| 1 | //! Provider model offerings (#3084). |
| 2 | //! |
| 3 | //! A [`ProviderModelOffering`] binds a provider to a canonical model, the |
| 4 | //! provider-owned wire id that serves it, and the endpoint key. This is the |
| 5 | //! seam that proves the #2608 invariant: the SAME canonical model can be served |
| 6 | //! by multiple providers under DIFFERENT wire ids (some aggregator-prefixed), |
| 7 | //! and a prefix never implies provider ownership. |
| 8 | //! |
| 9 | //! Catalog-derived offerings from [`crate::catalog::bundled_catalog_offerings`] |
| 10 | //! remain the general bundled source of truth. [`bundled_offerings`] contains |
| 11 | //! only transport facts that Models.dev cannot express, such as a single |
| 12 | //! provider routing different models over different wire protocols. |
| 13 | |
| 14 | use serde::{Deserialize, Serialize}; |
| 15 | |
| 16 | use super::candidate::PricingSku; |
| 17 | use super::capabilities::{CapabilityState, RouteCapabilities}; |
| 18 | use super::ids::{ModelId, ProviderId, WireModelId}; |
| 19 | |
| 20 | /// Token limits for one resolved route/offering. |
| 21 | /// |
| 22 | /// These are optional because hosted catalogs, local runtimes, and custom |
| 23 | /// endpoints can legitimately omit some or all limit facts. Callers should |
| 24 | /// treat `None` as unknown, not zero. |
| 25 | #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)] |
| 26 | pub struct RouteLimits { |
| 27 | /// Total context window (input + output), in tokens. |
| 28 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 29 | pub context_tokens: Option<u64>, |
| 30 | /// Input-token limit, when the provider reports it separately. |
| 31 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 32 | pub input_tokens: Option<u64>, |
| 33 | /// Output-token cap for the route/offering, when known. |
| 34 | #[serde(default, skip_serializing_if = "Option::is_none")] |
| 35 | pub output_tokens: Option<u64>, |
| 36 | } |
| 37 | |
| 38 | impl RouteLimits { |
| 39 | /// Whether at least one limit fact is known. |
| 40 | #[must_use] |
| 41 | pub const fn has_known_limit(self) -> bool { |
| 42 | self.context_tokens.is_some() || self.input_tokens.is_some() || self.output_tokens.is_some() |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | /// One provider's way of serving a (possibly canonical) model. |
| 47 | /// |
| 48 | /// `Eq` is intentionally NOT derived: [`PricingSku::Token`] carries `f64` rates, |
| 49 | /// so the offering is only `PartialEq`. No caller keys a set/map on offerings. |
| 50 | #[derive(Debug, Clone, PartialEq)] |
| 51 | pub struct ProviderModelOffering { |
| 52 | /// Provider serving this offering. |
| 53 | pub provider: ProviderId, |
| 54 | /// Canonical model identity, if this offering maps to one. |
| 55 | pub canonical_model: Option<ModelId>, |
| 56 | /// Provider-owned wire id sent on the request (verbatim). |
| 57 | pub wire_model_id: WireModelId, |
| 58 | /// Endpoint key the offering is served on. |
| 59 | pub endpoint_key: String, |
| 60 | /// Whether this is the provider's default offering. |
| 61 | pub default_for_provider: bool, |
| 62 | /// Provider/offering-scoped token limits, when known. |
| 63 | pub limits: RouteLimits, |
| 64 | /// Provider/model-scoped capability facts. Unknown is preserved rather |
| 65 | /// than inferred from the wire protocol. |
| 66 | pub capabilities: RouteCapabilities, |
| 67 | /// Coarse route-facing pricing meter for this offering (#3085). |
| 68 | /// |
| 69 | /// Projected from the offering's sourced cost at the layer that owns it |
| 70 | /// (`CatalogOffering::to_offering` → [`crate::pricing::route_pricing_sku`]). |
| 71 | /// The resolver carries this verbatim onto the candidate; it is |
| 72 | /// [`PricingSku::UnknownOrStale`] whenever no price was sourced — never a |
| 73 | /// fabricated zero (the #2608 / #3085 honesty rule). |
| 74 | pub pricing: PricingSku, |
| 75 | } |
| 76 | |
| 77 | // Transport snapshot verified against https://opencode.ai/docs/zen on |
| 78 | // 2026-07-17. Gemini rows are intentionally absent because they use Google's |
| 79 | // model-specific wire protocol, which CodeWhale does not currently implement. |
| 80 | /// Token Plan text models (Text Generation / Reasoning, coding scope). |
| 81 | /// |
| 82 | /// Available on both Token Plan Personal and Team. The same model set is also |
| 83 | /// available on the Coding Plan; rows are duplicated per provider id below. |
| 84 | /// Pay-as-you-go workspace-id templating is deferred to a follow-up. |
| 85 | const MODELSTUDIO_TEXT_MODELS: &[&str] = &[ |
| 86 | "qwen3.8-max", |
| 87 | "qwen3.8-max-preview", |
| 88 | "qwen3.7-plus", |
| 89 | "qwen3.7-max", |
| 90 | "qwen3.6-flash", |
| 91 | // DeepSeek models served under Model Studio are scoped to this provider; |
| 92 | // they do not collide with first-party DeepSeek routes. |
| 93 | "deepseek-v4-pro", |
| 94 | "deepseek-v4-flash-0731", |
| 95 | // GLM models served under Model Studio are scoped to this provider; |
| 96 | // they do not collide with first-party Zhipu / Z.ai routes. |
| 97 | // |
| 98 | // glm-5.3 is deliberately absent (2026-08-03): this list is a curated |
| 99 | // snapshot of what Model Studio's upstream roster actually serves, and |
| 100 | // Model Studio publishes no glm-5.3 entry. The direct Z.ai / OpenRouter |
| 101 | // glm-5.3 rows inherit their metadata from glm-5.2, but metadata |
| 102 | // inheritance is not evidence that a third-party gateway carries the |
| 103 | // model. Add it here only against a Model Studio console/roster listing. |
| 104 | "glm-5.2", |
| 105 | ]; |
| 106 | |
| 107 | pub(crate) const OPENCODE_ZEN_RESPONSES_MODELS: &[&str] = &[ |
| 108 | "gpt-5.6-sol", |
| 109 | "gpt-5.6-terra", |
| 110 | "gpt-5.6-luna", |
| 111 | "gpt-5.5", |
| 112 | "gpt-5.5-pro", |
| 113 | "gpt-5.4", |
| 114 | "gpt-5.4-pro", |
| 115 | "gpt-5.4-mini", |
| 116 | "gpt-5.4-nano", |
| 117 | "gpt-5.3-codex", |
| 118 | "gpt-5.3-codex-spark", |
| 119 | "gpt-5.2", |
| 120 | "gpt-5.2-codex", |
| 121 | "gpt-5.1", |
| 122 | "gpt-5.1-codex", |
| 123 | "gpt-5.1-codex-max", |
| 124 | "gpt-5.1-codex-mini", |
| 125 | "gpt-5", |
| 126 | "gpt-5-codex", |
| 127 | "gpt-5-nano", |
| 128 | // Muse Spark via OpenCode Zen gateway — Responses-only (reported |
| 129 | // 2026-08-29: muse-spark-1.2-contributor-free rejects Chat Completions). |
| 130 | "muse-spark-1.2", |
| 131 | "muse-spark-1.2-contributor", |
| 132 | "muse-spark-1.2-contributor-free", |
| 133 | ]; |
| 134 | |
| 135 | pub(crate) const OPENCODE_ZEN_MESSAGES_MODELS: &[&str] = &[ |
| 136 | "claude-fable-5", |
| 137 | "claude-opus-4-8", |
| 138 | "claude-opus-4-7", |
| 139 | "claude-opus-4-6", |
| 140 | "claude-opus-4-5", |
| 141 | "claude-sonnet-5", |
| 142 | "claude-sonnet-4-6", |
| 143 | "claude-sonnet-4-5", |
| 144 | "claude-haiku-4-5", |
| 145 | "qwen3.7-max", |
| 146 | "qwen3.7-plus", |
| 147 | "qwen3.6-plus", |
| 148 | "qwen3.5-plus", |
| 149 | ]; |
| 150 | |
| 151 | pub(crate) const OPENCODE_ZEN_CHAT_MODELS: &[&str] = &[ |
| 152 | "deepseek-v4-pro", |
| 153 | "deepseek-v4-flash", |
| 154 | "minimax-m3", |
| 155 | "minimax-m2.7", |
| 156 | "minimax-m2.5", |
| 157 | // glm-5.3 is deliberately absent (2026-08-03): this snapshot tracks the |
| 158 | // official OpenCode Zen endpoint table, which lists no glm-5.3 row. Zen |
| 159 | // fails closed on unknown models by design; registering a route Zen does |
| 160 | // not serve would convert that into a guaranteed upstream 404. |
| 161 | "glm-5.2", |
| 162 | "glm-5.1", |
| 163 | "glm-5", |
| 164 | "kimi-k2.5", |
| 165 | "kimi-k2.6", |
| 166 | "kimi-k2.7-code", |
| 167 | "grok-4.5", |
| 168 | "grok-build-0.1", |
| 169 | "big-pickle", |
| 170 | "mimo-v2.5-free", |
| 171 | "north-mini-code-free", |
| 172 | "nemotron-3-ultra-free", |
| 173 | "deepseek-v4-flash-free", |
| 174 | ]; |
| 175 | |
| 176 | /// Logical default plus every documented Zen wire id, for picker fallbacks |
| 177 | /// when Models.dev is stale or failed. `gpt-5.6` is the user-facing default; |
| 178 | /// `gpt-5.6-sol` is the proven Responses wire id. |
| 179 | #[must_use] |
| 180 | pub fn opencode_zen_picker_models() -> Vec<&'static str> { |
| 181 | let mut models = vec![crate::DEFAULT_OPENCODE_ZEN_MODEL]; |
| 182 | for model in OPENCODE_ZEN_RESPONSES_MODELS |
| 183 | .iter() |
| 184 | .chain(OPENCODE_ZEN_MESSAGES_MODELS) |
| 185 | .chain(OPENCODE_ZEN_CHAT_MODELS) |
| 186 | { |
| 187 | if !models |
| 188 | .iter() |
| 189 | .any(|existing| existing.eq_ignore_ascii_case(model)) |
| 190 | { |
| 191 | models.push(*model); |
| 192 | } |
| 193 | } |
| 194 | models |
| 195 | } |
| 196 | |
| 197 | /// Codewhale API bootstrap rows used only when the account's live |
| 198 | /// `GET {base}/models` cannot be fetched. |
| 199 | /// |
| 200 | /// The account catalog is authoritative: it lists exactly the providers the |
| 201 | /// customer connected, and each row states its own protocol. These three rows |
| 202 | /// exist so a route can still be selected offline; every consumer that shows |
| 203 | /// models must say the list is a fallback, not the account's catalog. |
| 204 | pub const CODEWHALE_FALLBACK_MODELS: &[&str] = &[ |
| 205 | "deepseek/deepseek-v4-pro", |
| 206 | "anthropic/claude-sonnet-5", |
| 207 | "openai/gpt-5.6", |
| 208 | ]; |
| 209 | |
| 210 | /// Endpoint key for one Codewhale API model id. |
| 211 | /// |
| 212 | /// The live catalog states the protocol per model in `codewhale.protocol`; |
| 213 | /// this is the offline inference used for the bootstrap rows and for a model |
| 214 | /// id the local catalog has never seen. Only the `anthropic/` namespace routes |
| 215 | /// to `{base}/messages`; everything else is OpenAI Chat Completions at |
| 216 | /// `{base}/chat/completions`. The id alone carries no signal for the |
| 217 | /// Responses surface — a `responses` row only ever comes from the catalog's |
| 218 | /// stated `codewhale.protocol`, never from a model name. |
| 219 | #[must_use] |
| 220 | pub fn codewhale_endpoint_key_for_model(model: &str) -> &'static str { |
| 221 | if model.trim().to_ascii_lowercase().starts_with("anthropic/") { |
| 222 | "messages" |
| 223 | } else { |
| 224 | "chat" |
| 225 | } |
| 226 | } |
| 227 | |
| 228 | /// Return curated provider/model transport facts as owned offering rows. |
| 229 | /// |
| 230 | /// OpenCode Zen's official catalog serves models over three protocol families. |
| 231 | /// These rows intentionally carry no inferred limits, pricing, or canonical |
| 232 | /// identity: their sole claim is the documented wire model and endpoint key. |
| 233 | #[must_use] |
| 234 | pub fn bundled_offerings() -> Vec<ProviderModelOffering> { |
| 235 | // DeepSeek's 2026-07-31 production Flash update added a native Responses |
| 236 | // endpoint without changing the model id, and the 2026-08 unversioned |
| 237 | // rename to `deepseek-flash` kept that wire: DeepSeek's own Codex |
| 238 | // integration documents the Responses API as the path for `deepseek-flash` |
| 239 | // (legacy `deepseek-v4-flash` ids are served by the same model). The |
| 240 | // shipped default therefore rides Responses; Pro remains Chat Completions |
| 241 | // until its announced Responses rollout. These exact-route transport facts |
| 242 | // cannot be represented by the Models.dev-shaped fallback asset. |
| 243 | let deepseek = ProviderId::from("deepseek"); |
| 244 | let documented_capabilities = RouteCapabilities { |
| 245 | image_input: CapabilityState::Unsupported, |
| 246 | reasoning: CapabilityState::Supported, |
| 247 | native_tool_calls: CapabilityState::Supported, |
| 248 | structured_output: CapabilityState::Supported, |
| 249 | parallel_tool_calls: CapabilityState::Supported, |
| 250 | streaming: CapabilityState::Supported, |
| 251 | prompt_caching: CapabilityState::Supported, |
| 252 | // Search execution uses a separate bounded Responses request rather |
| 253 | // than replaying `web_search_call` items in the main chat. |
| 254 | server_side_web_search: CapabilityState::Supported, |
| 255 | ..RouteCapabilities::default() |
| 256 | }; |
| 257 | let documented_limits = RouteLimits { |
| 258 | context_tokens: Some(1_000_000), |
| 259 | input_tokens: None, |
| 260 | output_tokens: Some(384_000), |
| 261 | }; |
| 262 | let mut offerings = vec![ |
| 263 | ProviderModelOffering { |
| 264 | provider: deepseek.clone(), |
| 265 | canonical_model: Some(ModelId::from("deepseek-flash")), |
| 266 | wire_model_id: WireModelId::from("deepseek-flash"), |
| 267 | endpoint_key: "responses".to_string(), |
| 268 | default_for_provider: true, |
| 269 | limits: documented_limits, |
| 270 | capabilities: documented_capabilities, |
| 271 | pricing: PricingSku::UnknownOrStale, |
| 272 | }, |
| 273 | ProviderModelOffering { |
| 274 | provider: deepseek.clone(), |
| 275 | canonical_model: Some(ModelId::from("deepseek-v4-pro")), |
| 276 | wire_model_id: WireModelId::from("deepseek-v4-pro"), |
| 277 | endpoint_key: "chat".to_string(), |
| 278 | default_for_provider: false, |
| 279 | limits: documented_limits, |
| 280 | capabilities: documented_capabilities, |
| 281 | pricing: PricingSku::UnknownOrStale, |
| 282 | }, |
| 283 | ProviderModelOffering { |
| 284 | provider: deepseek.clone(), |
| 285 | canonical_model: Some(ModelId::from("deepseek-v4-flash")), |
| 286 | wire_model_id: WireModelId::from("deepseek-v4-flash"), |
| 287 | endpoint_key: "responses".to_string(), |
| 288 | default_for_provider: false, |
| 289 | limits: documented_limits, |
| 290 | capabilities: documented_capabilities, |
| 291 | pricing: PricingSku::UnknownOrStale, |
| 292 | }, |
| 293 | // Vision-experimental sibling of v4-flash, verified live on |
| 294 | // api.deepseek.com /models (2026-08-21). Image input is the one |
| 295 | // documented difference; limits inherit the v4-flash row until |
| 296 | // DeepSeek publishes distinct numbers. |
| 297 | ProviderModelOffering { |
| 298 | provider: deepseek, |
| 299 | canonical_model: Some(ModelId::from("deepseek-v4-flash-vision-exp")), |
| 300 | wire_model_id: WireModelId::from("deepseek-v4-flash-vision-exp"), |
| 301 | endpoint_key: "chat".to_string(), |
| 302 | default_for_provider: false, |
| 303 | limits: documented_limits, |
| 304 | capabilities: RouteCapabilities { |
| 305 | image_input: CapabilityState::Supported, |
| 306 | ..documented_capabilities |
| 307 | }, |
| 308 | pricing: PricingSku::UnknownOrStale, |
| 309 | }, |
| 310 | ]; |
| 311 | |
| 312 | offerings.extend( |
| 313 | crate::opencode_go::MODEL_GROUPS |
| 314 | .iter() |
| 315 | .flat_map(|(endpoint_key, models)| { |
| 316 | models.iter().map(move |model| ProviderModelOffering { |
| 317 | provider: ProviderId::from("opencode-go"), |
| 318 | canonical_model: None, |
| 319 | wire_model_id: WireModelId::from(*model), |
| 320 | endpoint_key: (*endpoint_key).to_string(), |
| 321 | default_for_provider: *model == crate::DEFAULT_OPENCODE_GO_MODEL, |
| 322 | limits: RouteLimits::default(), |
| 323 | capabilities: RouteCapabilities::default(), |
| 324 | pricing: PricingSku::UnknownOrStale, |
| 325 | }) |
| 326 | }), |
| 327 | ); |
| 328 | |
| 329 | let provider = ProviderId::from("opencode-zen"); |
| 330 | let groups = [ |
| 331 | ("responses", OPENCODE_ZEN_RESPONSES_MODELS), |
| 332 | ("messages", OPENCODE_ZEN_MESSAGES_MODELS), |
| 333 | ("chat", OPENCODE_ZEN_CHAT_MODELS), |
| 334 | ]; |
| 335 | |
| 336 | offerings.extend(groups.into_iter().flat_map(|(endpoint_key, models)| { |
| 337 | let provider = provider.clone(); |
| 338 | models.iter().map(move |model| ProviderModelOffering { |
| 339 | provider: provider.clone(), |
| 340 | // The bundled catalog exposes `gpt-5.6` as the user-facing |
| 341 | // logical choice and records `gpt-5.6-sol` as its proven Zen |
| 342 | // wire id. Keep the generic choice honest by resolving it to the |
| 343 | // documented concrete Responses model rather than sending an |
| 344 | // unproven generic wire id to Zen. |
| 345 | canonical_model: (*model == "gpt-5.6-sol").then(|| ModelId::from("gpt-5.6")), |
| 346 | wire_model_id: WireModelId::from(*model), |
| 347 | endpoint_key: endpoint_key.to_string(), |
| 348 | default_for_provider: *model == "gpt-5.6-sol", |
| 349 | limits: RouteLimits::default(), |
| 350 | capabilities: RouteCapabilities::default(), |
| 351 | pricing: PricingSku::UnknownOrStale, |
| 352 | }) |
| 353 | })); |
| 354 | |
| 355 | // Codewhale API bootstrap rows. The account's authenticated |
| 356 | // `GET {base}/models` is the catalog authority and replaces these as soon |
| 357 | // as it is reachable; they exist so the route resolves offline. |
| 358 | let codewhale = ProviderId::from("codewhale"); |
| 359 | offerings.extend( |
| 360 | CODEWHALE_FALLBACK_MODELS |
| 361 | .iter() |
| 362 | .map(|model| ProviderModelOffering { |
| 363 | provider: codewhale.clone(), |
| 364 | canonical_model: None, |
| 365 | wire_model_id: WireModelId::from(*model), |
| 366 | endpoint_key: codewhale_endpoint_key_for_model(model).to_string(), |
| 367 | default_for_provider: *model == crate::DEFAULT_CODEWHALE_MODEL, |
| 368 | limits: RouteLimits::default(), |
| 369 | capabilities: RouteCapabilities::default(), |
| 370 | pricing: PricingSku::UnknownOrStale, |
| 371 | }), |
| 372 | ); |
| 373 | |
| 374 | // Alibaba Cloud Model Studio — one vendor identity in the hand seam |
| 375 | // (`modelstudio-token-plan`). Plan (token vs coding) and wire dialect |
| 376 | // (OpenAI Chat Completions vs Anthropic Messages) are config (`mode` / |
| 377 | // `wire`), not separate ProviderKinds — same product shape as Z.ai / |
| 378 | // Xiaomi for plans and a power-user toggle for dialect. Legacy provider |
| 379 | // ids still get catalog rows so old configs resolve, but the picker |
| 380 | // catalog surface only lists the primary id. |
| 381 | // |
| 382 | // Limits: owner's Token Plan console + curated models_dev rows |
| 383 | // (2026-08-03): qwen3.8-max is ~1M context / 128K output, NOT 128K |
| 384 | // total. Empty RouteLimits here used to win identity collisions over |
| 385 | // the asset catalog and fall through to the 128K legacy default. |
| 386 | fn ms_capabilities(model: &str) -> RouteCapabilities { |
| 387 | let image_input = match model { |
| 388 | "qwen3.8-max" | "qwen3.8-max-preview" | "qwen3.7-plus" | "qwen3.6-flash" => { |
| 389 | CapabilityState::Supported |
| 390 | } |
| 391 | _ => CapabilityState::Unsupported, |
| 392 | }; |
| 393 | RouteCapabilities { |
| 394 | reasoning: CapabilityState::Supported, |
| 395 | native_tool_calls: CapabilityState::Supported, |
| 396 | structured_output: CapabilityState::Supported, |
| 397 | streaming: CapabilityState::Supported, |
| 398 | image_input, |
| 399 | server_side_web_search: super::documented_server_side_web_search( |
| 400 | "modelstudio-token-plan", |
| 401 | model, |
| 402 | ), |
| 403 | ..RouteCapabilities::default() |
| 404 | } |
| 405 | } |
| 406 | fn ms_limits(model: &str) -> RouteLimits { |
| 407 | // Context/output from models_dev.bundled.json Model Studio rows and |
| 408 | // the owner console (verified 2026-08-03). Keep output separate from |
| 409 | // context so a 128K generation ceiling is never mistaken for the |
| 410 | // window. |
| 411 | let (context_tokens, output_tokens) = match model { |
| 412 | "qwen3.8-max" | "qwen3.8-max-preview" => (1_000_000, 131_072), |
| 413 | "qwen3.7-plus" | "qwen3.7-max" => (1_000_000, 65_536), |
| 414 | "qwen3.6-flash" => (1_000_000, 65_536), |
| 415 | "deepseek-v4-pro" | "deepseek-v4-flash-0731" => (1_000_000, 384_000), |
| 416 | "glm-5.2" => (1_000_000, 131_072), |
| 417 | _ => (1_000_000, 131_072), |
| 418 | }; |
| 419 | RouteLimits { |
| 420 | context_tokens: Some(context_tokens), |
| 421 | input_tokens: None, |
| 422 | output_tokens: Some(output_tokens), |
| 423 | } |
| 424 | } |
| 425 | // Primary vendor id only in the hand seam. Coding-plan / anthropic |
| 426 | // dialect endpoint selection is owned by config resolution (mode/wire), |
| 427 | // which rewrites base_url + request dialect without inventing kinds. |
| 428 | let plan = ProviderId::from("modelstudio-token-plan"); |
| 429 | offerings.extend( |
| 430 | MODELSTUDIO_TEXT_MODELS |
| 431 | .iter() |
| 432 | .enumerate() |
| 433 | .map(|(i, model)| ProviderModelOffering { |
| 434 | provider: plan.clone(), |
| 435 | canonical_model: None, |
| 436 | wire_model_id: WireModelId::from(*model), |
| 437 | endpoint_key: "chat".to_string(), |
| 438 | default_for_provider: i == 0, |
| 439 | limits: ms_limits(model), |
| 440 | capabilities: ms_capabilities(model), |
| 441 | pricing: PricingSku::UnknownOrStale, |
| 442 | }), |
| 443 | ); |
| 444 | |
| 445 | offerings |
| 446 | } |
| 447 |