llm/providers/bedrock/
provider.rs

1use super::mantle::{MantleAuth, MantleClient};
2use super::mappers::{default_cache_point, map_messages, map_tools};
3use super::streaming::process_bedrock_stream;
4use crate::catalog::transport::ModelTransport;
5use crate::provider::{LlmResponseStream, ProviderFactory, StreamingModelProvider, get_context_window, stream_from};
6use crate::providers::openai_responses::transport::process_connection;
7use crate::{Context, LlmError, ProviderAuthMode, ProviderConnectionConfig, ProviderError, Result};
8use aws_config::Region;
9use aws_credential_types::provider::SharedCredentialsProvider;
10use aws_sdk_bedrockruntime::config::{BehaviorVersion, Credentials};
11use aws_sdk_bedrockruntime::error::SdkError;
12use aws_sdk_bedrockruntime::operation::converse_stream::ConverseStreamError;
13use aws_sdk_bedrockruntime::primitives::event_stream::EventReceiver;
14use aws_sdk_bedrockruntime::types::error::ConverseStreamOutputError;
15use aws_sdk_bedrockruntime::types::{ConverseStreamOutput, InferenceConfiguration};
16use aws_sdk_bedrockruntime::{Client, Config};
17use tracing::{error, info, warn};
18
19const DEFAULT_MODEL: &str = "anthropic.claude-sonnet-4-5-20250929-v1:0";
20const DEFAULT_MAX_TOKENS: i32 = 16_384;
21const DEFAULT_REGION: &str = "us-east-1";
22
23/// AWS credentials for explicit authentication with Bedrock.
24#[derive(Clone)]
25pub struct AwsCredentials {
26    pub access_key_id: String,
27    pub secret_access_key: String,
28    pub session_token: Option<String>,
29}
30
31#[derive(Clone)]
32pub struct BedrockProvider {
33    client: Client,
34    mantle: MantleClient,
35    model: String,
36    inference_profile_arn: Option<String>,
37}
38
39impl BedrockProvider {
40    /// Create a provider using the default AWS credential chain
41    /// (env vars, `~/.aws/credentials`, IAM roles, SSO).
42    pub async fn new(connection: ProviderConnectionConfig) -> Self {
43        if connection.auth_mode == ProviderAuthMode::None {
44            return Self::from_config(None, region_from_env().as_deref(), connection);
45        }
46
47        let mut loader = aws_config::defaults(BehaviorVersion::latest());
48        if let Some(url) = &connection.base_url {
49            loader = loader.endpoint_url(url.clone());
50        }
51
52        let config = loader.load().await;
53        let region = config
54            .region()
55            .map(ToString::to_string)
56            .or_else(region_from_env)
57            .unwrap_or_else(|| DEFAULT_REGION.to_string());
58        let auth = mantle_auth(config.credentials_provider(), &region);
59        Self::assemble(Client::new(&config), region, auth, connection)
60    }
61
62    /// Create a provider from explicit configuration without async credential discovery.
63    pub fn from_config(
64        credentials: Option<AwsCredentials>,
65        region: Option<&str>,
66        connection: ProviderConnectionConfig,
67    ) -> Self {
68        let region = region.unwrap_or(DEFAULT_REGION).to_string();
69        let auth = match connection.auth_mode {
70            ProviderAuthMode::Default => mantle_auth(credentials.clone().map(shared_credentials), &region),
71            ProviderAuthMode::None => MantleAuth::None,
72        };
73        let client = build_client(credentials, &region, connection.base_url.as_deref(), connection.auth_mode);
74        Self::assemble(client, region, auth, connection)
75    }
76
77    fn assemble(client: Client, region: String, auth: MantleAuth, connection: ProviderConnectionConfig) -> Self {
78        let mantle = MantleClient::new(region, auth, connection.base_url.clone());
79        Self {
80            client,
81            mantle,
82            model: DEFAULT_MODEL.to_string(),
83            inference_profile_arn: connection.inference_profile_arn,
84        }
85    }
86
87    pub fn with_model(mut self, model: &str) -> Self {
88        self.model = model.to_string();
89        self
90    }
91
92    pub fn with_inference_profile_arn(mut self, arn: impl Into<String>) -> Self {
93        self.inference_profile_arn = Some(arn.into());
94        self
95    }
96
97    /// Authenticate Responses-transport requests with a Bedrock API key instead
98    /// of `SigV4`. Equivalent to setting `AWS_BEARER_TOKEN_BEDROCK`.
99    pub fn with_bearer_token(mut self, token: impl Into<String>) -> Self {
100        self.mantle = self.mantle.with_auth(MantleAuth::BearerToken(token.into()));
101        self
102    }
103
104    fn request_model_id(&self) -> &str {
105        self.inference_profile_arn.as_deref().unwrap_or(&self.model)
106    }
107
108    /// The Responses-API transport for the current model, when the catalog says
109    /// it is not served by the Converse API.
110    fn mantle_transport(&self) -> Option<ModelTransport> {
111        self.model().and_then(|model| model.transport())
112    }
113
114    async fn send_converse_stream(
115        &self,
116        context: &Context,
117    ) -> Result<EventReceiver<ConverseStreamOutput, ConverseStreamOutputError>> {
118        let cache_point =
119            self.model().is_some_and(|m| m.supports_prompt_caching()).then(default_cache_point).transpose()?;
120        let (system_blocks, messages) = map_messages(context.messages(), cache_point.as_ref())?;
121        let settings = context.model_settings();
122        let max_tokens = settings.max_tokens.and_then(|m| i32::try_from(m).ok()).unwrap_or(DEFAULT_MAX_TOKENS);
123        let mut inference_config = InferenceConfiguration::builder().max_tokens(max_tokens);
124
125        if let Some(temp) = settings.temperature {
126            inference_config = inference_config.temperature(temp);
127        }
128
129        if let Some(top_p) = settings.top_p {
130            inference_config = inference_config.top_p(top_p);
131        }
132
133        let inference_config = inference_config.build();
134
135        let mut request = self
136            .client
137            .converse_stream()
138            .model_id(self.request_model_id())
139            .set_messages(Some(messages))
140            .inference_config(inference_config);
141
142        if !system_blocks.is_empty() {
143            request = request.set_system(Some(system_blocks));
144        }
145
146        if !context.tools().is_empty() {
147            let tool_config = map_tools(context.tools(), cache_point.as_ref())?;
148            request = request.tool_config(tool_config);
149        }
150
151        if let Some(arn) = self.inference_profile_arn.as_deref() {
152            info!(model = %self.model, inference_profile_arn = %arn, "Sending Bedrock converse_stream request");
153        } else {
154            info!(model = %self.model, "Sending Bedrock converse_stream request");
155        }
156
157        let response = request.send().await.map_err(|e| {
158            error!(model = %self.model, error = ?e, "Bedrock API error");
159            LlmError::from(e)
160        })?;
161
162        Ok(response.stream)
163    }
164}
165
166impl ProviderFactory for BedrockProvider {
167    async fn from_env() -> Result<Self> {
168        Ok(Self::new(ProviderConnectionConfig::default()).await)
169    }
170
171    async fn from_env_with_connection(connection: ProviderConnectionConfig) -> Result<Self> {
172        Ok(Self::new(connection).await)
173    }
174
175    fn with_model(self, model: &str) -> Self {
176        self.with_model(model)
177    }
178}
179
180impl StreamingModelProvider for BedrockProvider {
181    fn model(&self) -> Option<crate::LlmModel> {
182        format!("bedrock:{}", self.model).parse().ok()
183    }
184
185    fn context_window(&self) -> Option<u32> {
186        get_context_window("bedrock", &self.model)
187    }
188
189    fn stream_response(&self, context: &Context) -> LlmResponseStream {
190        let provider = self.clone();
191        let context = context.clone();
192
193        let Some(transport) = self.mantle_transport() else {
194            return stream_from(async move { provider.send_converse_stream(&context).await }, process_bedrock_stream);
195        };
196
197        if let Some(arn) = self.inference_profile_arn.as_deref() {
198            warn!(
199                model = %self.model,
200                inference_profile_arn = %arn,
201                "Ignoring inferenceProfileArn: this model is served by the Responses API, which has no inference profiles"
202            );
203        }
204
205        stream_from(
206            async move { provider.mantle.stream(&provider.model, &transport, &context).await },
207            process_connection,
208        )
209    }
210
211    fn display_name(&self) -> String {
212        format!("Bedrock ({})", self.model)
213    }
214}
215
216impl From<SdkError<ConverseStreamError>> for LlmError {
217    fn from(e: SdkError<ConverseStreamError>) -> Self {
218        let message = format!("Bedrock API error: {e}");
219        let status = e.raw_response().map(|r| r.status().as_u16());
220        let request_id = e.raw_response().and_then(|r| r.headers().get("x-amzn-requestid")).map(str::to_string);
221        let mut provider = match &e {
222            SdkError::TimeoutError(_) => ProviderError::timeout(message),
223            SdkError::DispatchFailure(_) => ProviderError::network(message),
224            SdkError::ResponseError(_) => ProviderError::server(message),
225            SdkError::ServiceError(svc) => {
226                let inner = svc.err();
227                if inner.is_throttling_exception() {
228                    ProviderError::rate_limit(message)
229                } else if inner.is_service_unavailable_exception()
230                    || inner.is_internal_server_exception()
231                    || inner.is_model_stream_error_exception()
232                {
233                    ProviderError::server(message)
234                } else {
235                    ProviderError::api(message)
236                }
237            }
238            _ => ProviderError::api(message),
239        };
240        provider = provider.with_http_metadata(status, request_id);
241        Self::from(provider)
242    }
243}
244
245fn build_client(
246    credentials: Option<AwsCredentials>,
247    region: &str,
248    base_url: Option<&str>,
249    auth_mode: ProviderAuthMode,
250) -> Client {
251    let mut config =
252        Config::builder().behavior_version(BehaviorVersion::latest()).region(Region::new(region.to_string()));
253
254    if auth_mode == ProviderAuthMode::None {
255        config = config.allow_no_auth();
256    } else if let Some(credentials) = credentials {
257        config = config.credentials_provider(shared_credentials(credentials));
258    }
259    if let Some(url) = base_url {
260        config = config.endpoint_url(url);
261    }
262
263    Client::from_conf(config.build())
264}
265
266fn shared_credentials(credentials: AwsCredentials) -> SharedCredentialsProvider {
267    SharedCredentialsProvider::new(Credentials::new(
268        credentials.access_key_id,
269        credentials.secret_access_key,
270        credentials.session_token,
271        None,
272        "aether-bedrock-provider",
273    ))
274}
275
276/// Resolve the credential scheme for the Responses transport.
277///
278/// A Bedrock API key takes precedence over the credential chain because it is
279/// the scheme the model catalog advertises for these endpoints; `SigV4` keeps
280/// SSO and IAM-role users working without extra configuration.
281fn mantle_auth(credentials: Option<SharedCredentialsProvider>, region: &str) -> MantleAuth {
282    if let Some(token) = MantleAuth::bearer_token_from_env() {
283        return MantleAuth::BearerToken(token);
284    }
285    match credentials {
286        Some(credentials) => MantleAuth::SigV4 { credentials, region: region.to_string() },
287        None => MantleAuth::None,
288    }
289}
290
291fn region_from_env() -> Option<String> {
292    ["AWS_REGION", "AWS_DEFAULT_REGION"].into_iter().find_map(|name| match std::env::var(name) {
293        Ok(value) if !value.is_empty() => Some(value),
294        _ => None,
295    })
296}
297
298#[cfg(test)]
299mod tests {
300    use super::*;
301    use crate::catalog::Provider;
302    use crate::providers::test_capture_server::CaptureServer;
303    use crate::types::IsoString;
304    use crate::{AssistantReasoning, ChatMessage, EncryptedReasoningContent, LlmModel};
305    use axum::Router;
306    use axum::body::Body;
307    use axum::extract::State;
308    use axum::http::{HeaderMap, Method, Request, StatusCode};
309    use axum::response::IntoResponse;
310    use axum::routing::any;
311    use futures::StreamExt;
312    use std::sync::Arc;
313    use tokio::net::TcpListener;
314    use tokio::sync::{Mutex, oneshot};
315
316    fn inference_profile_arn(model: &str) -> String {
317        format!("arn:aws:bedrock:us-west-2:000000000000:inference-profile/{model}")
318    }
319
320    fn application_inference_profile_arn() -> &'static str {
321        "arn:aws:bedrock:us-west-2:000000000000:application-inference-profile/000000000000"
322    }
323
324    fn test_provider() -> BedrockProvider {
325        BedrockProvider::from_config(None, None, ProviderConnectionConfig::default())
326    }
327
328    /// A catalog model routed to the Responses transport, resolved from the
329    /// catalog so a models.dev sync that retires one model does not quietly
330    /// leave these tests exercising the Converse path instead.
331    fn mantle_model() -> String {
332        LlmModel::all()
333            .iter()
334            .find(|model| model.provider_enum() == Provider::Bedrock && model.transport().is_some())
335            .expect("catalog must expose at least one Responses-transport Bedrock model")
336            .model_id()
337            .to_string()
338    }
339
340    /// A provider talking to `server` over the Responses transport, unauthenticated.
341    async fn mantle_provider(server: &CaptureServer) -> BedrockProvider {
342        BedrockProvider::new(ProviderConnectionConfig {
343            base_url: Some(server.base_url.clone()),
344            auth_mode: ProviderAuthMode::None,
345            ..Default::default()
346        })
347        .await
348        .with_model(&mantle_model())
349    }
350
351    #[test]
352    fn test_display_name() {
353        assert_eq!(test_provider().display_name(), "Bedrock (anthropic.claude-sonnet-4-5-20250929-v1:0)");
354    }
355
356    #[test]
357    fn test_with_model() {
358        let provider = test_provider().with_model("anthropic.claude-opus-4-20250514-v1:0");
359        assert_eq!(provider.display_name(), "Bedrock (anthropic.claude-opus-4-20250514-v1:0)");
360    }
361
362    #[test]
363    fn test_default_values() {
364        let provider = test_provider();
365        assert_eq!(provider.model, "anthropic.claude-sonnet-4-5-20250929-v1:0");
366    }
367
368    #[tokio::test]
369    async fn auth_none_sends_unsigned_request_to_custom_endpoint() {
370        let endpoint = FakeBedrockEndpoint::start().await;
371        let provider = BedrockProvider::new(ProviderConnectionConfig {
372            base_url: Some(endpoint.url.clone()),
373            auth_mode: ProviderAuthMode::None,
374            ..Default::default()
375        })
376        .await;
377
378        let result = provider.send_converse_stream(&hello_context()).await;
379        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
380
381        assert!(result.is_err());
382        assert_eq!(request.method, Method::POST);
383        assert!(request.path.starts_with("/model/"), "{}", request.path);
384        assert!(!request.headers.contains_key("authorization"), "request was signed: {:?}", request.headers);
385        assert!(
386            !request.headers.contains_key("x-amz-security-token"),
387            "request included session token: {:?}",
388            request.headers
389        );
390    }
391
392    fn static_credentials() -> AwsCredentials {
393        AwsCredentials {
394            access_key_id: "AKIAIOSFODNN7EXAMPLE".to_string(),
395            secret_access_key: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY".to_string(),
396            session_token: None,
397        }
398    }
399
400    fn hello_context() -> Context {
401        Context::new(vec![ChatMessage::user("Hello")], vec![])
402    }
403
404    #[tokio::test]
405    async fn responses_shape_models_are_sent_to_the_responses_endpoint() {
406        let mut server = CaptureServer::start_responses().await;
407        let provider = mantle_provider(&server).await;
408        let mut context = hello_context();
409        context.set_reasoning_effort(Some(crate::ReasoningEffort::Xhigh));
410
411        let responses = provider.stream_response(&context).collect::<Vec<_>>().await;
412        let captured = server.captured().await;
413
414        assert!(!responses.is_empty());
415        assert_eq!(captured.path, "/responses");
416        assert_eq!(captured.body["model"], mantle_model());
417        assert_eq!(captured.body["stream"], true);
418        assert_eq!(captured.body["store"], false);
419        assert_eq!(captured.body["reasoning"]["effort"], "xhigh");
420        assert_eq!(captured.body["input"][0]["role"], "user");
421        assert_eq!(captured.body["input"][0]["content"][0]["type"], "input_text");
422        assert_eq!(captured.body["input"][0]["content"][0]["text"], "Hello");
423        assert!(captured.headers.get("authorization").is_none(), "{:?}", captured.headers);
424    }
425
426    #[tokio::test]
427    async fn http_200_failed_server_error_is_retryable_with_diagnostics() {
428        use crate::providers::test_capture_server::ResponseSpec;
429        let spec = ResponseSpec::sse(include_str!("../../../tests/fixtures/openai_responses/04_failed_server.sse"))
430            .with_header("x-amzn-requestid", "amzn-req-123");
431        let mut server = CaptureServer::start_with_spec(spec).await;
432        let provider = mantle_provider(&server).await;
433
434        let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
435        let _ = server.captured().await;
436
437        assert!(!responses.iter().any(|r| matches!(r, Ok(crate::LlmResponse::Done { .. }))));
438        let err = responses.iter().find_map(|r| r.as_ref().err()).expect("expected a failure");
439        assert!(err.is_retryable(), "server_error must be retryable: {err:?}");
440        let provider_error = err.provider().expect("expected provider error");
441        assert_eq!(provider_error.kind, crate::ProviderErrorKind::Server);
442        assert_eq!(provider_error.code.as_deref(), Some("server_error"));
443        assert_eq!(provider_error.http_status, Some(200));
444        assert_eq!(provider_error.request_id.as_deref(), Some("amzn-req-123"));
445    }
446
447    #[tokio::test]
448    async fn failed_event_with_unknown_code_is_terminal_without_done() {
449        let body = "event: response.created\ndata: {\"type\":\"response.created\",\"response\":{\"id\":\"r1\"}}\n\nevent: response.failed\ndata: {\"type\":\"response.failed\",\"response\":{\"error\":{\"code\":\"invalid_prompt\",\"message\":\"bad prompt\"}}}\n\n";
450        let mut server = CaptureServer::start_with_response(body).await;
451        let provider = mantle_provider(&server).await;
452
453        let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
454        let _ = server.captured().await;
455
456        assert!(!responses.iter().any(|r| matches!(r, Ok(crate::LlmResponse::Done { .. }))));
457        let err = responses.iter().find_map(|r| r.as_ref().err()).expect("expected a failure");
458        assert!(!err.is_retryable(), "invalid_prompt must be terminal: {err:?}");
459    }
460
461    #[tokio::test]
462    async fn responses_transport_rejects_malformed_and_truncated_streams() {
463        for response in [
464            "data: {not-json}\n\n",
465            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_1\"}}\n\ndata: [DONE]\n\n",
466        ] {
467            let mut server = CaptureServer::start_with_response(response).await;
468            let provider = mantle_provider(&server).await;
469
470            let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
471            let _ = server.captured().await;
472
473            assert!(responses.iter().any(|response| {
474                response.as_ref().err().and_then(LlmError::provider).map(|provider| provider.kind)
475                    == Some(crate::ProviderErrorKind::StreamInterrupted)
476            }));
477            assert!(!responses.iter().any(|response| matches!(response, Ok(crate::LlmResponse::Done { .. }))));
478        }
479    }
480
481    #[tokio::test]
482    async fn responses_transport_drops_encrypted_reasoning_from_another_model() {
483        let mut server = CaptureServer::start_responses().await;
484        let provider = mantle_provider(&server).await;
485        let context = Context::new(
486            vec![ChatMessage::Assistant {
487                content: "previous answer".to_string(),
488                reasoning: AssistantReasoning {
489                    summary_text: None,
490                    encrypted_content: Some(EncryptedReasoningContent {
491                        id: "reasoning-id".to_string(),
492                        model: "bedrock:openai.gpt-5.5".parse().unwrap(),
493                        content: "opaque-for-another-model".to_string(),
494                    }),
495                },
496                timestamp: IsoString::now(),
497                tool_calls: vec![],
498            }],
499            vec![],
500        );
501
502        let _ = provider.stream_response(&context).collect::<Vec<_>>().await;
503        let captured = server.captured().await;
504
505        assert!(
506            captured.body["input"].as_array().unwrap().iter().all(|item| item["type"] != "reasoning"),
507            "{}",
508            captured.body
509        );
510    }
511
512    #[tokio::test]
513    async fn converse_shape_models_do_not_use_the_responses_endpoint() {
514        let endpoint = FakeBedrockEndpoint::start().await;
515        let provider = BedrockProvider::new(ProviderConnectionConfig {
516            base_url: Some(endpoint.url.clone()),
517            auth_mode: ProviderAuthMode::None,
518            ..Default::default()
519        })
520        .await
521        .with_model(DEFAULT_MODEL);
522
523        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
524        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
525
526        assert!(request.path.starts_with("/model/"), "{}", request.path);
527    }
528
529    #[tokio::test]
530    async fn bearer_token_authenticates_responses_requests() {
531        let mut server = CaptureServer::start_responses().await;
532        let provider = BedrockProvider::from_config(
533            None,
534            Some("us-west-2"),
535            ProviderConnectionConfig { base_url: Some(server.base_url.clone()), ..Default::default() },
536        )
537        .with_bearer_token("test-token")
538        .with_model(&mantle_model());
539
540        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
541        let captured = server.captured().await;
542
543        assert_eq!(captured.headers.get("authorization").unwrap(), "Bearer test-token");
544    }
545
546    #[tokio::test]
547    async fn credential_chain_sigv4_signs_responses_requests() {
548        let mut server = CaptureServer::start_responses().await;
549        let provider = BedrockProvider::from_config(
550            Some(static_credentials()),
551            Some("us-west-2"),
552            ProviderConnectionConfig { base_url: Some(server.base_url.clone()), ..Default::default() },
553        )
554        .with_model(&mantle_model());
555
556        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
557        let captured = server.captured().await;
558
559        let authorization = captured.headers.get("authorization").expect("request was not signed").to_str().unwrap();
560        assert!(
561            authorization.starts_with("AWS4-HMAC-SHA256 Credential=AKIAIOSFODNN7EXAMPLE/"),
562            "unexpected authorization header: {authorization}"
563        );
564        assert!(authorization.contains("/us-west-2/bedrock/aws4_request"), "{authorization}");
565        assert!(captured.headers.contains_key("x-amz-date"), "{:?}", captured.headers);
566    }
567
568    #[test]
569    fn only_responses_shape_models_route_to_the_mantle_transport() {
570        let mantle = test_provider().with_model(&mantle_model());
571        let converse = test_provider().with_model(DEFAULT_MODEL);
572        let profile = test_provider().with_model("us.anthropic.claude-future-model-v99:0");
573
574        assert!(matches!(mantle.mantle_transport(), Some(ModelTransport::OpenAiResponses { .. })));
575        assert_eq!(converse.mantle_transport(), None);
576        assert_eq!(profile.mantle_transport(), None);
577    }
578
579    #[test]
580    fn explicit_connection_preserves_inference_profile() {
581        let provider = BedrockProvider::from_config(
582            None,
583            Some("us-west-2"),
584            ProviderConnectionConfig { inference_profile_arn: Some("arn:test".to_string()), ..Default::default() },
585        );
586
587        assert_eq!(provider.inference_profile_arn.as_deref(), Some("arn:test"));
588    }
589
590    #[test]
591    fn test_from_config_with_credentials() {
592        let provider =
593            BedrockProvider::from_config(Some(static_credentials()), None, ProviderConnectionConfig::default());
594        assert_eq!(provider.model, DEFAULT_MODEL);
595    }
596
597    #[test]
598    fn test_from_config_with_credentials_and_region() {
599        let credentials =
600            AwsCredentials { session_token: Some("FwoGZXIvYXdzEBYaD...".to_string()), ..static_credentials() };
601
602        let provider =
603            BedrockProvider::from_config(Some(credentials), Some("us-west-2"), ProviderConnectionConfig::default())
604                .with_model("anthropic.claude-opus-4-20250514-v1:0");
605
606        assert_eq!(provider.model, "anthropic.claude-opus-4-20250514-v1:0");
607    }
608
609    #[test]
610    fn test_from_config_with_region_only() {
611        let provider = BedrockProvider::from_config(None, Some("eu-west-1"), ProviderConnectionConfig::default());
612        assert_eq!(provider.model, DEFAULT_MODEL);
613    }
614
615    #[test]
616    fn catalog_foundation_id_resolves_context_window() {
617        let provider = test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0");
618        assert!(provider.context_window().is_some());
619        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
620    }
621
622    #[test]
623    fn cross_region_profile_id_in_catalog_resolves() {
624        let provider = test_provider().with_model("us.anthropic.claude-opus-4-6-v1");
625        assert!(provider.context_window().is_some());
626    }
627
628    #[test]
629    fn unknown_cross_region_profile_id_falls_through_to_profile() {
630        let id = "us.anthropic.claude-future-model-v99:0";
631        let provider = test_provider().with_model(id);
632        assert_eq!(provider.context_window(), None);
633        assert_eq!(provider.model().unwrap().to_string(), format!("bedrock:{id}"));
634        assert_eq!(provider.display_name(), format!("Bedrock ({id})"));
635    }
636
637    #[tokio::test]
638    async fn separate_inference_profile_arn_is_used_as_request_model_id() {
639        let endpoint = FakeBedrockEndpoint::start().await;
640        let provider = BedrockProvider::new(ProviderConnectionConfig {
641            base_url: Some(endpoint.url.clone()),
642            auth_mode: ProviderAuthMode::None,
643            request_model: None,
644            inference_profile_arn: Some(application_inference_profile_arn().to_string()),
645        })
646        .await
647        .with_model(DEFAULT_MODEL);
648
649        let result = provider.send_converse_stream(&hello_context()).await;
650        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
651
652        assert!(result.is_err());
653        assert!(
654            request.path.contains("arn%3Aaws%3Abedrock%3Aus-west-2%3A000000000000%3Aapplication-inference-profile"),
655            "{}",
656            request.path
657        );
658        assert_eq!(provider.context_window(), Some(200_000));
659        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
660    }
661
662    #[test]
663    fn with_inference_profile_arn_keeps_canonical_model_identity() {
664        let arn = inference_profile_arn("us.anthropic.claude-sonnet-4-5-20250929-v1:0");
665        let provider =
666            test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0").with_inference_profile_arn(&arn);
667
668        assert_eq!(provider.request_model_id(), arn);
669        assert_eq!(provider.context_window(), Some(200_000));
670        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
671    }
672
673    #[test]
674    fn prompt_caching_support_comes_from_canonical_model() {
675        let cached = test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0");
676        assert!(cached.model().unwrap().supports_prompt_caching());
677
678        let unknown_profile = test_provider().with_model("us.anthropic.claude-future-model-v99:0");
679        assert!(!unknown_profile.model().unwrap().supports_prompt_caching());
680    }
681
682    struct FakeBedrockEndpoint {
683        url: String,
684        request: oneshot::Receiver<CapturedRequest>,
685    }
686
687    struct CapturedRequest {
688        method: Method,
689        path: String,
690        headers: HeaderMap,
691    }
692
693    #[derive(Clone)]
694    struct FakeBedrockState {
695        request_tx: Arc<Mutex<Option<oneshot::Sender<CapturedRequest>>>>,
696        shutdown_tx: Arc<Mutex<Option<oneshot::Sender<()>>>>,
697    }
698
699    impl FakeBedrockEndpoint {
700        async fn start() -> Self {
701            let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind fake Bedrock endpoint");
702            let url = format!("http://{}", listener.local_addr().expect("fake Bedrock endpoint address"));
703            let (request_tx, request) = oneshot::channel();
704            let (shutdown_tx, shutdown) = oneshot::channel();
705            let state = FakeBedrockState {
706                request_tx: Arc::new(Mutex::new(Some(request_tx))),
707                shutdown_tx: Arc::new(Mutex::new(Some(shutdown_tx))),
708            };
709
710            let app = Router::new().fallback(any(capture_bedrock_request)).with_state(state);
711            tokio::spawn(async move {
712                axum::serve(listener, app)
713                    .with_graceful_shutdown(async {
714                        let _ = shutdown.await;
715                    })
716                    .await
717                    .expect("serve fake Bedrock endpoint");
718            });
719
720            Self { url, request }
721        }
722    }
723
724    async fn capture_bedrock_request(
725        State(state): State<FakeBedrockState>,
726        request: Request<Body>,
727    ) -> impl IntoResponse {
728        let (parts, _) = request.into_parts();
729        if let Some(tx) = state.request_tx.lock().await.take() {
730            let _ = tx.send(CapturedRequest {
731                method: parts.method,
732                path: parts.uri.path().to_string(),
733                headers: parts.headers,
734            });
735        }
736        if let Some(tx) = state.shutdown_tx.lock().await.take() {
737            let _ = tx.send(());
738        }
739        (StatusCode::FORBIDDEN, "{}")
740    }
741}
llm/providers/bedrock/provider.rs

llm/providers/bedrock/
provider.rs
