1pub(crate) mod account_routing;
8pub mod catalog;
9pub(crate) mod codex_wire;
10pub(crate) mod extension_route;
11pub mod net;
12pub mod ping;
13pub mod reasoning;
14pub mod registry;
15pub mod stream;
16pub mod translate;
17pub mod types;
18pub mod wire;
19
20use std::sync::Arc;
21
22use crate::extensions::manager::ExtensionManager;
23use crate::extensions::providers::ProviderRegistry;
24
25pub use types::{
26 ChatMessage, ChatOptions, ChatRequest, FunctionCall, FunctionDefinition, OaiEvent,
27 ProviderConfig, StreamOptions, ToolCall, ToolChoice, ToolDefinition,
28};
29pub use wire::StreamDecoder;
30
31static EXTENSION_MANAGER: std::sync::RwLock<Option<Arc<tokio::sync::RwLock<ExtensionManager>>>> =
32 std::sync::RwLock::new(None);
33
34pub fn set_extension_manager_for_routing(manager: Arc<tokio::sync::RwLock<ExtensionManager>>) {
35 *EXTENSION_MANAGER
36 .write()
37 .expect("extension manager routing lock poisoned") = Some(manager);
38}
39
40pub fn clear_extension_manager_for_routing() {
41 *EXTENSION_MANAGER
42 .write()
43 .expect("extension manager routing lock poisoned") = None;
44}
45
46pub fn extension_manager_for_routing() -> Option<Arc<tokio::sync::RwLock<ExtensionManager>>> {
47 EXTENSION_MANAGER
48 .read()
49 .expect("extension manager routing lock poisoned")
50 .clone()
51}
52
53#[derive(Debug, Clone, Copy, PartialEq, Eq)]
54pub enum AuthPolicy {
55 OAuthAccessToken(crate::auth::OAuthProviderId),
56 BrokerProxy,
57}
58#[derive(Debug, Clone, Copy, PartialEq, Eq)]
59pub enum WireProtocol {
60 AnthropicMessages,
61 OpenAiChatCompletions,
62 OpenAiResponses,
63 CodexResponses,
64 GoogleGeminiCodeAssist,
67}
68#[derive(Debug, Clone, PartialEq, Eq)]
69pub struct ResolvedRoute {
70 pub endpoint: String,
71 pub model: String,
72 pub provider: String,
73 pub auth: AuthPolicy,
74 pub wire: WireProtocol,
75}
76
77pub fn resolve_route(model: &str) -> Option<ResolvedRoute> {
79 if let Some((prefix, rest)) = model.split_once('/') {
80 if prefix == "anthropic" && rest.starts_with("claude-") {
81 return Some(ResolvedRoute {
82 endpoint: "https://api.anthropic.com".into(),
83 model: rest.into(),
84 provider: prefix.into(),
85 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Anthropic),
86 wire: WireProtocol::AnthropicMessages,
87 });
88 }
89 if prefix == "xai-auth" && catalog::xai_model(rest).is_some() {
90 return Some(ResolvedRoute {
91 endpoint: "https://api.x.ai/v1".into(),
92 model: rest.into(),
93 provider: prefix.into(),
94 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Xai),
95 wire: WireProtocol::OpenAiResponses,
96 });
97 }
98 if prefix == "github-copilot" {
99 let descriptor = catalog::github_copilot_runtime_model(rest)?;
100 return Some(ResolvedRoute {
101 endpoint: catalog::MODELS_BASE_URL.into(),
102 model: rest.into(),
103 provider: prefix.into(),
104 auth: AuthPolicy::BrokerProxy,
105 wire: descriptor,
106 });
107 }
108 if prefix == "google-gemini" {
109 catalog::google_gemini_model(rest)?;
113 return Some(ResolvedRoute {
114 endpoint: crate::auth::broker::GOOGLE_GEMINI_CODE_ASSIST_BASE_URL.into(),
115 model: rest.into(),
116 provider: prefix.into(),
117 auth: AuthPolicy::BrokerProxy,
118 wire: WireProtocol::GoogleGeminiCodeAssist,
119 });
120 }
121 if prefix == "openai-codex" {
122 let c = registry::resolve_codex_shorthand(model)?;
123 return Some(ResolvedRoute {
124 endpoint: c.base_url,
125 model: c.model,
126 provider: c.provider,
127 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::OpenAiCodex),
128 wire: WireProtocol::CodexResponses,
129 });
130 }
131 if prefix == "kimi-code" {
132 catalog::kimi_code_model(rest)?;
137 return Some(ResolvedRoute {
138 endpoint: crate::auth::kimi_code::API_BASE_URL.into(),
139 model: rest.into(),
140 provider: prefix.into(),
141 auth: AuthPolicy::BrokerProxy,
142 wire: WireProtocol::OpenAiChatCompletions,
143 });
144 }
145 if prefix == "muse" {
146 if !catalog::muse_routable_id(rest) {
152 return None;
153 }
154 return Some(ResolvedRoute {
155 endpoint: crate::auth::muse::API_BASE_URL.into(),
156 model: rest.into(),
157 provider: prefix.into(),
158 auth: AuthPolicy::BrokerProxy,
159 wire: WireProtocol::OpenAiResponses,
160 });
161 }
162 if prefix == "local" {
163 let c = registry::resolve_provider_model(prefix, rest)?;
164 return Some(ResolvedRoute {
165 endpoint: c.base_url,
166 model: c.model,
167 provider: c.provider,
168 auth: AuthPolicy::BrokerProxy,
169 wire: WireProtocol::OpenAiChatCompletions,
170 });
171 }
172 if let Some(spec) = registry::providers().iter().find(|s| s.key == prefix) {
173 return Some(ResolvedRoute {
174 endpoint: spec.base_url.into(),
175 model: rest.into(),
176 provider: prefix.into(),
177 auth: AuthPolicy::BrokerProxy,
178 wire: WireProtocol::OpenAiChatCompletions,
179 });
180 }
181 return None;
182 }
183 Some(ResolvedRoute {
184 endpoint: "https://api.anthropic.com".into(),
185 model: model.into(),
186 provider: "anthropic".into(),
187 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Anthropic),
188 wire: WireProtocol::AnthropicMessages,
189 })
190}
191fn provider_config(r: &ResolvedRoute) -> ProviderConfig {
192 ProviderConfig {
193 base_url: r.endpoint.clone(),
194 model: r.model.clone(),
195 provider: r.provider.clone(),
196 }
197}
198
199#[allow(clippy::too_many_arguments)]
210pub async fn try_route(
211 model: &str,
212 client: &reqwest::Client,
213 tools_schema: &std::sync::Arc<Vec<serde_json::Value>>,
214 system_prompt: &Option<String>,
215 messages: &[crate::SharedMessage],
216 tx: &tokio::sync::mpsc::UnboundedSender<crate::runtime::types::StreamEvent>,
217 temperature: Option<f32>,
218 max_tokens: Option<u32>,
219 thinking_budget: u32,
220 reasoning_level: agent_core::reasoning::ReasoningLevel,
221 cancel: &tokio_util::sync::CancellationToken,
222 source: &crate::auth::CredentialSource,
223 cache: &crate::auth::TokenCache,
224 max_retries: u32,
225 codex_request_role: catalog::CodexRequestRole,
226 tool_session_id: Option<&crate::tools::activation::SessionId>,
227 session_tool_set: Option<&crate::tools::activation::SharedSessionToolSet>,
228 trace: &crate::runtime::trace::TraceContext,
229 suppress_stream_deltas: bool,
230) -> Option<Result<serde_json::Value, Box<dyn std::error::Error + Send + Sync>>> {
231 try_route_with_memory_backend(
232 model,
233 client,
234 tools_schema,
235 system_prompt,
236 messages,
237 tx,
238 temperature,
239 max_tokens,
240 thinking_budget,
241 reasoning_level,
242 cancel,
243 source,
244 cache,
245 max_retries,
246 codex_request_role,
247 tool_session_id,
248 session_tool_set,
249 None,
250 trace,
251 suppress_stream_deltas,
252 )
253 .await
254}
255
256#[allow(clippy::too_many_arguments)]
259pub(crate) async fn try_route_with_memory_backend(
260 model: &str,
261 client: &reqwest::Client,
262 tools_schema: &std::sync::Arc<Vec<serde_json::Value>>,
263 system_prompt: &Option<String>,
264 messages: &[crate::SharedMessage],
265 tx: &tokio::sync::mpsc::UnboundedSender<crate::runtime::types::StreamEvent>,
266 temperature: Option<f32>,
267 max_tokens: Option<u32>,
268 thinking_budget: u32,
269 reasoning_level: agent_core::reasoning::ReasoningLevel,
270 cancel: &tokio_util::sync::CancellationToken,
271 source: &crate::auth::CredentialSource,
272 cache: &crate::auth::TokenCache,
273 max_retries: u32,
274 codex_request_role: catalog::CodexRequestRole,
275 tool_session_id: Option<&crate::tools::activation::SessionId>,
276 session_tool_set: Option<&crate::tools::activation::SharedSessionToolSet>,
277 memory_backend: Option<&crate::memory_backend::MemoryBinding>,
278 trace: &crate::runtime::trace::TraceContext,
279 suppress_stream_deltas: bool,
280) -> Option<Result<serde_json::Value, Box<dyn std::error::Error + Send + Sync>>> {
281 if model.starts_with("muse/") {
285 let broker = crate::auth::broker_from_source(source, cache, client.clone());
286 catalog::ensure_muse_catalog_for(model, &broker).await;
287 }
288 if let Err(error) = super::attachments::validate_history_for_request(model, messages) {
289 return Some(Err(error.into()));
290 }
291 if let Some((plugin_id, provider_id, model_id)) = ProviderRegistry::parse_model_id(model) {
292 if let Some(manager) = extension_manager_for_routing() {
293 return Some(
296 extension_route::route_extension_provider(
297 manager,
298 plugin_id,
299 provider_id,
300 model_id,
301 model,
302 tools_schema,
303 system_prompt,
304 messages,
305 tx,
306 temperature,
307 max_tokens,
308 thinking_budget,
309 cancel,
310 tool_session_id,
311 session_tool_set,
312 memory_backend,
313 trace,
314 )
315 .await,
316 );
317 }
318 return Some(Err(format!(
319 "Extension provider model '{}' is not available",
320 model
321 )
322 .into()));
323 }
324
325 let Some(route) = resolve_route(model) else {
326 return Some(Err(
327 format!("Unknown provider route for model '{model}'").into()
328 ));
329 };
330 let cfg = provider_config(&route);
331 let broker = crate::auth::broker_from_source(source, cache, client.clone());
332 let exact_wire_bytes = !source.is_remote();
336 match route.wire {
337 WireProtocol::OpenAiChatCompletions => Some(
338 stream::call_oai_stream_inner(
339 &cfg,
340 &broker,
341 tools_schema,
342 system_prompt,
343 messages,
344 tx,
345 temperature,
346 max_tokens,
347 thinking_budget,
348 reasoning_level,
349 cancel,
350 trace,
351 exact_wire_bytes,
352 suppress_stream_deltas,
353 max_retries,
354 )
355 .await,
356 ),
357 WireProtocol::OpenAiResponses => {
358 if suppress_stream_deltas {
359 return Some(Err(
360 "internal sync routing is unavailable for this streaming provider".into(),
361 ));
362 }
363 Some(
366 stream::call_responses_route_inner(
367 &route.provider,
368 &cfg,
369 &broker,
370 tools_schema,
371 system_prompt,
372 messages,
373 tx,
374 max_tokens,
375 reasoning_level,
376 cancel,
377 max_retries,
381 trace,
382 exact_wire_bytes,
383 )
384 .await,
385 )
386 }
387 WireProtocol::CodexResponses => {
388 if suppress_stream_deltas {
389 return Some(Err(
390 "internal sync routing is unavailable for this streaming provider".into(),
391 ));
392 }
393 Some(
394 stream::call_codex_stream_inner(
395 &cfg,
396 client,
397 &broker,
398 tools_schema,
399 system_prompt,
400 messages,
401 tx,
402 temperature,
403 max_tokens,
404 reasoning_level,
405 codex_request_role,
406 cancel,
407 stream::codex_retry_budget(max_retries),
411 trace,
412 )
413 .await,
414 )
415 }
416 WireProtocol::AnthropicMessages => None,
417 WireProtocol::GoogleGeminiCodeAssist => {
418 if suppress_stream_deltas {
419 return Some(Err(
420 "internal sync routing is unavailable for this streaming provider".into(),
421 ));
422 }
423 Some(
424 crate::runtime::google_gemini::runtime::call_google_gemini_stream_inner(
425 &cfg,
426 &broker,
427 tools_schema,
428 system_prompt,
429 messages,
430 tx,
431 cancel,
432 trace,
433 exact_wire_bytes,
434 )
435 .await,
436 )
437 }
438 }
439}
440
441#[cfg(test)]
442mod tests {
443 use super::*;
444
445 #[test]
446 fn anthropic_identity_is_explicit_while_legacy_bare_ids_still_route() {
447 for id in ["claude-sonnet-4-6", "claude-opus-4-7"] {
448 let qualified = resolve_route(&format!("anthropic/{id}"))
449 .expect("qualified Anthropic model must route");
450 assert_eq!(qualified.provider, "anthropic");
451 assert_eq!(qualified.model, id);
452 assert_eq!(qualified.wire, WireProtocol::AnthropicMessages);
453
454 let legacy = resolve_route(id).expect("legacy bare Claude model must route");
455 assert_eq!(legacy.provider, "anthropic");
456 assert_eq!(legacy.model, id);
457 }
458 assert_eq!(
459 resolve_route("github-copilot/claude-sonnet-4.6")
460 .unwrap()
461 .provider,
462 "github-copilot"
463 );
464 }
465
466 #[test]
467 fn brokered_provider_routing_is_typed_and_fail_closed() {
468 for descriptor in catalog::XAI_TEXT_MODELS {
469 let runtime_id = format!("xai-auth/{}", descriptor.id);
470 let xai = resolve_route(&runtime_id).unwrap();
471 assert_eq!(xai.model, descriptor.id);
472 assert_eq!(xai.provider, "xai-auth");
473 assert_eq!(xai.endpoint, "https://api.x.ai/v1");
474 assert_eq!(
475 xai.auth,
476 AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Xai)
477 );
478 assert_eq!(xai.wire, WireProtocol::OpenAiResponses);
479 }
480 let static_route = resolve_route("groq/llama-3.3-70b-versatile").unwrap();
481 assert_eq!(static_route.auth, AuthPolicy::BrokerProxy);
482 assert_eq!(static_route.wire, WireProtocol::OpenAiChatCompletions);
483
484 let copilot_chat = resolve_route("github-copilot/claude-sonnet-4.6").unwrap();
485 assert_eq!(copilot_chat.auth, AuthPolicy::BrokerProxy);
486 assert_eq!(copilot_chat.wire, WireProtocol::OpenAiChatCompletions);
487 let copilot_responses = resolve_route("github-copilot/gpt-5.3-codex").unwrap();
488 assert_eq!(copilot_responses.auth, AuthPolicy::BrokerProxy);
489 assert_eq!(copilot_responses.wire, WireProtocol::OpenAiResponses);
490 assert!(resolve_route("github-copilot/unverified-model").is_none());
491 assert!(resolve_route("unknown-provider/model").is_none());
492 assert!(resolve_route("xai-auth/not-a-real-model").is_none());
493 assert!(resolve_route("xai-auth/grok-build-0.1").is_none());
494 assert!(resolve_route("xai-auth/grok-build-latest").is_none());
495 assert!(resolve_route("xai-auth/grok-imagine-image").is_none());
496
497 for descriptor in catalog::GOOGLE_GEMINI_TEXT_MODELS {
499 let route = resolve_route(&format!("google-gemini/{}", descriptor.id)).unwrap();
500 assert_eq!(route.provider, "google-gemini");
501 assert_eq!(route.model, descriptor.id);
502 assert_eq!(route.endpoint, "https://cloudcode-pa.googleapis.com");
503 assert_eq!(route.auth, AuthPolicy::BrokerProxy);
504 assert_eq!(route.wire, WireProtocol::GoogleGeminiCodeAssist);
505 }
506 for banned in [
508 "auto-gemini-2.5",
509 "auto-gemini-3",
510 "text-embedding-004",
511 "gemini-2.5-image",
512 "gemini-2.5", ] {
514 assert!(
515 resolve_route(&format!("google-gemini/{banned}")).is_none(),
516 "{banned} must not resolve"
517 );
518 }
519 }
520
521 #[test]
525 fn does_not_resolve_public_api_alias_gemini_pro_latest() {
526 assert!(
527 resolve_route("google-gemini/gemini-pro-latest").is_none(),
528 "gemini-pro-latest is not a Code Assist wire ID (404 upstream)"
529 );
530 }
531
532 #[test]
533 fn muse_routing_is_broker_proxied_responses_and_fail_closed() {
534 let route = resolve_route("muse/muse-wonder-1.0").expect("documented seed routes");
535 assert_eq!(route.provider, "muse");
536 assert_eq!(route.model, "muse-wonder-1.0");
537 assert_eq!(route.endpoint, crate::auth::muse::API_BASE_URL);
538 assert_eq!(route.auth, AuthPolicy::BrokerProxy);
540 assert_eq!(route.wire, WireProtocol::OpenAiResponses);
541 let live_only = resolve_route("muse/muse-spark-9.9").expect("shape-valid id routes");
546 assert_eq!(live_only.model, "muse-spark-9.9");
547 assert_eq!(live_only.wire, WireProtocol::OpenAiResponses);
548 for banned in ["MUSE-WONDER-1.0", "", "a b", "../x", "-x", "a/b"] {
550 assert!(
551 resolve_route(&format!("muse/{banned}")).is_none(),
552 "{banned} must not resolve"
553 );
554 }
555 }
556
557 #[test]
558 fn kimi_code_routing_is_broker_proxied_and_fail_closed() {
559 for descriptor in catalog::KIMI_CODE_TEXT_MODELS {
560 let route = resolve_route(&format!("kimi-code/{}", descriptor.id)).unwrap();
561 assert_eq!(route.provider, "kimi-code");
562 assert_eq!(route.model, descriptor.id);
563 assert_eq!(route.endpoint, "https://api.kimi.com/coding/v1");
564 assert_eq!(route.auth, AuthPolicy::BrokerProxy);
567 assert_eq!(route.wire, WireProtocol::OpenAiChatCompletions);
568 }
569 assert!(resolve_route("kimi-code/unknown-model").is_none());
572 assert!(resolve_route("kimi-code/kimi-k3").is_none());
573 assert!(resolve_route("kimi-code/kimi-k2.7-code").is_none());
574 }
575
576 #[test]
577 fn resolves_openai_codex_without_requiring_eager_credentials() {
578 std::env::remove_var("OPENAI_CODEX_ACCESS_TOKEN");
579 let route = resolve_route("openai-codex/gpt-5.1-codex-mini").unwrap();
580 assert_eq!(route.provider, "openai-codex");
581 assert_eq!(route.wire, WireProtocol::CodexResponses);
582 }
583
584 #[tokio::test]
597 async fn try_route_google_gemini_is_not_none_and_does_not_fall_through_to_anthropic() {
598 let client = reqwest::Client::builder()
599 .connect_timeout(std::time::Duration::from_secs(2))
600 .build()
601 .expect("client");
602 let source = crate::auth::CredentialSource::Remote {
605 endpoint: "http://127.0.0.1:1".into(),
606 machine_token: "not-a-real-token".into(),
607 };
608 let cache = crate::auth::TokenCache::new();
609 let tools_schema: std::sync::Arc<Vec<serde_json::Value>> = std::sync::Arc::new(Vec::new());
610 let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
611 let cancel = tokio_util::sync::CancellationToken::new();
612
613 let messages: Vec<crate::SharedMessage> = vec![std::sync::Arc::new(serde_json::json!({
614 "role": "user",
615 "content": "hi"
616 }))];
617
618 let trace = crate::runtime::trace::TraceContext::disabled();
619 let fut = try_route(
620 "google-gemini/gemini-2.5-pro",
621 &client,
622 &tools_schema,
623 &None,
624 &messages,
625 &tx,
626 None,
627 None,
628 0,
629 agent_core::reasoning::ReasoningLevel::Adaptive,
630 &cancel,
631 &source,
632 &cache,
633 0,
634 catalog::CodexRequestRole::Foreground,
635 None,
636 None,
637 &trace,
638 false,
639 );
640 let result = tokio::time::timeout(std::time::Duration::from_secs(10), fut)
641 .await
642 .expect("try_route completed within budget");
643
644 assert!(
645 result.is_some(),
646 "google-gemini must route through try_route (was None → fell through to Anthropic)"
647 );
648 }
649
650 #[tokio::test]
656 async fn xai_off_is_rejected_pre_network_through_try_route() {
657 let client = reqwest::Client::builder()
658 .connect_timeout(std::time::Duration::from_secs(2))
659 .build()
660 .expect("client");
661 let source = crate::auth::CredentialSource::Remote {
662 endpoint: "http://127.0.0.1:1".into(),
663 machine_token: "not-a-real-token".into(),
664 };
665 let cache = crate::auth::TokenCache::new();
666 let tools_schema: std::sync::Arc<Vec<serde_json::Value>> = std::sync::Arc::new(Vec::new());
667 let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
668 let cancel = tokio_util::sync::CancellationToken::new();
669 let messages: Vec<crate::SharedMessage> = vec![std::sync::Arc::new(serde_json::json!({
670 "role": "user",
671 "content": "hi"
672 }))];
673
674 for (level, needle) in [
675 (
676 agent_core::reasoning::ReasoningLevel::Off,
677 "cannot be disabled",
678 ),
679 (
680 agent_core::reasoning::ReasoningLevel::XHigh,
681 "not supported",
682 ),
683 (
684 agent_core::reasoning::ReasoningLevel::Ultra,
685 "not supported",
686 ),
687 ] {
688 let result = try_route(
689 "xai-auth/grok-4.5",
690 &client,
691 &tools_schema,
692 &None,
693 &messages,
694 &tx,
695 None,
696 None,
697 0,
698 level,
699 &cancel,
700 &source,
701 &cache,
702 0,
703 catalog::CodexRequestRole::Foreground,
704 None,
705 None,
706 &crate::runtime::trace::TraceContext::disabled(),
707 false,
708 )
709 .await
710 .expect("xai-auth must route through try_route");
711 let err = result.expect_err("rejected level must error");
712 assert!(
713 err.to_string().contains(needle),
714 "{level}: pre-network capability rejection expected, got: {err}"
715 );
716 }
717 }
718
719 #[test]
720 fn set_extension_manager_for_routing_overwrites_previous_manager() {
721 clear_extension_manager_for_routing();
722 let first = Arc::new(tokio::sync::RwLock::new(ExtensionManager::new(Arc::new(
723 crate::extensions::hooks::HookBus::new(),
724 ))));
725 let second = Arc::new(tokio::sync::RwLock::new(ExtensionManager::new(Arc::new(
726 crate::extensions::hooks::HookBus::new(),
727 ))));
728
729 set_extension_manager_for_routing(first.clone());
730 assert!(Arc::ptr_eq(
731 &extension_manager_for_routing().unwrap(),
732 &first
733 ));
734
735 set_extension_manager_for_routing(second.clone());
736 assert!(Arc::ptr_eq(
737 &extension_manager_for_routing().unwrap(),
738 &second
739 ));
740
741 clear_extension_manager_for_routing();
742 assert!(extension_manager_for_routing().is_none());
743 }
744}