider,
376 message: error.message,
377 },
378 OverloadedError => Self::ServerOverloaded {
379 provider,
380 retry_after: None,
381 },
382 },
383 None => Self::Other(error.into()),
384 }
385 }
386}
387
388impl From<open_ai::RequestError> for LanguageModelCompletionError {
389 fn from(error: open_ai::RequestError) -> Self {
390 match error {
391 open_ai::RequestError::HttpResponseError {
392 provider,
393 status_code,
394 body,
395 headers,
396 } => {
397 let retry_after = headers
398 .get(http::header::RETRY_AFTER)
399 .and_then(|val| val.to_str().ok()?.parse::<u64>().ok())
400 .map(Duration::from_secs);
401
402 Self::from_http_status(provider.into(), status_code, body, retry_after)
403 }
404 open_ai::RequestError::Other(e) => Self::Other(e),
405 }
406 }
407}
408
409impl From<OpenRouterError> for LanguageModelCompletionError {
410 fn from(error: OpenRouterError) -> Self {
411 let provider = LanguageModelProviderName::new("OpenRouter");
412 match error {
413 OpenRouterError::SerializeRequest(error) => Self::SerializeRequest { provider, error },
414 OpenRouterError::BuildRequestBody(error) => Self::BuildRequestBody { provider, error },
415 OpenRouterError::HttpSend(error) => Self::HttpSend { provider, error },
416 OpenRouterError::DeserializeResponse(error) => {
417 Self::DeserializeResponse { provider, error }
418 }
419 OpenRouterError::ReadResponse(error) => Self::ApiReadResponseError { provider, error },
420 OpenRouterError::RateLimit { retry_after } => Self::RateLimitExceeded {
421 provider,
422 retry_after: Some(retry_after),
423 },
424 OpenRouterError::ServerOverloaded { retry_after } => Self::ServerOverloaded {
425 provider,
426 retry_after,
427 },
428 OpenRouterError::ApiError(api_error) => api_error.into(),
429 }
430 }
431}
432
433impl From<open_router::ApiError> for LanguageModelCompletionError {
434 fn from(error: open_router::ApiError) -> Self {
435 use open_router::ApiErrorCode::*;
436 let provider = LanguageModelProviderName::new("OpenRouter");
437 match error.code {
438 InvalidRequestError => Self::BadRequestFormat {
439 provider,
440 message: error.message,
441 },
442 AuthenticationError => Self::AuthenticationError {
443 provider,
444 message: error.message,
445 },
446 PaymentRequiredError => Self::AuthenticationError {
447 provider,
448 message: format!("Payment required: {}", error.message),
449 },
450 PermissionError => Self::PermissionError {
451 provider,
452 message: error.message,
453 },
454 RequestTimedOut => Self::HttpResponseError {
455 provider,
456 status_code: StatusCode::REQUEST_TIMEOUT,
457 message: error.message,
458 },
459 RateLimitError => Self::RateLimitExceeded {
460 provider,
461 retry_after: None,
462 },
463 ApiError => Self::ApiInternalServerError {
464 provider,
465 message: error.message,
466 },
467 OverloadedError => Self::ServerOverloaded {
468 provider,
469 retry_after: None,
470 },
471 }
472 }
473}
474
475#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize)]
476#[serde(rename_all = "snake_case")]
477pub enum StopReason {
478 EndTurn,
479 MaxTokens,
480 ToolUse,
481 Refusal,
482}
483
484#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize, Default)]
485pub struct TokenUsage {
486 #[serde(default, skip_serializing_if = "is_default")]
487 pub input_tokens: u64,
488 #[serde(default, skip_serializing_if = "is_default")]
489 pub output_tokens: u64,
490 #[serde(default, skip_serializing_if = "is_default")]
491 pub cache_creation_input_tokens: u64,
492 #[serde(default, skip_serializing_if = "is_default")]
493 pub cache_read_input_tokens: u64,
494}
495
496impl TokenUsage {
497 pub fn total_tokens(&self) -> u64 {
498 self.input_tokens
499 + self.output_tokens
500 + self.cache_read_input_tokens
501 + self.cache_creation_input_tokens
502 }
503}
504
505impl Add<TokenUsage> for TokenUsage {
506 type Output = Self;
507
508 fn add(self, other: Self) -> Self {
509 Self {
510 input_tokens: self.input_tokens + other.input_tokens,
511 output_tokens: self.output_tokens + other.output_tokens,
512 cache_creation_input_tokens: self.cache_creation_input_tokens
513 + other.cache_creation_input_tokens,
514 cache_read_input_tokens: self.cache_read_input_tokens + other.cache_read_input_tokens,
515 }
516 }
517}
518
519impl Sub<TokenUsage> for TokenUsage {
520 type Output = Self;
521
522 fn sub(self, other: Self) -> Self {
523 Self {
524 input_tokens: self.input_tokens - other.input_tokens,
525 output_tokens: self.output_tokens - other.output_tokens,
526 cache_creation_input_tokens: self.cache_creation_input_tokens
527 - other.cache_creation_input_tokens,
528 cache_read_input_tokens: self.cache_read_input_tokens - other.cache_read_input_tokens,
529 }
530 }
531}
532
533#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
534pub struct LanguageModelToolUseId(Arc<str>);
535
536impl fmt::Display for LanguageModelToolUseId {
537 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
538 write!(f, "{}", self.0)
539 }
540}
541
542impl<T> From<T> for LanguageModelToolUseId
543where
544 T: Into<Arc<str>>,
545{
546 fn from(value: T) -> Self {
547 Self(value.into())
548 }
549}
550
551#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
552pub struct LanguageModelToolUse {
553 pub id: LanguageModelToolUseId,
554 pub name: Arc<str>,
555 pub raw_input: String,
556 pub input: serde_json::Value,
557 pub is_input_complete: bool,
558 /// Thought signature the model sent us. Some models require that this
559 /// signature be preserved and sent back in conversation history for validation.
560 pub thought_signature: Option<String>,
561}
562
563pub struct LanguageModelTextStream {
564 pub message_id: Option<String>,
565 pub stream: BoxStream<'static, Result<String, LanguageModelCompletionError>>,
566 // Has complete token usage after the stream has finished
567 pub last_token_usage: Arc<Mutex<TokenUsage>>,
568}
569
570impl Default for LanguageModelTextStream {
571 fn default() -> Self {
572 Self {
573 message_id: None,
574 stream: Box::pin(futures::stream::empty()),
575 last_token_usage: Arc::new(Mutex::new(TokenUsage::default())),
576 }
577 }
578}
579
580#[derive(Debug, Clone)]
581pub struct LanguageModelEffortLevel {
582 pub name: SharedString,
583 pub value: SharedString,
584 pub is_default: bool,
585}
586
587pub trait LanguageModel: Send + Sync {
588 fn id(&self) -> LanguageModelId;
589 fn name(&self) -> LanguageModelName;
590 fn provider_id(&self) -> LanguageModelProviderId;
591 fn provider_name(&self) -> LanguageModelProviderName;
592 fn upstream_provider_id(&self) -> LanguageModelProviderId {
593 self.provider_id()
594 }
595 fn upstream_provider_name(&self) -> LanguageModelProviderName {
596 self.provider_name()
597 }
598
599 /// Returns whether this model is the "latest", so we can highlight it in the UI.
600 fn is_latest(&self) -> bool {
601 false
602 }
603
604 fn telemetry_id(&self) -> String;
605
606 fn api_key(&self, _cx: &App) -> Option<String> {
607 None
608 }
609
610 /// Information about the cost of using this model, if available.
611 fn model_cost_info(&self) -> Option<LanguageModelCostInfo> {
612 None
613 }
614
615 /// Whether this model supports thinking.
616 fn supports_thinking(&self) -> bool {
617 false
618 }
619
620 /// Returns the list of supported effort levels that can be used when thinking.
621 fn supported_effort_levels(&self) -> Vec<LanguageModelEffortLevel> {
622 Vec::new()
623 }
624
625 /// Returns the default effort level to use when thinking.
626 fn default_effort_level(&self) -> Option<LanguageModelEffortLevel> {
627 self.supported_effort_levels()
628 .into_iter()
629 .find(|effort_level| effort_level.is_default)
630 }
631
632 /// Whether this model supports images
633 fn supports_images(&self) -> bool;
634
635 /// Whether this model supports tools.
636 fn supports_tools(&self) -> bool;
637
638 /// Whether this model supports choosing which tool to use.
639 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool;
640
641 /// Returns whether this model or provider supports streaming tool calls;
642 fn supports_streaming_tools(&self) -> bool {
643 false
644 }
645
646 /// Returns whether this model/provider reports accurate split input/output token counts.
647 /// When true, the UI may show separate input/output token indicators.
648 fn supports_split_token_display(&self) -> bool {
649 false
650 }
651
652 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
653 LanguageModelToolSchemaFormat::JsonSchema
654 }
655
656 fn max_token_count(&self) -> u64;
657 fn max_output_tokens(&self) -> Option<u64> {
658 None
659 }
660
661 fn count_tokens(
662 &self,
663 request: LanguageModelRequest,
664 cx: &App,
665 ) -> BoxFuture<'static, Result<u64>>;
666
667 fn stream_completion(
668 &self,
669 request: LanguageModelRequest,
670 cx: &AsyncApp,
671 ) -> BoxFuture<
672 'static,
673 Result<
674 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
675 LanguageModelCompletionError,
676 >,
677 >;
678
679 fn stream_completion_text(
680 &self,
681 request: LanguageModelRequest,
682 cx: &AsyncApp,
683 ) -> BoxFuture<'static, Result<LanguageModelTextStream, LanguageModelCompletionError>> {
684 let future = self.stream_completion(request, cx);
685
686 async move {
687 let events = future.await?;
688 let mut events = events.fuse();
689 let mut message_id = None;
690 let mut first_item_text = None;
691 let last_token_usage = Arc::new(Mutex::new(TokenUsage::default()));
692
693 if let Some(first_event) = events.next().await {
694 match first_event {
695 Ok(LanguageModelCompletionEvent::StartMessage { message_id: id }) => {
696 message_id = Some(id);
697 }
698 Ok(LanguageModelCompletionEvent::Text(text)) => {
699 first_item_text = Some(text);
700 }
701 _ => (),
702 }
703 }
704
705 let stream = futures::stream::iter(first_item_text.map(Ok))
706 .chain(events.filter_map({
707 let last_token_usage = last_token_usage.clone();
708 move |result| {
709 let last_token_usage = last_token_usage.clone();
710 async move {
711 match result {
712 Ok(LanguageModelCompletionEvent::Queued { .. }) => None,
713 Ok(LanguageModelCompletionEvent::Started) => None,
714 Ok(LanguageModelCompletionEvent::StartMessage { .. }) => None,
715 Ok(LanguageModelCompletionEvent::Text(text)) => Some(Ok(text)),
716 Ok(LanguageModelCompletionEvent::Thinking { .. }) => None,
717 Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => None,
718 Ok(LanguageModelCompletionEvent::ReasoningDetails(_)) => None,
719 Ok(LanguageModelCompletionEvent::Stop(_)) => None,
720 Ok(LanguageModelCompletionEvent::ToolUse(_)) => None,
721 Ok(LanguageModelCompletionEvent::ToolUseJsonParseError {
722 ..
723 }) => None,
724 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {
725 *last_token_usage.lock() = token_usage;
726 None
727 }
728 Err(err) => Some(Err(err)),
729 }
730 }
731 }
732 }))
733 .boxed();
734
735 Ok(LanguageModelTextStream {
736 message_id,
737 stream,
738 last_token_usage,
739 })
740 }
741 .boxed()
742 }
743
744 fn stream_completion_tool(
745 &self,
746 request: LanguageModelRequest,
747 cx: &AsyncApp,
748 ) -> BoxFuture<'static, Result<LanguageModelToolUse, LanguageModelCompletionError>> {
749 let future = self.stream_completion(request, cx);
750
751 async move {
752 let events = future.await?;
753 let mut events = events.fuse();
754
755 // Iterate through events until we find a complete ToolUse
756 while let Some(event) = events.next().await {
757 match event {
758 Ok(LanguageModelCompletionEvent::ToolUse(tool_use))
759 if tool_use.is_input_complete =>
760 {
761 return Ok(tool_use);
762 }
763 Err(err) => {
764 return Err(err);
765 }
766 _ => {}
767 }
768 }
769
770 // Stream ended without a complete tool use
771 Err(LanguageModelCompletionError::Other(anyhow::anyhow!(
772 "Stream ended without receiving a complete tool use"
773 )))
774 }
775 .boxed()
776 }
777
778 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
779 None
780 }
781
782 #[cfg(any(test, feature = "test-support"))]
783 fn as_fake(&self) -> &fake_provider::FakeLanguageModel {
784 unimplemented!()
785 }
786}
787
788impl std::fmt::Debug for dyn LanguageModel {
789 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
790 f.debug_struct("<dyn LanguageModel>")
791 .field("id", &self.id())
792 .field("name", &self.name())
793 .field("provider_id", &self.provider_id())
794 .field("provider_name", &self.provider_name())
795 .field("upstream_provider_name", &self.upstream_provider_name())
796 .field("upstream_provider_id", &self.upstream_provider_id())
797 .field("upstream_provider_id", &self.upstream_provider_id())
798 .field("supports_streaming_tools", &self.supports_streaming_tools())
799 .finish()
800 }
801}
802
803/// An error that occurred when trying to authenticate the language model provider.
804#[derive(Debug, Error)]
805pub enum AuthenticateError {
806 #[error("connection refused")]
807 ConnectionRefused,
808 #[error("credentials not found")]
809 CredentialsNotFound,
810 #[error(transparent)]
811 Other(#[from] anyhow::Error),
812}
813
814/// Either a built-in icon name or a path to an external SVG.
815#[derive(Debug, Clone, PartialEq, Eq)]
816pub enum IconOrSvg {
817 /// A built-in icon from Zed's icon set.
818 Icon(IconName),
819 /// Path to a custom SVG icon file.
820 Svg(SharedString),
821}
822
823impl Default for IconOrSvg {
824 fn default() -> Self {
825 Self::Icon(IconName::ZedAssistant)
826 }
827}
828
829pub trait LanguageModelProvider: 'static {
830 fn id(&self) -> LanguageModelProviderId;
831 fn name(&self) -> LanguageModelProviderName;
832 fn icon(&self) -> IconOrSvg {
833 IconOrSvg::default()
834 }
835 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
836 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
837 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>>;
838 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
839 Vec::new()
840 }
841 fn is_authenticated(&self, cx: &App) -> bool;
842 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>>;
843 fn configuration_view(
844 &self,
845 target_agent: ConfigurationViewTargetAgent,
846 window: &mut Window,
847 cx: &mut App,
848 ) -> AnyView;
849 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>>;
850}
851
852#[derive(Default, Clone, PartialEq, Eq)]
853pub enum ConfigurationViewTargetAgent {
854 #[default]
855 ZedAgent,
856 Other(SharedString),
857}
858
859#[derive(PartialEq, Eq)]
860pub enum LanguageModelProviderTosView {
861 /// When there are some past interactions in the Agent Panel.
862 ThreadEmptyState,
863 /// When there are no past interactions in the Agent Panel.
864 ThreadFreshStart,
865 TextThreadPopup,
866 Configuration,
867}
868
869pub trait LanguageModelProviderState: 'static {
870 type ObservableEntity;
871
872 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>>;
873
874 fn subscribe<T: 'static>(
875 &self,
876 cx: &mut gpui::Context<T>,
877 callback: impl Fn(&mut T, &mut gpui::Context<T>) + 'static,
878 ) -> Option<gpui::Subscription> {
879 let entity = self.observable_entity()?;
880 Some(cx.observe(&entity, move |this, _, cx| {
881 callback(this, cx);
882 }))
883 }
884}
885
886#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd, Serialize, Deserialize)]
887pub struct LanguageModelId(pub SharedString);
888
889#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
890pub struct LanguageModelName(pub SharedString);
891
892#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
893pub struct LanguageModelProviderId(pub SharedString);
894
895#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
896pub struct LanguageModelProviderName(pub SharedString);
897
898#[derive(Clone, Debug, PartialEq)]
899pub enum LanguageModelCostInfo {
900 /// Cost per 1,000 input and output tokens
901 TokenCost {
902 input_token_cost_per_1m: f64,
903 output_token_cost_per_1m: f64,
904 },
905 /// Cost per request
906 RequestCost { cost_per_request: f64 },
907}
908
909impl LanguageModelCostInfo {
910 pub fn to_shared_string(&self) -> SharedString {
911 match self {
912 LanguageModelCostInfo::RequestCost { cost_per_request } => {
913 let cost_str = format!("{}×", Self::cost_value_to_string(cost_per_request));
914 SharedString::from(cost_str)
915 }
916 LanguageModelCostInfo::TokenCost {
917 input_token_cost_per_1m,
918 output_token_cost_per_1m,
919 } => {
920 let input_cost = Self::cost_value_to_string(input_token_cost_per_1m);
921 let output_cost = Self::cost_value_to_string(output_token_cost_per_1m);
922 SharedString::from(format!("{}$/{}$", input_cost, output_cost))
923 }
924 }
925 }
926
927 fn cost_value_to_string(cost: &f64) -> SharedString {
928 if (cost.fract() - 0.0).abs() < std::f64::EPSILON {
929 SharedString::from(format!("{:.0}", cost))
930 } else {
931 SharedString::from(format!("{:.2}", cost))
932 }
933 }
934}
935
936impl LanguageModelProviderId {
937 pub const fn new(id: &'static str) -> Self {
938 Self(SharedString::new_static(id))
939 }
940}
941
942impl LanguageModelProviderName {
943 pub const fn new(id: &'static str) -> Self {
944 Self(SharedString::new_static(id))
945 }
946}
947
948impl fmt::Display for LanguageModelProviderId {
949 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
950 write!(f, "{}", self.0)
951 }
952}
953
954impl fmt::Display for LanguageModelProviderName {
955 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
956 write!(f, "{}", self.0)
957 }
958}
959
960impl From<String> for LanguageModelId {
961 fn from(value: String) -> Self {
962 Self(SharedString::from(value))
963 }
964}
965
966impl From<String> for LanguageModelName {
967 fn from(value: String) -> Self {
968 Self(SharedString::from(value))
969 }
970}
971
972impl From<String> for LanguageModelProviderId {
973 fn from(value: String) -> Self {
974 Self(SharedString::from(value))
975 }
976}
977
978impl From<String> for LanguageModelProviderName {
979 fn from(value: String) -> Self {
980 Self(SharedString::from(value))
981 }
982}
983
984impl From<Arc<str>> for LanguageModelProviderId {
985 fn from(value: Arc<str>) -> Self {
986 Self(SharedString::from(value))
987 }
988}
989
990impl From<Arc<str>> for LanguageModelProviderName {
991 fn from(value: Arc<str>) -> Self {
992 Self(SharedString::from(value))
993 }
994}
995
996#[cfg(test)]
997mod tests {
998 use super::*;
999
1000 #[test]
1001 fn test_from_cloud_failure_with_upstream_http_error() {
1002 let error = LanguageModelCompletionError::from_cloud_failure(
1003 String::from("anthropic").into(),
1004 "upstream_http_error".to_string(),
1005 r#"{"code":"upstream_http_error","message":"Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers. reset reason: connection timeout","upstream_status":503}"#.to_string(),
1006 None,
1007 );
1008
1009 match error {
1010 LanguageModelCompletionError::ServerOverloaded { provider, .. } => {
1011 assert_eq!(provider.0, "anthropic");
1012 }
1013 _ => panic!(
1014 "Expected ServerOverloaded error for 503 status, got: {:?}",
1015 error
1016 ),
1017 }
1018
1019 let error = LanguageModelCompletionError::from_cloud_failure(
1020 String::from("anthropic").into(),
1021 "upstream_http_error".to_string(),
1022 r#"{"code":"upstream_http_error","message":"Internal server error","upstream_status":500}"#.to_string(),
1023 None,
1024 );
1025
1026 match error {
1027 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1028 assert_eq!(provider.0, "anthropic");
1029 assert_eq!(message, "Internal server error");
1030 }
1031 _ => panic!(
1032 "Expected ApiInternalServerError for 500 status, got: {:?}",
1033 error
1034 ),
1035 }
1036 }
1037
1038 #[test]
1039 fn test_from_cloud_failure_with_standard_format() {
1040 let error = LanguageModelCompletionError::from_cloud_failure(
1041 String::from("anthropic").into(),
1042 "upstream_http_503".to_string(),
1043 "Service unavailable".to_string(),
1044 None,
1045 );
1046
1047 match error {
1048 LanguageModelCompletionError::ServerOverloaded { provider, .. } => {
1049 assert_eq!(provider.0, "anthropic");
1050 }
1051 _ => panic!("Expected ServerOverloaded error for upstream_http_503"),
1052 }
1053 }
1054
1055 #[test]
1056 fn test_upstream_http_error_connection_timeout() {
1057 let error = LanguageModelCompletionError::from_cloud_failure(
1058 String::from("anthropic").into(),
1059 "upstream_http_error".to_string(),
1060 r#"{"code":"upstream_http_error","message":"Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers. reset reason: connection timeout","upstream_status":503}"#.to_string(),
1061 None,
1062 );
1063
1064 match error {
1065 LanguageModelCompletionError::ServerOverloaded { provider, .. } => {
1066 assert_eq!(provider.0, "anthropic");
1067 }
1068 _ => panic!(
1069 "Expected ServerOverloaded error for connection timeout with 503 status, got: {:?}",
1070 error
1071 ),
1072 }
1073
1074 let error = LanguageModelCompletionError::from_cloud_failure(
1075 String::from("anthropic").into(),
1076 "upstream_http_error".to_string(),
1077 r#"{"code":"upstream_http_error","message":"Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers. reset reason: connection timeout","upstream_status":500}"#.to_string(),
1078 None,
1079 );
1080
1081 match error {
1082 LanguageModelCompletionError::ApiInternalServerError { provider, message } => {
1083 assert_eq!(provider.0, "anthropic");
1084 assert_eq!(
1085 message,
1086 "Received an error from the Anthropic API: upstream connect error or disconnect/reset before headers. reset reason: connection timeout"
1087 );
1088 }
1089 _ => panic!(
1090 "Expected ApiInternalServerError for connection timeout with 500 status, got: {:?}",
1091 error
1092 ),
1093 }
1094 }
1095
1096 #[test]
1097 fn test_language_model_tool_use_serializes_with_signature() {
1098 use serde_json::json;
1099
1100 let tool_use = LanguageModelToolUse {
1101 id: LanguageModelToolUseId::from("test_id"),
1102 name: "test_tool".into(),
1103 raw_input: json!({"arg": "value"}).to_string(),
1104 input: json!({"arg": "value"}),
1105 is_input_complete: true,
1106 thought_signature: Some("test_signature".to_string()),
1107 };
1108
1109 let serialized = serde_json::to_value(&tool_use).unwrap();
1110
1111 assert_eq!(serialized["id"], "test_id");
1112 assert_eq!(serialized["name"], "test_tool");
1113 assert_eq!(serialized["thought_signature"], "test_signature");
1114 }
1115
1116 #[test]
1117 fn test_language_model_tool_use_deserializes_with_missing_signature() {
1118 use serde_json::json;
1119
1120 let json = json!({
1121 "id": "test_id",
1122 "name": "test_tool",
1123 "raw_input": "{\"arg\":\"value\"}",
1124 "input": {"arg": "value"},
1125 "is_input_complete": true
1126 });
1127
1128 let tool_use: LanguageModelToolUse = serde_json::from_value(json).unwrap();
1129
1130 assert_eq!(tool_use.id, LanguageModelToolUseId::from("test_id"));
1131 assert_eq!(tool_use.name.as_ref(), "test_tool");
1132 assert_eq!(tool_use.thought_signature, None);
1133 }
1134
1135 #[test]
1136 fn test_language_model_tool_use_round_trip_with_signature() {
1137 use serde_json::json;
1138
1139 let original = LanguageModelToolUse {
1140 id: LanguageModelToolUseId::from("round_trip_id"),
1141 name: "round_trip_tool".into(),
1142 raw_input: json!({"key": "value"}).to_string(),
1143 input: json!({"key": "value"}),
1144 is_input_complete: true,
1145 thought_signature: Some("round_trip_sig".to_string()),
1146 };
1147
1148 let serialized = serde_json::to_value(&original).unwrap();
1149 let deserialized: LanguageModelToolUse = serde_json::from_value(serialized).unwrap();
1150
1151 assert_eq!(deserialized.id, original.id);
1152 assert_eq!(deserialized.name, original.name);
1153 assert_eq!(deserialized.thought_signature, original.thought_signature);
1154 }
1155
1156 #[test]
1157 fn test_language_model_tool_use_round_trip_without_signature() {
1158 use serde_json::json;
1159
1160 let original = LanguageModelToolUse {
1161 id: LanguageModelToolUseId::from("no_sig_id"),
1162 name: "no_sig_tool".into(),
1163 raw_input: json!({"arg": "value"}).to_string(),
1164 input: json!({"arg": "value"}),
1165 is_input_complete: true,
1166 thought_signature: None,
1167 };
1168
1169 let serialized = serde_json::to_value(&original).unwrap();
1170 let deserialized: LanguageModelToolUse = serde_json::from_value(serialized).unwrap();
1171
1172 assert_eq!(deserialized.id, original.id);
1173 assert_eq!(deserialized.name, original.name);
1174 assert_eq!(deserialized.thought_signature, None);
1175 }
1176}