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 || status.is_some_and(|status| status >= 500)
276 || (status.is_none()
277 && matches!(
278 report.code.as_deref(),
279 Some("overloaded_error" | "server_error" | "api_error")
280 ))
281 {
282 Code::Transient
283 } else {
284 Code::ProviderError
285 };
286 Classified {
287 code,
288 message,
289 retry_after: retry_after(error),
290 }
291 }
292
293 pub async fn turn(
298 &self,
299 request: &CompletionRequest,
300 context: Turn<'_>,
301 attempts: &mut Vec<Attempt>,
302 ) -> Result<CompletionResponse, Failure> {
303 self.turn_observed(request, context, attempts, &mut |_| {})
304 .await
305 }
306
307 pub async fn turn_observed(
309 &self,
310 request: &CompletionRequest,
311 context: Turn<'_>,
312 attempts: &mut Vec<Attempt>,
313 observe: &mut impl FnMut(crate::agent::session::Delivery),
314 ) -> Result<CompletionResponse, Failure> {
315 for attempt in 1..=ATTEMPTS {
316 if context.cancellation.is_cancelled() {
317 return Err(Failure::cancelled());
318 }
319 if Instant::now() >= context.deadline {
320 return Err(Failure::timeout());
321 }
322 let usage = Arc::new(ReportedUsage::default());
323 let observed = AdapterContext::new(
324 usage.clone(),
325 Subject::default(),
326 format!("turn-{}", context.number),
327 );
328 let mut emitted = false;
329 let sink = delivery::Sink::default();
330 let (result, partial) = sink
332 .scope(async {
333 let cancelled = || Classified::from(Failure::cancelled());
334 let timeout = || Classified::from(Failure::timeout());
335 let deadline = context.deadline;
336 let opened = tokio::select! {
337 biased;
338 _ = context.cancellation.cancelled() => return (Err(cancelled()), None),
339 _ = tokio::time::sleep_until(deadline) => return (Err(timeout()), None),
340 opened = self.stream(request.clone(), observed) => opened,
341 };
342 let mut stream = match opened {
343 Ok(stream) => stream,
344 Err(failure) => return (Err(failure), None),
345 };
346 let mut cadence = tokio::time::interval(delivery::CADENCE);
347 cadence.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
348 let read = loop {
349 tokio::select! {
350 biased;
351 _ = context.cancellation.cancelled() => break Err(cancelled()),
352 _ = tokio::time::sleep_until(deadline) => break Err(timeout()),
353 _ = cadence.tick() => {
354 for event in sink.take(context.number, attempt) {
355 observe(event);
356 }
357 }
358 item = stream.next() => match item {
359 Some(Ok(item)) => {
360 emitted = true;
361 sink.normalized(&item);
362 if sink.ready() {
363 for event in sink.take(context.number, attempt) {
364 observe(event);
365 }
366 }
367 }
368 Some(Err(error)) => break Err(self.classify(&error)),
369 None => break Ok(()),
370 }
371 }
372 };
373 for event in sink.take(context.number, attempt) {
374 observe(event);
375 }
376 let partial = stream.partial();
377 let result = match read {
378 Ok(()) => stream.finish().await.map_err(|error| self.classify(&error)),
379 Err(error) => Err(error),
380 };
381 let final_response = result.as_ref().ok().unwrap_or(&partial);
382 if result.is_ok() {
383 sink.final_response(&final_response.choice);
384 }
385 for event in sink.take(context.number, attempt) {
386 observe(event);
387 }
388 if let Err(error) = &result {
389 observe(crate::agent::session::Delivery {
390 turn: context.number,
391 attempt,
392 block: crate::types::DeliveryBlock::attempt(),
393 kind: crate::agent::session::DeliveryKind::Text,
394 text: error.message.clone(),
395 state: crate::agent::session::DeliveryState::Interrupted,
396 });
397 }
398 (result, Some(partial))
399 })
400 .await;
401 let mut counters = usage.0.lock().unwrap().clone();
402 if let Some(partial) = &partial {
403 add_cache_usage(&mut counters, partial);
404 }
405 if let Ok(response) = &result {
406 add_cache_usage(&mut counters, response);
407 }
408 let used_tokens = counters.values().any(|value| *value > 0);
409 attempts.push(Attempt {
410 turn: context.number,
411 attempt,
412 usage: counters.into(),
413 error: result.as_ref().err().map(|failure| failure.message.clone()),
414 error_code: result.as_ref().err().map(|failure| failure.code),
415 });
416 let failure = match result {
417 Ok(response) => return Ok(response),
418 Err(failure) => failure,
419 };
420 if !failure.code.retryable() || emitted || used_tokens || attempt == ATTEMPTS {
421 return Err(failure.into());
422 }
423 let wait = failure
424 .retry_after
425 .unwrap_or(RETRY_BACKOFF_BASE * (1 << (attempt - 1)));
426 if failure.retry_after.is_some()
428 && Instant::now()
429 .checked_add(wait)
430 .is_none_or(|end| end >= context.deadline)
431 {
432 return Err(Failure::new(
433 failure.code,
434 format!(
435 "{}. The provider asked to retry after {:.1} s, past the review deadline.",
436 failure.message.trim_end_matches('.'),
437 wait.as_secs_f64()
438 ),
439 ));
440 }
441 tokio::select! {
442 _ = context.cancellation.cancelled() => return Err(Failure::cancelled()),
443 _ = tokio::time::sleep_until(context.deadline) => return Err(Failure::timeout()),
444 _ = tokio::time::sleep(wait) => {},
445 }
446 }
447 unreachable!()
448 }
449}
450
451pub struct Turn<'a> {
452 pub number: usize,
453 pub deadline: Instant,
454 pub cancellation: &'a CancellationToken,
455}
456
457fn add_cache_usage(
458 counters: &mut std::collections::BTreeMap<String, u64>,
459 response: &CompletionResponse,
460) {
461 insert_counter(
462 counters,
463 "cacheWriteTokens",
464 response.usage.cache_creation_input_tokens,
465 );
466 insert_counter(
467 counters,
468 "cacheWrite1hTokens",
469 response
470 .raw
471 .pointer("/usage/cache_creation/ephemeral_1h_input_tokens")
472 .and_then(Value::as_u64),
473 );
474}
475
476pub fn validate_response(response: &CompletionResponse, model: &str) -> Result<(), String> {
477 if response
478 .model
479 .as_deref()
480 .is_some_and(|actual| actual != model)
481 {
482 return Err(format!(
483 "Provider model mismatch: requested {model:?}, received {:?}.",
484 response.model
485 ));
486 }
487 if !matches!(
488 response.finish_reason(),
489 Some(FinishReason::Stop | FinishReason::ToolCalls)
490 ) {
491 return Err(format!(
492 "Provider response was incomplete: {:?}.",
493 response.finish_reason()
494 ));
495 }
496 Ok(())
497}
498
499const MILLIS_PER_SECOND: u32 = 1000;
501
502fn retry_after(error: &ProviderError) -> Option<Duration> {
505 wait_header(
506 error.provider_response_headers()?,
507 time::OffsetDateTime::now_utc(),
508 )
509}
510
511fn wait_header(
512 headers: &rig_core::http_client::HeaderMap,
513 now: time::OffsetDateTime,
514) -> Option<Duration> {
515 let header = |name: &str| headers.get(name)?.to_str().ok().map(str::trim);
516 let seconds = |value: &str| {
517 value
518 .parse::<f64>()
519 .ok()
520 .filter(|value| value.is_finite() && *value >= 0.0)
521 .and_then(|value| Duration::try_from_secs_f64(value).ok())
522 };
523 if let Some(wait) = header("retry-after-ms").and_then(seconds) {
524 return Some(wait / MILLIS_PER_SECOND);
525 }
526 let value = header("retry-after")?;
527 seconds(value).or_else(|| {
528 let at = time::OffsetDateTime::parse(value, &time::format_description::well_known::Rfc2822)
529 .ok()?;
530 Some(Duration::try_from(at - now).unwrap_or_default())
531 })
532}
533
534fn diagnostic_text(error: &ProviderError) -> String {
537 clean_text(&message(error))
538}
539
540fn diagnostic_hiding(error: &ProviderError, secrets: &[&str]) -> String {
543 crate::secret::server_text(&diagnostic_text(error), secrets, MAX_DIAGNOSTIC_CHARS)
544}
545
546fn message(error: &ProviderError) -> String {
547 error
548 .provider_response()
549 .and_then(|response| serde_json::from_str::<Value>(&response.body).ok())
550 .and_then(|body| {
551 body.pointer("/error/message")
552 .or_else(|| body.get("message"))
553 .and_then(Value::as_str)
554 .map(str::to_owned)
555 })
556 .unwrap_or_else(|| error.to_string())
557}
558
559pub fn key_variable(backend: Backend) -> Option<&'static str> {
561 match backend {
562 Backend::Openai => Some("OPENAI_API_KEY"),
563 Backend::Anthropic => Some("ANTHROPIC_API_KEY"),
564 Backend::Codex => None,
565 }
566}
567
568pub fn key_present(backend: Backend) -> bool {
571 key_variable(backend)
572 .and_then(crate::platform::environment::var_text)
573 .is_some_and(|key| !key.trim().is_empty())
574}
575
576pub fn base_url_variable(backend: Backend) -> &'static str {
578 match backend {
579 Backend::Openai => "ARTIFACTIZE_OPENAI_BASE_URL",
580 Backend::Anthropic => "ARTIFACTIZE_ANTHROPIC_BASE_URL",
581 Backend::Codex => "ARTIFACTIZE_CODEX_BASE_URL",
582 }
583}
584
585const TEST_ENDPOINTS: [&str; 4] = [
587 "ARTIFACTIZE_OPENAI_BASE_URL",
588 "ARTIFACTIZE_ANTHROPIC_BASE_URL",
589 "ARTIFACTIZE_CODEX_BASE_URL",
590 crate::auth::codex::AUTH_URL_VARIABLE,
591];
592
593pub fn test_endpoint(backend: Backend) -> Result<Option<String>, String> {
595 variable_endpoint(base_url_variable(backend))
596}
597
598pub fn variable_endpoint(variable: &str) -> Result<Option<String>, String> {
601 let Some(value) = crate::platform::environment::var(variable).filter(|value| !value.is_empty())
602 else {
603 return Ok(None);
604 };
605 value
606 .to_str()
607 .ok_or_else(|| "it is not UTF-8".to_owned())
608 .and_then(loopback_url)
609 .map(Some)
610 .map_err(|error| format!("{variable}: {error}."))
611}
612
613pub fn active_test_endpoint() -> Option<&'static str> {
615 TEST_ENDPOINTS.into_iter().find(|variable| {
616 crate::platform::environment::var(variable).is_some_and(|value| !value.is_empty())
617 })
618}
619
620fn loopback_url(value: &str) -> Result<String, String> {
623 let url = url::Url::parse(value).map_err(|_| "it is not a URL".to_owned())?;
624 let loopback = match url.host() {
625 Some(url::Host::Domain(host)) => host.eq_ignore_ascii_case("localhost"),
626 Some(url::Host::Ipv4(ip)) => IpAddr::V4(ip).is_loopback(),
627 Some(url::Host::Ipv6(ip)) => IpAddr::V6(ip).is_loopback(),
628 None => false,
629 };
630 if !matches!(url.scheme(), "http" | "https") || !loopback {
631 return Err("a test endpoint must be http(s) on localhost, 127.0.0.0/8 or [::1]".into());
632 }
633 if !url.username().is_empty()
634 || url.password().is_some()
635 || url.query().is_some()
636 || url.fragment().is_some()
637 {
638 return Err("a test endpoint has no credentials, query or fragment".into());
639 }
640 Ok(url.as_str().trim_end_matches('/').to_owned())
641}
642
643fn base_url(backend: Backend) -> Result<String, String> {
645 Ok(test_endpoint(backend)?.unwrap_or_else(|| {
646 match backend {
647 Backend::Openai => "https://api.openai.com/v1",
648 Backend::Anthropic => "https://api.anthropic.com",
649 Backend::Codex => codex::BASE_URL,
650 }
651 .into()
652 }))
653}
654
655fn api_key(backend: Backend, purpose: &str) -> Result<String, String> {
657 let variable = key_variable(backend).expect("an API-key backend");
658 crate::platform::environment::var_text(variable)
659 .filter(|key| !key.trim().is_empty())
660 .ok_or_else(|| format!("{variable} is required {purpose}."))
661}
662
663fn http_client() -> Result<reqwest::Client, String> {
664 reqwest::Client::builder()
665 .retry(reqwest::retry::never())
666 .redirect(reqwest::redirect::Policy::none())
667 .build()
668 .map_err(|error| error.to_string())
669}
670
671fn clean_text(message: &str) -> String {
673 String::from_utf8(crate::runtime::clean_output(message.as_bytes()))
674 .expect("clean output is UTF-8")
675}