@linxiraos/pi-catalog 1.1.4 → 1.1.6

package/CHANGELOG.md

@@ -2,11 +2,13 @@
 
 ## [Unreleased]
 
-## [1.1.4] - 2026-08-26
+## [1.1.6] - 2026-08-30
 
-### Changed
+- 同步上游 OMP v18.0.9（`cc14e04f075d`）。
+
+## [1.1.5] - 2026-08-26
 
-- 同步 1.1.4 发布线（与 1.1.3 无功能差异）。
+- 同步上游 OMP v18.0.5 / v18.0.6：新增 DeepInfra / Yolo-Auto 提供商标识。
 
 ## [1.1.3] - 2026-08-25
 

package/dist/types/identity/family.d.ts

@@ -145,24 +145,28 @@ export declare const supportsCodexReasoningSummary: (modelId: string) => boolean
 export declare const isOpenAISamplingRestrictedModelId: (modelId: string) => boolean;
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
- * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning
- * `-flash`/`-flashx`/`-preview` variants, and pre-4.5 ids. Matching the family
- * keeps newly-bumped integers (`glm-5.3`, `glm-6`, …) covered without a per-id
- * allowlist.
+ * `-turbo` lines, plus `-flash` from GLM-5.3-Flash on. Excludes the vision
+ * (`…v`) shape, the non-reasoning `-flashx`/`-preview` variants, and pre-4.5
+ * ids. Matching the family keeps newly-bumped integers (`glm-5.4`, `glm-6`, …)
+ * covered without a per-id allowlist.
  */
 export declare const isReasoningGlmModelId: (modelId: string) => boolean;
 /** GLM-5.2+ coding SKUs accept `reasoning_effort` in addition to binary thinking. */
 export declare const isGlm52ReasoningEffortModelId: (modelId: string) => boolean;
 /**
- * GLM-5.3+ coding SKUs. Unlike GLM-5.2 (whose reasoning_effort dialect is
- * host-specific), GLM-5.3+ exposes a uniform wire-exact `low`/`high`/`max`
- * ladder on every host, and thinking can no longer be disabled —
- * `thinking.type` must always be `enabled`. Matching the family keeps future
- * bumps (`glm-5.4`, `glm-6`, …) covered while excluding the vision (`…v`)
- * shape and the non-reasoning `-flash`/`-flashx`/`-preview` variants.
+ * GLM-5.3+ reasoning-effort SKUs. Unlike GLM-5.2 (whose dialect is
+ * host-specific), the base, `-air`, `-turbo`, and `-flash` variants expose a
+ * uniform wire-exact `low`/`high`/`max` ladder on every host, and thinking can
+ * no longer be disabled — `thinking.type` must always be `enabled`. Matching
+ * the family keeps future bumps (`glm-5.4`, `glm-6`, …) covered while excluding
+ * the vision (`…v`) shape and the non-reasoning `-flashx`/`-preview` variants.
  */
 export declare const isGlm53ReasoningEffortModelId: (modelId: string) => boolean;
-/** GLM vision SKUs — the `v` that attaches to the version (`glm-4v`, `glm-4.5v`). */
+/**
+ * GLM SKUs that accept image input: the `v` that attaches to the version
+ * (`glm-4v`, `glm-4.5v`) and the natively multimodal `-flash` line from
+ * GLM-5.3-Flash on, whose id carries no `v` marker.
+ */
 export declare const isGlmVisionModelId: (modelId: string) => boolean;
 /**
  * Coarse vendor-lineage token for "are two models the same family?" checks

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

@@ -11,6 +11,8 @@ export interface ModelsDevFallback<TApi extends Api = Api, TPayload = unknown> {
     fetch(): Promise<TPayload>;
     /** Maps payload into provider models. */
     map(payload: TPayload, providerId: Provider): readonly ModelSpec<TApi>[];
+    /** When true, mapped rows can add model ids but cannot replace static metadata. */
+    additiveOnly?: boolean;
 }
 /**
  * Configuration for provider model resolution.
@@ -45,18 +47,25 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
     /** Clock override for deterministic tests. */
     now?: () => number;
 }
+/**
+ * Catalog source that most recently refreshed the resolved provider snapshot.
+ */
+export type ModelResolutionSource = "bundled" | "cache" | "models.dev" | "provider";
 /**
  * Resolution result.
  *
  * `stale` is false when the resolved catalog is authoritative for the selected provider:
- * - a dynamic endpoint fetch succeeded in this call (an empty catalog is still
+ * - a provider endpoint fetch succeeded in this call (an empty catalog is still
  *   authoritative for the cycle, so downstream pruning of removed models runs),
+ * - a models.dev fetch succeeded for a provider without endpoint discovery,
  * - a still-fresh authoritative cache was reused in `online-if-uncached` mode, or
- * - the provider has no dynamic fetcher configured.
+ * - the provider has no remote fetcher configured.
  */
 export interface ModelResolutionResult<TApi extends Api = Api> {
     models: Model<TApi>[];
     stale: boolean;
+    source: ModelResolutionSource;
+    updatedAt?: number;
 }
 /**
  * Stateful facade over provider model resolution.
@@ -70,8 +79,15 @@ export interface ModelManager<TApi extends Api = Api> {
 export declare function createModelManager<TApi extends Api = Api, TModelsDevPayload = unknown>(options: ModelManagerOptions<TApi, TModelsDevPayload>): ModelManager<TApi>;
 /**
  * Resolves provider models with source precedence:
- * static -> stencil.so -> cache -> dynamic.
+ * static -> cached fallback -> stencil.so -> dynamic.
  *
- * Later sources override earlier ones by model id.
+ * Later sources override earlier ones by model id. Cached rows participate only
+ * when at least one configured remote source did not refresh successfully.
  */
 export declare function resolveProviderModels<TApi extends Api = Api, TModelsDevPayload = unknown>(options: ModelManagerOptions<TApi, TModelsDevPayload>, strategy?: ModelRefreshStrategy): Promise<ModelResolutionResult<TApi>>;
+/**
+ * Return the versioned, low-collision model-cache identity for a static provider
+ * slice. Results are cached by array reference so repeat cold-start paths skip
+ * the JSON serialization and hash.
+ */
+export declare function fingerprintStaticModels<TApi extends Api>(models: readonly (ModelSpec<TApi> | Model<TApi>)[], dynamicModelsAuthoritative?: boolean): string;

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/models-dev-policies.d.ts

@@ -0,0 +1,8 @@
+import type { Api, ModelSpec } from "../types.js";
+/**
+ * Remove models.dev rows that OMP cannot route successfully.
+ *
+ * Generation and runtime refresh share this policy so a live catalog cannot
+ * reintroduce selectors deliberately excluded from the bundled catalog.
+ */
+export declare function filterModelsDevCatalogRows<TApi extends Api>(models: readonly ModelSpec<TApi>[]): ModelSpec<TApi>[];

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

@@ -1,4 +1,4 @@
-import type { ModelManagerOptions } from "../model-manager.js";
+import type { ModelManagerOptions, ModelsDevFallback } from "../model-manager.js";
 import { getBundledModels } from "../models.js";
 import type { Api, FetchImpl, ModelSpec, Provider } from "../types.js";
 import { ALIBABA_TOKEN_PLAN_BASE_URL } from "../wire/alibaba-token-plan.js";
@@ -32,9 +32,12 @@ export interface ModelsDevModel {
  * The frame magic is sniffed rather than trusting content-type so plain-JSON
  * responses (test stubs, fallback mirrors) parse identically.
  *
- * Fetched fully once per process: concurrent callers share the in-flight
- * request, repeat callers send a conditional GET that the server answers
- * (and deliberately does not log) with `304`.
+ * Fetched fully once per fetch context: concurrent callers sharing a fetch
+ * implementation reuse one transport request, while repeat callers send a
+ * conditional GET that the server answers with `304`. Each subscriber may stop
+ * waiting independently; the shared transport retains its own hard deadline.
+ * Transient failures reuse the last in-memory payload for callers that only
+ * need best-effort metadata.
  */
 export declare function fetchWellKnownModels(fetchImpl?: FetchImpl, signal?: AbortSignal): Promise<unknown>;
 /**
@@ -161,6 +164,19 @@ 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;
@@ -528,6 +544,29 @@ 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;
@@ -649,3 +688,16 @@ export interface ModelsDevProviderDescriptor {
 export declare function mapModelsDevToModels(data: Record<string, unknown>, descriptors: readonly ModelsDevProviderDescriptor[]): ModelSpec<Api>[];
 /** All provider descriptors for models.dev data mapping in generate-models.ts. */
 export declare const MODELS_DEV_PROVIDER_DESCRIPTORS: readonly ModelsDevProviderDescriptor[];
+/** Providers whose bundled catalog can receive additive models.dev updates at runtime. */
+export declare const MODELS_DEV_CATALOG_PROVIDER_IDS: readonly string[];
+/**
+ * Build the shared models.dev fallback for one known provider.
+ *
+ * Provider managers sharing one fetch implementation reuse its conditional
+ * catalog session. Each mapped provider slice is persisted independently so
+ * startup can restore it without parsing the full catalog.
+ *
+ * `timeoutMs` bounds the catalog request. It is configurable for callers with a
+ * stricter startup budget and for deterministic timeout tests.
+ */
+export declare function modelsDevCatalogFallback(providerId: string, fetchImpl?: FetchImpl, timeoutMs?: number): ModelsDevFallback<Api> | undefined;

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.

package/dist/types/wire/cloudflare-ai-gateway.d.ts

@@ -0,0 +1,18 @@
+/** Provider-native AI Gateway prefix before the upstream route segment. */
+export declare const CLOUDFLARE_AI_GATEWAY_BASE_URL = "https://gateway.ai.cloudflare.com/v1/<account>/<gateway>";
+/** Anthropic Messages passthrough endpoint. */
+export declare const CLOUDFLARE_AI_GATEWAY_ANTHROPIC_BASE_URL = "https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/anthropic";
+/** OpenAI Responses passthrough endpoint. */
+export declare const CLOUDFLARE_AI_GATEWAY_OPENAI_BASE_URL = "https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/openai";
+/** OpenAI-compatible Workers AI endpoint. */
+export declare const CLOUDFLARE_AI_GATEWAY_COMPAT_BASE_URL = "https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/compat";
+/** Opaque API-key credential payload persisted by Cloudflare AI Gateway login. */
+export interface CloudflareAiGatewayCredential {
+    token: string;
+    accountId?: string;
+    gatewayId?: string;
+}
+/** Parse both structured login credentials and legacy plain gateway tokens. */
+export declare function parseCloudflareAiGatewayCredential(value: string): CloudflareAiGatewayCredential | null;
+/** Serialize the gateway token with the routing identifiers collected during login. */
+export declare function serializeCloudflareAiGatewayCredential(token: string, accountId: string, gatewayId: string): string;

package/package.json

@@ -1,8 +1,8 @@
 {
 	"type": "module",
 	"name": "@linxiraos/pi-catalog",
-	"version": "1.1.4",
-	"description": "Model catalog for Zeta: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
+	"version": "1.1.6",
+	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://linxira-os.github.io/zeta/",
 	"author": "Can Boluk",
 	"license": "MIT",
@@ -34,11 +34,11 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@linxiraos/pi-omptype": "1.1.4",
-		"@linxiraos/pi-utils": "1.1.4"
+		"@linxiraos/pi-omptype": "1.1.6",
+		"@linxiraos/pi-utils": "1.1.6"
 	},
 	"devDependencies": {
-		"@linxiraos/pi-ai": "1.1.4",
+		"@linxiraos/pi-ai": "1.1.6",
 		"@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/family.ts

@@ -9,6 +9,7 @@
 
 import {
 	bareModelId,
+	type GlmModel,
 	isAnthropicAdaptiveGenAtLeast,
 	isFableOrMythos,
 	parseAnthropicModel,
@@ -283,19 +284,35 @@ export const isOpenAISamplingRestrictedModelId = memo((modelId: string): boolean
 	return isOpenAIWireGen5Plus(modelId) || O_SERIES_REASONING_RE.test(bare);
 });
 
+/**
+ * GLM SKU shapes that carry the reasoning ladders. `base`/`air`/`turbo` have
+ * always been reasoning lines; `flash` joins at GLM-5.3-Flash, the first flash
+ * SKU whose thinking is mandatory (pre-5.3 `-flash` SKUs are non-reasoning).
+ * `flashx` and `preview` stay out.
+ */
+function isGlmReasoningVariant(glm: GlmModel): boolean {
+	switch (glm.variant) {
+		case "base":
+		case "air":
+		case "turbo":
+			return true;
+		case "flash":
+			return semverGte(glm.version, "5.3");
+		default:
+			return false;
+	}
+}
+
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
- * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning
- * `-flash`/`-flashx`/`-preview` variants, and pre-4.5 ids. Matching the family
- * keeps newly-bumped integers (`glm-5.3`, `glm-6`, …) covered without a per-id
- * allowlist.
+ * `-turbo` lines, plus `-flash` from GLM-5.3-Flash on. Excludes the vision
+ * (`…v`) shape, the non-reasoning `-flashx`/`-preview` variants, and pre-4.5
+ * ids. Matching the family keeps newly-bumped integers (`glm-5.4`, `glm-6`, …)
+ * covered without a per-id allowlist.
  */
 export const isReasoningGlmModelId = memo((modelId: string): boolean => {
 	const glm = parseGlmModel(bareModelId(modelId));
-	if (!glm || glm.vision) {
-		return false;
-	}
-	if (glm.variant !== "base" && glm.variant !== "air" && glm.variant !== "turbo") {
+	if (!glm || glm.vision || !isGlmReasoningVariant(glm)) {
 		return false;
 	}
 	return semverGte(glm.version, "4.5");
@@ -304,37 +321,39 @@ export const isReasoningGlmModelId = memo((modelId: string): boolean => {
 /** GLM-5.2+ coding SKUs accept `reasoning_effort` in addition to binary thinking. */
 export const isGlm52ReasoningEffortModelId = memo((modelId: string): boolean => {
 	const glm = parseGlmModel(bareModelId(modelId));
-	if (!glm || glm.vision) {
-		return false;
-	}
-	if (glm.variant !== "base" && glm.variant !== "air" && glm.variant !== "turbo") {
+	if (!glm || glm.vision || !isGlmReasoningVariant(glm)) {
 		return false;
 	}
 	return semverGte(glm.version, "5.2");
 });
 
 /**
- * GLM-5.3+ coding SKUs. Unlike GLM-5.2 (whose reasoning_effort dialect is
- * host-specific), GLM-5.3+ exposes a uniform wire-exact `low`/`high`/`max`
- * ladder on every host, and thinking can no longer be disabled —
- * `thinking.type` must always be `enabled`. Matching the family keeps future
- * bumps (`glm-5.4`, `glm-6`, …) covered while excluding the vision (`…v`)
- * shape and the non-reasoning `-flash`/`-flashx`/`-preview` variants.
+ * GLM-5.3+ reasoning-effort SKUs. Unlike GLM-5.2 (whose dialect is
+ * host-specific), the base, `-air`, `-turbo`, and `-flash` variants expose a
+ * uniform wire-exact `low`/`high`/`max` ladder on every host, and thinking can
+ * no longer be disabled — `thinking.type` must always be `enabled`. Matching
+ * the family keeps future bumps (`glm-5.4`, `glm-6`, …) covered while excluding
+ * the vision (`…v`) shape and the non-reasoning `-flashx`/`-preview` variants.
  */
 export const isGlm53ReasoningEffortModelId = memo((modelId: string): boolean => {
 	const glm = parseGlmModel(bareModelId(modelId));
-	if (!glm || glm.vision) {
-		return false;
-	}
-	if (glm.variant !== "base" && glm.variant !== "air" && glm.variant !== "turbo") {
+	if (!glm || glm.vision || !isGlmReasoningVariant(glm)) {
 		return false;
 	}
 	return semverGte(glm.version, "5.3");
 });
 
-/** GLM vision SKUs — the `v` that attaches to the version (`glm-4v`, `glm-4.5v`). */
+/**
+ * GLM SKUs that accept image input: the `v` that attaches to the version
+ * (`glm-4v`, `glm-4.5v`) and the natively multimodal `-flash` line from
+ * GLM-5.3-Flash on, whose id carries no `v` marker.
+ */
 export const isGlmVisionModelId = memo((modelId: string): boolean => {
-	return parseGlmModel(bareModelId(modelId))?.vision === true;
+	const glm = parseGlmModel(bareModelId(modelId));
+	if (!glm) {
+		return false;
+	}
+	return glm.vision || (glm.variant === "flash" && semverGte(glm.version, "5.3"));
 });
 
 /**

package/src/identity/priority.ts

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