1pub mod catalog;
8pub(crate) mod extension_route;
9pub mod net;
10pub mod ping;
11pub mod reasoning;
12pub mod registry;
13pub mod stream;
14pub mod translate;
15pub mod types;
16pub mod wire;
17
18use std::sync::Arc;
19
20use crate::extensions::manager::ExtensionManager;
21use crate::extensions::providers::ProviderRegistry;
22
23pub use types::{
24 ChatMessage, ChatOptions, ChatRequest, FunctionCall, FunctionDefinition, OaiEvent,
25 ProviderConfig, StreamOptions, ToolCall, ToolChoice, ToolDefinition,
26};
27pub use wire::StreamDecoder;
28
29static EXTENSION_MANAGER: std::sync::RwLock<Option<Arc<tokio::sync::RwLock<ExtensionManager>>>> =
30 std::sync::RwLock::new(None);
31
32pub fn set_extension_manager_for_routing(manager: Arc<tokio::sync::RwLock<ExtensionManager>>) {
33 *EXTENSION_MANAGER
34 .write()
35 .expect("extension manager routing lock poisoned") = Some(manager);
36}
37
38pub fn clear_extension_manager_for_routing() {
39 *EXTENSION_MANAGER
40 .write()
41 .expect("extension manager routing lock poisoned") = None;
42}
43
44pub fn extension_manager_for_routing() -> Option<Arc<tokio::sync::RwLock<ExtensionManager>>> {
45 EXTENSION_MANAGER
46 .read()
47 .expect("extension manager routing lock poisoned")
48 .clone()
49}
50
51#[derive(Debug, Clone, Copy, PartialEq, Eq)]
52pub enum AuthPolicy {
53 OAuthAccessToken(crate::auth::OAuthProviderId),
54 BrokerProxy,
55}
56#[derive(Debug, Clone, Copy, PartialEq, Eq)]
57pub enum WireProtocol {
58 AnthropicMessages,
59 OpenAiChatCompletions,
60 OpenAiResponses,
61 CodexResponses,
62 GoogleGeminiCodeAssist,
65}
66#[derive(Debug, Clone, PartialEq, Eq)]
67pub struct ResolvedRoute {
68 pub endpoint: String,
69 pub model: String,
70 pub provider: String,
71 pub auth: AuthPolicy,
72 pub wire: WireProtocol,
73}
74
75pub fn resolve_route(model: &str) -> Option<ResolvedRoute> {
77 if let Some((prefix, rest)) = model.split_once('/') {
78 if prefix == "anthropic" && rest.starts_with("claude-") {
79 return Some(ResolvedRoute {
80 endpoint: "https://api.anthropic.com".into(),
81 model: rest.into(),
82 provider: prefix.into(),
83 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Anthropic),
84 wire: WireProtocol::AnthropicMessages,
85 });
86 }
87 if prefix == "xai-auth" && catalog::xai_model(rest).is_some() {
88 return Some(ResolvedRoute {
89 endpoint: "https://api.x.ai/v1".into(),
90 model: rest.into(),
91 provider: prefix.into(),
92 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Xai),
93 wire: WireProtocol::OpenAiResponses,
94 });
95 }
96 if prefix == "github-copilot" {
97 let descriptor = catalog::github_copilot_runtime_model(rest)?;
98 return Some(ResolvedRoute {
99 endpoint: catalog::MODELS_BASE_URL.into(),
100 model: rest.into(),
101 provider: prefix.into(),
102 auth: AuthPolicy::BrokerProxy,
103 wire: descriptor,
104 });
105 }
106 if prefix == "google-gemini" {
107 catalog::google_gemini_model(rest)?;
111 return Some(ResolvedRoute {
112 endpoint: crate::auth::broker::GOOGLE_GEMINI_CODE_ASSIST_BASE_URL.into(),
113 model: rest.into(),
114 provider: prefix.into(),
115 auth: AuthPolicy::BrokerProxy,
116 wire: WireProtocol::GoogleGeminiCodeAssist,
117 });
118 }
119 if prefix == "openai-codex" {
120 let c = registry::resolve_codex_shorthand(model)?;
121 return Some(ResolvedRoute {
122 endpoint: c.base_url,
123 model: c.model,
124 provider: c.provider,
125 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::OpenAiCodex),
126 wire: WireProtocol::CodexResponses,
127 });
128 }
129 if prefix == "kimi-code" {
130 catalog::kimi_code_model(rest)?;
135 return Some(ResolvedRoute {
136 endpoint: crate::auth::kimi_code::API_BASE_URL.into(),
137 model: rest.into(),
138 provider: prefix.into(),
139 auth: AuthPolicy::BrokerProxy,
140 wire: WireProtocol::OpenAiChatCompletions,
141 });
142 }
143 if prefix == "local" {
144 let c = registry::resolve_provider_model(prefix, rest)?;
145 return Some(ResolvedRoute {
146 endpoint: c.base_url,
147 model: c.model,
148 provider: c.provider,
149 auth: AuthPolicy::BrokerProxy,
150 wire: WireProtocol::OpenAiChatCompletions,
151 });
152 }
153 if let Some(spec) = registry::providers().iter().find(|s| s.key == prefix) {
154 return Some(ResolvedRoute {
155 endpoint: spec.base_url.into(),
156 model: rest.into(),
157 provider: prefix.into(),
158 auth: AuthPolicy::BrokerProxy,
159 wire: WireProtocol::OpenAiChatCompletions,
160 });
161 }
162 return None;
163 }
164 Some(ResolvedRoute {
165 endpoint: "https://api.anthropic.com".into(),
166 model: model.into(),
167 provider: "anthropic".into(),
168 auth: AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Anthropic),
169 wire: WireProtocol::AnthropicMessages,
170 })
171}
172fn provider_config(r: &ResolvedRoute) -> ProviderConfig {
173 ProviderConfig {
174 base_url: r.endpoint.clone(),
175 model: r.model.clone(),
176 provider: r.provider.clone(),
177 }
178}
179
180#[allow(clippy::too_many_arguments)]
191pub async fn try_route(
192 model: &str,
193 client: &reqwest::Client,
194 tools_schema: &std::sync::Arc<Vec<serde_json::Value>>,
195 system_prompt: &Option<String>,
196 messages: &[crate::SharedMessage],
197 tx: &tokio::sync::mpsc::UnboundedSender<crate::runtime::types::StreamEvent>,
198 temperature: Option<f32>,
199 max_tokens: Option<u32>,
200 thinking_budget: u32,
201 reasoning_level: agent_core::reasoning::ReasoningLevel,
202 cancel: &tokio_util::sync::CancellationToken,
203 source: &crate::auth::CredentialSource,
204 cache: &crate::auth::TokenCache,
205 max_retries: u32,
206 codex_request_role: catalog::CodexRequestRole,
207 tool_session_id: Option<&crate::tools::activation::SessionId>,
208 session_tool_set: Option<&crate::tools::activation::SharedSessionToolSet>,
209 trace: &crate::runtime::trace::TraceContext,
210 suppress_stream_deltas: bool,
211) -> Option<Result<serde_json::Value, Box<dyn std::error::Error + Send + Sync>>> {
212 if let Some((plugin_id, provider_id, model_id)) = ProviderRegistry::parse_model_id(model) {
213 if let Some(manager) = extension_manager_for_routing() {
214 return Some(
217 extension_route::route_extension_provider(
218 manager,
219 plugin_id,
220 provider_id,
221 model_id,
222 model,
223 tools_schema,
224 system_prompt,
225 messages,
226 tx,
227 temperature,
228 max_tokens,
229 thinking_budget,
230 cancel,
231 tool_session_id,
232 session_tool_set,
233 trace,
234 )
235 .await,
236 );
237 }
238 return Some(Err(format!(
239 "Extension provider model '{}' is not available",
240 model
241 )
242 .into()));
243 }
244
245 let Some(route) = resolve_route(model) else {
246 return Some(Err(
247 format!("Unknown provider route for model '{model}'").into()
248 ));
249 };
250 let cfg = provider_config(&route);
251 let broker = crate::auth::broker_from_source(source, cache, client.clone());
252 let exact_wire_bytes = !source.is_remote();
256 match route.wire {
257 WireProtocol::OpenAiChatCompletions => Some(
258 stream::call_oai_stream_inner(
259 &cfg,
260 &broker,
261 tools_schema,
262 system_prompt,
263 messages,
264 tx,
265 temperature,
266 max_tokens,
267 thinking_budget,
268 reasoning_level,
269 cancel,
270 trace,
271 exact_wire_bytes,
272 suppress_stream_deltas,
273 )
274 .await,
275 ),
276 WireProtocol::OpenAiResponses => {
277 if suppress_stream_deltas {
278 return Some(Err(
279 "internal sync routing is unavailable for this streaming provider".into(),
280 ));
281 }
282 Some(
283 stream::call_xai_responses_stream_inner(
284 &cfg,
285 &broker,
286 tools_schema,
287 system_prompt,
288 messages,
289 tx,
290 max_tokens,
291 reasoning_level,
292 cancel,
293 max_retries,
297 trace,
298 exact_wire_bytes,
299 )
300 .await,
301 )
302 }
303 WireProtocol::CodexResponses => {
304 if suppress_stream_deltas {
305 return Some(Err(
306 "internal sync routing is unavailable for this streaming provider".into(),
307 ));
308 }
309 Some(
310 stream::call_codex_stream_inner(
311 &cfg,
312 client,
313 &broker,
314 tools_schema,
315 system_prompt,
316 messages,
317 tx,
318 temperature,
319 max_tokens,
320 reasoning_level,
321 codex_request_role,
322 cancel,
323 stream::codex_retry_budget(max_retries),
327 trace,
328 )
329 .await,
330 )
331 }
332 WireProtocol::AnthropicMessages => None,
333 WireProtocol::GoogleGeminiCodeAssist => {
334 if suppress_stream_deltas {
335 return Some(Err(
336 "internal sync routing is unavailable for this streaming provider".into(),
337 ));
338 }
339 Some(
340 crate::runtime::google_gemini::runtime::call_google_gemini_stream_inner(
341 &cfg,
342 &broker,
343 tools_schema,
344 system_prompt,
345 messages,
346 tx,
347 cancel,
348 trace,
349 exact_wire_bytes,
350 )
351 .await,
352 )
353 }
354 }
355}
356
357#[cfg(test)]
358mod tests {
359 use super::*;
360
361 #[test]
362 fn anthropic_identity_is_explicit_while_legacy_bare_ids_still_route() {
363 for id in ["claude-sonnet-4-6", "claude-opus-4-7"] {
364 let qualified = resolve_route(&format!("anthropic/{id}"))
365 .expect("qualified Anthropic model must route");
366 assert_eq!(qualified.provider, "anthropic");
367 assert_eq!(qualified.model, id);
368 assert_eq!(qualified.wire, WireProtocol::AnthropicMessages);
369
370 let legacy = resolve_route(id).expect("legacy bare Claude model must route");
371 assert_eq!(legacy.provider, "anthropic");
372 assert_eq!(legacy.model, id);
373 }
374 assert_eq!(
375 resolve_route("github-copilot/claude-sonnet-4.6")
376 .unwrap()
377 .provider,
378 "github-copilot"
379 );
380 }
381
382 #[test]
383 fn brokered_provider_routing_is_typed_and_fail_closed() {
384 for descriptor in catalog::XAI_TEXT_MODELS {
385 let runtime_id = format!("xai-auth/{}", descriptor.id);
386 let xai = resolve_route(&runtime_id).unwrap();
387 assert_eq!(xai.model, descriptor.id);
388 assert_eq!(xai.provider, "xai-auth");
389 assert_eq!(xai.endpoint, "https://api.x.ai/v1");
390 assert_eq!(
391 xai.auth,
392 AuthPolicy::OAuthAccessToken(crate::auth::OAuthProviderId::Xai)
393 );
394 assert_eq!(xai.wire, WireProtocol::OpenAiResponses);
395 }
396 let static_route = resolve_route("groq/llama-3.3-70b-versatile").unwrap();
397 assert_eq!(static_route.auth, AuthPolicy::BrokerProxy);
398 assert_eq!(static_route.wire, WireProtocol::OpenAiChatCompletions);
399
400 let copilot_chat = resolve_route("github-copilot/claude-sonnet-4.6").unwrap();
401 assert_eq!(copilot_chat.auth, AuthPolicy::BrokerProxy);
402 assert_eq!(copilot_chat.wire, WireProtocol::OpenAiChatCompletions);
403 let copilot_responses = resolve_route("github-copilot/gpt-5.3-codex").unwrap();
404 assert_eq!(copilot_responses.auth, AuthPolicy::BrokerProxy);
405 assert_eq!(copilot_responses.wire, WireProtocol::OpenAiResponses);
406 assert!(resolve_route("github-copilot/unverified-model").is_none());
407 assert!(resolve_route("unknown-provider/model").is_none());
408 assert!(resolve_route("xai-auth/not-a-real-model").is_none());
409 assert!(resolve_route("xai-auth/grok-build-0.1").is_none());
410 assert!(resolve_route("xai-auth/grok-build-latest").is_none());
411 assert!(resolve_route("xai-auth/grok-imagine-image").is_none());
412
413 for descriptor in catalog::GOOGLE_GEMINI_TEXT_MODELS {
415 let route = resolve_route(&format!("google-gemini/{}", descriptor.id)).unwrap();
416 assert_eq!(route.provider, "google-gemini");
417 assert_eq!(route.model, descriptor.id);
418 assert_eq!(route.endpoint, "https://cloudcode-pa.googleapis.com");
419 assert_eq!(route.auth, AuthPolicy::BrokerProxy);
420 assert_eq!(route.wire, WireProtocol::GoogleGeminiCodeAssist);
421 }
422 for banned in [
424 "auto-gemini-2.5",
425 "auto-gemini-3",
426 "text-embedding-004",
427 "gemini-2.5-image",
428 "gemini-2.5", ] {
430 assert!(
431 resolve_route(&format!("google-gemini/{banned}")).is_none(),
432 "{banned} must not resolve"
433 );
434 }
435 }
436
437 #[test]
441 fn does_not_resolve_public_api_alias_gemini_pro_latest() {
442 assert!(
443 resolve_route("google-gemini/gemini-pro-latest").is_none(),
444 "gemini-pro-latest is not a Code Assist wire ID (404 upstream)"
445 );
446 }
447
448 #[test]
449 fn kimi_code_routing_is_broker_proxied_and_fail_closed() {
450 for descriptor in catalog::KIMI_CODE_TEXT_MODELS {
451 let route = resolve_route(&format!("kimi-code/{}", descriptor.id)).unwrap();
452 assert_eq!(route.provider, "kimi-code");
453 assert_eq!(route.model, descriptor.id);
454 assert_eq!(route.endpoint, "https://api.kimi.com/coding/v1");
455 assert_eq!(route.auth, AuthPolicy::BrokerProxy);
458 assert_eq!(route.wire, WireProtocol::OpenAiChatCompletions);
459 }
460 assert!(resolve_route("kimi-code/unknown-model").is_none());
463 assert!(resolve_route("kimi-code/kimi-k3").is_none());
464 assert!(resolve_route("kimi-code/kimi-k2.7-code").is_none());
465 }
466
467 #[test]
468 fn resolves_openai_codex_without_requiring_eager_credentials() {
469 std::env::remove_var("OPENAI_CODEX_ACCESS_TOKEN");
470 let route = resolve_route("openai-codex/gpt-5.1-codex-mini").unwrap();
471 assert_eq!(route.provider, "openai-codex");
472 assert_eq!(route.wire, WireProtocol::CodexResponses);
473 }
474
475 #[tokio::test]
488 async fn try_route_google_gemini_is_not_none_and_does_not_fall_through_to_anthropic() {
489 let client = reqwest::Client::builder()
490 .connect_timeout(std::time::Duration::from_secs(2))
491 .build()
492 .expect("client");
493 let source = crate::auth::CredentialSource::Remote {
496 endpoint: "http://127.0.0.1:1".into(),
497 machine_token: "not-a-real-token".into(),
498 };
499 let cache = crate::auth::TokenCache::new();
500 let tools_schema: std::sync::Arc<Vec<serde_json::Value>> = std::sync::Arc::new(Vec::new());
501 let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
502 let cancel = tokio_util::sync::CancellationToken::new();
503
504 let messages: Vec<crate::SharedMessage> = vec![std::sync::Arc::new(serde_json::json!({
505 "role": "user",
506 "content": "hi"
507 }))];
508
509 let trace = crate::runtime::trace::TraceContext::disabled();
510 let fut = try_route(
511 "google-gemini/gemini-2.5-pro",
512 &client,
513 &tools_schema,
514 &None,
515 &messages,
516 &tx,
517 None,
518 None,
519 0,
520 agent_core::reasoning::ReasoningLevel::Adaptive,
521 &cancel,
522 &source,
523 &cache,
524 0,
525 catalog::CodexRequestRole::Foreground,
526 None,
527 None,
528 &trace,
529 false,
530 );
531 let result = tokio::time::timeout(std::time::Duration::from_secs(10), fut)
532 .await
533 .expect("try_route completed within budget");
534
535 assert!(
536 result.is_some(),
537 "google-gemini must route through try_route (was None → fell through to Anthropic)"
538 );
539 }
540
541 #[tokio::test]
547 async fn xai_off_is_rejected_pre_network_through_try_route() {
548 let client = reqwest::Client::builder()
549 .connect_timeout(std::time::Duration::from_secs(2))
550 .build()
551 .expect("client");
552 let source = crate::auth::CredentialSource::Remote {
553 endpoint: "http://127.0.0.1:1".into(),
554 machine_token: "not-a-real-token".into(),
555 };
556 let cache = crate::auth::TokenCache::new();
557 let tools_schema: std::sync::Arc<Vec<serde_json::Value>> = std::sync::Arc::new(Vec::new());
558 let (tx, _rx) = tokio::sync::mpsc::unbounded_channel();
559 let cancel = tokio_util::sync::CancellationToken::new();
560 let messages: Vec<crate::SharedMessage> = vec![std::sync::Arc::new(serde_json::json!({
561 "role": "user",
562 "content": "hi"
563 }))];
564
565 for (level, needle) in [
566 (
567 agent_core::reasoning::ReasoningLevel::Off,
568 "cannot be disabled",
569 ),
570 (
571 agent_core::reasoning::ReasoningLevel::XHigh,
572 "not supported",
573 ),
574 (
575 agent_core::reasoning::ReasoningLevel::Ultra,
576 "not supported",
577 ),
578 ] {
579 let result = try_route(
580 "xai-auth/grok-4.5",
581 &client,
582 &tools_schema,
583 &None,
584 &messages,
585 &tx,
586 None,
587 None,
588 0,
589 level,
590 &cancel,
591 &source,
592 &cache,
593 0,
594 catalog::CodexRequestRole::Foreground,
595 None,
596 None,
597 &crate::runtime::trace::TraceContext::disabled(),
598 false,
599 )
600 .await
601 .expect("xai-auth must route through try_route");
602 let err = result.expect_err("rejected level must error");
603 assert!(
604 err.to_string().contains(needle),
605 "{level}: pre-network capability rejection expected, got: {err}"
606 );
607 }
608 }
609
610 #[test]
611 fn set_extension_manager_for_routing_overwrites_previous_manager() {
612 clear_extension_manager_for_routing();
613 let first = Arc::new(tokio::sync::RwLock::new(ExtensionManager::new(Arc::new(
614 crate::extensions::hooks::HookBus::new(),
615 ))));
616 let second = Arc::new(tokio::sync::RwLock::new(ExtensionManager::new(Arc::new(
617 crate::extensions::hooks::HookBus::new(),
618 ))));
619
620 set_extension_manager_for_routing(first.clone());
621 assert!(Arc::ptr_eq(
622 &extension_manager_for_routing().unwrap(),
623 &first
624 ));
625
626 set_extension_manager_for_routing(second.clone());
627 assert!(Arc::ptr_eq(
628 &extension_manager_for_routing().unwrap(),
629 &second
630 ));
631
632 clear_extension_manager_for_routing();
633 assert!(extension_manager_for_routing().is_none());
634 }
635}