@linxiraos/pi-catalog 1.1.5 → 1.1.7

package/CHANGELOG.md

@@ -2,6 +2,14 @@
 
 ## [Unreleased]
 
+## [1.1.7] - 2026-09-01
+
+- 同步上游 OMP v18.0.11（`b8ce33a58911c26bed1d84f0db9a5e2e727c49a2`）。
+
+## [1.1.6] - 2026-08-30
+
+- 同步上游 OMP v18.0.9（`cc14e04f075d`）。
+
 ## [1.1.5] - 2026-08-26
 
 - 同步上游 OMP v18.0.5 / v18.0.6：新增 DeepInfra / Yolo-Auto 提供商标识。

package/dist/types/discovery/cursor.d.ts

@@ -21,3 +21,9 @@ export interface CursorModelDiscoveryOptions {
  * Returns `[]` only when the endpoint responds successfully with no usable models.
  */
 export declare function fetchCursorUsableModels(options: CursorModelDiscoveryOptions): Promise<ModelSpec<"cursor-agent">[] | null>;
+/**
+ * Resolves input modalities from a bundled reference when available, except
+ * for Cursor-only families whose image support is known independently. Without
+ * a reference, native multimodal families fall back to id classification.
+ */
+export declare function resolveCursorInput(id: string, referenceInput?: ("text" | "image")[]): ("text" | "image")[];

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/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>;
 /**
@@ -685,3 +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/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

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@linxiraos/pi-catalog",
-	"version": "1.1.5",
+	"version": "1.1.7",
 	"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",
@@ -34,11 +34,11 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@linxiraos/pi-omptype": "1.1.5",
-		"@linxiraos/pi-utils": "1.1.5"
+		"@linxiraos/pi-omptype": "1.1.7",
+		"@linxiraos/pi-utils": "1.1.7"
 	},
 	"devDependencies": {
-		"@linxiraos/pi-ai": "1.1.5",
+		"@linxiraos/pi-ai": "1.1.7",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/discovery/cursor.ts

@@ -33,14 +33,20 @@ const CURSOR_KIMI_K3_BARE_ID_PATTERN = /(^|\/)k3$/i;
  * reasoning models whose effort is carried in the per-tier sibling id.
  * `GetUsableModels` ships no `thinkingDetails` and the bundled references read
  * `reasoning: false`, so classification falls back to the id. The non-reasoning
- * `grok-code-*` coding models lack the version digit and stay out.
+ * `grok-code-fast-*` family deliberately stays out.
  */
 const CURSOR_GROK_REASONING_ID_PATTERN = /^cursor-grok-\d/i;
 
+/**
+ * Cursor-only families verified to accept `selectedImages` even though
+ * `GetUsableModels` does not advertise input modalities.
+ */
+const CURSOR_GROK_4_MULTIMODAL_ID_PATTERN = /^cursor-grok-4(?:[.:_-]|$)/i;
+const CURSOR_COMPOSER_25_MULTIMODAL_ID_PATTERN = /^composer-2\.5(?:[.:_-]|$)/i;
+
 /**
  * Model-id families whose native catalogs (anthropic, openai/openai-codex,
- * google) are multimodal. Cursor-only or text-only families (`composer-*`,
- * `grok-code-*`) intentionally stay outside this pattern.
+ * google) are multimodal. Cursor-only verified families are handled separately.
  */
 const CURSOR_MULTIMODAL_ID_PATTERN = /claude|gemini|gpt-|codex/;
 
@@ -319,6 +325,7 @@ function normalizeCursorModel(
 			name,
 			baseUrl: baseUrlOverride ?? reference.baseUrl,
 			reasoning,
+			input: resolveCursorInput(id, reference.input),
 			contextWindow: resolveCursorContextWindow(details, id, reference.contextWindow),
 			cursorMaxMode: details.maxMode,
 		};
@@ -330,7 +337,7 @@ function normalizeCursorModel(
 		provider: "cursor",
 		baseUrl: baseUrlOverride ?? CURSOR_DEFAULT_BASE_URL,
 		reasoning,
-		input: inferInputFromCursorId(id),
+		input: resolveCursorInput(id),
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 		contextWindow: resolveCursorContextWindow(details, id, DEFAULT_CONTEXT_WINDOW),
 		maxTokens: DEFAULT_MAX_TOKENS,
@@ -394,14 +401,22 @@ function pickModelDisplayName(model: CursorModelDetailsValue, fallbackId: string
 }
 
 /**
- * Infers input modalities for Cursor models without a bundled reference.
- *
- * `GetUsableModels` carries no per-model modality metadata, so classification
- * falls back to the model family: families that are multimodal in OMP's own
- * native catalogs accept images, everything else stays text-only. Mirrors
- * `inferInputFromGeminiId` in ./gemini.ts.
+ * Resolves input modalities from a bundled reference when available, except
+ * for Cursor-only families whose image support is known independently. Without
+ * a reference, native multimodal families fall back to id classification.
  */
-function inferInputFromCursorId(id: string): ("text" | "image")[] {
+export function resolveCursorInput(id: string, referenceInput?: ("text" | "image")[]): ("text" | "image")[] {
+	if (
+		isKimiK3ModelId(id) ||
+		CURSOR_KIMI_K3_BARE_ID_PATTERN.test(id) ||
+		CURSOR_GROK_4_MULTIMODAL_ID_PATTERN.test(id) ||
+		CURSOR_COMPOSER_25_MULTIMODAL_ID_PATTERN.test(id)
+	) {
+		return ["text", "image"];
+	}
+	if (referenceInput) {
+		return referenceInput;
+	}
 	if (CURSOR_MULTIMODAL_ID_PATTERN.test(id.toLowerCase())) {
 		return ["text", "image"];
 	}

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"));
 });
 
 /**