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