| 1 | //! Model selection and auto-routing. |
| 2 | //! |
| 3 | //! The CLI, TUI, runtime threads, subagents, and command handlers all need |
| 4 | //! this behavior, so it intentionally lives outside the command tree. |
| 5 | |
| 6 | use std::time::Duration; |
| 7 | |
| 8 | use anyhow::Result; |
| 9 | use serde::{Deserialize, Serialize}; |
| 10 | |
| 11 | use crate::client::CodewhaleClient; |
| 12 | use crate::config::{ApiProvider, Config, normalize_model_name_for_provider}; |
| 13 | use crate::cost_status::{ |
| 14 | EffectiveRouteEnvelope, EffectiveRouteUsage, RuntimeUsageDropRecord, RuntimeUsageRecord, |
| 15 | }; |
| 16 | use crate::llm_client::LlmClient; |
| 17 | use crate::model_inventory::ModelInventory; |
| 18 | use crate::reasoning_preference::ReasoningEffort; |
| 19 | use codewhale_models::Role; |
| 20 | use codewhale_models::{ContentBlock, Message, MessageRequest, MessageResponse, SystemPrompt}; |
| 21 | |
| 22 | /// Big/cheap model pair the auto-router may choose between for the active |
| 23 | /// provider (#3018). |
| 24 | /// |
| 25 | /// `cheap == None` means the provider has no known cheap tier: the local |
| 26 | /// fallback stays on the current model (only thinking effort varies) and the |
| 27 | /// network router is skipped entirely (#1549). |
| 28 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 29 | pub(crate) struct RouterCandidates { |
| 30 | pub(crate) big: String, |
| 31 | pub(crate) cheap: Option<String>, |
| 32 | } |
| 33 | |
| 34 | impl RouterCandidates { |
| 35 | pub(crate) fn deepseek() -> Self { |
| 36 | Self { |
| 37 | big: "deepseek-v4-pro".to_string(), |
| 38 | cheap: Some("deepseek-v4-flash".to_string()), |
| 39 | } |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | /// Return a provider-owned strong/fast pair for model families whose catalog |
| 44 | /// exposes more than one tier. The ids here are deliberately explicit: a |
| 45 | /// model name alone is not evidence that another provider can serve its |
| 46 | /// sibling, so unknown providers and unknown families remain single-tier. |
| 47 | fn catalog_family_candidates( |
| 48 | provider: ApiProvider, |
| 49 | current_model: &str, |
| 50 | ) -> Option<RouterCandidates> { |
| 51 | let normalized = normalize_model_name_for_provider(provider, current_model) |
| 52 | .unwrap_or_else(|| current_model.trim().to_string()); |
| 53 | let lower = normalized.to_ascii_lowercase(); |
| 54 | |
| 55 | let cheap = match provider { |
| 56 | ApiProvider::Openai | ApiProvider::OpenaiCodex |
| 57 | if matches!(lower.as_str(), "gpt-5.6" | "gpt-5.6-sol" | "gpt-5.6-terra") => |
| 58 | { |
| 59 | Some("gpt-5.6-luna".to_string()) |
| 60 | } |
| 61 | ApiProvider::Anthropic |
| 62 | if matches!( |
| 63 | lower.as_str(), |
| 64 | "claude-opus-4-8" | "claude-sonnet-4-6" | "claude-sonnet-5" |
| 65 | ) => |
| 66 | { |
| 67 | Some("claude-haiku-4-5".to_string()) |
| 68 | } |
| 69 | ApiProvider::XiaomiMimo if lower == "mimo-v2.5-pro" => Some("mimo-v2.5".to_string()), |
| 70 | ApiProvider::Arcee |
| 71 | if matches!( |
| 72 | lower.as_str(), |
| 73 | "trinity-large-thinking" | "trinity-large-preview" |
| 74 | ) => |
| 75 | { |
| 76 | Some("trinity-mini".to_string()) |
| 77 | } |
| 78 | ApiProvider::Moonshot if lower == "kimi-k2.7-code" => Some("kimi-k2.6".to_string()), |
| 79 | ApiProvider::Minimax | ApiProvider::MinimaxAnthropic if lower == "minimax-m2.7" => { |
| 80 | Some("MiniMax-M2.7-highspeed".to_string()) |
| 81 | } |
| 82 | ApiProvider::OpencodeGo if lower == "kimi-k3" => Some("kimi-k2.7-code".to_string()), |
| 83 | ApiProvider::Openrouter |
| 84 | if lower == "qwen/qwen3.6-max-preview" |
| 85 | || lower == "qwen/qwen3.6-plus" |
| 86 | || lower == "qwen/qwen3.6-27b" |
| 87 | || lower == "qwen/qwen3.6-35b-a3b" => |
| 88 | { |
| 89 | Some("qwen/qwen3.6-flash".to_string()) |
| 90 | } |
| 91 | ApiProvider::Openrouter if lower == "xiaomi/mimo-v2.5-pro" => { |
| 92 | Some("xiaomi/mimo-v2.5".to_string()) |
| 93 | } |
| 94 | ApiProvider::Openrouter |
| 95 | if matches!( |
| 96 | lower.as_str(), |
| 97 | "arcee-ai/trinity-large-thinking" | "arcee-ai/trinity-large-preview" |
| 98 | ) => |
| 99 | { |
| 100 | Some("arcee-ai/trinity-mini".to_string()) |
| 101 | } |
| 102 | ApiProvider::Openrouter if lower == "moonshotai/kimi-k2.7-code" => { |
| 103 | Some("moonshotai/kimi-k2.6".to_string()) |
| 104 | } |
| 105 | ApiProvider::Openrouter |
| 106 | if lower == "anthropic/claude-opus-4-8" |
| 107 | || lower == "anthropic/claude-sonnet-4-6" |
| 108 | || lower == "anthropic/claude-sonnet-5" => |
| 109 | { |
| 110 | Some("anthropic/claude-haiku-4-5".to_string()) |
| 111 | } |
| 112 | _ => None, |
| 113 | }?; |
| 114 | |
| 115 | Some(RouterCandidates { |
| 116 | big: normalized, |
| 117 | cheap: Some(cheap), |
| 118 | }) |
| 119 | } |
| 120 | |
| 121 | /// Derive the auto-router's candidate pair for the active provider (#3018). |
| 122 | /// |
| 123 | /// DeepSeek providers route between the canonical pro/flash pair. Hosted |
| 124 | /// routes with known wire ids for that pair (NVIDIA NIM, OpenRouter, Novita, |
| 125 | /// SiliconFlow, SGLang, vLLM, Wanjie Ark, Volcengine) use their provider |
| 126 | /// spellings. Every other provider has no known cheap tier: `big` is the |
| 127 | /// session model and `cheap` is `None`, so auto mode never fabricates a |
| 128 | /// DeepSeek id for a provider that cannot serve it. |
| 129 | pub(crate) fn provider_router_candidates( |
| 130 | provider: crate::config::ApiProvider, |
| 131 | current_model: &str, |
| 132 | ) -> RouterCandidates { |
| 133 | use crate::config::ApiProvider; |
| 134 | if let Some(candidates) = catalog_family_candidates(provider, current_model) { |
| 135 | return candidates; |
| 136 | } |
| 137 | |
| 138 | if provider == ApiProvider::Zai { |
| 139 | let normalized = crate::config::normalize_model_name_for_provider(provider, current_model) |
| 140 | .unwrap_or_else(|| current_model.to_string()); |
| 141 | return RouterCandidates { |
| 142 | // GLM-5.3 routes faster/explore children to GLM-5.3-Flash. |
| 143 | // GLM-5.2 still uses GLM-5-Turbo. Flash, Turbo, and 5.1 have no |
| 144 | // cheaper tier and keep children on the parent model. |
| 145 | cheap: if normalized == crate::config::ZAI_GLM_5_3_MODEL { |
| 146 | Some(crate::config::ZAI_GLM_5_3_FLASH_MODEL.to_string()) |
| 147 | } else if normalized == crate::config::ZAI_GLM_5_2_MODEL { |
| 148 | Some(crate::config::ZAI_GLM_5_TURBO_MODEL.to_string()) |
| 149 | } else { |
| 150 | None |
| 151 | }, |
| 152 | big: normalized, |
| 153 | }; |
| 154 | } |
| 155 | |
| 156 | if provider == ApiProvider::Openrouter |
| 157 | && let Some(normalized) = |
| 158 | crate::config::normalize_model_name_for_provider(provider, current_model) |
| 159 | && matches!( |
| 160 | normalized.as_str(), |
| 161 | crate::config::OPENROUTER_GLM_5_1_MODEL |
| 162 | | crate::config::OPENROUTER_GLM_5_2_MODEL |
| 163 | | crate::config::OPENROUTER_GLM_5_3_MODEL |
| 164 | | crate::config::OPENROUTER_GLM_5_3_FLASH_MODEL |
| 165 | | crate::config::OPENROUTER_GLM_5_TURBO_MODEL |
| 166 | ) |
| 167 | { |
| 168 | return RouterCandidates { |
| 169 | // z-ai/glm-5.3 routes faster children to z-ai/glm-5.3-flash; |
| 170 | // z-ai/glm-5.2 still uses z-ai/glm-5-turbo. Flash, turbo, and 5.1 |
| 171 | // have no cheaper tier and keep children on parent. |
| 172 | cheap: if normalized == crate::config::OPENROUTER_GLM_5_3_MODEL { |
| 173 | Some(crate::config::OPENROUTER_GLM_5_3_FLASH_MODEL.to_string()) |
| 174 | } else if normalized == crate::config::OPENROUTER_GLM_5_2_MODEL { |
| 175 | Some(crate::config::OPENROUTER_GLM_5_TURBO_MODEL.to_string()) |
| 176 | } else { |
| 177 | None |
| 178 | }, |
| 179 | big: normalized, |
| 180 | }; |
| 181 | } |
| 182 | |
| 183 | match provider { |
| 184 | ApiProvider::Deepseek | ApiProvider::DeepseekCN => RouterCandidates::deepseek(), |
| 185 | ApiProvider::NvidiaNim |
| 186 | | ApiProvider::Openrouter |
| 187 | | ApiProvider::Novita |
| 188 | | ApiProvider::Siliconflow |
| 189 | | ApiProvider::SiliconflowCn |
| 190 | | ApiProvider::Sglang |
| 191 | | ApiProvider::Vllm |
| 192 | | ApiProvider::WanjieArk |
| 193 | if current_model.to_ascii_lowercase().contains("deepseek") => |
| 194 | { |
| 195 | RouterCandidates { |
| 196 | big: crate::config::wire_model_for_provider(provider, "deepseek-v4-pro"), |
| 197 | cheap: Some(crate::config::wire_model_for_provider( |
| 198 | provider, |
| 199 | "deepseek-v4-flash", |
| 200 | )), |
| 201 | } |
| 202 | } |
| 203 | ApiProvider::Volcengine if current_model.to_ascii_lowercase().contains("deepseek") => { |
| 204 | RouterCandidates { |
| 205 | big: crate::config::DEFAULT_VOLCENGINE_MODEL.to_string(), |
| 206 | cheap: Some(crate::config::DEFAULT_VOLCENGINE_FLASH_MODEL.to_string()), |
| 207 | } |
| 208 | } |
| 209 | _ => RouterCandidates { |
| 210 | big: current_model.to_string(), |
| 211 | cheap: None, |
| 212 | }, |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 217 | pub(crate) enum AutoRouteSource { |
| 218 | FlashRouter, |
| 219 | Heuristic, |
| 220 | } |
| 221 | |
| 222 | impl AutoRouteSource { |
| 223 | #[must_use] |
| 224 | pub(crate) fn label(self) -> &'static str { |
| 225 | match self { |
| 226 | AutoRouteSource::FlashRouter => "classifier", |
| 227 | AutoRouteSource::Heuristic => "heuristic", |
| 228 | } |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | /// Provider-safe tier reported for the concrete Auto route. |
| 233 | /// |
| 234 | /// `Selected` is deliberately neutral: a classifier may choose a runnable |
| 235 | /// inventory model that is not part of a known strong/fast pair, and the UI |
| 236 | /// must not invent a tier from the model id. |
| 237 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 238 | #[serde(rename_all = "snake_case")] |
| 239 | pub(crate) enum AutoRouteTier { |
| 240 | Strong, |
| 241 | Fast, |
| 242 | Only, |
| 243 | Selected, |
| 244 | } |
| 245 | |
| 246 | impl AutoRouteTier { |
| 247 | #[must_use] |
| 248 | pub(crate) fn label(self) -> &'static str { |
| 249 | match self { |
| 250 | Self::Strong => "strong", |
| 251 | Self::Fast => "fast", |
| 252 | Self::Only => "only model", |
| 253 | Self::Selected => "selected", |
| 254 | } |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | /// Scope from which the concrete Auto route was selected. |
| 259 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 260 | #[serde(rename_all = "snake_case")] |
| 261 | pub(crate) enum AutoRouteScope { |
| 262 | /// The network classifier could choose any runnable provider/model pair in |
| 263 | /// the redacted inventory. Only reachable under the persisted |
| 264 | /// `[auto] cross_provider = true` opt-in (#4411). |
| 265 | RunnableProviders, |
| 266 | /// The network classifier saw only the active provider's runnable routes — |
| 267 | /// the default Auto scope (#4411). |
| 268 | ActiveProvider, |
| 269 | /// The local declared fallback selected within one resolved route. |
| 270 | ResolvedProvider, |
| 271 | } |
| 272 | |
| 273 | impl AutoRouteScope { |
| 274 | #[must_use] |
| 275 | pub(crate) fn label(self) -> &'static str { |
| 276 | match self { |
| 277 | Self::RunnableProviders => "runnable providers", |
| 278 | Self::ActiveProvider => "active provider only", |
| 279 | Self::ResolvedProvider => "resolved provider", |
| 280 | } |
| 281 | } |
| 282 | } |
| 283 | |
| 284 | /// Non-secret data path used to make an Auto decision. |
| 285 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 286 | #[serde(rename_all = "snake_case")] |
| 287 | pub(crate) enum AutoRouteDataPath { |
| 288 | LocalHeuristic, |
| 289 | Classifier { |
| 290 | provider: ApiProvider, |
| 291 | model: String, |
| 292 | }, |
| 293 | } |
| 294 | |
| 295 | impl AutoRouteDataPath { |
| 296 | #[must_use] |
| 297 | pub(crate) fn label(&self) -> String { |
| 298 | match self { |
| 299 | Self::LocalHeuristic => "local only (no router request)".to_string(), |
| 300 | Self::Classifier { provider, model } => format!( |
| 301 | "latest request + bounded recent context -> {} / {model}", |
| 302 | provider.display_name() |
| 303 | ), |
| 304 | } |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | /// Local signal that selected the provider-safe strong/fast candidate. |
| 309 | /// |
| 310 | /// Since the #6290 rework the local fallback never judges request content: |
| 311 | /// without the flash classifier there is no per-request signal, so the route |
| 312 | /// is the configured default (or the runnable fast sibling under the explicit |
| 313 | /// `[auto] cost_saving` opt-in). The content-derived variants below are never |
| 314 | /// constructed for new routes; they are retained so saved sessions from before |
| 315 | /// the rework still deserialize. |
| 316 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 317 | #[serde(rename_all = "snake_case")] |
| 318 | pub(crate) enum AutoRouteHeuristicReason { |
| 319 | /// Legacy: the deleted keyword/length classifier judged the request |
| 320 | /// complex. Retained for saved-session serde compat only. |
| 321 | ComplexRequest, |
| 322 | /// Legacy: the deleted length rule judged the request short. Retained |
| 323 | /// for saved-session serde compat only. |
| 324 | ShortRequest, |
| 325 | /// Legacy: the deleted length rule judged the request long. Retained |
| 326 | /// for saved-session serde compat only. |
| 327 | LongRequest, |
| 328 | CostSavingPolicy, |
| 329 | /// Legacy: the deleted classifier judged the request routine. Retained |
| 330 | /// for saved-session serde compat only. |
| 331 | RoutineRequest, |
| 332 | NoFastSibling, |
| 333 | NoRunnableCandidate, |
| 334 | /// The configured default model: no classifier was available and no |
| 335 | /// content signal was consulted. |
| 336 | DeclaredDefault, |
| 337 | } |
| 338 | |
| 339 | impl AutoRouteHeuristicReason { |
| 340 | #[must_use] |
| 341 | fn label(self) -> &'static str { |
| 342 | match self { |
| 343 | Self::ComplexRequest => "complex request", |
| 344 | Self::ShortRequest => "short request", |
| 345 | Self::LongRequest => "long request", |
| 346 | Self::CostSavingPolicy => "cost-saving policy", |
| 347 | Self::RoutineRequest => "routine request", |
| 348 | Self::NoFastSibling => "no runnable fast sibling", |
| 349 | Self::NoRunnableCandidate => "no runnable inventory candidate", |
| 350 | Self::DeclaredDefault => "configured default (no classifier)", |
| 351 | } |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | /// Why the route was selected. Classifier failures are intentionally |
| 356 | /// collapsed to a non-secret reason; provider errors and response bodies must |
| 357 | /// never enter diagnostics. |
| 358 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 359 | #[serde(rename_all = "snake_case")] |
| 360 | pub(crate) enum AutoRouteReason { |
| 361 | ClassifierRecommendation, |
| 362 | /// The `local_heuristic` alias keeps sessions saved before the #6290 |
| 363 | /// rework loadable; new sessions persist `local_fallback`. |
| 364 | #[serde(alias = "local_heuristic")] |
| 365 | LocalFallback(AutoRouteHeuristicReason), |
| 366 | ClassifierFallback(AutoRouteHeuristicReason), |
| 367 | } |
| 368 | |
| 369 | impl AutoRouteReason { |
| 370 | #[must_use] |
| 371 | pub(crate) fn label(self) -> String { |
| 372 | match self { |
| 373 | Self::ClassifierRecommendation => "classifier recommendation".to_string(), |
| 374 | Self::LocalFallback(reason) => format!("local fallback: {}", reason.label()), |
| 375 | Self::ClassifierFallback(reason) => { |
| 376 | format!("classifier fallback: {}", reason.label()) |
| 377 | } |
| 378 | } |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | /// Effective provider-scoped model pair used to classify the selected tier. |
| 383 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 384 | pub(crate) struct AutoRoutePair { |
| 385 | pub(crate) strong: String, |
| 386 | pub(crate) fast: Option<String>, |
| 387 | } |
| 388 | |
| 389 | /// Per-turn Auto routing diagnostics. Provider/model identity remains owned by |
| 390 | /// the authoritative runtime `TurnRoute`; this receipt only records how the |
| 391 | /// concrete route was chosen. |
| 392 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 393 | pub(crate) struct AutoRouteReceipt { |
| 394 | pub(crate) tier: AutoRouteTier, |
| 395 | pub(crate) pair: AutoRoutePair, |
| 396 | pub(crate) scope: AutoRouteScope, |
| 397 | pub(crate) data_path: AutoRouteDataPath, |
| 398 | pub(crate) reason: AutoRouteReason, |
| 399 | } |
| 400 | |
| 401 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 402 | pub(crate) struct AutoRouteSelection { |
| 403 | pub(crate) provider: ApiProvider, |
| 404 | pub(crate) model: String, |
| 405 | pub(crate) reasoning_effort: Option<ReasoningEffort>, |
| 406 | pub(crate) source: AutoRouteSource, |
| 407 | /// Present for Auto decisions; explicit inventory lookups intentionally do |
| 408 | /// not pretend to be Auto routing receipts. |
| 409 | pub(crate) receipt: Option<AutoRouteReceipt>, |
| 410 | /// Provider calls made to choose this route. These are deliberately kept |
| 411 | /// separate from the selected parent route: a classifier may run on a |
| 412 | /// different provider/model/quote, so pricing it under the eventual turn |
| 413 | /// would double-charge the parent and lose the classifier's real route. |
| 414 | /// |
| 415 | /// Auto currently admits at most one classifier request per selection. |
| 416 | pub(crate) routed_usage: Vec<RuntimeUsageRecord>, |
| 417 | /// Exact frozen routes for admitted classifier calls whose provider |
| 418 | /// response omitted usage. The count below remains authoritative and may |
| 419 | /// exceed this bounded vector after overflow. |
| 420 | pub(crate) routed_usage_drop_records: Vec<RuntimeUsageDropRecord>, |
| 421 | /// Classifier requests admitted to dispatch whose response usage could not |
| 422 | /// be recovered (timeout/transport failure). Consumers must surface this |
| 423 | /// as incomplete coverage rather than silently treating it as zero spend. |
| 424 | pub(crate) routed_usage_dropped_records: u64, |
| 425 | } |
| 426 | |
| 427 | fn extract_first_json_object(raw: &str) -> Option<&str> { |
| 428 | let start = raw.find('{')?; |
| 429 | let end = raw.rfind('}')?; |
| 430 | (end >= start).then_some(&raw[start..=end]) |
| 431 | } |
| 432 | |
| 433 | fn parse_auto_route_reasoning_effort(effort: &str) -> Option<ReasoningEffort> { |
| 434 | ReasoningEffort::parse_strict(effort).ok() |
| 435 | } |
| 436 | |
| 437 | /// Normalize an Auto-route effort when only the provider is known. |
| 438 | /// |
| 439 | /// This delegates to the one authoritative normalizer, |
| 440 | /// [`ReasoningEffort::normalize_for_route`], with an unresolved route (empty |
| 441 | /// endpoint and wire model), so the Auto path cannot carry a second copy of |
| 442 | /// the historic `low | medium -> high` provider collapse (Slice 4, D2). |
| 443 | #[must_use] |
| 444 | pub(crate) fn normalize_auto_route_effort_for_provider( |
| 445 | provider: ApiProvider, |
| 446 | effort: ReasoningEffort, |
| 447 | ) -> ReasoningEffort { |
| 448 | effort.normalize_for_route(provider, "", "") |
| 449 | } |
| 450 | |
| 451 | /// Select the reasoning request that accompanies an Auto-model route. |
| 452 | /// |
| 453 | /// Model routing and reasoning routing are independent. An explicit fixed |
| 454 | /// preference wins over the classifier's suggestion; an absent preference or |
| 455 | /// explicit `Auto` keeps reasoning under per-prompt control. |
| 456 | #[must_use] |
| 457 | pub(crate) fn resolve_auto_model_reasoning( |
| 458 | preference: Option<ReasoningEffort>, |
| 459 | routed: Option<ReasoningEffort>, |
| 460 | ) -> (Option<ReasoningEffort>, bool) { |
| 461 | match preference { |
| 462 | Some( |
| 463 | effort @ (ReasoningEffort::Off |
| 464 | | ReasoningEffort::Minimal |
| 465 | | ReasoningEffort::Low |
| 466 | | ReasoningEffort::Medium |
| 467 | | ReasoningEffort::High |
| 468 | | ReasoningEffort::XHigh |
| 469 | | ReasoningEffort::Ultra |
| 470 | | ReasoningEffort::Max), |
| 471 | ) => (Some(effort), false), |
| 472 | None | Some(ReasoningEffort::Auto) => (routed, true), |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | /// Route-aware equivalent of [`normalize_auto_route_effort_for_provider`]. |
| 477 | /// The inventory knows the selected provider/model, and the route resolver |
| 478 | /// supplies the endpoint needed to distinguish Kimi Code's official bare-K3 |
| 479 | /// contract from generic Moonshot. |
| 480 | #[must_use] |
| 481 | pub(crate) fn normalize_auto_route_effort_for_configured_route( |
| 482 | config: &Config, |
| 483 | provider: ApiProvider, |
| 484 | model: &str, |
| 485 | effort: ReasoningEffort, |
| 486 | ) -> ReasoningEffort { |
| 487 | crate::route_runtime::resolve_runtime_route(config, provider, Some(model)) |
| 488 | .map(|route| { |
| 489 | effort.normalize_for_route(provider, &route.candidate.endpoint().base_url, &route.model) |
| 490 | }) |
| 491 | .unwrap_or_else(|_| normalize_auto_route_effort_for_provider(provider, effort)) |
| 492 | } |
| 493 | |
| 494 | fn normalize_auto_route_selection_for_config( |
| 495 | config: &Config, |
| 496 | mut selection: AutoRouteSelection, |
| 497 | ) -> AutoRouteSelection { |
| 498 | selection.reasoning_effort = selection.reasoning_effort.map(|effort| { |
| 499 | normalize_auto_route_effort_for_configured_route( |
| 500 | config, |
| 501 | selection.provider, |
| 502 | &selection.model, |
| 503 | effort, |
| 504 | ) |
| 505 | }); |
| 506 | selection |
| 507 | } |
| 508 | |
| 509 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 510 | struct InventoryAutoRouteRecommendation { |
| 511 | provider: ApiProvider, |
| 512 | model: String, |
| 513 | reasoning_effort: Option<ReasoningEffort>, |
| 514 | } |
| 515 | |
| 516 | /// One provider-backed classifier attempt. A provider-success response always |
| 517 | /// reaches this shape before its content is interpreted, so invalid JSON and |
| 518 | /// provider-declared incomplete output retain their exact routed usage. |
| 519 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 520 | struct InventoryAutoRouteAttempt { |
| 521 | recommendation: Option<InventoryAutoRouteRecommendation>, |
| 522 | routed_usage: Vec<RuntimeUsageRecord>, |
| 523 | routed_usage_drop_records: Vec<RuntimeUsageDropRecord>, |
| 524 | routed_usage_dropped_records: u64, |
| 525 | } |
| 526 | |
| 527 | pub(crate) async fn resolve_auto_route_with_inventory( |
| 528 | config: &Config, |
| 529 | latest_request: &str, |
| 530 | recent_context: &str, |
| 531 | selected_model_mode: &str, |
| 532 | selected_thinking_mode: &str, |
| 533 | ) -> Result<AutoRouteSelection> { |
| 534 | resolve_auto_route_with_inventory_for_session( |
| 535 | config, |
| 536 | latest_request, |
| 537 | recent_context, |
| 538 | "agent", |
| 539 | selected_model_mode, |
| 540 | selected_thinking_mode, |
| 541 | ) |
| 542 | .await |
| 543 | } |
| 544 | |
| 545 | pub(crate) async fn resolve_auto_route_with_inventory_for_session( |
| 546 | config: &Config, |
| 547 | latest_request: &str, |
| 548 | recent_context: &str, |
| 549 | session_mode: &str, |
| 550 | selected_model_mode: &str, |
| 551 | selected_thinking_mode: &str, |
| 552 | ) -> Result<AutoRouteSelection> { |
| 553 | resolve_auto_route_with_inventory_for_session_and_cache_policy( |
| 554 | config, |
| 555 | latest_request, |
| 556 | recent_context, |
| 557 | session_mode, |
| 558 | selected_model_mode, |
| 559 | selected_thinking_mode, |
| 560 | true, |
| 561 | ) |
| 562 | .await |
| 563 | } |
| 564 | |
| 565 | pub(crate) async fn resolve_auto_route_with_inventory_for_session_and_cache_policy( |
| 566 | config: &Config, |
| 567 | latest_request: &str, |
| 568 | recent_context: &str, |
| 569 | session_mode: &str, |
| 570 | selected_model_mode: &str, |
| 571 | selected_thinking_mode: &str, |
| 572 | allow_response_cache: bool, |
| 573 | ) -> Result<AutoRouteSelection> { |
| 574 | let inventory = ModelInventory::from_config(config); |
| 575 | if !inventory.router_available { |
| 576 | // Fall back to declared-default auto routing when the flash router |
| 577 | // is unavailable (e.g. non-DeepSeek providers like wanjie-ark). |
| 578 | return Ok(normalize_auto_route_selection_for_config( |
| 579 | config, |
| 580 | auto_route_declared_fallback(config, &inventory), |
| 581 | )); |
| 582 | } |
| 583 | |
| 584 | let fallback = auto_route_declared_fallback(config, &inventory); |
| 585 | if cfg!(test) { |
| 586 | return Ok(normalize_auto_route_selection_for_config(config, fallback)); |
| 587 | } |
| 588 | |
| 589 | let selection = match auto_route_inventory_recommendation( |
| 590 | config, |
| 591 | &inventory, |
| 592 | latest_request, |
| 593 | recent_context, |
| 594 | session_mode, |
| 595 | selected_model_mode, |
| 596 | selected_thinking_mode, |
| 597 | allow_response_cache, |
| 598 | ) |
| 599 | .await |
| 600 | { |
| 601 | Ok(attempt) => auto_route_from_classifier_attempt(fallback, &inventory, attempt), |
| 602 | // Client construction/preparation failed before a provider request was |
| 603 | // admitted. There is no provider usage to invent and no dropped |
| 604 | // response receipt to claim. |
| 605 | Err(_) => auto_route_classifier_fallback(fallback, &inventory), |
| 606 | }; |
| 607 | Ok(normalize_auto_route_selection_for_config(config, selection)) |
| 608 | } |
| 609 | |
| 610 | pub(crate) fn resolve_explicit_route_with_inventory( |
| 611 | config: &Config, |
| 612 | requested_model: &str, |
| 613 | ) -> Option<AutoRouteSelection> { |
| 614 | let requested_model = requested_model.trim(); |
| 615 | if requested_model.is_empty() || requested_model.eq_ignore_ascii_case("auto") { |
| 616 | return None; |
| 617 | } |
| 618 | |
| 619 | let inventory = ModelInventory::from_config(config); |
| 620 | let active_provider = config.api_provider(); |
| 621 | |
| 622 | if let Some(candidate) = inventory.candidates.iter().find(|candidate| { |
| 623 | candidate.provider == active_provider |
| 624 | && explicit_model_matches_candidate(candidate, requested_model) |
| 625 | }) { |
| 626 | return Some(AutoRouteSelection { |
| 627 | provider: candidate.provider, |
| 628 | model: candidate.model.clone(), |
| 629 | reasoning_effort: config.reasoning_effort().map(|setting| { |
| 630 | normalize_auto_route_effort_for_configured_route( |
| 631 | config, |
| 632 | candidate.provider, |
| 633 | &candidate.model, |
| 634 | ReasoningEffort::from_setting(setting), |
| 635 | ) |
| 636 | }), |
| 637 | source: AutoRouteSource::Heuristic, |
| 638 | receipt: None, |
| 639 | routed_usage: Vec::new(), |
| 640 | routed_usage_drop_records: Vec::new(), |
| 641 | routed_usage_dropped_records: 0, |
| 642 | }); |
| 643 | } |
| 644 | |
| 645 | let mut matches = inventory |
| 646 | .candidates |
| 647 | .iter() |
| 648 | .filter(|candidate| explicit_model_matches_candidate(candidate, requested_model)); |
| 649 | let candidate = matches.next()?; |
| 650 | if matches.next().is_some() { |
| 651 | return None; |
| 652 | } |
| 653 | |
| 654 | Some(AutoRouteSelection { |
| 655 | provider: candidate.provider, |
| 656 | model: candidate.model.clone(), |
| 657 | reasoning_effort: config.reasoning_effort().map(|setting| { |
| 658 | normalize_auto_route_effort_for_configured_route( |
| 659 | config, |
| 660 | candidate.provider, |
| 661 | &candidate.model, |
| 662 | ReasoningEffort::from_setting(setting), |
| 663 | ) |
| 664 | }), |
| 665 | source: AutoRouteSource::Heuristic, |
| 666 | receipt: None, |
| 667 | routed_usage: Vec::new(), |
| 668 | routed_usage_drop_records: Vec::new(), |
| 669 | routed_usage_dropped_records: 0, |
| 670 | }) |
| 671 | } |
| 672 | |
| 673 | pub(crate) fn explicit_route_candidate_providers( |
| 674 | config: &Config, |
| 675 | requested_model: &str, |
| 676 | ) -> Vec<ApiProvider> { |
| 677 | let requested_model = requested_model.trim(); |
| 678 | if requested_model.is_empty() || requested_model.eq_ignore_ascii_case("auto") { |
| 679 | return Vec::new(); |
| 680 | } |
| 681 | |
| 682 | let inventory = ModelInventory::from_config(config); |
| 683 | let mut providers = Vec::new(); |
| 684 | for candidate in inventory |
| 685 | .candidates |
| 686 | .iter() |
| 687 | .filter(|candidate| explicit_model_matches_candidate(candidate, requested_model)) |
| 688 | { |
| 689 | if !providers.contains(&candidate.provider) { |
| 690 | providers.push(candidate.provider); |
| 691 | } |
| 692 | } |
| 693 | providers |
| 694 | } |
| 695 | |
| 696 | fn explicit_model_matches_candidate( |
| 697 | candidate: &crate::model_inventory::ModelRouteCandidate, |
| 698 | requested_model: &str, |
| 699 | ) -> bool { |
| 700 | candidate.model.eq_ignore_ascii_case(requested_model) |
| 701 | || normalize_model_name_for_provider(candidate.provider, requested_model) |
| 702 | .is_some_and(|model| candidate.model.eq_ignore_ascii_case(&model)) |
| 703 | } |
| 704 | |
| 705 | /// Declared local fallback for Auto routing when the flash classifier is |
| 706 | /// unavailable (or fails): the configured default model. |
| 707 | /// |
| 708 | /// There is no per-request signal here by design. Until the #6290 rework this |
| 709 | /// guessed cheap-vs-big from request wording (`COMPLEX_KEYWORDS` plus |
| 710 | /// char-length thresholds) — host-side semantic determinism that made cost |
| 711 | /// and quality depend on vocabulary. The only content-blind override is the |
| 712 | /// explicit `[auto] cost_saving` opt-in, which pins the runnable fast |
| 713 | /// sibling; providers without one stay on the default. |
| 714 | fn auto_route_declared_fallback(config: &Config, inventory: &ModelInventory) -> AutoRouteSelection { |
| 715 | let Some(active) = inventory.active_default() else { |
| 716 | let model = config.default_model(); |
| 717 | return AutoRouteSelection { |
| 718 | provider: config.api_provider(), |
| 719 | receipt: Some(auto_route_receipt( |
| 720 | inventory, |
| 721 | config.api_provider(), |
| 722 | &model, |
| 723 | AutoRouteScope::ResolvedProvider, |
| 724 | AutoRouteDataPath::LocalHeuristic, |
| 725 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::NoRunnableCandidate), |
| 726 | )), |
| 727 | model, |
| 728 | reasoning_effort: Some(crate::auto_reasoning::select()), |
| 729 | source: AutoRouteSource::Heuristic, |
| 730 | routed_usage: Vec::new(), |
| 731 | routed_usage_drop_records: Vec::new(), |
| 732 | routed_usage_dropped_records: 0, |
| 733 | }; |
| 734 | }; |
| 735 | let router_candidates = provider_router_candidates(active.provider, &active.model); |
| 736 | let runnable_fast = router_candidates.cheap.as_deref().filter(|model| { |
| 737 | inventory |
| 738 | .candidate(active.provider, model) |
| 739 | .is_some_and(|candidate| candidate.readiness.can_attempt()) |
| 740 | }); |
| 741 | let (model, reason) = if config.auto_cost_saving() { |
| 742 | match runnable_fast { |
| 743 | Some(cheap) => ( |
| 744 | cheap.to_string(), |
| 745 | AutoRouteHeuristicReason::CostSavingPolicy, |
| 746 | ), |
| 747 | None => ( |
| 748 | active.model.clone(), |
| 749 | AutoRouteHeuristicReason::NoFastSibling, |
| 750 | ), |
| 751 | } |
| 752 | } else { |
| 753 | ( |
| 754 | active.model.clone(), |
| 755 | AutoRouteHeuristicReason::DeclaredDefault, |
| 756 | ) |
| 757 | }; |
| 758 | AutoRouteSelection { |
| 759 | provider: active.provider, |
| 760 | receipt: Some(auto_route_receipt( |
| 761 | inventory, |
| 762 | active.provider, |
| 763 | &model, |
| 764 | AutoRouteScope::ResolvedProvider, |
| 765 | AutoRouteDataPath::LocalHeuristic, |
| 766 | AutoRouteReason::LocalFallback(reason), |
| 767 | )), |
| 768 | model, |
| 769 | reasoning_effort: Some(crate::auto_reasoning::select()), |
| 770 | source: AutoRouteSource::Heuristic, |
| 771 | routed_usage: Vec::new(), |
| 772 | routed_usage_drop_records: Vec::new(), |
| 773 | routed_usage_dropped_records: 0, |
| 774 | } |
| 775 | } |
| 776 | |
| 777 | fn auto_route_from_classifier( |
| 778 | inventory: &ModelInventory, |
| 779 | recommendation: InventoryAutoRouteRecommendation, |
| 780 | ) -> AutoRouteSelection { |
| 781 | let data_path = AutoRouteDataPath::Classifier { |
| 782 | provider: inventory.router_provider, |
| 783 | model: inventory.router_model.to_string(), |
| 784 | }; |
| 785 | // Report the scope the classifier actually had, not the widest one it |
| 786 | // could ever have (#4411). |
| 787 | let scope = if inventory.cross_provider_auto { |
| 788 | AutoRouteScope::RunnableProviders |
| 789 | } else { |
| 790 | AutoRouteScope::ActiveProvider |
| 791 | }; |
| 792 | AutoRouteSelection { |
| 793 | provider: recommendation.provider, |
| 794 | receipt: Some(auto_route_receipt( |
| 795 | inventory, |
| 796 | recommendation.provider, |
| 797 | &recommendation.model, |
| 798 | scope, |
| 799 | data_path, |
| 800 | AutoRouteReason::ClassifierRecommendation, |
| 801 | )), |
| 802 | model: recommendation.model, |
| 803 | reasoning_effort: recommendation.reasoning_effort, |
| 804 | source: AutoRouteSource::FlashRouter, |
| 805 | routed_usage: Vec::new(), |
| 806 | routed_usage_drop_records: Vec::new(), |
| 807 | routed_usage_dropped_records: 0, |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | fn auto_route_from_classifier_attempt( |
| 812 | fallback: AutoRouteSelection, |
| 813 | inventory: &ModelInventory, |
| 814 | attempt: InventoryAutoRouteAttempt, |
| 815 | ) -> AutoRouteSelection { |
| 816 | let InventoryAutoRouteAttempt { |
| 817 | recommendation, |
| 818 | routed_usage, |
| 819 | routed_usage_drop_records, |
| 820 | routed_usage_dropped_records, |
| 821 | } = attempt; |
| 822 | let mut selection = recommendation.map_or_else( |
| 823 | || auto_route_classifier_fallback(fallback, inventory), |
| 824 | |recommendation| auto_route_from_classifier(inventory, recommendation), |
| 825 | ); |
| 826 | selection.routed_usage = routed_usage; |
| 827 | selection.routed_usage_drop_records = routed_usage_drop_records; |
| 828 | selection.routed_usage_dropped_records = routed_usage_dropped_records; |
| 829 | selection |
| 830 | } |
| 831 | |
| 832 | fn auto_route_classifier_fallback( |
| 833 | mut fallback: AutoRouteSelection, |
| 834 | inventory: &ModelInventory, |
| 835 | ) -> AutoRouteSelection { |
| 836 | if let Some(receipt) = fallback.receipt.as_mut() { |
| 837 | let fallback_reason = match receipt.reason { |
| 838 | AutoRouteReason::LocalFallback(reason) |
| 839 | | AutoRouteReason::ClassifierFallback(reason) => reason, |
| 840 | AutoRouteReason::ClassifierRecommendation => AutoRouteHeuristicReason::DeclaredDefault, |
| 841 | }; |
| 842 | receipt.data_path = AutoRouteDataPath::Classifier { |
| 843 | provider: inventory.router_provider, |
| 844 | model: inventory.router_model.to_string(), |
| 845 | }; |
| 846 | receipt.reason = AutoRouteReason::ClassifierFallback(fallback_reason); |
| 847 | } |
| 848 | fallback |
| 849 | } |
| 850 | |
| 851 | fn auto_route_receipt( |
| 852 | inventory: &ModelInventory, |
| 853 | provider: ApiProvider, |
| 854 | selected_model: &str, |
| 855 | scope: AutoRouteScope, |
| 856 | data_path: AutoRouteDataPath, |
| 857 | reason: AutoRouteReason, |
| 858 | ) -> AutoRouteReceipt { |
| 859 | let pair = auto_route_pair(inventory, provider, selected_model); |
| 860 | let tier = if pair |
| 861 | .fast |
| 862 | .as_deref() |
| 863 | .is_some_and(|fast| fast.eq_ignore_ascii_case(selected_model)) |
| 864 | { |
| 865 | AutoRouteTier::Fast |
| 866 | } else if pair.strong.eq_ignore_ascii_case(selected_model) { |
| 867 | if pair.fast.is_some() { |
| 868 | AutoRouteTier::Strong |
| 869 | } else { |
| 870 | AutoRouteTier::Only |
| 871 | } |
| 872 | } else { |
| 873 | AutoRouteTier::Selected |
| 874 | }; |
| 875 | AutoRouteReceipt { |
| 876 | tier, |
| 877 | pair, |
| 878 | scope, |
| 879 | data_path, |
| 880 | reason, |
| 881 | } |
| 882 | } |
| 883 | |
| 884 | fn auto_route_pair( |
| 885 | inventory: &ModelInventory, |
| 886 | provider: ApiProvider, |
| 887 | selected_model: &str, |
| 888 | ) -> AutoRoutePair { |
| 889 | // A provider can expose several unrelated model families. Derive the pair |
| 890 | // from a runnable candidate that actually contains the selected model, |
| 891 | // preferring a cheap-tier match before a strong-tier match. Falling back |
| 892 | // to the provider default would report a truthful provider with a false |
| 893 | // model family (for example OpenRouter GLM reported as DeepSeek). |
| 894 | let matching_pair = inventory |
| 895 | .candidates |
| 896 | .iter() |
| 897 | .filter(|candidate| candidate.provider == provider && candidate.readiness.can_attempt()) |
| 898 | .map(|candidate| provider_router_candidates(provider, &candidate.model)) |
| 899 | .find(|pair| { |
| 900 | pair.cheap |
| 901 | .as_deref() |
| 902 | .is_some_and(|fast| fast.eq_ignore_ascii_case(selected_model)) |
| 903 | }) |
| 904 | .or_else(|| { |
| 905 | inventory |
| 906 | .candidates |
| 907 | .iter() |
| 908 | .filter(|candidate| { |
| 909 | candidate.provider == provider && candidate.readiness.can_attempt() |
| 910 | }) |
| 911 | .map(|candidate| provider_router_candidates(provider, &candidate.model)) |
| 912 | .find(|pair| pair.big.eq_ignore_ascii_case(selected_model)) |
| 913 | }); |
| 914 | let Some(candidates) = matching_pair else { |
| 915 | return AutoRoutePair { |
| 916 | strong: selected_model.to_string(), |
| 917 | fast: None, |
| 918 | }; |
| 919 | }; |
| 920 | let Some(strong) = inventory |
| 921 | .candidate(provider, &candidates.big) |
| 922 | .filter(|candidate| candidate.readiness.can_attempt()) |
| 923 | .map(|candidate| candidate.model.clone()) |
| 924 | else { |
| 925 | return AutoRoutePair { |
| 926 | strong: selected_model.to_string(), |
| 927 | fast: None, |
| 928 | }; |
| 929 | }; |
| 930 | let fast = candidates.cheap.as_deref().and_then(|model| { |
| 931 | inventory |
| 932 | .candidate(provider, model) |
| 933 | .filter(|candidate| candidate.readiness.can_attempt()) |
| 934 | .map(|candidate| candidate.model.clone()) |
| 935 | }); |
| 936 | AutoRoutePair { strong, fast } |
| 937 | } |
| 938 | |
| 939 | fn auto_route_usage_has_reported_data(usage: &codewhale_models::Usage) -> bool { |
| 940 | usage.input_tokens > 0 |
| 941 | || usage.output_tokens > 0 |
| 942 | || usage.prompt_cache_hit_tokens.is_some() |
| 943 | || usage.prompt_cache_miss_tokens.is_some() |
| 944 | || usage.prompt_cache_write_tokens.is_some() |
| 945 | || usage.reasoning_tokens.is_some() |
| 946 | || usage.reasoning_replay_tokens.is_some() |
| 947 | || usage.server_tool_use.is_some() |
| 948 | } |
| 949 | |
| 950 | /// Stable, persistence-safe identity for one classifier response. The raw |
| 951 | /// provider response id is hashed with the frozen dispatch route and instant; |
| 952 | /// neither it nor any custom route label crosses into telemetry/persistence. |
| 953 | fn auto_route_usage_source_id(route: &EffectiveRouteEnvelope, response_id: &str) -> String { |
| 954 | use sha2::{Digest as _, Sha256}; |
| 955 | |
| 956 | let mut digest = Sha256::new(); |
| 957 | let dispatched_at = route.dispatched_at.to_rfc3339(); |
| 958 | for part in [ |
| 959 | b"codewhale:auto-route-classifier:v1".as_slice(), |
| 960 | route.provider.as_str().as_bytes(), |
| 961 | route.provider_identity.as_bytes(), |
| 962 | route.model.as_bytes(), |
| 963 | route |
| 964 | .endpoint_fingerprint |
| 965 | .as_deref() |
| 966 | .unwrap_or_default() |
| 967 | .as_bytes(), |
| 968 | dispatched_at.as_bytes(), |
| 969 | response_id.as_bytes(), |
| 970 | ] { |
| 971 | digest.update((part.len() as u64).to_le_bytes()); |
| 972 | digest.update(part); |
| 973 | } |
| 974 | format!( |
| 975 | "auto-router:{}", |
| 976 | crate::hashing::hex_bytes(digest.finalize()) |
| 977 | ) |
| 978 | } |
| 979 | |
| 980 | fn auto_route_attempt_from_response( |
| 981 | request_route: EffectiveRouteEnvelope, |
| 982 | response: &MessageResponse, |
| 983 | inventory: &ModelInventory, |
| 984 | ) -> InventoryAutoRouteAttempt { |
| 985 | // All-zero usage cannot price a routed segment. The dispatch caller owns |
| 986 | // the stronger cache/provenance context and must explicitly classify this |
| 987 | // as either a proven cache replay or missing provider billing evidence. |
| 988 | let routed_usage = auto_route_usage_has_reported_data(&response.usage) |
| 989 | .then(|| RuntimeUsageRecord { |
| 990 | source_id: auto_route_usage_source_id(&request_route, &response.id), |
| 991 | usage: EffectiveRouteUsage { |
| 992 | route: request_route.sanitized_for_persistence(), |
| 993 | usage: response.usage.clone(), |
| 994 | }, |
| 995 | }) |
| 996 | .into_iter() |
| 997 | .collect(); |
| 998 | let recommendation = |
| 999 | (!codewhale_models::is_incomplete_stop_reason(response.stop_reason.as_deref())) |
| 1000 | .then(|| { |
| 1001 | parse_inventory_auto_route_recommendation( |
| 1002 | &message_response_text(response), |
| 1003 | inventory, |
| 1004 | ) |
| 1005 | }) |
| 1006 | .flatten(); |
| 1007 | InventoryAutoRouteAttempt { |
| 1008 | recommendation, |
| 1009 | routed_usage, |
| 1010 | routed_usage_drop_records: Vec::new(), |
| 1011 | routed_usage_dropped_records: 0, |
| 1012 | } |
| 1013 | } |
| 1014 | |
| 1015 | fn auto_route_attempt_from_provider_response( |
| 1016 | request_route: EffectiveRouteEnvelope, |
| 1017 | response: &MessageResponse, |
| 1018 | inventory: &ModelInventory, |
| 1019 | ) -> InventoryAutoRouteAttempt { |
| 1020 | let drop_route = request_route.sanitized_for_persistence(); |
| 1021 | let mut attempt = auto_route_attempt_from_response(request_route, response, inventory); |
| 1022 | // This classifier request is currently not cacheable (temperature is |
| 1023 | // provider-default, not the deterministic Some(0.0) cache contract), so a |
| 1024 | // decoded all-zero response is missing provider billing evidence even when |
| 1025 | // the caller permits response-cache use. Do not silently reinterpret the |
| 1026 | // policy boolean as cache-hit provenance. |
| 1027 | if attempt.routed_usage.is_empty() { |
| 1028 | attempt.routed_usage_drop_records = vec![RuntimeUsageDropRecord { |
| 1029 | source_id: auto_route_usage_source_id( |
| 1030 | &drop_route, |
| 1031 | &format!("missing-usage:{}", response.id), |
| 1032 | ), |
| 1033 | route: drop_route, |
| 1034 | }]; |
| 1035 | attempt.routed_usage_dropped_records = 1; |
| 1036 | } |
| 1037 | attempt |
| 1038 | } |
| 1039 | |
| 1040 | fn auto_route_attempt_with_dropped_response( |
| 1041 | request_route: EffectiveRouteEnvelope, |
| 1042 | ) -> InventoryAutoRouteAttempt { |
| 1043 | let request_route = request_route.sanitized_for_persistence(); |
| 1044 | InventoryAutoRouteAttempt { |
| 1045 | recommendation: None, |
| 1046 | routed_usage: Vec::new(), |
| 1047 | routed_usage_drop_records: vec![RuntimeUsageDropRecord { |
| 1048 | source_id: auto_route_usage_source_id(&request_route, "transport-error"), |
| 1049 | route: request_route, |
| 1050 | }], |
| 1051 | routed_usage_dropped_records: 1, |
| 1052 | } |
| 1053 | } |
| 1054 | |
| 1055 | /// Prove that the deterministic request seam accepts this classifier request |
| 1056 | /// before capturing a quote or entering any provider permit/network path. |
| 1057 | fn preflight_auto_route_request(client: &CodewhaleClient, request: &MessageRequest) -> Result<()> { |
| 1058 | client.prepare_outbound_request(request.clone(), false)?; |
| 1059 | Ok(()) |
| 1060 | } |
| 1061 | |
| 1062 | #[allow(clippy::too_many_arguments)] |
| 1063 | async fn auto_route_inventory_recommendation( |
| 1064 | config: &Config, |
| 1065 | inventory: &ModelInventory, |
| 1066 | latest_request: &str, |
| 1067 | recent_context: &str, |
| 1068 | session_mode: &str, |
| 1069 | selected_model_mode: &str, |
| 1070 | selected_thinking_mode: &str, |
| 1071 | allow_response_cache: bool, |
| 1072 | ) -> Result<InventoryAutoRouteAttempt> { |
| 1073 | let mut router_config = config.clone(); |
| 1074 | // The classifier runs on the inventory's router route: the explicit |
| 1075 | // [auto.router] route when configured, else the DeepSeek flash default. |
| 1076 | router_config.provider = Some(inventory.router_provider.as_str().to_string()); |
| 1077 | router_config.default_text_model = Some(inventory.router_model.clone()); |
| 1078 | |
| 1079 | let client = CodewhaleClient::new(&router_config)?; |
| 1080 | let router_system = inventory_auto_router_system_prompt(inventory, config.auto_cost_saving()); |
| 1081 | let router_prompt = classifier_prompt( |
| 1082 | &client, |
| 1083 | latest_request, |
| 1084 | recent_context, |
| 1085 | session_mode, |
| 1086 | selected_model_mode, |
| 1087 | selected_thinking_mode, |
| 1088 | ); |
| 1089 | let max_tokens = client.effective_max_output_tokens(&inventory.router_model); |
| 1090 | let request = MessageRequest { |
| 1091 | model: inventory.router_model.to_string(), |
| 1092 | messages: vec![Message { |
| 1093 | role: Role::User, |
| 1094 | content: vec![ContentBlock::Text { |
| 1095 | text: router_prompt, |
| 1096 | cache_control: None, |
| 1097 | }], |
| 1098 | }], |
| 1099 | max_tokens, |
| 1100 | system: Some(SystemPrompt::Text(router_system)), |
| 1101 | tools: None, |
| 1102 | tool_choice: None, |
| 1103 | metadata: None, |
| 1104 | thinking: None, |
| 1105 | reasoning_effort: Some( |
| 1106 | inventory |
| 1107 | .router_thinking |
| 1108 | .clone() |
| 1109 | .unwrap_or_else(|| "off".to_string()), |
| 1110 | ), |
| 1111 | stream: Some(false), |
| 1112 | temperature: None, |
| 1113 | top_p: None, |
| 1114 | }; |
| 1115 | |
| 1116 | // Freeze pricing at the last application seam before the provider future |
| 1117 | // starts. Prompt shaping above may be slow and may overlap a catalog |
| 1118 | // refresh; completion-time mutable catalog state must never reprice this |
| 1119 | // already-admitted classifier request. |
| 1120 | preflight_auto_route_request(&client, &request)?; |
| 1121 | let request_route = |
| 1122 | client.effective_route_envelope(&inventory.router_model, chrono::Utc::now()); |
| 1123 | let response = if allow_response_cache { |
| 1124 | tokio::time::timeout( |
| 1125 | Duration::from_secs(inventory.router_timeout_secs), |
| 1126 | client.create_message(request), |
| 1127 | ) |
| 1128 | .await |
| 1129 | } else { |
| 1130 | tokio::time::timeout( |
| 1131 | Duration::from_secs(inventory.router_timeout_secs), |
| 1132 | client.create_message_without_response_cache(request), |
| 1133 | ) |
| 1134 | .await |
| 1135 | }; |
| 1136 | let response = match response { |
| 1137 | Ok(Ok(response)) => response, |
| 1138 | // The request crossed Codewhale's dispatch boundary, but no exact |
| 1139 | // provider usage came back. Preserve the fallback while explicitly |
| 1140 | // failing cost coverage closed. |
| 1141 | Ok(Err(_)) => return Ok(auto_route_attempt_with_dropped_response(request_route)), |
| 1142 | // The local deadline cancels the future and can fire while the request |
| 1143 | // is still waiting on an application/provider permit. With no response |
| 1144 | // evidence we must not invent a provider call or a missing-usage |
| 1145 | // receipt. Transport errors returned by the client remain the |
| 1146 | // conservative explicit-dropped path above. |
| 1147 | Err(_) => { |
| 1148 | return Ok(InventoryAutoRouteAttempt { |
| 1149 | recommendation: None, |
| 1150 | routed_usage: Vec::new(), |
| 1151 | routed_usage_drop_records: Vec::new(), |
| 1152 | routed_usage_dropped_records: 0, |
| 1153 | }); |
| 1154 | } |
| 1155 | }; |
| 1156 | Ok(auto_route_attempt_from_provider_response( |
| 1157 | request_route, |
| 1158 | &response, |
| 1159 | inventory, |
| 1160 | )) |
| 1161 | } |
| 1162 | |
| 1163 | fn inventory_auto_router_system_prompt(inventory: &ModelInventory, cost_saving: bool) -> String { |
| 1164 | let mut prompt = if inventory.cross_provider_auto { |
| 1165 | String::new() |
| 1166 | } else { |
| 1167 | // The inventory JSON below is already scoped to the active provider |
| 1168 | // (#4411); say so, so the classifier does not try to name one it was |
| 1169 | // never shown. |
| 1170 | format!( |
| 1171 | "Auto routing is scoped to the active provider `{}`. Every model in the inventory \ |
| 1172 | below belongs to it; never select another provider.\n\n", |
| 1173 | inventory.active_provider.as_str() |
| 1174 | ) |
| 1175 | }; |
| 1176 | prompt.push_str(&format!( |
| 1177 | "You are the codewhale model-routing classifier. Return only compact JSON: \ |
| 1178 | {{\"provider\":\"<provider>\",\"model\":\"<model>\",\"thinking\":\"off|high|max\"}}.\n\ |
| 1179 | Choose only provider/model pairs present in the inventory JSON. Use off only for trivial no-tool answers, \ |
| 1180 | high for ordinary reasoning, and max for agentic, coding, multi-file, release, architecture, debugging, \ |
| 1181 | security, tool-heavy, or uncertain work.\n\nInventory JSON:\n{}", |
| 1182 | inventory.router_context_json() |
| 1183 | )); |
| 1184 | |
| 1185 | if cost_saving { |
| 1186 | let active_pair = inventory.active_default().and_then(|active| { |
| 1187 | let candidates = provider_router_candidates(active.provider, &active.model); |
| 1188 | let fast = candidates.cheap.as_deref()?; |
| 1189 | (inventory |
| 1190 | .candidate(active.provider, &candidates.big) |
| 1191 | .is_some_and(|candidate| candidate.readiness.can_attempt()) |
| 1192 | && inventory |
| 1193 | .candidate(active.provider, fast) |
| 1194 | .is_some_and(|candidate| candidate.readiness.can_attempt())) |
| 1195 | .then_some((active.provider, candidates.big, fast.to_string())) |
| 1196 | }); |
| 1197 | |
| 1198 | if let Some((provider, strong, fast)) = active_pair { |
| 1199 | prompt.push_str(&format!( |
| 1200 | "\n\nCost-saving mode is ON. For the active provider `{}`, `{fast}` is the fast tier \ |
| 1201 | and `{strong}` is the strong tier. Prefer `{fast}` for ambiguous, routine, or single-step work. \ |
| 1202 | Select `{strong}` only when the request is unmistakably agentic, multi-step, architecture/design, \ |
| 1203 | security review, debugging, or otherwise clearly beyond the fast tier. Keep the selected model paired \ |
| 1204 | with provider `{}`.", |
| 1205 | provider.as_str(), |
| 1206 | provider.as_str() |
| 1207 | )); |
| 1208 | } else { |
| 1209 | prompt.push_str( |
| 1210 | "\n\nCost-saving mode is ON, but the active provider has no known runnable fast sibling. \ |
| 1211 | Do not invent a model or cross-provider downgrade solely to save cost.", |
| 1212 | ); |
| 1213 | } |
| 1214 | } |
| 1215 | |
| 1216 | prompt |
| 1217 | } |
| 1218 | |
| 1219 | fn parse_inventory_auto_route_recommendation( |
| 1220 | raw: &str, |
| 1221 | inventory: &ModelInventory, |
| 1222 | ) -> Option<InventoryAutoRouteRecommendation> { |
| 1223 | let json = extract_first_json_object(raw)?; |
| 1224 | let value: serde_json::Value = serde_json::from_str(json).ok()?; |
| 1225 | let provider = value |
| 1226 | .get("provider") |
| 1227 | .and_then(serde_json::Value::as_str) |
| 1228 | .and_then(ApiProvider::parse)?; |
| 1229 | // Defense in depth for #4411: the payload already hides other providers, |
| 1230 | // but a hallucinated (or stale) provider must not become a route either. |
| 1231 | if !inventory.auto_scope_allows(provider) { |
| 1232 | return None; |
| 1233 | } |
| 1234 | let model = value.get("model").and_then(serde_json::Value::as_str)?; |
| 1235 | let candidate = inventory |
| 1236 | .candidate(provider, model) |
| 1237 | .filter(|candidate| candidate.readiness.can_attempt())?; |
| 1238 | let reasoning_effort = value |
| 1239 | .get("thinking") |
| 1240 | .or_else(|| value.get("reasoning_effort")) |
| 1241 | .or_else(|| value.get("effort")) |
| 1242 | .and_then(serde_json::Value::as_str) |
| 1243 | .and_then(parse_auto_route_reasoning_effort); |
| 1244 | |
| 1245 | Some(InventoryAutoRouteRecommendation { |
| 1246 | provider, |
| 1247 | model: candidate.model.clone(), |
| 1248 | reasoning_effort, |
| 1249 | }) |
| 1250 | } |
| 1251 | |
| 1252 | fn auto_route_prompt( |
| 1253 | latest_request: &str, |
| 1254 | recent_context: &str, |
| 1255 | session_mode: &str, |
| 1256 | selected_model_mode: &str, |
| 1257 | selected_thinking_mode: &str, |
| 1258 | ) -> String { |
| 1259 | format!( |
| 1260 | "Session mode: {}\nSelected model mode: {}\nSelected thinking mode: {}\n\nRecent context:\n{}\n\nLatest user request:\n{}\n\nReturn JSON only.", |
| 1261 | session_mode, |
| 1262 | selected_model_mode, |
| 1263 | selected_thinking_mode, |
| 1264 | if recent_context.trim().is_empty() { |
| 1265 | "No prior context." |
| 1266 | } else { |
| 1267 | recent_context |
| 1268 | }, |
| 1269 | truncate_for_auto_router(latest_request, 4_000) |
| 1270 | ) |
| 1271 | } |
| 1272 | |
| 1273 | fn classifier_prompt( |
| 1274 | client: &CodewhaleClient, |
| 1275 | latest_request: &str, |
| 1276 | recent_context: &str, |
| 1277 | session_mode: &str, |
| 1278 | selected_model_mode: &str, |
| 1279 | selected_thinking_mode: &str, |
| 1280 | ) -> String { |
| 1281 | client.redact_model_bound_text(&auto_route_prompt( |
| 1282 | latest_request, |
| 1283 | recent_context, |
| 1284 | session_mode, |
| 1285 | selected_model_mode, |
| 1286 | selected_thinking_mode, |
| 1287 | )) |
| 1288 | } |
| 1289 | |
| 1290 | fn message_response_text(response: &MessageResponse) -> String { |
| 1291 | let mut out = String::new(); |
| 1292 | for block in &response.content { |
| 1293 | match block { |
| 1294 | ContentBlock::Text { text, .. } | ContentBlock::ToolResult { content: text, .. } => { |
| 1295 | append_router_text(&mut out, text); |
| 1296 | } |
| 1297 | ContentBlock::Thinking { thinking, .. } => { |
| 1298 | append_router_text(&mut out, thinking); |
| 1299 | } |
| 1300 | ContentBlock::ToolUse { name, .. } => { |
| 1301 | append_router_text(&mut out, &format!("[tool call: {name}]")); |
| 1302 | } |
| 1303 | _ => {} |
| 1304 | } |
| 1305 | } |
| 1306 | out |
| 1307 | } |
| 1308 | |
| 1309 | fn append_router_text(out: &mut String, text: &str) { |
| 1310 | if !out.is_empty() { |
| 1311 | out.push('\n'); |
| 1312 | } |
| 1313 | out.push_str(text); |
| 1314 | } |
| 1315 | |
| 1316 | fn truncate_for_auto_router(text: &str, max_chars: usize) -> String { |
| 1317 | let mut chars = text.chars(); |
| 1318 | let truncated: String = chars.by_ref().take(max_chars).collect(); |
| 1319 | if chars.next().is_some() { |
| 1320 | format!("{truncated}...") |
| 1321 | } else { |
| 1322 | truncated |
| 1323 | } |
| 1324 | } |
| 1325 | |
| 1326 | #[cfg(test)] |
| 1327 | mod tests { |
| 1328 | use super::*; |
| 1329 | |
| 1330 | struct ProviderCatalogReset; |
| 1331 | |
| 1332 | impl Drop for ProviderCatalogReset { |
| 1333 | fn drop(&mut self) { |
| 1334 | crate::provider_catalog_live::reset_cache_for_test(); |
| 1335 | crate::provider_lake::clear_live_snapshot(); |
| 1336 | } |
| 1337 | } |
| 1338 | |
| 1339 | fn priced_openrouter_delta( |
| 1340 | model: &str, |
| 1341 | fingerprint: &str, |
| 1342 | fetched_at: u64, |
| 1343 | input: f64, |
| 1344 | output: f64, |
| 1345 | ) -> codewhale_config::catalog::ProviderCatalogDelta { |
| 1346 | use codewhale_config::catalog::{CatalogOffering, CatalogSource, ProviderCatalogDelta}; |
| 1347 | |
| 1348 | ProviderCatalogDelta { |
| 1349 | provider: ApiProvider::Openrouter.as_str().to_string(), |
| 1350 | base_url_fingerprint: fingerprint.to_string(), |
| 1351 | fetched_at, |
| 1352 | offerings: vec![CatalogOffering { |
| 1353 | provider: ApiProvider::Openrouter.as_str().to_string(), |
| 1354 | wire_model_id: model.to_string(), |
| 1355 | endpoint_key: "chat".to_string(), |
| 1356 | source: CatalogSource::Live { |
| 1357 | base_url_fingerprint: fingerprint.to_string(), |
| 1358 | fetched_at, |
| 1359 | }, |
| 1360 | cost: Some(codewhale_config::models_dev::ModelsDevCost { |
| 1361 | input: Some(input), |
| 1362 | output: Some(output), |
| 1363 | cache_read: Some(input / 2.0), |
| 1364 | cache_write: None, |
| 1365 | }), |
| 1366 | ..CatalogOffering::default() |
| 1367 | }], |
| 1368 | } |
| 1369 | } |
| 1370 | |
| 1371 | fn classifier_response( |
| 1372 | id: &str, |
| 1373 | text: &str, |
| 1374 | stop_reason: &str, |
| 1375 | usage: codewhale_models::Usage, |
| 1376 | ) -> MessageResponse { |
| 1377 | MessageResponse { |
| 1378 | id: id.to_string(), |
| 1379 | r#type: "message".to_string(), |
| 1380 | role: "assistant".to_string(), |
| 1381 | content: vec![ContentBlock::Text { |
| 1382 | text: text.to_string(), |
| 1383 | cache_control: None, |
| 1384 | }], |
| 1385 | model: "router-response-alias-must-not-price".to_string(), |
| 1386 | stop_reason: Some(stop_reason.to_string()), |
| 1387 | stop_sequence: None, |
| 1388 | container: None, |
| 1389 | usage, |
| 1390 | } |
| 1391 | } |
| 1392 | |
| 1393 | #[test] |
| 1394 | fn classifier_semantic_fallbacks_keep_exact_quotes_and_replay_once() { |
| 1395 | let _env_lock = crate::test_support::lock_test_env(); |
| 1396 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 1397 | let home = tempfile::tempdir().expect("test home"); |
| 1398 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 1399 | let _reset = ProviderCatalogReset; |
| 1400 | crate::provider_catalog_live::reset_cache_for_test(); |
| 1401 | crate::provider_lake::clear_live_snapshot(); |
| 1402 | |
| 1403 | let model = "synthetic/openrouter-auto-classifier"; |
| 1404 | let config = Config { |
| 1405 | provider: Some("openrouter".to_string()), |
| 1406 | providers: Some(crate::config::ProvidersConfig { |
| 1407 | openrouter: crate::config::ProviderConfig { |
| 1408 | api_key: Some("test-openrouter-key".to_string()), |
| 1409 | base_url: Some(crate::config::DEFAULT_OPENROUTER_BASE_URL.to_string()), |
| 1410 | model: Some(model.to_string()), |
| 1411 | ..Default::default() |
| 1412 | }, |
| 1413 | ..Default::default() |
| 1414 | }), |
| 1415 | auto: Some(crate::config::AutoConfig { |
| 1416 | cost_saving: None, |
| 1417 | cross_provider: None, |
| 1418 | router: Some(crate::config::AutoRouterConfig { |
| 1419 | provider: Some("openrouter".to_string()), |
| 1420 | model: Some(model.to_string()), |
| 1421 | thinking: Some("off".to_string()), |
| 1422 | timeout_secs: None, |
| 1423 | }), |
| 1424 | }), |
| 1425 | ..Default::default() |
| 1426 | }; |
| 1427 | let inventory = ModelInventory::from_config(&config); |
| 1428 | assert!( |
| 1429 | inventory |
| 1430 | .candidate(ApiProvider::Openrouter, model) |
| 1431 | .is_some() |
| 1432 | ); |
| 1433 | let client = CodewhaleClient::new(&config).expect("OpenRouter classifier client"); |
| 1434 | let fingerprint = codewhale_config::catalog::base_url_fingerprint( |
| 1435 | crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 1436 | ); |
| 1437 | let first_at = chrono::Utc::now(); |
| 1438 | let fetched_at = u64::try_from(first_at.timestamp()).expect("nonnegative timestamp"); |
| 1439 | |
| 1440 | crate::provider_catalog_live::record_success(priced_openrouter_delta( |
| 1441 | model, |
| 1442 | &fingerprint, |
| 1443 | fetched_at, |
| 1444 | 1.0, |
| 1445 | 4.0, |
| 1446 | )); |
| 1447 | let first_route = client.effective_route_envelope(model, first_at); |
| 1448 | let valid = auto_route_attempt_from_response( |
| 1449 | first_route, |
| 1450 | &classifier_response( |
| 1451 | "same-provider-response-id", |
| 1452 | &format!(r#"{{"provider":"openrouter","model":"{model}","thinking":"off"}}"#), |
| 1453 | "stop", |
| 1454 | codewhale_models::Usage { |
| 1455 | input_tokens: 10, |
| 1456 | output_tokens: 2, |
| 1457 | prompt_cache_hit_tokens: Some(3), |
| 1458 | ..Default::default() |
| 1459 | }, |
| 1460 | ), |
| 1461 | &inventory, |
| 1462 | ); |
| 1463 | |
| 1464 | // Replace the live row in the same Unix second. The first routed |
| 1465 | // record must keep its old immutable revision and the new attempt must |
| 1466 | // freeze a distinct one at its own dispatch boundary. |
| 1467 | crate::provider_catalog_live::record_success(priced_openrouter_delta( |
| 1468 | model, |
| 1469 | &fingerprint, |
| 1470 | fetched_at, |
| 1471 | 9.0, |
| 1472 | 19.0, |
| 1473 | )); |
| 1474 | let second_at = first_at + chrono::Duration::nanoseconds(1); |
| 1475 | let invalid = auto_route_attempt_from_response( |
| 1476 | client.effective_route_envelope(model, second_at), |
| 1477 | &classifier_response( |
| 1478 | "same-provider-response-id", |
| 1479 | "not valid route json", |
| 1480 | "stop", |
| 1481 | codewhale_models::Usage { |
| 1482 | input_tokens: 11, |
| 1483 | output_tokens: 3, |
| 1484 | ..Default::default() |
| 1485 | }, |
| 1486 | ), |
| 1487 | &inventory, |
| 1488 | ); |
| 1489 | let incomplete = auto_route_attempt_from_response( |
| 1490 | client.effective_route_envelope(model, second_at + chrono::Duration::nanoseconds(1)), |
| 1491 | &classifier_response( |
| 1492 | "same-provider-response-id", |
| 1493 | &format!(r#"{{"provider":"openrouter","model":"{model}"}}"#), |
| 1494 | "length", |
| 1495 | codewhale_models::Usage { |
| 1496 | input_tokens: 12, |
| 1497 | output_tokens: 4, |
| 1498 | ..Default::default() |
| 1499 | }, |
| 1500 | ), |
| 1501 | &inventory, |
| 1502 | ); |
| 1503 | let missing_usage_route = |
| 1504 | client.effective_route_envelope(model, second_at + chrono::Duration::nanoseconds(2)); |
| 1505 | let missing_usage = auto_route_attempt_from_provider_response( |
| 1506 | missing_usage_route.clone(), |
| 1507 | &classifier_response( |
| 1508 | "missing-usage-response-id", |
| 1509 | "not valid route json", |
| 1510 | "stop", |
| 1511 | codewhale_models::Usage::default(), |
| 1512 | ), |
| 1513 | &inventory, |
| 1514 | ); |
| 1515 | assert!(missing_usage.routed_usage.is_empty()); |
| 1516 | assert_eq!(missing_usage.routed_usage_dropped_records, 1); |
| 1517 | assert_eq!(missing_usage.routed_usage_drop_records.len(), 1); |
| 1518 | assert_eq!( |
| 1519 | missing_usage.routed_usage_drop_records[0].route, |
| 1520 | missing_usage_route.sanitized_for_persistence() |
| 1521 | ); |
| 1522 | assert!( |
| 1523 | missing_usage.routed_usage_drop_records[0] |
| 1524 | .source_id |
| 1525 | .starts_with("auto-router:") |
| 1526 | ); |
| 1527 | assert!( |
| 1528 | !missing_usage.routed_usage_drop_records[0] |
| 1529 | .source_id |
| 1530 | .contains("missing-usage-response-id") |
| 1531 | ); |
| 1532 | |
| 1533 | let transport = auto_route_attempt_with_dropped_response( |
| 1534 | client.effective_route_envelope(model, second_at + chrono::Duration::nanoseconds(3)), |
| 1535 | ); |
| 1536 | assert_eq!(transport.routed_usage_dropped_records, 1); |
| 1537 | assert_eq!(transport.routed_usage_drop_records.len(), 1); |
| 1538 | assert!(transport.routed_usage.is_empty()); |
| 1539 | |
| 1540 | let fallback = auto_route_declared_fallback(&config, &inventory); |
| 1541 | let valid = auto_route_from_classifier_attempt(fallback.clone(), &inventory, valid); |
| 1542 | let invalid = auto_route_from_classifier_attempt(fallback.clone(), &inventory, invalid); |
| 1543 | let incomplete = auto_route_from_classifier_attempt(fallback, &inventory, incomplete); |
| 1544 | assert_eq!(valid.source, AutoRouteSource::FlashRouter); |
| 1545 | for fallback in [&invalid, &incomplete] { |
| 1546 | assert_eq!(fallback.source, AutoRouteSource::Heuristic); |
| 1547 | assert!(matches!( |
| 1548 | fallback.receipt.as_ref().map(|receipt| receipt.reason), |
| 1549 | Some(AutoRouteReason::ClassifierFallback(_)) |
| 1550 | )); |
| 1551 | assert_eq!(fallback.routed_usage.len(), 1); |
| 1552 | assert_eq!(fallback.routed_usage_dropped_records, 0); |
| 1553 | } |
| 1554 | assert_eq!(valid.routed_usage.len(), 1); |
| 1555 | assert_eq!(valid.routed_usage[0].usage.usage.input_tokens, 10); |
| 1556 | assert_eq!( |
| 1557 | valid.routed_usage[0].usage.usage.prompt_cache_hit_tokens, |
| 1558 | Some(3) |
| 1559 | ); |
| 1560 | |
| 1561 | let first_quote = valid.routed_usage[0] |
| 1562 | .usage |
| 1563 | .route |
| 1564 | .provider_live_pricing |
| 1565 | .as_ref() |
| 1566 | .expect("first exact quote"); |
| 1567 | let second_quote = invalid.routed_usage[0] |
| 1568 | .usage |
| 1569 | .route |
| 1570 | .provider_live_pricing |
| 1571 | .as_ref() |
| 1572 | .expect("replacement exact quote"); |
| 1573 | assert_ne!(first_quote.catalog_revision, second_quote.catalog_revision); |
| 1574 | assert_eq!(first_quote.input_per_million.as_deref(), Some("1")); |
| 1575 | assert_eq!(second_quote.input_per_million.as_deref(), Some("9")); |
| 1576 | |
| 1577 | let records = valid |
| 1578 | .routed_usage |
| 1579 | .iter() |
| 1580 | .chain(&invalid.routed_usage) |
| 1581 | .chain(&incomplete.routed_usage) |
| 1582 | .cloned() |
| 1583 | .collect::<Vec<_>>(); |
| 1584 | assert_eq!(records.len(), 3); |
| 1585 | assert!(records.iter().all(|record| { |
| 1586 | record.source_id.starts_with("auto-router:") |
| 1587 | && record.source_id.len() == "auto-router:".len() + 64 |
| 1588 | && !record.source_id.contains("same-provider-response-id") |
| 1589 | })); |
| 1590 | |
| 1591 | // Exercise the canonical sink exactly as selection consumers do: |
| 1592 | // replaying any record cannot add parent-route spend or a second |
| 1593 | // routed segment, while distinct dispatches remain distinct. |
| 1594 | let _cost_scope = crate::cost_status::test_scope(); |
| 1595 | let owner = "auto-router-selection-test-owner"; |
| 1596 | crate::cost_status::register_interactive_runtime_usage_sink( |
| 1597 | owner, |
| 1598 | crate::cost_status::scope_token(), |
| 1599 | ); |
| 1600 | let lease = crate::cost_status::acquire_runtime_usage_lease(owner) |
| 1601 | .expect("runtime usage owner lease"); |
| 1602 | for record in &records { |
| 1603 | for _ in 0..2 { |
| 1604 | crate::cost_status::report_effective_route_for_runtime( |
| 1605 | crate::cost_status::scope_token(), |
| 1606 | Some(lease.owner()), |
| 1607 | &record.source_id, |
| 1608 | &record.usage.route, |
| 1609 | &record.usage.usage, |
| 1610 | ); |
| 1611 | } |
| 1612 | } |
| 1613 | crate::cost_status::finish_runtime_usage_owner(owner); |
| 1614 | drop(lease); |
| 1615 | let pending = crate::cost_status::drain(); |
| 1616 | assert_eq!(pending.usage_source_fingerprints.len(), records.len()); |
| 1617 | assert_eq!(pending.priced_turns, records.len() as u32); |
| 1618 | assert_eq!(pending.unpriced_turns, 0); |
| 1619 | } |
| 1620 | |
| 1621 | #[test] |
| 1622 | fn auto_model_reasoning_keeps_model_and_thinking_choices_independent() { |
| 1623 | assert_eq!( |
| 1624 | resolve_auto_model_reasoning(Some(ReasoningEffort::Low), Some(ReasoningEffort::Max)), |
| 1625 | (Some(ReasoningEffort::Low), false) |
| 1626 | ); |
| 1627 | assert_eq!( |
| 1628 | resolve_auto_model_reasoning(Some(ReasoningEffort::Auto), Some(ReasoningEffort::Max)), |
| 1629 | (Some(ReasoningEffort::Max), true) |
| 1630 | ); |
| 1631 | assert_eq!( |
| 1632 | resolve_auto_model_reasoning(None, Some(ReasoningEffort::High)), |
| 1633 | (Some(ReasoningEffort::High), true) |
| 1634 | ); |
| 1635 | } |
| 1636 | |
| 1637 | #[test] |
| 1638 | fn auto_route_prompt_uses_current_session_mode() { |
| 1639 | let prompt = auto_route_prompt( |
| 1640 | "Please explain the change before editing files.", |
| 1641 | "No prior context.", |
| 1642 | "plan", |
| 1643 | "auto", |
| 1644 | "auto", |
| 1645 | ); |
| 1646 | |
| 1647 | assert!( |
| 1648 | prompt.starts_with("Session mode: plan\n"), |
| 1649 | "auto-route prompt should reflect the active session mode, got: {prompt}" |
| 1650 | ); |
| 1651 | } |
| 1652 | |
| 1653 | #[test] |
| 1654 | fn classifier_prompt_redacts_secret_after_tool_result_flattening() { |
| 1655 | let secret = "cw-router-secret-should-never-leave-process"; |
| 1656 | let config = Config { |
| 1657 | api_key: Some(secret.to_string()), |
| 1658 | ..Default::default() |
| 1659 | }; |
| 1660 | let client = CodewhaleClient::new(&config).expect("classifier client"); |
| 1661 | // `recent_auto_router_context` converts ToolResult blocks into ordinary |
| 1662 | // text before this boundary. Exercise that exact flattened shape. |
| 1663 | let recent_context = format!("assistant: [tool result] token={secret}"); |
| 1664 | |
| 1665 | let prompt = classifier_prompt( |
| 1666 | &client, |
| 1667 | "continue the investigation", |
| 1668 | &recent_context, |
| 1669 | "agent", |
| 1670 | "auto", |
| 1671 | "auto", |
| 1672 | ); |
| 1673 | |
| 1674 | assert!( |
| 1675 | !prompt.contains(secret), |
| 1676 | "flattened tool-result secret leaked" |
| 1677 | ); |
| 1678 | assert!( |
| 1679 | prompt.contains(codewhale_config::persistence::REDACTED), |
| 1680 | "secret should be visibly redacted" |
| 1681 | ); |
| 1682 | assert!(prompt.contains("continue the investigation")); |
| 1683 | } |
| 1684 | |
| 1685 | #[test] |
| 1686 | fn inventory_auto_router_prompt_names_cost_saving_zai_pair() { |
| 1687 | let _env_lock = crate::test_support::lock_test_env(); |
| 1688 | let _deepseek = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY"); |
| 1689 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 1690 | let config = Config { |
| 1691 | provider: Some("zai".to_string()), |
| 1692 | ..Default::default() |
| 1693 | }; |
| 1694 | let inventory = ModelInventory::from_config(&config); |
| 1695 | |
| 1696 | let balanced = inventory_auto_router_system_prompt(&inventory, false); |
| 1697 | let cost_saving = inventory_auto_router_system_prompt(&inventory, true); |
| 1698 | |
| 1699 | assert!(!balanced.contains("Cost-saving mode is ON")); |
| 1700 | assert!( |
| 1701 | cost_saving.contains( |
| 1702 | "For the active provider `zai`, `GLM-5.3-Flash` is the fast tier and `GLM-5.3` is the strong tier" |
| 1703 | ), |
| 1704 | "cost-saving classifier policy must name the provider-safe pair: {cost_saving}" |
| 1705 | ); |
| 1706 | assert!( |
| 1707 | cost_saving.contains("Keep the selected model paired with provider `zai`"), |
| 1708 | "cost-saving policy must preserve provider/model validation: {cost_saving}" |
| 1709 | ); |
| 1710 | } |
| 1711 | |
| 1712 | #[test] |
| 1713 | fn auto_route_effort_normalization_is_provider_aware() { |
| 1714 | // Slice 4, D2: the Auto path delegates to the canonical route |
| 1715 | // normalizer with an unresolved route. Two providers where the deleted |
| 1716 | // local copy disagreed with the authority are pinned explicitly. |
| 1717 | assert_eq!( |
| 1718 | normalize_auto_route_effort_for_provider(ApiProvider::Deepseek, ReasoningEffort::Low), |
| 1719 | ReasoningEffort::Low, |
| 1720 | "first-party DeepSeek documents low|high|max, so the canonical normalizer keeps low" |
| 1721 | ); |
| 1722 | assert_eq!( |
| 1723 | normalize_auto_route_effort_for_provider( |
| 1724 | ApiProvider::Deepseek, |
| 1725 | ReasoningEffort::Medium |
| 1726 | ), |
| 1727 | ReasoningEffort::High |
| 1728 | ); |
| 1729 | assert_eq!( |
| 1730 | normalize_auto_route_effort_for_provider( |
| 1731 | ApiProvider::OllamaCloud, |
| 1732 | ReasoningEffort::Minimal |
| 1733 | ), |
| 1734 | ReasoningEffort::Low, |
| 1735 | "OllamaCloud folds the Codewhale-only `minimal` spelling onto low" |
| 1736 | ); |
| 1737 | assert_eq!( |
| 1738 | normalize_auto_route_effort_for_provider( |
| 1739 | ApiProvider::OllamaCloud, |
| 1740 | ReasoningEffort::Ultra |
| 1741 | ), |
| 1742 | ReasoningEffort::Max |
| 1743 | ); |
| 1744 | // A provider with no exact-route rule keeps the historic collapse. |
| 1745 | assert_eq!( |
| 1746 | normalize_auto_route_effort_for_provider(ApiProvider::Moonshot, ReasoningEffort::Low), |
| 1747 | ReasoningEffort::High |
| 1748 | ); |
| 1749 | assert_eq!( |
| 1750 | normalize_auto_route_effort_for_provider(ApiProvider::Moonshot, ReasoningEffort::Auto), |
| 1751 | ReasoningEffort::Auto |
| 1752 | ); |
| 1753 | assert_eq!( |
| 1754 | normalize_auto_route_effort_for_provider(ApiProvider::Moonshot, ReasoningEffort::Max), |
| 1755 | ReasoningEffort::Max |
| 1756 | ); |
| 1757 | // Codex keeps its provider-level mapping (off -> low, auto -> medium). |
| 1758 | assert_eq!( |
| 1759 | normalize_auto_route_effort_for_provider( |
| 1760 | ApiProvider::OpenaiCodex, |
| 1761 | ReasoningEffort::Low |
| 1762 | ), |
| 1763 | ReasoningEffort::Low |
| 1764 | ); |
| 1765 | assert_eq!( |
| 1766 | normalize_auto_route_effort_for_provider( |
| 1767 | ApiProvider::OpenaiCodex, |
| 1768 | ReasoningEffort::Medium |
| 1769 | ), |
| 1770 | ReasoningEffort::Medium |
| 1771 | ); |
| 1772 | assert_eq!( |
| 1773 | normalize_auto_route_effort_for_provider( |
| 1774 | ApiProvider::OpenaiCodex, |
| 1775 | ReasoningEffort::Off |
| 1776 | ), |
| 1777 | ReasoningEffort::Low |
| 1778 | ); |
| 1779 | } |
| 1780 | |
| 1781 | #[test] |
| 1782 | fn configured_route_effort_normalizer_keeps_kimi_code_low_medium_local() { |
| 1783 | let mut config = Config { |
| 1784 | provider: Some("moonshot".to_string()), |
| 1785 | providers: Some(crate::config::ProvidersConfig { |
| 1786 | moonshot: crate::config::ProviderConfig { |
| 1787 | base_url: Some(crate::config::DEFAULT_KIMI_CODE_BASE_URL.to_string()), |
| 1788 | model: Some("k3".to_string()), |
| 1789 | ..Default::default() |
| 1790 | }, |
| 1791 | ..Default::default() |
| 1792 | }), |
| 1793 | ..Default::default() |
| 1794 | }; |
| 1795 | assert_eq!( |
| 1796 | normalize_auto_route_effort_for_configured_route( |
| 1797 | &config, |
| 1798 | ApiProvider::Moonshot, |
| 1799 | "k3", |
| 1800 | ReasoningEffort::Low, |
| 1801 | ), |
| 1802 | ReasoningEffort::Low |
| 1803 | ); |
| 1804 | assert_eq!( |
| 1805 | normalize_auto_route_effort_for_configured_route( |
| 1806 | &config, |
| 1807 | ApiProvider::Moonshot, |
| 1808 | "k3", |
| 1809 | ReasoningEffort::Medium, |
| 1810 | ), |
| 1811 | ReasoningEffort::Medium |
| 1812 | ); |
| 1813 | |
| 1814 | config |
| 1815 | .providers |
| 1816 | .as_mut() |
| 1817 | .expect("providers") |
| 1818 | .moonshot |
| 1819 | .base_url = Some(crate::config::DEFAULT_MOONSHOT_BASE_URL.to_string()); |
| 1820 | assert_eq!( |
| 1821 | normalize_auto_route_effort_for_configured_route( |
| 1822 | &config, |
| 1823 | ApiProvider::Moonshot, |
| 1824 | "k3", |
| 1825 | ReasoningEffort::Low, |
| 1826 | ), |
| 1827 | ReasoningEffort::High |
| 1828 | ); |
| 1829 | } |
| 1830 | |
| 1831 | #[test] |
| 1832 | fn inventory_auto_route_recommendation_requires_runnable_pair() { |
| 1833 | let _env_lock = crate::test_support::lock_test_env(); |
| 1834 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 1835 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 1836 | let config = Config { |
| 1837 | provider: Some("zai".to_string()), |
| 1838 | default_text_model: Some(crate::config::DEFAULT_TEXT_MODEL.to_string()), |
| 1839 | ..Default::default() |
| 1840 | }; |
| 1841 | let inventory = ModelInventory::from_config(&config); |
| 1842 | |
| 1843 | let route = parse_inventory_auto_route_recommendation( |
| 1844 | r#"{"provider":"zai","model":"GLM-5.2","thinking":"max"}"#, |
| 1845 | &inventory, |
| 1846 | ) |
| 1847 | .expect("valid inventory route should parse"); |
| 1848 | assert_eq!(route.provider, ApiProvider::Zai); |
| 1849 | assert_eq!(route.model, crate::config::ZAI_GLM_5_2_MODEL); |
| 1850 | assert_eq!(route.reasoning_effort, Some(ReasoningEffort::Max)); |
| 1851 | |
| 1852 | assert!( |
| 1853 | parse_inventory_auto_route_recommendation( |
| 1854 | r#"{"provider":"zai","model":"deepseek-v4-pro","thinking":"max"}"#, |
| 1855 | &inventory, |
| 1856 | ) |
| 1857 | .is_none(), |
| 1858 | "router must not pair a DeepSeek model with the Z.ai provider" |
| 1859 | ); |
| 1860 | |
| 1861 | let wrapped = parse_inventory_auto_route_recommendation( |
| 1862 | r#"route: {"provider":"zai","model":"GLM-5-Turbo","reasoning_effort":"medium"}"#, |
| 1863 | &inventory, |
| 1864 | ) |
| 1865 | .expect("wrapped inventory route should parse"); |
| 1866 | assert_eq!(wrapped.provider, ApiProvider::Zai); |
| 1867 | assert_eq!(wrapped.model, crate::config::ZAI_GLM_5_TURBO_MODEL); |
| 1868 | // Parsing is strict and literal; the historic Medium->High coercion |
| 1869 | // is applied downstream by normalize_auto_route_selection_for_config |
| 1870 | // so route-specific contracts (Kimi Code K3) can keep Medium. |
| 1871 | assert_eq!(wrapped.reasoning_effort, Some(ReasoningEffort::Medium)); |
| 1872 | } |
| 1873 | |
| 1874 | #[test] |
| 1875 | fn inventory_auto_route_recommendation_rejects_unready_candidate() { |
| 1876 | let _env_lock = crate::test_support::lock_test_env(); |
| 1877 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 1878 | let config = Config { |
| 1879 | provider: Some("zai".to_string()), |
| 1880 | ..Default::default() |
| 1881 | }; |
| 1882 | let mut inventory = ModelInventory::from_config(&config); |
| 1883 | let candidate = inventory |
| 1884 | .candidates |
| 1885 | .iter_mut() |
| 1886 | .find(|candidate| { |
| 1887 | candidate.provider == ApiProvider::Zai |
| 1888 | && candidate.model == crate::config::ZAI_GLM_5_2_MODEL |
| 1889 | }) |
| 1890 | .expect("Z.ai strong candidate"); |
| 1891 | candidate.readiness = crate::provider_readiness::ResolvedProviderReadiness::InvalidRoute; |
| 1892 | |
| 1893 | assert!( |
| 1894 | parse_inventory_auto_route_recommendation( |
| 1895 | r#"{"provider":"zai","model":"GLM-5.2","thinking":"max"}"#, |
| 1896 | &inventory, |
| 1897 | ) |
| 1898 | .is_none(), |
| 1899 | "classifier output must not revive an unsupported route" |
| 1900 | ); |
| 1901 | } |
| 1902 | |
| 1903 | #[test] |
| 1904 | fn inventory_auto_route_recommendation_accepts_wanjie_v4_ids() { |
| 1905 | let _env_lock = crate::test_support::lock_test_env(); |
| 1906 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 1907 | let _wanjie = crate::test_support::EnvVarGuard::set("WANJIE_ARK_API_KEY", "wanjie-key"); |
| 1908 | let config = Config { |
| 1909 | provider: Some("wanjie-ark".to_string()), |
| 1910 | ..Default::default() |
| 1911 | }; |
| 1912 | let inventory = ModelInventory::from_config(&config); |
| 1913 | |
| 1914 | let route = parse_inventory_auto_route_recommendation( |
| 1915 | r#"{"provider":"wanjie-ark","model":"deepseek-v4-pro","thinking":"max"}"#, |
| 1916 | &inventory, |
| 1917 | ) |
| 1918 | .expect("Wanjie V4 Pro inventory route should parse"); |
| 1919 | assert_eq!(route.provider, ApiProvider::WanjieArk); |
| 1920 | assert_eq!(route.model, "deepseek-v4-pro"); |
| 1921 | assert_eq!(route.reasoning_effort, Some(ReasoningEffort::Max)); |
| 1922 | |
| 1923 | let route = parse_inventory_auto_route_recommendation( |
| 1924 | r#"{"provider":"wanjie-ark","model":"deepseek-v4-flash","thinking":"off"}"#, |
| 1925 | &inventory, |
| 1926 | ) |
| 1927 | .expect("Wanjie V4 Flash inventory route should parse"); |
| 1928 | assert_eq!(route.provider, ApiProvider::WanjieArk); |
| 1929 | assert_eq!(route.model, "deepseek-v4-flash"); |
| 1930 | assert_eq!(route.reasoning_effort, Some(ReasoningEffort::Off)); |
| 1931 | } |
| 1932 | |
| 1933 | #[test] |
| 1934 | fn explicit_route_to_nonactive_provider_uses_that_providers_effort() { |
| 1935 | // Active provider is DeepSeek (whose effort floor is low/medium), but the |
| 1936 | // explicit model `GLM-5.2` only routes to Z.ai. The resolved effort must |
| 1937 | // be normalized for Z.ai — not left at DeepSeek's raw `low` setting. |
| 1938 | let _env_lock = crate::test_support::lock_test_env(); |
| 1939 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 1940 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 1941 | let config = Config { |
| 1942 | provider: Some("deepseek".to_string()), |
| 1943 | reasoning_effort: Some("low".to_string()), |
| 1944 | ..Default::default() |
| 1945 | }; |
| 1946 | |
| 1947 | let route = resolve_explicit_route_with_inventory(&config, "GLM-5.2") |
| 1948 | .expect("explicit GLM route should resolve to its provider"); |
| 1949 | |
| 1950 | assert_eq!( |
| 1951 | route.provider, |
| 1952 | ApiProvider::Zai, |
| 1953 | "GLM-5.2 must route to Z.ai, not the active DeepSeek provider" |
| 1954 | ); |
| 1955 | assert_eq!( |
| 1956 | route.reasoning_effort, |
| 1957 | Some(ReasoningEffort::High), |
| 1958 | "low must be normalized up to high for the Z.ai route, not passed through" |
| 1959 | ); |
| 1960 | |
| 1961 | // GLM-5.3 is the default and a first-class route: same provider |
| 1962 | // ownership, same effort normalization, and it resolves to its own id. |
| 1963 | let route_53 = resolve_explicit_route_with_inventory(&config, "GLM-5.3") |
| 1964 | .expect("explicit GLM-5.3 route should resolve to its provider"); |
| 1965 | assert_eq!( |
| 1966 | route_53.provider, |
| 1967 | ApiProvider::Zai, |
| 1968 | "GLM-5.3 must route to Z.ai, not the active DeepSeek provider" |
| 1969 | ); |
| 1970 | assert_eq!( |
| 1971 | route_53.model, |
| 1972 | crate::config::ZAI_GLM_5_3_MODEL, |
| 1973 | "GLM-5.3 must resolve to its own id" |
| 1974 | ); |
| 1975 | assert_eq!(route_53.reasoning_effort, Some(ReasoningEffort::High)); |
| 1976 | } |
| 1977 | |
| 1978 | #[tokio::test] |
| 1979 | #[allow(clippy::await_holding_lock)] |
| 1980 | async fn inventory_auto_route_resolves_active_authenticated_provider() { |
| 1981 | let _env_lock = crate::test_support::lock_test_env(); |
| 1982 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 1983 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 1984 | let config = Config { |
| 1985 | provider: Some("zai".to_string()), |
| 1986 | ..Default::default() |
| 1987 | }; |
| 1988 | |
| 1989 | // #6290 rework: without the flash classifier there is no |
| 1990 | // per-request signal, so every wording resolves the same declared |
| 1991 | // default — the short chat and the complex ask below must agree. |
| 1992 | for prompt in [ |
| 1993 | "quick status check", |
| 1994 | "please refactor this architecture and audit its security boundaries", |
| 1995 | ] { |
| 1996 | let route = resolve_auto_route_with_inventory(&config, prompt, "", "auto", "auto") |
| 1997 | .await |
| 1998 | .expect("inventory route should resolve with authenticated active provider"); |
| 1999 | |
| 2000 | assert_eq!(route.provider, ApiProvider::Zai); |
| 2001 | assert_eq!(route.model, crate::config::DEFAULT_ZAI_MODEL); |
| 2002 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2003 | let receipt = route.receipt.expect("Auto route receipt"); |
| 2004 | assert_eq!(receipt.tier, AutoRouteTier::Strong); |
| 2005 | assert_eq!(receipt.scope, AutoRouteScope::ResolvedProvider); |
| 2006 | assert_eq!(receipt.data_path, AutoRouteDataPath::LocalHeuristic); |
| 2007 | assert_eq!( |
| 2008 | receipt.reason, |
| 2009 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::DeclaredDefault), |
| 2010 | "prompt {prompt:?} must take the declared default, not a content judgment" |
| 2011 | ); |
| 2012 | assert_eq!(receipt.pair.strong, crate::config::DEFAULT_ZAI_MODEL); |
| 2013 | assert_eq!( |
| 2014 | receipt.pair.fast.as_deref(), |
| 2015 | Some(crate::config::ZAI_GLM_5_3_FLASH_MODEL) |
| 2016 | ); |
| 2017 | } |
| 2018 | } |
| 2019 | |
| 2020 | #[test] |
| 2021 | fn classifier_receipt_discloses_active_provider_scope_and_data_path() { |
| 2022 | let _env_lock = crate::test_support::lock_test_env(); |
| 2023 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2024 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2025 | let config = Config { |
| 2026 | provider: Some("zai".to_string()), |
| 2027 | ..Default::default() |
| 2028 | }; |
| 2029 | let inventory = ModelInventory::from_config(&config); |
| 2030 | let recommendation = parse_inventory_auto_route_recommendation( |
| 2031 | r#"{"provider":"zai","model":"GLM-5-Turbo","thinking":"off"}"#, |
| 2032 | &inventory, |
| 2033 | ) |
| 2034 | .expect("runnable classifier recommendation"); |
| 2035 | |
| 2036 | let route = auto_route_from_classifier(&inventory, recommendation); |
| 2037 | |
| 2038 | assert_eq!(route.provider, ApiProvider::Zai); |
| 2039 | assert_eq!(route.model, crate::config::ZAI_GLM_5_TURBO_MODEL); |
| 2040 | assert_eq!(route.source, AutoRouteSource::FlashRouter); |
| 2041 | let receipt = route.receipt.expect("classifier receipt"); |
| 2042 | assert_eq!(receipt.tier, AutoRouteTier::Fast); |
| 2043 | // #4411: the classifier only saw Z.ai routes, so the receipt says so |
| 2044 | // instead of claiming the wider runnable-providers scope. |
| 2045 | assert_eq!(receipt.scope, AutoRouteScope::ActiveProvider); |
| 2046 | assert_eq!( |
| 2047 | receipt.data_path, |
| 2048 | AutoRouteDataPath::Classifier { |
| 2049 | provider: ApiProvider::Deepseek, |
| 2050 | model: "deepseek-v4-flash".to_string(), |
| 2051 | } |
| 2052 | ); |
| 2053 | assert_eq!(receipt.reason, AutoRouteReason::ClassifierRecommendation); |
| 2054 | } |
| 2055 | |
| 2056 | #[test] |
| 2057 | fn classifier_recommendation_for_another_provider_is_refused_by_default() { |
| 2058 | // #4411: the payload never named DeepSeek, but a classifier can still |
| 2059 | // emit one. The recommendation must not become a route unless the |
| 2060 | // persisted cross-provider opt-in is set. |
| 2061 | let _env_lock = crate::test_support::lock_test_env(); |
| 2062 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2063 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2064 | let scoped = Config { |
| 2065 | provider: Some("zai".to_string()), |
| 2066 | ..Default::default() |
| 2067 | }; |
| 2068 | let scoped_inventory = ModelInventory::from_config(&scoped); |
| 2069 | let raw = r#"{"provider":"deepseek","model":"deepseek-v4-flash","thinking":"off"}"#; |
| 2070 | |
| 2071 | assert!( |
| 2072 | parse_inventory_auto_route_recommendation(raw, &scoped_inventory).is_none(), |
| 2073 | "cross-provider classifier output must be refused by default" |
| 2074 | ); |
| 2075 | // The same inventory still accepts an in-scope active-provider route. |
| 2076 | assert!( |
| 2077 | parse_inventory_auto_route_recommendation( |
| 2078 | r#"{"provider":"zai","model":"GLM-5.2","thinking":"max"}"#, |
| 2079 | &scoped_inventory, |
| 2080 | ) |
| 2081 | .is_some() |
| 2082 | ); |
| 2083 | |
| 2084 | let opted_in = Config { |
| 2085 | auto: Some(crate::config::AutoConfig { |
| 2086 | cost_saving: None, |
| 2087 | cross_provider: Some(true), |
| 2088 | router: None, |
| 2089 | }), |
| 2090 | ..scoped.clone() |
| 2091 | }; |
| 2092 | let opted_in_route = |
| 2093 | parse_inventory_auto_route_recommendation(raw, &ModelInventory::from_config(&opted_in)) |
| 2094 | .expect("opt-in admits the cross-provider recommendation"); |
| 2095 | assert_eq!(opted_in_route.provider, ApiProvider::Deepseek); |
| 2096 | } |
| 2097 | |
| 2098 | #[test] |
| 2099 | fn classifier_prompt_declares_active_provider_scope_by_default() { |
| 2100 | let _env_lock = crate::test_support::lock_test_env(); |
| 2101 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2102 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2103 | let config = Config { |
| 2104 | provider: Some("zai".to_string()), |
| 2105 | ..Default::default() |
| 2106 | }; |
| 2107 | |
| 2108 | let prompt = |
| 2109 | inventory_auto_router_system_prompt(&ModelInventory::from_config(&config), false); |
| 2110 | |
| 2111 | assert!( |
| 2112 | prompt.contains("Auto routing is scoped to the active provider `zai`"), |
| 2113 | "{prompt}" |
| 2114 | ); |
| 2115 | assert!(!prompt.contains("deepseek"), "{prompt}"); |
| 2116 | } |
| 2117 | |
| 2118 | #[tokio::test] |
| 2119 | #[allow(clippy::await_holding_lock)] |
| 2120 | async fn active_provider_declared_default_survives_scoping() { |
| 2121 | // #4411: scoping keeps Auto on the active provider. #6290 rework: |
| 2122 | // without the flash classifier there is no per-request tier signal, |
| 2123 | // so both wordings resolve the same declared default on Zai. |
| 2124 | let _env_lock = crate::test_support::lock_test_env(); |
| 2125 | let _deepseek = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY"); |
| 2126 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2127 | let config = Config { |
| 2128 | provider: Some("zai".to_string()), |
| 2129 | ..Default::default() |
| 2130 | }; |
| 2131 | |
| 2132 | for prompt in [ |
| 2133 | "refactor the routing module and audit its security boundaries", |
| 2134 | "hi", |
| 2135 | ] { |
| 2136 | let route = resolve_auto_route_with_inventory(&config, prompt, "", "auto", "auto") |
| 2137 | .await |
| 2138 | .expect("scoped Auto route"); |
| 2139 | assert_eq!(route.provider, ApiProvider::Zai); |
| 2140 | assert_eq!(route.model, crate::config::DEFAULT_ZAI_MODEL); |
| 2141 | let receipt = route.receipt.expect("scoped receipt"); |
| 2142 | assert_eq!(receipt.tier, AutoRouteTier::Strong); |
| 2143 | assert_eq!(receipt.scope, AutoRouteScope::ResolvedProvider); |
| 2144 | assert_eq!( |
| 2145 | receipt.reason, |
| 2146 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::DeclaredDefault), |
| 2147 | "prompt {prompt:?}" |
| 2148 | ); |
| 2149 | } |
| 2150 | } |
| 2151 | |
| 2152 | #[test] |
| 2153 | fn config_auto_cross_provider_defaults_to_false() { |
| 2154 | assert!(!Config::default().auto_cross_provider()); |
| 2155 | let opted_in = Config { |
| 2156 | auto: Some(crate::config::AutoConfig { |
| 2157 | cost_saving: None, |
| 2158 | cross_provider: Some(true), |
| 2159 | router: None, |
| 2160 | }), |
| 2161 | ..Default::default() |
| 2162 | }; |
| 2163 | assert!(opted_in.auto_cross_provider()); |
| 2164 | } |
| 2165 | |
| 2166 | #[test] |
| 2167 | fn classifier_receipt_never_reports_openrouter_default_for_another_family() { |
| 2168 | let _env_lock = crate::test_support::lock_test_env(); |
| 2169 | let _openrouter = |
| 2170 | crate::test_support::EnvVarGuard::set("OPENROUTER_API_KEY", "openrouter-key"); |
| 2171 | let config = Config { |
| 2172 | provider: Some("openrouter".to_string()), |
| 2173 | ..Default::default() |
| 2174 | }; |
| 2175 | let inventory = ModelInventory::from_config(&config); |
| 2176 | let recommendation = parse_inventory_auto_route_recommendation( |
| 2177 | r#"{"provider":"openrouter","model":"z-ai/glm-5.2","thinking":"max"}"#, |
| 2178 | &inventory, |
| 2179 | ) |
| 2180 | .expect("runnable non-default OpenRouter family"); |
| 2181 | |
| 2182 | let route = auto_route_from_classifier(&inventory, recommendation); |
| 2183 | let receipt = route.receipt.expect("classifier receipt"); |
| 2184 | |
| 2185 | assert_eq!(route.model, crate::config::OPENROUTER_GLM_5_2_MODEL); |
| 2186 | assert_eq!(receipt.pair.strong, crate::config::OPENROUTER_GLM_5_2_MODEL); |
| 2187 | assert_ne!( |
| 2188 | receipt.pair.fast.as_deref(), |
| 2189 | Some(crate::config::DEFAULT_OPENROUTER_FLASH_MODEL), |
| 2190 | "a GLM selection must not be described as the DeepSeek default pair" |
| 2191 | ); |
| 2192 | assert!(matches!( |
| 2193 | receipt.tier, |
| 2194 | AutoRouteTier::Strong | AutoRouteTier::Only |
| 2195 | )); |
| 2196 | } |
| 2197 | |
| 2198 | #[test] |
| 2199 | fn classifier_fallback_preserves_attempted_data_path_without_error_text() { |
| 2200 | let _env_lock = crate::test_support::lock_test_env(); |
| 2201 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2202 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2203 | let config = Config { |
| 2204 | provider: Some("zai".to_string()), |
| 2205 | ..Default::default() |
| 2206 | }; |
| 2207 | let inventory = ModelInventory::from_config(&config); |
| 2208 | let fallback = auto_route_declared_fallback(&config, &inventory); |
| 2209 | |
| 2210 | let route = auto_route_classifier_fallback(fallback, &inventory); |
| 2211 | |
| 2212 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2213 | let receipt = route.receipt.expect("fallback receipt"); |
| 2214 | assert_eq!(receipt.scope, AutoRouteScope::ResolvedProvider); |
| 2215 | assert!(matches!( |
| 2216 | receipt.data_path, |
| 2217 | AutoRouteDataPath::Classifier { |
| 2218 | provider: ApiProvider::Deepseek, |
| 2219 | ref model, |
| 2220 | } if model == "deepseek-v4-flash" |
| 2221 | )); |
| 2222 | assert_eq!( |
| 2223 | receipt.reason, |
| 2224 | AutoRouteReason::ClassifierFallback(AutoRouteHeuristicReason::DeclaredDefault) |
| 2225 | ); |
| 2226 | assert!(!receipt.reason.label().contains("secret-provider-error")); |
| 2227 | } |
| 2228 | |
| 2229 | #[test] |
| 2230 | fn pre_rework_receipt_shape_still_deserializes() { |
| 2231 | // Sessions saved before the #6290 rework persist |
| 2232 | // `local_heuristic` + content-derived reasons. They must keep |
| 2233 | // loading: the wrapper arrives via serde alias, the legacy reasons |
| 2234 | // are retained variants. |
| 2235 | let reason: AutoRouteReason = |
| 2236 | serde_json::from_str(r#"{"local_heuristic":"complex_request"}"#) |
| 2237 | .expect("pre-rework receipt reason deserializes"); |
| 2238 | assert_eq!( |
| 2239 | reason, |
| 2240 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::ComplexRequest) |
| 2241 | ); |
| 2242 | let receipt: AutoRouteReceipt = serde_json::from_str( |
| 2243 | r#"{"tier":"fast","pair":{"strong":"GLM-5.3","fast":"GLM-5.3-Flash"},"scope":"resolved_provider","data_path":"local_heuristic","reason":{"local_heuristic":"short_request"}}"#, |
| 2244 | ) |
| 2245 | .expect("pre-rework receipt deserializes"); |
| 2246 | assert_eq!(receipt.reason.label(), "local fallback: short request"); |
| 2247 | } |
| 2248 | |
| 2249 | #[tokio::test] |
| 2250 | #[allow(clippy::await_holding_lock)] |
| 2251 | async fn inventory_auto_route_never_falls_back_across_providers_by_default() { |
| 2252 | // #4411: the active provider has no usable credential, but another |
| 2253 | // provider does. Auto must stay on the active provider and report a |
| 2254 | // no-runnable-candidate fallback instead of silently spending the |
| 2255 | // other provider's key. |
| 2256 | let _env_lock = crate::test_support::lock_test_env(); |
| 2257 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2258 | let _zai = crate::test_support::EnvVarGuard::remove("ZAI_API_KEY"); |
| 2259 | let config = Config { |
| 2260 | provider: Some("zai".to_string()), |
| 2261 | ..Default::default() |
| 2262 | }; |
| 2263 | |
| 2264 | let route = |
| 2265 | resolve_auto_route_with_inventory(&config, "quick status check", "", "auto", "auto") |
| 2266 | .await |
| 2267 | .expect("inventory route should resolve without leaving the active provider"); |
| 2268 | |
| 2269 | assert_eq!(route.provider, ApiProvider::Zai); |
| 2270 | assert_ne!(route.provider, ApiProvider::Deepseek); |
| 2271 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2272 | let receipt = route.receipt.expect("Auto route receipt"); |
| 2273 | assert_eq!(receipt.scope, AutoRouteScope::ResolvedProvider); |
| 2274 | assert_eq!( |
| 2275 | receipt.reason, |
| 2276 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::NoRunnableCandidate) |
| 2277 | ); |
| 2278 | } |
| 2279 | |
| 2280 | #[tokio::test] |
| 2281 | #[allow(clippy::await_holding_lock)] |
| 2282 | async fn inventory_auto_route_crosses_providers_only_under_persisted_opt_in() { |
| 2283 | // The same configuration as above, plus the persisted |
| 2284 | // `[auto] cross_provider = true` opt-in (#4411). |
| 2285 | let _env_lock = crate::test_support::lock_test_env(); |
| 2286 | let _deepseek = crate::test_support::EnvVarGuard::set("DEEPSEEK_API_KEY", "ds-key"); |
| 2287 | let _zai = crate::test_support::EnvVarGuard::remove("ZAI_API_KEY"); |
| 2288 | let config = Config { |
| 2289 | provider: Some("zai".to_string()), |
| 2290 | auto: Some(crate::config::AutoConfig { |
| 2291 | cost_saving: None, |
| 2292 | cross_provider: Some(true), |
| 2293 | router: None, |
| 2294 | }), |
| 2295 | ..Default::default() |
| 2296 | }; |
| 2297 | |
| 2298 | let route = |
| 2299 | resolve_auto_route_with_inventory(&config, "quick status check", "", "auto", "auto") |
| 2300 | .await |
| 2301 | .expect("opted-in route should fall back to an authenticated provider"); |
| 2302 | |
| 2303 | assert_eq!(route.provider, ApiProvider::Deepseek); |
| 2304 | assert_eq!(route.model, "deepseek-flash"); |
| 2305 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2306 | } |
| 2307 | |
| 2308 | #[tokio::test] |
| 2309 | #[allow(clippy::await_holding_lock)] |
| 2310 | async fn inventory_auto_route_cost_saving_pins_fast_sibling() { |
| 2311 | let _env_lock = crate::test_support::lock_test_env(); |
| 2312 | let _deepseek = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY"); |
| 2313 | let _zai = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2314 | let balanced = Config { |
| 2315 | provider: Some("zai".to_string()), |
| 2316 | ..Default::default() |
| 2317 | }; |
| 2318 | let cost_saving = Config { |
| 2319 | auto: Some(crate::config::AutoConfig { |
| 2320 | cost_saving: Some(true), |
| 2321 | cross_provider: None, |
| 2322 | router: None, |
| 2323 | }), |
| 2324 | ..balanced.clone() |
| 2325 | }; |
| 2326 | |
| 2327 | let balanced_route = resolve_auto_route_with_inventory( |
| 2328 | &balanced, |
| 2329 | "Please implement a binary search", |
| 2330 | "", |
| 2331 | "auto", |
| 2332 | "auto", |
| 2333 | ) |
| 2334 | .await |
| 2335 | .expect("balanced Auto route should resolve"); |
| 2336 | let cost_saving_route = resolve_auto_route_with_inventory( |
| 2337 | &cost_saving, |
| 2338 | "Please implement a binary search", |
| 2339 | "", |
| 2340 | "auto", |
| 2341 | "auto", |
| 2342 | ) |
| 2343 | .await |
| 2344 | .expect("cost-saving Auto route should resolve"); |
| 2345 | |
| 2346 | assert_eq!(balanced_route.provider, ApiProvider::Zai); |
| 2347 | assert_eq!(balanced_route.model, crate::config::DEFAULT_ZAI_MODEL); |
| 2348 | assert_eq!(cost_saving_route.provider, ApiProvider::Zai); |
| 2349 | assert_eq!( |
| 2350 | cost_saving_route.model, |
| 2351 | crate::config::ZAI_GLM_5_3_FLASH_MODEL |
| 2352 | ); |
| 2353 | assert_eq!(cost_saving_route.source, AutoRouteSource::Heuristic); |
| 2354 | assert_eq!( |
| 2355 | balanced_route |
| 2356 | .receipt |
| 2357 | .as_ref() |
| 2358 | .map(|receipt| (receipt.tier, receipt.reason)), |
| 2359 | Some(( |
| 2360 | AutoRouteTier::Strong, |
| 2361 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::DeclaredDefault), |
| 2362 | )) |
| 2363 | ); |
| 2364 | assert_eq!( |
| 2365 | cost_saving_route |
| 2366 | .receipt |
| 2367 | .as_ref() |
| 2368 | .map(|receipt| (receipt.tier, receipt.reason)), |
| 2369 | Some(( |
| 2370 | AutoRouteTier::Fast, |
| 2371 | AutoRouteReason::LocalFallback(AutoRouteHeuristicReason::CostSavingPolicy), |
| 2372 | )) |
| 2373 | ); |
| 2374 | } |
| 2375 | |
| 2376 | #[tokio::test] |
| 2377 | #[allow(clippy::await_holding_lock)] |
| 2378 | async fn inventory_auto_route_uses_wanjie_v4_pair_without_deepseek_router() { |
| 2379 | let _env_lock = crate::test_support::lock_test_env(); |
| 2380 | let _deepseek = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY"); |
| 2381 | let _wanjie = crate::test_support::EnvVarGuard::set("WANJIE_ARK_API_KEY", "wanjie-key"); |
| 2382 | let config = Config { |
| 2383 | provider: Some("wanjie-ark".to_string()), |
| 2384 | default_text_model: Some("auto".to_string()), |
| 2385 | ..Default::default() |
| 2386 | }; |
| 2387 | |
| 2388 | // #6290 rework: no classifier, no content signal — both wordings |
| 2389 | // take the declared Wanjie default. |
| 2390 | for prompt in ["quick status check", "please refactor this architecture"] { |
| 2391 | let route = resolve_auto_route_with_inventory(&config, prompt, "", "auto", "auto") |
| 2392 | .await |
| 2393 | .expect("declared-default Wanjie route should resolve"); |
| 2394 | assert_eq!(route.provider, ApiProvider::WanjieArk); |
| 2395 | assert_eq!(route.model, "deepseek-reasoner"); |
| 2396 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2397 | } |
| 2398 | } |
| 2399 | |
| 2400 | #[tokio::test] |
| 2401 | #[allow(clippy::await_holding_lock)] |
| 2402 | async fn inventory_auto_route_uses_volcengine_v4_pair_without_deepseek_router() { |
| 2403 | let _env_lock = crate::test_support::lock_test_env(); |
| 2404 | let _deepseek = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY"); |
| 2405 | let _volcengine = |
| 2406 | crate::test_support::EnvVarGuard::set("VOLCENGINE_API_KEY", "volcengine-key"); |
| 2407 | let config = Config { |
| 2408 | provider: Some("volcengine".to_string()), |
| 2409 | default_text_model: Some("auto".to_string()), |
| 2410 | ..Default::default() |
| 2411 | }; |
| 2412 | |
| 2413 | // #6290 rework: no classifier, no content signal — both wordings |
| 2414 | // take the declared Volcengine default. |
| 2415 | for prompt in ["quick status check", "please refactor this architecture"] { |
| 2416 | let route = resolve_auto_route_with_inventory(&config, prompt, "", "auto", "auto") |
| 2417 | .await |
| 2418 | .expect("declared-default Volcengine route should resolve"); |
| 2419 | assert_eq!(route.provider, ApiProvider::Volcengine); |
| 2420 | assert_eq!(route.model, "deepseek-v4-pro"); |
| 2421 | assert_eq!(route.source, AutoRouteSource::Heuristic); |
| 2422 | } |
| 2423 | } |
| 2424 | |
| 2425 | #[test] |
| 2426 | fn provider_router_candidates_cover_known_provider_classes() { |
| 2427 | use crate::config::ApiProvider; |
| 2428 | |
| 2429 | let deepseek = provider_router_candidates(ApiProvider::Deepseek, "deepseek-v4-pro"); |
| 2430 | assert_eq!(deepseek.big, "deepseek-v4-pro"); |
| 2431 | assert_eq!(deepseek.cheap.as_deref(), Some("deepseek-v4-flash")); |
| 2432 | |
| 2433 | let openrouter = |
| 2434 | provider_router_candidates(ApiProvider::Openrouter, "deepseek/deepseek-v4-pro"); |
| 2435 | assert_eq!(openrouter.big, "deepseek/deepseek-v4-pro"); |
| 2436 | assert_eq!( |
| 2437 | openrouter.cheap.as_deref(), |
| 2438 | Some("deepseek/deepseek-v4-flash") |
| 2439 | ); |
| 2440 | |
| 2441 | let wanjie = provider_router_candidates(ApiProvider::WanjieArk, "deepseek-reasoner"); |
| 2442 | assert_eq!(wanjie.big, "deepseek-v4-pro"); |
| 2443 | assert_eq!(wanjie.cheap.as_deref(), Some("deepseek-v4-flash")); |
| 2444 | |
| 2445 | let volcengine = provider_router_candidates(ApiProvider::Volcengine, "DeepSeek-V4-Pro"); |
| 2446 | assert_eq!(volcengine.big, "DeepSeek-V4-Pro"); |
| 2447 | assert_eq!(volcengine.cheap.as_deref(), Some("DeepSeek-V4-Flash")); |
| 2448 | |
| 2449 | let zai = provider_router_candidates(ApiProvider::Zai, "GLM-5.2"); |
| 2450 | assert_eq!(zai.big, "GLM-5.2"); |
| 2451 | // GLM-5.2 faster/explore children route to GLM-5-Turbo (same-family fast |
| 2452 | // sibling), not back down to GLM-5.1. |
| 2453 | assert_eq!(zai.cheap.as_deref(), Some("GLM-5-Turbo")); |
| 2454 | |
| 2455 | let openrouter_glm = provider_router_candidates(ApiProvider::Openrouter, "z-ai/glm-5.2"); |
| 2456 | assert_eq!(openrouter_glm.big, "z-ai/glm-5.2"); |
| 2457 | assert_eq!(openrouter_glm.cheap.as_deref(), Some("z-ai/glm-5-turbo")); |
| 2458 | |
| 2459 | // GLM-5.3's fast sibling is Flash; GLM-5.2 still uses Turbo. |
| 2460 | let zai_53 = provider_router_candidates(ApiProvider::Zai, "GLM-5.3"); |
| 2461 | assert_eq!(zai_53.big, "GLM-5.3"); |
| 2462 | assert_eq!(zai_53.cheap.as_deref(), Some("GLM-5.3-Flash")); |
| 2463 | |
| 2464 | let openrouter_glm_53 = provider_router_candidates(ApiProvider::Openrouter, "z-ai/glm-5.3"); |
| 2465 | assert_eq!(openrouter_glm_53.big, "z-ai/glm-5.3"); |
| 2466 | assert_eq!( |
| 2467 | openrouter_glm_53.cheap.as_deref(), |
| 2468 | Some("z-ai/glm-5.3-flash") |
| 2469 | ); |
| 2470 | |
| 2471 | let zai_flash = provider_router_candidates(ApiProvider::Zai, "GLM-5.3-Flash"); |
| 2472 | assert_eq!(zai_flash.big, "GLM-5.3-Flash"); |
| 2473 | assert_eq!(zai_flash.cheap, None); |
| 2474 | |
| 2475 | // GLM-5.1 has no cheaper tier; faster children stay on the parent. |
| 2476 | let zai_51 = provider_router_candidates(ApiProvider::Zai, "GLM-5.1"); |
| 2477 | assert_eq!(zai_51.big, "GLM-5.1"); |
| 2478 | assert_eq!(zai_51.cheap, None); |
| 2479 | |
| 2480 | // GLM-5-Turbo is itself the cheap tier; no further downgrade. |
| 2481 | let zai_turbo = provider_router_candidates(ApiProvider::Zai, "GLM-5-Turbo"); |
| 2482 | assert_eq!(zai_turbo.big, "GLM-5-Turbo"); |
| 2483 | assert_eq!(zai_turbo.cheap, None); |
| 2484 | |
| 2485 | // Providers without a known cheap tier: big = session model, no cheap. |
| 2486 | let ollama = provider_router_candidates(ApiProvider::Ollama, "qwen3:32b"); |
| 2487 | assert_eq!(ollama.big, "qwen3:32b"); |
| 2488 | assert_eq!(ollama.cheap, None); |
| 2489 | |
| 2490 | let moonshot = provider_router_candidates(ApiProvider::Moonshot, "kimi-k2.6"); |
| 2491 | assert_eq!(moonshot.big, "kimi-k2.6"); |
| 2492 | assert_eq!(moonshot.cheap, None); |
| 2493 | } |
| 2494 | |
| 2495 | #[test] |
| 2496 | fn provider_router_candidates_cover_catalog_fast_siblings() { |
| 2497 | use crate::config::ApiProvider; |
| 2498 | |
| 2499 | let cases = [ |
| 2500 | (ApiProvider::OpenaiCodex, "gpt-5.6-sol", "gpt-5.6-luna"), |
| 2501 | ( |
| 2502 | ApiProvider::Anthropic, |
| 2503 | "claude-sonnet-4-6", |
| 2504 | "claude-haiku-4-5", |
| 2505 | ), |
| 2506 | (ApiProvider::XiaomiMimo, "mimo-v2.5-pro", "mimo-v2.5"), |
| 2507 | (ApiProvider::Arcee, "trinity-large-thinking", "trinity-mini"), |
| 2508 | (ApiProvider::Moonshot, "kimi-k2.7-code", "kimi-k2.6"), |
| 2509 | ( |
| 2510 | ApiProvider::Minimax, |
| 2511 | "MiniMax-M2.7", |
| 2512 | "MiniMax-M2.7-highspeed", |
| 2513 | ), |
| 2514 | (ApiProvider::OpencodeGo, "kimi-k3", "kimi-k2.7-code"), |
| 2515 | ( |
| 2516 | ApiProvider::Openrouter, |
| 2517 | "qwen/qwen3.6-max-preview", |
| 2518 | "qwen/qwen3.6-flash", |
| 2519 | ), |
| 2520 | ( |
| 2521 | ApiProvider::Openrouter, |
| 2522 | "anthropic/claude-sonnet-4-6", |
| 2523 | "anthropic/claude-haiku-4-5", |
| 2524 | ), |
| 2525 | ]; |
| 2526 | |
| 2527 | for (provider, strong, fast) in cases { |
| 2528 | let candidates = provider_router_candidates(provider, strong); |
| 2529 | assert_eq!(candidates.big, strong); |
| 2530 | assert_eq!(candidates.cheap.as_deref(), Some(fast)); |
| 2531 | assert_eq!( |
| 2532 | provider_router_candidates(provider, fast).cheap, |
| 2533 | None, |
| 2534 | "already-fast model must not downgrade again: {provider:?}/{fast}" |
| 2535 | ); |
| 2536 | } |
| 2537 | |
| 2538 | for (provider, model) in [ |
| 2539 | (ApiProvider::Ollama, "qwen3:32b"), |
| 2540 | (ApiProvider::Custom, "gpt-5.6-sol"), |
| 2541 | (ApiProvider::OpenaiCodex, "gpt-5.6-luna"), |
| 2542 | ] { |
| 2543 | assert_eq!(provider_router_candidates(provider, model).cheap, None); |
| 2544 | } |
| 2545 | } |
| 2546 | |
| 2547 | #[tokio::test] |
| 2548 | #[allow(clippy::await_holding_lock)] |
| 2549 | async fn declared_fallback_without_cheap_tier_stays_on_default_model() { |
| 2550 | // #3018 AC: Ollama + auto must never fabricate a DeepSeek id. The |
| 2551 | // declared fallback returns the configured default verbatim, so no |
| 2552 | // sibling id can be invented regardless of request wording. |
| 2553 | let _env_lock = crate::test_support::lock_test_env(); |
| 2554 | let config = Config { |
| 2555 | provider: Some("ollama".to_string()), |
| 2556 | ..Default::default() |
| 2557 | }; |
| 2558 | for prompt in ["hi", "please refactor the auth module for security"] { |
| 2559 | let route = resolve_auto_route_with_inventory(&config, prompt, "", "auto", "auto") |
| 2560 | .await |
| 2561 | .expect("ollama Auto route should resolve"); |
| 2562 | assert_eq!(route.provider, ApiProvider::Ollama); |
| 2563 | assert_eq!(route.model, config.default_model()); |
| 2564 | assert!( |
| 2565 | !route.model.to_ascii_lowercase().contains("deepseek"), |
| 2566 | "no DeepSeek id may be fabricated: {}", |
| 2567 | route.model |
| 2568 | ); |
| 2569 | } |
| 2570 | } |
| 2571 | |
| 2572 | #[test] |
| 2573 | fn config_auto_cost_saving_defaults_to_false() { |
| 2574 | let cfg = Config::default(); |
| 2575 | assert!(!cfg.auto_cost_saving()); |
| 2576 | } |
| 2577 | |
| 2578 | #[test] |
| 2579 | fn config_auto_cost_saving_reads_table() { |
| 2580 | let cfg = Config { |
| 2581 | auto: Some(crate::config::AutoConfig { |
| 2582 | cost_saving: Some(true), |
| 2583 | cross_provider: None, |
| 2584 | router: None, |
| 2585 | }), |
| 2586 | ..Default::default() |
| 2587 | }; |
| 2588 | assert!(cfg.auto_cost_saving()); |
| 2589 | } |
| 2590 | } |
| 2591 |