@gajae-code/ai 0.13.2 → 0.14.0

package/src/core.ts

@@ -23,6 +23,7 @@ export {
 	PROVIDER_RUNTIME_DESCRIPTORS,
 	type ProviderRuntimeDescriptor,
 } from "./providers/register-builtins";
+export { hasAdjacentPrivateThinkingBlocks } from "./providers/transform-messages";
 export * from "./rate-limit-utils";
 export * from "./stream";
 export * from "./types";

package/src/model-cache.ts

@@ -6,7 +6,7 @@ import { Database } from "bun:sqlite";
 import { getModelDbPath } from "@gajae-code/utils/dirs";
 import type { Api, Model } from "./types";
 
-const CACHE_SCHEMA_VERSION = 3;
+const CACHE_SCHEMA_VERSION = 5;
 
 interface CacheRow {
 	provider_id: string;
@@ -14,6 +14,8 @@ interface CacheRow {
 	updated_at: number;
 	authoritative: number;
 	static_fingerprint: string;
+	dynamic_model_ids: string | null;
+	dynamic_model_provenance: string | null;
 	models: string;
 }
 
@@ -33,6 +35,9 @@ interface CacheEntry<TApi extends Api = Api> {
 	 * match — the cache already incorporates the same static state.
 	 */
 	staticFingerprint: string;
+	/** IDs returned by the authoritative dynamic provider catalog, when retained. */
+	dynamicModelIds: string[] | undefined;
+	dynamicModelProvenance: string | undefined;
 }
 
 let sharedDb: Database | null = null;
@@ -56,6 +61,8 @@ function getDb(dbPath?: string): Database {
 			updated_at INTEGER NOT NULL,
 			authoritative INTEGER NOT NULL DEFAULT 0,
 			static_fingerprint TEXT NOT NULL DEFAULT '',
+			dynamic_model_ids TEXT,
+			dynamic_model_provenance TEXT,
 			models TEXT NOT NULL
 		)
 	`);
@@ -81,7 +88,13 @@ function migrateCacheSchema(db: Database): void {
 	if (!columns.some(column => column.name === "static_fingerprint")) {
 		db.run("ALTER TABLE model_cache ADD COLUMN static_fingerprint TEXT NOT NULL DEFAULT ''");
 	}
-	db.run("UPDATE model_cache SET version = ? WHERE version = 2", [CACHE_SCHEMA_VERSION]);
+	if (!columns.some(column => column.name === "dynamic_model_ids")) {
+		db.run("ALTER TABLE model_cache ADD COLUMN dynamic_model_ids TEXT");
+	}
+	if (!columns.some(column => column.name === "dynamic_model_provenance")) {
+		db.run("ALTER TABLE model_cache ADD COLUMN dynamic_model_provenance TEXT");
+	}
+	db.run("UPDATE model_cache SET version = ? WHERE version IN (2, 3, 4)", [CACHE_SCHEMA_VERSION]);
 }
 
 export function readModelCache<TApi extends Api>(
@@ -105,6 +118,8 @@ export function readModelCache<TApi extends Api>(
 			authoritative: row.authoritative === 1,
 			updatedAt: row.updated_at,
 			staticFingerprint: row.static_fingerprint ?? "",
+			dynamicModelIds: row.dynamic_model_ids === null ? undefined : (JSON.parse(row.dynamic_model_ids) as string[]),
+			dynamicModelProvenance: row.dynamic_model_provenance ?? undefined,
 		};
 	} catch {
 		return null;
@@ -118,18 +133,22 @@ export function writeModelCache<TApi extends Api>(
 	authoritative: boolean,
 	staticFingerprint: string,
 	dbPath?: string,
+	dynamicModelIds?: readonly string[],
+	dynamicModelProvenance?: string,
 ): void {
 	try {
 		const db = getDb(dbPath);
 		db.run(
-			`INSERT OR REPLACE INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, models)
-			 VALUES (?, ?, ?, ?, ?, ?)`,
+			`INSERT OR REPLACE INTO model_cache (provider_id, version, updated_at, authoritative, static_fingerprint, dynamic_model_ids, dynamic_model_provenance, models)
+			 VALUES (?, ?, ?, ?, ?, ?, ?, ?)`,
 			[
 				providerId,
 				CACHE_SCHEMA_VERSION,
 				updatedAt,
 				authoritative ? 1 : 0,
 				staticFingerprint,
+				dynamicModelIds === undefined ? null : JSON.stringify(dynamicModelIds),
+				dynamicModelProvenance ?? null,
 				JSON.stringify(models),
 			],
 		);

package/src/model-manager.ts

@@ -43,6 +43,8 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
 	now?: () => number;
 	/** Optional guard that must permit cache publication. Default: writes are permitted. */
 	canPublishCache?: () => boolean;
+	/** Credential-and-endpoint identity required to reuse dynamic catalog IDs. */
+	cacheDynamicModelProvenance?: string;
 }
 
 /**
@@ -56,8 +58,17 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
 export interface ModelResolutionResult<TApi extends Api = Api> {
 	models: Model<TApi>[];
 	stale: boolean;
+	/** Whether the cache row consulted for this resolution was still within its TTL. */
+	cacheFresh: boolean;
+	/** Whether the consulted cache row was authoritative. */
+	cacheAuthoritative: boolean;
 	/** Whether this resolution successfully fetched dynamic models. */
 	fetched: boolean;
+	/**
+	 * IDs returned by a current authoritative dynamic provider catalog. This is
+	 * deliberately distinct from `models`, which merges static and cached data.
+	 */
+	dynamicModelIds?: readonly string[];
 }
 
 /**
@@ -125,14 +136,19 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	const cache = readModelCache<TApi>(options.providerId, ttlMs, now, dbPath);
 	const dynamicFetcher = options.fetchDynamicModels;
 	const hasDynamicFetcher = typeof dynamicFetcher === "function";
-	const hasAuthoritativeCache = (cache?.authoritative ?? false) || !hasDynamicFetcher;
+	const cacheDynamicModelIdsCurrent =
+		cache?.dynamicModelIds !== undefined &&
+		cache.dynamicModelProvenance !== undefined &&
+		cache.dynamicModelProvenance === options.cacheDynamicModelProvenance;
+	const cacheProvenanceMismatch = cache?.dynamicModelIds !== undefined && !cacheDynamicModelIdsCurrent;
+	const hasAuthoritativeCache =
+		!hasDynamicFetcher ||
+		((cache?.authoritative ?? false) && (cache?.dynamicModelIds === undefined || cacheDynamicModelIdsCurrent));
 	const cacheAgeMs = cache ? now() - cache.updatedAt : Number.POSITIVE_INFINITY;
-	const shouldFetchFromNetwork = shouldFetchRemoteSources(
-		strategy,
-		cache?.fresh ?? false,
-		hasAuthoritativeCache,
-		cacheAgeMs,
-	);
+	const shouldFetchFromNetwork =
+		cacheProvenanceMismatch && strategy !== "offline"
+			? true
+			: shouldFetchRemoteSources(strategy, cache?.fresh ?? false, hasAuthoritativeCache, cacheAgeMs);
 	const staticFingerprint = fingerprintStatic(staticModels);
 
 	// Cold-start fast path: when a fresh, authoritative cache exists, the network
@@ -149,13 +165,38 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	) {
 		const cachedModels = passModelList<TApi>(cache.models);
 		if (!hasStaticTransportDrift(staticModels, cachedModels)) {
-			return { models: cachedModels, stale: false, fetched: false };
+			return {
+				models: cachedModels,
+				stale: false,
+				cacheFresh: true,
+				cacheAuthoritative: true,
+				fetched: false,
+				dynamicModelIds:
+					strategy === "online-if-uncached" && cacheDynamicModelIdsCurrent ? cache.dynamicModelIds : undefined,
+			};
 		}
 		const repairedModels = mergeDynamicModels(staticModels, cachedModels);
 		if (options.canPublishCache?.() ?? true) {
-			writeModelCache(options.providerId, now(), repairedModels, true, staticFingerprint, dbPath);
+			writeModelCache(
+				options.providerId,
+				now(),
+				repairedModels,
+				true,
+				staticFingerprint,
+				dbPath,
+				cache.dynamicModelIds,
+				cache.dynamicModelProvenance,
+			);
 		}
-		return { models: repairedModels, stale: false, fetched: false };
+		return {
+			models: repairedModels,
+			stale: false,
+			cacheFresh: true,
+			cacheAuthoritative: true,
+			fetched: false,
+			dynamicModelIds:
+				strategy === "online-if-uncached" && cacheDynamicModelIdsCurrent ? cache.dynamicModelIds : undefined,
+		};
 	}
 
 	const [fetchedModelsDevModels, fetchedDynamicModels] = shouldFetchFromNetwork
@@ -176,7 +217,16 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 				mergeDynamicModels(mergeModelSources(staticModels, modelsDevModels), dynamicModels),
 			);
 			if (options.canPublishCache?.() ?? true) {
-				writeModelCache(options.providerId, now(), snapshotModels, true, staticFingerprint, dbPath);
+				writeModelCache(
+					options.providerId,
+					now(),
+					snapshotModels,
+					true,
+					staticFingerprint,
+					dbPath,
+					dynamicModels.map(model => model.id),
+					options.cacheDynamicModelProvenance,
+				);
 			}
 		} else {
 			// Dynamic fetch failed — update cache with a non-authoritative snapshot so
@@ -202,7 +252,16 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	return {
 		models,
 		stale: !dynamicAuthoritative,
+		cacheFresh: cache?.fresh ?? false,
+		cacheAuthoritative: cache?.authoritative ?? false,
 		fetched: shouldFetchFromNetwork && dynamicFetchSucceeded,
+		dynamicModelIds: dynamicFetchSucceeded
+			? dynamicModels.map(model => model.id)
+			: shouldUseFreshCacheAsAuthoritative
+				? cacheDynamicModelIdsCurrent
+					? cache?.dynamicModelIds
+					: undefined
+				: undefined,
 	};
 }
 

package/src/model-thinking.ts

@@ -53,11 +53,16 @@ const DEFAULT_REASONING_EFFORTS_WITH_XHIGH_AND_MAX: readonly Effort[] = [
 ];
 const GEMINI_3_PRO_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High];
 const GEMINI_3_FLASH_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High];
+// Gemini 3.7 Flash dropped `minimal`; the official API returns an error for it.
+// https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash
+const GEMINI_3_7_FLASH_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
 const GPT_5_2_PLUS_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
 const GPT_5_6_PLUS_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh, Effort.Max];
 const GPT_5_5_DEFAULT_EFFORT = Effort.XHigh;
 const KIMI_K3_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
 const DEEPSEEK_V4_FLASH_0731_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High, Effort.Max];
+const GROK_4_5_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
+const GROK_4_6_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High, Effort.XHigh];
 
 const GPT_5_1_CODEX_MINI_EFFORTS: readonly Effort[] = [Effort.Medium, Effort.High];
 const CLOUDFLARE_AI_GATEWAY_BASE_URL = "https://gateway.ai.cloudflare.com/v1/<account>/<gateway>/anthropic";
@@ -206,10 +211,16 @@ export function refreshModelThinking<TApi extends Api>(model: ApiModel<TApi>): A
 export function applyGeneratedModelPolicies(models: ApiModel<Api>[]): void {
 	for (let index = 0; index < models.length; index++) {
 		const source = models[index]!;
+		if (source.provider === "xai" && (source.id === "grok-4.5" || source.id === "grok-4.6")) {
+			source.reasoning = true;
+		}
 		if (source.provider === "alibaba-token-plan" && source.id === "deepseek-v4-flash-0731") {
 			source.reasoning = true;
 			source.name = "DeepSeek V4 Flash 0731";
 		}
+		if (source.id.split("/").at(-1)?.toLowerCase() === "muse-spark-1.2") {
+			source.reasoning = true;
+		}
 		const model = refreshModelThinking(source);
 		applyGeneratedModelPolicy(model);
 		models[index] = model;
@@ -460,6 +471,17 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	if (model.provider === "zai" && model.id === "glm-5.2") {
 		model.contextWindow = 1_000_000;
 	}
+	// GLM-5.3 always thinks and exposes only low/high/max reasoning_effort.
+	// https://z.ai/blog/glm-5.3#api-changes-in-glm-5-3
+	if (model.provider === "zai" && model.id === "glm-5.3") {
+		model.thinking = {
+			mode: "effort",
+			minLevel: Effort.Low,
+			maxLevel: Effort.Max,
+			defaultLevel: Effort.Max,
+			levels: [Effort.Low, Effort.High, Effort.Max],
+		};
+	}
 	if (model.provider === "alibaba-token-plan" && model.id === "deepseek-v4-flash-0731") {
 		model.contextWindow = 1_000_000;
 		model.maxTokens = 384_000;
@@ -471,6 +493,9 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 			requiresReasoningContentForToolCalls: true,
 		};
 	}
+	if (model.provider === "xai" && (model.id === "grok-4.5" || model.id === "grok-4.6")) {
+		model.maxTokens = Math.min(model.maxTokens, 64_000);
+	}
 	// MiniMax-M3's official Token Plan routes expose a 1M context window.
 	// Scope the correction to the four first-class regional MiniMax routes
 	// (canonical id plus the Anthropic Token Plan `[1m]` id); unrelated
@@ -664,6 +689,12 @@ function expandEffortRange(thinking: ThinkingConfig): readonly Effort[] {
 }
 
 function inferSupportedEfforts<TApi extends Api>(parsedModel: ParsedModel, model: ApiModel<TApi>): readonly Effort[] {
+	if (model.provider === "xai" && model.id === "grok-4.5") {
+		return GROK_4_5_EFFORTS;
+	}
+	if (model.provider === "xai" && model.id === "grok-4.6") {
+		return GROK_4_6_EFFORTS;
+	}
 	if (model.provider === "kimi-code" && model.id === "k3") {
 		return KIMI_K3_EFFORTS;
 	}
@@ -699,7 +730,13 @@ function inferGeminiSupportedEfforts(model: GeminiModel): readonly Effort[] {
 	if (!semverGte(model.version, "3.0")) {
 		return DEFAULT_REASONING_EFFORTS;
 	}
-	return model.kind === "pro" ? GEMINI_3_PRO_EFFORTS : GEMINI_3_FLASH_EFFORTS;
+	if (model.kind === "pro") {
+		return GEMINI_3_PRO_EFFORTS;
+	}
+	if (semverGte(model.version, "3.7")) {
+		return GEMINI_3_7_FLASH_EFFORTS;
+	}
+	return GEMINI_3_FLASH_EFFORTS;
 }
 
 function inferAnthropicSupportedEfforts<TApi extends Api>(
@@ -730,6 +767,13 @@ function inferAnthropicSupportedEfforts<TApi extends Api>(
 }
 
 function inferFallbackEfforts<TApi extends Api>(model: ApiModel<TApi>): readonly Effort[] {
+	// Meta documents Muse Spark 1.2 as accepting the full minimal..xhigh
+	// reasoning range. Keep that capability provider-independent so runtime
+	// model discovery/merge cannot downgrade the bundled OpenRouter entry to
+	// the generic openai-completions ceiling of `high`.
+	if (model.id.split("/").at(-1)?.toLowerCase() === "muse-spark-1.2") {
+		return DEFAULT_REASONING_EFFORTS_WITH_XHIGH;
+	}
 	if (model.api === "anthropic-messages") {
 		return DEFAULT_REASONING_EFFORTS_WITH_XHIGH;
 	}