| 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 | ]; |
| 129 | |
| 130 | pub(crate) const OPENCODE_ZEN_MESSAGES_MODELS: &[&str] = &[ |
| 131 | "claude-fable-5", |
| 132 | "claude-opus-4-8", |
| 133 | "claude-opus-4-7", |
| 134 | "claude-opus-4-6", |
| 135 | "claude-opus-4-5", |
| 136 | "claude-sonnet-5", |
| 137 | "claude-sonnet-4-6", |
| 138 | "claude-sonnet-4-5", |
| 139 | "claude-haiku-4-5", |
| 140 | "qwen3.7-max", |
| 141 | "qwen3.7-plus", |
| 142 | "qwen3.6-plus", |
| 143 | "qwen3.5-plus", |
| 144 | ]; |
| 145 | |
| 146 | pub(crate) const OPENCODE_ZEN_CHAT_MODELS: &[&str] = &[ |
| 147 | "deepseek-v4-pro", |
| 148 | "deepseek-v4-flash", |
| 149 | "minimax-m3", |
| 150 | "minimax-m2.7", |
| 151 | "minimax-m2.5", |
| 152 | // glm-5.3 is deliberately absent (2026-08-03): this snapshot tracks the |
| 153 | // official OpenCode Zen endpoint table, which lists no glm-5.3 row. Zen |
| 154 | // fails closed on unknown models by design; registering a route Zen does |
| 155 | // not serve would convert that into a guaranteed upstream 404. |
| 156 | "glm-5.2", |
| 157 | "glm-5.1", |
| 158 | "glm-5", |
| 159 | "kimi-k2.5", |
| 160 | "kimi-k2.6", |
| 161 | "kimi-k2.7-code", |
| 162 | "grok-4.5", |
| 163 | "grok-build-0.1", |
| 164 | "big-pickle", |
| 165 | "mimo-v2.5-free", |
| 166 | "north-mini-code-free", |
| 167 | "nemotron-3-ultra-free", |
| 168 | "deepseek-v4-flash-free", |
| 169 | ]; |
| 170 | |
| 171 | /// Return curated provider/model transport facts as owned offering rows. |
| 172 | /// |
| 173 | /// OpenCode Zen's official catalog serves models over three protocol families. |
| 174 | /// These rows intentionally carry no inferred limits, pricing, or canonical |
| 175 | /// identity: their sole claim is the documented wire model and endpoint key. |
| 176 | #[must_use] |
| 177 | pub fn bundled_offerings() -> Vec<ProviderModelOffering> { |
| 178 | // DeepSeek's 2026-07-31 production Flash update added a native Responses |
| 179 | // endpoint without changing the model id. Pro remains Chat Completions |
| 180 | // until its announced Responses rollout. These exact-route transport facts |
| 181 | // cannot be represented by the Models.dev-shaped fallback asset. |
| 182 | let deepseek = ProviderId::from("deepseek"); |
| 183 | let documented_capabilities = RouteCapabilities { |
| 184 | image_input: CapabilityState::Unsupported, |
| 185 | reasoning: CapabilityState::Supported, |
| 186 | native_tool_calls: CapabilityState::Supported, |
| 187 | structured_output: CapabilityState::Supported, |
| 188 | parallel_tool_calls: CapabilityState::Supported, |
| 189 | streaming: CapabilityState::Supported, |
| 190 | prompt_caching: CapabilityState::Supported, |
| 191 | // The endpoint supports native web search, but Codewhale does not yet |
| 192 | // replay `web_search_call` items on this stateless route. Keep the |
| 193 | // executable capability honest until that loop is implemented. |
| 194 | server_side_web_search: CapabilityState::Unknown, |
| 195 | ..RouteCapabilities::default() |
| 196 | }; |
| 197 | let documented_limits = RouteLimits { |
| 198 | context_tokens: Some(1_000_000), |
| 199 | input_tokens: None, |
| 200 | output_tokens: Some(384_000), |
| 201 | }; |
| 202 | let mut offerings = vec![ |
| 203 | ProviderModelOffering { |
| 204 | provider: deepseek.clone(), |
| 205 | canonical_model: Some(ModelId::from("deepseek-v4-pro")), |
| 206 | wire_model_id: WireModelId::from("deepseek-v4-pro"), |
| 207 | endpoint_key: "chat".to_string(), |
| 208 | default_for_provider: true, |
| 209 | limits: documented_limits, |
| 210 | capabilities: documented_capabilities, |
| 211 | pricing: PricingSku::UnknownOrStale, |
| 212 | }, |
| 213 | ProviderModelOffering { |
| 214 | provider: deepseek, |
| 215 | canonical_model: Some(ModelId::from("deepseek-v4-flash")), |
| 216 | wire_model_id: WireModelId::from("deepseek-v4-flash"), |
| 217 | endpoint_key: "responses".to_string(), |
| 218 | default_for_provider: false, |
| 219 | limits: documented_limits, |
| 220 | capabilities: documented_capabilities, |
| 221 | pricing: PricingSku::UnknownOrStale, |
| 222 | }, |
| 223 | ]; |
| 224 | |
| 225 | let provider = ProviderId::from("opencode-zen"); |
| 226 | let groups = [ |
| 227 | ("responses", OPENCODE_ZEN_RESPONSES_MODELS), |
| 228 | ("messages", OPENCODE_ZEN_MESSAGES_MODELS), |
| 229 | ("chat", OPENCODE_ZEN_CHAT_MODELS), |
| 230 | ]; |
| 231 | |
| 232 | offerings.extend(groups.into_iter().flat_map(|(endpoint_key, models)| { |
| 233 | let provider = provider.clone(); |
| 234 | models.iter().map(move |model| ProviderModelOffering { |
| 235 | provider: provider.clone(), |
| 236 | // The bundled catalog exposes `gpt-5.6` as the user-facing |
| 237 | // logical choice and records `gpt-5.6-sol` as its proven Zen |
| 238 | // wire id. Keep the generic choice honest by resolving it to the |
| 239 | // documented concrete Responses model rather than sending an |
| 240 | // unproven generic wire id to Zen. |
| 241 | canonical_model: (*model == "gpt-5.6-sol").then(|| ModelId::from("gpt-5.6")), |
| 242 | wire_model_id: WireModelId::from(*model), |
| 243 | endpoint_key: endpoint_key.to_string(), |
| 244 | default_for_provider: *model == "gpt-5.6-sol", |
| 245 | limits: RouteLimits::default(), |
| 246 | capabilities: RouteCapabilities::default(), |
| 247 | pricing: PricingSku::UnknownOrStale, |
| 248 | }) |
| 249 | })); |
| 250 | |
| 251 | // Alibaba Cloud Model Studio — one vendor identity in the hand seam |
| 252 | // (`modelstudio-token-plan`). Plan (token vs coding) and wire dialect |
| 253 | // (OpenAI Chat Completions vs Anthropic Messages) are config (`mode` / |
| 254 | // `wire`), not separate ProviderKinds — same product shape as Z.ai / |
| 255 | // Xiaomi for plans and a power-user toggle for dialect. Legacy provider |
| 256 | // ids still get catalog rows so old configs resolve, but the picker |
| 257 | // catalog surface only lists the primary id. |
| 258 | // |
| 259 | // Limits: owner's Token Plan console + curated models_dev rows |
| 260 | // (2026-08-03): qwen3.8-max is ~1M context / 128K output, NOT 128K |
| 261 | // total. Empty RouteLimits here used to win identity collisions over |
| 262 | // the asset catalog and fall through to the 128K legacy default. |
| 263 | fn ms_capabilities(model: &str) -> RouteCapabilities { |
| 264 | let image_input = match model { |
| 265 | "qwen3.8-max" | "qwen3.8-max-preview" | "qwen3.7-plus" | "qwen3.6-flash" => { |
| 266 | CapabilityState::Supported |
| 267 | } |
| 268 | _ => CapabilityState::Unsupported, |
| 269 | }; |
| 270 | RouteCapabilities { |
| 271 | reasoning: CapabilityState::Supported, |
| 272 | native_tool_calls: CapabilityState::Supported, |
| 273 | structured_output: CapabilityState::Supported, |
| 274 | streaming: CapabilityState::Supported, |
| 275 | image_input, |
| 276 | ..RouteCapabilities::default() |
| 277 | } |
| 278 | } |
| 279 | fn ms_limits(model: &str) -> RouteLimits { |
| 280 | // Context/output from models_dev.bundled.json Model Studio rows and |
| 281 | // the owner console (verified 2026-08-03). Keep output separate from |
| 282 | // context so a 128K generation ceiling is never mistaken for the |
| 283 | // window. |
| 284 | let (context_tokens, output_tokens) = match model { |
| 285 | "qwen3.8-max" | "qwen3.8-max-preview" => (1_000_000, 131_072), |
| 286 | "qwen3.7-plus" | "qwen3.7-max" => (1_000_000, 65_536), |
| 287 | "qwen3.6-flash" => (1_000_000, 65_536), |
| 288 | "deepseek-v4-pro" | "deepseek-v4-flash-0731" => (1_000_000, 384_000), |
| 289 | "glm-5.2" => (1_000_000, 131_072), |
| 290 | _ => (1_000_000, 131_072), |
| 291 | }; |
| 292 | RouteLimits { |
| 293 | context_tokens: Some(context_tokens), |
| 294 | input_tokens: None, |
| 295 | output_tokens: Some(output_tokens), |
| 296 | } |
| 297 | } |
| 298 | // Primary vendor id only in the hand seam. Coding-plan / anthropic |
| 299 | // dialect endpoint selection is owned by config resolution (mode/wire), |
| 300 | // which rewrites base_url + request dialect without inventing kinds. |
| 301 | let plan = ProviderId::from("modelstudio-token-plan"); |
| 302 | offerings.extend( |
| 303 | MODELSTUDIO_TEXT_MODELS |
| 304 | .iter() |
| 305 | .enumerate() |
| 306 | .map(|(i, model)| ProviderModelOffering { |
| 307 | provider: plan.clone(), |
| 308 | canonical_model: None, |
| 309 | wire_model_id: WireModelId::from(*model), |
| 310 | endpoint_key: "chat".to_string(), |
| 311 | default_for_provider: i == 0, |
| 312 | limits: ms_limits(model), |
| 313 | capabilities: ms_capabilities(model), |
| 314 | pricing: PricingSku::UnknownOrStale, |
| 315 | }), |
| 316 | ); |
| 317 | |
| 318 | offerings |
| 319 | } |
| 320 |