| 1 | //! Process-wide cost-accrual side-channel (#526). |
| 2 | //! |
| 3 | //! Background LLM calls outside the main turn-complete path |
| 4 | //! (compaction summaries) used |
| 5 | //! to drop their token usage on the floor — the dashboard's |
| 6 | //! session-cost only saw the parent turn's tokens, so a long |
| 7 | //! session that triggered compaction under-reported |
| 8 | //! cost by however many tokens those background calls consumed. |
| 9 | //! |
| 10 | //! Mirrors the [`crate::retry_status`] pattern: background callers |
| 11 | //! call [`crate::cost_status::report_effective_route`] after each |
| 12 | //! `client.create_message`, the TUI |
| 13 | //! render loop calls [`drain`] every frame, and any drained amount |
| 14 | //! gets folded into `App::accrue_subagent_cost_estimate`. |
| 15 | //! |
| 16 | //! Why a side-channel and not a plumbed callback: the leaky callers |
| 17 | //! (`compaction.rs`) are |
| 18 | //! engine-internal machinery without a direct handle to `App` or |
| 19 | //! the engine's event channel. A side-channel keeps the change |
| 20 | //! surface tiny — one new `report` line per call site — and any |
| 21 | //! future background caller (summarizers, retrieval helpers) gets |
| 22 | //! accrued for free without further plumbing. |
| 23 | //! |
| 24 | //! ## One pool, not a pile of counters (#4318) |
| 25 | //! |
| 26 | //! Money and the *completeness* of that money are one fact, so they live in one |
| 27 | //! mutex-guarded [`PendingBackgroundCost`] that [`drain`] takes atomically. |
| 28 | //! Splitting them across free-standing atomics made two things go wrong at once: |
| 29 | //! a drain could observe a total without the counters that explain it, and every |
| 30 | //! new global was another piece of state a parallel test had to remember to |
| 31 | //! reset. There is exactly one *drainable cost pool*. The runtime-owner journal |
| 32 | //! below is a separate route/usage copy (never another money counter), and the |
| 33 | //! shared test reset clears both stores. |
| 34 | |
| 35 | use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet, VecDeque}; |
| 36 | use std::sync::{Arc, Mutex, OnceLock}; |
| 37 | |
| 38 | use chrono::{DateTime, Utc}; |
| 39 | |
| 40 | use crate::config::ProviderKind; |
| 41 | use crate::pricing::{CostEstimate, TurnCostAudit}; |
| 42 | use crate::route_billing::BillingPresentation; |
| 43 | use codewhale_models::Usage; |
| 44 | |
| 45 | /// Everything a drained background accrual needs to be explained. |
| 46 | /// |
| 47 | /// The money and the coverage/provenance that qualify it are drained together, |
| 48 | /// so `/cost` can never show a background subtotal whose completeness came from |
| 49 | /// a different observation. |
| 50 | #[derive(Debug, Clone, Default, PartialEq)] |
| 51 | pub struct PendingBackgroundCost { |
| 52 | /// Summed cost of the background turns that were priced. |
| 53 | pub estimate: CostEstimate, |
| 54 | /// Background turns that produced an authoritative price. |
| 55 | pub priced_turns: u32, |
| 56 | /// Background turns that were money-metered (or of unknown basis) but |
| 57 | /// produced no authoritative price, so their spend is missing. |
| 58 | pub unpriced_turns: u32, |
| 59 | /// Money-metered turns authoritatively priced in CNY. |
| 60 | pub cny_priced_turns: u32, |
| 61 | /// Money-metered turns missing authoritative CNY pricing. |
| 62 | pub cny_unpriced_turns: u32, |
| 63 | /// Stable reason labels for the unpriced turns. |
| 64 | pub unpriced_reasons: BTreeSet<&'static str>, |
| 65 | pub cny_unpriced_reasons: BTreeSet<&'static str>, |
| 66 | /// Token classes used on a background route that carry no published price. |
| 67 | pub unpriced_classes: BTreeSet<&'static str>, |
| 68 | /// Provenance labels of the pricing rows that were applied or attempted. |
| 69 | pub pricing_provenances: BTreeSet<&'static str>, |
| 70 | /// Live-pricing downgrade receipts, when a live catalog row could not be |
| 71 | /// verified for the endpoint that served the turn. |
| 72 | pub live_pricing_defects: BTreeSet<&'static str>, |
| 73 | /// Live pricing failed and no bundled row could price the turn. Kept |
| 74 | /// separate so `/cost` never claims a bundled fallback was used when the |
| 75 | /// result is actually unavailable. |
| 76 | pub live_pricing_unusable_defects: BTreeSet<&'static str>, |
| 77 | /// One redacted receipt per distinct background route that reported. |
| 78 | /// |
| 79 | /// See [`EffectiveRouteEnvelope::receipt`] for the exact contents; these carry |
| 80 | /// provider identity, endpoint *fingerprint*, billing surface, wire model, |
| 81 | /// and currency — never a URL, key, token, or filesystem path. |
| 82 | pub route_receipts: BTreeSet<String>, |
| 83 | /// Durable, redacted identities of provider responses folded into this |
| 84 | /// batch. These travel with the money so a session snapshot can make a |
| 85 | /// replay idempotent after reload. |
| 86 | pub usage_source_fingerprints: BTreeSet<String>, |
| 87 | /// Exact unresolved coverage in this same source ledger, bounded to 64. |
| 88 | pub missing_usage_sources: BTreeMap<String, MissingUsageCoverage>, |
| 89 | /// A late real receipt supersedes only this exact unresolved source. |
| 90 | pub resolved_missing_usage_sources: BTreeSet<String>, |
| 91 | pub missing_usage_overflowed: bool, |
| 92 | /// Prompt-cache classes the background routes reported, through |
| 93 | /// [`crate::pricing::token_usage_for_pricing`] so they never exceed the |
| 94 | /// input they partition (#6565). `None` until a child reports cache |
| 95 | /// telemetry at all: no report is not a 0% hit rate. |
| 96 | pub cache_hit_tokens: Option<u64>, |
| 97 | pub cache_miss_tokens: Option<u64>, |
| 98 | pub cache_write_tokens: Option<u64>, |
| 99 | } |
| 100 | |
| 101 | /// Immutable, non-secret route evidence captured before a provider request. |
| 102 | /// It contains enough information to audit the eventual usage without reading |
| 103 | /// mutable parent/app config at completion time. |
| 104 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 105 | pub struct EffectiveRouteEnvelope { |
| 106 | pub provider: ProviderKind, |
| 107 | pub provider_identity: String, |
| 108 | pub model: String, |
| 109 | /// Requested OpenRouter upstream, frozen with the client that dispatched. |
| 110 | pub openrouter_vendor: Option<String>, |
| 111 | pub billing_surface: Option<String>, |
| 112 | pub endpoint_fingerprint: Option<String>, |
| 113 | /// Frozen provider-live or signed cloud rates captured from the exact catalog scope |
| 114 | /// at CodeWhale's pre-permit application-dispatch boundary. Legacy |
| 115 | /// receipts omit this and therefore cannot meter a reviewed custom route |
| 116 | /// retroactively. |
| 117 | pub provider_live_pricing: Option<crate::provider_catalog_live::ProviderLivePricingQuote>, |
| 118 | pub billing_mode: RouteBillingMode, |
| 119 | pub dispatched_at: DateTime<Utc>, |
| 120 | } |
| 121 | |
| 122 | #[derive(serde::Deserialize)] |
| 123 | struct EffectiveRouteEnvelopeWire { |
| 124 | provider: String, |
| 125 | provider_identity: String, |
| 126 | model: String, |
| 127 | /// Requested OpenRouter upstream, frozen with the client that dispatched. |
| 128 | #[serde(default)] |
| 129 | openrouter_vendor: Option<String>, |
| 130 | billing_surface: Option<String>, |
| 131 | endpoint_fingerprint: Option<String>, |
| 132 | /// Frozen provider-live or signed cloud rates captured from the exact catalog scope |
| 133 | /// at CodeWhale's pre-permit application-dispatch boundary. Legacy |
| 134 | /// receipts omit this and therefore cannot meter a reviewed custom route |
| 135 | /// retroactively. |
| 136 | #[serde( |
| 137 | default, |
| 138 | deserialize_with = "crate::provider_catalog_live::deserialize_optional_provider_live_pricing" |
| 139 | )] |
| 140 | provider_live_pricing: Option<crate::provider_catalog_live::ProviderLivePricingQuote>, |
| 141 | #[serde(default)] |
| 142 | billing_mode: RouteBillingMode, |
| 143 | dispatched_at: DateTime<Utc>, |
| 144 | } |
| 145 | |
| 146 | impl<'de> serde::Deserialize<'de> for EffectiveRouteEnvelope { |
| 147 | fn deserialize<D>(deserializer: D) -> Result<Self, D::Error> |
| 148 | where |
| 149 | D: serde::Deserializer<'de>, |
| 150 | { |
| 151 | let wire = <EffectiveRouteEnvelopeWire as serde::Deserialize>::deserialize(deserializer)?; |
| 152 | let provider = codewhale_config::descriptors::kind_from_tui_wire_tag( |
| 153 | &wire.provider, |
| 154 | &wire.provider_identity, |
| 155 | ) |
| 156 | .ok_or_else(|| serde::de::Error::custom("contradictory usage provider identity"))?; |
| 157 | Ok(Self { |
| 158 | provider, |
| 159 | provider_identity: wire.provider_identity, |
| 160 | model: wire.model, |
| 161 | openrouter_vendor: wire.openrouter_vendor, |
| 162 | billing_surface: wire.billing_surface, |
| 163 | endpoint_fingerprint: wire.endpoint_fingerprint, |
| 164 | provider_live_pricing: wire.provider_live_pricing, |
| 165 | billing_mode: wire.billing_mode, |
| 166 | dispatched_at: wire.dispatched_at, |
| 167 | }) |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | impl serde::Serialize for EffectiveRouteEnvelope { |
| 172 | fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error> |
| 173 | where |
| 174 | S: serde::Serializer, |
| 175 | { |
| 176 | use serde::ser::SerializeStruct as _; |
| 177 | |
| 178 | let route = self.sanitized_for_persistence(); |
| 179 | let mut state = serializer.serialize_struct( |
| 180 | "EffectiveRouteEnvelope", |
| 181 | 8 + usize::from(route.openrouter_vendor.is_some()), |
| 182 | )?; |
| 183 | let provider = codewhale_config::descriptors::tui_wire_tag_for_route( |
| 184 | route.provider, |
| 185 | &route.provider_identity, |
| 186 | ) |
| 187 | .ok_or_else(|| serde::ser::Error::custom("contradictory usage provider identity"))?; |
| 188 | state.serialize_field("provider", provider)?; |
| 189 | state.serialize_field("provider_identity", &route.provider_identity)?; |
| 190 | state.serialize_field("model", &route.model)?; |
| 191 | if let Some(vendor) = &route.openrouter_vendor { |
| 192 | state.serialize_field("openrouter_vendor", vendor)?; |
| 193 | } |
| 194 | state.serialize_field("billing_surface", &route.billing_surface)?; |
| 195 | state.serialize_field("endpoint_fingerprint", &route.endpoint_fingerprint)?; |
| 196 | state.serialize_field("provider_live_pricing", &route.provider_live_pricing)?; |
| 197 | state.serialize_field("billing_mode", &route.billing_mode)?; |
| 198 | state.serialize_field("dispatched_at", &route.dispatched_at)?; |
| 199 | state.end() |
| 200 | } |
| 201 | } |
| 202 | |
| 203 | /// One provider usage payload paired with the immutable route that served it. |
| 204 | /// |
| 205 | /// Runtime hosts persist these for model calls made below the parent turn |
| 206 | /// (sub-agents, review/verify/RLM tools, and compaction). Keeping route and |
| 207 | /// usage together makes the record independently auditable and prevents a |
| 208 | /// later provider/model selection from changing its price. |
| 209 | #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 210 | pub struct EffectiveRouteUsage { |
| 211 | pub route: EffectiveRouteEnvelope, |
| 212 | pub usage: Usage, |
| 213 | } |
| 214 | |
| 215 | #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 216 | #[serde(rename_all = "snake_case")] |
| 217 | pub enum RouteBillingMode { |
| 218 | Metered, |
| 219 | Subscription, |
| 220 | Local, |
| 221 | #[default] |
| 222 | Unknown, |
| 223 | } |
| 224 | |
| 225 | impl From<BillingPresentation> for RouteBillingMode { |
| 226 | fn from(value: BillingPresentation) -> Self { |
| 227 | match value { |
| 228 | BillingPresentation::Metered => Self::Metered, |
| 229 | BillingPresentation::Subscription(_) => Self::Subscription, |
| 230 | BillingPresentation::Local => Self::Local, |
| 231 | BillingPresentation::Unknown => Self::Unknown, |
| 232 | } |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | impl EffectiveRouteEnvelope { |
| 237 | /// Provider-neutral observation. This never opens configuration or credentials. |
| 238 | #[must_use] |
| 239 | pub fn capture_observed( |
| 240 | provider: ProviderKind, |
| 241 | provider_identity: impl Into<String>, |
| 242 | model: impl Into<String>, |
| 243 | base_url: Option<&str>, |
| 244 | dispatched_at: DateTime<Utc>, |
| 245 | ) -> Self { |
| 246 | let key = provider_identity.into(); |
| 247 | Self::from_facts(None, (provider, &key), model, base_url, dispatched_at) |
| 248 | } |
| 249 | |
| 250 | #[cfg(test)] |
| 251 | pub fn capture( |
| 252 | config: Option<&crate::config::Config>, |
| 253 | provider: ProviderKind, |
| 254 | provider_identity: impl Into<String>, |
| 255 | model: impl Into<String>, |
| 256 | base_url: Option<&str>, |
| 257 | dispatched_at: DateTime<Utc>, |
| 258 | ) -> Self { |
| 259 | let key = provider_identity.into(); |
| 260 | if let Some(config) = config |
| 261 | && let Ok(identity) = |
| 262 | config.resolve_persisted_provider_identity(Some(provider.as_str()), Some(&key)) |
| 263 | { |
| 264 | Self::from_admitted(Some(config), &identity, model, base_url, dispatched_at) |
| 265 | } else { |
| 266 | Self::capture_observed(provider, key, model, base_url, dispatched_at) |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | #[must_use] |
| 271 | pub fn from_admitted( |
| 272 | config: Option<&crate::config::Config>, |
| 273 | identity: &crate::config::ProviderIdentity, |
| 274 | model: impl Into<String>, |
| 275 | base_url: Option<&str>, |
| 276 | dispatched_at: DateTime<Utc>, |
| 277 | ) -> Self { |
| 278 | let admission = config |
| 279 | .filter(|config| config.verify_provider_identity(identity).is_ok()) |
| 280 | .map(|config| (config, identity)); |
| 281 | Self::from_facts( |
| 282 | admission, |
| 283 | (identity.provider, identity.key.as_str()), |
| 284 | model, |
| 285 | base_url, |
| 286 | dispatched_at, |
| 287 | ) |
| 288 | } |
| 289 | |
| 290 | fn from_facts( |
| 291 | admission: Option<(&crate::config::Config, &crate::config::ProviderIdentity)>, |
| 292 | facts: (ProviderKind, &str), |
| 293 | model: impl Into<String>, |
| 294 | base_url: Option<&str>, |
| 295 | dispatched_at: DateTime<Utc>, |
| 296 | ) -> Self { |
| 297 | let (provider, provider_identity) = facts; |
| 298 | let config = admission.map(|(config, _)| config); |
| 299 | let model = model.into(); |
| 300 | let billing = admission.map_or_else( |
| 301 | || crate::route_billing::for_endpoint_without_config(provider, base_url), |
| 302 | |(config, identity)| { |
| 303 | base_url.map_or_else( |
| 304 | || crate::route_billing::for_route(config, identity), |
| 305 | |endpoint| { |
| 306 | crate::route_billing::for_route_with_endpoint(config, identity, endpoint) |
| 307 | }, |
| 308 | ) |
| 309 | }, |
| 310 | ); |
| 311 | let endpoint_fingerprint = base_url.and_then(endpoint_fingerprint); |
| 312 | let provider_live_pricing = base_url.and_then(|base_url| { |
| 313 | u64::try_from(dispatched_at.timestamp()) |
| 314 | .ok() |
| 315 | .and_then(|at| { |
| 316 | crate::provider_catalog_live::declared_or_catalog_quote( |
| 317 | config.and_then(|config| { |
| 318 | crate::provider_catalog_live::configured_dispatch_pricing_quote_at( |
| 319 | config.custom_models.as_deref().unwrap_or_default(), |
| 320 | provider, |
| 321 | provider_identity, |
| 322 | &model, |
| 323 | base_url, |
| 324 | at, |
| 325 | ) |
| 326 | }), |
| 327 | || { |
| 328 | crate::provider_catalog_live::fresh_dispatch_pricing_quote_at( |
| 329 | provider, |
| 330 | provider_identity, |
| 331 | &model, |
| 332 | base_url, |
| 333 | at, |
| 334 | ) |
| 335 | }, |
| 336 | ) |
| 337 | }) |
| 338 | }); |
| 339 | Self { |
| 340 | provider, |
| 341 | provider_identity: sanitize_persisted_route_label(provider_identity), |
| 342 | model: sanitize_persisted_route_label(&model), |
| 343 | openrouter_vendor: admission |
| 344 | .filter(|_| provider == ProviderKind::Openrouter) |
| 345 | .and_then(|(config, identity)| config.provider_config_for(identity)) |
| 346 | .and_then(|entry| entry.vendor.as_deref()) |
| 347 | .map(str::trim) |
| 348 | .filter(|vendor| !vendor.is_empty()) |
| 349 | .map(sanitize_persisted_route_label), |
| 350 | billing_surface: admission |
| 351 | .map_or_else( |
| 352 | || crate::pricing::billing_surface_for_route(provider, base_url), |
| 353 | |(config, identity)| { |
| 354 | crate::route_billing::billing_surface_for_dispatch( |
| 355 | Some(config), |
| 356 | identity, |
| 357 | base_url, |
| 358 | ) |
| 359 | }, |
| 360 | ) |
| 361 | .map(str::to_string), |
| 362 | endpoint_fingerprint, |
| 363 | provider_live_pricing, |
| 364 | billing_mode: billing.into(), |
| 365 | dispatched_at, |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | #[must_use] |
| 370 | pub fn audit(&self, usage: &Usage) -> TurnCostAudit { |
| 371 | let reviewed_custom_metered = crate::pricing::reviewed_custom_route_is_metered( |
| 372 | self.provider, |
| 373 | self.endpoint_fingerprint.as_deref(), |
| 374 | ); |
| 375 | let declared_quote = self.provider_live_pricing.as_ref().filter(|quote| { |
| 376 | quote.provenance == codewhale_config::pricing::PricingProvenance::UserOverride |
| 377 | && self |
| 378 | .endpoint_fingerprint |
| 379 | .as_deref() |
| 380 | .zip(u64::try_from(self.dispatched_at.timestamp()).ok()) |
| 381 | .is_some_and(|(fingerprint, at)| { |
| 382 | quote |
| 383 | .pricing_for_route( |
| 384 | self.provider, |
| 385 | &self.provider_identity, |
| 386 | &self.model, |
| 387 | fingerprint, |
| 388 | at, |
| 389 | ) |
| 390 | .is_some() |
| 391 | }) |
| 392 | }); |
| 393 | let declared_estimate = declared_quote.is_some(); |
| 394 | match self.billing_mode { |
| 395 | RouteBillingMode::Subscription | RouteBillingMode::Local => { |
| 396 | return TurnCostAudit::unpriced(crate::pricing::UnpricedReason::NotMoneyMetered); |
| 397 | } |
| 398 | RouteBillingMode::Unknown if !reviewed_custom_metered && !declared_estimate => { |
| 399 | return TurnCostAudit::unpriced( |
| 400 | crate::pricing::UnpricedReason::UnknownBillingBasis, |
| 401 | ); |
| 402 | } |
| 403 | RouteBillingMode::Metered | RouteBillingMode::Unknown => {} |
| 404 | } |
| 405 | // The OpenRouter model catalog does not identify a pinned upstream's |
| 406 | // price. An endpoint match alone must not promote that aggregate rate. |
| 407 | // An operator's own declared rate for this exact route is not the |
| 408 | // aggregate catalog, so it still prices the pinned turn. |
| 409 | if self.provider == ProviderKind::Openrouter |
| 410 | && self.openrouter_vendor.is_some() |
| 411 | && !declared_quote.is_some_and(|quote| quote.carries_rates()) |
| 412 | { |
| 413 | return TurnCostAudit::unpriced(crate::pricing::UnpricedReason::RoutingDependentPrice); |
| 414 | } |
| 415 | crate::pricing::audit_turn_cost_for_route_on_endpoint_for_identity_at( |
| 416 | self.provider, |
| 417 | Some(&self.provider_identity), |
| 418 | &self.model, |
| 419 | self.billing_surface.as_deref(), |
| 420 | self.endpoint_fingerprint.as_deref(), |
| 421 | self.provider_live_pricing.as_ref(), |
| 422 | usage, |
| 423 | self.dispatched_at, |
| 424 | ) |
| 425 | } |
| 426 | |
| 427 | #[must_use] |
| 428 | pub fn receipt(&self, audit: &TurnCostAudit) -> String { |
| 429 | let route = self.sanitized_for_persistence(); |
| 430 | let mut receipt = route_receipt( |
| 431 | route.provider, |
| 432 | Some(&route.provider_identity), |
| 433 | &route.model, |
| 434 | route.billing_surface.as_deref(), |
| 435 | route.endpoint_fingerprint.as_deref(), |
| 436 | route.billing_mode, |
| 437 | currency_tag(audit), |
| 438 | ); |
| 439 | if let Some(vendor) = route.openrouter_vendor.as_deref() { |
| 440 | receipt.push_str(" openrouter_vendor="); |
| 441 | receipt.push_str(&safe_receipt_field(vendor)); |
| 442 | } |
| 443 | receipt |
| 444 | } |
| 445 | |
| 446 | /// Redact filesystem-like labels before a route crosses a persistence or |
| 447 | /// metadata boundary. Ordinary provider model namespaces such as |
| 448 | /// `anthropic/claude-*` remain intact; absolute/local path forms do not. |
| 449 | #[must_use] |
| 450 | pub fn sanitized_for_persistence(&self) -> Self { |
| 451 | let mut route = self.clone(); |
| 452 | route.provider_identity = sanitize_persisted_route_label(&route.provider_identity); |
| 453 | route.model = sanitize_persisted_route_label(&route.model); |
| 454 | route.openrouter_vendor = route |
| 455 | .openrouter_vendor |
| 456 | .as_deref() |
| 457 | .map(sanitize_persisted_route_label); |
| 458 | route.billing_surface = route |
| 459 | .billing_surface |
| 460 | .as_deref() |
| 461 | .map(sanitize_persisted_route_label); |
| 462 | route.endpoint_fingerprint = |
| 463 | route |
| 464 | .endpoint_fingerprint |
| 465 | .as_deref() |
| 466 | .and_then(|fingerprint| { |
| 467 | let fingerprint = fingerprint.trim(); |
| 468 | (fingerprint.len() == 64 |
| 469 | && fingerprint.bytes().all(|byte| byte.is_ascii_hexdigit())) |
| 470 | .then(|| fingerprint.to_ascii_lowercase()) |
| 471 | }); |
| 472 | let quote_is_valid = route |
| 473 | .provider_live_pricing |
| 474 | .as_ref() |
| 475 | .zip(route.endpoint_fingerprint.as_deref()) |
| 476 | .and_then(|(quote, fingerprint)| { |
| 477 | u64::try_from(route.dispatched_at.timestamp()) |
| 478 | .ok() |
| 479 | .and_then(|dispatched_at_unix| { |
| 480 | quote.pricing_for_route( |
| 481 | route.provider, |
| 482 | &route.provider_identity, |
| 483 | &route.model, |
| 484 | fingerprint, |
| 485 | dispatched_at_unix, |
| 486 | ) |
| 487 | }) |
| 488 | }) |
| 489 | .is_some(); |
| 490 | if !quote_is_valid { |
| 491 | route.provider_live_pricing = None; |
| 492 | } |
| 493 | route |
| 494 | } |
| 495 | } |
| 496 | |
| 497 | fn receipt_with_usage_classes(mut receipt: String, usage: &Usage) -> String { |
| 498 | let classes = crate::pricing::token_usage_for_pricing(usage); |
| 499 | if classes.cache_write > 0 { |
| 500 | receipt.push_str(" cache_write=yes"); |
| 501 | } |
| 502 | if usage.reasoning_tokens.unwrap_or(0) > 0 { |
| 503 | receipt.push_str(" reasoning=yes"); |
| 504 | } |
| 505 | receipt |
| 506 | } |
| 507 | |
| 508 | /// Canonical redacted route receipt for one exact usage payload. |
| 509 | #[must_use] |
| 510 | pub fn effective_route_usage_receipt( |
| 511 | route: &EffectiveRouteEnvelope, |
| 512 | audit: &TurnCostAudit, |
| 513 | usage: &Usage, |
| 514 | ) -> String { |
| 515 | receipt_with_usage_classes(route.receipt(audit), usage) |
| 516 | } |
| 517 | |
| 518 | /// Canonical `child_*` token and route metadata for tools that make their own |
| 519 | /// LLM calls (`review`, `verify`, and `rlm`). Keeping this next to the immutable |
| 520 | /// route envelope prevents the pure model types from depending on app config. |
| 521 | #[must_use] |
| 522 | pub fn child_usage_metadata_fields( |
| 523 | route: &EffectiveRouteEnvelope, |
| 524 | usage: &Usage, |
| 525 | ) -> serde_json::Map<String, serde_json::Value> { |
| 526 | let route = route.sanitized_for_persistence(); |
| 527 | let mut fields = serde_json::Map::new(); |
| 528 | fields.insert("child_provider".into(), serde_json::json!(route.provider)); |
| 529 | fields.insert( |
| 530 | "child_provider_identity".into(), |
| 531 | serde_json::json!(route.provider_identity), |
| 532 | ); |
| 533 | fields.insert("child_model".into(), serde_json::json!(route.model)); |
| 534 | fields.insert( |
| 535 | "child_openrouter_vendor".into(), |
| 536 | serde_json::json!(route.openrouter_vendor), |
| 537 | ); |
| 538 | fields.insert( |
| 539 | "child_billing_surface".into(), |
| 540 | serde_json::json!(route.billing_surface), |
| 541 | ); |
| 542 | fields.insert( |
| 543 | "child_endpoint_fingerprint".into(), |
| 544 | serde_json::json!(route.endpoint_fingerprint), |
| 545 | ); |
| 546 | fields.insert( |
| 547 | "child_provider_live_pricing".into(), |
| 548 | serde_json::json!(route.provider_live_pricing), |
| 549 | ); |
| 550 | fields.insert( |
| 551 | "child_billing_mode".into(), |
| 552 | serde_json::json!(route.billing_mode), |
| 553 | ); |
| 554 | fields.insert( |
| 555 | "child_dispatched_at".into(), |
| 556 | serde_json::json!(route.dispatched_at), |
| 557 | ); |
| 558 | fields.insert( |
| 559 | "child_input_tokens".into(), |
| 560 | serde_json::json!(usage.input_tokens), |
| 561 | ); |
| 562 | fields.insert( |
| 563 | "child_output_tokens".into(), |
| 564 | serde_json::json!(usage.output_tokens), |
| 565 | ); |
| 566 | fields.insert( |
| 567 | "child_prompt_cache_hit_tokens".into(), |
| 568 | serde_json::json!(usage.prompt_cache_hit_tokens), |
| 569 | ); |
| 570 | fields.insert( |
| 571 | "child_prompt_cache_miss_tokens".into(), |
| 572 | serde_json::json!(usage.prompt_cache_miss_tokens), |
| 573 | ); |
| 574 | fields.insert( |
| 575 | "child_prompt_cache_write_tokens".into(), |
| 576 | serde_json::json!(usage.prompt_cache_write_tokens), |
| 577 | ); |
| 578 | // Informational: reasoning tokens are already included in output tokens. |
| 579 | fields.insert( |
| 580 | "child_reasoning_tokens".into(), |
| 581 | serde_json::json!(usage.reasoning_tokens), |
| 582 | ); |
| 583 | fields.insert( |
| 584 | "child_reasoning_replay_tokens".into(), |
| 585 | serde_json::json!(usage.reasoning_replay_tokens), |
| 586 | ); |
| 587 | fields.insert( |
| 588 | "child_server_tool_use".into(), |
| 589 | serde_json::json!(usage.server_tool_use), |
| 590 | ); |
| 591 | fields |
| 592 | } |
| 593 | |
| 594 | /// Merge canonical child usage into a tool metadata object. |
| 595 | pub fn attach_child_usage_metadata( |
| 596 | metadata: &mut serde_json::Value, |
| 597 | route: &EffectiveRouteEnvelope, |
| 598 | usage: &Usage, |
| 599 | ) { |
| 600 | if let Some(object) = metadata.as_object_mut() { |
| 601 | object.extend(child_usage_metadata_fields(route, usage)); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | /// Maximum number of distinct routed-usage segments accepted from one tool |
| 606 | /// result. RLM reserves against the same bound before dispatch, so a valid |
| 607 | /// producer never has to discard a provider receipt after doing the work. |
| 608 | pub const MAX_CHILD_USAGE_RECORDS: usize = 64; |
| 609 | |
| 610 | const CHILD_USAGE_RECORDS_KEY: &str = "child_usage_records"; |
| 611 | const CHILD_DECISION_RECEIPTS_KEY: &str = "child_decision_receipts"; |
| 612 | const CHILD_USAGE_DROP_RECORDS_KEY: &str = "child_usage_drop_records"; |
| 613 | const CHILD_USAGE_DROPPED_RECORDS_KEY: &str = "child_usage_dropped_records"; |
| 614 | |
| 615 | /// Attach a bounded batch of routed child usage to tool metadata. |
| 616 | /// |
| 617 | /// The source identity is reduced to a one-way fingerprint before metadata |
| 618 | /// can enter a transcript. Routes pass through their persistence sanitizer, |
| 619 | /// so neither a raw response id nor an endpoint/credential can hitch a ride. |
| 620 | /// New consumers prefer this batch over the legacy single `child_*` fields. |
| 621 | /// Attach a bounded batch containing both exact usage receipts and exact |
| 622 | /// provider-success/missing-usage route receipts. |
| 623 | pub fn attach_child_usage_batch_metadata( |
| 624 | metadata: &mut serde_json::Value, |
| 625 | batch: &RuntimeUsageBatch, |
| 626 | ) { |
| 627 | let Some(object) = metadata.as_object_mut() else { |
| 628 | return; |
| 629 | }; |
| 630 | object.insert( |
| 631 | CHILD_DECISION_RECEIPTS_KEY.into(), |
| 632 | serde_json::json!( |
| 633 | batch |
| 634 | .decisions |
| 635 | .iter() |
| 636 | .filter(|receipt| receipt.is_bounded()) |
| 637 | .take(MAX_CHILD_USAGE_RECORDS) |
| 638 | .map(RuntimeDecisionReceipt::sanitized) |
| 639 | .collect::<Vec<_>>() |
| 640 | ), |
| 641 | ); |
| 642 | let retained_records = batch |
| 643 | .records |
| 644 | .iter() |
| 645 | .take(MAX_CHILD_USAGE_RECORDS) |
| 646 | .map(|record| { |
| 647 | serde_json::json!({ |
| 648 | "source_id": format!( |
| 649 | "routed:{}", |
| 650 | usage_source_fingerprint(&record.source_id) |
| 651 | ), |
| 652 | "route": record.usage.route.sanitized_for_persistence(), |
| 653 | "usage": record.usage.usage, |
| 654 | }) |
| 655 | }) |
| 656 | .collect::<Vec<_>>(); |
| 657 | let remaining = MAX_CHILD_USAGE_RECORDS.saturating_sub(retained_records.len()); |
| 658 | let retained_drops = batch |
| 659 | .drop_records |
| 660 | .iter() |
| 661 | .take(remaining) |
| 662 | .map(|record| RuntimeUsageDropRecord { |
| 663 | reason: record.reason, |
| 664 | source_id: format!("routed:{}", usage_source_fingerprint(&record.source_id)), |
| 665 | route: record.route.sanitized_for_persistence(), |
| 666 | }) |
| 667 | .collect::<Vec<_>>(); |
| 668 | object.insert( |
| 669 | CHILD_USAGE_RECORDS_KEY.into(), |
| 670 | serde_json::json!(retained_records), |
| 671 | ); |
| 672 | object.insert( |
| 673 | CHILD_USAGE_DROP_RECORDS_KEY.into(), |
| 674 | serde_json::json!(retained_drops), |
| 675 | ); |
| 676 | let usage_overflow = batch.records.len().saturating_sub(MAX_CHILD_USAGE_RECORDS); |
| 677 | let decision_overflow = batch |
| 678 | .decisions |
| 679 | .len() |
| 680 | .saturating_sub(MAX_CHILD_USAGE_RECORDS) |
| 681 | .saturating_add( |
| 682 | batch |
| 683 | .decisions |
| 684 | .iter() |
| 685 | .take(MAX_CHILD_USAGE_RECORDS) |
| 686 | .filter(|r| !r.is_bounded()) |
| 687 | .count(), |
| 688 | ); |
| 689 | let dropped_records = batch |
| 690 | .dropped_records |
| 691 | .max(u64::try_from(batch.drop_records.len()).unwrap_or(u64::MAX)) |
| 692 | .saturating_add(u64::try_from(usage_overflow).unwrap_or(u64::MAX)) |
| 693 | .saturating_add(u64::try_from(decision_overflow).unwrap_or(u64::MAX)); |
| 694 | if dropped_records > 0 { |
| 695 | object.insert( |
| 696 | CHILD_USAGE_DROPPED_RECORDS_KEY.into(), |
| 697 | serde_json::json!(dropped_records), |
| 698 | ); |
| 699 | } else { |
| 700 | object.remove(CHILD_USAGE_DROPPED_RECORDS_KEY); |
| 701 | } |
| 702 | } |
| 703 | |
| 704 | /// Parse the preferred routed child-usage batch. |
| 705 | /// |
| 706 | /// `None` means the batch key was absent and callers may use the legacy |
| 707 | /// single-record parser. Once the key is present, malformed/overflow entries |
| 708 | /// are represented by `dropped_records` instead of falling back and risking a |
| 709 | /// partial subtotal being presented as complete. |
| 710 | #[must_use] |
| 711 | pub fn child_usage_records_from_metadata( |
| 712 | metadata: &serde_json::Value, |
| 713 | ) -> Option<RuntimeUsageBatch> { |
| 714 | let value = metadata.get(CHILD_USAGE_RECORDS_KEY)?; |
| 715 | let drop_values = metadata |
| 716 | .get(CHILD_USAGE_DROP_RECORDS_KEY) |
| 717 | .and_then(serde_json::Value::as_array) |
| 718 | .map(Vec::as_slice) |
| 719 | .unwrap_or_default(); |
| 720 | let declared_dropped = metadata |
| 721 | .get(CHILD_USAGE_DROPPED_RECORDS_KEY) |
| 722 | .and_then(serde_json::Value::as_u64) |
| 723 | .unwrap_or(0); |
| 724 | let Some(values) = value.as_array() else { |
| 725 | return Some(RuntimeUsageBatch { |
| 726 | decisions: Vec::new(), |
| 727 | records: Vec::new(), |
| 728 | drop_records: Vec::new(), |
| 729 | dropped_records: declared_dropped.saturating_add(1), |
| 730 | }); |
| 731 | }; |
| 732 | |
| 733 | let overflow = values.len().saturating_sub(MAX_CHILD_USAGE_RECORDS); |
| 734 | let mut batch = RuntimeUsageBatch { |
| 735 | decisions: Vec::new(), |
| 736 | records: Vec::with_capacity(values.len().min(MAX_CHILD_USAGE_RECORDS)), |
| 737 | drop_records: Vec::with_capacity(drop_values.len().min(MAX_CHILD_USAGE_RECORDS)), |
| 738 | dropped_records: declared_dropped |
| 739 | .max(u64::try_from(drop_values.len()).unwrap_or(u64::MAX)) |
| 740 | .saturating_add(u64::try_from(overflow).unwrap_or(u64::MAX)), |
| 741 | }; |
| 742 | for value in values.iter().take(MAX_CHILD_USAGE_RECORDS) { |
| 743 | let parsed = (|| { |
| 744 | let source_id = value.get("source_id")?.as_str()?; |
| 745 | let route = |
| 746 | serde_json::from_value::<EffectiveRouteEnvelope>(value.get("route")?.clone()) |
| 747 | .ok()? |
| 748 | .sanitized_for_persistence(); |
| 749 | let usage = serde_json::from_value::<Usage>(value.get("usage")?.clone()).ok()?; |
| 750 | Some(RuntimeUsageRecord { |
| 751 | // Treat metadata as an untrusted persistence boundary. A |
| 752 | // stable hash preserves idempotence without retaining the |
| 753 | // producer's raw identifier. |
| 754 | source_id: usage_source_fingerprint(source_id), |
| 755 | usage: EffectiveRouteUsage { route, usage }, |
| 756 | }) |
| 757 | })(); |
| 758 | if let Some(record) = parsed { |
| 759 | batch.records.push(record); |
| 760 | } else { |
| 761 | batch.dropped_records = batch.dropped_records.saturating_add(1); |
| 762 | } |
| 763 | } |
| 764 | let remaining = MAX_CHILD_USAGE_RECORDS.saturating_sub(batch.records.len()); |
| 765 | for value in drop_values.iter().take(remaining) { |
| 766 | let parsed = (|| { |
| 767 | let source_id = value.get("source_id")?.as_str()?; |
| 768 | let route = |
| 769 | serde_json::from_value::<EffectiveRouteEnvelope>(value.get("route")?.clone()) |
| 770 | .ok()? |
| 771 | .sanitized_for_persistence(); |
| 772 | Some(RuntimeUsageDropRecord { |
| 773 | reason: value |
| 774 | .get("reason") |
| 775 | .map(|value| serde_json::from_value(value.clone())) |
| 776 | .transpose() |
| 777 | .ok()? |
| 778 | .unwrap_or_default(), |
| 779 | source_id: usage_source_fingerprint(source_id), |
| 780 | route, |
| 781 | }) |
| 782 | })(); |
| 783 | if let Some(record) = parsed { |
| 784 | batch.drop_records.push(record); |
| 785 | } |
| 786 | // Every declared drop slot already contributes to dropped_records, |
| 787 | // including malformed entries; do not count the same gap twice. |
| 788 | } |
| 789 | if let Some(decisions) = metadata.get(CHILD_DECISION_RECEIPTS_KEY) { |
| 790 | if let Some(values) = decisions.as_array() { |
| 791 | for value in values.iter().take(MAX_CHILD_USAGE_RECORDS) { |
| 792 | let bounded = serde_json::to_vec(value).is_ok_and(|bytes| bytes.len() <= 8 * 1024); |
| 793 | let parsed = bounded |
| 794 | .then(|| serde_json::from_value::<RuntimeDecisionReceipt>(value.clone()).ok()) |
| 795 | .flatten(); |
| 796 | if let Some(receipt) = parsed.filter(RuntimeDecisionReceipt::is_bounded) { |
| 797 | batch.decisions.push(receipt.sanitized()); |
| 798 | } else { |
| 799 | batch.dropped_records = batch.dropped_records.saturating_add(1); |
| 800 | } |
| 801 | } |
| 802 | batch.dropped_records = batch.dropped_records.saturating_add( |
| 803 | u64::try_from(values.len().saturating_sub(MAX_CHILD_USAGE_RECORDS)) |
| 804 | .unwrap_or(u64::MAX), |
| 805 | ); |
| 806 | } else { |
| 807 | batch.dropped_records = batch.dropped_records.saturating_add(1); |
| 808 | } |
| 809 | } |
| 810 | Some(batch) |
| 811 | } |
| 812 | |
| 813 | /// Rehydrate the immutable route envelope emitted with child usage. Legacy or |
| 814 | /// incomplete metadata becomes an explicitly unknown route and never borrows |
| 815 | /// mutable parent-session facts. |
| 816 | #[must_use] |
| 817 | pub fn child_route_envelope_from_metadata( |
| 818 | metadata: &serde_json::Value, |
| 819 | ) -> Option<EffectiveRouteEnvelope> { |
| 820 | let model = metadata.get("child_model")?.as_str()?.to_string(); |
| 821 | let provider = metadata |
| 822 | .get("child_provider") |
| 823 | .cloned() |
| 824 | .and_then(|value| serde_json::from_value(value).ok()); |
| 825 | let provider_identity = metadata |
| 826 | .get("child_provider_identity") |
| 827 | .and_then(serde_json::Value::as_str) |
| 828 | .map(str::to_string); |
| 829 | let billing_mode = metadata |
| 830 | .get("child_billing_mode") |
| 831 | .cloned() |
| 832 | .and_then(|value| serde_json::from_value(value).ok()); |
| 833 | let dispatched_at = metadata |
| 834 | .get("child_dispatched_at") |
| 835 | .cloned() |
| 836 | .and_then(|value| serde_json::from_value(value).ok()); |
| 837 | |
| 838 | let complete = provider.is_some() |
| 839 | && provider_identity.is_some() |
| 840 | && billing_mode.is_some() |
| 841 | && dispatched_at.is_some(); |
| 842 | Some( |
| 843 | EffectiveRouteEnvelope { |
| 844 | provider: provider.unwrap_or(ProviderKind::Custom), |
| 845 | provider_identity: provider_identity.unwrap_or_else(|| "legacy-unreported".to_string()), |
| 846 | model, |
| 847 | openrouter_vendor: metadata |
| 848 | .get("child_openrouter_vendor") |
| 849 | .and_then(serde_json::Value::as_str) |
| 850 | .map(str::to_string), |
| 851 | billing_surface: metadata |
| 852 | .get("child_billing_surface") |
| 853 | .and_then(serde_json::Value::as_str) |
| 854 | .map(str::to_string), |
| 855 | endpoint_fingerprint: metadata |
| 856 | .get("child_endpoint_fingerprint") |
| 857 | .and_then(serde_json::Value::as_str) |
| 858 | .map(str::to_string), |
| 859 | provider_live_pricing: metadata |
| 860 | .get("child_provider_live_pricing") |
| 861 | .cloned() |
| 862 | .and_then(|value| serde_json::from_value(value).ok()), |
| 863 | billing_mode: billing_mode |
| 864 | .filter(|_| complete) |
| 865 | .unwrap_or(RouteBillingMode::Unknown), |
| 866 | dispatched_at: dispatched_at.unwrap_or_else(|| { |
| 867 | DateTime::<Utc>::from_timestamp(0, 0).expect("Unix epoch is representable") |
| 868 | }), |
| 869 | } |
| 870 | .sanitized_for_persistence(), |
| 871 | ) |
| 872 | } |
| 873 | |
| 874 | /// Rehydrate the complete child usage payload emitted by |
| 875 | /// [`attach_child_usage_metadata`]. The presence of a canonical child token |
| 876 | /// field is significant even when every value is zero: a zero-usage provider |
| 877 | /// call still needs a route receipt and coverage classification. |
| 878 | #[must_use] |
| 879 | pub fn child_usage_from_metadata(metadata: &serde_json::Value) -> Option<Usage> { |
| 880 | const TOKEN_FIELDS: &[&str] = &[ |
| 881 | "child_input_tokens", |
| 882 | "child_output_tokens", |
| 883 | "child_prompt_cache_hit_tokens", |
| 884 | "child_prompt_cache_miss_tokens", |
| 885 | "child_prompt_cache_write_tokens", |
| 886 | "child_reasoning_tokens", |
| 887 | "child_reasoning_replay_tokens", |
| 888 | ]; |
| 889 | if !TOKEN_FIELDS |
| 890 | .iter() |
| 891 | .any(|field| metadata.get(field).is_some()) |
| 892 | { |
| 893 | return None; |
| 894 | } |
| 895 | |
| 896 | fn u32_field(metadata: &serde_json::Value, field: &str) -> Option<u32> { |
| 897 | metadata |
| 898 | .get(field) |
| 899 | .and_then(serde_json::Value::as_u64) |
| 900 | .map(|value| u32::try_from(value).unwrap_or(u32::MAX)) |
| 901 | } |
| 902 | |
| 903 | Some(Usage { |
| 904 | input_tokens: u32_field(metadata, "child_input_tokens").unwrap_or(0), |
| 905 | output_tokens: u32_field(metadata, "child_output_tokens").unwrap_or(0), |
| 906 | prompt_cache_hit_tokens: u32_field(metadata, "child_prompt_cache_hit_tokens"), |
| 907 | prompt_cache_miss_tokens: u32_field(metadata, "child_prompt_cache_miss_tokens"), |
| 908 | prompt_cache_write_tokens: u32_field(metadata, "child_prompt_cache_write_tokens"), |
| 909 | reasoning_tokens: u32_field(metadata, "child_reasoning_tokens"), |
| 910 | reasoning_replay_tokens: u32_field(metadata, "child_reasoning_replay_tokens"), |
| 911 | server_tool_use: metadata |
| 912 | .get("child_server_tool_use") |
| 913 | .cloned() |
| 914 | .and_then(|value| serde_json::from_value(value).ok()), |
| 915 | }) |
| 916 | } |
| 917 | |
| 918 | impl PendingBackgroundCost { |
| 919 | /// Whether anything at all was accrued. |
| 920 | /// |
| 921 | /// Compared against `Default` rather than checking a subset of fields, so a |
| 922 | /// field added later cannot be silently left out of the emptiness test. |
| 923 | #[must_use] |
| 924 | pub fn is_empty(&self) -> bool { |
| 925 | *self == Self::default() |
| 926 | } |
| 927 | } |
| 928 | |
| 929 | #[derive(Default)] |
| 930 | struct ScopedPendingBackgroundCost { |
| 931 | generation: u64, |
| 932 | pending: PendingBackgroundCost, |
| 933 | /// All provider responses accepted in this session generation, including |
| 934 | /// batches already drained into the live session projection. |
| 935 | seen_usage_source_fingerprints: HashSet<String>, |
| 936 | missing_usage_sources: BTreeMap<String, MissingUsageCoverage>, |
| 937 | missing_usage_overflowed: bool, |
| 938 | } |
| 939 | |
| 940 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 941 | pub struct CostScopeToken(u64); |
| 942 | |
| 943 | #[cfg(not(test))] |
| 944 | static PENDING: OnceLock<Mutex<ScopedPendingBackgroundCost>> = OnceLock::new(); |
| 945 | |
| 946 | #[cfg(test)] |
| 947 | static TEST_PENDING: OnceLock< |
| 948 | Mutex<std::collections::HashMap<std::thread::ThreadId, ScopedPendingBackgroundCost>>, |
| 949 | > = OnceLock::new(); |
| 950 | |
| 951 | fn with_pending_state_mut<R>(f: impl FnOnce(&mut ScopedPendingBackgroundCost) -> R) -> R { |
| 952 | #[cfg(not(test))] |
| 953 | { |
| 954 | let mut pending = PENDING |
| 955 | .get_or_init(|| Mutex::new(ScopedPendingBackgroundCost::default())) |
| 956 | .lock() |
| 957 | .unwrap_or_else(|e| e.into_inner()); |
| 958 | f(&mut pending) |
| 959 | } |
| 960 | #[cfg(test)] |
| 961 | { |
| 962 | // Rust tests run concurrently. A test-local collector prevents a UI |
| 963 | // drain or successful purge in one test from stealing another test's |
| 964 | // accounting. Tokio's default test runtime is current-thread, so async |
| 965 | // helpers retain this scope across awaits. |
| 966 | let mut by_thread = TEST_PENDING |
| 967 | .get_or_init(|| Mutex::new(std::collections::HashMap::new())) |
| 968 | .lock() |
| 969 | .unwrap_or_else(|e| e.into_inner()); |
| 970 | f(by_thread.entry(test_cost_scope_id()).or_default()) |
| 971 | } |
| 972 | } |
| 973 | |
| 974 | /// Runtime accounting gets a cloned, owner-scoped copy of compaction usage. |
| 975 | /// This journal is deliberately separate from the TUI pending-money pool: |
| 976 | /// taking one runtime owner's records cannot steal or reset the foreground |
| 977 | /// session's `/cost` state. |
| 978 | const MAX_RUNTIME_USAGE_RECORDS_PER_OWNER: usize = 64; |
| 979 | |
| 980 | #[derive(Default)] |
| 981 | struct OwnerRuntimeUsageJournal { |
| 982 | records: VecDeque<RuntimeUsageRecord>, |
| 983 | decisions: VecDeque<RuntimeDecisionReceipt>, |
| 984 | drop_records: VecDeque<RuntimeUsageDropRecord>, |
| 985 | dropped_records: u64, |
| 986 | dropped_source_fingerprints: HashSet<String>, |
| 987 | dropped_fingerprint_overflowed: bool, |
| 988 | } |
| 989 | |
| 990 | type RuntimeUsageJournal = HashMap<String, OwnerRuntimeUsageJournal>; |
| 991 | |
| 992 | /// Bounded fallback batch returned when no synchronous runtime sink was |
| 993 | /// available. `dropped_records` is persisted into the turn so aggregates fail |
| 994 | /// closed instead of silently presenting a partial cost as complete. |
| 995 | #[derive(Debug, Clone, Default, PartialEq, Eq)] |
| 996 | pub struct RuntimeUsageBatch { |
| 997 | /// Bounded decision evidence, owned and retired with the same usage ledger. |
| 998 | pub decisions: Vec<RuntimeDecisionReceipt>, |
| 999 | pub records: Vec<RuntimeUsageRecord>, |
| 1000 | /// Exact provider-success calls whose usage payload was absent. The |
| 1001 | /// bounded records retain route billing truth; `dropped_records` remains |
| 1002 | /// the authoritative total and may exceed this vector after overflow. |
| 1003 | pub drop_records: Vec<RuntimeUsageDropRecord>, |
| 1004 | pub dropped_records: u64, |
| 1005 | } |
| 1006 | |
| 1007 | /// One owner-scoped usage report with the stable provider-call identity used |
| 1008 | /// to make durable replay idempotent. |
| 1009 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 1010 | pub struct RuntimeUsageRecord { |
| 1011 | pub source_id: String, |
| 1012 | pub usage: EffectiveRouteUsage, |
| 1013 | } |
| 1014 | |
| 1015 | /// One provider-success response that omitted usage metadata. |
| 1016 | /// |
| 1017 | /// The frozen route is required to distinguish money-metered calls from |
| 1018 | /// subscription/local calls without consulting mutable completion-time config. |
| 1019 | #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 1020 | pub struct RuntimeUsageDropRecord { |
| 1021 | #[serde(default, skip_serializing_if = "RuntimeUsageMissingReason::is_success")] |
| 1022 | pub reason: RuntimeUsageMissingReason, |
| 1023 | pub source_id: String, |
| 1024 | pub route: EffectiveRouteEnvelope, |
| 1025 | } |
| 1026 | |
| 1027 | /// Why this dispatched request has no usable usage receipt. Legacy bytes |
| 1028 | /// mean a successful response omitted usage, never an inferred failure charge. |
| 1029 | #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 1030 | #[serde(rename_all = "snake_case")] |
| 1031 | pub enum RuntimeUsageMissingReason { |
| 1032 | #[default] |
| 1033 | SuccessWithoutUsage, |
| 1034 | RequestOutcomeUnknown, |
| 1035 | } |
| 1036 | impl RuntimeUsageMissingReason { |
| 1037 | pub fn is_success(&self) -> bool { |
| 1038 | *self == Self::SuccessWithoutUsage |
| 1039 | } |
| 1040 | pub(crate) fn label(self) -> &'static str { |
| 1041 | match self { |
| 1042 | Self::SuccessWithoutUsage => "provider_success_missing_usage", |
| 1043 | Self::RequestOutcomeUnknown => "request_outcome_unknown", |
| 1044 | } |
| 1045 | } |
| 1046 | } |
| 1047 | /// Minimal, redacted unresolved-source metadata carried by the existing cost |
| 1048 | /// snapshot. It retains no endpoint/model/secret and never infers token usage. |
| 1049 | #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 1050 | pub struct MissingUsageCoverage { |
| 1051 | pub reason: RuntimeUsageMissingReason, |
| 1052 | pub money_metered: bool, |
| 1053 | pub route_sha256: String, |
| 1054 | } |
| 1055 | pub(crate) const MAX_MISSING_USAGE_SOURCES: usize = 64; |
| 1056 | pub(crate) fn deserialize_missing_usage_sources<'de, D>( |
| 1057 | d: D, |
| 1058 | ) -> Result<BTreeMap<String, MissingUsageCoverage>, D::Error> |
| 1059 | where |
| 1060 | D: serde::Deserializer<'de>, |
| 1061 | { |
| 1062 | use serde::Deserialize as _; |
| 1063 | let values = BTreeMap::<String, MissingUsageCoverage>::deserialize(d)?; |
| 1064 | if values.len() > MAX_MISSING_USAGE_SOURCES |
| 1065 | || values |
| 1066 | .keys() |
| 1067 | .any(|key| key.len() != 64 || !key.bytes().all(|b| b.is_ascii_hexdigit())) |
| 1068 | || values.values().any(|value| { |
| 1069 | !value.route_sha256.is_empty() |
| 1070 | && (value.route_sha256.len() != 64 |
| 1071 | || !value |
| 1072 | .route_sha256 |
| 1073 | .bytes() |
| 1074 | .all(|byte| byte.is_ascii_hexdigit())) |
| 1075 | }) |
| 1076 | { |
| 1077 | return Err(serde::de::Error::custom( |
| 1078 | "unbounded or invalid missing-usage source ledger", |
| 1079 | )); |
| 1080 | } |
| 1081 | Ok(values) |
| 1082 | } |
| 1083 | impl MissingUsageCoverage { |
| 1084 | pub(crate) fn for_route( |
| 1085 | route: &EffectiveRouteEnvelope, |
| 1086 | reason: RuntimeUsageMissingReason, |
| 1087 | ) -> Self { |
| 1088 | Self { |
| 1089 | reason, |
| 1090 | money_metered: !matches!( |
| 1091 | route.billing_mode, |
| 1092 | RouteBillingMode::Subscription | RouteBillingMode::Local |
| 1093 | ), |
| 1094 | route_sha256: serde_json::to_string(&route.sanitized_for_persistence()) |
| 1095 | .map(|value| usage_source_fingerprint(&value)) |
| 1096 | .unwrap_or_default(), |
| 1097 | } |
| 1098 | } |
| 1099 | pub(crate) fn matches_route(&self, route: &EffectiveRouteEnvelope) -> bool { |
| 1100 | !self.route_sha256.is_empty() |
| 1101 | && self.route_sha256 == Self::for_route(route, self.reason).route_sha256 |
| 1102 | } |
| 1103 | } |
| 1104 | impl RuntimeUsageDropRecord { |
| 1105 | pub(crate) fn coverage(&self) -> MissingUsageCoverage { |
| 1106 | MissingUsageCoverage::for_route(&self.route, self.reason) |
| 1107 | } |
| 1108 | } |
| 1109 | /// Provider decision evidence attached to the existing origin-turn ledger. |
| 1110 | /// It is diagnostic evidence, never an instruction to change the model route. |
| 1111 | #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)] |
| 1112 | pub struct RuntimeDecisionReceipt { |
| 1113 | pub source_id: String, |
| 1114 | pub route: EffectiveRouteEnvelope, |
| 1115 | pub usage: Option<Usage>, |
| 1116 | #[serde(default)] |
| 1117 | pub usage_complete: bool, |
| 1118 | pub shadow: bool, |
| 1119 | pub valid_answers: bool, |
| 1120 | pub evidence: crate::model_routing::AutoRouteDecisionEvidence, |
| 1121 | } |
| 1122 | |
| 1123 | impl RuntimeDecisionReceipt { |
| 1124 | pub(crate) fn is_bounded(&self) -> bool { |
| 1125 | !self.source_id.trim().is_empty() |
| 1126 | && self.source_id.len() <= 128 |
| 1127 | && self.evidence.choice.len() <= 128 |
| 1128 | && self.evidence.probabilities_bp.len() <= 64 |
| 1129 | && self |
| 1130 | .evidence |
| 1131 | .probabilities_bp |
| 1132 | .iter() |
| 1133 | .all(|(key, value)| key.len() <= 128 && *value <= 10_000) |
| 1134 | && self.evidence.confidence_bp <= 10_000 |
| 1135 | && self.evidence.min_confidence_bp <= 10_000 |
| 1136 | && self |
| 1137 | .evidence |
| 1138 | .thinking |
| 1139 | .as_ref() |
| 1140 | .is_none_or(|v| v.len() <= 128) |
| 1141 | && self |
| 1142 | .evidence |
| 1143 | .response_model |
| 1144 | .as_ref() |
| 1145 | .is_none_or(|v| v.len() <= 128) |
| 1146 | && self |
| 1147 | .evidence |
| 1148 | .provider_reported_cost_usd |
| 1149 | .as_ref() |
| 1150 | .is_none_or(|v| { |
| 1151 | v.len() <= 128 |
| 1152 | && v.parse::<f64>() |
| 1153 | .is_ok_and(|cost| cost.is_finite() && cost >= 0.0) |
| 1154 | }) |
| 1155 | && serde_json::to_vec(self).is_ok_and(|bytes| bytes.len() <= 8 * 1024) |
| 1156 | } |
| 1157 | |
| 1158 | pub(crate) fn sanitized(&self) -> Self { |
| 1159 | let mut receipt = self.clone(); |
| 1160 | receipt.source_id = usage_source_fingerprint(&receipt.source_id); |
| 1161 | receipt.route = receipt.route.sanitized_for_persistence(); |
| 1162 | receipt.evidence.choice = sanitize_persisted_route_label(&receipt.evidence.choice); |
| 1163 | receipt.evidence.thinking = receipt |
| 1164 | .evidence |
| 1165 | .thinking |
| 1166 | .as_deref() |
| 1167 | .map(sanitize_persisted_route_label); |
| 1168 | receipt.evidence.response_model = receipt |
| 1169 | .evidence |
| 1170 | .response_model |
| 1171 | .as_deref() |
| 1172 | .map(sanitize_persisted_route_label); |
| 1173 | receipt.evidence.probabilities_bp = receipt |
| 1174 | .evidence |
| 1175 | .probabilities_bp |
| 1176 | .into_iter() |
| 1177 | .map(|(key, value)| (sanitize_persisted_route_label(&key), value)) |
| 1178 | .collect(); |
| 1179 | receipt |
| 1180 | } |
| 1181 | |
| 1182 | pub(crate) fn diagnostic_receipt(&self) -> String { |
| 1183 | // Reuse cost_status's persisted diagnostic receipt surface. No prompt, |
| 1184 | // response id, URL, auth header or unoffered choice enters this string. |
| 1185 | format!( |
| 1186 | "decision:{}", |
| 1187 | serde_json::to_string(&self.sanitized()).unwrap_or_default() |
| 1188 | ) |
| 1189 | } |
| 1190 | } |
| 1191 | |
| 1192 | pub(crate) type RuntimeDecisionSink = Arc<dyn Fn(RuntimeDecisionReceipt) -> bool + Send + Sync>; |
| 1193 | |
| 1194 | pub(crate) type RuntimeUsageSink = Arc<dyn Fn(RuntimeUsageRecord) -> bool + Send + Sync>; |
| 1195 | pub(crate) type RuntimeUsageDropSink = Arc<dyn Fn(RuntimeUsageDropRecord) -> bool + Send + Sync>; |
| 1196 | |
| 1197 | struct RuntimeUsageSinkEntry { |
| 1198 | sink: RuntimeUsageSink, |
| 1199 | dropped_sink: Option<RuntimeUsageDropSink>, |
| 1200 | decision_sink: Option<RuntimeDecisionSink>, |
| 1201 | leases: usize, |
| 1202 | terminal: bool, |
| 1203 | } |
| 1204 | |
| 1205 | #[cfg(test)] |
| 1206 | pub(crate) fn decision_receipt_fixture(source_id: &str) -> RuntimeDecisionReceipt { |
| 1207 | let mut route = EffectiveRouteEnvelope::capture( |
| 1208 | None, |
| 1209 | crate::config::ProviderKind::Custom, |
| 1210 | "typesafe", |
| 1211 | "jev-latest", |
| 1212 | Some("https://api.typesafe.ai/v1"), |
| 1213 | Utc::now(), |
| 1214 | ); |
| 1215 | route.billing_mode = RouteBillingMode::Unknown; |
| 1216 | RuntimeDecisionReceipt { |
| 1217 | source_id: source_id.to_string(), |
| 1218 | route, |
| 1219 | usage: Some(Usage { |
| 1220 | input_tokens: 9, |
| 1221 | output_tokens: 4, |
| 1222 | ..Default::default() |
| 1223 | }), |
| 1224 | usage_complete: true, |
| 1225 | shadow: true, |
| 1226 | valid_answers: false, |
| 1227 | evidence: crate::model_routing::AutoRouteDecisionEvidence { |
| 1228 | choice: "invalid".to_string(), |
| 1229 | probabilities_bp: Default::default(), |
| 1230 | confidence_bp: 0, |
| 1231 | min_confidence_bp: 5_000, |
| 1232 | cost_saving_kept_fast: false, |
| 1233 | thinking: None, |
| 1234 | provider_reported_cost_usd: Some("0.000012054".to_string()), |
| 1235 | latency_ms: 12, |
| 1236 | response_model: Some("jev-latest".to_string()), |
| 1237 | }, |
| 1238 | } |
| 1239 | } |
| 1240 | |
| 1241 | /// Keeps an owner sink alive while a detached child can still report usage. |
| 1242 | /// The runtime turn may already be terminal; the last child release retires |
| 1243 | /// the sink only after its final provider response has been durably appended. |
| 1244 | #[derive(Debug)] |
| 1245 | pub(crate) struct RuntimeUsageLease { |
| 1246 | owner: String, |
| 1247 | active: bool, |
| 1248 | } |
| 1249 | |
| 1250 | #[cfg(not(test))] |
| 1251 | static RUNTIME_USAGE_JOURNAL: OnceLock<Mutex<RuntimeUsageJournal>> = OnceLock::new(); |
| 1252 | |
| 1253 | #[cfg(test)] |
| 1254 | static TEST_RUNTIME_USAGE_JOURNAL: OnceLock< |
| 1255 | Mutex<std::collections::HashMap<std::thread::ThreadId, RuntimeUsageJournal>>, |
| 1256 | > = OnceLock::new(); |
| 1257 | |
| 1258 | #[cfg(not(test))] |
| 1259 | static RUNTIME_USAGE_SINKS: OnceLock<Mutex<HashMap<String, RuntimeUsageSinkEntry>>> = |
| 1260 | OnceLock::new(); |
| 1261 | |
| 1262 | /// Sinks are keyed by owner id, and owner ids in tests are short fixture |
| 1263 | /// strings that repeat across tests. Under the default parallel test harness a |
| 1264 | /// process-global map let one test's `register_runtime_usage_sink` replace |
| 1265 | /// another's live sink, and let one test's `finish_runtime_usage_owner` retire |
| 1266 | /// it — turning exactly-once child accounting into an order-dependent race. |
| 1267 | /// Scoping by thread matches the pending-cost pool and the runtime journal, |
| 1268 | /// which are already thread-scoped for the same reason. |
| 1269 | #[cfg(test)] |
| 1270 | #[allow(clippy::type_complexity)] |
| 1271 | static TEST_RUNTIME_USAGE_SINKS: OnceLock< |
| 1272 | Mutex<HashMap<std::thread::ThreadId, HashMap<String, RuntimeUsageSinkEntry>>>, |
| 1273 | > = OnceLock::new(); |
| 1274 | |
| 1275 | /// Run `f` against this scope's sink registry. |
| 1276 | fn with_runtime_usage_sinks<R>( |
| 1277 | f: impl FnOnce(&mut HashMap<String, RuntimeUsageSinkEntry>) -> R, |
| 1278 | ) -> R { |
| 1279 | #[cfg(not(test))] |
| 1280 | { |
| 1281 | let mut sinks = RUNTIME_USAGE_SINKS |
| 1282 | .get_or_init(|| Mutex::new(HashMap::new())) |
| 1283 | .lock() |
| 1284 | .unwrap_or_else(|error| error.into_inner()); |
| 1285 | f(&mut sinks) |
| 1286 | } |
| 1287 | #[cfg(test)] |
| 1288 | { |
| 1289 | let mut by_thread = TEST_RUNTIME_USAGE_SINKS |
| 1290 | .get_or_init(|| Mutex::new(HashMap::new())) |
| 1291 | .lock() |
| 1292 | .unwrap_or_else(|error| error.into_inner()); |
| 1293 | f(by_thread.entry(test_cost_scope_id()).or_default()) |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | /// Like [`with_runtime_usage_sinks`], but does not create the registry when it |
| 1298 | /// has never been initialized. Used on drop paths, where allocating a registry |
| 1299 | /// to then find it empty would be pointless. |
| 1300 | fn with_existing_runtime_usage_sinks<R>( |
| 1301 | f: impl FnOnce(&mut HashMap<String, RuntimeUsageSinkEntry>) -> R, |
| 1302 | ) -> Option<R> { |
| 1303 | #[cfg(not(test))] |
| 1304 | { |
| 1305 | let sinks = RUNTIME_USAGE_SINKS.get()?; |
| 1306 | let mut sinks = sinks.lock().unwrap_or_else(|error| error.into_inner()); |
| 1307 | Some(f(&mut sinks)) |
| 1308 | } |
| 1309 | #[cfg(test)] |
| 1310 | { |
| 1311 | let by_thread = TEST_RUNTIME_USAGE_SINKS.get()?; |
| 1312 | let mut by_thread = by_thread.lock().unwrap_or_else(|error| error.into_inner()); |
| 1313 | let sinks = by_thread.get_mut(&test_cost_scope_id())?; |
| 1314 | Some(f(sinks)) |
| 1315 | } |
| 1316 | } |
| 1317 | |
| 1318 | fn with_runtime_usage_journal_mut<R>(f: impl FnOnce(&mut RuntimeUsageJournal) -> R) -> R { |
| 1319 | #[cfg(not(test))] |
| 1320 | { |
| 1321 | let mut journal = RUNTIME_USAGE_JOURNAL |
| 1322 | .get_or_init(|| Mutex::new(HashMap::new())) |
| 1323 | .lock() |
| 1324 | .unwrap_or_else(|error| error.into_inner()); |
| 1325 | f(&mut journal) |
| 1326 | } |
| 1327 | #[cfg(test)] |
| 1328 | { |
| 1329 | let mut by_thread = TEST_RUNTIME_USAGE_JOURNAL |
| 1330 | .get_or_init(|| Mutex::new(std::collections::HashMap::new())) |
| 1331 | .lock() |
| 1332 | .unwrap_or_else(|error| error.into_inner()); |
| 1333 | f(by_thread.entry(test_cost_scope_id()).or_default()) |
| 1334 | } |
| 1335 | } |
| 1336 | |
| 1337 | fn record_runtime_usage( |
| 1338 | owner: &str, |
| 1339 | source_id: &str, |
| 1340 | route: &EffectiveRouteEnvelope, |
| 1341 | usage: &Usage, |
| 1342 | ) { |
| 1343 | if usage == &Usage::default() { |
| 1344 | record_runtime_usage_drop(owner, source_id, route); |
| 1345 | return; |
| 1346 | } |
| 1347 | let owner = owner.trim(); |
| 1348 | if owner.is_empty() { |
| 1349 | return; |
| 1350 | } |
| 1351 | let record = RuntimeUsageRecord { |
| 1352 | source_id: source_id.to_string(), |
| 1353 | usage: EffectiveRouteUsage { |
| 1354 | route: route.sanitized_for_persistence(), |
| 1355 | usage: usage.clone(), |
| 1356 | }, |
| 1357 | }; |
| 1358 | let sink = |
| 1359 | with_runtime_usage_sinks(|sinks| sinks.get(owner).map(|entry| Arc::clone(&entry.sink))); |
| 1360 | if sink.is_some_and(|sink| sink(record.clone())) { |
| 1361 | return; |
| 1362 | } |
| 1363 | with_runtime_usage_journal_mut(|journal| { |
| 1364 | let owner_journal = journal.entry(owner.to_string()).or_default(); |
| 1365 | if owner_journal.records.iter().any(|old| { |
| 1366 | usage_source_fingerprint(&old.source_id) == usage_source_fingerprint(source_id) |
| 1367 | }) { |
| 1368 | return; |
| 1369 | } |
| 1370 | let fingerprint = usage_source_fingerprint(source_id); |
| 1371 | if let Some(index) = owner_journal |
| 1372 | .drop_records |
| 1373 | .iter() |
| 1374 | .position(|old| usage_source_fingerprint(&old.source_id) == fingerprint) |
| 1375 | { |
| 1376 | if owner_journal.drop_records[index].route != record.usage.route { |
| 1377 | return; |
| 1378 | } |
| 1379 | owner_journal.drop_records.remove(index); |
| 1380 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_sub(1); |
| 1381 | owner_journal |
| 1382 | .dropped_source_fingerprints |
| 1383 | .remove(&fingerprint); |
| 1384 | } |
| 1385 | if owner_journal.records.len() == MAX_RUNTIME_USAGE_RECORDS_PER_OWNER { |
| 1386 | owner_journal.records.pop_front(); |
| 1387 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_add(1); |
| 1388 | } |
| 1389 | owner_journal.records.push_back(record); |
| 1390 | }); |
| 1391 | } |
| 1392 | |
| 1393 | fn record_runtime_usage_drop(owner: &str, source_id: &str, route: &EffectiveRouteEnvelope) { |
| 1394 | record_runtime_usage_missing( |
| 1395 | owner, |
| 1396 | source_id, |
| 1397 | route, |
| 1398 | RuntimeUsageMissingReason::SuccessWithoutUsage, |
| 1399 | ); |
| 1400 | } |
| 1401 | fn record_runtime_usage_missing( |
| 1402 | owner: &str, |
| 1403 | source_id: &str, |
| 1404 | route: &EffectiveRouteEnvelope, |
| 1405 | reason: RuntimeUsageMissingReason, |
| 1406 | ) { |
| 1407 | let owner = owner.trim(); |
| 1408 | if owner.is_empty() { |
| 1409 | return; |
| 1410 | } |
| 1411 | let fingerprint = usage_source_fingerprint(source_id); |
| 1412 | let sink = with_runtime_usage_sinks(|sinks| { |
| 1413 | sinks |
| 1414 | .get(owner) |
| 1415 | .and_then(|entry| entry.dropped_sink.as_ref().map(Arc::clone)) |
| 1416 | }); |
| 1417 | let record = RuntimeUsageDropRecord { |
| 1418 | reason, |
| 1419 | source_id: source_id.to_string(), |
| 1420 | route: route.sanitized_for_persistence(), |
| 1421 | }; |
| 1422 | if sink.is_some_and(|sink| sink(record.clone())) { |
| 1423 | return; |
| 1424 | } |
| 1425 | with_runtime_usage_journal_mut(|journal| { |
| 1426 | let owner_journal = journal.entry(owner.to_string()).or_default(); |
| 1427 | if owner_journal |
| 1428 | .records |
| 1429 | .iter() |
| 1430 | .any(|old| usage_source_fingerprint(&old.source_id) == fingerprint) |
| 1431 | { |
| 1432 | return; |
| 1433 | } |
| 1434 | if owner_journal |
| 1435 | .dropped_source_fingerprints |
| 1436 | .contains(&fingerprint) |
| 1437 | { |
| 1438 | return; |
| 1439 | } |
| 1440 | if owner_journal.dropped_source_fingerprints.len() < MAX_RUNTIME_USAGE_RECORDS_PER_OWNER { |
| 1441 | owner_journal |
| 1442 | .dropped_source_fingerprints |
| 1443 | .insert(fingerprint); |
| 1444 | owner_journal.drop_records.push_back(record); |
| 1445 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_add(1); |
| 1446 | } else if !owner_journal.dropped_fingerprint_overflowed { |
| 1447 | // Preserve a bounded fail-closed overflow marker. Once the exact |
| 1448 | // identity ledger is full, further unknown ids share this one |
| 1449 | // marker so replays cannot grow the count without bound. |
| 1450 | owner_journal.dropped_fingerprint_overflowed = true; |
| 1451 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_add(1); |
| 1452 | } |
| 1453 | }); |
| 1454 | } |
| 1455 | |
| 1456 | fn record_runtime_usage_drop_count(owner: &str, source_id: &str, count: u64) { |
| 1457 | let owner = owner.trim(); |
| 1458 | if owner.is_empty() || count == 0 { |
| 1459 | return; |
| 1460 | } |
| 1461 | let fingerprint = usage_source_fingerprint(source_id); |
| 1462 | with_runtime_usage_journal_mut(|journal| { |
| 1463 | let owner_journal = journal.entry(owner.to_string()).or_default(); |
| 1464 | if owner_journal |
| 1465 | .dropped_source_fingerprints |
| 1466 | .contains(&fingerprint) |
| 1467 | { |
| 1468 | return; |
| 1469 | } |
| 1470 | if owner_journal.dropped_source_fingerprints.len() < MAX_RUNTIME_USAGE_RECORDS_PER_OWNER { |
| 1471 | owner_journal |
| 1472 | .dropped_source_fingerprints |
| 1473 | .insert(fingerprint); |
| 1474 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_add(count); |
| 1475 | } else if !owner_journal.dropped_fingerprint_overflowed { |
| 1476 | owner_journal.dropped_fingerprint_overflowed = true; |
| 1477 | owner_journal.dropped_records = owner_journal.dropped_records.saturating_add(1); |
| 1478 | } |
| 1479 | }); |
| 1480 | } |
| 1481 | |
| 1482 | /// Install a synchronous durability sink for one active runtime turn. |
| 1483 | /// Compaction calls invoke this before they return to the engine, so a process |
| 1484 | /// crash cannot erase already-reported usage from an in-memory journal. |
| 1485 | #[cfg(test)] |
| 1486 | pub(crate) fn register_runtime_usage_sink(owner: &str, sink: RuntimeUsageSink) { |
| 1487 | register_runtime_usage_sink_with_drop(owner, sink, None); |
| 1488 | } |
| 1489 | |
| 1490 | pub(crate) fn register_runtime_usage_sink_with_drop( |
| 1491 | owner: &str, |
| 1492 | sink: RuntimeUsageSink, |
| 1493 | dropped_sink: Option<RuntimeUsageDropSink>, |
| 1494 | ) { |
| 1495 | let owner = owner.trim(); |
| 1496 | if owner.is_empty() { |
| 1497 | return; |
| 1498 | } |
| 1499 | with_runtime_usage_sinks(|sinks| { |
| 1500 | sinks.insert( |
| 1501 | owner.to_string(), |
| 1502 | RuntimeUsageSinkEntry { |
| 1503 | sink, |
| 1504 | dropped_sink, |
| 1505 | decision_sink: None, |
| 1506 | leases: 0, |
| 1507 | terminal: false, |
| 1508 | }, |
| 1509 | ); |
| 1510 | }); |
| 1511 | } |
| 1512 | |
| 1513 | /// Extend the existing owner entry; this is not another sink registry. |
| 1514 | pub(crate) fn register_runtime_decision_sink(owner: &str, sink: RuntimeDecisionSink) { |
| 1515 | with_runtime_usage_sinks(|sinks| { |
| 1516 | if let Some(entry) = sinks.get_mut(owner.trim()) { |
| 1517 | entry.decision_sink = Some(sink); |
| 1518 | } |
| 1519 | }); |
| 1520 | } |
| 1521 | |
| 1522 | fn record_runtime_decision(owner: &str, record: &RuntimeDecisionReceipt) { |
| 1523 | if !record.is_bounded() { |
| 1524 | record_runtime_usage_drop(owner, &record.source_id, &record.route); |
| 1525 | return; |
| 1526 | } |
| 1527 | let record = record.sanitized(); |
| 1528 | let sink = with_runtime_usage_sinks(|sinks| { |
| 1529 | sinks |
| 1530 | .get(owner) |
| 1531 | .and_then(|entry| entry.decision_sink.clone()) |
| 1532 | }); |
| 1533 | if sink.is_some_and(|sink| sink(record.clone())) { |
| 1534 | return; |
| 1535 | } |
| 1536 | with_runtime_usage_journal_mut(|journal| { |
| 1537 | let entry = journal.entry(owner.to_string()).or_default(); |
| 1538 | if entry |
| 1539 | .decisions |
| 1540 | .iter() |
| 1541 | .any(|old| old.source_id == record.source_id) |
| 1542 | { |
| 1543 | return; |
| 1544 | } |
| 1545 | if entry.decisions.len() == MAX_RUNTIME_USAGE_RECORDS_PER_OWNER { |
| 1546 | entry.decisions.pop_front(); |
| 1547 | entry.dropped_records = entry.dropped_records.saturating_add(1); |
| 1548 | } |
| 1549 | entry.decisions.push_back(record); |
| 1550 | }); |
| 1551 | } |
| 1552 | |
| 1553 | fn report_interactive_decision(scope: CostScopeToken, receipt: &RuntimeDecisionReceipt) { |
| 1554 | if !receipt.is_bounded() { |
| 1555 | return; |
| 1556 | } |
| 1557 | with_pending_state_mut(|state| { |
| 1558 | if state.generation == scope.0 { |
| 1559 | if state |
| 1560 | .pending |
| 1561 | .route_receipts |
| 1562 | .iter() |
| 1563 | .filter(|v| v.starts_with("decision:")) |
| 1564 | .count() |
| 1565 | < MAX_RUNTIME_USAGE_RECORDS_PER_OWNER |
| 1566 | { |
| 1567 | state |
| 1568 | .pending |
| 1569 | .route_receipts |
| 1570 | .insert(receipt.diagnostic_receipt()); |
| 1571 | } else { |
| 1572 | state |
| 1573 | .pending |
| 1574 | .route_receipts |
| 1575 | .insert("decision:diagnostic_receipt_bound_reached".to_string()); |
| 1576 | } |
| 1577 | } |
| 1578 | }); |
| 1579 | } |
| 1580 | |
| 1581 | /// Redacted durable identity shared by runtime-turn, worker, and interactive |
| 1582 | /// session accounting. Raw response ids never need to be persisted merely to |
| 1583 | /// make replay idempotent. |
| 1584 | #[must_use] |
| 1585 | pub(crate) fn usage_source_fingerprint(source_id: &str) -> String { |
| 1586 | let source_id = source_id.trim(); |
| 1587 | let fingerprint = source_id.strip_prefix("routed:").unwrap_or(source_id); |
| 1588 | if fingerprint.len() == 64 && fingerprint.bytes().all(|byte| byte.is_ascii_hexdigit()) { |
| 1589 | return fingerprint.to_ascii_lowercase(); |
| 1590 | } |
| 1591 | codewhale_config::catalog::base_url_fingerprint(source_id) |
| 1592 | } |
| 1593 | |
| 1594 | /// Install the interactive session's synchronous runtime sink. A detached |
| 1595 | /// child may report after the parent mailbox has sealed; its owner lease keeps |
| 1596 | /// this sink alive, while the captured scope prevents a later session from |
| 1597 | /// inheriting the spend. |
| 1598 | #[cfg(test)] |
| 1599 | pub(crate) fn register_interactive_runtime_usage_sink(owner: &str, scope: CostScopeToken) { |
| 1600 | register_runtime_usage_sink_with_drop( |
| 1601 | owner, |
| 1602 | Arc::new(move |record| record_interactive_runtime_usage(scope, record)), |
| 1603 | Some(Arc::new(move |record| { |
| 1604 | record_interactive_runtime_usage_drop(scope, record) |
| 1605 | })), |
| 1606 | ); |
| 1607 | } |
| 1608 | |
| 1609 | /// Install an interactive sink whose stale-scope fallback is an origin-session |
| 1610 | /// sidecar. `/new` may close the foreground pool while a detached provider call |
| 1611 | /// is still running; the sidecar keeps that exact response with the old saved |
| 1612 | /// session instead of either dropping it or contaminating the new one. |
| 1613 | pub(crate) fn register_persistent_interactive_runtime_usage_sink( |
| 1614 | owner: &str, |
| 1615 | scope: CostScopeToken, |
| 1616 | session_id: &str, |
| 1617 | turn_id: &str, |
| 1618 | ) { |
| 1619 | let Ok(manager) = crate::session_manager::SessionManager::default_location() else { |
| 1620 | // With no durable origin gate, leave the owner on the bounded journal |
| 1621 | // fallback. An in-memory-only sink could keep accepting a deleted |
| 1622 | // session's responses after its directory becomes available again. |
| 1623 | return; |
| 1624 | }; |
| 1625 | register_persistent_interactive_runtime_usage_sink_at( |
| 1626 | owner, |
| 1627 | scope, |
| 1628 | session_id, |
| 1629 | turn_id, |
| 1630 | manager.sessions_dir().to_path_buf(), |
| 1631 | ); |
| 1632 | } |
| 1633 | |
| 1634 | fn register_persistent_interactive_runtime_usage_sink_at( |
| 1635 | owner: &str, |
| 1636 | scope: CostScopeToken, |
| 1637 | session_id: &str, |
| 1638 | turn_id: &str, |
| 1639 | sessions_dir: std::path::PathBuf, |
| 1640 | ) { |
| 1641 | let usage_session_id = session_id.to_string(); |
| 1642 | let usage_turn_id = turn_id.to_string(); |
| 1643 | let drop_session_id = usage_session_id.clone(); |
| 1644 | let drop_turn_id = usage_turn_id.clone(); |
| 1645 | let usage_sessions_dir = sessions_dir.clone(); |
| 1646 | let decision_sessions_dir = sessions_dir.clone(); |
| 1647 | let decision_session_id = session_id.to_string(); |
| 1648 | let decision_turn_id = turn_id.to_string(); |
| 1649 | register_runtime_usage_sink_with_drop( |
| 1650 | owner, |
| 1651 | Arc::new(move |record| { |
| 1652 | crate::session_manager::SessionManager::new(usage_sessions_dir.clone()) |
| 1653 | .map(|manager| { |
| 1654 | report_effective_route_for_interactive_origin_with_manager( |
| 1655 | scope, |
| 1656 | &usage_session_id, |
| 1657 | &usage_turn_id, |
| 1658 | &record.source_id, |
| 1659 | &record.usage.route, |
| 1660 | &record.usage.usage, |
| 1661 | &manager, |
| 1662 | ) |
| 1663 | }) |
| 1664 | .unwrap_or(false) |
| 1665 | }), |
| 1666 | Some(Arc::new(move |record| { |
| 1667 | crate::session_manager::SessionManager::new(sessions_dir.clone()) |
| 1668 | .map(|manager| { |
| 1669 | report_missing_usage_for_interactive_origin_with_manager( |
| 1670 | scope, |
| 1671 | &drop_session_id, |
| 1672 | &drop_turn_id, |
| 1673 | &record.source_id, |
| 1674 | &record.route, |
| 1675 | record.reason, |
| 1676 | &manager, |
| 1677 | ) |
| 1678 | }) |
| 1679 | .unwrap_or(false) |
| 1680 | })), |
| 1681 | ); |
| 1682 | register_runtime_decision_sink( |
| 1683 | owner, |
| 1684 | Arc::new(move |receipt| { |
| 1685 | let Ok(manager) = |
| 1686 | crate::session_manager::SessionManager::new(decision_sessions_dir.clone()) |
| 1687 | else { |
| 1688 | return false; |
| 1689 | }; |
| 1690 | let accepted = manager |
| 1691 | .persist_late_decision_receipt(&decision_session_id, &decision_turn_id, &receipt) |
| 1692 | .unwrap_or(false); |
| 1693 | if accepted { |
| 1694 | // The append may have been handled by a deletion tombstone. Only |
| 1695 | // a live origin may enter the foreground diagnostic projection. |
| 1696 | let _ = manager.with_live_session_origin(&decision_session_id, || { |
| 1697 | report_interactive_decision(scope, &receipt); |
| 1698 | true |
| 1699 | }); |
| 1700 | } |
| 1701 | accepted |
| 1702 | }), |
| 1703 | ); |
| 1704 | } |
| 1705 | |
| 1706 | #[cfg(test)] |
| 1707 | pub(crate) fn register_persistent_interactive_runtime_usage_sink_for_test( |
| 1708 | owner: &str, |
| 1709 | scope: CostScopeToken, |
| 1710 | session_id: &str, |
| 1711 | turn_id: &str, |
| 1712 | manager: &crate::session_manager::SessionManager, |
| 1713 | ) { |
| 1714 | register_persistent_interactive_runtime_usage_sink_at( |
| 1715 | owner, |
| 1716 | scope, |
| 1717 | session_id, |
| 1718 | turn_id, |
| 1719 | manager.sessions_dir().to_path_buf(), |
| 1720 | ); |
| 1721 | } |
| 1722 | |
| 1723 | /// Acquire an owner lease for a root sub-agent runtime. Runtime clones inherit |
| 1724 | /// the lease, so top-level detached children can outlive the parent mailbox |
| 1725 | /// without losing their accounting path. |
| 1726 | pub(crate) fn acquire_runtime_usage_lease(owner: &str) -> Option<RuntimeUsageLease> { |
| 1727 | let owner = owner.trim(); |
| 1728 | if owner.is_empty() { |
| 1729 | return None; |
| 1730 | } |
| 1731 | with_runtime_usage_sinks(|sinks| { |
| 1732 | let entry = sinks.get_mut(owner)?; |
| 1733 | entry.leases = entry.leases.saturating_add(1); |
| 1734 | Some(RuntimeUsageLease { |
| 1735 | owner: owner.to_string(), |
| 1736 | active: true, |
| 1737 | }) |
| 1738 | }) |
| 1739 | } |
| 1740 | |
| 1741 | impl RuntimeUsageLease { |
| 1742 | #[must_use] |
| 1743 | pub(crate) fn owner(&self) -> &str { |
| 1744 | &self.owner |
| 1745 | } |
| 1746 | } |
| 1747 | |
| 1748 | impl Clone for RuntimeUsageLease { |
| 1749 | fn clone(&self) -> Self { |
| 1750 | if self.active { |
| 1751 | let cloned = with_runtime_usage_sinks(|sinks| { |
| 1752 | sinks.get_mut(&self.owner).map(|entry| { |
| 1753 | entry.leases = entry.leases.saturating_add(1); |
| 1754 | }) |
| 1755 | }); |
| 1756 | if cloned.is_some() { |
| 1757 | return Self { |
| 1758 | owner: self.owner.clone(), |
| 1759 | active: true, |
| 1760 | }; |
| 1761 | } |
| 1762 | } |
| 1763 | Self { |
| 1764 | owner: self.owner.clone(), |
| 1765 | active: false, |
| 1766 | } |
| 1767 | } |
| 1768 | } |
| 1769 | |
| 1770 | impl Drop for RuntimeUsageLease { |
| 1771 | fn drop(&mut self) { |
| 1772 | if !self.active { |
| 1773 | return; |
| 1774 | } |
| 1775 | with_existing_runtime_usage_sinks(|sinks| { |
| 1776 | let should_remove = sinks.get_mut(&self.owner).is_some_and(|entry| { |
| 1777 | entry.leases = entry.leases.saturating_sub(1); |
| 1778 | entry.terminal && entry.leases == 0 |
| 1779 | }); |
| 1780 | if should_remove { |
| 1781 | sinks.remove(&self.owner); |
| 1782 | } |
| 1783 | }); |
| 1784 | } |
| 1785 | } |
| 1786 | |
| 1787 | /// Mark the parent turn terminal. An owner with detached children stays live |
| 1788 | /// until their cloned leases drop; owners without children retire now. |
| 1789 | pub(crate) fn finish_runtime_usage_owner(owner: &str) { |
| 1790 | with_existing_runtime_usage_sinks(|sinks| { |
| 1791 | let should_remove = sinks.get_mut(owner).is_some_and(|entry| { |
| 1792 | entry.terminal = true; |
| 1793 | entry.leases == 0 |
| 1794 | }); |
| 1795 | if should_remove { |
| 1796 | sinks.remove(owner); |
| 1797 | } |
| 1798 | }); |
| 1799 | } |
| 1800 | |
| 1801 | /// Take only the background usage assigned to one runtime turn. |
| 1802 | /// Other runtime turns and the TUI pending pool remain untouched. |
| 1803 | #[must_use] |
| 1804 | pub fn take_runtime_usage(owner: &str) -> RuntimeUsageBatch { |
| 1805 | with_runtime_usage_journal_mut(|journal| { |
| 1806 | journal |
| 1807 | .remove(owner) |
| 1808 | .map_or_else(RuntimeUsageBatch::default, |entry| RuntimeUsageBatch { |
| 1809 | decisions: entry.decisions.into_iter().collect(), |
| 1810 | records: entry.records.into_iter().collect(), |
| 1811 | drop_records: entry.drop_records.into_iter().collect(), |
| 1812 | dropped_records: entry.dropped_records, |
| 1813 | }) |
| 1814 | }) |
| 1815 | } |
| 1816 | |
| 1817 | #[cfg(test)] |
| 1818 | thread_local! { |
| 1819 | static TEST_COST_ORIGIN: std::cell::Cell<Option<std::thread::ThreadId>> = const { std::cell::Cell::new(None) }; |
| 1820 | } |
| 1821 | |
| 1822 | #[cfg(test)] |
| 1823 | pub(crate) fn test_cost_scope_id() -> std::thread::ThreadId { |
| 1824 | TEST_COST_ORIGIN |
| 1825 | .with(|scope| scope.get()) |
| 1826 | .unwrap_or_else(|| std::thread::current().id()) |
| 1827 | } |
| 1828 | |
| 1829 | #[cfg(test)] |
| 1830 | pub(crate) struct TestCostScopeBinding(Option<std::thread::ThreadId>); |
| 1831 | |
| 1832 | #[cfg(test)] |
| 1833 | pub(crate) fn bind_test_cost_scope(origin: std::thread::ThreadId) -> TestCostScopeBinding { |
| 1834 | TestCostScopeBinding(TEST_COST_ORIGIN.with(|scope| scope.replace(Some(origin)))) |
| 1835 | } |
| 1836 | |
| 1837 | #[cfg(test)] |
| 1838 | impl Drop for TestCostScopeBinding { |
| 1839 | fn drop(&mut self) { |
| 1840 | TEST_COST_ORIGIN.with(|scope| scope.set(self.0)); |
| 1841 | } |
| 1842 | } |
| 1843 | |
| 1844 | /// Capture the current session/run generation before starting a background |
| 1845 | /// provider request. The same token must be supplied when its usage returns. |
| 1846 | #[must_use] |
| 1847 | pub fn scope_token() -> CostScopeToken { |
| 1848 | with_pending_state_mut(|state| CostScopeToken(state.generation)) |
| 1849 | } |
| 1850 | |
| 1851 | /// Atomically close the current cost scope and start a fresh generation. |
| 1852 | /// Reports from old in-flight requests are rejected after this returns, so |
| 1853 | /// `/new` and session load cannot inherit another session's spend. |
| 1854 | #[must_use] |
| 1855 | pub fn close_current_scope() -> PendingBackgroundCost { |
| 1856 | with_pending_state_mut(|state| { |
| 1857 | let pending = std::mem::take(&mut state.pending); |
| 1858 | state.generation = state.generation.wrapping_add(1); |
| 1859 | state.seen_usage_source_fingerprints.clear(); |
| 1860 | state.missing_usage_sources.clear(); |
| 1861 | state.missing_usage_overflowed = false; |
| 1862 | pending |
| 1863 | }) |
| 1864 | } |
| 1865 | |
| 1866 | /// Restore the durable response identities belonging to the newly loaded |
| 1867 | /// session. Callers close the previous scope before loading, so replacing the |
| 1868 | /// set cannot make another session's usage visible here. |
| 1869 | #[cfg(test)] |
| 1870 | pub(crate) fn restore_usage_source_fingerprints(fingerprints: impl IntoIterator<Item = String>) { |
| 1871 | with_pending_state_mut(|state| { |
| 1872 | state.seen_usage_source_fingerprints = fingerprints.into_iter().collect(); |
| 1873 | }) |
| 1874 | } |
| 1875 | |
| 1876 | pub(crate) fn restore_usage_source_ledger( |
| 1877 | fingerprints: impl IntoIterator<Item = String>, |
| 1878 | missing: &BTreeMap<String, MissingUsageCoverage>, |
| 1879 | overflowed: bool, |
| 1880 | ) { |
| 1881 | with_pending_state_mut(|state| { |
| 1882 | state.missing_usage_sources = missing |
| 1883 | .iter() |
| 1884 | .take(MAX_MISSING_USAGE_SOURCES) |
| 1885 | .map(|(key, value)| (key.clone(), value.clone())) |
| 1886 | .collect(); |
| 1887 | state.missing_usage_overflowed = overflowed || missing.len() > MAX_MISSING_USAGE_SOURCES; |
| 1888 | state.seen_usage_source_fingerprints = fingerprints |
| 1889 | .into_iter() |
| 1890 | .filter(|key| !state.missing_usage_sources.contains_key(key)) |
| 1891 | .collect(); |
| 1892 | }); |
| 1893 | } |
| 1894 | |
| 1895 | /// Mark a deleted origin's response handled in its original live generation. |
| 1896 | /// This suppresses legacy mailbox fallback without putting deleted usage or |
| 1897 | /// even its fingerprint into the pending pool or a durable session snapshot. |
| 1898 | fn acknowledge_retired_usage_source(scope: CostScopeToken, source_id: &str) { |
| 1899 | with_pending_state_mut(|state| { |
| 1900 | if state.generation == scope.0 { |
| 1901 | state |
| 1902 | .seen_usage_source_fingerprints |
| 1903 | .insert(usage_source_fingerprint(source_id)); |
| 1904 | } |
| 1905 | }); |
| 1906 | } |
| 1907 | |
| 1908 | /// Whether this session generation already accepted or retired a response. |
| 1909 | /// Used by mailbox delivery to avoid pricing a response that the synchronous |
| 1910 | /// runtime sink already handled. |
| 1911 | #[must_use] |
| 1912 | pub(crate) fn usage_source_seen(source_id: &str) -> bool { |
| 1913 | let fingerprint = usage_source_fingerprint(source_id); |
| 1914 | with_pending_state_mut(|state| { |
| 1915 | state.seen_usage_source_fingerprints.contains(&fingerprint) |
| 1916 | || state.missing_usage_sources.contains_key(&fingerprint) |
| 1917 | }) |
| 1918 | } |
| 1919 | |
| 1920 | /// The non-secret identity of a background LLM call's route. |
| 1921 | /// |
| 1922 | /// Background helpers run off a bare client with no app `Config`, so they cannot |
| 1923 | /// resolve credential-derived billing. They *can* report what they actually know |
| 1924 | /// — which provider, which configured route, which wire model, which endpoint — |
| 1925 | /// and this type carries exactly that, so the pricing decision is made from |
| 1926 | /// evidence instead of from a provider name. |
| 1927 | #[derive(Debug, Clone, Copy)] |
| 1928 | #[cfg(test)] |
| 1929 | pub struct BackgroundRoute<'a> { |
| 1930 | /// Provider kind serving the call. |
| 1931 | pub provider: ProviderKind, |
| 1932 | /// Configured route identity (the `[providers.<name>]` key), when the |
| 1933 | /// caller has one. This is a user-chosen label, not a credential. |
| 1934 | pub provider_identity: Option<&'a str>, |
| 1935 | /// Wire model id as sent on the request. |
| 1936 | pub wire_model: &'a str, |
| 1937 | /// Concrete base URL the request went to, when the client exposes one. |
| 1938 | /// |
| 1939 | /// Only ever used to derive a billing-surface classification and a |
| 1940 | /// SHA-256 fingerprint; the URL itself never leaves this struct. |
| 1941 | pub base_url: Option<&'a str>, |
| 1942 | } |
| 1943 | |
| 1944 | #[cfg(test)] |
| 1945 | impl<'a> BackgroundRoute<'a> { |
| 1946 | /// A route with no endpoint information. |
| 1947 | #[must_use] |
| 1948 | pub fn new(provider: ProviderKind, wire_model: &'a str) -> Self { |
| 1949 | Self { |
| 1950 | provider, |
| 1951 | provider_identity: None, |
| 1952 | wire_model, |
| 1953 | base_url: None, |
| 1954 | } |
| 1955 | } |
| 1956 | |
| 1957 | #[must_use] |
| 1958 | pub fn with_base_url(mut self, base_url: Option<&'a str>) -> Self { |
| 1959 | self.base_url = base_url; |
| 1960 | self |
| 1961 | } |
| 1962 | |
| 1963 | /// Non-secret billing-surface classification for this endpoint. |
| 1964 | #[must_use] |
| 1965 | pub fn billing_surface(&self) -> Option<&'static str> { |
| 1966 | crate::pricing::billing_surface_for_route(self.provider, self.base_url) |
| 1967 | } |
| 1968 | |
| 1969 | /// SHA-256 fingerprint of the normalized base URL, or `None` when unknown. |
| 1970 | /// |
| 1971 | /// This is the same digest the catalog scopes live rows on, so a live |
| 1972 | /// pricing row can be proven to price *this* endpoint. |
| 1973 | #[must_use] |
| 1974 | pub fn endpoint_fingerprint(&self) -> Option<String> { |
| 1975 | self.base_url.and_then(endpoint_fingerprint) |
| 1976 | } |
| 1977 | |
| 1978 | /// Billing presentation derivable without app config. |
| 1979 | #[must_use] |
| 1980 | pub fn billing(&self) -> BillingPresentation { |
| 1981 | crate::route_billing::for_endpoint_without_config(self.provider, self.base_url) |
| 1982 | } |
| 1983 | |
| 1984 | /// A redacted, stable receipt describing this route. |
| 1985 | #[must_use] |
| 1986 | pub fn receipt(&self, currency: &str) -> String { |
| 1987 | route_receipt( |
| 1988 | self.provider, |
| 1989 | self.provider_identity, |
| 1990 | self.wire_model, |
| 1991 | self.billing_surface(), |
| 1992 | self.endpoint_fingerprint().as_deref(), |
| 1993 | self.billing().into(), |
| 1994 | currency, |
| 1995 | ) |
| 1996 | } |
| 1997 | } |
| 1998 | |
| 1999 | /// Format one redacted route receipt. |
| 2000 | /// |
| 2001 | /// Contains only: provider kind, configured route label, wire model, |
| 2002 | /// billing-surface classification, endpoint fingerprint, billing mode, and the currency the |
| 2003 | /// estimate is denominated in. It deliberately contains no URL, no credential, |
| 2004 | /// and no filesystem path, so it is safe to persist into a saved session and to |
| 2005 | /// log. This is the single formatter, so the foreground turn path and the |
| 2006 | /// background pool cannot describe the same route two different ways. |
| 2007 | #[must_use] |
| 2008 | pub fn route_receipt( |
| 2009 | provider: ProviderKind, |
| 2010 | provider_identity: Option<&str>, |
| 2011 | wire_model: &str, |
| 2012 | billing_surface: Option<&str>, |
| 2013 | endpoint_fingerprint: Option<&str>, |
| 2014 | billing_mode: RouteBillingMode, |
| 2015 | currency: &str, |
| 2016 | ) -> String { |
| 2017 | format!( |
| 2018 | "provider={} identity={} model={} surface={} endpoint_fp={} billing_mode={} currency={currency}", |
| 2019 | provider.as_str(), |
| 2020 | safe_receipt_field(provider_identity.unwrap_or("-")), |
| 2021 | safe_receipt_field(wire_model), |
| 2022 | safe_receipt_field(billing_surface.unwrap_or("unreported")), |
| 2023 | safe_receipt_field(endpoint_fingerprint.unwrap_or("unreported")), |
| 2024 | match billing_mode { |
| 2025 | RouteBillingMode::Metered => "metered", |
| 2026 | RouteBillingMode::Subscription => "subscription", |
| 2027 | RouteBillingMode::Local => "local", |
| 2028 | RouteBillingMode::Unknown => "unknown", |
| 2029 | }, |
| 2030 | ) |
| 2031 | } |
| 2032 | |
| 2033 | const MAX_RECEIPT_FIELD_CHARS: usize = 96; |
| 2034 | |
| 2035 | fn safe_receipt_field(raw: &str) -> String { |
| 2036 | let sanitized = sanitize_persisted_route_label(raw); |
| 2037 | let mut out = String::with_capacity(raw.len().min(MAX_RECEIPT_FIELD_CHARS)); |
| 2038 | let mut previous_separator = false; |
| 2039 | for ch in sanitized.chars() { |
| 2040 | if out.chars().count() >= MAX_RECEIPT_FIELD_CHARS { |
| 2041 | break; |
| 2042 | } |
| 2043 | let safe = if ch.is_alphanumeric() || matches!(ch, '.' | '_' | '-' | '/' | ':' | '+') { |
| 2044 | ch |
| 2045 | } else { |
| 2046 | '_' |
| 2047 | }; |
| 2048 | let separator = safe == '_'; |
| 2049 | if separator && previous_separator { |
| 2050 | continue; |
| 2051 | } |
| 2052 | out.push(safe); |
| 2053 | previous_separator = separator; |
| 2054 | } |
| 2055 | if out.is_empty() { "-".to_string() } else { out } |
| 2056 | } |
| 2057 | |
| 2058 | pub(crate) fn sanitize_persisted_route_label(raw: &str) -> String { |
| 2059 | const MAX_PERSISTED_ROUTE_LABEL_CHARS: usize = 256; |
| 2060 | let value = raw.trim(); |
| 2061 | let lower = value.to_ascii_lowercase(); |
| 2062 | |
| 2063 | if value.is_empty() { |
| 2064 | return "-".to_string(); |
| 2065 | } |
| 2066 | |
| 2067 | // URLs are not route labels. Endpoints have a dedicated, validated hash |
| 2068 | // field; persisting a URL here risks leaking userinfo, query credentials, |
| 2069 | // or fragments through a custom provider/model name. |
| 2070 | if value.contains("://") { |
| 2071 | return "redacted-url".to_string(); |
| 2072 | } |
| 2073 | |
| 2074 | let authorization_value = ["bearer ", "basic ", "digest ", "token ", "apikey "] |
| 2075 | .iter() |
| 2076 | .any(|scheme| lower.starts_with(scheme)) |
| 2077 | || lower.contains("authorization:") |
| 2078 | || lower.contains("proxy-authorization:"); |
| 2079 | if authorization_value { |
| 2080 | return "redacted-credential".to_string(); |
| 2081 | } |
| 2082 | |
| 2083 | // Reject credential assignments regardless of common casing or separator: |
| 2084 | // FOO_API_KEY=..., access-token:..., password = .... |
| 2085 | for (index, ch) in value.char_indices() { |
| 2086 | if !matches!(ch, '=' | ':') { |
| 2087 | continue; |
| 2088 | } |
| 2089 | let name = lower[..index] |
| 2090 | .trim() |
| 2091 | .trim_matches(|ch: char| matches!(ch, '"' | '\'' | '{' | '[' | ',')); |
| 2092 | let name = name.rsplit([' ', ',', ';']).next().unwrap_or(name); |
| 2093 | let normalized = name.replace('-', "_"); |
| 2094 | if normalized.ends_with("api_key") |
| 2095 | || normalized.ends_with("token") |
| 2096 | || normalized.ends_with("secret") |
| 2097 | || normalized.ends_with("password") |
| 2098 | || normalized.ends_with("passwd") |
| 2099 | { |
| 2100 | return "redacted-credential".to_string(); |
| 2101 | } |
| 2102 | } |
| 2103 | |
| 2104 | // Common credential token prefixes. These are intentionally checked at |
| 2105 | // word boundaries so model ids containing an incidental "sk" survive. |
| 2106 | let credential_prefix = lower |
| 2107 | .split(|ch: char| ch.is_whitespace() || matches!(ch, '=' | ':' | ',' | ';' | '"' | '\'')) |
| 2108 | .filter(|part| !part.is_empty()) |
| 2109 | .any(|part| { |
| 2110 | [ |
| 2111 | "sk-", |
| 2112 | "sk_", |
| 2113 | "rk-", |
| 2114 | "pk-", |
| 2115 | "ghp_", |
| 2116 | "gho_", |
| 2117 | "ghu_", |
| 2118 | "ghs_", |
| 2119 | "github_pat_", |
| 2120 | "hf_", |
| 2121 | "glpat-", |
| 2122 | "xoxb-", |
| 2123 | "xoxp-", |
| 2124 | "xoxa-", |
| 2125 | "akia", |
| 2126 | "aiza", |
| 2127 | "eyj", |
| 2128 | ] |
| 2129 | .iter() |
| 2130 | .any(|prefix| part.starts_with(prefix)) |
| 2131 | }); |
| 2132 | if credential_prefix { |
| 2133 | return "redacted-credential".to_string(); |
| 2134 | } |
| 2135 | |
| 2136 | let windows_absolute = value.as_bytes().get(1) == Some(&b':') |
| 2137 | && value |
| 2138 | .as_bytes() |
| 2139 | .get(2) |
| 2140 | .is_some_and(|separator| matches!(separator, b'/' | b'\\')); |
| 2141 | let contains_local_root = [ |
| 2142 | "/users/", |
| 2143 | "/volumes/", |
| 2144 | "/home/", |
| 2145 | "/private/", |
| 2146 | "\\users\\", |
| 2147 | "file://", |
| 2148 | "/.ssh/", |
| 2149 | "\\.ssh\\", |
| 2150 | ] |
| 2151 | .iter() |
| 2152 | .any(|needle| lower.contains(needle)); |
| 2153 | let looks_like_relative_path = value.contains('\\') |
| 2154 | || lower.starts_with(".ssh/") |
| 2155 | || lower.starts_with(".ssh\\") |
| 2156 | || lower.split('/').any(|segment| { |
| 2157 | matches!( |
| 2158 | segment, |
| 2159 | "." | ".." |
| 2160 | | ".ssh" |
| 2161 | | ".config" |
| 2162 | | "secrets" |
| 2163 | | "secret" |
| 2164 | | "credentials" |
| 2165 | | "credential" |
| 2166 | | "relative" |
| 2167 | | "workspace" |
| 2168 | | "tmp" |
| 2169 | ) |
| 2170 | }); |
| 2171 | if std::path::Path::new(value).is_absolute() |
| 2172 | || windows_absolute |
| 2173 | || value.starts_with("~/") |
| 2174 | || value.starts_with("./") |
| 2175 | || value.starts_with("../") |
| 2176 | || contains_local_root |
| 2177 | || looks_like_relative_path |
| 2178 | { |
| 2179 | return "redacted-local-path".to_string(); |
| 2180 | } |
| 2181 | let bounded: String = value |
| 2182 | .chars() |
| 2183 | .filter(|ch| !ch.is_control()) |
| 2184 | .take(MAX_PERSISTED_ROUTE_LABEL_CHARS) |
| 2185 | .collect(); |
| 2186 | if bounded.is_empty() { |
| 2187 | "-".to_string() |
| 2188 | } else { |
| 2189 | bounded |
| 2190 | } |
| 2191 | } |
| 2192 | |
| 2193 | /// Validate and canonicalize an endpoint before producing the cryptographic |
| 2194 | /// fingerprint persisted in a receipt. Secret-bearing/malformed URLs receive |
| 2195 | /// no fingerprint at all; userinfo, query strings, and fragments are never fed |
| 2196 | /// to the hash function. |
| 2197 | #[must_use] |
| 2198 | pub fn endpoint_fingerprint(base_url: &str) -> Option<String> { |
| 2199 | let mut parsed = reqwest::Url::parse(base_url.trim()).ok()?; |
| 2200 | if !matches!(parsed.scheme(), "http" | "https") |
| 2201 | || !parsed.username().is_empty() |
| 2202 | || parsed.password().is_some() |
| 2203 | || parsed.query().is_some() |
| 2204 | || parsed.fragment().is_some() |
| 2205 | || parsed.host_str().is_none() |
| 2206 | { |
| 2207 | return None; |
| 2208 | } |
| 2209 | parsed.set_query(None); |
| 2210 | parsed.set_fragment(None); |
| 2211 | let canonical = parsed.as_str().trim_end_matches('/'); |
| 2212 | Some(codewhale_config::catalog::base_url_fingerprint(canonical)) |
| 2213 | } |
| 2214 | |
| 2215 | /// Currency tag for a receipt, derived from authoritative currency coverage — |
| 2216 | /// not from a positive amount, because a zero-usage priced turn is still a |
| 2217 | /// valid zero in its published currency. |
| 2218 | #[must_use] |
| 2219 | pub fn currency_tag(audit: &TurnCostAudit) -> &'static str { |
| 2220 | match (audit.usd_priced, audit.cny_priced) { |
| 2221 | (true, true) => "usd+cny", |
| 2222 | (true, false) => "usd", |
| 2223 | (false, true) => "cny", |
| 2224 | (false, false) => "unpriced", |
| 2225 | } |
| 2226 | } |
| 2227 | |
| 2228 | /// Background callers report their LLM usage here. |
| 2229 | /// |
| 2230 | /// The route is priced through the same [`crate::pricing::audit_turn_cost_for_route_on_endpoint`] |
| 2231 | /// the foreground turn path uses, so a background turn cannot be counted under |
| 2232 | /// different rules than a parent turn. Adds no money when the route is exactly |
| 2233 | /// non-metered (a local runtime, an OAuth broker, a named plan endpoint), and |
| 2234 | /// counts the turn as *missing spend* whenever it is money-metered or of unknown |
| 2235 | /// basis but could not be priced — an unknown basis is never waved through as a |
| 2236 | /// subscription (#4318). |
| 2237 | #[cfg(test)] |
| 2238 | pub fn report(scope: CostScopeToken, route: &BackgroundRoute<'_>, usage: &Usage) { |
| 2239 | let billing_surface = route.billing_surface(); |
| 2240 | let fingerprint = route.endpoint_fingerprint(); |
| 2241 | let audit = crate::pricing::audit_turn_cost_for_route_on_endpoint( |
| 2242 | route.provider, |
| 2243 | route.wire_model, |
| 2244 | billing_surface, |
| 2245 | fingerprint.as_deref(), |
| 2246 | usage, |
| 2247 | chrono::Utc::now(), |
| 2248 | route.billing(), |
| 2249 | ); |
| 2250 | record(scope, route.receipt(currency_tag(&audit)), &audit, usage); |
| 2251 | } |
| 2252 | |
| 2253 | /// Report background usage to exactly one accounting owner. |
| 2254 | /// |
| 2255 | /// Runtime-owned calls go only to the durable runtime sink. Calls without a |
| 2256 | /// runtime owner belong to the interactive TUI pool. Mixing both paths would |
| 2257 | /// count one provider response twice in hosts that expose both projections. |
| 2258 | pub fn report_effective_route_for_runtime( |
| 2259 | scope: CostScopeToken, |
| 2260 | runtime_owner: Option<&str>, |
| 2261 | source_id: &str, |
| 2262 | route: &EffectiveRouteEnvelope, |
| 2263 | usage: &Usage, |
| 2264 | ) { |
| 2265 | if let Some(owner) = runtime_owner { |
| 2266 | record_runtime_usage(owner, source_id, route, usage); |
| 2267 | } else { |
| 2268 | record_interactive_runtime_usage( |
| 2269 | scope, |
| 2270 | RuntimeUsageRecord { |
| 2271 | source_id: source_id.to_string(), |
| 2272 | usage: EffectiveRouteUsage { |
| 2273 | route: route.sanitized_for_persistence(), |
| 2274 | usage: usage.clone(), |
| 2275 | }, |
| 2276 | }, |
| 2277 | ); |
| 2278 | } |
| 2279 | } |
| 2280 | |
| 2281 | /// Report an interactive auxiliary response against its immutable origin. |
| 2282 | /// A stale foreground scope is not an error: it means `/new` or session load |
| 2283 | /// already moved on, so the exact receipt is appended to the old session's |
| 2284 | /// durable sidecar instead of being redirected to the active session. |
| 2285 | pub(crate) fn report_effective_route_for_interactive_origin( |
| 2286 | scope: CostScopeToken, |
| 2287 | session_id: &str, |
| 2288 | turn_id: &str, |
| 2289 | source_id: &str, |
| 2290 | route: &EffectiveRouteEnvelope, |
| 2291 | usage: &Usage, |
| 2292 | ) { |
| 2293 | let persisted = |
| 2294 | crate::session_manager::SessionManager::default_location().is_ok_and(|manager| { |
| 2295 | report_effective_route_for_interactive_origin_with_manager( |
| 2296 | scope, session_id, turn_id, source_id, route, usage, &manager, |
| 2297 | ) |
| 2298 | }); |
| 2299 | if !persisted { |
| 2300 | tracing::warn!("late interactive usage could not be persisted for its origin session"); |
| 2301 | } |
| 2302 | } |
| 2303 | |
| 2304 | fn report_effective_route_for_interactive_origin_with_manager( |
| 2305 | scope: CostScopeToken, |
| 2306 | session_id: &str, |
| 2307 | turn_id: &str, |
| 2308 | source_id: &str, |
| 2309 | route: &EffectiveRouteEnvelope, |
| 2310 | usage: &Usage, |
| 2311 | manager: &crate::session_manager::SessionManager, |
| 2312 | ) -> bool { |
| 2313 | let record = RuntimeUsageRecord { |
| 2314 | source_id: source_id.to_string(), |
| 2315 | usage: EffectiveRouteUsage { |
| 2316 | route: route.sanitized_for_persistence(), |
| 2317 | usage: usage.clone(), |
| 2318 | }, |
| 2319 | }; |
| 2320 | match manager.with_live_session_origin(session_id, || { |
| 2321 | record_interactive_runtime_usage(scope, record.clone()) |
| 2322 | }) { |
| 2323 | Ok(None) => { |
| 2324 | acknowledge_retired_usage_source(scope, source_id); |
| 2325 | return true; |
| 2326 | } |
| 2327 | Ok(Some(true)) => return true, |
| 2328 | Ok(Some(false)) => {} |
| 2329 | Err(_) => return false, |
| 2330 | } |
| 2331 | manager |
| 2332 | .persist_late_runtime_usage(session_id, turn_id, &record) |
| 2333 | .unwrap_or(false) |
| 2334 | } |
| 2335 | |
| 2336 | pub(crate) fn report_unreceipted_for_interactive_origin( |
| 2337 | scope: CostScopeToken, |
| 2338 | session_id: &str, |
| 2339 | turn_id: &str, |
| 2340 | source_id: &str, |
| 2341 | route: &EffectiveRouteEnvelope, |
| 2342 | ) { |
| 2343 | report_missing_usage_for_interactive_origin( |
| 2344 | scope, |
| 2345 | session_id, |
| 2346 | turn_id, |
| 2347 | source_id, |
| 2348 | route, |
| 2349 | RuntimeUsageMissingReason::SuccessWithoutUsage, |
| 2350 | ); |
| 2351 | } |
| 2352 | pub(crate) fn report_missing_usage_for_interactive_origin( |
| 2353 | scope: CostScopeToken, |
| 2354 | session_id: &str, |
| 2355 | turn_id: &str, |
| 2356 | source_id: &str, |
| 2357 | route: &EffectiveRouteEnvelope, |
| 2358 | reason: RuntimeUsageMissingReason, |
| 2359 | ) { |
| 2360 | let persisted = |
| 2361 | crate::session_manager::SessionManager::default_location().is_ok_and(|manager| { |
| 2362 | report_missing_usage_for_interactive_origin_with_manager( |
| 2363 | scope, session_id, turn_id, source_id, route, reason, &manager, |
| 2364 | ) |
| 2365 | }); |
| 2366 | if !persisted { |
| 2367 | tracing::warn!( |
| 2368 | "late interactive missing-usage receipt could not be persisted for its origin session" |
| 2369 | ); |
| 2370 | } |
| 2371 | } |
| 2372 | |
| 2373 | #[cfg(test)] |
| 2374 | fn report_unreceipted_for_interactive_origin_with_manager( |
| 2375 | scope: CostScopeToken, |
| 2376 | session_id: &str, |
| 2377 | turn_id: &str, |
| 2378 | source_id: &str, |
| 2379 | route: &EffectiveRouteEnvelope, |
| 2380 | manager: &crate::session_manager::SessionManager, |
| 2381 | ) -> bool { |
| 2382 | report_missing_usage_for_interactive_origin_with_manager( |
| 2383 | scope, |
| 2384 | session_id, |
| 2385 | turn_id, |
| 2386 | source_id, |
| 2387 | route, |
| 2388 | RuntimeUsageMissingReason::SuccessWithoutUsage, |
| 2389 | manager, |
| 2390 | ) |
| 2391 | } |
| 2392 | fn report_missing_usage_for_interactive_origin_with_manager( |
| 2393 | scope: CostScopeToken, |
| 2394 | session_id: &str, |
| 2395 | turn_id: &str, |
| 2396 | source_id: &str, |
| 2397 | route: &EffectiveRouteEnvelope, |
| 2398 | reason: RuntimeUsageMissingReason, |
| 2399 | manager: &crate::session_manager::SessionManager, |
| 2400 | ) -> bool { |
| 2401 | let record = RuntimeUsageDropRecord { |
| 2402 | reason, |
| 2403 | source_id: source_id.to_string(), |
| 2404 | route: route.sanitized_for_persistence(), |
| 2405 | }; |
| 2406 | match manager.with_live_session_origin(session_id, || { |
| 2407 | record_interactive_runtime_usage_drop(scope, record.clone()) |
| 2408 | }) { |
| 2409 | Ok(None) => { |
| 2410 | acknowledge_retired_usage_source(scope, source_id); |
| 2411 | return true; |
| 2412 | } |
| 2413 | Ok(Some(true)) => return true, |
| 2414 | Ok(Some(false)) => {} |
| 2415 | Err(_) => return false, |
| 2416 | } |
| 2417 | manager |
| 2418 | .persist_late_runtime_drop(session_id, turn_id, &record) |
| 2419 | .unwrap_or(false) |
| 2420 | } |
| 2421 | |
| 2422 | /// Record one provider-success response whose usage payload was absent. |
| 2423 | /// |
| 2424 | /// Callers must supply the same fixed-length, non-secret source identity they |
| 2425 | /// would use for a normal routed usage receipt. Runtime owners persist one |
| 2426 | /// bounded dropped-coverage marker; ownerless/interactive calls add one |
| 2427 | /// unpriced coverage turn to the captured session scope. Replays are |
| 2428 | /// idempotent, and a stale scope cannot contaminate a later session. |
| 2429 | pub(crate) fn report_unreceipted_provider_success( |
| 2430 | scope: CostScopeToken, |
| 2431 | runtime_owner: Option<&str>, |
| 2432 | source_id: &str, |
| 2433 | route: &EffectiveRouteEnvelope, |
| 2434 | ) { |
| 2435 | report_missing_runtime_usage( |
| 2436 | scope, |
| 2437 | runtime_owner, |
| 2438 | source_id, |
| 2439 | route, |
| 2440 | RuntimeUsageMissingReason::SuccessWithoutUsage, |
| 2441 | ); |
| 2442 | } |
| 2443 | pub(crate) fn report_missing_runtime_usage( |
| 2444 | scope: CostScopeToken, |
| 2445 | runtime_owner: Option<&str>, |
| 2446 | source_id: &str, |
| 2447 | route: &EffectiveRouteEnvelope, |
| 2448 | reason: RuntimeUsageMissingReason, |
| 2449 | ) { |
| 2450 | if let Some(owner) = runtime_owner { |
| 2451 | record_runtime_usage_missing(owner, source_id, route, reason); |
| 2452 | } else { |
| 2453 | record_interactive_runtime_usage_drop( |
| 2454 | scope, |
| 2455 | RuntimeUsageDropRecord { |
| 2456 | reason, |
| 2457 | source_id: source_id.to_string(), |
| 2458 | route: route.sanitized_for_persistence(), |
| 2459 | }, |
| 2460 | ); |
| 2461 | } |
| 2462 | } |
| 2463 | |
| 2464 | /// Settle one bounded routed-usage batch without repricing or losing exact |
| 2465 | /// missing-usage route evidence. Replaying the same batch is idempotent by the |
| 2466 | /// stable per-response source ids. Any residual count whose exact record was |
| 2467 | /// truncated remains an explicit fail-closed coverage gap. |
| 2468 | pub(crate) fn report_runtime_usage_batch( |
| 2469 | scope: CostScopeToken, |
| 2470 | runtime_owner: Option<&str>, |
| 2471 | batch: &RuntimeUsageBatch, |
| 2472 | ) { |
| 2473 | for receipt in &batch.decisions { |
| 2474 | if let Some(owner) = runtime_owner { |
| 2475 | record_runtime_decision(owner, receipt); |
| 2476 | } else { |
| 2477 | report_interactive_decision(scope, receipt); |
| 2478 | } |
| 2479 | } |
| 2480 | for record in &batch.records { |
| 2481 | report_effective_route_for_runtime( |
| 2482 | scope, |
| 2483 | runtime_owner, |
| 2484 | &record.source_id, |
| 2485 | &record.usage.route, |
| 2486 | &record.usage.usage, |
| 2487 | ); |
| 2488 | } |
| 2489 | for record in &batch.drop_records { |
| 2490 | report_missing_runtime_usage( |
| 2491 | scope, |
| 2492 | runtime_owner, |
| 2493 | &record.source_id, |
| 2494 | &record.route, |
| 2495 | record.reason, |
| 2496 | ); |
| 2497 | } |
| 2498 | |
| 2499 | let residual = batch |
| 2500 | .dropped_records |
| 2501 | .saturating_sub(u64::try_from(batch.drop_records.len()).unwrap_or(u64::MAX)); |
| 2502 | if residual == 0 { |
| 2503 | return; |
| 2504 | } |
| 2505 | let mut identities = batch |
| 2506 | .records |
| 2507 | .iter() |
| 2508 | .map(|record| usage_source_fingerprint(&record.source_id)) |
| 2509 | .chain( |
| 2510 | batch |
| 2511 | .drop_records |
| 2512 | .iter() |
| 2513 | .map(|record| usage_source_fingerprint(&record.source_id)), |
| 2514 | ) |
| 2515 | .take(MAX_RUNTIME_USAGE_RECORDS_PER_OWNER) |
| 2516 | .collect::<Vec<_>>(); |
| 2517 | identities.sort_unstable(); |
| 2518 | let residual_source = format!( |
| 2519 | "runtime-usage-batch-residual:{}", |
| 2520 | usage_source_fingerprint(&format!( |
| 2521 | "{}:{}:{}:{}", |
| 2522 | batch.records.len(), |
| 2523 | batch.drop_records.len(), |
| 2524 | batch.dropped_records, |
| 2525 | identities.join(":") |
| 2526 | )) |
| 2527 | ); |
| 2528 | if let Some(owner) = runtime_owner { |
| 2529 | record_runtime_usage_drop_count(owner, &residual_source, residual); |
| 2530 | } else { |
| 2531 | record_interactive_runtime_usage_drop_count(scope, &residual_source, residual); |
| 2532 | } |
| 2533 | } |
| 2534 | |
| 2535 | #[must_use] |
| 2536 | pub(crate) fn background_cost_for_runtime_usage( |
| 2537 | record: &RuntimeUsageRecord, |
| 2538 | ) -> PendingBackgroundCost { |
| 2539 | if record.usage.usage == Usage::default() { |
| 2540 | return background_cost_for_runtime_drop(&RuntimeUsageDropRecord { |
| 2541 | reason: crate::cost_status::RuntimeUsageMissingReason::default(), |
| 2542 | source_id: record.source_id.clone(), |
| 2543 | route: record.usage.route.clone(), |
| 2544 | }); |
| 2545 | } |
| 2546 | let mut pending = PendingBackgroundCost::default(); |
| 2547 | let fingerprint = usage_source_fingerprint(&record.source_id); |
| 2548 | let audit = record.usage.route.audit(&record.usage.usage); |
| 2549 | let receipt = record.usage.route.receipt(&audit); |
| 2550 | pending |
| 2551 | .usage_source_fingerprints |
| 2552 | .insert(fingerprint.clone()); |
| 2553 | pending.resolved_missing_usage_sources.insert(fingerprint); |
| 2554 | fold_audit_into_pending(&mut pending, receipt, &audit, &record.usage.usage); |
| 2555 | pending |
| 2556 | } |
| 2557 | |
| 2558 | #[must_use] |
| 2559 | pub(crate) fn background_cost_for_runtime_drop( |
| 2560 | record: &RuntimeUsageDropRecord, |
| 2561 | ) -> PendingBackgroundCost { |
| 2562 | let mut pending = PendingBackgroundCost::default(); |
| 2563 | let fingerprint = usage_source_fingerprint(&record.source_id); |
| 2564 | pending |
| 2565 | .usage_source_fingerprints |
| 2566 | .insert(fingerprint.clone()); |
| 2567 | pending |
| 2568 | .missing_usage_sources |
| 2569 | .insert(fingerprint, record.coverage()); |
| 2570 | if record.coverage().money_metered { |
| 2571 | pending.unpriced_turns = 1; |
| 2572 | pending.cny_unpriced_turns = 1; |
| 2573 | pending.unpriced_reasons.insert(record.reason.label()); |
| 2574 | pending.cny_unpriced_reasons.insert(record.reason.label()); |
| 2575 | } |
| 2576 | pending |
| 2577 | } |
| 2578 | /// Reconcile exact missing slots in a projection before adding real usage. |
| 2579 | /// Overflow and legacy unattributed gaps are never erased by this operation. |
| 2580 | pub(crate) fn reconcile_missing_usage_sources( |
| 2581 | missing: &mut BTreeMap<String, MissingUsageCoverage>, |
| 2582 | resolved: &BTreeSet<String>, |
| 2583 | unpriced: &mut u32, |
| 2584 | cny_unpriced: &mut u32, |
| 2585 | ) { |
| 2586 | for fingerprint in resolved { |
| 2587 | if missing |
| 2588 | .remove(fingerprint) |
| 2589 | .is_some_and(|coverage| coverage.money_metered) |
| 2590 | { |
| 2591 | *unpriced = unpriced.saturating_sub(1); |
| 2592 | *cny_unpriced = cny_unpriced.saturating_sub(1); |
| 2593 | } |
| 2594 | } |
| 2595 | } |
| 2596 | |
| 2597 | /// Apply bounded unresolved metadata to an existing cost projection. The |
| 2598 | /// returned batch omits exact identities whose missing slots overflowed; a |
| 2599 | /// real receipt for such an id can still settle once, but cannot erase the gap. |
| 2600 | pub(crate) fn project_missing_usage_ledger( |
| 2601 | missing: &mut BTreeMap<String, MissingUsageCoverage>, |
| 2602 | overflowed: &mut bool, |
| 2603 | unpriced: &mut u32, |
| 2604 | cny_unpriced: &mut u32, |
| 2605 | pool: &PendingBackgroundCost, |
| 2606 | ) -> PendingBackgroundCost { |
| 2607 | reconcile_missing_usage_sources( |
| 2608 | missing, |
| 2609 | &pool.resolved_missing_usage_sources, |
| 2610 | unpriced, |
| 2611 | cny_unpriced, |
| 2612 | ); |
| 2613 | let mut pool = pool.clone(); |
| 2614 | for (fingerprint, coverage) in &pool.missing_usage_sources { |
| 2615 | if missing.contains_key(fingerprint) { |
| 2616 | continue; |
| 2617 | } |
| 2618 | if missing.len() < MAX_MISSING_USAGE_SOURCES { |
| 2619 | missing.insert(fingerprint.clone(), coverage.clone()); |
| 2620 | } else { |
| 2621 | pool.usage_source_fingerprints.remove(fingerprint); |
| 2622 | if coverage.money_metered { |
| 2623 | pool.unpriced_turns = pool.unpriced_turns.saturating_sub(1); |
| 2624 | pool.cny_unpriced_turns = pool.cny_unpriced_turns.saturating_sub(1); |
| 2625 | } |
| 2626 | if !*overflowed && !pool.missing_usage_overflowed { |
| 2627 | pool.missing_usage_overflowed = true; |
| 2628 | pool.unpriced_turns = pool.unpriced_turns.saturating_add(1); |
| 2629 | pool.cny_unpriced_turns = pool.cny_unpriced_turns.saturating_add(1); |
| 2630 | pool.unpriced_reasons |
| 2631 | .insert("missing_usage_source_overflow"); |
| 2632 | pool.cny_unpriced_reasons |
| 2633 | .insert("missing_usage_source_overflow"); |
| 2634 | } |
| 2635 | } |
| 2636 | } |
| 2637 | *overflowed |= pool.missing_usage_overflowed; |
| 2638 | pool |
| 2639 | } |
| 2640 | |
| 2641 | /// Fold one already-computed audit into the pending pool. |
| 2642 | #[cfg(test)] |
| 2643 | fn record(scope: CostScopeToken, route_receipt: String, audit: &TurnCostAudit, usage: &Usage) { |
| 2644 | with_pending_state_mut(|state| { |
| 2645 | if state.generation != scope.0 { |
| 2646 | return; |
| 2647 | } |
| 2648 | fold_audit_into_pending(&mut state.pending, route_receipt, audit, usage); |
| 2649 | }); |
| 2650 | } |
| 2651 | |
| 2652 | fn record_interactive_runtime_usage(scope: CostScopeToken, record: RuntimeUsageRecord) -> bool { |
| 2653 | if record.usage.usage == Usage::default() { |
| 2654 | return record_interactive_runtime_usage_drop( |
| 2655 | scope, |
| 2656 | RuntimeUsageDropRecord { |
| 2657 | reason: RuntimeUsageMissingReason::SuccessWithoutUsage, |
| 2658 | source_id: record.source_id, |
| 2659 | route: record.usage.route, |
| 2660 | }, |
| 2661 | ); |
| 2662 | } |
| 2663 | with_pending_state_mut(|state| { |
| 2664 | if state.generation != scope.0 { |
| 2665 | return false; |
| 2666 | } |
| 2667 | let fingerprint = usage_source_fingerprint(&record.source_id); |
| 2668 | if state |
| 2669 | .missing_usage_sources |
| 2670 | .get(&fingerprint) |
| 2671 | .is_some_and(|coverage| !coverage.matches_route(&record.usage.route)) |
| 2672 | { |
| 2673 | return true; |
| 2674 | } |
| 2675 | if !state |
| 2676 | .seen_usage_source_fingerprints |
| 2677 | .insert(fingerprint.clone()) |
| 2678 | { |
| 2679 | return true; |
| 2680 | } |
| 2681 | if let Some(coverage) = state.missing_usage_sources.remove(&fingerprint) { |
| 2682 | if state |
| 2683 | .pending |
| 2684 | .missing_usage_sources |
| 2685 | .remove(&fingerprint) |
| 2686 | .is_some() |
| 2687 | { |
| 2688 | if coverage.money_metered { |
| 2689 | state.pending.unpriced_turns = state.pending.unpriced_turns.saturating_sub(1); |
| 2690 | state.pending.cny_unpriced_turns = |
| 2691 | state.pending.cny_unpriced_turns.saturating_sub(1); |
| 2692 | } |
| 2693 | } else { |
| 2694 | state |
| 2695 | .pending |
| 2696 | .resolved_missing_usage_sources |
| 2697 | .insert(fingerprint.clone()); |
| 2698 | } |
| 2699 | } |
| 2700 | let audit = record.usage.route.audit(&record.usage.usage); |
| 2701 | let receipt = record.usage.route.receipt(&audit); |
| 2702 | state.pending.usage_source_fingerprints.insert(fingerprint); |
| 2703 | fold_audit_into_pending(&mut state.pending, receipt, &audit, &record.usage.usage); |
| 2704 | true |
| 2705 | }) |
| 2706 | } |
| 2707 | fn record_interactive_runtime_usage_drop( |
| 2708 | scope: CostScopeToken, |
| 2709 | record: RuntimeUsageDropRecord, |
| 2710 | ) -> bool { |
| 2711 | with_pending_state_mut(|state| { |
| 2712 | if state.generation != scope.0 { |
| 2713 | return false; |
| 2714 | } |
| 2715 | let fingerprint = usage_source_fingerprint(&record.source_id); |
| 2716 | if state.seen_usage_source_fingerprints.contains(&fingerprint) |
| 2717 | || state.missing_usage_sources.contains_key(&fingerprint) |
| 2718 | { |
| 2719 | return true; |
| 2720 | } |
| 2721 | if state.missing_usage_sources.len() == MAX_MISSING_USAGE_SOURCES { |
| 2722 | if !state.missing_usage_overflowed { |
| 2723 | state.missing_usage_overflowed = true; |
| 2724 | state.pending.missing_usage_overflowed = true; |
| 2725 | state.pending.unpriced_turns = state.pending.unpriced_turns.saturating_add(1); |
| 2726 | state.pending.cny_unpriced_turns = |
| 2727 | state.pending.cny_unpriced_turns.saturating_add(1); |
| 2728 | state |
| 2729 | .pending |
| 2730 | .unpriced_reasons |
| 2731 | .insert("missing_usage_source_overflow"); |
| 2732 | state |
| 2733 | .pending |
| 2734 | .cny_unpriced_reasons |
| 2735 | .insert("missing_usage_source_overflow"); |
| 2736 | } |
| 2737 | return true; |
| 2738 | } |
| 2739 | let coverage = record.coverage(); |
| 2740 | state |
| 2741 | .missing_usage_sources |
| 2742 | .insert(fingerprint.clone(), coverage.clone()); |
| 2743 | state |
| 2744 | .pending |
| 2745 | .missing_usage_sources |
| 2746 | .insert(fingerprint.clone(), coverage.clone()); |
| 2747 | state.pending.usage_source_fingerprints.insert(fingerprint); |
| 2748 | if coverage.money_metered { |
| 2749 | state.pending.unpriced_turns = state.pending.unpriced_turns.saturating_add(1); |
| 2750 | state.pending.cny_unpriced_turns = state.pending.cny_unpriced_turns.saturating_add(1); |
| 2751 | state.pending.unpriced_reasons.insert(record.reason.label()); |
| 2752 | state |
| 2753 | .pending |
| 2754 | .cny_unpriced_reasons |
| 2755 | .insert(record.reason.label()); |
| 2756 | } |
| 2757 | true |
| 2758 | }) |
| 2759 | } |
| 2760 | |
| 2761 | fn record_interactive_runtime_usage_drop_count( |
| 2762 | scope: CostScopeToken, |
| 2763 | source_id: &str, |
| 2764 | count: u64, |
| 2765 | ) -> bool { |
| 2766 | with_pending_state_mut(|state| { |
| 2767 | if state.generation != scope.0 { |
| 2768 | return false; |
| 2769 | } |
| 2770 | let fingerprint = usage_source_fingerprint(source_id); |
| 2771 | if !state |
| 2772 | .seen_usage_source_fingerprints |
| 2773 | .insert(fingerprint.clone()) |
| 2774 | { |
| 2775 | return true; |
| 2776 | } |
| 2777 | state.pending.usage_source_fingerprints.insert(fingerprint); |
| 2778 | let count = u32::try_from(count).unwrap_or(u32::MAX); |
| 2779 | state.pending.unpriced_turns = state.pending.unpriced_turns.saturating_add(count); |
| 2780 | state.pending.cny_unpriced_turns = state.pending.cny_unpriced_turns.saturating_add(count); |
| 2781 | state |
| 2782 | .pending |
| 2783 | .unpriced_reasons |
| 2784 | .insert("routed_usage_receipt_missing"); |
| 2785 | state |
| 2786 | .pending |
| 2787 | .cny_unpriced_reasons |
| 2788 | .insert("routed_usage_receipt_missing"); |
| 2789 | true |
| 2790 | }) |
| 2791 | } |
| 2792 | |
| 2793 | fn fold_audit_into_pending( |
| 2794 | pending: &mut PendingBackgroundCost, |
| 2795 | route_receipt: String, |
| 2796 | audit: &TurnCostAudit, |
| 2797 | usage: &Usage, |
| 2798 | ) { |
| 2799 | if let Some(provenance) = audit.provenance.as_ref() { |
| 2800 | pending.pricing_provenances.insert(provenance.label()); |
| 2801 | } |
| 2802 | if let Some(defect) = audit.live_pricing_defect.as_ref() { |
| 2803 | if audit.estimate.is_some() { |
| 2804 | pending.live_pricing_defects.insert(defect.label()); |
| 2805 | } else { |
| 2806 | pending.live_pricing_unusable_defects.insert(defect.label()); |
| 2807 | } |
| 2808 | } |
| 2809 | if let Some(cost) = audit.estimate { |
| 2810 | pending.estimate = pending.estimate.saturating_add(cost); |
| 2811 | } |
| 2812 | if usage.prompt_cache_hit_tokens.is_some() |
| 2813 | || usage.prompt_cache_miss_tokens.is_some() |
| 2814 | || usage.prompt_cache_write_tokens.is_some() |
| 2815 | { |
| 2816 | let classes = crate::pricing::token_usage_for_pricing(usage); |
| 2817 | let add = |slot: &mut Option<u64>, tokens: u64| { |
| 2818 | *slot = Some(slot.unwrap_or(0).saturating_add(tokens)); |
| 2819 | }; |
| 2820 | add(&mut pending.cache_hit_tokens, classes.cache_read); |
| 2821 | add(&mut pending.cache_miss_tokens, classes.input); |
| 2822 | add(&mut pending.cache_write_tokens, classes.cache_write); |
| 2823 | } |
| 2824 | |
| 2825 | // Only money-metered/unknown-basis turns belong in missing-money coverage |
| 2826 | // or its reason list. A subscription/local receipt is still audited below, |
| 2827 | // but `not_money_metered` must never be presented as a gap in a subtotal. |
| 2828 | if audit.counts_toward_money_coverage() { |
| 2829 | if audit.usd_priced { |
| 2830 | pending.priced_turns = pending.priced_turns.saturating_add(1); |
| 2831 | } else { |
| 2832 | pending.unpriced_turns = pending.unpriced_turns.saturating_add(1); |
| 2833 | } |
| 2834 | if audit.cny_priced { |
| 2835 | pending.cny_priced_turns = pending.cny_priced_turns.saturating_add(1); |
| 2836 | } else { |
| 2837 | pending.cny_unpriced_turns = pending.cny_unpriced_turns.saturating_add(1); |
| 2838 | } |
| 2839 | for class in &audit.unpriced_classes { |
| 2840 | pending.unpriced_classes.insert(class.label()); |
| 2841 | } |
| 2842 | if !audit.usd_priced |
| 2843 | && let Some(reason) = audit.unpriced_reason |
| 2844 | { |
| 2845 | pending.unpriced_reasons.insert(reason.label()); |
| 2846 | } |
| 2847 | if !audit.cny_priced { |
| 2848 | pending.cny_unpriced_reasons.insert( |
| 2849 | audit |
| 2850 | .unpriced_reason |
| 2851 | .map_or("currency_not_published", |reason| reason.label()), |
| 2852 | ); |
| 2853 | } |
| 2854 | } |
| 2855 | |
| 2856 | // Record which token classes this route actually billed on, so a receipt |
| 2857 | // shows whether cache-write/reasoning telemetry was even present. |
| 2858 | pending |
| 2859 | .route_receipts |
| 2860 | .insert(receipt_with_usage_classes(route_receipt, usage)); |
| 2861 | } |
| 2862 | |
| 2863 | /// Drain the pending pool, returning it and resetting to zero. |
| 2864 | /// |
| 2865 | /// Money and its completeness leave together, so a caller can never fold a |
| 2866 | /// subtotal into a session total without the counters that qualify it. |
| 2867 | #[must_use] |
| 2868 | pub fn drain() -> PendingBackgroundCost { |
| 2869 | with_pending_state_mut(|state| std::mem::take(&mut state.pending)) |
| 2870 | } |
| 2871 | |
| 2872 | /// Reset the pool to zero without consuming. Test-only helper for |
| 2873 | /// suites that share the static and need to start from a known |
| 2874 | /// state. Production code should always use [`drain`]. |
| 2875 | #[cfg(test)] |
| 2876 | pub fn reset_for_tests() { |
| 2877 | with_pending_state_mut(|state| { |
| 2878 | state.pending = PendingBackgroundCost::default(); |
| 2879 | state.seen_usage_source_fingerprints.clear(); |
| 2880 | state.missing_usage_sources.clear(); |
| 2881 | state.missing_usage_overflowed = false; |
| 2882 | }); |
| 2883 | with_runtime_usage_journal_mut(HashMap::clear); |
| 2884 | } |
| 2885 | |
| 2886 | #[cfg(test)] |
| 2887 | pub(crate) struct TestCostScope; |
| 2888 | |
| 2889 | #[cfg(test)] |
| 2890 | impl Drop for TestCostScope { |
| 2891 | fn drop(&mut self) { |
| 2892 | reset_for_tests(); |
| 2893 | } |
| 2894 | } |
| 2895 | |
| 2896 | #[cfg(test)] |
| 2897 | pub(crate) fn test_scope() -> TestCostScope { |
| 2898 | reset_for_tests(); |
| 2899 | TestCostScope |
| 2900 | } |
| 2901 | |
| 2902 | #[cfg(test)] |
| 2903 | mod tests { |
| 2904 | use super::*; |
| 2905 | |
| 2906 | fn configured_fixture_receipt() -> (crate::config::Config, EffectiveRouteEnvelope, Usage) { |
| 2907 | let config = toml::from_str(include_str!( |
| 2908 | "../../config/tests/fixtures/custom_models.toml" |
| 2909 | )) |
| 2910 | .unwrap(); |
| 2911 | let receipt = EffectiveRouteEnvelope::capture( |
| 2912 | Some(&config), |
| 2913 | ProviderKind::Deepseek, |
| 2914 | "deepseek", |
| 2915 | "deepseek-v4.1-flash-expires-on-0910", |
| 2916 | Some("https://models.example.test/v1"), |
| 2917 | Utc::now(), |
| 2918 | ); |
| 2919 | let usage = Usage { |
| 2920 | input_tokens: 1_000_000, |
| 2921 | output_tokens: 1_000_000, |
| 2922 | ..Usage::default() |
| 2923 | }; |
| 2924 | (config, receipt, usage) |
| 2925 | } |
| 2926 | |
| 2927 | #[test] |
| 2928 | fn configured_model_estimate_is_frozen_and_exactly_bound() { |
| 2929 | let (mut config, receipt, usage) = configured_fixture_receipt(); |
| 2930 | let audit = receipt.audit(&usage); |
| 2931 | assert_eq!( |
| 2932 | audit.provenance, |
| 2933 | Some(codewhale_config::pricing::PricingProvenance::UserOverride) |
| 2934 | ); |
| 2935 | assert!((audit.estimate.unwrap().usd - 2.0).abs() < 1e-12); |
| 2936 | let frozen: EffectiveRouteEnvelope = |
| 2937 | serde_json::from_str(&serde_json::to_string(&receipt).unwrap()).unwrap(); |
| 2938 | config.custom_models.as_mut().unwrap()[0] |
| 2939 | .cost |
| 2940 | .as_mut() |
| 2941 | .unwrap() |
| 2942 | .input = Some(9.0); |
| 2943 | assert!((frozen.audit(&usage).estimate.unwrap().usd - 2.0).abs() < 1e-12); |
| 2944 | for (field, value) in [ |
| 2945 | ("model", "deepseek-v4.1-flash"), |
| 2946 | ("identity", "other-provider"), |
| 2947 | ("endpoint", "https://other.example.test/v1"), |
| 2948 | ] { |
| 2949 | let mut wrong = frozen.clone(); |
| 2950 | match field { |
| 2951 | "model" => wrong.model = value.into(), |
| 2952 | "identity" => wrong.provider_identity = value.into(), |
| 2953 | _ => wrong.endpoint_fingerprint = endpoint_fingerprint(value), |
| 2954 | } |
| 2955 | assert!(wrong.audit(&usage).estimate.is_none(), "{field}"); |
| 2956 | } |
| 2957 | for billing in [RouteBillingMode::Local, RouteBillingMode::Subscription] { |
| 2958 | let mut nonmoney = frozen.clone(); |
| 2959 | nonmoney.billing_mode = billing; |
| 2960 | assert_eq!( |
| 2961 | nonmoney.audit(&usage).unpriced_reason, |
| 2962 | Some(crate::pricing::UnpricedReason::NotMoneyMetered) |
| 2963 | ); |
| 2964 | } |
| 2965 | let mut cached = usage.clone(); |
| 2966 | cached.prompt_cache_write_tokens = Some(500); |
| 2967 | assert!(frozen.audit(&cached).estimate.is_none()); |
| 2968 | } |
| 2969 | |
| 2970 | #[test] |
| 2971 | fn configured_model_missing_prices_stay_unknown_and_vendor_pin_still_wins() { |
| 2972 | let (mut config, receipt, usage) = configured_fixture_receipt(); |
| 2973 | config.custom_models.as_mut().unwrap()[0].cost = None; |
| 2974 | let unknown = EffectiveRouteEnvelope::capture( |
| 2975 | Some(&config), |
| 2976 | receipt.provider, |
| 2977 | receipt.provider_identity.clone(), |
| 2978 | receipt.model.clone(), |
| 2979 | Some("https://models.example.test/v1"), |
| 2980 | receipt.dispatched_at, |
| 2981 | ); |
| 2982 | assert!(unknown.provider_live_pricing.is_some()); |
| 2983 | assert!(unknown.audit(&usage).estimate.is_none()); |
| 2984 | let mut pinned = receipt; |
| 2985 | pinned.provider = ProviderKind::Openrouter; |
| 2986 | pinned.openrouter_vendor = Some("exact-upstream".into()); |
| 2987 | pinned.billing_mode = RouteBillingMode::Metered; |
| 2988 | assert_eq!( |
| 2989 | pinned.audit(&usage).unpriced_reason, |
| 2990 | Some(crate::pricing::UnpricedReason::RoutingDependentPrice) |
| 2991 | ); |
| 2992 | } |
| 2993 | |
| 2994 | #[test] |
| 2995 | fn configured_model_client_keeps_its_metadata_snapshot_after_reload() { |
| 2996 | let (mut config, _, usage) = configured_fixture_receipt(); |
| 2997 | config.set_legacy_root(Some("fixture-not-a-provider-credential".into()), None); |
| 2998 | let id = "deepseek-v4.1-flash-expires-on-0910"; |
| 2999 | let route = |
| 3000 | crate::route_runtime::resolve_runtime_route(&config, ProviderKind::Deepseek, Some(id)) |
| 3001 | .unwrap(); |
| 3002 | let client = |
| 3003 | crate::client::CodewhaleClient::from_candidate(&config, &route.candidate).unwrap(); |
| 3004 | config.custom_models.as_mut().unwrap()[0] |
| 3005 | .cost |
| 3006 | .as_mut() |
| 3007 | .unwrap() |
| 3008 | .input = Some(9.0); |
| 3009 | let envelope = client.effective_route_envelope(id, Utc::now()); |
| 3010 | assert!((envelope.audit(&usage).estimate.unwrap().usd - 2.0).abs() < 1e-12); |
| 3011 | assert_eq!( |
| 3012 | client |
| 3013 | .effective_route_envelope("other-model", Utc::now()) |
| 3014 | .provider_live_pricing, |
| 3015 | None |
| 3016 | ); |
| 3017 | } |
| 3018 | |
| 3019 | const DECLARED_OPENROUTER_MODEL: &str = "synthetic/declared-model"; |
| 3020 | |
| 3021 | /// An OpenRouter config on the official endpoint with one |
| 3022 | /// `[[custom_models]]` row; `extra` adds fields such as `cost`. |
| 3023 | fn openrouter_declared_config(extra: &str, vendor: Option<&str>) -> crate::config::Config { |
| 3024 | let vendor = vendor.map_or_else(String::new, |vendor| format!("vendor = \"{vendor}\"\n")); |
| 3025 | toml::from_str(&format!( |
| 3026 | "provider = \"openrouter\"\ntelemetry = false\n\n\ |
| 3027 | [[custom_models]]\nprovider = \"openrouter\"\n\ |
| 3028 | base_url = \"{base}\"\nid = \"{DECLARED_OPENROUTER_MODEL}\"\n{extra}\n\n\ |
| 3029 | [providers.openrouter]\nbase_url = \"{base}\"\n\ |
| 3030 | api_key = \"fixture-not-a-provider-credential\"\n{vendor}", |
| 3031 | base = crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 3032 | )) |
| 3033 | .expect("declared OpenRouter config") |
| 3034 | } |
| 3035 | |
| 3036 | fn declared_openrouter_client( |
| 3037 | config: &crate::config::Config, |
| 3038 | ) -> crate::client::CodewhaleClient { |
| 3039 | let route = crate::route_runtime::resolve_runtime_route( |
| 3040 | config, |
| 3041 | ProviderKind::Openrouter, |
| 3042 | Some(DECLARED_OPENROUTER_MODEL), |
| 3043 | ) |
| 3044 | .expect("declared OpenRouter route"); |
| 3045 | crate::client::CodewhaleClient::from_candidate(config, &route.candidate) |
| 3046 | .expect("declared OpenRouter client") |
| 3047 | } |
| 3048 | |
| 3049 | /// #6690: a `[[custom_models]]` row declared only to add a model (no |
| 3050 | /// rates) used to freeze a rate-less quote on the main turn and hide the |
| 3051 | /// endpoint's own catalog price. A declared rate still wins, and with no |
| 3052 | /// catalog row the rate-less declaration stays frozen so no same-named |
| 3053 | /// bundled price can fill it. |
| 3054 | #[test] |
| 3055 | fn main_turn_rateless_declaration_yields_to_the_endpoint_catalog_price() { |
| 3056 | let _env = crate::test_support::lock_test_env(); |
| 3057 | let home = tempfile::tempdir().expect("isolated catalog home"); |
| 3058 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3059 | let _reset = ProviderCatalogTestReset; |
| 3060 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3061 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3062 | crate::provider_lake::clear_live_snapshot(); |
| 3063 | |
| 3064 | let fingerprint = endpoint_fingerprint(crate::config::DEFAULT_OPENROUTER_BASE_URL) |
| 3065 | .expect("official endpoint"); |
| 3066 | let now = u64::try_from(Utc::now().timestamp()).expect("timestamp"); |
| 3067 | let rateless = declared_openrouter_client(&openrouter_declared_config("", None)); |
| 3068 | let declared = declared_openrouter_client(&openrouter_declared_config( |
| 3069 | "cost = { input = 0.4, output = 1.6 }", |
| 3070 | None, |
| 3071 | )); |
| 3072 | let quote_for = |client: &crate::client::CodewhaleClient| { |
| 3073 | crate::client::main_turn_pricing_quote_at( |
| 3074 | Some(client), |
| 3075 | ProviderKind::Openrouter, |
| 3076 | "openrouter", |
| 3077 | DECLARED_OPENROUTER_MODEL, |
| 3078 | &fingerprint, |
| 3079 | now, |
| 3080 | ) |
| 3081 | .expect("a frozen main-turn quote") |
| 3082 | }; |
| 3083 | |
| 3084 | // No catalog row yet: the rate-less declaration is frozen as-is. |
| 3085 | let frozen = quote_for(&rateless); |
| 3086 | assert_eq!( |
| 3087 | frozen.provenance, |
| 3088 | codewhale_config::pricing::PricingProvenance::UserOverride |
| 3089 | ); |
| 3090 | assert!(!frozen.carries_rates()); |
| 3091 | |
| 3092 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3093 | "openrouter", |
| 3094 | DECLARED_OPENROUTER_MODEL, |
| 3095 | &fingerprint, |
| 3096 | now, |
| 3097 | )); |
| 3098 | let catalog = quote_for(&rateless); |
| 3099 | assert_eq!( |
| 3100 | catalog.provenance, |
| 3101 | codewhale_config::pricing::PricingProvenance::ProviderLive |
| 3102 | ); |
| 3103 | assert_eq!(catalog.input_per_million.as_deref(), Some("1.25")); |
| 3104 | assert_eq!(catalog.output_per_million.as_deref(), Some("5")); |
| 3105 | assert_eq!(catalog.cache_read_per_million.as_deref(), Some("0.25")); |
| 3106 | |
| 3107 | let explicit = quote_for(&declared); |
| 3108 | assert_eq!( |
| 3109 | explicit.provenance, |
| 3110 | codewhale_config::pricing::PricingProvenance::UserOverride |
| 3111 | ); |
| 3112 | assert_eq!(explicit.input_per_million.as_deref(), Some("0.4")); |
| 3113 | assert_eq!(explicit.output_per_million.as_deref(), Some("1.6")); |
| 3114 | } |
| 3115 | |
| 3116 | /// #6690 review: a pinned OpenRouter vendor blocks the aggregate catalog |
| 3117 | /// price, but an operator's own declared rate for the exact route is not |
| 3118 | /// that aggregate and must still price the turn. |
| 3119 | #[test] |
| 3120 | fn openrouter_vendor_pin_is_priced_by_an_explicit_declared_rate() { |
| 3121 | let _env = crate::test_support::lock_test_env(); |
| 3122 | let home = tempfile::tempdir().expect("isolated catalog home"); |
| 3123 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3124 | let _reset = ProviderCatalogTestReset; |
| 3125 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3126 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3127 | crate::provider_lake::clear_live_snapshot(); |
| 3128 | |
| 3129 | let usage = Usage { |
| 3130 | input_tokens: 1_000_000, |
| 3131 | output_tokens: 1_000_000, |
| 3132 | ..Usage::default() |
| 3133 | }; |
| 3134 | let capture = |extra: &str| { |
| 3135 | EffectiveRouteEnvelope::capture( |
| 3136 | Some(&openrouter_declared_config(extra, Some("cerebras"))), |
| 3137 | ProviderKind::Openrouter, |
| 3138 | "openrouter", |
| 3139 | DECLARED_OPENROUTER_MODEL, |
| 3140 | Some(crate::config::DEFAULT_OPENROUTER_BASE_URL), |
| 3141 | Utc::now(), |
| 3142 | ) |
| 3143 | }; |
| 3144 | |
| 3145 | let declared = capture("cost = { input = 0.4, output = 1.6 }"); |
| 3146 | assert_eq!(declared.openrouter_vendor.as_deref(), Some("cerebras")); |
| 3147 | let audit = declared.audit(&usage); |
| 3148 | assert_eq!(audit.unpriced_reason, None, "{audit:?}"); |
| 3149 | assert_eq!( |
| 3150 | audit.provenance, |
| 3151 | Some(codewhale_config::pricing::PricingProvenance::UserOverride) |
| 3152 | ); |
| 3153 | let usd = audit.estimate.expect("declared estimate").usd; |
| 3154 | assert!((usd - 2.0).abs() < 1e-12, "{usd}"); |
| 3155 | |
| 3156 | // A rate-less declaration is no explicit rate: the pin still wins. |
| 3157 | let rateless = capture(""); |
| 3158 | assert_eq!( |
| 3159 | rateless.audit(&usage).unpriced_reason, |
| 3160 | Some(crate::pricing::UnpricedReason::RoutingDependentPrice) |
| 3161 | ); |
| 3162 | } |
| 3163 | |
| 3164 | struct ProviderCatalogTestReset; |
| 3165 | |
| 3166 | impl Drop for ProviderCatalogTestReset { |
| 3167 | fn drop(&mut self) { |
| 3168 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3169 | crate::provider_lake::clear_live_snapshot(); |
| 3170 | } |
| 3171 | } |
| 3172 | |
| 3173 | fn priced_provider_delta( |
| 3174 | provider: &str, |
| 3175 | model: &str, |
| 3176 | fingerprint: &str, |
| 3177 | fetched_at: u64, |
| 3178 | ) -> codewhale_config::catalog::ProviderCatalogDelta { |
| 3179 | priced_provider_delta_with_rates(provider, model, fingerprint, fetched_at, 1.25, 5.0) |
| 3180 | } |
| 3181 | |
| 3182 | fn priced_provider_delta_with_rates( |
| 3183 | provider: &str, |
| 3184 | model: &str, |
| 3185 | fingerprint: &str, |
| 3186 | fetched_at: u64, |
| 3187 | input: f64, |
| 3188 | output: f64, |
| 3189 | ) -> codewhale_config::catalog::ProviderCatalogDelta { |
| 3190 | codewhale_config::catalog::ProviderCatalogDelta { |
| 3191 | provider: provider.to_string(), |
| 3192 | base_url_fingerprint: fingerprint.to_string(), |
| 3193 | fetched_at, |
| 3194 | offerings: vec![codewhale_config::catalog::CatalogOffering { |
| 3195 | provider: provider.to_string(), |
| 3196 | wire_model_id: model.to_string(), |
| 3197 | endpoint_key: "chat".to_string(), |
| 3198 | cost: Some(codewhale_config::models_dev::ModelsDevCost { |
| 3199 | input: Some(input), |
| 3200 | output: Some(output), |
| 3201 | cache_read: Some(0.25), |
| 3202 | cache_write: None, |
| 3203 | }), |
| 3204 | ..Default::default() |
| 3205 | }], |
| 3206 | } |
| 3207 | } |
| 3208 | |
| 3209 | fn custom_usage_envelope( |
| 3210 | identity: &str, |
| 3211 | model: &str, |
| 3212 | fingerprint: &str, |
| 3213 | billing_mode: RouteBillingMode, |
| 3214 | dispatched_at: DateTime<Utc>, |
| 3215 | ) -> EffectiveRouteEnvelope { |
| 3216 | provider_live_usage_envelope( |
| 3217 | ProviderKind::Custom, |
| 3218 | identity, |
| 3219 | model, |
| 3220 | fingerprint, |
| 3221 | Some(crate::pricing::UNCLASSIFIED_BILLING_SURFACE), |
| 3222 | billing_mode, |
| 3223 | dispatched_at, |
| 3224 | ) |
| 3225 | } |
| 3226 | |
| 3227 | fn provider_live_usage_envelope( |
| 3228 | provider: ProviderKind, |
| 3229 | identity: &str, |
| 3230 | model: &str, |
| 3231 | fingerprint: &str, |
| 3232 | billing_surface: Option<&str>, |
| 3233 | billing_mode: RouteBillingMode, |
| 3234 | dispatched_at: DateTime<Utc>, |
| 3235 | ) -> EffectiveRouteEnvelope { |
| 3236 | let provider_live_pricing = |
| 3237 | u64::try_from(dispatched_at.timestamp()) |
| 3238 | .ok() |
| 3239 | .and_then(|dispatched_at_unix| { |
| 3240 | crate::provider_catalog_live::fresh_provider_live_pricing_quote_at( |
| 3241 | provider, |
| 3242 | identity, |
| 3243 | model, |
| 3244 | fingerprint, |
| 3245 | dispatched_at_unix, |
| 3246 | ) |
| 3247 | }); |
| 3248 | EffectiveRouteEnvelope { |
| 3249 | provider, |
| 3250 | provider_identity: identity.to_string(), |
| 3251 | model: model.to_string(), |
| 3252 | openrouter_vendor: None, |
| 3253 | billing_surface: billing_surface.map(str::to_string), |
| 3254 | endpoint_fingerprint: Some(fingerprint.to_string()), |
| 3255 | provider_live_pricing, |
| 3256 | billing_mode, |
| 3257 | dispatched_at, |
| 3258 | } |
| 3259 | } |
| 3260 | |
| 3261 | fn small_usage() -> Usage { |
| 3262 | Usage { |
| 3263 | input_tokens: 1_000, |
| 3264 | output_tokens: 500, |
| 3265 | ..Default::default() |
| 3266 | } |
| 3267 | } |
| 3268 | |
| 3269 | #[test] |
| 3270 | fn baseten_usage_prices_only_the_reviewed_identity_on_the_official_endpoint() { |
| 3271 | let _env = crate::test_support::lock_test_env(); |
| 3272 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3273 | let home = tempfile::tempdir().expect("test home"); |
| 3274 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3275 | let _reset = ProviderCatalogTestReset; |
| 3276 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3277 | crate::provider_lake::clear_live_snapshot(); |
| 3278 | |
| 3279 | let now = Utc::now(); |
| 3280 | let fetched_at = u64::try_from(now.timestamp()).expect("nonnegative timestamp"); |
| 3281 | let model = "synthetic-baseten-priced-model"; |
| 3282 | let fingerprint = codewhale_config::catalog::base_url_fingerprint( |
| 3283 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3284 | ); |
| 3285 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3286 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3287 | model, |
| 3288 | &fingerprint, |
| 3289 | fetched_at, |
| 3290 | )); |
| 3291 | let usage = Usage { |
| 3292 | input_tokens: 1_000_000, |
| 3293 | ..Usage::default() |
| 3294 | }; |
| 3295 | |
| 3296 | let exact = custom_usage_envelope( |
| 3297 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3298 | model, |
| 3299 | &fingerprint, |
| 3300 | RouteBillingMode::Unknown, |
| 3301 | now, |
| 3302 | ) |
| 3303 | .audit(&usage); |
| 3304 | assert!(exact.is_priced(), "{exact:?}"); |
| 3305 | assert_eq!( |
| 3306 | exact.provenance, |
| 3307 | Some(codewhale_config::pricing::PricingProvenance::ProviderLive) |
| 3308 | ); |
| 3309 | assert_eq!(exact.estimate.expect("priced").usd, 1.25); |
| 3310 | |
| 3311 | // A reviewed schema alias remains a distinct custom ownership scope. |
| 3312 | // It becomes billable only after that exact identity refreshed its own |
| 3313 | // catalog; it cannot borrow the canonical `baseten` partition above. |
| 3314 | let alias = "base-ten"; |
| 3315 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3316 | alias, |
| 3317 | model, |
| 3318 | &fingerprint, |
| 3319 | fetched_at, |
| 3320 | )); |
| 3321 | let alias_audit = |
| 3322 | custom_usage_envelope(alias, model, &fingerprint, RouteBillingMode::Unknown, now) |
| 3323 | .audit(&usage); |
| 3324 | assert!(alias_audit.is_priced(), "{alias_audit:?}"); |
| 3325 | assert_eq!( |
| 3326 | alias_audit.provenance, |
| 3327 | Some(codewhale_config::pricing::PricingProvenance::ProviderLive) |
| 3328 | ); |
| 3329 | assert_eq!(alias_audit.estimate.expect("priced").usd, 1.25); |
| 3330 | |
| 3331 | let generic = custom_usage_envelope( |
| 3332 | "custom-lab", |
| 3333 | model, |
| 3334 | &fingerprint, |
| 3335 | RouteBillingMode::Metered, |
| 3336 | now, |
| 3337 | ) |
| 3338 | .audit(&usage); |
| 3339 | assert!(!generic.is_priced(), "{generic:?}"); |
| 3340 | // The endpoint fingerprint establishes Baseten's billing contract no |
| 3341 | // matter the table name, so the failure is an unverified price for |
| 3342 | // this identity — not an unknown basis (#6289). |
| 3343 | assert_eq!( |
| 3344 | generic.unpriced_reason, |
| 3345 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3346 | ); |
| 3347 | |
| 3348 | let wrong_fingerprint = |
| 3349 | codewhale_config::catalog::base_url_fingerprint("https://proxy.example/v1"); |
| 3350 | let wrong_endpoint = custom_usage_envelope( |
| 3351 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3352 | model, |
| 3353 | &wrong_fingerprint, |
| 3354 | RouteBillingMode::Metered, |
| 3355 | now, |
| 3356 | ) |
| 3357 | .audit(&usage); |
| 3358 | assert!(!wrong_endpoint.is_priced(), "{wrong_endpoint:?}"); |
| 3359 | assert_eq!( |
| 3360 | wrong_endpoint.unpriced_reason, |
| 3361 | Some(crate::pricing::UnpricedReason::UnknownBillingBasis) |
| 3362 | ); |
| 3363 | } |
| 3364 | |
| 3365 | #[test] |
| 3366 | fn baseten_usage_rejects_unknown_stale_and_failed_live_catalogs() { |
| 3367 | let _env = crate::test_support::lock_test_env(); |
| 3368 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3369 | let home = tempfile::tempdir().expect("test home"); |
| 3370 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3371 | let _reset = ProviderCatalogTestReset; |
| 3372 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3373 | crate::provider_lake::clear_live_snapshot(); |
| 3374 | |
| 3375 | let now = Utc::now(); |
| 3376 | let now_unix = u64::try_from(now.timestamp()).expect("nonnegative timestamp"); |
| 3377 | let model = "synthetic-baseten-status-model"; |
| 3378 | let fingerprint = codewhale_config::catalog::base_url_fingerprint( |
| 3379 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3380 | ); |
| 3381 | let unknown_route = custom_usage_envelope( |
| 3382 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3383 | model, |
| 3384 | &fingerprint, |
| 3385 | RouteBillingMode::Unknown, |
| 3386 | now, |
| 3387 | ); |
| 3388 | assert!(unknown_route.provider_live_pricing.is_none()); |
| 3389 | let usage = Usage { |
| 3390 | input_tokens: 1_000_000, |
| 3391 | ..Usage::default() |
| 3392 | }; |
| 3393 | |
| 3394 | // A same-model price owned by another custom partition cannot price a |
| 3395 | // Baseten receipt whose exact catalog was never refreshed. |
| 3396 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3397 | "other-custom", |
| 3398 | model, |
| 3399 | &fingerprint, |
| 3400 | now_unix, |
| 3401 | )); |
| 3402 | let unknown = unknown_route.audit(&usage); |
| 3403 | assert!(!unknown.is_priced(), "{unknown:?}"); |
| 3404 | assert_eq!( |
| 3405 | unknown.unpriced_reason, |
| 3406 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3407 | ); |
| 3408 | |
| 3409 | let stale_at = now_unix |
| 3410 | .saturating_sub(crate::provider_catalog_live::DEFAULT_PROVIDER_CATALOG_TTL_SECS) |
| 3411 | .saturating_sub(1); |
| 3412 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3413 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3414 | model, |
| 3415 | &fingerprint, |
| 3416 | stale_at, |
| 3417 | )); |
| 3418 | let stale_route = custom_usage_envelope( |
| 3419 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3420 | model, |
| 3421 | &fingerprint, |
| 3422 | RouteBillingMode::Unknown, |
| 3423 | now, |
| 3424 | ); |
| 3425 | assert!(stale_route.provider_live_pricing.is_none()); |
| 3426 | let stale = stale_route.audit(&usage); |
| 3427 | assert!(!stale.is_priced(), "{stale:?}"); |
| 3428 | assert_eq!( |
| 3429 | stale.unpriced_reason, |
| 3430 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3431 | ); |
| 3432 | |
| 3433 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3434 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3435 | model, |
| 3436 | &fingerprint, |
| 3437 | now_unix, |
| 3438 | )); |
| 3439 | crate::provider_catalog_live::record_failure( |
| 3440 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3441 | &fingerprint, |
| 3442 | codewhale_config::catalog::CatalogRefreshError::Network, |
| 3443 | ); |
| 3444 | let failed_route = custom_usage_envelope( |
| 3445 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3446 | model, |
| 3447 | &fingerprint, |
| 3448 | RouteBillingMode::Unknown, |
| 3449 | now, |
| 3450 | ); |
| 3451 | assert!(failed_route.provider_live_pricing.is_none()); |
| 3452 | let failed = failed_route.audit(&usage); |
| 3453 | assert!(!failed.is_priced(), "{failed:?}"); |
| 3454 | assert_eq!( |
| 3455 | failed.unpriced_reason, |
| 3456 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3457 | ); |
| 3458 | } |
| 3459 | |
| 3460 | #[test] |
| 3461 | fn reviewed_provider_live_quotes_survive_same_second_refresh_and_key_state_changes() { |
| 3462 | let _env = crate::test_support::lock_test_env(); |
| 3463 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3464 | let home = tempfile::tempdir().expect("test home"); |
| 3465 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3466 | let _reset = ProviderCatalogTestReset; |
| 3467 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3468 | crate::provider_lake::clear_live_snapshot(); |
| 3469 | |
| 3470 | let now = Utc::now(); |
| 3471 | let fetched_at = u64::try_from(now.timestamp()).expect("nonnegative timestamp"); |
| 3472 | let cases = [ |
| 3473 | ( |
| 3474 | ProviderKind::Openrouter, |
| 3475 | ProviderKind::Openrouter.as_str(), |
| 3476 | "synthetic-openrouter-frozen-price", |
| 3477 | codewhale_config::catalog::base_url_fingerprint( |
| 3478 | crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 3479 | ), |
| 3480 | crate::pricing::AGGREGATOR_BILLING_SURFACE, |
| 3481 | RouteBillingMode::Metered, |
| 3482 | ), |
| 3483 | ( |
| 3484 | ProviderKind::Custom, |
| 3485 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3486 | "synthetic-baseten-frozen-price", |
| 3487 | codewhale_config::catalog::base_url_fingerprint( |
| 3488 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3489 | ), |
| 3490 | crate::pricing::UNCLASSIFIED_BILLING_SURFACE, |
| 3491 | RouteBillingMode::Unknown, |
| 3492 | ), |
| 3493 | ]; |
| 3494 | let usage = Usage { |
| 3495 | input_tokens: 1_000_000, |
| 3496 | ..Usage::default() |
| 3497 | }; |
| 3498 | |
| 3499 | for (provider, identity, model, fingerprint, surface, mode) in cases { |
| 3500 | crate::provider_catalog_live::record_success(priced_provider_delta_with_rates( |
| 3501 | identity, |
| 3502 | model, |
| 3503 | &fingerprint, |
| 3504 | fetched_at, |
| 3505 | 1.25, |
| 3506 | 5.0, |
| 3507 | )); |
| 3508 | let first = provider_live_usage_envelope( |
| 3509 | provider, |
| 3510 | identity, |
| 3511 | model, |
| 3512 | &fingerprint, |
| 3513 | Some(surface), |
| 3514 | mode, |
| 3515 | now, |
| 3516 | ); |
| 3517 | let first_quote = first |
| 3518 | .provider_live_pricing |
| 3519 | .as_ref() |
| 3520 | .expect("fresh exact scope freezes a quote"); |
| 3521 | |
| 3522 | // A second refresh in the same Unix second must still be a distinct |
| 3523 | // catalog revision and must not retroactively change `first`. |
| 3524 | crate::provider_catalog_live::record_success(priced_provider_delta_with_rates( |
| 3525 | identity, |
| 3526 | model, |
| 3527 | &fingerprint, |
| 3528 | fetched_at, |
| 3529 | 9.5, |
| 3530 | 19.0, |
| 3531 | )); |
| 3532 | let second = provider_live_usage_envelope( |
| 3533 | provider, |
| 3534 | identity, |
| 3535 | model, |
| 3536 | &fingerprint, |
| 3537 | Some(surface), |
| 3538 | mode, |
| 3539 | now, |
| 3540 | ); |
| 3541 | let second_quote = second |
| 3542 | .provider_live_pricing |
| 3543 | .as_ref() |
| 3544 | .expect("replacement fresh scope freezes a quote"); |
| 3545 | assert_ne!( |
| 3546 | first_quote.catalog_revision, second_quote.catalog_revision, |
| 3547 | "same-second price changes need distinct revisions" |
| 3548 | ); |
| 3549 | |
| 3550 | crate::provider_catalog_live::record_failure( |
| 3551 | identity, |
| 3552 | &fingerprint, |
| 3553 | codewhale_config::catalog::CatalogRefreshError::Unauthorized, |
| 3554 | ); |
| 3555 | if provider == ProviderKind::Custom { |
| 3556 | // Baseten's same URL can represent another account after a key |
| 3557 | // switch. Starting that refresh clears the mutable old scope. |
| 3558 | let _new_key_refresh = crate::provider_catalog_live::begin_refresh_for_identity( |
| 3559 | provider, |
| 3560 | identity, |
| 3561 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3562 | ); |
| 3563 | } |
| 3564 | |
| 3565 | let first_audit = first.audit(&usage); |
| 3566 | let second_audit = second.audit(&usage); |
| 3567 | assert_eq!(first_audit.estimate.expect("first quote priced").usd, 1.25); |
| 3568 | assert_eq!(second_audit.estimate.expect("second quote priced").usd, 9.5); |
| 3569 | |
| 3570 | let after_mutation = provider_live_usage_envelope( |
| 3571 | provider, |
| 3572 | identity, |
| 3573 | model, |
| 3574 | &fingerprint, |
| 3575 | Some(surface), |
| 3576 | mode, |
| 3577 | now, |
| 3578 | ); |
| 3579 | assert!( |
| 3580 | after_mutation.provider_live_pricing.is_none(), |
| 3581 | "failed or cleared mutable state cannot mint a new quote" |
| 3582 | ); |
| 3583 | } |
| 3584 | } |
| 3585 | |
| 3586 | #[test] |
| 3587 | fn legacy_no_quote_receipts_cannot_be_retro_priced_by_a_later_refresh() { |
| 3588 | let _env = crate::test_support::lock_test_env(); |
| 3589 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3590 | let home = tempfile::tempdir().expect("test home"); |
| 3591 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3592 | let _reset = ProviderCatalogTestReset; |
| 3593 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3594 | crate::provider_lake::clear_live_snapshot(); |
| 3595 | |
| 3596 | let now = Utc::now(); |
| 3597 | let fetched_at = u64::try_from(now.timestamp()).expect("nonnegative timestamp"); |
| 3598 | let routes = [ |
| 3599 | provider_live_usage_envelope( |
| 3600 | ProviderKind::Openrouter, |
| 3601 | ProviderKind::Openrouter.as_str(), |
| 3602 | "synthetic-openrouter-legacy", |
| 3603 | &codewhale_config::catalog::base_url_fingerprint( |
| 3604 | crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 3605 | ), |
| 3606 | Some(crate::pricing::AGGREGATOR_BILLING_SURFACE), |
| 3607 | RouteBillingMode::Metered, |
| 3608 | now, |
| 3609 | ), |
| 3610 | custom_usage_envelope( |
| 3611 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3612 | "synthetic-baseten-legacy", |
| 3613 | &codewhale_config::catalog::base_url_fingerprint( |
| 3614 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3615 | ), |
| 3616 | RouteBillingMode::Unknown, |
| 3617 | now, |
| 3618 | ), |
| 3619 | ]; |
| 3620 | assert!( |
| 3621 | routes |
| 3622 | .iter() |
| 3623 | .all(|route| route.provider_live_pricing.is_none()) |
| 3624 | ); |
| 3625 | |
| 3626 | for route in &routes { |
| 3627 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3628 | &route.provider_identity, |
| 3629 | &route.model, |
| 3630 | route.endpoint_fingerprint.as_deref().expect("fingerprint"), |
| 3631 | fetched_at, |
| 3632 | )); |
| 3633 | let audit = route.audit(&Usage { |
| 3634 | input_tokens: 1_000_000, |
| 3635 | ..Usage::default() |
| 3636 | }); |
| 3637 | assert_eq!( |
| 3638 | audit.unpriced_reason, |
| 3639 | Some(if route.provider == ProviderKind::Openrouter { |
| 3640 | crate::pricing::UnpricedReason::NoPricingRow |
| 3641 | } else { |
| 3642 | crate::pricing::UnpricedReason::UnverifiedLivePricing |
| 3643 | }), |
| 3644 | "a completion-time refresh must not price {route:?}" |
| 3645 | ); |
| 3646 | } |
| 3647 | } |
| 3648 | |
| 3649 | #[test] |
| 3650 | fn openrouter_offline_bundled_price_is_immutable_after_dispatch() { |
| 3651 | let _env = crate::test_support::lock_test_env(); |
| 3652 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3653 | let home = tempfile::tempdir().expect("test home"); |
| 3654 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3655 | let _reset = ProviderCatalogTestReset; |
| 3656 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3657 | crate::provider_lake::clear_live_snapshot(); |
| 3658 | |
| 3659 | let dispatched_at = Utc::now(); |
| 3660 | let fetched_at = u64::try_from(dispatched_at.timestamp()).expect("timestamp"); |
| 3661 | let model = "qwen/qwen3.8-flash"; |
| 3662 | let fingerprint = codewhale_config::catalog::base_url_fingerprint( |
| 3663 | crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 3664 | ); |
| 3665 | let route = provider_live_usage_envelope( |
| 3666 | ProviderKind::Openrouter, |
| 3667 | ProviderKind::Openrouter.as_str(), |
| 3668 | model, |
| 3669 | &fingerprint, |
| 3670 | Some(crate::pricing::AGGREGATOR_BILLING_SURFACE), |
| 3671 | RouteBillingMode::Metered, |
| 3672 | dispatched_at, |
| 3673 | ); |
| 3674 | assert!(route.provider_live_pricing.is_none()); |
| 3675 | |
| 3676 | let usage = Usage { |
| 3677 | input_tokens: 1_000_000, |
| 3678 | ..Usage::default() |
| 3679 | }; |
| 3680 | let offline = route.audit(&usage); |
| 3681 | assert_eq!( |
| 3682 | offline.estimate.expect("bundled OpenRouter price").usd, |
| 3683 | 0.15 |
| 3684 | ); |
| 3685 | assert_eq!( |
| 3686 | offline.provenance, |
| 3687 | Some(codewhale_config::pricing::PricingProvenance::ModelsDevBundled) |
| 3688 | ); |
| 3689 | |
| 3690 | // A later mutable refresh cannot change a turn that had no quote at |
| 3691 | // the application-dispatch boundary. |
| 3692 | crate::provider_catalog_live::record_success(priced_provider_delta_with_rates( |
| 3693 | ProviderKind::Openrouter.as_str(), |
| 3694 | model, |
| 3695 | &fingerprint, |
| 3696 | fetched_at, |
| 3697 | 19.0, |
| 3698 | 29.0, |
| 3699 | )); |
| 3700 | let after_refresh = route.audit(&usage); |
| 3701 | assert_eq!(after_refresh, offline); |
| 3702 | |
| 3703 | // Admission without provider usage does not create a charge. |
| 3704 | let no_usage = route.audit(&Usage::default()); |
| 3705 | let no_usage_estimate = no_usage.estimate.expect("known zero usage is priced"); |
| 3706 | assert_eq!(no_usage_estimate.usd, 0.0); |
| 3707 | assert_eq!(no_usage_estimate.cny, 0.0); |
| 3708 | } |
| 3709 | |
| 3710 | #[test] |
| 3711 | fn provider_live_quotes_reject_future_prices_and_every_route_binding_mismatch() { |
| 3712 | let _env = crate::test_support::lock_test_env(); |
| 3713 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3714 | let home = tempfile::tempdir().expect("test home"); |
| 3715 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3716 | let _reset = ProviderCatalogTestReset; |
| 3717 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3718 | crate::provider_lake::clear_live_snapshot(); |
| 3719 | |
| 3720 | let dispatched_at = Utc::now(); |
| 3721 | let dispatch_unix = u64::try_from(dispatched_at.timestamp()).expect("timestamp"); |
| 3722 | let future_at = dispatched_at + chrono::Duration::seconds(1); |
| 3723 | let future_unix = dispatch_unix.saturating_add(1); |
| 3724 | let cases = [ |
| 3725 | ( |
| 3726 | ProviderKind::Openrouter, |
| 3727 | ProviderKind::Openrouter.as_str(), |
| 3728 | "synthetic-openrouter-future", |
| 3729 | codewhale_config::catalog::base_url_fingerprint( |
| 3730 | crate::config::DEFAULT_OPENROUTER_BASE_URL, |
| 3731 | ), |
| 3732 | crate::pricing::AGGREGATOR_BILLING_SURFACE, |
| 3733 | RouteBillingMode::Metered, |
| 3734 | ), |
| 3735 | ( |
| 3736 | ProviderKind::Custom, |
| 3737 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3738 | "synthetic-baseten-future", |
| 3739 | codewhale_config::catalog::base_url_fingerprint( |
| 3740 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3741 | ), |
| 3742 | crate::pricing::UNCLASSIFIED_BILLING_SURFACE, |
| 3743 | RouteBillingMode::Unknown, |
| 3744 | ), |
| 3745 | ]; |
| 3746 | let usage = Usage { |
| 3747 | input_tokens: 1_000_000, |
| 3748 | ..Usage::default() |
| 3749 | }; |
| 3750 | |
| 3751 | for (provider, identity, model, fingerprint, surface, mode) in cases { |
| 3752 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3753 | identity, |
| 3754 | model, |
| 3755 | &fingerprint, |
| 3756 | future_unix, |
| 3757 | )); |
| 3758 | let no_future_quote = provider_live_usage_envelope( |
| 3759 | provider, |
| 3760 | identity, |
| 3761 | model, |
| 3762 | &fingerprint, |
| 3763 | Some(surface), |
| 3764 | mode, |
| 3765 | dispatched_at, |
| 3766 | ); |
| 3767 | assert!(no_future_quote.provider_live_pricing.is_none()); |
| 3768 | assert_eq!( |
| 3769 | no_future_quote.audit(&usage).unpriced_reason, |
| 3770 | Some(if provider == ProviderKind::Openrouter { |
| 3771 | crate::pricing::UnpricedReason::NoPricingRow |
| 3772 | } else { |
| 3773 | crate::pricing::UnpricedReason::UnverifiedLivePricing |
| 3774 | }) |
| 3775 | ); |
| 3776 | |
| 3777 | let captured = provider_live_usage_envelope( |
| 3778 | provider, |
| 3779 | identity, |
| 3780 | model, |
| 3781 | &fingerprint, |
| 3782 | Some(surface), |
| 3783 | mode, |
| 3784 | future_at, |
| 3785 | ); |
| 3786 | assert!(captured.provider_live_pricing.is_some()); |
| 3787 | |
| 3788 | let mut future_relative_to_dispatch = captured.clone(); |
| 3789 | future_relative_to_dispatch.dispatched_at = dispatched_at; |
| 3790 | assert_eq!( |
| 3791 | future_relative_to_dispatch.audit(&usage).unpriced_reason, |
| 3792 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3793 | ); |
| 3794 | let persisted = serde_json::to_value(&future_relative_to_dispatch) |
| 3795 | .expect("invalid future quote serializes only as absent"); |
| 3796 | assert!(persisted["provider_live_pricing"].is_null()); |
| 3797 | |
| 3798 | let mut wrong_model = captured.clone(); |
| 3799 | wrong_model.model.push_str("-other"); |
| 3800 | assert_eq!( |
| 3801 | wrong_model.audit(&usage).unpriced_reason, |
| 3802 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3803 | ); |
| 3804 | |
| 3805 | let mut wrong_identity = captured.clone(); |
| 3806 | wrong_identity.provider_identity.push_str("-other"); |
| 3807 | // The endpoint fingerprint still establishes the billing contract, |
| 3808 | // so a renamed identity fails quote verification (#6289). |
| 3809 | assert_eq!( |
| 3810 | wrong_identity.audit(&usage).unpriced_reason, |
| 3811 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3812 | ); |
| 3813 | |
| 3814 | let mut wrong_endpoint = captured; |
| 3815 | wrong_endpoint.endpoint_fingerprint = Some( |
| 3816 | codewhale_config::catalog::base_url_fingerprint("https://proxy.example/v1"), |
| 3817 | ); |
| 3818 | assert_eq!( |
| 3819 | wrong_endpoint.audit(&usage).unpriced_reason, |
| 3820 | Some(if provider == ProviderKind::Custom { |
| 3821 | crate::pricing::UnpricedReason::UnknownBillingBasis |
| 3822 | } else { |
| 3823 | crate::pricing::UnpricedReason::UnverifiedLivePricing |
| 3824 | }) |
| 3825 | ); |
| 3826 | } |
| 3827 | } |
| 3828 | |
| 3829 | #[test] |
| 3830 | fn provider_live_quote_serialization_is_secret_free_and_legacy_compatible() { |
| 3831 | let _env = crate::test_support::lock_test_env(); |
| 3832 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 3833 | let home = tempfile::tempdir().expect("test home"); |
| 3834 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3835 | let _reset = ProviderCatalogTestReset; |
| 3836 | crate::provider_catalog_live::reset_cache_for_test(); |
| 3837 | crate::provider_lake::clear_live_snapshot(); |
| 3838 | |
| 3839 | let now = Utc::now(); |
| 3840 | let fetched_at = u64::try_from(now.timestamp()).expect("nonnegative timestamp"); |
| 3841 | let model = "synthetic-baseten-serialized-quote"; |
| 3842 | let fingerprint = codewhale_config::catalog::base_url_fingerprint( |
| 3843 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3844 | ); |
| 3845 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 3846 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3847 | model, |
| 3848 | &fingerprint, |
| 3849 | fetched_at, |
| 3850 | )); |
| 3851 | let route = custom_usage_envelope( |
| 3852 | codewhale_config::catalog::BASETEN_PROVIDER_ID, |
| 3853 | model, |
| 3854 | &fingerprint, |
| 3855 | RouteBillingMode::Unknown, |
| 3856 | now, |
| 3857 | ); |
| 3858 | assert!(route.provider_live_pricing.is_some()); |
| 3859 | |
| 3860 | let serialized = serde_json::to_string(&route).expect("serialize frozen route"); |
| 3861 | assert!(serialized.contains("provider_live_pricing")); |
| 3862 | assert!(serialized.contains("catalog_revision")); |
| 3863 | assert!(serialized.contains("input_per_million")); |
| 3864 | for secret in [ |
| 3865 | codewhale_config::catalog::BASETEN_BASE_URL, |
| 3866 | "api_key", |
| 3867 | "Bearer ", |
| 3868 | ] { |
| 3869 | // The assertion message must not itself become a logging sink for |
| 3870 | // the credential fragment it checks for — name the check, not the |
| 3871 | // secret. |
| 3872 | assert!( |
| 3873 | !serialized.contains(secret), |
| 3874 | "frozen route serialization leaked a credential fragment" |
| 3875 | ); |
| 3876 | } |
| 3877 | |
| 3878 | let mut child = serde_json::json!({}); |
| 3879 | attach_child_usage_metadata(&mut child, &route, &Usage::default()); |
| 3880 | let child_route = child_route_envelope_from_metadata(&child).expect("child route"); |
| 3881 | assert_eq!(child_route, route.sanitized_for_persistence()); |
| 3882 | |
| 3883 | let mut legacy: serde_json::Value = |
| 3884 | serde_json::from_str(&serialized).expect("route JSON value"); |
| 3885 | legacy |
| 3886 | .as_object_mut() |
| 3887 | .expect("route object") |
| 3888 | .remove("provider_live_pricing"); |
| 3889 | let legacy: EffectiveRouteEnvelope = |
| 3890 | serde_json::from_value(legacy).expect("legacy route remains readable"); |
| 3891 | assert!(legacy.provider_live_pricing.is_none()); |
| 3892 | let audit = legacy.audit(&Usage { |
| 3893 | input_tokens: 1_000_000, |
| 3894 | ..Usage::default() |
| 3895 | }); |
| 3896 | assert_eq!( |
| 3897 | audit.unpriced_reason, |
| 3898 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3899 | ); |
| 3900 | |
| 3901 | let mut wrong_model = route.clone(); |
| 3902 | wrong_model.model.push_str("-other"); |
| 3903 | assert_eq!( |
| 3904 | wrong_model.audit(&Usage::default()).unpriced_reason, |
| 3905 | Some(crate::pricing::UnpricedReason::UnverifiedLivePricing) |
| 3906 | ); |
| 3907 | } |
| 3908 | |
| 3909 | #[test] |
| 3910 | fn routed_child_batch_is_preferred_bounded_and_sanitized() { |
| 3911 | let route = deepseek_envelope(); |
| 3912 | let records = vec![ |
| 3913 | RuntimeUsageRecord { |
| 3914 | source_id: "raw-provider-response-id-one".to_string(), |
| 3915 | usage: EffectiveRouteUsage { |
| 3916 | route: route.clone(), |
| 3917 | usage: Usage { |
| 3918 | input_tokens: 11, |
| 3919 | ..Usage::default() |
| 3920 | }, |
| 3921 | }, |
| 3922 | }, |
| 3923 | RuntimeUsageRecord { |
| 3924 | source_id: "raw-provider-response-id-two".to_string(), |
| 3925 | usage: EffectiveRouteUsage { |
| 3926 | route: route.clone(), |
| 3927 | usage: Usage { |
| 3928 | output_tokens: 7, |
| 3929 | ..Usage::default() |
| 3930 | }, |
| 3931 | }, |
| 3932 | }, |
| 3933 | ]; |
| 3934 | let mut metadata = serde_json::json!({}); |
| 3935 | attach_child_usage_metadata(&mut metadata, &route, &Usage::default()); |
| 3936 | attach_child_usage_batch_metadata( |
| 3937 | &mut metadata, |
| 3938 | &RuntimeUsageBatch { |
| 3939 | decisions: Vec::new(), |
| 3940 | records, |
| 3941 | drop_records: Vec::new(), |
| 3942 | dropped_records: 0, |
| 3943 | }, |
| 3944 | ); |
| 3945 | |
| 3946 | let serialized = serde_json::to_string(&metadata).expect("batch metadata"); |
| 3947 | assert!(!serialized.contains("raw-provider-response-id")); |
| 3948 | let batch = child_usage_records_from_metadata(&metadata).expect("preferred batch"); |
| 3949 | assert_eq!(batch.records.len(), 2); |
| 3950 | assert_eq!(batch.records[0].usage.usage.input_tokens, 11); |
| 3951 | assert_eq!(batch.records[1].usage.usage.output_tokens, 7); |
| 3952 | assert_eq!(batch.dropped_records, 0); |
| 3953 | |
| 3954 | metadata[CHILD_USAGE_RECORDS_KEY] = serde_json::json!([{"bad": true}]); |
| 3955 | let malformed = child_usage_records_from_metadata(&metadata).expect("batch key wins"); |
| 3956 | assert!(malformed.records.is_empty()); |
| 3957 | assert_eq!(malformed.dropped_records, 1); |
| 3958 | } |
| 3959 | |
| 3960 | #[test] |
| 3961 | fn decision_receipts_metadata_round_trip_is_bounded_and_legacy_compatible() { |
| 3962 | let receipt = decision_receipt_fixture("raw-child-decision-id"); |
| 3963 | let mut metadata = serde_json::json!({}); |
| 3964 | attach_child_usage_batch_metadata( |
| 3965 | &mut metadata, |
| 3966 | &RuntimeUsageBatch { |
| 3967 | decisions: vec![receipt.clone()], |
| 3968 | ..Default::default() |
| 3969 | }, |
| 3970 | ); |
| 3971 | assert!(!metadata.to_string().contains("raw-child-decision-id")); |
| 3972 | let decoded = child_usage_records_from_metadata(&metadata).expect("batch"); |
| 3973 | assert_eq!(decoded.decisions, vec![receipt.sanitized()]); |
| 3974 | assert_eq!(decoded.dropped_records, 0); |
| 3975 | metadata[CHILD_DECISION_RECEIPTS_KEY][0]["evidence"]["response_model"] = |
| 3976 | serde_json::json!("x".repeat(129)); |
| 3977 | let invalid = child_usage_records_from_metadata(&metadata).expect("batch"); |
| 3978 | assert!(invalid.decisions.is_empty()); |
| 3979 | assert_eq!(invalid.dropped_records, 1); |
| 3980 | metadata |
| 3981 | .as_object_mut() |
| 3982 | .expect("metadata") |
| 3983 | .remove(CHILD_DECISION_RECEIPTS_KEY); |
| 3984 | assert!( |
| 3985 | child_usage_records_from_metadata(&metadata) |
| 3986 | .expect("old batch") |
| 3987 | .decisions |
| 3988 | .is_empty() |
| 3989 | ); |
| 3990 | let mut oversized = receipt; |
| 3991 | oversized.evidence.response_model = Some("x".repeat(129)); |
| 3992 | attach_child_usage_batch_metadata( |
| 3993 | &mut metadata, |
| 3994 | &RuntimeUsageBatch { |
| 3995 | decisions: vec![oversized], |
| 3996 | ..Default::default() |
| 3997 | }, |
| 3998 | ); |
| 3999 | let invalid = child_usage_records_from_metadata(&metadata).expect("bounded batch"); |
| 4000 | assert!(invalid.decisions.is_empty()); |
| 4001 | assert_eq!( |
| 4002 | invalid.dropped_records, 1, |
| 4003 | "discarded evidence must leave an explicit coverage gap" |
| 4004 | ); |
| 4005 | } |
| 4006 | |
| 4007 | fn deepseek() -> BackgroundRoute<'static> { |
| 4008 | BackgroundRoute::new(ProviderKind::Deepseek, "deepseek-v4-flash") |
| 4009 | .with_base_url(Some(crate::config::DEFAULT_DEEPSEEK_BASE_URL)) |
| 4010 | } |
| 4011 | |
| 4012 | fn deepseek_envelope() -> EffectiveRouteEnvelope { |
| 4013 | let config = crate::config::Config::default(); |
| 4014 | let identity = config |
| 4015 | .active_provider_identity() |
| 4016 | .expect("captured DeepSeek identity"); |
| 4017 | EffectiveRouteEnvelope::from_admitted( |
| 4018 | Some(&config), |
| 4019 | &identity, |
| 4020 | "deepseek-v4-flash", |
| 4021 | Some(crate::config::DEFAULT_DEEPSEEK_BASE_URL), |
| 4022 | Utc::now(), |
| 4023 | ) |
| 4024 | } |
| 4025 | |
| 4026 | #[test] |
| 4027 | fn child_cache_classes_reach_the_background_pool_only_when_reported() { |
| 4028 | // #6565: sub-agent cache was missing from session totals. |
| 4029 | let reported = background_cost_for_runtime_usage(&RuntimeUsageRecord { |
| 4030 | source_id: "child-cache-reported".into(), |
| 4031 | usage: EffectiveRouteUsage { |
| 4032 | route: deepseek_envelope(), |
| 4033 | usage: Usage { |
| 4034 | input_tokens: 1_000, |
| 4035 | output_tokens: 50, |
| 4036 | prompt_cache_hit_tokens: Some(700), |
| 4037 | prompt_cache_miss_tokens: Some(300), |
| 4038 | ..Usage::default() |
| 4039 | }, |
| 4040 | }, |
| 4041 | }); |
| 4042 | assert_eq!(reported.cache_hit_tokens, Some(700)); |
| 4043 | assert_eq!(reported.cache_miss_tokens, Some(300)); |
| 4044 | assert_eq!(reported.cache_write_tokens, Some(0)); |
| 4045 | |
| 4046 | let silent = background_cost_for_runtime_usage(&RuntimeUsageRecord { |
| 4047 | source_id: "child-cache-silent".into(), |
| 4048 | usage: EffectiveRouteUsage { |
| 4049 | route: deepseek_envelope(), |
| 4050 | usage: Usage { |
| 4051 | input_tokens: 1_000, |
| 4052 | output_tokens: 50, |
| 4053 | ..Usage::default() |
| 4054 | }, |
| 4055 | }, |
| 4056 | }); |
| 4057 | assert_eq!(silent.cache_hit_tokens, None, "no report is not 0%"); |
| 4058 | assert_eq!(silent.cache_miss_tokens, None); |
| 4059 | } |
| 4060 | |
| 4061 | #[test] |
| 4062 | fn background_cache_write_only_telemetry_is_recorded() { |
| 4063 | for written in [0, 400] { |
| 4064 | let pending = background_cost_for_runtime_usage(&RuntimeUsageRecord { |
| 4065 | source_id: "child-cache-write-only".into(), |
| 4066 | usage: EffectiveRouteUsage { |
| 4067 | route: deepseek_envelope(), |
| 4068 | usage: Usage { |
| 4069 | input_tokens: 1_000, |
| 4070 | prompt_cache_write_tokens: Some(written), |
| 4071 | ..Usage::default() |
| 4072 | }, |
| 4073 | }, |
| 4074 | }); |
| 4075 | assert_eq!(pending.cache_hit_tokens, Some(0)); |
| 4076 | assert_eq!(pending.cache_miss_tokens, Some(u64::from(1_000 - written))); |
| 4077 | assert_eq!(pending.cache_write_tokens, Some(u64::from(written))); |
| 4078 | } |
| 4079 | } |
| 4080 | |
| 4081 | #[test] |
| 4082 | fn default_usage_is_one_missing_receipt_across_canonical_replay_and_owners() { |
| 4083 | let _g = test_scope(); |
| 4084 | let route = deepseek_envelope(); |
| 4085 | let raw = "compaction:turn:response"; |
| 4086 | let fingerprint = usage_source_fingerprint(raw); |
| 4087 | let encoded = format!("routed:{fingerprint}"); |
| 4088 | for source in [raw, fingerprint.as_str(), encoded.as_str()] { |
| 4089 | report_effective_route_for_runtime( |
| 4090 | scope_token(), |
| 4091 | None, |
| 4092 | source, |
| 4093 | &route, |
| 4094 | &Usage::default(), |
| 4095 | ); |
| 4096 | } |
| 4097 | let missing = drain(); |
| 4098 | assert_eq!(missing.priced_turns, 0); |
| 4099 | assert_eq!(missing.unpriced_turns, 1); |
| 4100 | assert_eq!(missing.cny_unpriced_turns, 1); |
| 4101 | assert_eq!(missing.estimate, CostEstimate::default()); |
| 4102 | assert_eq!( |
| 4103 | missing.usage_source_fingerprints, |
| 4104 | BTreeSet::from([fingerprint.clone()]) |
| 4105 | ); |
| 4106 | assert!( |
| 4107 | missing |
| 4108 | .unpriced_reasons |
| 4109 | .contains("provider_success_missing_usage") |
| 4110 | ); |
| 4111 | report_effective_route_for_runtime( |
| 4112 | scope_token(), |
| 4113 | None, |
| 4114 | &encoded, |
| 4115 | &route, |
| 4116 | &Usage::default(), |
| 4117 | ); |
| 4118 | assert!( |
| 4119 | drain().is_empty(), |
| 4120 | "replayed metadata must stay consumed after drain" |
| 4121 | ); |
| 4122 | |
| 4123 | let owner = "runtime-default-usage-owner"; |
| 4124 | for source in [raw, fingerprint.as_str(), encoded.as_str()] { |
| 4125 | report_effective_route_for_runtime( |
| 4126 | scope_token(), |
| 4127 | Some(owner), |
| 4128 | source, |
| 4129 | &route, |
| 4130 | &Usage::default(), |
| 4131 | ); |
| 4132 | } |
| 4133 | let batch = take_runtime_usage(owner); |
| 4134 | assert!(batch.records.is_empty()); |
| 4135 | assert_eq!(batch.drop_records.len(), 1); |
| 4136 | assert_eq!(batch.dropped_records, 1); |
| 4137 | assert!(drain().is_empty()); |
| 4138 | let replay = background_cost_for_runtime_usage(&RuntimeUsageRecord { |
| 4139 | source_id: encoded, |
| 4140 | usage: EffectiveRouteUsage { |
| 4141 | route: route.clone(), |
| 4142 | usage: Usage::default(), |
| 4143 | }, |
| 4144 | }); |
| 4145 | assert_eq!(replay.unpriced_turns, missing.unpriced_turns); |
| 4146 | assert_eq!(replay.cny_unpriced_turns, missing.cny_unpriced_turns); |
| 4147 | assert_eq!( |
| 4148 | replay.usage_source_fingerprints, |
| 4149 | missing.usage_source_fingerprints |
| 4150 | ); |
| 4151 | |
| 4152 | for billing_mode in [RouteBillingMode::Subscription, RouteBillingMode::Local] { |
| 4153 | let mut nonmetered = route.clone(); |
| 4154 | nonmetered.billing_mode = billing_mode; |
| 4155 | let cost = background_cost_for_runtime_usage(&RuntimeUsageRecord { |
| 4156 | source_id: raw.into(), |
| 4157 | usage: EffectiveRouteUsage { |
| 4158 | route: nonmetered, |
| 4159 | usage: Usage::default(), |
| 4160 | }, |
| 4161 | }); |
| 4162 | assert_eq!(cost.unpriced_turns, 0); |
| 4163 | assert_eq!(cost.cny_unpriced_turns, 0); |
| 4164 | assert_eq!(cost.usage_source_fingerprints.len(), 1); |
| 4165 | } |
| 4166 | |
| 4167 | let tmp = tempfile::tempdir().unwrap(); |
| 4168 | let manager = |
| 4169 | crate::session_manager::SessionManager::new(tmp.path().join("sessions")).unwrap(); |
| 4170 | let session = crate::session_manager::create_saved_session_with_id_and_mode( |
| 4171 | "missing-origin".into(), |
| 4172 | &[], |
| 4173 | "deepseek-v4-flash", |
| 4174 | tmp.path(), |
| 4175 | 0, |
| 4176 | None, |
| 4177 | Some("agent"), |
| 4178 | ); |
| 4179 | manager.save_session(&session).unwrap(); |
| 4180 | let origin_scope = scope_token(); |
| 4181 | assert!(close_current_scope().is_empty()); |
| 4182 | assert!(report_effective_route_for_interactive_origin_with_manager( |
| 4183 | origin_scope, |
| 4184 | "missing-origin", |
| 4185 | "turn", |
| 4186 | raw, |
| 4187 | &route, |
| 4188 | &Usage::default(), |
| 4189 | &manager, |
| 4190 | )); |
| 4191 | for _ in 0..2 { |
| 4192 | let snapshot = manager.load_session_snapshot("missing-origin").unwrap(); |
| 4193 | assert_eq!(snapshot.metadata.total_tokens, 0); |
| 4194 | assert_eq!(snapshot.metadata.cost.priced_turns, 0); |
| 4195 | assert_eq!(snapshot.metadata.cost.unpriced_turns, 1); |
| 4196 | assert_eq!(snapshot.metadata.cost.cny_unpriced_turns, 1); |
| 4197 | assert_eq!(snapshot.metadata.cost.usage_source_fingerprints.len(), 1); |
| 4198 | manager.save_session(&snapshot).unwrap(); |
| 4199 | } |
| 4200 | assert!(drain().is_empty()); |
| 4201 | } |
| 4202 | |
| 4203 | #[test] |
| 4204 | fn openrouter_vendor_pin_does_not_inherit_aggregate_catalog_price() { |
| 4205 | let _env = crate::test_support::lock_test_env(); |
| 4206 | let home = tempfile::tempdir().expect("isolated catalog home"); |
| 4207 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 4208 | let _reset = ProviderCatalogTestReset; |
| 4209 | crate::provider_catalog_live::reset_cache_for_test(); |
| 4210 | let _live = crate::provider_lake::lock_live_snapshot(); |
| 4211 | crate::provider_lake::clear_live_snapshot(); |
| 4212 | let mut route = EffectiveRouteEnvelope::capture( |
| 4213 | None, |
| 4214 | ProviderKind::Openrouter, |
| 4215 | "openrouter", |
| 4216 | "qwen/qwen3.7-plus", |
| 4217 | Some(ProviderKind::Openrouter.provider().default_base_url()), |
| 4218 | Utc::now(), |
| 4219 | ); |
| 4220 | let usage = small_usage(); |
| 4221 | let aggregate = route.audit(&usage); |
| 4222 | assert!( |
| 4223 | aggregate.is_priced(), |
| 4224 | "aggregate fixture must be priced: {aggregate:?}" |
| 4225 | ); |
| 4226 | |
| 4227 | route.openrouter_vendor = Some("cerebras".to_string()); |
| 4228 | let audit = route.audit(&usage); |
| 4229 | assert_eq!( |
| 4230 | audit.unpriced_reason, |
| 4231 | Some(crate::pricing::UnpricedReason::RoutingDependentPrice) |
| 4232 | ); |
| 4233 | assert!(audit.estimate.is_none()); |
| 4234 | assert!(audit.counts_toward_money_coverage()); |
| 4235 | assert!(route.receipt(&audit).contains("openrouter_vendor=cerebras")); |
| 4236 | |
| 4237 | // Even a valid, frozen aggregate quote has no upstream-vendor dimension. |
| 4238 | let fingerprint = route |
| 4239 | .endpoint_fingerprint |
| 4240 | .clone() |
| 4241 | .expect("official endpoint"); |
| 4242 | let dispatched_at = u64::try_from(route.dispatched_at.timestamp()).expect("timestamp"); |
| 4243 | crate::provider_catalog_live::record_success(priced_provider_delta( |
| 4244 | "openrouter", |
| 4245 | &route.model, |
| 4246 | &fingerprint, |
| 4247 | dispatched_at, |
| 4248 | )); |
| 4249 | route.provider_live_pricing = |
| 4250 | crate::provider_catalog_live::fresh_provider_live_pricing_quote_at( |
| 4251 | route.provider, |
| 4252 | &route.provider_identity, |
| 4253 | &route.model, |
| 4254 | &fingerprint, |
| 4255 | dispatched_at, |
| 4256 | ); |
| 4257 | assert!(route.provider_live_pricing.is_some()); |
| 4258 | let saved: EffectiveRouteEnvelope = |
| 4259 | serde_json::from_value(serde_json::to_value(&route).unwrap()).unwrap(); |
| 4260 | let child = child_route_envelope_from_metadata(&serde_json::Value::Object( |
| 4261 | child_usage_metadata_fields(&saved, &usage), |
| 4262 | )) |
| 4263 | .expect("child envelope"); |
| 4264 | for receipt in [&route, &saved, &child] { |
| 4265 | assert_eq!( |
| 4266 | receipt.audit(&usage).unpriced_reason, |
| 4267 | Some(crate::pricing::UnpricedReason::RoutingDependentPrice) |
| 4268 | ); |
| 4269 | } |
| 4270 | |
| 4271 | for (billing_mode, reason) in [ |
| 4272 | ( |
| 4273 | RouteBillingMode::Subscription, |
| 4274 | crate::pricing::UnpricedReason::NotMoneyMetered, |
| 4275 | ), |
| 4276 | ( |
| 4277 | RouteBillingMode::Local, |
| 4278 | crate::pricing::UnpricedReason::NotMoneyMetered, |
| 4279 | ), |
| 4280 | ( |
| 4281 | RouteBillingMode::Unknown, |
| 4282 | crate::pricing::UnpricedReason::UnknownBillingBasis, |
| 4283 | ), |
| 4284 | ] { |
| 4285 | route.billing_mode = billing_mode; |
| 4286 | assert_eq!(route.audit(&usage).unpriced_reason, Some(reason)); |
| 4287 | } |
| 4288 | } |
| 4289 | |
| 4290 | #[test] |
| 4291 | fn openrouter_vendor_pin_survives_envelope_and_child_metadata_persistence() { |
| 4292 | let mut config = crate::config::Config { |
| 4293 | provider: Some("openrouter".to_string()), |
| 4294 | ..Default::default() |
| 4295 | }; |
| 4296 | config |
| 4297 | .provider_config_for_mut(&config.test_identity_for_kind(ProviderKind::Openrouter)) |
| 4298 | .unwrap() |
| 4299 | .vendor = Some("cerebras".to_string()); |
| 4300 | let route = EffectiveRouteEnvelope::capture( |
| 4301 | Some(&config), |
| 4302 | ProviderKind::Openrouter, |
| 4303 | "openrouter", |
| 4304 | "qwen/qwen3.7-plus", |
| 4305 | Some(ProviderKind::Openrouter.provider().default_base_url()), |
| 4306 | Utc::now(), |
| 4307 | ); |
| 4308 | config |
| 4309 | .provider_config_for_mut(&config.test_identity_for_kind(ProviderKind::Openrouter)) |
| 4310 | .unwrap() |
| 4311 | .vendor = None; |
| 4312 | assert_eq!(route.openrouter_vendor.as_deref(), Some("cerebras")); |
| 4313 | |
| 4314 | let mut json = serde_json::to_value(&route).expect("serialize route"); |
| 4315 | let restored: EffectiveRouteEnvelope = |
| 4316 | serde_json::from_value(json.clone()).expect("restore route"); |
| 4317 | assert_eq!(restored, route); |
| 4318 | let metadata = |
| 4319 | serde_json::Value::Object(child_usage_metadata_fields(&route, &small_usage())); |
| 4320 | assert_eq!(child_route_envelope_from_metadata(&metadata), Some(route)); |
| 4321 | |
| 4322 | json.as_object_mut() |
| 4323 | .expect("route object") |
| 4324 | .remove("openrouter_vendor"); |
| 4325 | let legacy: EffectiveRouteEnvelope = serde_json::from_value(json).expect("legacy route"); |
| 4326 | assert_eq!(legacy.openrouter_vendor, None); |
| 4327 | } |
| 4328 | |
| 4329 | #[test] |
| 4330 | fn child_metadata_round_trip_preserves_zero_and_reasoning_usage() { |
| 4331 | let route = deepseek_envelope(); |
| 4332 | let usage = Usage { |
| 4333 | input_tokens: 0, |
| 4334 | output_tokens: 9, |
| 4335 | reasoning_tokens: Some(7), |
| 4336 | reasoning_replay_tokens: Some(3), |
| 4337 | ..Usage::default() |
| 4338 | }; |
| 4339 | let mut metadata = serde_json::json!({"tool": "rlm_eval"}); |
| 4340 | attach_child_usage_metadata(&mut metadata, &route, &usage); |
| 4341 | |
| 4342 | assert_eq!(child_route_envelope_from_metadata(&metadata), Some(route)); |
| 4343 | assert_eq!(child_usage_from_metadata(&metadata), Some(usage)); |
| 4344 | |
| 4345 | let mut zero_metadata = serde_json::json!({}); |
| 4346 | let zero = Usage::default(); |
| 4347 | attach_child_usage_metadata(&mut zero_metadata, &deepseek_envelope(), &zero); |
| 4348 | assert_eq!(child_usage_from_metadata(&zero_metadata), Some(zero)); |
| 4349 | } |
| 4350 | |
| 4351 | #[test] |
| 4352 | fn runtime_owned_usage_is_isolated_from_tui_pool() { |
| 4353 | let _g = test_scope(); |
| 4354 | let route = deepseek_envelope(); |
| 4355 | let usage = small_usage(); |
| 4356 | report_effective_route_for_runtime( |
| 4357 | scope_token(), |
| 4358 | Some("turn-a"), |
| 4359 | "response-a", |
| 4360 | &route, |
| 4361 | &usage, |
| 4362 | ); |
| 4363 | report_effective_route_for_runtime( |
| 4364 | scope_token(), |
| 4365 | Some("turn-b"), |
| 4366 | "response-b", |
| 4367 | &route, |
| 4368 | &usage, |
| 4369 | ); |
| 4370 | |
| 4371 | assert_eq!(take_runtime_usage("turn-a").records.len(), 1); |
| 4372 | assert!(take_runtime_usage("turn-a").records.is_empty()); |
| 4373 | assert_eq!(take_runtime_usage("turn-b").records.len(), 1); |
| 4374 | assert!( |
| 4375 | drain().is_empty(), |
| 4376 | "runtime-owned usage must not enter TUI cost" |
| 4377 | ); |
| 4378 | |
| 4379 | report_effective_route_for_runtime(scope_token(), None, "response-tui", &route, &usage); |
| 4380 | assert_eq!(drain().priced_turns, 1, "ownerless usage belongs to TUI"); |
| 4381 | } |
| 4382 | |
| 4383 | /// Every piece of shared cost accounting is scoped to the test that owns |
| 4384 | /// it, including the durability sink registry. |
| 4385 | /// |
| 4386 | /// Sinks are keyed by owner id, and owner ids in tests are short fixture |
| 4387 | /// strings that repeat. A process-global registry let one test's |
| 4388 | /// `register_runtime_usage_sink` overwrite another's live sink, and let one |
| 4389 | /// test's `finish_runtime_usage_owner` retire it mid-flight — so a passing |
| 4390 | /// exactly-once assertion depended on which tests happened to run |
| 4391 | /// concurrently. This pins the isolation directly: a sink registered on |
| 4392 | /// another thread must be invisible here, and usage reported here must not |
| 4393 | /// reach it. |
| 4394 | #[test] |
| 4395 | fn runtime_usage_sinks_do_not_leak_across_test_threads() { |
| 4396 | let _g = test_scope(); |
| 4397 | let owner = "shared-owner"; |
| 4398 | let other_thread_deliveries = Arc::new(std::sync::atomic::AtomicUsize::new(0)); |
| 4399 | |
| 4400 | // A concurrent test, standing in for any other test in the binary that |
| 4401 | // happens to use the same owner id. |
| 4402 | let deliveries = Arc::clone(&other_thread_deliveries); |
| 4403 | let (ready_tx, ready_rx) = std::sync::mpsc::channel(); |
| 4404 | let (done_tx, done_rx) = std::sync::mpsc::channel(); |
| 4405 | let other = std::thread::spawn(move || { |
| 4406 | register_runtime_usage_sink( |
| 4407 | owner, |
| 4408 | Arc::new(move |_record| { |
| 4409 | deliveries.fetch_add(1, std::sync::atomic::Ordering::SeqCst); |
| 4410 | true |
| 4411 | }), |
| 4412 | ); |
| 4413 | ready_tx.send(()).expect("signal registration"); |
| 4414 | // Hold the registration open across this thread's assertions. |
| 4415 | done_rx.recv().expect("wait for the other test to finish"); |
| 4416 | // The other thread's own reports still reach its own sink. |
| 4417 | report_effective_route_for_runtime( |
| 4418 | scope_token(), |
| 4419 | Some(owner), |
| 4420 | "response-other", |
| 4421 | &deepseek_envelope(), |
| 4422 | &small_usage(), |
| 4423 | ); |
| 4424 | }); |
| 4425 | ready_rx.recv().expect("other test registered its sink"); |
| 4426 | |
| 4427 | // This thread never registered a sink, so its usage must fall through |
| 4428 | // to this thread's journal — not into the other test's sink. |
| 4429 | report_effective_route_for_runtime( |
| 4430 | scope_token(), |
| 4431 | Some(owner), |
| 4432 | "response-mine", |
| 4433 | &deepseek_envelope(), |
| 4434 | &small_usage(), |
| 4435 | ); |
| 4436 | assert_eq!( |
| 4437 | other_thread_deliveries.load(std::sync::atomic::Ordering::SeqCst), |
| 4438 | 0, |
| 4439 | "another test's sink received this test's usage" |
| 4440 | ); |
| 4441 | let mine = take_runtime_usage(owner); |
| 4442 | assert_eq!(mine.records.len(), 1); |
| 4443 | assert_eq!(mine.records[0].source_id, "response-mine"); |
| 4444 | assert_eq!(mine.dropped_records, 0); |
| 4445 | |
| 4446 | // Retiring the owner here must not retire the other test's sink. |
| 4447 | finish_runtime_usage_owner(owner); |
| 4448 | done_tx.send(()).expect("release the other test"); |
| 4449 | other.join().expect("other test thread"); |
| 4450 | assert_eq!( |
| 4451 | other_thread_deliveries.load(std::sync::atomic::Ordering::SeqCst), |
| 4452 | 1, |
| 4453 | "the other test's sink was retired by an unrelated test" |
| 4454 | ); |
| 4455 | } |
| 4456 | |
| 4457 | #[test] |
| 4458 | fn runtime_usage_fallback_is_bounded_and_reports_truncation() { |
| 4459 | let _g = test_scope(); |
| 4460 | let route = deepseek_envelope(); |
| 4461 | for index in 0..(MAX_RUNTIME_USAGE_RECORDS_PER_OWNER + 3) { |
| 4462 | report_effective_route_for_runtime( |
| 4463 | scope_token(), |
| 4464 | Some("turn-bounded"), |
| 4465 | &format!("response-{index}"), |
| 4466 | &route, |
| 4467 | &small_usage(), |
| 4468 | ); |
| 4469 | } |
| 4470 | |
| 4471 | let batch = take_runtime_usage("turn-bounded"); |
| 4472 | assert_eq!(batch.records.len(), MAX_RUNTIME_USAGE_RECORDS_PER_OWNER); |
| 4473 | assert_eq!(batch.dropped_records, 3); |
| 4474 | assert!(drain().is_empty(), "runtime fallback must stay out of TUI"); |
| 4475 | } |
| 4476 | |
| 4477 | #[test] |
| 4478 | fn route_labels_redact_local_paths_but_preserve_model_namespaces() { |
| 4479 | let route = EffectiveRouteEnvelope { |
| 4480 | openrouter_vendor: None, |
| 4481 | provider: ProviderKind::Openrouter, |
| 4482 | provider_identity: "/Users/alice/.config/provider-secret".to_string(), |
| 4483 | model: "/Volumes/private/checkpoints/model.gguf".to_string(), |
| 4484 | billing_surface: None, |
| 4485 | endpoint_fingerprint: None, |
| 4486 | provider_live_pricing: None, |
| 4487 | billing_mode: RouteBillingMode::Metered, |
| 4488 | dispatched_at: Utc::now(), |
| 4489 | }; |
| 4490 | let sanitized = route.sanitized_for_persistence(); |
| 4491 | assert_eq!(sanitized.provider_identity, "redacted-local-path"); |
| 4492 | assert_eq!(sanitized.model, "redacted-local-path"); |
| 4493 | let receipt = route.receipt(&TurnCostAudit::unpriced( |
| 4494 | crate::pricing::UnpricedReason::NoPricingRow, |
| 4495 | )); |
| 4496 | assert!(!receipt.contains("alice")); |
| 4497 | assert!(!receipt.contains("Volumes")); |
| 4498 | |
| 4499 | assert_eq!( |
| 4500 | sanitize_persisted_route_label("anthropic/claude-sonnet-5"), |
| 4501 | "anthropic/claude-sonnet-5" |
| 4502 | ); |
| 4503 | } |
| 4504 | |
| 4505 | #[test] |
| 4506 | fn route_label_sanitizer_rejects_credentials_urls_and_relative_paths() { |
| 4507 | for credential in [ |
| 4508 | "Bearer secret-token", |
| 4509 | "Authorization: Basic abc123", |
| 4510 | "OPENAI_API_KEY=sk-secret", |
| 4511 | "service_token: ghp_secret", |
| 4512 | "hf_secret-token", |
| 4513 | "glpat-secret-token", |
| 4514 | "db-password=hunter2", |
| 4515 | "sk-live-secret", |
| 4516 | "https://alice:password@example.test/v1?api_key=secret#fragment", |
| 4517 | ] { |
| 4518 | let sanitized = sanitize_persisted_route_label(credential); |
| 4519 | assert!( |
| 4520 | sanitized.starts_with("redacted-"), |
| 4521 | "credential was not redacted: {credential:?} -> {sanitized:?}" |
| 4522 | ); |
| 4523 | } |
| 4524 | for path in [ |
| 4525 | ".ssh/id_ed25519", |
| 4526 | "../secrets/provider.key", |
| 4527 | "workspace/.ssh/config", |
| 4528 | "relative/path/to/credential", |
| 4529 | r"relative\path\credential", |
| 4530 | ] { |
| 4531 | assert_eq!( |
| 4532 | sanitize_persisted_route_label(path), |
| 4533 | "redacted-local-path", |
| 4534 | "path was not redacted: {path:?}" |
| 4535 | ); |
| 4536 | } |
| 4537 | assert_eq!( |
| 4538 | sanitize_persisted_route_label("moonshot/kimi-k3"), |
| 4539 | "moonshot/kimi-k3" |
| 4540 | ); |
| 4541 | } |
| 4542 | |
| 4543 | #[test] |
| 4544 | fn serialized_route_envelopes_records_and_child_receipts_are_secret_free() { |
| 4545 | let route = EffectiveRouteEnvelope { |
| 4546 | openrouter_vendor: Some("Authorization: Bearer vendor-secret".to_string()), |
| 4547 | provider: ProviderKind::Custom, |
| 4548 | provider_identity: "Authorization: Bearer provider-secret".to_string(), |
| 4549 | model: "MODEL_API_KEY=sk-model-secret".to_string(), |
| 4550 | billing_surface: Some( |
| 4551 | "https://alice:password@example.test/v1?token=secret#fragment".to_string(), |
| 4552 | ), |
| 4553 | endpoint_fingerprint: Some("../.ssh/provider_key".to_string()), |
| 4554 | provider_live_pricing: None, |
| 4555 | billing_mode: RouteBillingMode::Metered, |
| 4556 | dispatched_at: Utc::now(), |
| 4557 | }; |
| 4558 | let usage = Usage { |
| 4559 | input_tokens: 7, |
| 4560 | output_tokens: 3, |
| 4561 | ..Usage::default() |
| 4562 | }; |
| 4563 | |
| 4564 | let envelope_json = serde_json::to_string(&route).expect("serialize envelope"); |
| 4565 | let record_json = serde_json::to_string(&EffectiveRouteUsage { |
| 4566 | route: route.clone(), |
| 4567 | usage: usage.clone(), |
| 4568 | }) |
| 4569 | .expect("serialize route usage"); |
| 4570 | let child_json = serde_json::to_string(&child_usage_metadata_fields(&route, &usage)) |
| 4571 | .expect("serialize child receipt"); |
| 4572 | for serialized in [&envelope_json, &record_json, &child_json] { |
| 4573 | for secret in [ |
| 4574 | "vendor-secret", |
| 4575 | "provider-secret", |
| 4576 | "sk-model-secret", |
| 4577 | "alice", |
| 4578 | "password", |
| 4579 | "token=secret", |
| 4580 | ".ssh", |
| 4581 | ] { |
| 4582 | assert!( |
| 4583 | !serialized.contains(secret), |
| 4584 | "serialized route leaked {secret:?}: {serialized}" |
| 4585 | ); |
| 4586 | } |
| 4587 | } |
| 4588 | } |
| 4589 | |
| 4590 | #[test] |
| 4591 | fn report_adds_to_pool_and_drain_returns_then_resets() { |
| 4592 | let _g = test_scope(); |
| 4593 | report(scope_token(), &deepseek(), &small_usage()); |
| 4594 | let first = drain(); |
| 4595 | assert!( |
| 4596 | first.estimate.usd > 0.0, |
| 4597 | "expected positive USD cost, got {first:?}" |
| 4598 | ); |
| 4599 | assert!( |
| 4600 | first.estimate.cny > 0.0, |
| 4601 | "expected positive CNY cost, got {first:?}" |
| 4602 | ); |
| 4603 | assert_eq!(first.priced_turns, 1); |
| 4604 | assert_eq!(first.unpriced_turns, 0); |
| 4605 | assert_eq!(first.cny_priced_turns, 1); |
| 4606 | assert_eq!(first.cny_unpriced_turns, 0); |
| 4607 | // The receipt names the route without leaking the endpoint URL. |
| 4608 | assert_eq!(first.route_receipts.len(), 1); |
| 4609 | let receipt = first.route_receipts.iter().next().expect("receipt"); |
| 4610 | assert!(receipt.contains("provider=deepseek"), "{receipt}"); |
| 4611 | assert!(receipt.contains("model=deepseek-v4-flash"), "{receipt}"); |
| 4612 | assert!(receipt.contains("currency=usd+cny"), "{receipt}"); |
| 4613 | assert!(!receipt.contains("http"), "{receipt}"); |
| 4614 | |
| 4615 | let second = drain(); |
| 4616 | assert!(second.is_empty(), "drain must zero the pool: {second:?}"); |
| 4617 | } |
| 4618 | |
| 4619 | #[test] |
| 4620 | fn reports_from_a_closed_session_scope_are_discarded() { |
| 4621 | let _g = test_scope(); |
| 4622 | let old_scope = scope_token(); |
| 4623 | let settled = close_current_scope(); |
| 4624 | assert!(settled.is_empty()); |
| 4625 | |
| 4626 | report(old_scope, &deepseek(), &small_usage()); |
| 4627 | assert!(drain().is_empty(), "old session usage crossed the boundary"); |
| 4628 | |
| 4629 | report(scope_token(), &deepseek(), &small_usage()); |
| 4630 | assert_eq!(drain().priced_turns, 1); |
| 4631 | } |
| 4632 | |
| 4633 | #[test] |
| 4634 | fn retired_origin_acknowledges_sources_without_accrual_or_scope_leak() { |
| 4635 | let _g = test_scope(); |
| 4636 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 4637 | let manager = crate::session_manager::SessionManager::new(tmp.path().join("sessions")) |
| 4638 | .expect("manager"); |
| 4639 | let session = crate::session_manager::create_saved_session_with_id_and_mode( |
| 4640 | "retired-origin".to_string(), |
| 4641 | &[], |
| 4642 | "deepseek-v4-flash", |
| 4643 | tmp.path(), |
| 4644 | 0, |
| 4645 | None, |
| 4646 | Some("agent"), |
| 4647 | ); |
| 4648 | manager.save_session(&session).expect("save origin"); |
| 4649 | let origin_scope = scope_token(); |
| 4650 | manager |
| 4651 | .delete_session("retired-origin") |
| 4652 | .expect("delete origin"); |
| 4653 | let route = deepseek_envelope(); |
| 4654 | for _ in 0..2 { |
| 4655 | assert!(report_effective_route_for_interactive_origin_with_manager( |
| 4656 | origin_scope, |
| 4657 | "retired-origin", |
| 4658 | "origin-turn", |
| 4659 | "retired-usage", |
| 4660 | &route, |
| 4661 | &small_usage(), |
| 4662 | &manager, |
| 4663 | )); |
| 4664 | assert!(report_unreceipted_for_interactive_origin_with_manager( |
| 4665 | origin_scope, |
| 4666 | "retired-origin", |
| 4667 | "origin-turn", |
| 4668 | "retired-drop", |
| 4669 | &route, |
| 4670 | &manager, |
| 4671 | )); |
| 4672 | } |
| 4673 | assert!(usage_source_seen("retired-usage")); |
| 4674 | assert!(usage_source_seen("retired-drop")); |
| 4675 | assert!( |
| 4676 | drain().is_empty(), |
| 4677 | "retirement must not create any pending projection" |
| 4678 | ); |
| 4679 | assert!( |
| 4680 | !manager |
| 4681 | .sessions_dir() |
| 4682 | .join(".late-usage/retired-origin.json") |
| 4683 | .exists() |
| 4684 | ); |
| 4685 | |
| 4686 | assert!(close_current_scope().is_empty()); |
| 4687 | assert!(!usage_source_seen("retired-usage")); |
| 4688 | assert!(report_effective_route_for_interactive_origin_with_manager( |
| 4689 | origin_scope, |
| 4690 | "retired-origin", |
| 4691 | "origin-turn", |
| 4692 | "after-scope-change", |
| 4693 | &route, |
| 4694 | &small_usage(), |
| 4695 | &manager, |
| 4696 | )); |
| 4697 | assert!( |
| 4698 | !usage_source_seen("after-scope-change"), |
| 4699 | "old retirement cannot poison a new scope" |
| 4700 | ); |
| 4701 | assert!(drain().is_empty()); |
| 4702 | report_effective_route_for_runtime( |
| 4703 | scope_token(), |
| 4704 | None, |
| 4705 | "after-scope-change", |
| 4706 | &route, |
| 4707 | &small_usage(), |
| 4708 | ); |
| 4709 | assert_eq!( |
| 4710 | drain().priced_turns, |
| 4711 | 1, |
| 4712 | "the replacement scope still admits its own response" |
| 4713 | ); |
| 4714 | } |
| 4715 | |
| 4716 | #[test] |
| 4717 | fn detached_advisor_and_translation_receipts_survive_new_exactly_once() { |
| 4718 | let _g = test_scope(); |
| 4719 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 4720 | let manager = crate::session_manager::SessionManager::new(tmp.path().join("sessions")) |
| 4721 | .expect("session manager"); |
| 4722 | let old_session_id = "origin-session"; |
| 4723 | let new_session_id = "replacement-session"; |
| 4724 | for session_id in [old_session_id, new_session_id] { |
| 4725 | let session = crate::session_manager::create_saved_session_with_id_and_mode( |
| 4726 | session_id.to_string(), |
| 4727 | &[], |
| 4728 | "deepseek-v4-flash", |
| 4729 | tmp.path(), |
| 4730 | 0, |
| 4731 | None, |
| 4732 | Some("agent"), |
| 4733 | ); |
| 4734 | manager.save_session(&session).expect("save session"); |
| 4735 | } |
| 4736 | |
| 4737 | let origin_scope = scope_token(); |
| 4738 | let owner = "interactive:origin-session:origin-turn"; |
| 4739 | register_persistent_interactive_runtime_usage_sink_at( |
| 4740 | owner, |
| 4741 | origin_scope, |
| 4742 | old_session_id, |
| 4743 | "origin-turn", |
| 4744 | manager.sessions_dir().to_path_buf(), |
| 4745 | ); |
| 4746 | let advisor_lease = acquire_runtime_usage_lease(owner).expect("advisor owner lease"); |
| 4747 | finish_runtime_usage_owner(owner); |
| 4748 | |
| 4749 | // `/new` closes the old foreground generation while the detached |
| 4750 | // advisor and translation requests are still in flight. |
| 4751 | assert!(close_current_scope().is_empty()); |
| 4752 | let route = deepseek_envelope(); |
| 4753 | let usage = Usage { |
| 4754 | input_tokens: 17, |
| 4755 | output_tokens: 5, |
| 4756 | ..Usage::default() |
| 4757 | }; |
| 4758 | for _ in 0..2 { |
| 4759 | report_effective_route_for_runtime( |
| 4760 | origin_scope, |
| 4761 | Some(owner), |
| 4762 | "advisor:origin-turn:response", |
| 4763 | &route, |
| 4764 | &usage, |
| 4765 | ); |
| 4766 | report_unreceipted_provider_success( |
| 4767 | origin_scope, |
| 4768 | Some(owner), |
| 4769 | "advisor:origin-turn:missing-usage", |
| 4770 | &route, |
| 4771 | ); |
| 4772 | assert!(report_effective_route_for_interactive_origin_with_manager( |
| 4773 | origin_scope, |
| 4774 | old_session_id, |
| 4775 | "origin-turn", |
| 4776 | "translation:origin-turn:assistant", |
| 4777 | &route, |
| 4778 | &usage, |
| 4779 | &manager, |
| 4780 | )); |
| 4781 | assert!(report_unreceipted_for_interactive_origin_with_manager( |
| 4782 | origin_scope, |
| 4783 | old_session_id, |
| 4784 | "origin-turn", |
| 4785 | "translation:origin-turn:thinking-missing-usage", |
| 4786 | &route, |
| 4787 | &manager, |
| 4788 | )); |
| 4789 | } |
| 4790 | drop(advisor_lease); |
| 4791 | |
| 4792 | let fallback = take_runtime_usage(owner); |
| 4793 | assert!(fallback.records.is_empty()); |
| 4794 | assert!(fallback.drop_records.is_empty()); |
| 4795 | assert_eq!(fallback.dropped_records, 0); |
| 4796 | assert!(drain().is_empty(), "late receipts polluted the new scope"); |
| 4797 | |
| 4798 | let old = manager |
| 4799 | .load_session_snapshot(old_session_id) |
| 4800 | .expect("load origin session"); |
| 4801 | assert_eq!(old.metadata.total_tokens, 44); |
| 4802 | assert_eq!(old.metadata.cost.priced_turns, 2); |
| 4803 | assert_eq!(old.metadata.cost.unpriced_turns, 2); |
| 4804 | assert_eq!(old.metadata.cost.cny_unpriced_turns, 2); |
| 4805 | assert_eq!(old.metadata.cost.usage_source_fingerprints.len(), 4); |
| 4806 | |
| 4807 | let replay = manager |
| 4808 | .load_session_snapshot(old_session_id) |
| 4809 | .expect("replay origin session"); |
| 4810 | assert_eq!(replay.metadata.total_tokens, 44); |
| 4811 | assert_eq!(replay.metadata.cost.usage_source_fingerprints.len(), 4); |
| 4812 | |
| 4813 | let replacement = manager |
| 4814 | .load_session_snapshot(new_session_id) |
| 4815 | .expect("load replacement session"); |
| 4816 | assert_eq!(replacement.metadata.total_tokens, 0); |
| 4817 | assert_eq!(replacement.metadata.cost.priced_turns, 0); |
| 4818 | assert_eq!(replacement.metadata.cost.unpriced_turns, 0); |
| 4819 | assert!( |
| 4820 | replacement |
| 4821 | .metadata |
| 4822 | .cost |
| 4823 | .usage_source_fingerprints |
| 4824 | .is_empty() |
| 4825 | ); |
| 4826 | } |
| 4827 | |
| 4828 | #[test] |
| 4829 | fn report_counts_unknown_models_as_missing_spend_not_as_free() { |
| 4830 | let _g = test_scope(); |
| 4831 | // NIM-hosted models intentionally have no DeepSeek pricing, but the |
| 4832 | // route *is* money-metered — so the turn is missing spend, not absent. |
| 4833 | report( |
| 4834 | scope_token(), |
| 4835 | &BackgroundRoute::new(ProviderKind::NvidiaNim, "deepseek-ai/deepseek-v4-pro"), |
| 4836 | &small_usage(), |
| 4837 | ); |
| 4838 | let drained = drain(); |
| 4839 | assert_eq!(drained.estimate, CostEstimate::default()); |
| 4840 | assert_eq!(drained.priced_turns, 0); |
| 4841 | assert_eq!(drained.unpriced_turns, 1); |
| 4842 | assert!(!drained.unpriced_reasons.is_empty()); |
| 4843 | } |
| 4844 | |
| 4845 | #[test] |
| 4846 | fn report_skips_codex_oauth_pricing_without_calling_it_incomplete() { |
| 4847 | let _g = test_scope(); |
| 4848 | report( |
| 4849 | scope_token(), |
| 4850 | &BackgroundRoute::new(ProviderKind::OpenaiCodex, "gpt-5.5") |
| 4851 | .with_base_url(Some("https://chatgpt.com/backend-api/codex")), |
| 4852 | &small_usage(), |
| 4853 | ); |
| 4854 | let drained = drain(); |
| 4855 | assert_eq!(drained.estimate, CostEstimate::default()); |
| 4856 | // Exactly non-metered: not counted in either coverage bucket. |
| 4857 | assert_eq!(drained.priced_turns, 0); |
| 4858 | assert_eq!(drained.unpriced_turns, 0); |
| 4859 | assert!(drained.unpriced_reasons.is_empty()); |
| 4860 | assert!(drained.cny_unpriced_reasons.is_empty()); |
| 4861 | } |
| 4862 | |
| 4863 | #[test] |
| 4864 | fn report_skips_stepfun_without_billing_surface() { |
| 4865 | let _g = test_scope(); |
| 4866 | report( |
| 4867 | scope_token(), |
| 4868 | &BackgroundRoute::new(ProviderKind::Stepfun, "step-3.7-flash"), |
| 4869 | &small_usage(), |
| 4870 | ); |
| 4871 | report( |
| 4872 | scope_token(), |
| 4873 | &BackgroundRoute::new(ProviderKind::Openrouter, "step-3.7-flash"), |
| 4874 | &small_usage(), |
| 4875 | ); |
| 4876 | let drained = drain(); |
| 4877 | assert_eq!(drained.estimate, CostEstimate::default()); |
| 4878 | // Both are metered-or-unknown routes that could not be priced, so both |
| 4879 | // are reported as missing rather than dropped. |
| 4880 | assert_eq!(drained.unpriced_turns, 2); |
| 4881 | } |
| 4882 | |
| 4883 | /// A local runtime and a plan endpoint must never be guessed into public |
| 4884 | /// per-token dollars just because the provider also sells a paid API. |
| 4885 | #[test] |
| 4886 | fn local_and_plan_endpoints_are_never_treated_as_public_payg() { |
| 4887 | let _g = test_scope(); |
| 4888 | report( |
| 4889 | scope_token(), |
| 4890 | &BackgroundRoute::new(ProviderKind::Ollama, "llama3.2"), |
| 4891 | &small_usage(), |
| 4892 | ); |
| 4893 | report( |
| 4894 | scope_token(), |
| 4895 | &BackgroundRoute::new(ProviderKind::Zai, "glm-5.2") |
| 4896 | .with_base_url(Some("https://api.z.ai/api/coding/paas/v4")), |
| 4897 | &small_usage(), |
| 4898 | ); |
| 4899 | report( |
| 4900 | scope_token(), |
| 4901 | &BackgroundRoute::new(ProviderKind::Moonshot, "kimi-for-coding") |
| 4902 | .with_base_url(Some(crate::config::DEFAULT_KIMI_CODE_BASE_URL)), |
| 4903 | &small_usage(), |
| 4904 | ); |
| 4905 | let drained = drain(); |
| 4906 | assert_eq!(drained.estimate, CostEstimate::default()); |
| 4907 | assert_eq!(drained.priced_turns, 0); |
| 4908 | assert_eq!( |
| 4909 | drained.unpriced_turns, 0, |
| 4910 | "exactly non-metered routes are not missing dollars: {drained:?}" |
| 4911 | ); |
| 4912 | assert!(drained.unpriced_reasons.is_empty()); |
| 4913 | assert!(drained.cny_unpriced_reasons.is_empty()); |
| 4914 | assert!( |
| 4915 | drained |
| 4916 | .route_receipts |
| 4917 | .iter() |
| 4918 | .any(|receipt| receipt.contains("surface=zai-coding-plan")), |
| 4919 | "{drained:?}" |
| 4920 | ); |
| 4921 | assert!( |
| 4922 | drained |
| 4923 | .route_receipts |
| 4924 | .iter() |
| 4925 | .any(|receipt| receipt.contains("surface=local-no-bill")), |
| 4926 | "{drained:?}" |
| 4927 | ); |
| 4928 | assert!( |
| 4929 | drained |
| 4930 | .route_receipts |
| 4931 | .iter() |
| 4932 | .any(|receipt| receipt.contains("surface=moonshot-kimi-code")), |
| 4933 | "{drained:?}" |
| 4934 | ); |
| 4935 | } |
| 4936 | |
| 4937 | /// The receipt carries an endpoint *fingerprint*, never the URL. |
| 4938 | #[test] |
| 4939 | fn route_receipts_fingerprint_the_endpoint_and_keep_secrets_out() { |
| 4940 | let _g = test_scope(); |
| 4941 | let base_url = "https://api.deepseek.com/v1"; |
| 4942 | report( |
| 4943 | scope_token(), |
| 4944 | &deepseek().with_base_url(Some(base_url)), |
| 4945 | &small_usage(), |
| 4946 | ); |
| 4947 | let drained = drain(); |
| 4948 | let receipt = drained.route_receipts.iter().next().expect("receipt"); |
| 4949 | let expected_fp = endpoint_fingerprint(base_url).expect("valid endpoint fingerprint"); |
| 4950 | assert!( |
| 4951 | receipt.contains(&format!("endpoint_fp={expected_fp}")), |
| 4952 | "{receipt}" |
| 4953 | ); |
| 4954 | for needle in ["http", "api.deepseek.com", "sk-", "/Users/", "/home/"] { |
| 4955 | assert!(!receipt.contains(needle), "{needle} leaked into {receipt}"); |
| 4956 | } |
| 4957 | } |
| 4958 | |
| 4959 | #[test] |
| 4960 | fn receipt_fields_are_bounded_and_secret_bearing_urls_are_not_hashed() { |
| 4961 | let hostile = format!("model\nAuthorization: bearer {}", "x".repeat(400)); |
| 4962 | let receipt = route_receipt( |
| 4963 | ProviderKind::Deepseek, |
| 4964 | Some("identity\r\nforged=yes"), |
| 4965 | &hostile, |
| 4966 | Some(crate::pricing::FIRST_PARTY_PAYG_BILLING_SURFACE), |
| 4967 | None, |
| 4968 | RouteBillingMode::Metered, |
| 4969 | "usd+cny", |
| 4970 | ); |
| 4971 | assert!(!receipt.contains('\n'), "{receipt}"); |
| 4972 | assert!(!receipt.contains('\r'), "{receipt}"); |
| 4973 | assert!( |
| 4974 | receipt.len() < 420, |
| 4975 | "receipt was not bounded: {}", |
| 4976 | receipt.len() |
| 4977 | ); |
| 4978 | |
| 4979 | for secret_url in [ |
| 4980 | "https://user:secret@api.example.com/v1", |
| 4981 | "https://api.example.com/v1?api_key=secret", |
| 4982 | "https://api.example.com/v1#secret", |
| 4983 | ] { |
| 4984 | assert_eq!(endpoint_fingerprint(secret_url), None, "{secret_url}"); |
| 4985 | } |
| 4986 | assert_eq!( |
| 4987 | endpoint_fingerprint("https://API.Example.com/v1/") |
| 4988 | .expect("valid endpoint") |
| 4989 | .len(), |
| 4990 | 64 |
| 4991 | ); |
| 4992 | } |
| 4993 | |
| 4994 | #[test] |
| 4995 | fn report_accumulates_across_multiple_calls() { |
| 4996 | let _g = test_scope(); |
| 4997 | report(scope_token(), &deepseek(), &small_usage()); |
| 4998 | report(scope_token(), &deepseek(), &small_usage()); |
| 4999 | let total = drain(); |
| 5000 | // Two equal reports — total must be 2× a single report. |
| 5001 | let single = crate::pricing::calculate_turn_cost_estimate_from_usage( |
| 5002 | "deepseek-v4-flash", |
| 5003 | &small_usage(), |
| 5004 | ) |
| 5005 | .unwrap(); |
| 5006 | assert!((total.estimate.usd - 2.0 * single.usd).abs() < 1e-12); |
| 5007 | assert!((total.estimate.cny - 2.0 * single.cny).abs() < 1e-12); |
| 5008 | assert_eq!(total.priced_turns, 2); |
| 5009 | // Identical routes collapse to one receipt rather than growing without |
| 5010 | // bound across a long session. |
| 5011 | assert_eq!(total.route_receipts.len(), 1); |
| 5012 | } |
| 5013 | |
| 5014 | /// A cache-write turn on a route with no published write rate must show up |
| 5015 | /// as missing spend naming the class, not as a discounted total. |
| 5016 | #[test] |
| 5017 | fn unpriced_cache_write_class_is_reported_not_absorbed() { |
| 5018 | let _g = test_scope(); |
| 5019 | let write_heavy = Usage { |
| 5020 | input_tokens: 1_000_000, |
| 5021 | output_tokens: 100_000, |
| 5022 | prompt_cache_hit_tokens: Some(200_000), |
| 5023 | prompt_cache_write_tokens: Some(100_000), |
| 5024 | ..Default::default() |
| 5025 | }; |
| 5026 | report( |
| 5027 | scope_token(), |
| 5028 | &BackgroundRoute::new(ProviderKind::Moonshot, "kimi-k2.7-code") |
| 5029 | .with_base_url(Some("https://api.moonshot.ai/v1")), |
| 5030 | &write_heavy, |
| 5031 | ); |
| 5032 | let drained = drain(); |
| 5033 | assert_eq!(drained.estimate, CostEstimate::default()); |
| 5034 | assert_eq!(drained.unpriced_turns, 1); |
| 5035 | assert!(drained.unpriced_reasons.contains("missing_class_price")); |
| 5036 | assert!(drained.unpriced_classes.contains("cache_write")); |
| 5037 | assert!( |
| 5038 | drained |
| 5039 | .route_receipts |
| 5040 | .iter() |
| 5041 | .any(|receipt| receipt.contains("cache_write=yes")), |
| 5042 | "{drained:?}" |
| 5043 | ); |
| 5044 | } |
| 5045 | |
| 5046 | #[test] |
| 5047 | fn missing_usage_reason_preserves_legacy_bytes_and_metadata_identity() { |
| 5048 | let route = deepseek_envelope(); |
| 5049 | let old = serde_json::json!({"source_id":"opaque-response", "route":route}); |
| 5050 | let legacy: RuntimeUsageDropRecord = serde_json::from_value(old.clone()).unwrap(); |
| 5051 | assert_eq!( |
| 5052 | legacy.reason, |
| 5053 | RuntimeUsageMissingReason::SuccessWithoutUsage |
| 5054 | ); |
| 5055 | assert_eq!(serde_json::to_value(&legacy).unwrap(), old); |
| 5056 | let unknown = RuntimeUsageDropRecord { |
| 5057 | reason: RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5058 | ..legacy |
| 5059 | }; |
| 5060 | let mut metadata = serde_json::json!({}); |
| 5061 | attach_child_usage_batch_metadata( |
| 5062 | &mut metadata, |
| 5063 | &RuntimeUsageBatch { |
| 5064 | drop_records: vec![unknown], |
| 5065 | dropped_records: 1, |
| 5066 | ..Default::default() |
| 5067 | }, |
| 5068 | ); |
| 5069 | let recovered = child_usage_records_from_metadata(&metadata).unwrap(); |
| 5070 | assert_eq!( |
| 5071 | recovered.drop_records[0].reason, |
| 5072 | RuntimeUsageMissingReason::RequestOutcomeUnknown |
| 5073 | ); |
| 5074 | assert_eq!( |
| 5075 | recovered.drop_records[0].source_id, |
| 5076 | usage_source_fingerprint("opaque-response") |
| 5077 | ); |
| 5078 | assert_eq!(recovered.dropped_records, 1); |
| 5079 | assert!( |
| 5080 | !serde_json::to_string(&metadata) |
| 5081 | .unwrap() |
| 5082 | .contains("opaque-response") |
| 5083 | ); |
| 5084 | } |
| 5085 | |
| 5086 | #[test] |
| 5087 | fn exact_missing_usage_promotes_once_before_drain_and_rejects_changed_route() { |
| 5088 | let _scope = test_scope(); |
| 5089 | let route = deepseek_envelope(); |
| 5090 | report_missing_runtime_usage( |
| 5091 | scope_token(), |
| 5092 | None, |
| 5093 | "attempt", |
| 5094 | &route, |
| 5095 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5096 | ); |
| 5097 | let mut changed = route.clone(); |
| 5098 | changed.model = "another-model".into(); |
| 5099 | report_effective_route_for_runtime( |
| 5100 | scope_token(), |
| 5101 | None, |
| 5102 | "attempt", |
| 5103 | &changed, |
| 5104 | &small_usage(), |
| 5105 | ); |
| 5106 | report_effective_route_for_runtime(scope_token(), None, "attempt", &route, &small_usage()); |
| 5107 | report_effective_route_for_runtime(scope_token(), None, "attempt", &route, &small_usage()); |
| 5108 | report_missing_runtime_usage( |
| 5109 | scope_token(), |
| 5110 | None, |
| 5111 | "attempt", |
| 5112 | &route, |
| 5113 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5114 | ); |
| 5115 | let pool = drain(); |
| 5116 | assert_eq!(pool.priced_turns, 1); |
| 5117 | assert_eq!(pool.unpriced_turns, 0); |
| 5118 | assert_eq!(pool.cny_unpriced_turns, 0); |
| 5119 | assert!(pool.missing_usage_sources.is_empty()); |
| 5120 | assert!(pool.estimate.usd > 0.0); |
| 5121 | assert_eq!(pool.usage_source_fingerprints.len(), 1); |
| 5122 | } |
| 5123 | |
| 5124 | #[test] |
| 5125 | fn drained_missing_slot_restores_and_late_receipt_resolves_only_its_origin() { |
| 5126 | let _scope = test_scope(); |
| 5127 | let route = deepseek_envelope(); |
| 5128 | report_missing_runtime_usage( |
| 5129 | scope_token(), |
| 5130 | None, |
| 5131 | "restored-attempt", |
| 5132 | &route, |
| 5133 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5134 | ); |
| 5135 | let missing = drain(); |
| 5136 | let mut slots = missing.missing_usage_sources.clone(); |
| 5137 | let mut overflow = false; |
| 5138 | let mut unpriced = missing.unpriced_turns; |
| 5139 | let mut cny_unpriced = missing.cny_unpriced_turns; |
| 5140 | assert!(close_current_scope().is_empty()); |
| 5141 | restore_usage_source_ledger(missing.usage_source_fingerprints.clone(), &slots, overflow); |
| 5142 | report_effective_route_for_runtime( |
| 5143 | scope_token(), |
| 5144 | None, |
| 5145 | "restored-attempt", |
| 5146 | &route, |
| 5147 | &small_usage(), |
| 5148 | ); |
| 5149 | let known = drain(); |
| 5150 | let projected = project_missing_usage_ledger( |
| 5151 | &mut slots, |
| 5152 | &mut overflow, |
| 5153 | &mut unpriced, |
| 5154 | &mut cny_unpriced, |
| 5155 | &known, |
| 5156 | ); |
| 5157 | assert!(slots.is_empty()); |
| 5158 | assert_eq!((unpriced, cny_unpriced), (0, 0)); |
| 5159 | assert_eq!(projected.priced_turns, 1); |
| 5160 | report_effective_route_for_runtime( |
| 5161 | scope_token(), |
| 5162 | None, |
| 5163 | "restored-attempt", |
| 5164 | &route, |
| 5165 | &small_usage(), |
| 5166 | ); |
| 5167 | assert!(drain().is_empty()); |
| 5168 | // A legacy snapshot retained only consumed identities. Its unattributed |
| 5169 | // coverage remains conservative rather than being erased by a replay. |
| 5170 | assert!(close_current_scope().is_empty()); |
| 5171 | restore_usage_source_fingerprints(missing.usage_source_fingerprints); |
| 5172 | report_effective_route_for_runtime( |
| 5173 | scope_token(), |
| 5174 | None, |
| 5175 | "restored-attempt", |
| 5176 | &route, |
| 5177 | &small_usage(), |
| 5178 | ); |
| 5179 | assert!(drain().is_empty()); |
| 5180 | } |
| 5181 | |
| 5182 | #[test] |
| 5183 | fn missing_usage_overflow_is_bounded_and_cannot_be_erased_by_late_receipts() { |
| 5184 | let _scope = test_scope(); |
| 5185 | let route = deepseek_envelope(); |
| 5186 | for index in 0..65 { |
| 5187 | for _ in 0..2 { |
| 5188 | report_missing_runtime_usage( |
| 5189 | scope_token(), |
| 5190 | None, |
| 5191 | &format!("attempt-{index}"), |
| 5192 | &route, |
| 5193 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5194 | ); |
| 5195 | } |
| 5196 | } |
| 5197 | let missing = drain(); |
| 5198 | assert_eq!( |
| 5199 | missing.missing_usage_sources.len(), |
| 5200 | MAX_MISSING_USAGE_SOURCES |
| 5201 | ); |
| 5202 | assert_eq!(missing.unpriced_turns, 65); |
| 5203 | assert!(missing.missing_usage_overflowed); |
| 5204 | let mut slots = missing.missing_usage_sources.clone(); |
| 5205 | let mut overflow = missing.missing_usage_overflowed; |
| 5206 | let mut unpriced = missing.unpriced_turns; |
| 5207 | let mut cny_unpriced = missing.cny_unpriced_turns; |
| 5208 | for index in 0..65 { |
| 5209 | for _ in 0..2 { |
| 5210 | report_effective_route_for_runtime( |
| 5211 | scope_token(), |
| 5212 | None, |
| 5213 | &format!("attempt-{index}"), |
| 5214 | &route, |
| 5215 | &small_usage(), |
| 5216 | ); |
| 5217 | } |
| 5218 | } |
| 5219 | let known = drain(); |
| 5220 | let projected = project_missing_usage_ledger( |
| 5221 | &mut slots, |
| 5222 | &mut overflow, |
| 5223 | &mut unpriced, |
| 5224 | &mut cny_unpriced, |
| 5225 | &known, |
| 5226 | ); |
| 5227 | assert_eq!(projected.priced_turns, 65); |
| 5228 | assert_eq!((unpriced, cny_unpriced), (1, 1)); |
| 5229 | assert!(slots.is_empty()); |
| 5230 | assert!(overflow); |
| 5231 | assert!(projected.estimate.usd > 0.0); |
| 5232 | } |
| 5233 | |
| 5234 | #[test] |
| 5235 | fn runtime_unknown_batch_keeps_reason_and_promotes_without_new_response_identity() { |
| 5236 | let _scope = test_scope(); |
| 5237 | let route = deepseek_envelope(); |
| 5238 | let owner = "unknown-batch-owner"; |
| 5239 | report_missing_runtime_usage( |
| 5240 | scope_token(), |
| 5241 | Some(owner), |
| 5242 | "batch-attempt", |
| 5243 | &route, |
| 5244 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5245 | ); |
| 5246 | let batch = take_runtime_usage(owner); |
| 5247 | assert_eq!( |
| 5248 | batch.drop_records[0].reason, |
| 5249 | RuntimeUsageMissingReason::RequestOutcomeUnknown |
| 5250 | ); |
| 5251 | let mut metadata = serde_json::json!({}); |
| 5252 | attach_child_usage_batch_metadata(&mut metadata, &batch); |
| 5253 | let batch = child_usage_records_from_metadata(&metadata).unwrap(); |
| 5254 | report_runtime_usage_batch(scope_token(), None, &batch); |
| 5255 | report_runtime_usage_batch(scope_token(), None, &batch); |
| 5256 | let missing = drain(); |
| 5257 | assert_eq!(missing.unpriced_turns, 1); |
| 5258 | assert!(missing.unpriced_reasons.contains("request_outcome_unknown")); |
| 5259 | report_effective_route_for_runtime( |
| 5260 | scope_token(), |
| 5261 | None, |
| 5262 | "batch-attempt", |
| 5263 | &route, |
| 5264 | &small_usage(), |
| 5265 | ); |
| 5266 | let known = drain(); |
| 5267 | assert_eq!(known.priced_turns, 1); |
| 5268 | assert_eq!( |
| 5269 | known.resolved_missing_usage_sources, |
| 5270 | BTreeSet::from([usage_source_fingerprint("batch-attempt")]) |
| 5271 | ); |
| 5272 | } |
| 5273 | |
| 5274 | #[test] |
| 5275 | fn late_unknown_receipt_promotes_after_restart_without_charging_replacement_session() { |
| 5276 | let _scope = test_scope(); |
| 5277 | let tmp = tempfile::tempdir().unwrap(); |
| 5278 | let sessions = tmp.path().join("sessions"); |
| 5279 | let manager = crate::session_manager::SessionManager::new(sessions.clone()).unwrap(); |
| 5280 | for id in ["unknown-origin", "replacement"] { |
| 5281 | let session = crate::session_manager::create_saved_session_with_id_and_mode( |
| 5282 | id.into(), |
| 5283 | &[], |
| 5284 | "deepseek-v4-flash", |
| 5285 | tmp.path(), |
| 5286 | 0, |
| 5287 | None, |
| 5288 | Some("agent"), |
| 5289 | ); |
| 5290 | manager.save_session(&session).unwrap(); |
| 5291 | } |
| 5292 | let origin = scope_token(); |
| 5293 | assert!(close_current_scope().is_empty()); |
| 5294 | let route = deepseek_envelope(); |
| 5295 | assert!(report_missing_usage_for_interactive_origin_with_manager( |
| 5296 | origin, |
| 5297 | "unknown-origin", |
| 5298 | "origin-turn", |
| 5299 | "late-attempt", |
| 5300 | &route, |
| 5301 | RuntimeUsageMissingReason::RequestOutcomeUnknown, |
| 5302 | &manager |
| 5303 | )); |
| 5304 | let missing = manager.load_session_snapshot("unknown-origin").unwrap(); |
| 5305 | assert_eq!(missing.metadata.cost.unpriced_turns, 1); |
| 5306 | assert_eq!(missing.metadata.cost.missing_usage_sources.len(), 1); |
| 5307 | assert!( |
| 5308 | missing |
| 5309 | .metadata |
| 5310 | .cost |
| 5311 | .unpriced_reasons |
| 5312 | .contains("request_outcome_unknown") |
| 5313 | ); |
| 5314 | manager.save_session(&missing).unwrap(); |
| 5315 | drop(manager); |
| 5316 | let manager = crate::session_manager::SessionManager::new(sessions).unwrap(); |
| 5317 | for _ in 0..2 { |
| 5318 | assert!(report_effective_route_for_interactive_origin_with_manager( |
| 5319 | origin, |
| 5320 | "unknown-origin", |
| 5321 | "origin-turn", |
| 5322 | "late-attempt", |
| 5323 | &route, |
| 5324 | &small_usage(), |
| 5325 | &manager |
| 5326 | )); |
| 5327 | let known = manager.load_session_snapshot("unknown-origin").unwrap(); |
| 5328 | assert_eq!(known.metadata.cost.priced_turns, 1); |
| 5329 | assert_eq!(known.metadata.cost.unpriced_turns, 0); |
| 5330 | assert!(known.metadata.cost.missing_usage_sources.is_empty()); |
| 5331 | assert_eq!( |
| 5332 | known.metadata.total_tokens, |
| 5333 | u64::from(small_usage().input_tokens) + u64::from(small_usage().output_tokens) |
| 5334 | ); |
| 5335 | manager.save_session(&known).unwrap(); |
| 5336 | } |
| 5337 | assert!(!report_effective_route_for_interactive_origin_with_manager( |
| 5338 | origin, |
| 5339 | "unknown-origin", |
| 5340 | "different-turn", |
| 5341 | "late-attempt", |
| 5342 | &route, |
| 5343 | &small_usage(), |
| 5344 | &manager |
| 5345 | )); |
| 5346 | let replacement = manager.load_session_snapshot("replacement").unwrap(); |
| 5347 | assert_eq!(replacement.metadata.total_tokens, 0); |
| 5348 | assert_eq!( |
| 5349 | ( |
| 5350 | replacement.metadata.cost.priced_turns, |
| 5351 | replacement.metadata.cost.unpriced_turns |
| 5352 | ), |
| 5353 | (0, 0) |
| 5354 | ); |
| 5355 | assert!(drain().is_empty()); |
| 5356 | } |
| 5357 | } |
| 5358 |