1pub mod codex;
5mod delivery;
6#[cfg(test)]
7mod delivery_tests;
8pub mod models;
9#[cfg(test)]
10pub(crate) mod tests;
11mod usage;
12
13use std::{
14 net::IpAddr,
15 path::Path,
16 sync::{Arc, Mutex},
17 time::Duration,
18};
19
20use futures_util::StreamExt;
21use rig_core::{
22 Model,
23 completion::{CompletionRequest, CompletionResponse, FinishReason},
24 error::ProviderError,
25 observe::{Action, AdapterContext, AdapterEvent, Observation, Subject, Witness},
26 providers::{anthropic, openai},
27 streaming::CompletionStream,
28};
29use serde::{Deserialize, Serialize};
30use serde_json::{Value, json};
31use tokio::time::Instant;
32use tokio_util::sync::CancellationToken;
33
34use crate::{
35 agent::error::{Code, Failure},
36 config::Backend,
37};
38
39const ATTEMPTS: usize = 3;
41const RETRY_BACKOFF_BASE: Duration = Duration::from_millis(250);
44const MAX_DIAGNOSTIC_CHARS: usize = 4096;
46const ANTHROPIC_OUTPUT_TOKENS: u64 = 16_384;
49
50pub enum Client {
53 Openai(Box<Model<openai::responses_api::wire::Responses>>, String),
54 Anthropic(Box<Model<anthropic::wire::Messages>>, String),
55 Codex(codex::Codex),
56}
57
58#[derive(Debug, Clone, Serialize, Deserialize)]
59#[serde(rename_all = "camelCase")]
60pub struct Attempt {
61 #[serde(default)]
62 pub turn: usize,
63 #[serde(default)]
64 pub attempt: usize,
65 #[serde(default)]
66 pub usage: usage::Usage,
67 #[serde(skip_serializing_if = "Option::is_none")]
68 pub error: Option<String>,
69 #[serde(default, skip_serializing_if = "Option::is_none")]
70 pub error_code: Option<Code>,
71}
72
73struct Classified {
75 code: Code,
76 message: String,
77 retry_after: Option<Duration>,
78}
79
80impl From<Classified> for Failure {
81 fn from(classified: Classified) -> Self {
82 Self::new(classified.code, classified.message)
83 }
84}
85
86impl From<Failure> for Classified {
87 fn from(failure: Failure) -> Self {
88 Self {
89 code: failure.code,
90 message: failure.message,
91 retry_after: None,
92 }
93 }
94}
95
96#[derive(Default)]
98struct ReportedUsage(Mutex<std::collections::BTreeMap<String, u64>>);
99
100impl Witness for ReportedUsage {
101 fn observe(&self, observation: Observation) {
102 if let Action::Adapter { observation } = observation.action
103 && let AdapterEvent::Usage { usage } = observation.event
104 {
105 let mut counters = self.0.lock().unwrap();
106 for (name, value) in [
107 ("inputTokens", usage.input_tokens),
108 ("outputTokens", usage.output_tokens),
109 ("totalTokens", usage.total_tokens),
110 ("cacheReadTokens", usage.cached_input_tokens),
111 ("reasoningTokens", usage.reasoning_tokens),
112 ] {
113 insert_counter(&mut counters, name, value);
114 }
115 }
116 }
117}
118
119fn insert_counter(
120 counters: &mut std::collections::BTreeMap<String, u64>,
121 name: &str,
122 value: Option<u64>,
123) {
124 if let Some(value) = value.filter(|value| *value <= crate::types::MAX_SAFE_JSON_INTEGER) {
125 counters.insert(name.into(), value);
126 }
127}
128
129impl Client {
130 pub fn new(
132 backend: Backend,
133 model: &crate::config::ModelId,
134 state: &Path,
135 repo: &Path,
136 ) -> Result<Self, Failure> {
137 let base = base_url(backend)?;
138 if backend == Backend::Codex {
139 return Ok(Self::Codex(codex::Codex::new(model, &base, state, repo)?));
140 }
141 let key = api_key(backend, "for this Agent backend")
142 .map_err(|message| Failure::new(Code::Authentication, message))?;
143 let http = delivery::Tap::new(http_client()?);
144 Ok(match backend {
145 Backend::Openai => Self::Openai(
146 Box::new(
147 openai::OpenAIConfig::new(key.clone())
148 .with_base_url(base)
149 .connect(http)
150 .responses(model.as_str()),
151 ),
152 key,
153 ),
154 Backend::Anthropic => Self::Anthropic(
155 Box::new(
156 anthropic::AnthropicConfig::new(key.clone())
157 .with_base_url(base)
158 .connect(http)
159 .completion(model.as_str()),
160 ),
161 key,
162 ),
163 Backend::Codex => unreachable!("built above"),
164 })
165 }
166
167 pub fn parameters(
170 backend: Backend,
171 reasoning: Option<&str>,
172 session: &crate::types::SessionId,
173 ) -> Result<Value, String> {
174 if let Some(reasoning) = reasoning {
175 backend.validate_reasoning(reasoning)?;
176 }
177 Ok(match backend {
178 Backend::Openai => {
179 let mut value = json!({
180 "store":false,
181 "parallel_tool_calls":false,
182 "include":["reasoning.encrypted_content"],
183 "prompt_cache_key":session,
184 });
185 if let Some(reasoning) = reasoning {
186 value["reasoning"] = json!({"effort":reasoning});
187 }
188 value
189 }
190 Backend::Anthropic => reasoning.map_or(
191 json!({}),
192 |effort| json!({"thinking":{"type":"adaptive"}, "output_config":{"effort":effort}}),
193 ),
194 Backend::Codex => {
197 let mut value = json!({"prompt_cache_key":session});
198 if let Some(effort) = reasoning {
199 value["reasoning"] = json!({"effort":effort, "summary":"auto"});
200 }
201 value
202 }
203 })
204 }
205
206 async fn stream(
207 &self,
208 mut request: CompletionRequest,
209 observed: AdapterContext,
210 ) -> Result<CompletionStream, Classified> {
211 match self {
212 Self::Openai(model, _) => model.stream_observed(request, observed),
213 Self::Anthropic(model, _) => {
214 request.max_tokens = Some(ANTHROPIC_OUTPUT_TOKENS);
215 model.stream_observed(request, observed)
216 }
217 Self::Codex(client) => {
219 let model = client.model().await.map_err(|error| Classified {
220 code: if error.transient {
221 Code::Transient
222 } else {
223 Code::Authentication
224 },
225 message: error.message,
226 retry_after: None,
227 })?;
228 model.stream_observed(request, observed)
229 }
230 }
231 .map_err(|error| self.classify(&error))
232 }
233
234 fn sent(&self) -> Vec<String> {
236 match self {
237 Self::Openai(_, key) | Self::Anthropic(_, key) => vec![key.clone()],
238 Self::Codex(client) => client.sent(),
239 }
240 }
241
242 fn classify(&self, error: &ProviderError) -> Classified {
246 let text = match self {
247 Self::Codex(_) => codex::diagnostic_text(error),
248 _ => diagnostic_text(error),
249 };
250 let sent = self.sent();
251 let sent: Vec<&str> = sent.iter().map(String::as_str).collect();
252 let message = crate::secret::server_text(&text, &sent, MAX_DIAGNOSTIC_CHARS);
253 let report = error.report();
254 let status = report.http_status;
255 let text = format!("{} {text}", report.code.as_deref().unwrap_or_default()).to_lowercase();
257 let any = |terms: &[&str]| terms.iter().any(|term| text.contains(term));
258 let code = if matches!(self, Self::Codex(_)) && codex::usage_limited(error)
259 || any(&["quota", "billing", "credit balance", "usage limit"])
260 {
261 Code::Quota
262 } else if matches!(status, Some(401 | 403))
263 || any(&[
264 "authentication",
265 "unauthorized",
266 "permission",
267 "invalid_api_key",
268 "api key",
269 ])
270 {
271 Code::Authentication
272 } else if status == Some(429) || any(&["rate_limit", "rate limit"]) {
273 Code::RateLimit
274 } else if error.is_retryable()
275 || matches!(self, Self::Openai(..) | Self::Codex(_)) && codex::overloaded(error)
276 || status.is_some_and(|status| status >= 500)
277 || (status.is_none()
278 && matches!(
279 report.code.as_deref(),
280 Some(
283 "overloaded_error" | "server_is_overloaded" | "server_error" | "api_error"
284 )
285 ))
286 {
287 Code::Transient
288 } else {
289 Code::ProviderError
290 };
291 Classified {
292 code,
293 message,
294 retry_after: retry_after(error),
295 }
296 }
297
298 pub async fn turn(
303 &self,
304 request: &CompletionRequest,
305 context: Turn<'_>,
306 attempts: &mut Vec<Attempt>,
307 ) -> Result<CompletionResponse, Failure> {
308 self.turn_observed(request, context, attempts, &mut |_| {})
309 .await
310 }
311
312 pub async fn turn_observed(
314 &self,
315 request: &CompletionRequest,
316 context: Turn<'_>,
317 attempts: &mut Vec<Attempt>,
318 observe: &mut impl FnMut(crate::agent::session::Delivery),
319 ) -> Result<CompletionResponse, Failure> {
320 for attempt in 1..=ATTEMPTS {
321 if context.cancellation.is_cancelled() {
322 return Err(Failure::cancelled());
323 }
324 if Instant::now() >= context.deadline {
325 return Err(Failure::timeout());
326 }
327 let usage = Arc::new(ReportedUsage::default());
328 let observed = AdapterContext::new(
329 usage.clone(),
330 Subject::default(),
331 format!("turn-{}", context.number),
332 );
333 let mut emitted = false;
334 let sink = delivery::Sink::default();
335 let (result, partial) = sink
337 .scope(async {
338 let cancelled = || Classified::from(Failure::cancelled());
339 let timeout = || Classified::from(Failure::timeout());
340 let deadline = context.deadline;
341 let opened = tokio::select! {
342 biased;
343 _ = context.cancellation.cancelled() => return (Err(cancelled()), None),
344 _ = tokio::time::sleep_until(deadline) => return (Err(timeout()), None),
345 opened = self.stream(request.clone(), observed) => opened,
346 };
347 let mut stream = match opened {
348 Ok(stream) => stream,
349 Err(failure) => return (Err(failure), None),
350 };
351 let mut cadence = tokio::time::interval(delivery::CADENCE);
352 cadence.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
353 let read = loop {
354 tokio::select! {
355 biased;
356 _ = context.cancellation.cancelled() => break Err(cancelled()),
357 _ = tokio::time::sleep_until(deadline) => break Err(timeout()),
358 _ = cadence.tick() => {
359 for event in sink.take(context.number, attempt) {
360 observe(event);
361 }
362 }
363 item = stream.next() => match item {
364 Some(Ok(item)) => {
365 emitted = true;
366 sink.normalized(&item);
367 if sink.ready() {
368 for event in sink.take(context.number, attempt) {
369 observe(event);
370 }
371 }
372 }
373 Some(Err(error)) => break Err(self.classify(&error)),
374 None => break Ok(()),
375 }
376 }
377 };
378 for event in sink.take(context.number, attempt) {
379 observe(event);
380 }
381 let partial = stream.partial();
382 let result = match read {
383 Ok(()) => stream.finish().await.map_err(|error| self.classify(&error)),
384 Err(error) => Err(error),
385 };
386 let final_response = result.as_ref().ok().unwrap_or(&partial);
387 if result.is_ok() {
388 sink.final_response(&final_response.choice);
389 }
390 for event in sink.take(context.number, attempt) {
391 observe(event);
392 }
393 if let Err(error) = &result {
394 observe(crate::agent::session::Delivery {
395 turn: context.number,
396 attempt,
397 block: crate::types::DeliveryBlock::attempt(),
398 kind: crate::agent::session::DeliveryKind::Text,
399 text: error.message.clone(),
400 state: crate::agent::session::DeliveryState::Interrupted,
401 });
402 }
403 (result, Some(partial))
404 })
405 .await;
406 let mut counters = usage.0.lock().unwrap().clone();
407 if let Some(partial) = &partial {
408 add_cache_usage(&mut counters, partial);
409 }
410 if let Ok(response) = &result {
411 add_cache_usage(&mut counters, response);
412 }
413 let used_tokens = counters.values().any(|value| *value > 0);
414 attempts.push(Attempt {
415 turn: context.number,
416 attempt,
417 usage: counters.into(),
418 error: result.as_ref().err().map(|failure| failure.message.clone()),
419 error_code: result.as_ref().err().map(|failure| failure.code),
420 });
421 let failure = match result {
422 Ok(response) => return Ok(response),
423 Err(failure) => failure,
424 };
425 if !failure.code.retryable() || emitted || used_tokens || attempt == ATTEMPTS {
426 return Err(failure.into());
427 }
428 let wait = failure
429 .retry_after
430 .unwrap_or(RETRY_BACKOFF_BASE * (1 << (attempt - 1)));
431 if failure.retry_after.is_some()
433 && Instant::now()
434 .checked_add(wait)
435 .is_none_or(|end| end >= context.deadline)
436 {
437 return Err(Failure::new(
438 failure.code,
439 format!(
440 "{}. The provider asked to retry after {:.1} s, past the review deadline.",
441 failure.message.trim_end_matches('.'),
442 wait.as_secs_f64()
443 ),
444 ));
445 }
446 tokio::select! {
447 _ = context.cancellation.cancelled() => return Err(Failure::cancelled()),
448 _ = tokio::time::sleep_until(context.deadline) => return Err(Failure::timeout()),
449 _ = tokio::time::sleep(wait) => {},
450 }
451 }
452 unreachable!()
453 }
454}
455
456pub struct Turn<'a> {
457 pub number: usize,
458 pub deadline: Instant,
459 pub cancellation: &'a CancellationToken,
460}
461
462fn add_cache_usage(
463 counters: &mut std::collections::BTreeMap<String, u64>,
464 response: &CompletionResponse,
465) {
466 insert_counter(
467 counters,
468 "cacheWriteTokens",
469 response.usage.cache_creation_input_tokens,
470 );
471 insert_counter(
472 counters,
473 "cacheWrite1hTokens",
474 response
475 .raw
476 .pointer("/usage/cache_creation/ephemeral_1h_input_tokens")
477 .and_then(Value::as_u64),
478 );
479}
480
481pub fn validate_response(response: &CompletionResponse, model: &str) -> Result<(), String> {
482 if response
483 .model
484 .as_deref()
485 .is_some_and(|actual| actual != model)
486 {
487 return Err(format!(
488 "Provider model mismatch: requested {model:?}, received {:?}.",
489 response.model
490 ));
491 }
492 if !matches!(
493 response.finish_reason(),
494 Some(FinishReason::Stop | FinishReason::ToolCalls)
495 ) {
496 return Err(format!(
497 "Provider response was incomplete: {:?}.",
498 response.finish_reason()
499 ));
500 }
501 Ok(())
502}
503
504const MILLIS_PER_SECOND: u32 = 1000;
506
507fn retry_after(error: &ProviderError) -> Option<Duration> {
510 wait_header(
511 error.provider_response_headers()?,
512 time::OffsetDateTime::now_utc(),
513 )
514}
515
516fn wait_header(
517 headers: &rig_core::http_client::HeaderMap,
518 now: time::OffsetDateTime,
519) -> Option<Duration> {
520 let header = |name: &str| headers.get(name)?.to_str().ok().map(str::trim);
521 let seconds = |value: &str| {
522 value
523 .parse::<f64>()
524 .ok()
525 .filter(|value| value.is_finite() && *value >= 0.0)
526 .and_then(|value| Duration::try_from_secs_f64(value).ok())
527 };
528 if let Some(wait) = header("retry-after-ms").and_then(seconds) {
529 return Some(wait / MILLIS_PER_SECOND);
530 }
531 let value = header("retry-after")?;
532 seconds(value).or_else(|| {
533 let at = time::OffsetDateTime::parse(value, &time::format_description::well_known::Rfc2822)
534 .ok()?;
535 Some(Duration::try_from(at - now).unwrap_or_default())
536 })
537}
538
539fn diagnostic_text(error: &ProviderError) -> String {
542 clean_text(&message(error))
543}
544
545fn diagnostic_hiding(error: &ProviderError, secrets: &[&str]) -> String {
548 crate::secret::server_text(&diagnostic_text(error), secrets, MAX_DIAGNOSTIC_CHARS)
549}
550
551fn message(error: &ProviderError) -> String {
552 error
553 .provider_response()
554 .and_then(|response| serde_json::from_str::<Value>(&response.body).ok())
555 .and_then(|body| {
556 body.pointer("/error/message")
557 .or_else(|| body.get("message"))
558 .and_then(Value::as_str)
559 .map(str::to_owned)
560 })
561 .unwrap_or_else(|| error.to_string())
562}
563
564pub fn key_variable(backend: Backend) -> Option<&'static str> {
566 match backend {
567 Backend::Openai => Some("OPENAI_API_KEY"),
568 Backend::Anthropic => Some("ANTHROPIC_API_KEY"),
569 Backend::Codex => None,
570 }
571}
572
573pub fn key_present(backend: Backend) -> bool {
576 key_variable(backend)
577 .and_then(crate::platform::environment::var_text)
578 .is_some_and(|key| !key.trim().is_empty())
579}
580
581pub fn base_url_variable(backend: Backend) -> &'static str {
583 match backend {
584 Backend::Openai => "ARTIFACTIZE_OPENAI_BASE_URL",
585 Backend::Anthropic => "ARTIFACTIZE_ANTHROPIC_BASE_URL",
586 Backend::Codex => "ARTIFACTIZE_CODEX_BASE_URL",
587 }
588}
589
590const TEST_ENDPOINTS: [&str; 4] = [
592 "ARTIFACTIZE_OPENAI_BASE_URL",
593 "ARTIFACTIZE_ANTHROPIC_BASE_URL",
594 "ARTIFACTIZE_CODEX_BASE_URL",
595 crate::auth::codex::AUTH_URL_VARIABLE,
596];
597
598pub fn test_endpoint(backend: Backend) -> Result<Option<String>, String> {
600 variable_endpoint(base_url_variable(backend))
601}
602
603pub fn variable_endpoint(variable: &str) -> Result<Option<String>, String> {
606 let Some(value) = crate::platform::environment::var(variable).filter(|value| !value.is_empty())
607 else {
608 return Ok(None);
609 };
610 value
611 .to_str()
612 .ok_or_else(|| "it is not UTF-8".to_owned())
613 .and_then(loopback_url)
614 .map(Some)
615 .map_err(|error| format!("{variable}: {error}."))
616}
617
618pub fn active_test_endpoint() -> Option<&'static str> {
620 TEST_ENDPOINTS.into_iter().find(|variable| {
621 crate::platform::environment::var(variable).is_some_and(|value| !value.is_empty())
622 })
623}
624
625fn loopback_url(value: &str) -> Result<String, String> {
628 let url = url::Url::parse(value).map_err(|_| "it is not a URL".to_owned())?;
629 let loopback = match url.host() {
630 Some(url::Host::Domain(host)) => host.eq_ignore_ascii_case("localhost"),
631 Some(url::Host::Ipv4(ip)) => IpAddr::V4(ip).is_loopback(),
632 Some(url::Host::Ipv6(ip)) => IpAddr::V6(ip).is_loopback(),
633 None => false,
634 };
635 if !matches!(url.scheme(), "http" | "https") || !loopback {
636 return Err("a test endpoint must be http(s) on localhost, 127.0.0.0/8 or [::1]".into());
637 }
638 if !url.username().is_empty()
639 || url.password().is_some()
640 || url.query().is_some()
641 || url.fragment().is_some()
642 {
643 return Err("a test endpoint has no credentials, query or fragment".into());
644 }
645 Ok(url.as_str().trim_end_matches('/').to_owned())
646}
647
648fn base_url(backend: Backend) -> Result<String, String> {
650 Ok(test_endpoint(backend)?.unwrap_or_else(|| {
651 match backend {
652 Backend::Openai => "https://api.openai.com/v1",
653 Backend::Anthropic => "https://api.anthropic.com",
654 Backend::Codex => codex::BASE_URL,
655 }
656 .into()
657 }))
658}
659
660fn api_key(backend: Backend, purpose: &str) -> Result<String, String> {
662 let variable = key_variable(backend).expect("an API-key backend");
663 crate::platform::environment::var_text(variable)
664 .filter(|key| !key.trim().is_empty())
665 .ok_or_else(|| format!("{variable} is required {purpose}."))
666}
667
668fn http_client() -> Result<reqwest::Client, String> {
669 reqwest::Client::builder()
670 .retry(reqwest::retry::never())
671 .redirect(reqwest::redirect::Policy::none())
672 .build()
673 .map_err(|error| error.to_string())
674}
675
676fn clean_text(message: &str) -> String {
678 String::from_utf8(crate::runtime::clean_output(message.as_bytes()))
679 .expect("clean output is UTF-8")
680}