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        if let Err(error) = crate::provider::validate_reasoning(context, self.model().as_ref()) {
191            return crate::provider::error_stream(error);
192        }
193        let provider = self.clone();
194        let context = context.clone();
195
196        let Some(transport) = self.mantle_transport() else {
197            return stream_from(async move { provider.send_converse_stream(&context).await }, process_bedrock_stream);
198        };
199
200        if let Some(arn) = self.inference_profile_arn.as_deref() {
201            warn!(
202                model = %self.model,
203                inference_profile_arn = %arn,
204                "Ignoring inferenceProfileArn: this model is served by the Responses API, which has no inference profiles"
205            );
206        }
207
208        stream_from(
209            async move { provider.mantle.stream(&provider.model, &transport, &context).await },
210            process_connection,
211        )
212    }
213
214    fn display_name(&self) -> String {
215        format!("Bedrock ({})", self.model)
216    }
217}
218
219impl From<SdkError<ConverseStreamError>> for LlmError {
220    fn from(e: SdkError<ConverseStreamError>) -> Self {
221        let message = format!("Bedrock API error: {e}");
222        let status = e.raw_response().map(|r| r.status().as_u16());
223        let request_id = e.raw_response().and_then(|r| r.headers().get("x-amzn-requestid")).map(str::to_string);
224        let mut provider = match &e {
225            SdkError::TimeoutError(_) => ProviderError::timeout(message),
226            SdkError::DispatchFailure(_) => ProviderError::network(message),
227            SdkError::ResponseError(_) => ProviderError::server(message),
228            SdkError::ServiceError(svc) => {
229                let inner = svc.err();
230                if inner.is_throttling_exception() {
231                    ProviderError::rate_limit(message)
232                } else if inner.is_service_unavailable_exception()
233                    || inner.is_internal_server_exception()
234                    || inner.is_model_stream_error_exception()
235                {
236                    ProviderError::server(message)
237                } else {
238                    ProviderError::api(message)
239                }
240            }
241            _ => ProviderError::api(message),
242        };
243        provider = provider.with_http_metadata(status, request_id);
244        Self::from(provider)
245    }
246}
247
248fn build_client(
249    credentials: Option<AwsCredentials>,
250    region: &str,
251    base_url: Option<&str>,
252    auth_mode: ProviderAuthMode,
253) -> Client {
254    let mut config =
255        Config::builder().behavior_version(BehaviorVersion::latest()).region(Region::new(region.to_string()));
256
257    if auth_mode == ProviderAuthMode::None {
258        config = config.allow_no_auth();
259    } else if let Some(credentials) = credentials {
260        config = config.credentials_provider(shared_credentials(credentials));
261    }
262    if let Some(url) = base_url {
263        config = config.endpoint_url(url);
264    }
265
266    Client::from_conf(config.build())
267}
268
269fn shared_credentials(credentials: AwsCredentials) -> SharedCredentialsProvider {
270    SharedCredentialsProvider::new(Credentials::new(
271        credentials.access_key_id,
272        credentials.secret_access_key,
273        credentials.session_token,
274        None,
275        "aether-bedrock-provider",
276    ))
277}
278
279/// Resolve the credential scheme for the Responses transport.
280///
281/// A Bedrock API key takes precedence over the credential chain because it is
282/// the scheme the model catalog advertises for these endpoints; `SigV4` keeps
283/// SSO and IAM-role users working without extra configuration.
284fn mantle_auth(credentials: Option<SharedCredentialsProvider>, region: &str) -> MantleAuth {
285    if let Some(token) = MantleAuth::bearer_token_from_env() {
286        return MantleAuth::BearerToken(token);
287    }
288    match credentials {
289        Some(credentials) => MantleAuth::SigV4 { credentials, region: region.to_string() },
290        None => MantleAuth::None,
291    }
292}
293
294fn region_from_env() -> Option<String> {
295    ["AWS_REGION", "AWS_DEFAULT_REGION"].into_iter().find_map(|name| match std::env::var(name) {
296        Ok(value) if !value.is_empty() => Some(value),
297        _ => None,
298    })
299}
300
301#[cfg(test)]
302mod tests {
303    use super::*;
304    use crate::catalog::Provider;
305    use crate::providers::test_capture_server::CaptureServer;
306    use crate::types::IsoString;
307    use crate::{AssistantReasoning, ChatMessage, EncryptedReasoningContent, LlmModel, MessageId};
308    use axum::Router;
309    use axum::body::Body;
310    use axum::extract::State;
311    use axum::http::{HeaderMap, Method, Request, StatusCode};
312    use axum::response::IntoResponse;
313    use axum::routing::any;
314    use futures::StreamExt;
315    use std::sync::Arc;
316    use tokio::net::TcpListener;
317    use tokio::sync::{Mutex, oneshot};
318    use utils::ReasoningEffort;
319
320    fn inference_profile_arn(model: &str) -> String {
321        format!("arn:aws:bedrock:us-west-2:000000000000:inference-profile/{model}")
322    }
323
324    fn application_inference_profile_arn() -> &'static str {
325        "arn:aws:bedrock:us-west-2:000000000000:application-inference-profile/000000000000"
326    }
327
328    fn test_provider() -> BedrockProvider {
329        BedrockProvider::from_config(None, None, ProviderConnectionConfig::default())
330    }
331
332    /// A catalog model routed to the Responses transport, resolved from the
333    /// catalog so a models.dev sync that retires one model does not quietly
334    /// leave these tests exercising the Converse path instead.
335    fn mantle_model() -> String {
336        LlmModel::all()
337            .iter()
338            .find(|model| model.provider_enum() == Provider::Bedrock && model.transport().is_some())
339            .expect("catalog must expose at least one Responses-transport Bedrock model")
340            .model_id()
341            .to_string()
342    }
343
344    /// A provider talking to `server` over the Responses transport, unauthenticated.
345    async fn mantle_provider(server: &CaptureServer) -> BedrockProvider {
346        BedrockProvider::new(ProviderConnectionConfig {
347            base_url: Some(server.base_url.clone()),
348            auth_mode: ProviderAuthMode::None,
349            ..Default::default()
350        })
351        .await
352        .with_model(&mantle_model())
353    }
354
355    #[test]
356    fn test_display_name() {
357        assert_eq!(test_provider().display_name(), "Bedrock (anthropic.claude-sonnet-4-5-20250929-v1:0)");
358    }
359
360    #[test]
361    fn test_with_model() {
362        let provider = test_provider().with_model("anthropic.claude-opus-4-20250514-v1:0");
363        assert_eq!(provider.display_name(), "Bedrock (anthropic.claude-opus-4-20250514-v1:0)");
364    }
365
366    #[test]
367    fn test_default_values() {
368        let provider = test_provider();
369        assert_eq!(provider.model, "anthropic.claude-sonnet-4-5-20250929-v1:0");
370    }
371
372    #[tokio::test]
373    async fn auth_none_sends_unsigned_request_to_custom_endpoint() {
374        let endpoint = FakeBedrockEndpoint::start().await;
375        let provider = BedrockProvider::new(ProviderConnectionConfig {
376            base_url: Some(endpoint.url.clone()),
377            auth_mode: ProviderAuthMode::None,
378            ..Default::default()
379        })
380        .await;
381
382        let result = provider.send_converse_stream(&hello_context()).await;
383        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
384
385        assert!(result.is_err());
386        assert_eq!(request.method, Method::POST);
387        assert!(request.path.starts_with("/model/"), "{}", request.path);
388        assert!(!request.headers.contains_key("authorization"), "request was signed: {:?}", request.headers);
389        assert!(
390            !request.headers.contains_key("x-amz-security-token"),
391            "request included session token: {:?}",
392            request.headers
393        );
394    }
395
396    fn static_credentials() -> AwsCredentials {
397        AwsCredentials {
398            access_key_id: "AKIAIOSFODNN7EXAMPLE".to_string(),
399            secret_access_key: "wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY".to_string(),
400            session_token: None,
401        }
402    }
403
404    fn hello_context() -> Context {
405        Context::new(vec![ChatMessage::user("Hello")], vec![])
406    }
407
408    #[tokio::test]
409    async fn mantle_disabled_uses_none_without_summary() {
410        let model = LlmModel::all()
411            .iter()
412            .find(|model| {
413                model.provider_enum() == Provider::Bedrock
414                    && model.transport().is_some()
415                    && model.supports_reasoning_off()
416            })
417            .unwrap();
418        let mut server = CaptureServer::start_responses().await;
419        let provider = mantle_provider(&server).await.with_model(&model.model_id());
420        let mut context = hello_context();
421        context.set_reasoning_effort(ReasoningEffort::Disabled);
422        let responses = provider.stream_response(&context).collect::<Vec<_>>().await;
423        assert!(responses.iter().all(Result::is_ok), "{responses:?}");
424        let body = server.captured().await.body;
425        assert_eq!(body["reasoning"]["effort"], "none");
426        assert!(body["reasoning"]["summary"].is_null());
427    }
428
429    #[tokio::test]
430    async fn converse_disabled_is_a_non_retryable_transport_error() {
431        let model = LlmModel::all()
432            .iter()
433            .find(|model| {
434                model.provider_enum() == Provider::Bedrock
435                    && model.transport().is_none()
436                    && model.supports_reasoning_off()
437            })
438            .unwrap();
439        let provider =
440            BedrockProvider::new(ProviderConnectionConfig { auth_mode: ProviderAuthMode::None, ..Default::default() })
441                .await
442                .with_model(&model.model_id());
443        let mut context = hello_context();
444        context.set_reasoning_effort(crate::ReasoningEffort::Disabled);
445        let responses = provider.stream_response(&context).collect::<Vec<_>>().await;
446        assert_eq!(responses.len(), 1);
447        assert!(matches!(&responses[0], Err(LlmError::UnsupportedDisableTransport { .. })));
448        assert!(!responses[0].as_ref().unwrap_err().is_retryable());
449    }
450
451    #[tokio::test]
452    async fn responses_shape_models_are_sent_to_the_responses_endpoint() {
453        let mut server = CaptureServer::start_responses().await;
454        let provider = mantle_provider(&server).await;
455        let mut context = hello_context();
456        context.set_reasoning_effort(crate::ReasoningEffort::Xhigh);
457
458        let responses = provider.stream_response(&context).collect::<Vec<_>>().await;
459        let captured = server.captured().await;
460
461        assert!(!responses.is_empty());
462        assert_eq!(captured.path, "/responses");
463        assert_eq!(captured.body["model"], mantle_model());
464        assert_eq!(captured.body["stream"], true);
465        assert_eq!(captured.body["store"], false);
466        assert_eq!(captured.body["reasoning"]["effort"], "xhigh");
467        assert_eq!(captured.body["input"][0]["role"], "user");
468        assert_eq!(captured.body["input"][0]["content"][0]["type"], "input_text");
469        assert_eq!(captured.body["input"][0]["content"][0]["text"], "Hello");
470        assert!(captured.headers.get("authorization").is_none(), "{:?}", captured.headers);
471    }
472
473    #[tokio::test]
474    async fn http_200_failed_server_error_is_retryable_with_diagnostics() {
475        use crate::providers::test_capture_server::ResponseSpec;
476        let spec = ResponseSpec::sse(include_str!("../../../tests/fixtures/openai_responses/04_failed_server.sse"))
477            .with_header("x-amzn-requestid", "amzn-req-123");
478        let mut server = CaptureServer::start_with_spec(spec).await;
479        let provider = mantle_provider(&server).await;
480
481        let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
482        let _ = server.captured().await;
483
484        assert!(!responses.iter().any(|r| matches!(r, Ok(crate::LlmResponse::Done { .. }))));
485        let err = responses.iter().find_map(|r| r.as_ref().err()).expect("expected a failure");
486        assert!(err.is_retryable(), "server_error must be retryable: {err:?}");
487        let provider_error = err.provider().expect("expected provider error");
488        assert_eq!(provider_error.kind, crate::ProviderErrorKind::Server);
489        assert_eq!(provider_error.code.as_deref(), Some("server_error"));
490        assert_eq!(provider_error.http_status, Some(200));
491        assert_eq!(provider_error.request_id.as_deref(), Some("amzn-req-123"));
492    }
493
494    #[tokio::test]
495    async fn failed_event_with_unknown_code_is_terminal_without_done() {
496        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";
497        let mut server = CaptureServer::start_with_response(body).await;
498        let provider = mantle_provider(&server).await;
499
500        let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
501        let _ = server.captured().await;
502
503        assert!(!responses.iter().any(|r| matches!(r, Ok(crate::LlmResponse::Done { .. }))));
504        let err = responses.iter().find_map(|r| r.as_ref().err()).expect("expected a failure");
505        assert!(!err.is_retryable(), "invalid_prompt must be terminal: {err:?}");
506    }
507
508    #[tokio::test]
509    async fn responses_transport_rejects_malformed_and_truncated_streams() {
510        for response in [
511            "data: {not-json}\n\n",
512            "data: {\"type\":\"response.created\",\"response\":{\"id\":\"resp_1\"}}\n\ndata: [DONE]\n\n",
513        ] {
514            let mut server = CaptureServer::start_with_response(response).await;
515            let provider = mantle_provider(&server).await;
516
517            let responses = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
518            let _ = server.captured().await;
519
520            assert!(responses.iter().any(|response| {
521                response.as_ref().err().and_then(LlmError::provider).map(|provider| provider.kind)
522                    == Some(crate::ProviderErrorKind::StreamInterrupted)
523            }));
524            assert!(!responses.iter().any(|response| matches!(response, Ok(crate::LlmResponse::Done { .. }))));
525        }
526    }
527
528    #[tokio::test]
529    async fn responses_transport_drops_encrypted_reasoning_from_another_model() {
530        let mut server = CaptureServer::start_responses().await;
531        let provider = mantle_provider(&server).await;
532        let context = Context::new(
533            vec![ChatMessage::Assistant {
534                message_id: MessageId::new(),
535                content: "previous answer".to_string(),
536                reasoning: AssistantReasoning {
537                    summary_text: None,
538                    encrypted_content: Some(EncryptedReasoningContent {
539                        id: "reasoning-id".to_string(),
540                        model: "bedrock:openai.gpt-5.5".parse().unwrap(),
541                        content: "opaque-for-another-model".to_string(),
542                    }),
543                },
544                timestamp: IsoString::now(),
545                tool_calls: vec![],
546            }],
547            vec![],
548        );
549
550        let _ = provider.stream_response(&context).collect::<Vec<_>>().await;
551        let captured = server.captured().await;
552
553        assert!(
554            captured.body["input"].as_array().unwrap().iter().all(|item| item["type"] != "reasoning"),
555            "{}",
556            captured.body
557        );
558    }
559
560    #[tokio::test]
561    async fn converse_shape_models_do_not_use_the_responses_endpoint() {
562        let endpoint = FakeBedrockEndpoint::start().await;
563        let provider = BedrockProvider::new(ProviderConnectionConfig {
564            base_url: Some(endpoint.url.clone()),
565            auth_mode: ProviderAuthMode::None,
566            ..Default::default()
567        })
568        .await
569        .with_model(DEFAULT_MODEL);
570
571        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
572        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
573
574        assert!(request.path.starts_with("/model/"), "{}", request.path);
575    }
576
577    #[tokio::test]
578    async fn bearer_token_authenticates_responses_requests() {
579        let mut server = CaptureServer::start_responses().await;
580        let provider = BedrockProvider::from_config(
581            None,
582            Some("us-west-2"),
583            ProviderConnectionConfig { base_url: Some(server.base_url.clone()), ..Default::default() },
584        )
585        .with_bearer_token("test-token")
586        .with_model(&mantle_model());
587
588        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
589        let captured = server.captured().await;
590
591        assert_eq!(captured.headers.get("authorization").unwrap(), "Bearer test-token");
592    }
593
594    #[tokio::test]
595    async fn credential_chain_sigv4_signs_responses_requests() {
596        let mut server = CaptureServer::start_responses().await;
597        let provider = BedrockProvider::from_config(
598            Some(static_credentials()),
599            Some("us-west-2"),
600            ProviderConnectionConfig { base_url: Some(server.base_url.clone()), ..Default::default() },
601        )
602        .with_model(&mantle_model());
603
604        let _ = provider.stream_response(&hello_context()).collect::<Vec<_>>().await;
605        let captured = server.captured().await;
606
607        let authorization = captured.headers.get("authorization").expect("request was not signed").to_str().unwrap();
608        assert!(
609            authorization.starts_with("AWS4-HMAC-SHA256 Credential=AKIAIOSFODNN7EXAMPLE/"),
610            "unexpected authorization header: {authorization}"
611        );
612        assert!(authorization.contains("/us-west-2/bedrock/aws4_request"), "{authorization}");
613        assert!(captured.headers.contains_key("x-amz-date"), "{:?}", captured.headers);
614    }
615
616    #[test]
617    fn only_responses_shape_models_route_to_the_mantle_transport() {
618        let mantle = test_provider().with_model(&mantle_model());
619        let converse = test_provider().with_model(DEFAULT_MODEL);
620        let profile = test_provider().with_model("us.anthropic.claude-future-model-v99:0");
621
622        assert!(matches!(mantle.mantle_transport(), Some(ModelTransport::OpenAiResponses { .. })));
623        assert_eq!(converse.mantle_transport(), None);
624        assert_eq!(profile.mantle_transport(), None);
625    }
626
627    #[test]
628    fn explicit_connection_preserves_inference_profile() {
629        let provider = BedrockProvider::from_config(
630            None,
631            Some("us-west-2"),
632            ProviderConnectionConfig { inference_profile_arn: Some("arn:test".to_string()), ..Default::default() },
633        );
634
635        assert_eq!(provider.inference_profile_arn.as_deref(), Some("arn:test"));
636    }
637
638    #[test]
639    fn test_from_config_with_credentials() {
640        let provider =
641            BedrockProvider::from_config(Some(static_credentials()), None, ProviderConnectionConfig::default());
642        assert_eq!(provider.model, DEFAULT_MODEL);
643    }
644
645    #[test]
646    fn test_from_config_with_credentials_and_region() {
647        let credentials =
648            AwsCredentials { session_token: Some("FwoGZXIvYXdzEBYaD...".to_string()), ..static_credentials() };
649
650        let provider =
651            BedrockProvider::from_config(Some(credentials), Some("us-west-2"), ProviderConnectionConfig::default())
652                .with_model("anthropic.claude-opus-4-20250514-v1:0");
653
654        assert_eq!(provider.model, "anthropic.claude-opus-4-20250514-v1:0");
655    }
656
657    #[test]
658    fn test_from_config_with_region_only() {
659        let provider = BedrockProvider::from_config(None, Some("eu-west-1"), ProviderConnectionConfig::default());
660        assert_eq!(provider.model, DEFAULT_MODEL);
661    }
662
663    #[test]
664    fn catalog_foundation_id_resolves_context_window() {
665        let provider = test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0");
666        assert!(provider.context_window().is_some());
667        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
668    }
669
670    #[test]
671    fn cross_region_profile_id_in_catalog_resolves() {
672        let provider = test_provider().with_model("us.anthropic.claude-opus-4-6-v1");
673        assert!(provider.context_window().is_some());
674    }
675
676    #[test]
677    fn unknown_cross_region_profile_id_falls_through_to_profile() {
678        let id = "us.anthropic.claude-future-model-v99:0";
679        let provider = test_provider().with_model(id);
680        assert_eq!(provider.context_window(), None);
681        assert_eq!(provider.model().unwrap().to_string(), format!("bedrock:{id}"));
682        assert_eq!(provider.display_name(), format!("Bedrock ({id})"));
683    }
684
685    #[tokio::test]
686    async fn separate_inference_profile_arn_is_used_as_request_model_id() {
687        let endpoint = FakeBedrockEndpoint::start().await;
688        let provider = BedrockProvider::new(ProviderConnectionConfig {
689            base_url: Some(endpoint.url.clone()),
690            auth_mode: ProviderAuthMode::None,
691            request_model: None,
692            inference_profile_arn: Some(application_inference_profile_arn().to_string()),
693        })
694        .await
695        .with_model(DEFAULT_MODEL);
696
697        let result = provider.send_converse_stream(&hello_context()).await;
698        let request = endpoint.request.await.expect("fake Bedrock endpoint received no request");
699
700        assert!(result.is_err());
701        assert!(
702            request.path.contains("arn%3Aaws%3Abedrock%3Aus-west-2%3A000000000000%3Aapplication-inference-profile"),
703            "{}",
704            request.path
705        );
706        assert_eq!(provider.context_window(), Some(200_000));
707        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
708    }
709
710    #[test]
711    fn with_inference_profile_arn_keeps_canonical_model_identity() {
712        let arn = inference_profile_arn("us.anthropic.claude-sonnet-4-5-20250929-v1:0");
713        let provider =
714            test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0").with_inference_profile_arn(&arn);
715
716        assert_eq!(provider.request_model_id(), arn);
717        assert_eq!(provider.context_window(), Some(200_000));
718        assert_eq!(provider.model().unwrap().to_string(), "bedrock:anthropic.claude-sonnet-4-5-20250929-v1:0");
719    }
720
721    #[test]
722    fn prompt_caching_support_comes_from_canonical_model() {
723        let cached = test_provider().with_model("anthropic.claude-sonnet-4-5-20250929-v1:0");
724        assert!(cached.model().unwrap().supports_prompt_caching());
725
726        let unknown_profile = test_provider().with_model("us.anthropic.claude-future-model-v99:0");
727        assert!(!unknown_profile.model().unwrap().supports_prompt_caching());
728    }
729
730    struct FakeBedrockEndpoint {
731        url: String,
732        request: oneshot::Receiver<CapturedRequest>,
733    }
734
735    struct CapturedRequest {
736        method: Method,
737        path: String,
738        headers: HeaderMap,
739    }
740
741    #[derive(Clone)]
742    struct FakeBedrockState {
743        request_tx: Arc<Mutex<Option<oneshot::Sender<CapturedRequest>>>>,
744        shutdown_tx: Arc<Mutex<Option<oneshot::Sender<()>>>>,
745    }
746
747    impl FakeBedrockEndpoint {
748        async fn start() -> Self {
749            let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind fake Bedrock endpoint");
750            let url = format!("http://{}", listener.local_addr().expect("fake Bedrock endpoint address"));
751            let (request_tx, request) = oneshot::channel();
752            let (shutdown_tx, shutdown) = oneshot::channel();
753            let state = FakeBedrockState {
754                request_tx: Arc::new(Mutex::new(Some(request_tx))),
755                shutdown_tx: Arc::new(Mutex::new(Some(shutdown_tx))),
756            };
757
758            let app = Router::new().fallback(any(capture_bedrock_request)).with_state(state);
759            tokio::spawn(async move {
760                axum::serve(listener, app)
761                    .with_graceful_shutdown(async {
762                        let _ = shutdown.await;
763                    })
764                    .await
765                    .expect("serve fake Bedrock endpoint");
766            });
767
768            Self { url, request }
769        }
770    }
771
772    async fn capture_bedrock_request(
773        State(state): State<FakeBedrockState>,
774        request: Request<Body>,
775    ) -> impl IntoResponse {
776        let (parts, _) = request.into_parts();
777        if let Some(tx) = state.request_tx.lock().await.take() {
778            let _ = tx.send(CapturedRequest {
779                method: parts.method,
780                path: parts.uri.path().to_string(),
781                headers: parts.headers,
782            });
783        }
784        if let Some(tx) = state.shutdown_tx.lock().await.take() {
785            let _ = tx.send(());
786        }
787        (StatusCode::FORBIDDEN, "{}")
788    }
789}
llm/providers/bedrock/provider.rs

llm/providers/bedrock/
provider.rs
