@oh-my-pi/pi-catalog 18.0.3 → 18.0.5

package/CHANGELOG.md

@@ -2,6 +2,25 @@
 
 ## [Unreleased]
 
+## [18.0.5] - 2026-08-25
+
+### Added
+
+- Added built-in DeepInfra provider support (`deepinfra`, `DEEPINFRA_API_KEY`) with live model discovery, including chat models, context windows, pricing, cache-read pricing, vision input, and reasoning-effort capabilities.
+- Added built-in Yolo-Auto provider support with the flat-rate `deepseek-flash-v4` model and live model discovery.
+
+### Fixed
+
+- Fixed the Synthetic provider’s default model to use `hf:zai-org/GLM-5.2` instead of the retired `hf:zai-org/GLM-5.1`.
+
+## [18.0.4] - 2026-08-24
+
+### Fixed
+
+- Fixed default reasoning effort for `cursor/cursor-grok-4.5` and `cursor/cursor-grok-4.6` so requests without an explicit effort setting default to `-medium` instead of `-low`, preventing rate limit rejections on Cursor's Start plan ([#9478](https://github.com/can1357/oh-my-pi/issues/9478)).
+- Fixed aliased OpenCode Zen Ox Alpha models exposing incorrect effort levels, ensuring the gateway's native `low`, `high`, and `max` tiers are correctly mapped and reachable ([#9349](https://github.com/can1357/oh-my-pi/issues/9349)).
+- Fixed missing rate card tier for public xAI and SuperGrok models with prompt lengths exceeding 200K tokens ([#9512](https://github.com/can1357/oh-my-pi/issues/9512)).
+
 ## [18.0.2] - 2026-08-23
 
 ### Fixed

package/dist/types/model-thinking.d.ts

@@ -67,4 +67,15 @@ export declare function resolveWireModelId<TApi extends Api>(model: ApiModel<TAp
  * thinking-off requests on `thinking.requiresEffort` models.
  */
 export declare function minimumSupportedEffort<TApi extends Api>(model: ApiModel<TApi>): Effort | undefined;
+/**
+ * Clamp target for effort-less requests on `thinking.requiresEffort` models:
+ * the effort whose wire route equals the model's default wire id
+ * (`requestModelId`), so a collapsed row clamps to the tier it already
+ * advertises as its default rather than the numerically lowest supported tier
+ * (e.g. Cursor Grok 4.5/4.6 default to `medium`, the only tier the Start plan
+ * serves). Falls back to {@link minimumSupportedEffort} when no route matches
+ * the default id — families whose default already is the minimum, or that
+ * expose no routing, are unaffected.
+ */
+export declare function defaultSupportedEffort<TApi extends Api>(model: ApiModel<TApi>): Effort | undefined;
 export {};

package/dist/types/models.d.ts

@@ -1,11 +1,14 @@
 import MODELS from "./models.json";
-import type { Api, KnownProvider, Model, Usage } from "./types.js";
+import type { Api, KnownProvider, Model, ModelCost, Usage } from "./types.js";
 export type GeneratedProvider = keyof typeof MODELS;
 export declare function getBundledModel<TApi extends Api = Api>(provider: GeneratedProvider, modelId: string): Model<TApi>;
 export declare function getBundledProviders(): KnownProvider[];
 export declare function getBundledModels(provider: GeneratedProvider): Model<Api>[];
 /** Price a prompt as fully uncached input under its active context-length tier. */
-export declare function calculateUncachedInputCost(cost: Model["cost"], promptInputTokens: number): number;
+export declare function calculateUncachedInputCost(cost: ModelCost, promptInputTokens: number): number;
+/** Price one usage record from a token rate card, including active context tiers. */
+export declare function calculateUsageCost(cost: ModelCost, usage: Usage): Usage["cost"];
+/** Price one usage record from its model's token rate card. */
 export declare function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage["cost"];
 /**
  * Check if two models are equal by comparing both their id and provider.

package/dist/types/provider-models/descriptors.d.ts

@@ -97,6 +97,16 @@ export declare const CATALOG_PROVIDERS: readonly [{
         readonly envVars: readonly ["CURSOR_API_KEY"];
         readonly oauthProvider: "cursor";
     };
+}, {
+    readonly id: "deepinfra";
+    readonly defaultModel: "deepseek-ai/DeepSeek-V4-Flash-0731";
+    readonly envVars: readonly ["DEEPINFRA_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "DeepInfra";
+        readonly allowUnauthenticated: true;
+    };
 }, {
     readonly id: "deepseek";
     readonly defaultModel: "deepseek-v4-pro";
@@ -361,7 +371,7 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly dynamicModelsAuthoritative: true;
 }, {
     readonly id: "synthetic";
-    readonly defaultModel: "hf:zai-org/GLM-5.1";
+    readonly defaultModel: "hf:zai-org/GLM-5.2";
     readonly envVars: readonly ["SYNTHETIC_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
     readonly dynamicModelsAuthoritative: true;
@@ -469,6 +479,15 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly defaultModel: "mimo-v2.5";
     readonly envVars: readonly ["XIAOMI_TOKEN_PLAN_SGP_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+}, {
+    readonly id: "yolo-auto";
+    readonly defaultModel: "deepseek-flash-v4";
+    readonly envVars: readonly ["YOLO_AUTO_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("../index.js").ModelManagerOptions<"openai-completions", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "Yolo-Auto";
+    };
 }, {
     readonly id: "zai";
     readonly defaultModel: "glm-5.3";

package/dist/types/provider-models/openai-compat.d.ts

@@ -161,11 +161,29 @@ export interface NovitaModelManagerConfig {
 }
 /** Builds Novita's public model-discovery manager. */
 export declare function novitaModelManagerOptions(config?: NovitaModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export declare const DEEPINFRA_BASE_URL = "https://api.deepinfra.com/v1/openai";
+/** DeepInfra OpenAI-compatible discovery configuration. */
+export interface DeepinfraModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+/**
+ * Builds DeepInfra's model-discovery manager. The catalog endpoint is public,
+ * so discovery (and keyless `gen:models` generation) works without an API key;
+ * model availability at runtime is still gated on credentials by the registry.
+ */
+export declare function deepinfraModelManagerOptions(config?: DeepinfraModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface XaiModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+/**
+ * Applies xAI's long-context rate card and mirrors exact public-model prices
+ * onto matching SuperGrok catalog rows.
+ */
+export declare function applyXaiCatalogPricing(models: readonly ModelSpec[]): ModelSpec[];
 export declare function xaiModelManagerOptions(config?: XaiModelManagerConfig): ModelManagerOptions<"openai-responses">;
 export interface XaiOAuthModelManagerConfig {
     apiKey?: string;
@@ -523,6 +541,29 @@ export interface AiandModelManagerConfig {
  * pricing metadata, so discovery is authoritative over the bundled seed.
  */
 export declare function aiandModelManagerOptions(config?: AiandModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Documented Yolo-Auto catalog (yolo-auto.com/docs, 2026-08) bundled so the
+ * provider is usable when generation and first boot have no live key. The
+ * flat-rate `/v1/models` response is authoritative once discovery runs.
+ * The compat block mirrors the provider's documented wire surface: the API
+ * speaks the generic chat template with `reasoning_effort` support and rejects
+ * the `developer` role and `store` param.
+ */
+export declare const YOLO_AUTO_STATIC_MODELS: readonly ModelSpec<"openai-completions">[];
+export interface YoloAutoModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+/**
+ * Yolo-Auto model manager: OpenAI-compatible chat completions at the
+ * flat-rate `deepseek-flash-v4` endpoint. Live `/v1/models` discovery replaces the
+ * bundled seed once a key is stored or present in `YOLO_AUTO_API_KEY`. Models
+ * the provider adds later inherit metadata (reasoning, thinking, context)
+ * from the global bundled reference index, so new ids work without a per-id
+ * code change.
+ */
+export declare function yoloAutoModelManagerOptions(config?: YoloAutoModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface QwenPortalModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -121,8 +121,8 @@ export interface Usage {
         total: number;
     };
 }
-export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "kimi" | "qwen" | "qwen-chat-template";
-export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "venice-disable-thinking" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false";
+export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "kimi" | "qwen" | "qwen-chat-template" | "chat-template";
+export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "none-effort" | "openrouter-enabled-false" | "venice-disable-thinking" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false" | "chat-template-thinking-false";
 export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "qwen" | "thinking";
 /**
  * Compatibility settings for openai-completions API.
@@ -705,12 +705,15 @@ export interface TokenCost {
     cacheWrite: number;
 }
 /**
- * Rates applied to the full request when its prompt exceeds `inputThreshold`.
- * Prompt input is the sum of uncached, cached-read, cache-write, and
- * provider-orchestration input tokens.
+ * Rates applied to the full request when its prompt exceeds `inputThreshold`,
+ * or reaches it when `inputThresholdInclusive` is true. Prompt input is the
+ * sum of uncached, cached-read, cache-write, and provider-orchestration input
+ * tokens.
  */
 export interface LongContextTokenCost extends TokenCost {
     inputThreshold: number;
+    /** Whether the long-context tier starts exactly at `inputThreshold`. */
+    inputThresholdInclusive?: boolean;
 }
 /** Base token rates plus an optional long-context tier. */
 export interface ModelCost extends TokenCost {

package/dist/types/variant-collapse.d.ts

@@ -21,6 +21,15 @@ export interface EffortVariantFamily {
      * when it equals the logical id).
      */
     members: readonly string[];
+    /**
+     * Preferred default wire id: overrides the member-order default for the
+     * collapsed spec's `requestModelId` when this member is live. Lets a
+     * mandatory-reasoning family advertise a canonical tier that is not its
+     * numeric floor (Cursor Grok 4.5/4.6 default to `-medium`, the only tier the
+     * Start plan serves; the `-low` floor is refused). Ignored when absent from
+     * the input or retired.
+     */
+    defaultMember?: string;
     /**
      * Wire ids upstream no longer serves (e.g. a deployment killed while
      * discovery still advertises it). Fresh collapsing never routes to them,

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.0.3",
+	"version": "18.0.5",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -34,11 +34,11 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.0.3",
-		"@oh-my-pi/pi-utils": "18.0.3"
+		"@oh-my-pi/omptype": "18.0.5",
+		"@oh-my-pi/pi-utils": "18.0.5"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "18.0.3",
+		"@oh-my-pi/pi-ai": "18.0.5",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -94,6 +94,8 @@ function resolveReasoningDisableMode(
 			return "qwen-enable-thinking-false";
 		case "qwen-chat-template":
 			return "qwen-template-false";
+		case "chat-template":
+			return "chat-template-thinking-false";
 		default:
 			return "lowest-effort";
 	}

package/src/identity/priority.ts

@@ -22,6 +22,7 @@ const DEFAULT_MODEL_PROVIDER_ORDER = [
 	"fireworks",
 	"cerebras",
 	"baseten",
+	"deepinfra",
 	"openrouter",
 	"aimlapi",
 	"together",

package/src/model-manager.ts

@@ -492,9 +492,17 @@ function mergeDynamicModel<TApi extends Api>(existingModel: Model<TApi>, dynamic
 	// dynamic discovery also pre-applies the correct image fallback for omitted
 	// `supports.vision`, so its explicit `false` must not be OR-upgraded by the
 	// canonical bundled model.
+	// DeepInfra's discovery is authoritative (`dynamicModelsAuthoritative`) and
+	// its `vision`/`vlm` tags are the catalog's whole truth for modality. Every
+	// row shares the single DeepInfra endpoint, so `endpointChanged` never fires
+	// there: without this carve-out a model that dropped those tags would keep
+	// the bundled reference's image support and the agent would go on sending
+	// images to a now text-only route.
 	const endpointChanged = existingModel.baseUrl !== dynamicModel.baseUrl;
 	const dynamicInputAuthoritative =
-		endpointChanged || (existingModel.provider === "github-copilot" && dynamicModel.provider === "github-copilot");
+		endpointChanged ||
+		(existingModel.provider === "github-copilot" && dynamicModel.provider === "github-copilot") ||
+		(existingModel.provider === "deepinfra" && dynamicModel.provider === "deepinfra");
 	const supportsImage = dynamicInputAuthoritative
 		? dynamicModel.input.includes("image")
 		: existingModel.input.includes("image") || dynamicModel.input.includes("image");

package/src/model-thinking.ts

@@ -329,6 +329,14 @@ function getModelDefinedEfforts<TApi extends Api>(
 	spec: ModelSpec<TApi>,
 	compat: CompatOf<TApi>,
 ): readonly Effort[] | undefined {
+	if (
+		isOpenAICompatReasoningApi(spec.api) &&
+		isChatTemplateThinkingFormat(compat) &&
+		spec.thinking?.mode === "effort" &&
+		spec.thinking.efforts.length > 0
+	) {
+		return spec.thinking.efforts;
+	}
 	if (isGlm53ReasoningEffortModelId(spec.id)) {
 		// GLM-5.3+ exposes a uniform wire-exact low/high/max ladder on every
 		// host — unlike GLM-5.2, whose reasoning_effort dialect is
@@ -519,6 +527,9 @@ function isQwenTemplateReasoningEffortCompat(compat: CompatOf<Api>): boolean {
 		compat !== undefined && "qwenTemplateReasoningEffort" in compat && compat.qwenTemplateReasoningEffort === true
 	);
 }
+function isChatTemplateThinkingFormat(compat: CompatOf<Api>): boolean {
+	return compat !== undefined && "thinkingFormat" in compat && compat.thinkingFormat === "chat-template";
+}
 
 function inferDetectedEffortMap<TApi extends Api>(
 	spec: ModelSpec<TApi>,
@@ -559,15 +570,19 @@ function isSakanaFuguReasoningModel<TApi extends Api>(spec: ModelSpec<TApi>): bo
  * `opencode-zen`) reason through the wire-exact `low`/`high`/`max` ladder with
  * mandatory thinking: the gateway rejects `minimal`/`medium`/`xhigh`
  * (`[1210] ... please use low, high, or max`), the same dialect it already
- * serves for GLM-5.3 and Kimi K3. Other hosts proxying an `ox-alpha` SKU
- * (Kilo, NanoGPT, Venice, OpenRouter) expose their own vocabularies and are
- * left untouched. See issue #9349.
+ * serves for GLM-5.3 and Kimi K3. The SKU also ships under unrelated aliased
+ * ids (`opencode-zen/x-preview-f-free`), so the stencil display name is
+ * matched as well. Other hosts proxying an `ox-alpha` SKU (Kilo, NanoGPT,
+ * Venice, OpenRouter) expose their own vocabularies and are left untouched.
+ * See issue #9349.
  */
 function isOpenCodeGatewayOxAlphaModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
-	return (
-		(spec.provider === "opencode-go" || spec.provider === "opencode-zen") &&
-		/(?:^|\/)ox-alpha(?:-|$)/i.test(bareModelId(spec.id))
-	);
+	if (spec.provider !== "opencode-go" && spec.provider !== "opencode-zen") return false;
+	if (/(?:^|\/)ox-alpha(?:-|$)/i.test(bareModelId(spec.id))) return true;
+	// Aliased gateway ids surface the same SKU under an unrelated id; the
+	// display name ("Ox Alpha …") is the stable cross-id signal. Anchored on
+	// both sides so "Box Alpha"/"Ox Alphabet" cannot false-positive.
+	return /\box[ _-]?alpha\b/i.test(spec.name ?? "");
 }
 
 function isDeepseekReasoningModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
@@ -923,3 +938,25 @@ export function minimumSupportedEffort<TApi extends Api>(model: ApiModel<TApi>):
 	}
 	return efforts[0];
 }
+
+/**
+ * Clamp target for effort-less requests on `thinking.requiresEffort` models:
+ * the effort whose wire route equals the model's default wire id
+ * (`requestModelId`), so a collapsed row clamps to the tier it already
+ * advertises as its default rather than the numerically lowest supported tier
+ * (e.g. Cursor Grok 4.5/4.6 default to `medium`, the only tier the Start plan
+ * serves). Falls back to {@link minimumSupportedEffort} when no route matches
+ * the default id — families whose default already is the minimum, or that
+ * expose no routing, are unaffected.
+ */
+export function defaultSupportedEffort<TApi extends Api>(model: ApiModel<TApi>): Effort | undefined {
+	const routing = model.thinking?.effortRouting;
+	const defaultWireId = model.requestModelId;
+	if (routing !== undefined && defaultWireId !== undefined) {
+		const efforts = model.thinking?.efforts;
+		for (const effort of THINKING_EFFORTS) {
+			if (efforts?.includes(effort) && routing[effort] === defaultWireId) return effort;
+		}
+	}
+	return minimumSupportedEffort(model);
+}