| 1 | //! LLM Client Trait and Retry Logic |
| 2 | //! |
| 3 | //! This module provides a unified interface for LLM providers with robust retry logic, |
| 4 | //! exponential backoff, and proper error classification. |
| 5 | //! |
| 6 | //! # Architecture |
| 7 | //! |
| 8 | //! - `LlmClient` trait: Async interface for LLM providers (DeepSeek, `OpenAI`, etc.) |
| 9 | //! - `RetryConfig`: Configurable retry behavior with exponential backoff and jitter |
| 10 | //! - `LlmError`: Classified errors with retryability information |
| 11 | |
| 12 | //! - `with_retry`: Generic retry wrapper for any async operation |
| 13 | //! |
| 14 | //! # Example |
| 15 | //! |
| 16 | //! ```ignore |
| 17 | //! use crate::llm_client::{LlmClient, RetryConfig, with_retry}; |
| 18 | //! |
| 19 | //! let config = RetryConfig::default(); |
| 20 | //! let result = with_retry(&config, || async { |
| 21 | //! client.create_message(request).await |
| 22 | //! }, None).await; |
| 23 | //! ``` |
| 24 | |
| 25 | use crate::config::RetryPolicy; |
| 26 | use anyhow::Result; |
| 27 | use codewhale_models::{MessageRequest, MessageResponse, StreamEvent}; |
| 28 | use serde_json::Value; |
| 29 | use std::future::Future; |
| 30 | use std::pin::Pin; |
| 31 | use std::time::{Duration, Instant}; |
| 32 | use uuid::Uuid; |
| 33 | |
| 34 | #[cfg(test)] |
| 35 | pub mod mock; |
| 36 | |
| 37 | // === LlmClient Trait === |
| 38 | |
| 39 | /// Type alias for boxed stream of SSE events |
| 40 | pub type StreamEventBox = |
| 41 | Pin<Box<dyn futures_util::Stream<Item = Result<StreamEvent>> + Send + 'static>>; |
| 42 | |
| 43 | /// Unified interface for LLM providers. |
| 44 | /// |
| 45 | /// This trait abstracts over different LLM APIs (DeepSeek, `OpenAI`, etc.) |
| 46 | /// allowing the agent to work with any provider that implements this interface. |
| 47 | /// |
| 48 | /// # Implementation Notes |
| 49 | /// |
| 50 | /// - All methods are async and require `Send + Sync` for thread safety |
| 51 | /// - The `create_message_stream` method returns a pinned boxed stream for SSE |
| 52 | /// - Implementations should handle their own authentication and base URL configuration |
| 53 | #[allow(async_fn_in_trait, dead_code)] // Trait methods are part of the LLM provider interface |
| 54 | pub trait LlmClient: Send + Sync { |
| 55 | /// Returns the provider name (e.g., "openai", "deepseek") |
| 56 | fn provider_name(&self) -> &'static str; |
| 57 | |
| 58 | /// Returns the model identifier being used |
| 59 | fn model(&self) -> &str; |
| 60 | |
| 61 | /// Creates a non-streaming message completion |
| 62 | fn create_message( |
| 63 | &self, |
| 64 | request: MessageRequest, |
| 65 | ) -> impl Future<Output = Result<MessageResponse>> + Send; |
| 66 | |
| 67 | /// Dispatch a fresh request. Clients with a local response cache must |
| 68 | /// override this; authorization decisions cannot reuse earlier answers. |
| 69 | fn create_message_uncached( |
| 70 | &self, |
| 71 | request: MessageRequest, |
| 72 | ) -> impl Future<Output = Result<MessageResponse>> + Send { |
| 73 | self.create_message(request) |
| 74 | } |
| 75 | |
| 76 | /// Creates a streaming message completion |
| 77 | /// |
| 78 | /// Returns a stream of SSE events that should be consumed until completion. |
| 79 | fn create_message_stream( |
| 80 | &self, |
| 81 | request: MessageRequest, |
| 82 | ) -> impl Future<Output = Result<StreamEventBox>> + Send; |
| 83 | |
| 84 | /// Optional health check to verify API connectivity |
| 85 | fn health_check(&self) -> impl Future<Output = Result<bool>> + Send { |
| 86 | async { Ok(true) } |
| 87 | } |
| 88 | |
| 89 | /// The concrete base URL requests go to, when the implementation knows it. |
| 90 | /// |
| 91 | /// Background cost accrual uses this for billing provenance only: it is |
| 92 | /// reduced to a non-secret surface classification and a SHA-256 fingerprint |
| 93 | /// before being recorded, and the URL itself is never persisted or logged |
| 94 | /// (#4318). The default is `None` so an implementation that cannot report a |
| 95 | /// stable endpoint yields "unknown endpoint" — which fails closed — rather |
| 96 | /// than being assumed to be the provider's public API. |
| 97 | fn billing_base_url(&self) -> Option<&str> { |
| 98 | None |
| 99 | } |
| 100 | |
| 101 | /// Non-secret limits frozen with the resolved route, when available. |
| 102 | fn route_limits(&self) -> Option<codewhale_config::route::RouteLimits> { |
| 103 | None |
| 104 | } |
| 105 | |
| 106 | /// Output cap for a request sent through this exact client route. |
| 107 | fn effective_max_output_tokens(&self, requested_model: &str) -> u32 { |
| 108 | let route = self.effective_route_envelope(requested_model, chrono::Utc::now()); |
| 109 | crate::route_budget::effective_max_output_tokens_for_route( |
| 110 | route.provider, |
| 111 | &route.model, |
| 112 | self.route_limits(), |
| 113 | ) |
| 114 | } |
| 115 | |
| 116 | /// Freeze the non-secret effective route immediately before a request is |
| 117 | /// dispatched. Implementations with richer configured identity/billing |
| 118 | /// facts should override this fail-closed default. |
| 119 | fn effective_route_envelope( |
| 120 | &self, |
| 121 | requested_model: &str, |
| 122 | dispatched_at: chrono::DateTime<chrono::Utc>, |
| 123 | ) -> crate::cost_status::EffectiveRouteEnvelope { |
| 124 | let provider = crate::config::ProviderKind::parse(self.provider_name()) |
| 125 | .unwrap_or(crate::config::ProviderKind::Custom); |
| 126 | crate::cost_status::EffectiveRouteEnvelope::capture_observed( |
| 127 | provider, |
| 128 | self.provider_name(), |
| 129 | requested_model, |
| 130 | self.billing_base_url(), |
| 131 | dispatched_at, |
| 132 | ) |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | // === Authentication diagnostics === |
| 137 | |
| 138 | #[derive(Debug, Clone, PartialEq, Eq, Default)] |
| 139 | pub struct AuthenticationErrorContext { |
| 140 | pub provider: Option<String>, |
| 141 | pub base_url_authority: Option<String>, |
| 142 | pub model: Option<String>, |
| 143 | pub key_source: Option<String>, |
| 144 | pub key_fingerprint: Option<String>, |
| 145 | pub key_kind: Option<String>, |
| 146 | /// The one command or action that replaces the rejected credential. |
| 147 | pub fix: Option<String>, |
| 148 | } |
| 149 | |
| 150 | impl AuthenticationErrorContext { |
| 151 | #[must_use] |
| 152 | pub fn from_parts( |
| 153 | provider: Option<&str>, |
| 154 | base_url: Option<&str>, |
| 155 | model: Option<&str>, |
| 156 | key_source: Option<&str>, |
| 157 | api_key: Option<&str>, |
| 158 | ) -> Self { |
| 159 | let api_key = api_key.and_then(non_empty_trimmed); |
| 160 | Self { |
| 161 | provider: provider.and_then(non_empty_trimmed).map(str::to_string), |
| 162 | base_url_authority: base_url.and_then(base_url_authority), |
| 163 | model: model.and_then(non_empty_trimmed).map(str::to_string), |
| 164 | key_source: key_source.and_then(non_empty_trimmed).map(str::to_string), |
| 165 | key_fingerprint: api_key.map(redacted_key_fingerprint), |
| 166 | key_kind: api_key.map(classify_api_key_prefix).map(str::to_string), |
| 167 | fix: None, |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | #[must_use] |
| 172 | pub fn with_fix(mut self, fix: impl Into<String>) -> Self { |
| 173 | self.fix = Some(fix.into()).filter(|fix: &String| !fix.trim().is_empty()); |
| 174 | self |
| 175 | } |
| 176 | |
| 177 | fn is_empty(&self) -> bool { |
| 178 | self.provider.is_none() |
| 179 | && self.base_url_authority.is_none() |
| 180 | && self.model.is_none() |
| 181 | && self.key_source.is_none() |
| 182 | && self.key_fingerprint.is_none() |
| 183 | && self.key_kind.is_none() |
| 184 | && self.fix.is_none() |
| 185 | } |
| 186 | |
| 187 | fn detail_segments(&self) -> Vec<String> { |
| 188 | let mut segments = Vec::new(); |
| 189 | if let Some(provider) = self.provider.as_deref() { |
| 190 | segments.push(format!("provider: {provider}")); |
| 191 | } |
| 192 | if let Some(authority) = self.base_url_authority.as_deref() { |
| 193 | segments.push(format!("base URL authority: {authority}")); |
| 194 | } |
| 195 | if let Some(model) = self.model.as_deref() { |
| 196 | segments.push(format!("model: {model}")); |
| 197 | } |
| 198 | if let Some(source) = self.key_source.as_deref() { |
| 199 | segments.push(format!("key source: {source}")); |
| 200 | } |
| 201 | if let Some(fingerprint) = self.key_fingerprint.as_deref() { |
| 202 | segments.push(format!("key fingerprint: {fingerprint}")); |
| 203 | } |
| 204 | if let Some(kind) = self.key_kind.as_deref() { |
| 205 | segments.push(format!("key type: {kind}")); |
| 206 | } |
| 207 | if let Some(fix) = self.fix.as_deref() { |
| 208 | segments.push(format!("fix: {fix}")); |
| 209 | } |
| 210 | segments |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 215 | pub struct AuthenticationErrorDetail { |
| 216 | message: String, |
| 217 | context: Option<AuthenticationErrorContext>, |
| 218 | } |
| 219 | |
| 220 | impl AuthenticationErrorDetail { |
| 221 | #[must_use] |
| 222 | pub fn new(message: impl Into<String>) -> Self { |
| 223 | Self { |
| 224 | message: message.into(), |
| 225 | context: None, |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | #[must_use] |
| 230 | pub fn with_context( |
| 231 | message: impl Into<String>, |
| 232 | context: Option<AuthenticationErrorContext>, |
| 233 | ) -> Self { |
| 234 | let context = context.filter(|context| !context.is_empty()); |
| 235 | Self { |
| 236 | message: message.into(), |
| 237 | context, |
| 238 | } |
| 239 | } |
| 240 | |
| 241 | #[must_use] |
| 242 | pub fn to_user_message(&self) -> String { |
| 243 | let Some(context) = self.context.as_ref() else { |
| 244 | return self.message.clone(); |
| 245 | }; |
| 246 | let segments = context.detail_segments(); |
| 247 | if segments.is_empty() { |
| 248 | self.message.clone() |
| 249 | } else { |
| 250 | format!("{} ({})", self.message, segments.join(", ")) |
| 251 | } |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | impl From<String> for AuthenticationErrorDetail { |
| 256 | fn from(message: String) -> Self { |
| 257 | Self::new(message) |
| 258 | } |
| 259 | } |
| 260 | |
| 261 | impl From<&str> for AuthenticationErrorDetail { |
| 262 | fn from(message: &str) -> Self { |
| 263 | Self::new(message) |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | #[must_use] |
| 268 | pub fn classify_api_key_prefix(api_key: &str) -> &'static str { |
| 269 | if api_key.starts_with("tp-") { |
| 270 | "Xiaomi MiMo Token Plan key" |
| 271 | } else { |
| 272 | "API key" |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | fn non_empty_trimmed(value: &str) -> Option<&str> { |
| 277 | let value = value.trim(); |
| 278 | if value.is_empty() { None } else { Some(value) } |
| 279 | } |
| 280 | |
| 281 | pub(crate) fn base_url_authority(base_url: &str) -> Option<String> { |
| 282 | let base_url = non_empty_trimmed(base_url)?; |
| 283 | let without_scheme = base_url |
| 284 | .split_once("://") |
| 285 | .map_or(base_url, |(_, rest)| rest); |
| 286 | let authority = without_scheme.split('/').next().unwrap_or(without_scheme); |
| 287 | let authority = authority |
| 288 | .rsplit_once('@') |
| 289 | .map_or(authority, |(_, authority)| authority); |
| 290 | non_empty_trimmed(authority).map(str::to_string) |
| 291 | } |
| 292 | |
| 293 | fn redacted_key_fingerprint(api_key: &str) -> String { |
| 294 | let api_key = api_key.trim(); |
| 295 | let len = api_key.chars().count(); |
| 296 | match public_key_prefix(api_key) { |
| 297 | Some(prefix) => format!("{prefix}... (len={len})"), |
| 298 | None => format!("unprefixed (len={len})"), |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | fn public_key_prefix(api_key: &str) -> Option<&str> { |
| 303 | ["tp-", "sk-", "hf_", "hf-", "ak-", "rk-"] |
| 304 | .into_iter() |
| 305 | .find(|prefix| api_key.starts_with(prefix)) |
| 306 | } |
| 307 | |
| 308 | #[cfg(test)] |
| 309 | fn redact_api_key_from_message(message: &str, api_key: Option<&str>) -> String { |
| 310 | let Some(api_key) = api_key.and_then(non_empty_trimmed) else { |
| 311 | return message.to_string(); |
| 312 | }; |
| 313 | message.replace(api_key, "[redacted API key]") |
| 314 | } |
| 315 | |
| 316 | // === LlmError - Classified Error Types === |
| 317 | |
| 318 | /// Evidence captured when a provider response explicitly identifies plan quota |
| 319 | /// exhaustion. The private field prevents callers outside this parser module |
| 320 | /// from manufacturing the durable classification from arbitrary text. |
| 321 | #[derive(Debug)] |
| 322 | pub struct QuotaExhaustionError { |
| 323 | message: String, |
| 324 | } |
| 325 | |
| 326 | impl QuotaExhaustionError { |
| 327 | fn from_http_message(message: String) -> Self { |
| 328 | Self { message } |
| 329 | } |
| 330 | |
| 331 | pub(crate) fn into_message(self) -> String { |
| 332 | self.message |
| 333 | } |
| 334 | |
| 335 | /// Append route guidance (which account hit the limit, how to switch) |
| 336 | /// to already-classified evidence. Cannot manufacture the class. |
| 337 | #[must_use] |
| 338 | pub(crate) fn with_guidance(mut self, guidance: &str) -> Self { |
| 339 | self.message = format!("{}\n{guidance}", self.message); |
| 340 | self |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | /// Classified LLM errors with retryability information. |
| 345 | /// |
| 346 | /// This enum categorizes API errors to enable smart retry decisions. |
| 347 | /// Some errors (rate limits, transient server errors) are retryable, |
| 348 | /// while others (auth failures, invalid requests) should fail immediately. |
| 349 | #[derive(Debug)] |
| 350 | pub enum LlmError { |
| 351 | /// Rate limit exceeded (HTTP 429) |
| 352 | /// Contains optional Retry-After duration from server |
| 353 | RateLimited { |
| 354 | message: String, |
| 355 | retry_after: Option<Duration>, |
| 356 | }, |
| 357 | |
| 358 | /// The provider explicitly reported that the account's plan quota is exhausted. |
| 359 | /// |
| 360 | /// Unlike an ordinary 429 rate limit, retrying the same request after a short |
| 361 | /// backoff cannot resolve this condition. This variant is constructed only at |
| 362 | /// a provider HTTP or structured stream-event boundary from explicit quota evidence. |
| 363 | QuotaExhausted(QuotaExhaustionError), |
| 364 | |
| 365 | /// Server error (HTTP 5xx) |
| 366 | ServerError { status: u16, message: String }, |
| 367 | |
| 368 | /// Network connectivity error |
| 369 | NetworkError(String), |
| 370 | |
| 371 | /// Request timed out |
| 372 | Timeout(Duration), |
| 373 | |
| 374 | /// Authentication failed (HTTP 401, selected HTTP 403) |
| 375 | AuthenticationError(AuthenticationErrorDetail), |
| 376 | |
| 377 | /// Authorization or provider-side blocking failed (HTTP 403) |
| 378 | AuthorizationError(String), |
| 379 | |
| 380 | /// Invalid request parameters (HTTP 400) |
| 381 | InvalidRequest { status: u16, message: String }, |
| 382 | |
| 383 | /// Model-specific error (model not found, etc.) |
| 384 | ModelError(String), |
| 385 | |
| 386 | /// Content policy violation (safety filters) |
| 387 | ContentPolicyError(String), |
| 388 | |
| 389 | /// Failed to parse API response |
| 390 | ParseError(String), |
| 391 | |
| 392 | /// Context length exceeded |
| 393 | ContextLengthError(String), |
| 394 | |
| 395 | /// Catch-all for other errors |
| 396 | Other(String), |
| 397 | } |
| 398 | |
| 399 | impl std::fmt::Display for LlmError { |
| 400 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 401 | match self { |
| 402 | LlmError::RateLimited { message, .. } => write!(f, "Rate limit exceeded: {message}"), |
| 403 | LlmError::QuotaExhausted(error) => { |
| 404 | write!(f, "Provider plan quota exhausted: {}", error.message) |
| 405 | } |
| 406 | LlmError::ServerError { status, message } => { |
| 407 | write!(f, "Server error ({status}): {message}") |
| 408 | } |
| 409 | LlmError::NetworkError(msg) => write!(f, "Network error: {msg}"), |
| 410 | LlmError::Timeout(d) => write!(f, "Request timed out after {d:?}"), |
| 411 | LlmError::AuthenticationError(auth) => { |
| 412 | write!(f, "Authentication failed: {}", auth.to_user_message()) |
| 413 | } |
| 414 | LlmError::AuthorizationError(msg) => write!(f, "Authorization failed: {msg}"), |
| 415 | LlmError::InvalidRequest { status, message } => { |
| 416 | write!(f, "Invalid request ({status}): {message}") |
| 417 | } |
| 418 | LlmError::ModelError(msg) => write!(f, "Model error: {msg}"), |
| 419 | LlmError::ContentPolicyError(msg) => write!(f, "Content policy violation: {msg}"), |
| 420 | LlmError::ParseError(msg) => write!(f, "Response parsing error: {msg}"), |
| 421 | LlmError::ContextLengthError(msg) => write!(f, "Context length exceeded: {msg}"), |
| 422 | LlmError::Other(msg) => write!(f, "LLM error: {msg}"), |
| 423 | } |
| 424 | } |
| 425 | } |
| 426 | |
| 427 | impl std::error::Error for LlmError {} |
| 428 | |
| 429 | impl LlmError { |
| 430 | /// Determines if this error is potentially transient and worth retrying. |
| 431 | /// |
| 432 | /// Retryable errors: |
| 433 | /// - Rate limits (with backoff) |
| 434 | /// - Server errors (5xx) |
| 435 | /// - Network errors (connection issues) |
| 436 | /// - Timeouts |
| 437 | /// |
| 438 | /// Non-retryable errors: |
| 439 | /// - Provider plan quota exhaustion |
| 440 | /// - Authentication failures |
| 441 | /// - Invalid requests |
| 442 | /// - Content policy violations |
| 443 | /// - Context length errors |
| 444 | pub fn is_retryable(&self) -> bool { |
| 445 | matches!( |
| 446 | self, |
| 447 | LlmError::RateLimited { .. } |
| 448 | | LlmError::ServerError { .. } |
| 449 | | LlmError::NetworkError(_) |
| 450 | | LlmError::Timeout(_) |
| 451 | ) |
| 452 | } |
| 453 | |
| 454 | /// Returns the server-suggested retry delay if available. |
| 455 | /// |
| 456 | /// This is typically present for rate limit errors when the server |
| 457 | /// provides a Retry-After header. |
| 458 | pub fn suggested_retry_delay(&self) -> Option<Duration> { |
| 459 | match self { |
| 460 | LlmError::RateLimited { retry_after, .. } => *retry_after, |
| 461 | _ => None, |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | /// Constructs an `LlmError` from HTTP status code and response body. |
| 466 | /// |
| 467 | /// Performs heuristic classification based on: |
| 468 | /// - Status code (429 = rate limit, 401/403 = auth, 499/5xx = transient upstream error) |
| 469 | /// - Response body keywords (`context_length`, `content_policy`, safety, etc.) |
| 470 | pub fn from_http_response(status: u16, body: &str) -> Self { |
| 471 | if let Some(error) = explicit_quota_code(body) |
| 472 | .or_else(|| explicit_quota_code_marker(body)) |
| 473 | .as_deref() |
| 474 | .and_then(Self::from_subscription_sharing_error_code) |
| 475 | { |
| 476 | return error; |
| 477 | } |
| 478 | if matches!(status, 400 | 402 | 429) && has_explicit_quota_evidence(body) { |
| 479 | return LlmError::QuotaExhausted(QuotaExhaustionError::from_http_message( |
| 480 | body.to_string(), |
| 481 | )); |
| 482 | } |
| 483 | |
| 484 | match status { |
| 485 | 429 => LlmError::RateLimited { |
| 486 | message: body.to_string(), |
| 487 | retry_after: None, |
| 488 | }, |
| 489 | 401 => Self::authentication_error(body), |
| 490 | 403 => { |
| 491 | if looks_like_authentication_failure(body) { |
| 492 | Self::authentication_error(body) |
| 493 | } else { |
| 494 | LlmError::AuthorizationError(body.to_string()) |
| 495 | } |
| 496 | } |
| 497 | 400 => { |
| 498 | // Classify 400 errors by examining the response body |
| 499 | let body_lower = body.to_lowercase(); |
| 500 | // An "unsupported parameter" 400 names the offending field |
| 501 | // (often `max_output_tokens` or another *token* field), which |
| 502 | // the generic keyword rules below would misread as a context |
| 503 | // window overflow. Parameter shape errors are invalid |
| 504 | // requests, not prompt-size errors, so they get their own |
| 505 | // branch ahead of the heuristic. |
| 506 | if body_lower.contains("unsupported parameter") |
| 507 | || body_lower.contains("invalid_request_error") |
| 508 | && body_lower.contains("parameter") |
| 509 | { |
| 510 | LlmError::InvalidRequest { |
| 511 | status, |
| 512 | message: body.to_string(), |
| 513 | } |
| 514 | } else if is_context_length_message(&body_lower) { |
| 515 | LlmError::ContextLengthError(body.to_string()) |
| 516 | } else if body_lower.contains("content_policy") |
| 517 | || body_lower.contains("safety") |
| 518 | || body_lower.contains("harmful") |
| 519 | || body_lower.contains("inappropriate") |
| 520 | { |
| 521 | LlmError::ContentPolicyError(body.to_string()) |
| 522 | } else if body_lower.contains("model") && body_lower.contains("not found") { |
| 523 | LlmError::ModelError(body.to_string()) |
| 524 | } else { |
| 525 | LlmError::InvalidRequest { |
| 526 | status, |
| 527 | message: body.to_string(), |
| 528 | } |
| 529 | } |
| 530 | } |
| 531 | 404 => { |
| 532 | if body.to_lowercase().contains("model") { |
| 533 | LlmError::ModelError(body.to_string()) |
| 534 | } else { |
| 535 | LlmError::InvalidRequest { |
| 536 | status, |
| 537 | message: body.to_string(), |
| 538 | } |
| 539 | } |
| 540 | } |
| 541 | // Several OpenAI-compatible gateways use nginx's non-standard |
| 542 | // 499 for an upstream request that was cancelled before response |
| 543 | // streaming began. At this boundary no response body stream has |
| 544 | // been exposed, so it is eligible for the same bounded retry |
| 545 | // policy as a 5xx gateway failure. |
| 546 | 499..=599 => LlmError::ServerError { |
| 547 | status, |
| 548 | message: body.to_string(), |
| 549 | }, |
| 550 | _ => LlmError::Other(format!("HTTP {status}: {body}")), |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | /// Official structured ChatGPT plan errors are terminal account states. |
| 555 | /// Plain text containing quota words is never enough to mint this type. |
| 556 | #[must_use] |
| 557 | pub(crate) fn from_subscription_sharing_error_code(code: &str) -> Option<Self> { |
| 558 | let message = match code { |
| 559 | "subscription_sharing_usage_limit_exceeded" => { |
| 560 | "ChatGPT plan usage limit reached. Check ChatGPT Settings > Usage for your remaining allowance and reset time." |
| 561 | } |
| 562 | "subscription_sharing_usage_unavailable" => { |
| 563 | "ChatGPT plan usage is unavailable. Check ChatGPT Settings > Usage and reconnect if needed." |
| 564 | } |
| 565 | _ => return None, |
| 566 | }; |
| 567 | Some(Self::QuotaExhausted( |
| 568 | QuotaExhaustionError::from_http_message(message.to_string()), |
| 569 | )) |
| 570 | } |
| 571 | |
| 572 | #[must_use] |
| 573 | pub fn authentication_error(message: impl Into<String>) -> Self { |
| 574 | LlmError::AuthenticationError(AuthenticationErrorDetail::new(message)) |
| 575 | } |
| 576 | |
| 577 | #[must_use] |
| 578 | pub fn authentication_error_with_context( |
| 579 | message: impl Into<String>, |
| 580 | context: Option<AuthenticationErrorContext>, |
| 581 | ) -> Self { |
| 582 | LlmError::AuthenticationError(AuthenticationErrorDetail::with_context(message, context)) |
| 583 | } |
| 584 | |
| 585 | /// Constructs an `LlmError` from HTTP response data plus request context |
| 586 | /// that is safe to display when authentication fails. |
| 587 | #[must_use] |
| 588 | #[cfg(test)] |
| 589 | pub fn from_http_response_with_request_context( |
| 590 | status: u16, |
| 591 | body: &str, |
| 592 | provider: Option<&str>, |
| 593 | base_url: Option<&str>, |
| 594 | model: Option<&str>, |
| 595 | key_source: Option<&str>, |
| 596 | api_key: Option<&str>, |
| 597 | ) -> Self { |
| 598 | let body = redact_api_key_from_message(body, api_key); |
| 599 | let context = |
| 600 | AuthenticationErrorContext::from_parts(provider, base_url, model, key_source, api_key); |
| 601 | Self::from_http_response_with_auth_context(status, &body, Some(context)) |
| 602 | } |
| 603 | |
| 604 | /// Constructs an `LlmError` from HTTP status code and response body, with |
| 605 | /// optional structured details for authentication failures. |
| 606 | /// |
| 607 | /// The `body` passed here must already be safe for user display. Prefer |
| 608 | /// [`Self::from_http_response_with_request_context`] when the raw API key is |
| 609 | /// available so the response body can be redacted before rendering. |
| 610 | #[must_use] |
| 611 | pub fn from_http_response_with_auth_context( |
| 612 | status: u16, |
| 613 | body: &str, |
| 614 | auth_context: Option<AuthenticationErrorContext>, |
| 615 | ) -> Self { |
| 616 | let classified = Self::from_http_response(status, body); |
| 617 | if matches!(classified, Self::QuotaExhausted(_)) { |
| 618 | return classified; |
| 619 | } |
| 620 | match status { |
| 621 | 401 => Self::authentication_error_with_context(body, auth_context), |
| 622 | 403 => { |
| 623 | if looks_like_authentication_failure(body) { |
| 624 | Self::authentication_error_with_context(body, auth_context) |
| 625 | } else { |
| 626 | LlmError::AuthorizationError(body.to_string()) |
| 627 | } |
| 628 | } |
| 629 | _ => classified, |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | /// Constructs an `LlmError` from HTTP status code, body, and optional Retry-After header. |
| 634 | pub fn from_http_response_with_retry_after( |
| 635 | status: u16, |
| 636 | body: &str, |
| 637 | retry_after: Option<Duration>, |
| 638 | ) -> Self { |
| 639 | let mut error = Self::from_http_response(status, body); |
| 640 | if let LlmError::RateLimited { |
| 641 | retry_after: ref mut ra, |
| 642 | .. |
| 643 | } = error |
| 644 | { |
| 645 | *ra = retry_after; |
| 646 | } |
| 647 | error |
| 648 | } |
| 649 | |
| 650 | /// Constructs an `LlmError` from a reqwest error. |
| 651 | pub fn from_reqwest(err: &reqwest::Error) -> Self { |
| 652 | if err.is_timeout() { |
| 653 | LlmError::Timeout(Duration::from_secs(0)) |
| 654 | } else if err.is_connect() { |
| 655 | LlmError::NetworkError(format!("Connection failed: {err}")) |
| 656 | } else if err.is_request() { |
| 657 | LlmError::NetworkError(format!("Request failed: {err}")) |
| 658 | } else { |
| 659 | LlmError::Other(err.to_string()) |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | /// Format provider HTTP error bodies before they are surfaced in the TUI. |
| 665 | /// |
| 666 | /// Providers sometimes return whole HTML error pages for gateway/WAF blocks. |
| 667 | /// Passing those pages through raw floods the transcript and can also make a |
| 668 | /// provider-side 403 look like a broken API key. Keep the useful details and |
| 669 | /// cap everything else. |
| 670 | #[must_use] |
| 671 | pub(crate) fn sanitize_http_error_body( |
| 672 | provider_label: Option<&str>, |
| 673 | status: u16, |
| 674 | body: &str, |
| 675 | ) -> String { |
| 676 | let json_message = extract_json_error_message(body); |
| 677 | let message = json_message.as_deref().unwrap_or(body); |
| 678 | // Gate on Google's actual rejection, not the selected provider or model: |
| 679 | // compatible gateways may manage signatures themselves (#6048). This |
| 680 | // shared boundary covers both streaming and non-streaming HTTP failures. |
| 681 | const SIGNATURE_HINT: &str = "Gemini rejected tool-call replay because a thought signature is missing. \ |
| 682 | Use the built-in `google` provider with its default endpoint, or a gateway that preserves \ |
| 683 | Google thought signatures, then start a new session before using tools. \ |
| 684 | Changing reasoning settings will not restore missing signatures."; |
| 685 | if status == 400 |
| 686 | && !is_probably_html(message) |
| 687 | && explicit_quota_code(body).is_none() |
| 688 | && !message.contains(SIGNATURE_HINT) |
| 689 | { |
| 690 | let lower = collapse_whitespace(message).to_ascii_lowercase(); |
| 691 | if lower.contains("missing a thought_signature") |
| 692 | || lower.contains("missing thought_signature") |
| 693 | || lower.contains("thought_signature is missing") |
| 694 | { |
| 695 | let detail = truncate_for_error(&collapse_whitespace(message), 900); |
| 696 | return format!("{SIGNATURE_HINT} Provider error: {detail}"); |
| 697 | } |
| 698 | } |
| 699 | |
| 700 | if let Some(message) = json_message { |
| 701 | let message = truncate_for_error(&collapse_whitespace(&message), 2_000); |
| 702 | if let Some(code) = explicit_quota_code(body) { |
| 703 | return format!("{message} (provider error code: {code})"); |
| 704 | } |
| 705 | return message; |
| 706 | } |
| 707 | |
| 708 | if is_probably_html(body) { |
| 709 | let text = html_to_text(body); |
| 710 | let lower = text.to_ascii_lowercase(); |
| 711 | let provider = provider_label.unwrap_or("Provider"); |
| 712 | |
| 713 | // Cloudflare's "Access Denied" interstitial strips the literal word |
| 714 | // "cloudflare" once tags are removed (it only survives in `<meta>` |
| 715 | // attributes and the `<style>`/`<script>` blocks we discard). Arcee's |
| 716 | // 403 page is exactly this shape, so also key off the WAF's stock copy |
| 717 | // ("security alert", "contact support") and a Cloudflare error/ray ID. |
| 718 | let error_id = extract_cloudflare_error_id(&text); |
| 719 | let is_cloudflare = lower.contains("cloudflare"); |
| 720 | let looks_like_access_denied = lower.contains("access denied") |
| 721 | && (is_cloudflare |
| 722 | || lower.contains("security alert") |
| 723 | || lower.contains("contact support") |
| 724 | || lower.contains("contact us") |
| 725 | || error_id.is_some()); |
| 726 | if looks_like_access_denied { |
| 727 | let label = if is_cloudflare { |
| 728 | "Cloudflare Access Denied" |
| 729 | } else { |
| 730 | "Access Denied" |
| 731 | }; |
| 732 | let mut message = format!( |
| 733 | "{provider} API returned {label} (HTTP {status}). \ |
| 734 | The request was blocked before it reached the model; retry with a \ |
| 735 | smaller request or fewer tools, or contact provider support" |
| 736 | ); |
| 737 | if let Some(id) = error_id { |
| 738 | message.push_str(&format!(" with ID {id}")); |
| 739 | } |
| 740 | message.push('.'); |
| 741 | return message; |
| 742 | } |
| 743 | |
| 744 | let text = truncate_for_error(&collapse_whitespace(&text), 900); |
| 745 | return format!("{provider} API returned an HTML error page (HTTP {status}): {text}"); |
| 746 | } |
| 747 | |
| 748 | truncate_for_error(&collapse_whitespace(body), 2_000) |
| 749 | } |
| 750 | |
| 751 | fn looks_like_authentication_failure(body: &str) -> bool { |
| 752 | let lower = body.to_ascii_lowercase(); |
| 753 | lower.contains("authentication") |
| 754 | || lower.contains("unauthorized") |
| 755 | || lower.contains("api key") |
| 756 | || lower.contains("invalid key") |
| 757 | || lower.contains("invalid token") |
| 758 | || lower.contains("bearer token") |
| 759 | || lower.contains("missing token") |
| 760 | } |
| 761 | |
| 762 | /// A provider error is a context overflow only when it says so. Bare |
| 763 | /// "token", "too long" or "maximum" also appear in ordinary invalid-request |
| 764 | /// errors (`max_tokens must be ...`, a field value too long), which compaction |
| 765 | /// or a bigger window cannot fix. This is the one phrase list: the typed 400 |
| 766 | /// classification here and the engine's string classifier both read it. |
| 767 | /// `lower` must already be lowercase. |
| 768 | pub(crate) fn is_context_length_message(lower: &str) -> bool { |
| 769 | [ |
| 770 | "context_length", |
| 771 | "context length", |
| 772 | "context window", |
| 773 | "context limit", |
| 774 | "maximum context", |
| 775 | "prompt is too long", |
| 776 | "input is too long", |
| 777 | "maximum prompt length", |
| 778 | "exceeded model token limit", |
| 779 | "tokens exceed", |
| 780 | "exceeds the maximum number of tokens", |
| 781 | // llama.cpp: "the request exceeds the available context size". |
| 782 | "available context size", |
| 783 | ] |
| 784 | .iter() |
| 785 | .any(|phrase| lower.contains(phrase)) |
| 786 | } |
| 787 | |
| 788 | /// Quota exhaustion is a durable account state, not a generic rate-limit |
| 789 | /// synonym. Accept only explicit provider evidence at the HTTP/parser boundary; |
| 790 | /// callers holding a stringified error must never promote it to this type. |
| 791 | fn has_explicit_quota_evidence(body: &str) -> bool { |
| 792 | explicit_quota_code(body).is_some() |
| 793 | || explicit_quota_code_marker(body).is_some() |
| 794 | || has_explicit_quota_phrase(body) |
| 795 | } |
| 796 | |
| 797 | fn explicit_quota_code(body: &str) -> Option<String> { |
| 798 | let value: Value = serde_json::from_str(body).ok()?; |
| 799 | [ |
| 800 | "/error/code", |
| 801 | "/error/type", |
| 802 | "/error/error_code", |
| 803 | "/code", |
| 804 | "/type", |
| 805 | "/error_code", |
| 806 | ] |
| 807 | .into_iter() |
| 808 | .filter_map(|pointer| value.pointer(pointer).and_then(Value::as_str)) |
| 809 | .find(|code| is_explicit_quota_code(code)) |
| 810 | .map(ToOwned::to_owned) |
| 811 | } |
| 812 | |
| 813 | fn is_explicit_quota_code(code: &str) -> bool { |
| 814 | let normalized: String = code |
| 815 | .chars() |
| 816 | .filter(|ch| ch.is_ascii_alphanumeric()) |
| 817 | .map(|ch| ch.to_ascii_lowercase()) |
| 818 | .collect(); |
| 819 | matches!( |
| 820 | normalized.as_str(), |
| 821 | "insufficientquota" |
| 822 | | "quotaexceeded" |
| 823 | | "quotaexhausted" |
| 824 | | "billinghardlimitreached" |
| 825 | | "billinglimitreached" |
| 826 | | "creditbalanceexhausted" |
| 827 | // ChatGPT/Codex subscription window (HTTP 429, `error.type`), |
| 828 | // as openai/codex `api_bridge.rs` maps it. It resets on the |
| 829 | // plan's schedule, not after a short backoff. |
| 830 | | "usagelimitreached" |
| 831 | // Same backend, same branch: the signed-in plan does not include |
| 832 | // Codex. Retrying cannot help; switching accounts can. |
| 833 | | "usagenotincluded" |
| 834 | | "subscriptionsharingusagelimitexceeded" |
| 835 | | "subscriptionsharingusageunavailable" |
| 836 | ) |
| 837 | } |
| 838 | |
| 839 | fn explicit_quota_code_marker(body: &str) -> Option<String> { |
| 840 | let lower = body.to_ascii_lowercase(); |
| 841 | let (_, suffix) = lower.split_once("provider error code:")?; |
| 842 | let code = suffix |
| 843 | .trim_start() |
| 844 | .split(|ch: char| !(ch.is_ascii_alphanumeric() || matches!(ch, '_' | '-'))) |
| 845 | .next() |
| 846 | .unwrap_or_default(); |
| 847 | is_explicit_quota_code(code).then(|| code.to_string()) |
| 848 | } |
| 849 | |
| 850 | fn has_explicit_quota_phrase(body: &str) -> bool { |
| 851 | let lower = body.to_ascii_lowercase(); |
| 852 | let current_quota_exhausted = lower.contains("exceeded your current quota") |
| 853 | || lower.contains("current quota has been exceeded"); |
| 854 | let plan_and_billing_guidance = lower.contains("plan") && lower.contains("billing"); |
| 855 | let durable_scope_exhausted = [ |
| 856 | "billing quota exceeded", |
| 857 | "billing quota exhausted", |
| 858 | "billing quota is exhausted", |
| 859 | "billing quota has been exceeded", |
| 860 | "billing quota has been exhausted", |
| 861 | "account quota exceeded", |
| 862 | "account quota exhausted", |
| 863 | "account quota is exhausted", |
| 864 | "account quota has been exceeded", |
| 865 | "account quota has been exhausted", |
| 866 | "plan quota exceeded", |
| 867 | "plan quota exhausted", |
| 868 | "plan quota is exhausted", |
| 869 | "plan quota has been exceeded", |
| 870 | "plan quota has been exhausted", |
| 871 | ] |
| 872 | .into_iter() |
| 873 | .any(|phrase| lower.contains(phrase)); |
| 874 | |
| 875 | lower.contains("billing hard limit has been reached") |
| 876 | || lower.contains("credit balance exhausted") |
| 877 | || lower.contains("credit balance is exhausted") |
| 878 | || durable_scope_exhausted |
| 879 | || (current_quota_exhausted && plan_and_billing_guidance) |
| 880 | } |
| 881 | |
| 882 | fn extract_json_error_message(body: &str) -> Option<String> { |
| 883 | let value: Value = serde_json::from_str(body).ok()?; |
| 884 | // Flat gateway bodies (`{"error":"Bad Request","message":"Invalid model |
| 885 | // name: 'x'"}` — Concentrate, among others) keep the class in `error` and |
| 886 | // the detail in `message`. Surfacing only the class hid the one line the |
| 887 | // person needed, so carry both when both are present and differ. |
| 888 | if let (Some(class), Some(detail)) = ( |
| 889 | value.pointer("/error").and_then(Value::as_str), |
| 890 | value.pointer("/message").and_then(Value::as_str), |
| 891 | ) && !class.trim().is_empty() |
| 892 | && !detail.trim().is_empty() |
| 893 | && !class.trim().eq_ignore_ascii_case(detail.trim()) |
| 894 | { |
| 895 | return Some(format!("{}: {}", class.trim(), detail.trim())); |
| 896 | } |
| 897 | for pointer in [ |
| 898 | "/error/message", |
| 899 | "/error", |
| 900 | "/message", |
| 901 | "/detail", |
| 902 | "/error_description", |
| 903 | ] { |
| 904 | let Some(value) = value.pointer(pointer) else { |
| 905 | continue; |
| 906 | }; |
| 907 | if let Some(message) = value.as_str() { |
| 908 | if !message.trim().is_empty() { |
| 909 | return Some(message.to_string()); |
| 910 | } |
| 911 | } else if value.is_object() || value.is_array() { |
| 912 | return Some(value.to_string()); |
| 913 | } |
| 914 | } |
| 915 | None |
| 916 | } |
| 917 | |
| 918 | fn is_probably_html(body: &str) -> bool { |
| 919 | let prefix = body |
| 920 | .chars() |
| 921 | .take(512) |
| 922 | .collect::<String>() |
| 923 | .to_ascii_lowercase(); |
| 924 | prefix.contains("<!doctype html") || prefix.contains("<html") || prefix.contains("<head") |
| 925 | } |
| 926 | |
| 927 | fn html_to_text(html: &str) -> String { |
| 928 | let without_scripts = strip_html_block(html, "script"); |
| 929 | let without_styles = strip_html_block(&without_scripts, "style"); |
| 930 | let mut text = String::with_capacity(without_styles.len().min(4096)); |
| 931 | let mut in_tag = false; |
| 932 | for ch in without_styles.chars() { |
| 933 | match ch { |
| 934 | '<' => { |
| 935 | in_tag = true; |
| 936 | text.push(' '); |
| 937 | } |
| 938 | '>' => { |
| 939 | in_tag = false; |
| 940 | text.push(' '); |
| 941 | } |
| 942 | _ if !in_tag => text.push(ch), |
| 943 | _ => {} |
| 944 | } |
| 945 | } |
| 946 | decode_basic_html_entities(&collapse_whitespace(&text)) |
| 947 | } |
| 948 | |
| 949 | fn strip_html_block(input: &str, tag: &str) -> String { |
| 950 | let mut out = String::with_capacity(input.len()); |
| 951 | let mut cursor = 0usize; |
| 952 | let lower = input.to_ascii_lowercase(); |
| 953 | let start_marker = format!("<{tag}"); |
| 954 | let end_marker = format!("</{tag}>"); |
| 955 | |
| 956 | while let Some(relative_start) = lower[cursor..].find(&start_marker) { |
| 957 | let start = cursor + relative_start; |
| 958 | out.push_str(&input[cursor..start]); |
| 959 | let after_start = start + start_marker.len(); |
| 960 | let Some(relative_end) = lower[after_start..].find(&end_marker) else { |
| 961 | cursor = input.len(); |
| 962 | break; |
| 963 | }; |
| 964 | cursor = after_start + relative_end + end_marker.len(); |
| 965 | out.push(' '); |
| 966 | } |
| 967 | out.push_str(&input[cursor..]); |
| 968 | out |
| 969 | } |
| 970 | |
| 971 | fn decode_basic_html_entities(input: &str) -> String { |
| 972 | input |
| 973 | .replace(" ", " ") |
| 974 | .replace("&", "&") |
| 975 | .replace("<", "<") |
| 976 | .replace(">", ">") |
| 977 | .replace(""", "\"") |
| 978 | .replace("'", "'") |
| 979 | .replace("'", "'") |
| 980 | } |
| 981 | |
| 982 | fn collapse_whitespace(input: &str) -> String { |
| 983 | input.split_whitespace().collect::<Vec<_>>().join(" ") |
| 984 | } |
| 985 | |
| 986 | fn truncate_for_error(input: &str, max_chars: usize) -> String { |
| 987 | let mut out = String::with_capacity(input.len().min(max_chars + 32)); |
| 988 | for (count, ch) in input.chars().enumerate() { |
| 989 | if count >= max_chars { |
| 990 | out.push_str("..."); |
| 991 | return out; |
| 992 | } |
| 993 | out.push(ch); |
| 994 | } |
| 995 | out |
| 996 | } |
| 997 | |
| 998 | fn extract_cloudflare_error_id(text: &str) -> Option<String> { |
| 999 | let mut last = None; |
| 1000 | for token in text.split(|ch: char| !ch.is_ascii_hexdigit()) { |
| 1001 | if (16..=64).contains(&token.len()) && token.bytes().any(|b| b.is_ascii_alphabetic()) { |
| 1002 | last = Some(token.to_string()); |
| 1003 | } |
| 1004 | } |
| 1005 | last |
| 1006 | } |
| 1007 | |
| 1008 | impl From<reqwest::Error> for LlmError { |
| 1009 | fn from(err: reqwest::Error) -> Self { |
| 1010 | LlmError::from_reqwest(&err) |
| 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | impl From<serde_json::Error> for LlmError { |
| 1015 | fn from(err: serde_json::Error) -> Self { |
| 1016 | LlmError::ParseError(err.to_string()) |
| 1017 | } |
| 1018 | } |
| 1019 | |
| 1020 | // === RetryConfig - Exponential Backoff Configuration === |
| 1021 | |
| 1022 | /// Configuration for retry behavior with exponential backoff. |
| 1023 | /// |
| 1024 | /// This struct controls how retries are performed: |
| 1025 | /// - Number of retry attempts |
| 1026 | /// - Delay calculation (exponential backoff with optional jitter) |
| 1027 | /// - Which HTTP status codes are retryable |
| 1028 | /// - Timeout handling |
| 1029 | /// |
| 1030 | /// # Default Values |
| 1031 | /// |
| 1032 | /// - `enabled`: true |
| 1033 | /// - `max_retries`: 3 |
| 1034 | /// - `initial_delay`: 1.0 seconds |
| 1035 | /// - `max_delay`: 60.0 seconds |
| 1036 | /// - `exponential_base`: 2.0 |
| 1037 | /// - `jitter`: true (adds randomness to prevent thundering herd) |
| 1038 | /// - `jitter_factor`: 0.1 (10% variation) |
| 1039 | /// - `retryable_status_codes`: [429, 499, 500, 502, 503, 504] |
| 1040 | #[derive(Debug, Clone)] |
| 1041 | pub struct RetryConfig { |
| 1042 | /// Whether retry logic is enabled |
| 1043 | pub enabled: bool, |
| 1044 | |
| 1045 | /// Maximum number of retry attempts (0 = no retries, 3 = up to 4 total attempts) |
| 1046 | pub max_retries: u32, |
| 1047 | |
| 1048 | /// Initial delay before first retry (seconds) |
| 1049 | pub initial_delay: f64, |
| 1050 | |
| 1051 | /// Maximum delay between retries (seconds) |
| 1052 | pub max_delay: f64, |
| 1053 | |
| 1054 | /// Base for exponential backoff (delay = initial * base^attempt) |
| 1055 | pub exponential_base: f64, |
| 1056 | |
| 1057 | /// Whether to add random jitter to delays |
| 1058 | pub jitter: bool, |
| 1059 | |
| 1060 | /// Jitter factor (0.1 = +/- 10% variation) |
| 1061 | pub jitter_factor: f64, |
| 1062 | |
| 1063 | /// Whether to respect server's Retry-After header |
| 1064 | pub respect_retry_after: bool, |
| 1065 | |
| 1066 | /// HTTP status codes that should trigger a retry |
| 1067 | #[allow(dead_code)] // Used in tests via is_retryable_status() |
| 1068 | pub retryable_status_codes: Vec<u16>, |
| 1069 | |
| 1070 | /// Timeout for individual requests (seconds, 0 = no timeout) |
| 1071 | #[allow(dead_code)] // Configuration field for retry consumers |
| 1072 | pub request_timeout: f64, |
| 1073 | |
| 1074 | /// Total timeout for all retry attempts (seconds, 0 = no total timeout) |
| 1075 | pub total_timeout: f64, |
| 1076 | } |
| 1077 | |
| 1078 | impl Default for RetryConfig { |
| 1079 | fn default() -> Self { |
| 1080 | Self { |
| 1081 | enabled: true, |
| 1082 | max_retries: 3, |
| 1083 | initial_delay: 1.0, |
| 1084 | max_delay: 60.0, |
| 1085 | exponential_base: 2.0, |
| 1086 | jitter: true, |
| 1087 | jitter_factor: 0.1, |
| 1088 | respect_retry_after: true, |
| 1089 | retryable_status_codes: vec![429, 499, 500, 502, 503, 504], |
| 1090 | request_timeout: 120.0, |
| 1091 | total_timeout: 0.0, // No total timeout by default |
| 1092 | } |
| 1093 | } |
| 1094 | } |
| 1095 | |
| 1096 | #[allow(dead_code)] // Public builder API, used in tests |
| 1097 | impl RetryConfig { |
| 1098 | /// Creates a new `RetryConfig` with default values |
| 1099 | pub fn new() -> Self { |
| 1100 | Self::default() |
| 1101 | } |
| 1102 | |
| 1103 | /// Creates a config with retry disabled |
| 1104 | pub fn disabled() -> Self { |
| 1105 | Self { |
| 1106 | enabled: false, |
| 1107 | ..Default::default() |
| 1108 | } |
| 1109 | } |
| 1110 | |
| 1111 | /// Builder method to set max retries |
| 1112 | pub fn with_max_retries(mut self, max_retries: u32) -> Self { |
| 1113 | self.max_retries = max_retries; |
| 1114 | self |
| 1115 | } |
| 1116 | |
| 1117 | /// Builder method to set initial delay |
| 1118 | pub fn with_initial_delay(mut self, delay: f64) -> Self { |
| 1119 | self.initial_delay = delay; |
| 1120 | self |
| 1121 | } |
| 1122 | |
| 1123 | /// Builder method to set max delay |
| 1124 | pub fn with_max_delay(mut self, delay: f64) -> Self { |
| 1125 | self.max_delay = delay; |
| 1126 | self |
| 1127 | } |
| 1128 | |
| 1129 | /// Builder method to enable/disable jitter |
| 1130 | pub fn with_jitter(mut self, enabled: bool) -> Self { |
| 1131 | self.jitter = enabled; |
| 1132 | self |
| 1133 | } |
| 1134 | |
| 1135 | /// Calculates the delay for a given retry attempt. |
| 1136 | /// |
| 1137 | /// Uses exponential backoff: delay = `initial_delay` * `exponential_base^attempt` |
| 1138 | /// The result is capped at `max_delay` and optionally has jitter applied. |
| 1139 | /// |
| 1140 | /// # Arguments |
| 1141 | /// |
| 1142 | /// * `attempt` - Zero-based attempt number (0 = first retry) |
| 1143 | /// |
| 1144 | /// # Returns |
| 1145 | /// |
| 1146 | /// Duration to wait before the next retry attempt |
| 1147 | pub fn delay_for_attempt(&self, attempt: u32) -> Duration { |
| 1148 | let exponent = i32::try_from(attempt).unwrap_or(i32::MAX); |
| 1149 | let base_delay = self.initial_delay * self.exponential_base.powi(exponent); |
| 1150 | let capped_delay = base_delay.min(self.max_delay); |
| 1151 | |
| 1152 | let final_delay = if self.jitter { |
| 1153 | // Add random jitter to prevent thundering herd problem |
| 1154 | let jitter_range = capped_delay * self.jitter_factor; |
| 1155 | // Use UUID v4 entropy for jitter randomness. |
| 1156 | let bytes = *Uuid::new_v4().as_bytes(); |
| 1157 | let sample = u16::from_le_bytes([bytes[0], bytes[1]]); |
| 1158 | let random_factor = f64::from(sample) / f64::from(u16::MAX); // 0.0 to 1.0 |
| 1159 | let jitter = jitter_range * (2.0 * random_factor - 1.0); // -range to +range |
| 1160 | |
| 1161 | (capped_delay + jitter).max(0.0) |
| 1162 | } else { |
| 1163 | capped_delay |
| 1164 | }; |
| 1165 | |
| 1166 | Duration::from_secs_f64(final_delay) |
| 1167 | } |
| 1168 | |
| 1169 | /// Checks if a given HTTP status code should trigger a retry |
| 1170 | pub fn is_retryable_status(&self, status: u16) -> bool { |
| 1171 | self.retryable_status_codes.contains(&status) |
| 1172 | } |
| 1173 | } |
| 1174 | |
| 1175 | /// Converts from the existing `RetryPolicy` in config |
| 1176 | impl From<RetryPolicy> for RetryConfig { |
| 1177 | fn from(policy: RetryPolicy) -> Self { |
| 1178 | Self { |
| 1179 | enabled: policy.enabled, |
| 1180 | max_retries: policy.max_retries, |
| 1181 | initial_delay: policy.initial_delay, |
| 1182 | max_delay: policy.max_delay, |
| 1183 | exponential_base: policy.exponential_base, |
| 1184 | jitter: policy.jitter, |
| 1185 | jitter_factor: policy.jitter_factor, |
| 1186 | respect_retry_after: policy.respect_retry_after, |
| 1187 | ..Default::default() |
| 1188 | } |
| 1189 | } |
| 1190 | } |
| 1191 | |
| 1192 | /// Converts back to `RetryPolicy` for compatibility |
| 1193 | impl From<RetryConfig> for RetryPolicy { |
| 1194 | fn from(config: RetryConfig) -> Self { |
| 1195 | Self { |
| 1196 | enabled: config.enabled, |
| 1197 | max_retries: config.max_retries, |
| 1198 | initial_delay: config.initial_delay, |
| 1199 | max_delay: config.max_delay, |
| 1200 | exponential_base: config.exponential_base, |
| 1201 | jitter: config.jitter, |
| 1202 | jitter_factor: config.jitter_factor, |
| 1203 | respect_retry_after: config.respect_retry_after, |
| 1204 | } |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | // === Retry Error and Result Types === |
| 1209 | |
| 1210 | /// Error returned when all retry attempts have been exhausted. |
| 1211 | #[derive(Debug)] |
| 1212 | pub struct RetryError { |
| 1213 | /// The last error encountered |
| 1214 | pub last_error: LlmError, |
| 1215 | |
| 1216 | /// Total number of attempts made |
| 1217 | pub attempts: u32, |
| 1218 | |
| 1219 | /// Total time spent across all attempts |
| 1220 | pub total_time: Duration, |
| 1221 | } |
| 1222 | |
| 1223 | impl std::fmt::Display for RetryError { |
| 1224 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 1225 | write!( |
| 1226 | f, |
| 1227 | "Retry exhausted after {} attempts ({:?}): {}", |
| 1228 | self.attempts, self.total_time, self.last_error |
| 1229 | ) |
| 1230 | } |
| 1231 | } |
| 1232 | |
| 1233 | impl std::error::Error for RetryError { |
| 1234 | fn source(&self) -> Option<&(dyn std::error::Error + 'static)> { |
| 1235 | Some(&self.last_error) |
| 1236 | } |
| 1237 | } |
| 1238 | |
| 1239 | /// Result type for retry operations |
| 1240 | pub type RetryResult<T> = Result<T, RetryError>; |
| 1241 | |
| 1242 | /// Callback type for retry notifications |
| 1243 | /// |
| 1244 | /// Called before each retry with: |
| 1245 | /// - The error that triggered the retry |
| 1246 | /// - The attempt number (0-based) |
| 1247 | /// - The delay before the next attempt |
| 1248 | pub type RetryCallback = Box<dyn Fn(&LlmError, u32, Duration) + Send + Sync>; |
| 1249 | |
| 1250 | /// An observation of the existing request, never another retry driver. Its |
| 1251 | /// lexical scope captures the producing Engine queue, not an ambient session. |
| 1252 | pub(crate) type RetryStatusEmitter = |
| 1253 | std::sync::Arc<dyn Fn(String) -> Pin<Box<dyn Future<Output = ()> + Send>> + Send + Sync>; |
| 1254 | |
| 1255 | #[derive(Clone)] |
| 1256 | pub(crate) struct RequestRetryObservation { |
| 1257 | pub(crate) retries: std::sync::Arc<std::sync::atomic::AtomicU32>, |
| 1258 | pub(crate) emit: RetryStatusEmitter, |
| 1259 | } |
| 1260 | |
| 1261 | tokio::task_local! { |
| 1262 | static REQUEST_RETRY_OBSERVATION: Option<RequestRetryObservation>; |
| 1263 | } |
| 1264 | |
| 1265 | pub(crate) async fn observe_request_retries<F: Future>( |
| 1266 | observation: Option<RequestRetryObservation>, |
| 1267 | future: F, |
| 1268 | ) -> F::Output { |
| 1269 | REQUEST_RETRY_OBSERVATION.scope(observation, future).await |
| 1270 | } |
| 1271 | |
| 1272 | async fn observe_transport_status(message: String) { |
| 1273 | let observation = REQUEST_RETRY_OBSERVATION |
| 1274 | .try_with(Clone::clone) |
| 1275 | .ok() |
| 1276 | .flatten(); |
| 1277 | if let Some(observation) = observation { |
| 1278 | (observation.emit)(message).await; |
| 1279 | } |
| 1280 | } |
| 1281 | |
| 1282 | fn observe_transport_attempt() { |
| 1283 | let _ = REQUEST_RETRY_OBSERVATION.try_with(|observation| { |
| 1284 | if let Some(observation) = observation { |
| 1285 | let mut count = observation |
| 1286 | .retries |
| 1287 | .load(std::sync::atomic::Ordering::Relaxed); |
| 1288 | loop { |
| 1289 | match observation.retries.compare_exchange_weak( |
| 1290 | count, |
| 1291 | count.saturating_add(1), |
| 1292 | std::sync::atomic::Ordering::Relaxed, |
| 1293 | std::sync::atomic::Ordering::Relaxed, |
| 1294 | ) { |
| 1295 | Ok(_) => break, |
| 1296 | Err(current) => count = current, |
| 1297 | } |
| 1298 | } |
| 1299 | } |
| 1300 | }); |
| 1301 | } |
| 1302 | |
| 1303 | /// A public receipt must never include the provider-controlled error payload. |
| 1304 | /// Raw errors remain intact in the original result and diagnostic/log paths. |
| 1305 | pub(crate) fn retry_reason_summary(error: &LlmError) -> String { |
| 1306 | match error { |
| 1307 | LlmError::RateLimited { .. } => "rate limited".into(), |
| 1308 | LlmError::ServerError { status, .. } => format!("upstream {status}"), |
| 1309 | LlmError::NetworkError(_) => "network error".into(), |
| 1310 | LlmError::Timeout(_) => "timeout".into(), |
| 1311 | _ => "non-retryable provider failure".into(), |
| 1312 | } |
| 1313 | } |
| 1314 | |
| 1315 | async fn observe_transport_stopped(retries: u32, error: &LlmError, exhausted: bool) { |
| 1316 | if retries > 0 { |
| 1317 | let disposition = if exhausted { |
| 1318 | "Retry exhaustion" |
| 1319 | } else { |
| 1320 | "Retry stopped" |
| 1321 | }; |
| 1322 | observe_transport_status(format!( |
| 1323 | "{disposition}: transport request stopped after {retries} retries; {}", |
| 1324 | retry_reason_summary(error), |
| 1325 | )) |
| 1326 | .await; |
| 1327 | } |
| 1328 | } |
| 1329 | |
| 1330 | // === with_retry - Generic Retry Wrapper === |
| 1331 | |
| 1332 | /// Executes an async operation with configurable retry logic. |
| 1333 | /// |
| 1334 | /// This function wraps any async operation that returns `Result<T, LlmError>` |
| 1335 | /// and automatically retries on transient failures using exponential backoff. |
| 1336 | /// |
| 1337 | /// # Arguments |
| 1338 | /// |
| 1339 | /// * `config` - Retry configuration (delays, max attempts, etc.) |
| 1340 | /// * `operation` - Async closure to execute (will be called multiple times on retry) |
| 1341 | /// * `callback` - Optional callback for retry notifications (logging, metrics, etc.) |
| 1342 | /// |
| 1343 | /// # Returns |
| 1344 | /// |
| 1345 | /// * `Ok(T)` - The successful result from the operation |
| 1346 | /// * `Err(RetryError)` - All retries exhausted or non-retryable error encountered |
| 1347 | /// |
| 1348 | /// # Known limitation: ambiguous failures are replayed |
| 1349 | /// |
| 1350 | /// A timeout or connection loss after the request was written is retried |
| 1351 | /// like a connect failure, so a provider that already accepted the first |
| 1352 | /// attempt may bill a second completion. Every caller sends model inference |
| 1353 | /// (messages, FIM, translation, speech, provider web search): a replay costs |
| 1354 | /// compute but has no external side effect, and the provider APIs used here |
| 1355 | /// expose no idempotency key for these requests that could dedupe it. An |
| 1356 | /// operation with an external side effect must not be retried through this |
| 1357 | /// helper without an idempotency key the server honors. |
| 1358 | /// |
| 1359 | /// # Example |
| 1360 | /// |
| 1361 | /// ```ignore |
| 1362 | /// let result = with_retry( |
| 1363 | /// &config, |
| 1364 | /// || async { client.send_request(&req).await }, |
| 1365 | /// Some(Box::new(|err, attempt, delay| { |
| 1366 | /// eprintln!("Retry {} after {:?}: {}", attempt, delay, err); |
| 1367 | /// })), |
| 1368 | /// ).await; |
| 1369 | /// ``` |
| 1370 | // Keep the structured error inline: this is a public compatibility surface and |
| 1371 | // boxing it in a patch release would force every caller to change ownership |
| 1372 | // handling for `last_error`. |
| 1373 | #[allow(clippy::result_large_err)] |
| 1374 | pub async fn with_retry<F, Fut, T>( |
| 1375 | config: &RetryConfig, |
| 1376 | mut operation: F, |
| 1377 | callback: Option<RetryCallback>, |
| 1378 | ) -> RetryResult<T> |
| 1379 | where |
| 1380 | F: FnMut() -> Fut, |
| 1381 | Fut: Future<Output = Result<T, LlmError>>, |
| 1382 | { |
| 1383 | // If retries are disabled, just run once |
| 1384 | if !config.enabled { |
| 1385 | return operation().await.map_err(|e| RetryError { |
| 1386 | last_error: e, |
| 1387 | attempts: 1, |
| 1388 | total_time: Duration::ZERO, |
| 1389 | }); |
| 1390 | } |
| 1391 | |
| 1392 | let start_time = Instant::now(); |
| 1393 | let total_timeout = if config.total_timeout > 0.0 { |
| 1394 | Some(Duration::from_secs_f64(config.total_timeout)) |
| 1395 | } else { |
| 1396 | None |
| 1397 | }; |
| 1398 | |
| 1399 | let mut last_error: Option<LlmError> = None; |
| 1400 | |
| 1401 | // Attempt 0 is the first try, then up to max_retries additional attempts |
| 1402 | for attempt in 0..=config.max_retries { |
| 1403 | // Check total timeout |
| 1404 | if let Some(timeout) = total_timeout |
| 1405 | && start_time.elapsed() >= timeout |
| 1406 | { |
| 1407 | let error = last_error.unwrap_or(LlmError::Timeout(timeout)); |
| 1408 | observe_transport_stopped(attempt.saturating_sub(1), &error, true).await; |
| 1409 | return Err(RetryError { |
| 1410 | last_error: error, |
| 1411 | attempts: attempt, |
| 1412 | total_time: start_time.elapsed(), |
| 1413 | }); |
| 1414 | } |
| 1415 | |
| 1416 | if attempt > 0 { |
| 1417 | observe_transport_attempt(); |
| 1418 | } |
| 1419 | match operation().await { |
| 1420 | Ok(result) => { |
| 1421 | if attempt > 0 { |
| 1422 | observe_transport_status(format!( |
| 1423 | "Retry recovery: transport request recovered after {attempt} retries" |
| 1424 | )) |
| 1425 | .await; |
| 1426 | } |
| 1427 | return Ok(result); |
| 1428 | } |
| 1429 | Err(err) => { |
| 1430 | // Non-retryable errors fail immediately |
| 1431 | if !err.is_retryable() { |
| 1432 | observe_transport_stopped(attempt, &err, false).await; |
| 1433 | return Err(RetryError { |
| 1434 | last_error: err, |
| 1435 | attempts: attempt + 1, |
| 1436 | total_time: start_time.elapsed(), |
| 1437 | }); |
| 1438 | } |
| 1439 | |
| 1440 | // Last attempt - no more retries |
| 1441 | if attempt >= config.max_retries { |
| 1442 | observe_transport_stopped(attempt, &err, true).await; |
| 1443 | return Err(RetryError { |
| 1444 | last_error: err, |
| 1445 | attempts: attempt + 1, |
| 1446 | total_time: start_time.elapsed(), |
| 1447 | }); |
| 1448 | } |
| 1449 | |
| 1450 | // Calculate delay |
| 1451 | // Use server's Retry-After if available and configured |
| 1452 | let base_delay = config.delay_for_attempt(attempt); |
| 1453 | let delay = if config.respect_retry_after { |
| 1454 | err.suggested_retry_delay().unwrap_or(base_delay) |
| 1455 | } else { |
| 1456 | base_delay |
| 1457 | }; |
| 1458 | |
| 1459 | // Notify callback if provided |
| 1460 | if let Some(ref cb) = callback { |
| 1461 | cb(&err, attempt, delay); |
| 1462 | } |
| 1463 | |
| 1464 | observe_transport_status(format!( |
| 1465 | "Retry attempt: transport {}/{}; {}; waiting {:.2}s", |
| 1466 | attempt + 1, |
| 1467 | config.max_retries, |
| 1468 | retry_reason_summary(&err), |
| 1469 | delay.as_secs_f64(), |
| 1470 | )) |
| 1471 | .await; |
| 1472 | last_error = Some(err); |
| 1473 | |
| 1474 | // Wait before retrying |
| 1475 | tokio::time::sleep(delay).await; |
| 1476 | } |
| 1477 | } |
| 1478 | } |
| 1479 | |
| 1480 | // Should not reach here, but handle gracefully |
| 1481 | Err(RetryError { |
| 1482 | last_error: last_error.unwrap_or(LlmError::Other("Unknown retry error".to_string())), |
| 1483 | attempts: config.max_retries + 1, |
| 1484 | total_time: start_time.elapsed(), |
| 1485 | }) |
| 1486 | } |
| 1487 | |
| 1488 | // === Utility Functions === |
| 1489 | |
| 1490 | /// The longest a `Retry-After` value is ever believed. A server (or a proxy |
| 1491 | /// in front of it) can send an arbitrarily large delay; without a ceiling a |
| 1492 | /// single `Retry-After: 86400` would wedge the turn for a day. One hour is |
| 1493 | /// well past any legitimate rate-limit window. |
| 1494 | const RETRY_AFTER_MAX: Duration = Duration::from_secs(3600); |
| 1495 | |
| 1496 | /// Parses the Retry-After header value into a Duration. |
| 1497 | /// |
| 1498 | /// Supports both: |
| 1499 | /// - Seconds as integer: "120" -> 120 seconds |
| 1500 | /// - HTTP-date format: "Wed, 21 Oct 2015 07:28:00 GMT" (not implemented, returns None) |
| 1501 | /// |
| 1502 | /// The value is server-controlled, so this never panics and never returns an |
| 1503 | /// unbounded delay: negative / NaN / infinite / absurd floats are rejected |
| 1504 | /// (`Duration::from_secs_f64` panics on a negative — a remote-triggerable |
| 1505 | /// crash before this guard), and any result is clamped to [`RETRY_AFTER_MAX`]. |
| 1506 | pub fn parse_retry_after(value: &str) -> Option<Duration> { |
| 1507 | // Try parsing as seconds |
| 1508 | if let Ok(seconds) = value.parse::<u64>() { |
| 1509 | return Some(Duration::from_secs(seconds).min(RETRY_AFTER_MAX)); |
| 1510 | } |
| 1511 | |
| 1512 | // Try parsing as float seconds. Only a finite, non-negative value is a |
| 1513 | // meaningful delay; everything else (`-5`, `nan`, `inf`) is "no usable |
| 1514 | // hint". Clamp to the ceiling BEFORE `from_secs_f64` so an out-of-range |
| 1515 | // float can never reach its overflow-panic path, while keeping the |
| 1516 | // sub-second precision a legitimate `1.5` carries. |
| 1517 | if let Ok(seconds) = value.parse::<f64>() |
| 1518 | && seconds.is_finite() |
| 1519 | && seconds >= 0.0 |
| 1520 | { |
| 1521 | let clamped = seconds.min(RETRY_AFTER_MAX.as_secs_f64()); |
| 1522 | return Some(Duration::from_secs_f64(clamped)); |
| 1523 | } |
| 1524 | |
| 1525 | // HTTP-date format not supported yet |
| 1526 | // Could use chrono or httpdate crate if needed |
| 1527 | None |
| 1528 | } |
| 1529 | |
| 1530 | /// Extracts Retry-After duration from response headers |
| 1531 | pub fn extract_retry_after(headers: &reqwest::header::HeaderMap) -> Option<Duration> { |
| 1532 | headers |
| 1533 | .get(reqwest::header::RETRY_AFTER) |
| 1534 | .and_then(|v| v.to_str().ok()) |
| 1535 | .and_then(parse_retry_after) |
| 1536 | } |
| 1537 | |
| 1538 | #[cfg(test)] |
| 1539 | #[path = "tests.rs"] |
| 1540 | mod quota_tests; |
| 1541 | |
| 1542 | // === Tests === |
| 1543 | |
| 1544 | #[cfg(test)] |
| 1545 | mod tests { |
| 1546 | use super::*; |
| 1547 | |
| 1548 | #[test] |
| 1549 | fn official_chatgpt_usage_codes_are_terminal_across_http_boundaries() { |
| 1550 | for code in [ |
| 1551 | "subscription_sharing_usage_limit_exceeded", |
| 1552 | "subscription_sharing_usage_unavailable", |
| 1553 | ] { |
| 1554 | let body = serde_json::json!({"error":{"code":code,"message":"opaque"}}).to_string(); |
| 1555 | for status in [400, 403, 429] { |
| 1556 | for safe_body in [body.clone(), sanitize_http_error_body(None, status, &body)] { |
| 1557 | let error = |
| 1558 | LlmError::from_http_response_with_auth_context(status, &safe_body, None); |
| 1559 | assert!(matches!(error, LlmError::QuotaExhausted(_)), "{error:?}"); |
| 1560 | assert!(!error.is_retryable()); |
| 1561 | assert!(error.to_string().contains("ChatGPT Settings > Usage")); |
| 1562 | let envelope = crate::error_taxonomy::envelope_for_llm_error( |
| 1563 | error.into(), |
| 1564 | "allowance unavailable".into(), |
| 1565 | ); |
| 1566 | assert!(!envelope.recoverable); |
| 1567 | assert_eq!(envelope.code, "llm_quota_exhausted"); |
| 1568 | } |
| 1569 | } |
| 1570 | } |
| 1571 | for body in [ |
| 1572 | "usage limit reached", |
| 1573 | r#"{"error":{"code":"subscription_sharing_usage_limit_exceeded_later"}}"#, |
| 1574 | ] { |
| 1575 | assert!(matches!( |
| 1576 | LlmError::from_http_response(429, body), |
| 1577 | LlmError::RateLimited { .. } |
| 1578 | )); |
| 1579 | assert!(LlmError::from_http_response(429, body).is_retryable()); |
| 1580 | } |
| 1581 | } |
| 1582 | |
| 1583 | fn assert_f64_eq(actual: f64, expected: f64) { |
| 1584 | assert!( |
| 1585 | (actual - expected).abs() < f64::EPSILON, |
| 1586 | "expected {expected}, got {actual}" |
| 1587 | ); |
| 1588 | } |
| 1589 | |
| 1590 | fn auth_user_message(error: LlmError) -> String { |
| 1591 | match error { |
| 1592 | LlmError::AuthenticationError(auth) => auth.to_user_message(), |
| 1593 | other => panic!("expected authentication error, got {other}"), |
| 1594 | } |
| 1595 | } |
| 1596 | |
| 1597 | #[test] |
| 1598 | fn test_retry_config_defaults() { |
| 1599 | let config = RetryConfig::default(); |
| 1600 | assert!(config.enabled); |
| 1601 | assert_eq!(config.max_retries, 3); |
| 1602 | assert_f64_eq(config.initial_delay, 1.0); |
| 1603 | assert_f64_eq(config.max_delay, 60.0); |
| 1604 | assert_f64_eq(config.exponential_base, 2.0); |
| 1605 | assert!(config.jitter); |
| 1606 | } |
| 1607 | |
| 1608 | #[test] |
| 1609 | fn test_retry_config_disabled() { |
| 1610 | let config = RetryConfig::disabled(); |
| 1611 | assert!(!config.enabled); |
| 1612 | } |
| 1613 | |
| 1614 | #[test] |
| 1615 | fn test_retry_config_builder() { |
| 1616 | let config = RetryConfig::new() |
| 1617 | .with_max_retries(5) |
| 1618 | .with_initial_delay(2.0) |
| 1619 | .with_max_delay(120.0) |
| 1620 | .with_jitter(false); |
| 1621 | |
| 1622 | assert_eq!(config.max_retries, 5); |
| 1623 | assert_f64_eq(config.initial_delay, 2.0); |
| 1624 | assert_f64_eq(config.max_delay, 120.0); |
| 1625 | assert!(!config.jitter); |
| 1626 | } |
| 1627 | |
| 1628 | #[test] |
| 1629 | fn test_delay_for_attempt_exponential() { |
| 1630 | let config = RetryConfig::new().with_jitter(false); |
| 1631 | |
| 1632 | // delay = initial * base^attempt |
| 1633 | // 1.0 * 2^0 = 1.0 |
| 1634 | let d0 = config.delay_for_attempt(0); |
| 1635 | assert_eq!(d0, Duration::from_secs_f64(1.0)); |
| 1636 | |
| 1637 | // 1.0 * 2^1 = 2.0 |
| 1638 | let d1 = config.delay_for_attempt(1); |
| 1639 | assert_eq!(d1, Duration::from_secs_f64(2.0)); |
| 1640 | |
| 1641 | // 1.0 * 2^2 = 4.0 |
| 1642 | let d2 = config.delay_for_attempt(2); |
| 1643 | assert_eq!(d2, Duration::from_secs_f64(4.0)); |
| 1644 | |
| 1645 | // 1.0 * 2^3 = 8.0 |
| 1646 | let d3 = config.delay_for_attempt(3); |
| 1647 | assert_eq!(d3, Duration::from_secs_f64(8.0)); |
| 1648 | } |
| 1649 | |
| 1650 | #[test] |
| 1651 | fn test_delay_for_attempt_capped() { |
| 1652 | let config = RetryConfig::new().with_jitter(false).with_max_delay(5.0); |
| 1653 | |
| 1654 | // 1.0 * 2^3 = 8.0, but capped at 5.0 |
| 1655 | let d3 = config.delay_for_attempt(3); |
| 1656 | assert_eq!(d3, Duration::from_secs_f64(5.0)); |
| 1657 | } |
| 1658 | |
| 1659 | #[test] |
| 1660 | fn test_delay_for_attempt_with_jitter() { |
| 1661 | let config = RetryConfig::new().with_jitter(true); |
| 1662 | |
| 1663 | // With jitter, delays should vary slightly |
| 1664 | let d1 = config.delay_for_attempt(1); |
| 1665 | let d2 = config.delay_for_attempt(1); |
| 1666 | |
| 1667 | // Both should be close to 2.0 seconds (within 10% jitter) |
| 1668 | let base = 2.0; |
| 1669 | let range = base * 0.1; |
| 1670 | assert!(d1.as_secs_f64() >= base - range); |
| 1671 | assert!(d1.as_secs_f64() <= base + range); |
| 1672 | assert!(d2.as_secs_f64() >= base - range); |
| 1673 | assert!(d2.as_secs_f64() <= base + range); |
| 1674 | } |
| 1675 | |
| 1676 | #[test] |
| 1677 | fn test_is_retryable_status() { |
| 1678 | let config = RetryConfig::default(); |
| 1679 | |
| 1680 | assert!(config.is_retryable_status(429)); // Rate limit |
| 1681 | assert!(config.is_retryable_status(499)); // Upstream request cancelled |
| 1682 | assert!(config.is_retryable_status(500)); // Internal server error |
| 1683 | assert!(config.is_retryable_status(502)); // Bad gateway |
| 1684 | assert!(config.is_retryable_status(503)); // Service unavailable |
| 1685 | assert!(config.is_retryable_status(504)); // Gateway timeout |
| 1686 | |
| 1687 | assert!(!config.is_retryable_status(400)); // Bad request |
| 1688 | assert!(!config.is_retryable_status(401)); // Unauthorized |
| 1689 | assert!(!config.is_retryable_status(403)); // Forbidden |
| 1690 | assert!(!config.is_retryable_status(404)); // Not found |
| 1691 | } |
| 1692 | |
| 1693 | #[test] |
| 1694 | fn auth_error_with_context_includes_provider_authority_model_and_key_source() { |
| 1695 | let err = LlmError::from_http_response_with_request_context( |
| 1696 | 401, |
| 1697 | "Invalid API Key", |
| 1698 | Some("Xiaomi MiMo"), |
| 1699 | Some("https://token-plan-sgp.xiaomimimo.com/v1"), |
| 1700 | Some("mimo-v2.5"), |
| 1701 | Some("env"), |
| 1702 | Some("tp-secret-token-plan-value"), |
| 1703 | ); |
| 1704 | let message = auth_user_message(err); |
| 1705 | |
| 1706 | assert!(message.contains("Invalid API Key")); |
| 1707 | assert!(message.contains("provider: Xiaomi MiMo")); |
| 1708 | assert!(message.contains("base URL authority: token-plan-sgp.xiaomimimo.com")); |
| 1709 | assert!(message.contains("model: mimo-v2.5")); |
| 1710 | assert!(message.contains("key source: env")); |
| 1711 | assert!(message.contains("key fingerprint: tp-... (len=26)")); |
| 1712 | } |
| 1713 | |
| 1714 | #[test] |
| 1715 | fn auth_error_redacts_full_api_key_from_body_and_context() { |
| 1716 | let api_key = "tp-secret-token-plan-value"; |
| 1717 | let err = LlmError::from_http_response_with_request_context( |
| 1718 | 401, |
| 1719 | &format!("Invalid API Key: {api_key}"), |
| 1720 | Some("Xiaomi MiMo"), |
| 1721 | Some("https://token-plan-sgp.xiaomimimo.com/v1"), |
| 1722 | Some("mimo-v2.5"), |
| 1723 | Some("config-file"), |
| 1724 | Some(api_key), |
| 1725 | ); |
| 1726 | let message = auth_user_message(err); |
| 1727 | |
| 1728 | assert!(!message.contains(api_key)); |
| 1729 | assert!(!message.contains("secret-token-plan-value")); |
| 1730 | assert!(message.contains("[redacted API key]")); |
| 1731 | assert!(message.contains("key fingerprint: tp-... (len=26)")); |
| 1732 | } |
| 1733 | |
| 1734 | #[test] |
| 1735 | fn auth_error_classifies_xiaomi_token_plan_key_prefix() { |
| 1736 | let token_plan = AuthenticationErrorContext::from_parts( |
| 1737 | None, |
| 1738 | None, |
| 1739 | None, |
| 1740 | Some("session"), |
| 1741 | Some("tp-secret-token-plan-value"), |
| 1742 | ); |
| 1743 | let generic = AuthenticationErrorContext::from_parts( |
| 1744 | None, |
| 1745 | None, |
| 1746 | None, |
| 1747 | Some("session"), |
| 1748 | Some("sk-other"), |
| 1749 | ); |
| 1750 | let unprefixed = AuthenticationErrorContext::from_parts( |
| 1751 | None, |
| 1752 | None, |
| 1753 | None, |
| 1754 | Some("session"), |
| 1755 | Some("plainsecretvalue"), |
| 1756 | ); |
| 1757 | |
| 1758 | assert_eq!( |
| 1759 | token_plan.key_kind.as_deref(), |
| 1760 | Some("Xiaomi MiMo Token Plan key") |
| 1761 | ); |
| 1762 | assert_eq!(generic.key_kind.as_deref(), Some("API key")); |
| 1763 | assert_eq!(unprefixed.key_kind.as_deref(), Some("API key")); |
| 1764 | assert_eq!( |
| 1765 | unprefixed.key_fingerprint.as_deref(), |
| 1766 | Some("unprefixed (len=16)") |
| 1767 | ); |
| 1768 | } |
| 1769 | |
| 1770 | #[test] |
| 1771 | fn authorization_403_is_not_reclassified_by_auth_context() { |
| 1772 | let err = LlmError::from_http_response_with_request_context( |
| 1773 | 403, |
| 1774 | "forbidden", |
| 1775 | Some("Arcee AI"), |
| 1776 | Some("https://api.arcee.ai/v1"), |
| 1777 | Some("auto"), |
| 1778 | Some("env"), |
| 1779 | Some("sk-arcee-secret"), |
| 1780 | ); |
| 1781 | |
| 1782 | assert!(matches!(err, LlmError::AuthorizationError(_))); |
| 1783 | } |
| 1784 | |
| 1785 | #[test] |
| 1786 | fn auth_error_without_context_preserves_bare_message() { |
| 1787 | let err = LlmError::from_http_response_with_auth_context( |
| 1788 | 401, |
| 1789 | "Invalid API Key", |
| 1790 | Some(AuthenticationErrorContext::default()), |
| 1791 | ); |
| 1792 | |
| 1793 | assert_eq!(auth_user_message(err), "Invalid API Key"); |
| 1794 | } |
| 1795 | |
| 1796 | /// A flat `{"error","message"}` body keeps its detail: the class alone |
| 1797 | /// ("Bad Request") told the person nothing about the rejected model. |
| 1798 | #[test] |
| 1799 | fn flat_error_and_message_body_surfaces_both_halves() { |
| 1800 | let body = r#"{"error":"Bad Request","message":"Invalid model name: 'invalid-model-xyz'"}"#; |
| 1801 | assert_eq!( |
| 1802 | sanitize_http_error_body(Some("Concentrate"), 400, body), |
| 1803 | "Bad Request: Invalid model name: 'invalid-model-xyz'" |
| 1804 | ); |
| 1805 | // Identical halves are not doubled, and the nested OpenAI shape is |
| 1806 | // untouched. |
| 1807 | assert_eq!( |
| 1808 | sanitize_http_error_body( |
| 1809 | Some("Concentrate"), |
| 1810 | 401, |
| 1811 | r#"{"error":"Unauthorized","message":"Unauthorized"}"# |
| 1812 | ), |
| 1813 | "Unauthorized" |
| 1814 | ); |
| 1815 | assert_eq!( |
| 1816 | sanitize_http_error_body( |
| 1817 | Some("fixture"), |
| 1818 | 400, |
| 1819 | r#"{"error":{"message":"nested detail"}}"# |
| 1820 | ), |
| 1821 | "nested detail" |
| 1822 | ); |
| 1823 | } |
| 1824 | |
| 1825 | #[test] |
| 1826 | fn cloudflare_html_error_is_summarized_without_raw_markup() { |
| 1827 | let body = r#"<!DOCTYPE html><html><head><title>Access Denied</title><style> |
| 1828 | .hidden { display: none; } |
| 1829 | </style></head><body> |
| 1830 | <h1>Access Denied</h1> |
| 1831 | <p>The action you just performed triggered a security alert.</p> |
| 1832 | <script>window.noisy = true;</script> |
| 1833 | <span>2600:1700:467:d410:f137:b94f:1dd0:d1e4</span> |
| 1834 | <span>a059a2873f3fdf82</span> |
| 1835 | <div>Cloudflare Error Pages</div> |
| 1836 | </body></html>"#; |
| 1837 | |
| 1838 | let message = sanitize_http_error_body(Some("Arcee AI"), 403, body); |
| 1839 | |
| 1840 | assert!(message.contains("Arcee AI API returned Cloudflare Access Denied")); |
| 1841 | assert!(message.contains("ID a059a2873f3fdf82")); |
| 1842 | assert!(!message.contains("<!DOCTYPE")); |
| 1843 | assert!(!message.contains("tailwindcss")); |
| 1844 | assert!(message.len() < 300); |
| 1845 | } |
| 1846 | |
| 1847 | #[test] |
| 1848 | fn cloudflare_access_denied_403_is_authorization_not_authentication() { |
| 1849 | let message = sanitize_http_error_body( |
| 1850 | Some("Arcee AI"), |
| 1851 | 403, |
| 1852 | r#"<!doctype html><html><body><h1>Access Denied</h1><p>Cloudflare Error Pages</p></body></html>"#, |
| 1853 | ); |
| 1854 | let err = LlmError::from_http_response(403, &message); |
| 1855 | |
| 1856 | assert!(matches!(err, LlmError::AuthorizationError(_))); |
| 1857 | } |
| 1858 | |
| 1859 | #[test] |
| 1860 | fn arcee_access_denied_without_literal_cloudflare_is_still_summarized() { |
| 1861 | // Mirrors api.arcee.ai's real 403 page: "Cloudflare" appears only in a |
| 1862 | // `<meta>` attribute and the `<style>` block, both stripped, so the |
| 1863 | // visible text never contains it. The summary must still fire from the |
| 1864 | // WAF's stock "security alert" / "Contact Support" copy + error ID. |
| 1865 | let body = r#"<!DOCTYPE html><html lang="en"><head> |
| 1866 | <meta name="description" content="Cloudflare Error Pages"> |
| 1867 | <title>Access Denied</title> |
| 1868 | <style>:root{--accent:cloudflare}</style></head><body> |
| 1869 | <h1>Access Denied</h1> |
| 1870 | <p>The action you just performed triggered a security alert.</p> |
| 1871 | <p>Please contact us if this was a mistake.</p> |
| 1872 | <a>Contact Support</a> |
| 1873 | <span>2600:1700:467:d410:f137:b94f:1dd0:d1e4</span> |
| 1874 | <span>a059c0d4caf1f9cc</span> |
| 1875 | </body></html>"#; |
| 1876 | |
| 1877 | let message = sanitize_http_error_body(Some("Arcee AI"), 403, body); |
| 1878 | |
| 1879 | assert!( |
| 1880 | message.contains("Arcee AI API returned Access Denied"), |
| 1881 | "got: {message}" |
| 1882 | ); |
| 1883 | assert!(message.contains("ID a059c0d4caf1f9cc"), "got: {message}"); |
| 1884 | assert!( |
| 1885 | !message.to_ascii_lowercase().contains("cloudflare"), |
| 1886 | "stripped Arcee page has no literal Cloudflare: {message}" |
| 1887 | ); |
| 1888 | assert!(!message.contains('<'), "no raw markup: {message}"); |
| 1889 | assert!(message.len() < 300, "stays concise: {message}"); |
| 1890 | |
| 1891 | // A WAF block is authorization, not a bad API key. |
| 1892 | let err = LlmError::from_http_response(403, &message); |
| 1893 | assert!(matches!(err, LlmError::AuthorizationError(_))); |
| 1894 | } |
| 1895 | |
| 1896 | #[test] |
| 1897 | fn test_llm_error_suggested_retry_delay() { |
| 1898 | let err = LlmError::RateLimited { |
| 1899 | message: "slow down".to_string(), |
| 1900 | retry_after: Some(Duration::from_secs(60)), |
| 1901 | }; |
| 1902 | assert_eq!(err.suggested_retry_delay(), Some(Duration::from_secs(60))); |
| 1903 | |
| 1904 | let err = LlmError::ServerError { |
| 1905 | status: 500, |
| 1906 | message: "error".to_string(), |
| 1907 | }; |
| 1908 | assert_eq!(err.suggested_retry_delay(), None); |
| 1909 | } |
| 1910 | |
| 1911 | #[test] |
| 1912 | fn test_parse_retry_after() { |
| 1913 | // Integer seconds |
| 1914 | assert_eq!(parse_retry_after("120"), Some(Duration::from_secs(120))); |
| 1915 | assert_eq!(parse_retry_after("0"), Some(Duration::from_secs(0))); |
| 1916 | |
| 1917 | // Float seconds keep sub-second precision |
| 1918 | assert_eq!(parse_retry_after("1.5"), Some(Duration::from_secs_f64(1.5))); |
| 1919 | |
| 1920 | // Invalid |
| 1921 | assert_eq!(parse_retry_after("invalid"), None); |
| 1922 | assert_eq!(parse_retry_after(""), None); |
| 1923 | } |
| 1924 | |
| 1925 | /// A `Retry-After` value is server-controlled. Malformed floats used to |
| 1926 | /// reach `Duration::from_secs_f64`, which panics on a negative — a |
| 1927 | /// remote-triggerable crash in the request path (2026-08-04 review). |
| 1928 | #[test] |
| 1929 | fn parse_retry_after_never_panics_and_is_bounded_on_hostile_input() { |
| 1930 | // None of these may panic. |
| 1931 | assert_eq!(parse_retry_after("-5"), None, "negative is not a delay"); |
| 1932 | assert_eq!(parse_retry_after("nan"), None); |
| 1933 | assert_eq!(parse_retry_after("inf"), None); |
| 1934 | assert_eq!(parse_retry_after("-inf"), None); |
| 1935 | // Absurdly large values clamp to the ceiling rather than overflowing |
| 1936 | // or wedging the turn for a day. |
| 1937 | assert_eq!(parse_retry_after("1e300"), Some(RETRY_AFTER_MAX)); |
| 1938 | assert_eq!(parse_retry_after("86400"), Some(RETRY_AFTER_MAX)); |
| 1939 | assert_eq!( |
| 1940 | parse_retry_after("999999999999"), |
| 1941 | Some(RETRY_AFTER_MAX), |
| 1942 | "integer path is clamped too" |
| 1943 | ); |
| 1944 | // A normal value still passes through untouched. |
| 1945 | assert_eq!(parse_retry_after("30"), Some(Duration::from_secs(30))); |
| 1946 | } |
| 1947 | |
| 1948 | #[test] |
| 1949 | fn test_retry_policy_conversion() { |
| 1950 | let policy = RetryPolicy { |
| 1951 | enabled: true, |
| 1952 | max_retries: 5, |
| 1953 | initial_delay: 2.0, |
| 1954 | max_delay: 30.0, |
| 1955 | exponential_base: 3.0, |
| 1956 | jitter: false, |
| 1957 | jitter_factor: 0.25, |
| 1958 | respect_retry_after: false, |
| 1959 | }; |
| 1960 | |
| 1961 | let config: RetryConfig = policy.clone().into(); |
| 1962 | assert_eq!(config.enabled, policy.enabled); |
| 1963 | assert_eq!(config.max_retries, policy.max_retries); |
| 1964 | assert_f64_eq(config.initial_delay, policy.initial_delay); |
| 1965 | assert_f64_eq(config.max_delay, policy.max_delay); |
| 1966 | assert_f64_eq(config.exponential_base, policy.exponential_base); |
| 1967 | |
| 1968 | // #6700: the jitter and Retry-After knobs survive the conversion |
| 1969 | // instead of silently resetting to `RetryConfig::default()`. |
| 1970 | assert!(!config.jitter); |
| 1971 | assert_f64_eq(config.jitter_factor, 0.25); |
| 1972 | assert!(!config.respect_retry_after); |
| 1973 | |
| 1974 | // Convert back |
| 1975 | let policy2: RetryPolicy = config.into(); |
| 1976 | assert_eq!(policy2.enabled, policy.enabled); |
| 1977 | assert_eq!(policy2.max_retries, policy.max_retries); |
| 1978 | assert_eq!(policy2.jitter, policy.jitter); |
| 1979 | assert_f64_eq(policy2.jitter_factor, policy.jitter_factor); |
| 1980 | assert_eq!(policy2.respect_retry_after, policy.respect_retry_after); |
| 1981 | } |
| 1982 | |
| 1983 | #[tokio::test] |
| 1984 | async fn test_with_retry_success_first_attempt() { |
| 1985 | let config = RetryConfig::default(); |
| 1986 | let mut call_count = 0; |
| 1987 | |
| 1988 | let result = with_retry( |
| 1989 | &config, |
| 1990 | || { |
| 1991 | call_count += 1; |
| 1992 | async { Ok::<_, LlmError>(42) } |
| 1993 | }, |
| 1994 | None, |
| 1995 | ) |
| 1996 | .await; |
| 1997 | |
| 1998 | assert!(result.is_ok()); |
| 1999 | assert_eq!(result.unwrap(), 42); |
| 2000 | assert_eq!(call_count, 1); |
| 2001 | } |
| 2002 | |
| 2003 | #[tokio::test] |
| 2004 | async fn test_with_retry_disabled() { |
| 2005 | let config = RetryConfig::disabled(); |
| 2006 | let mut call_count = 0; |
| 2007 | |
| 2008 | let result: RetryResult<i32> = with_retry( |
| 2009 | &config, |
| 2010 | || { |
| 2011 | call_count += 1; |
| 2012 | async { |
| 2013 | Err(LlmError::ServerError { |
| 2014 | status: 500, |
| 2015 | message: "error".to_string(), |
| 2016 | }) |
| 2017 | } |
| 2018 | }, |
| 2019 | None, |
| 2020 | ) |
| 2021 | .await; |
| 2022 | |
| 2023 | assert!(result.is_err()); |
| 2024 | assert_eq!(call_count, 1); // No retries when disabled |
| 2025 | } |
| 2026 | |
| 2027 | #[tokio::test] |
| 2028 | async fn test_with_retry_eventual_success() { |
| 2029 | let config = RetryConfig::new() |
| 2030 | .with_max_retries(3) |
| 2031 | .with_initial_delay(0.01); // Fast for testing |
| 2032 | |
| 2033 | let call_count = std::sync::Arc::new(std::sync::atomic::AtomicU32::new(0)); |
| 2034 | let cc = call_count.clone(); |
| 2035 | |
| 2036 | let result = with_retry( |
| 2037 | &config, |
| 2038 | || { |
| 2039 | let count = cc.fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 2040 | async move { |
| 2041 | if count < 2 { |
| 2042 | Err(LlmError::ServerError { |
| 2043 | status: 500, |
| 2044 | message: "temporary error".to_string(), |
| 2045 | }) |
| 2046 | } else { |
| 2047 | Ok::<_, LlmError>(42) |
| 2048 | } |
| 2049 | } |
| 2050 | }, |
| 2051 | None, |
| 2052 | ) |
| 2053 | .await; |
| 2054 | |
| 2055 | assert!(result.is_ok()); |
| 2056 | assert_eq!(result.unwrap(), 42); |
| 2057 | assert_eq!(call_count.load(std::sync::atomic::Ordering::SeqCst), 3); // 2 failures + 1 success |
| 2058 | } |
| 2059 | |
| 2060 | #[tokio::test] |
| 2061 | async fn test_with_retry_exhausted() { |
| 2062 | let config = RetryConfig::new() |
| 2063 | .with_max_retries(2) |
| 2064 | .with_initial_delay(0.01); |
| 2065 | |
| 2066 | let call_count = std::sync::Arc::new(std::sync::atomic::AtomicU32::new(0)); |
| 2067 | let cc = call_count.clone(); |
| 2068 | |
| 2069 | let result: RetryResult<i32> = with_retry( |
| 2070 | &config, |
| 2071 | || { |
| 2072 | cc.fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 2073 | async { |
| 2074 | Err(LlmError::ServerError { |
| 2075 | status: 500, |
| 2076 | message: "persistent error".to_string(), |
| 2077 | }) |
| 2078 | } |
| 2079 | }, |
| 2080 | None, |
| 2081 | ) |
| 2082 | .await; |
| 2083 | |
| 2084 | assert!(result.is_err()); |
| 2085 | let err = result.unwrap_err(); |
| 2086 | assert_eq!(err.attempts, 3); // 1 initial + 2 retries |
| 2087 | assert_eq!(call_count.load(std::sync::atomic::Ordering::SeqCst), 3); |
| 2088 | } |
| 2089 | |
| 2090 | #[tokio::test] |
| 2091 | async fn test_with_retry_callback() { |
| 2092 | let config = RetryConfig::new() |
| 2093 | .with_max_retries(2) |
| 2094 | .with_initial_delay(0.01); |
| 2095 | |
| 2096 | let callback_count = std::sync::Arc::new(std::sync::atomic::AtomicU32::new(0)); |
| 2097 | let cc = callback_count.clone(); |
| 2098 | |
| 2099 | let _: RetryResult<i32> = with_retry( |
| 2100 | &config, |
| 2101 | || async { |
| 2102 | Err(LlmError::ServerError { |
| 2103 | status: 500, |
| 2104 | message: "error".to_string(), |
| 2105 | }) |
| 2106 | }, |
| 2107 | Some(Box::new(move |_err, _attempt, _delay| { |
| 2108 | cc.fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 2109 | })), |
| 2110 | ) |
| 2111 | .await; |
| 2112 | |
| 2113 | // Callback called once per retry (not for the final failure) |
| 2114 | assert_eq!(callback_count.load(std::sync::atomic::Ordering::SeqCst), 2); |
| 2115 | } |
| 2116 | |
| 2117 | #[test] |
| 2118 | fn test_retry_error_display() { |
| 2119 | let err = RetryError { |
| 2120 | last_error: LlmError::ServerError { |
| 2121 | status: 500, |
| 2122 | message: "internal error".to_string(), |
| 2123 | }, |
| 2124 | attempts: 4, |
| 2125 | total_time: Duration::from_secs(10), |
| 2126 | }; |
| 2127 | |
| 2128 | let display = format!("{err}"); |
| 2129 | assert!(display.contains("4 attempts")); |
| 2130 | assert!(display.contains("10")); |
| 2131 | assert!(display.contains("Server error")); |
| 2132 | } |
| 2133 | } |
| 2134 |