@tanstack/ai-gemini 0.20.0 → 0.21.0

package/dist/esm/usage.js

@@ -1,94 +1,88 @@
 import { buildBaseUsage } from "@tanstack/ai";
+//#region src/usage.ts
+/**
+* Flattens Gemini's ModalityTokenCount array into individual token fields.
+* Extracts TEXT, IMAGE, AUDIO, VIDEO, DOCUMENT modality counts into a
+* normalized structure.
+*/
 function flattenModalityTokenCounts(modalities) {
-  if (!modalities || modalities.length === 0) {
-    return {};
-  }
-  const result = {};
-  for (const item of modalities) {
-    if (!item.modality || item.tokenCount === void 0) {
-      continue;
-    }
-    const modality = item.modality.toUpperCase();
-    const count = item.tokenCount;
-    switch (modality) {
-      case "TEXT":
-        result.textTokens = (result.textTokens ?? 0) + count;
-        break;
-      case "IMAGE":
-        result.imageTokens = (result.imageTokens ?? 0) + count;
-        break;
-      case "AUDIO":
-        result.audioTokens = (result.audioTokens ?? 0) + count;
-        break;
-      case "VIDEO":
-        result.videoTokens = (result.videoTokens ?? 0) + count;
-        break;
-      case "DOCUMENT":
-        result.documentTokens = (result.documentTokens ?? 0) + count;
-        break;
-    }
-  }
-  return result;
+	if (!modalities || modalities.length === 0) return {};
+	const result = {};
+	for (const item of modalities) {
+		if (!item.modality || item.tokenCount === void 0) continue;
+		const modality = item.modality.toUpperCase();
+		const count = item.tokenCount;
+		switch (modality) {
+			case "TEXT":
+				result.textTokens = (result.textTokens ?? 0) + count;
+				break;
+			case "IMAGE":
+				result.imageTokens = (result.imageTokens ?? 0) + count;
+				break;
+			case "AUDIO":
+				result.audioTokens = (result.audioTokens ?? 0) + count;
+				break;
+			case "VIDEO":
+				result.videoTokens = (result.videoTokens ?? 0) + count;
+				break;
+			case "DOCUMENT":
+				result.documentTokens = (result.documentTokens ?? 0) + count;
+				break;
+		}
+	}
+	return result;
 }
+/**
+* Checks if a FlattenedModalityTokens object has any values set.
+*/
 function hasModalityTokens(tokens) {
-  return tokens.textTokens !== void 0 || tokens.imageTokens !== void 0 || tokens.audioTokens !== void 0 || tokens.videoTokens !== void 0 || tokens.documentTokens !== void 0;
+	return tokens.textTokens !== void 0 || tokens.imageTokens !== void 0 || tokens.audioTokens !== void 0 || tokens.videoTokens !== void 0 || tokens.documentTokens !== void 0;
 }
+/**
+* Build normalized TokenUsage from Gemini's usageMetadata.
+* Handles modality breakdowns and thinking tokens. Returns `undefined` when the
+* provider reported no usage metadata, so callers omit the field rather than
+* fabricating zeroed totals.
+*/
 function buildGeminiUsage(usageMetadata) {
-  if (!usageMetadata) return void 0;
-  const promptTokens = usageMetadata.promptTokenCount ?? 0;
-  const completionTokens = usageMetadata.candidatesTokenCount ?? 0;
-  const result = buildBaseUsage({
-    promptTokens,
-    completionTokens,
-    totalTokens: usageMetadata.totalTokenCount ?? promptTokens + completionTokens
-  });
-  const promptModalities = flattenModalityTokenCounts(
-    usageMetadata.promptTokensDetails
-  );
-  const cachedTokens = usageMetadata.cachedContentTokenCount;
-  const promptTokensDetails = {
-    ...hasModalityTokens(promptModalities) ? promptModalities : {},
-    ...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {}
-  };
-  const completionModalities = flattenModalityTokenCounts(
-    usageMetadata.candidatesTokensDetails
-  );
-  const thoughtsTokens = usageMetadata.thoughtsTokenCount;
-  const completionTokensDetails = {
-    ...hasModalityTokens(completionModalities) ? completionModalities : {},
-    // Map thoughtsTokenCount to reasoningTokens for consistency with OpenAI
-    ...thoughtsTokens !== void 0 && thoughtsTokens > 0 ? { reasoningTokens: thoughtsTokens } : {}
-  };
-  const providerDetails = {
-    ...usageMetadata.trafficType ? { trafficType: usageMetadata.trafficType } : {},
-    ...usageMetadata.toolUsePromptTokenCount !== void 0 && usageMetadata.toolUsePromptTokenCount > 0 ? { toolUsePromptTokenCount: usageMetadata.toolUsePromptTokenCount } : {},
-    ...usageMetadata.toolUsePromptTokensDetails && usageMetadata.toolUsePromptTokensDetails.length > 0 ? {
-      toolUsePromptTokensDetails: usageMetadata.toolUsePromptTokensDetails.map((item) => ({
-        modality: item.modality || "UNKNOWN",
-        tokenCount: item.tokenCount ?? 0
-      }))
-    } : {},
-    ...usageMetadata.cacheTokensDetails && usageMetadata.cacheTokensDetails.length > 0 ? {
-      cacheTokensDetails: usageMetadata.cacheTokensDetails.map((item) => ({
-        modality: item.modality || "UNKNOWN",
-        tokenCount: item.tokenCount ?? 0
-      }))
-    } : {}
-  };
-  if (Object.keys(promptTokensDetails).length > 0) {
-    result.promptTokensDetails = promptTokensDetails;
-  }
-  if (Object.keys(providerDetails).length > 0) {
-    result.providerUsageDetails = providerDetails;
-  }
-  if (Object.keys(completionTokensDetails).length > 0) {
-    result.completionTokensDetails = completionTokensDetails;
-  }
-  return result;
+	if (!usageMetadata) return void 0;
+	const promptTokens = usageMetadata.promptTokenCount ?? 0;
+	const completionTokens = usageMetadata.candidatesTokenCount ?? 0;
+	const result = buildBaseUsage({
+		promptTokens,
+		completionTokens,
+		totalTokens: usageMetadata.totalTokenCount ?? promptTokens + completionTokens
+	});
+	const promptModalities = flattenModalityTokenCounts(usageMetadata.promptTokensDetails);
+	const cachedTokens = usageMetadata.cachedContentTokenCount;
+	const promptTokensDetails = {
+		...hasModalityTokens(promptModalities) ? promptModalities : {},
+		...cachedTokens !== void 0 && cachedTokens > 0 ? { cachedTokens } : {}
+	};
+	const completionModalities = flattenModalityTokenCounts(usageMetadata.candidatesTokensDetails);
+	const thoughtsTokens = usageMetadata.thoughtsTokenCount;
+	const completionTokensDetails = {
+		...hasModalityTokens(completionModalities) ? completionModalities : {},
+		...thoughtsTokens !== void 0 && thoughtsTokens > 0 ? { reasoningTokens: thoughtsTokens } : {}
+	};
+	const providerDetails = {
+		...usageMetadata.trafficType ? { trafficType: usageMetadata.trafficType } : {},
+		...usageMetadata.toolUsePromptTokenCount !== void 0 && usageMetadata.toolUsePromptTokenCount > 0 ? { toolUsePromptTokenCount: usageMetadata.toolUsePromptTokenCount } : {},
+		...usageMetadata.toolUsePromptTokensDetails && usageMetadata.toolUsePromptTokensDetails.length > 0 ? { toolUsePromptTokensDetails: usageMetadata.toolUsePromptTokensDetails.map((item) => ({
+			modality: item.modality || "UNKNOWN",
+			tokenCount: item.tokenCount ?? 0
+		})) } : {},
+		...usageMetadata.cacheTokensDetails && usageMetadata.cacheTokensDetails.length > 0 ? { cacheTokensDetails: usageMetadata.cacheTokensDetails.map((item) => ({
+			modality: item.modality || "UNKNOWN",
+			tokenCount: item.tokenCount ?? 0
+		})) } : {}
+	};
+	if (Object.keys(promptTokensDetails).length > 0) result.promptTokensDetails = promptTokensDetails;
+	if (Object.keys(providerDetails).length > 0) result.providerUsageDetails = providerDetails;
+	if (Object.keys(completionTokensDetails).length > 0) result.completionTokensDetails = completionTokensDetails;
+	return result;
 }
-export {
-  buildGeminiUsage,
-  flattenModalityTokenCounts,
-  hasModalityTokens
-};
-//# sourceMappingURL=usage.js.map
+//#endregion
+export { buildGeminiUsage, flattenModalityTokenCounts, hasModalityTokens };
+
+//# sourceMappingURL=usage.js.map

package/dist/esm/usage.js.map

@@ -1 +1 @@
-{"version":3,"file":"usage.js","sources":["../../src/usage.ts"],"sourcesContent":["import { buildBaseUsage } from '@tanstack/ai'\nimport type { TokenUsage } from '@tanstack/ai'\nimport type {\n  GenerateContentResponseUsageMetadata,\n  ModalityTokenCount,\n} from '@google/genai'\n\n/**\n * Flattened modality token counts for normalized usage reporting.\n * Maps Gemini's ModalityTokenCount array to individual fields.\n */\nexport interface FlattenedModalityTokens {\n  /** Text tokens */\n  textTokens?: number\n  /** Image tokens */\n  imageTokens?: number\n  /** Audio tokens */\n  audioTokens?: number\n  /** Video tokens */\n  videoTokens?: number\n  /** Document tokens (e.g. PDF inputs) */\n  documentTokens?: number\n}\n\n/**\n * Flattens Gemini's ModalityTokenCount array into individual token fields.\n * Extracts TEXT, IMAGE, AUDIO, VIDEO, DOCUMENT modality counts into a\n * normalized structure.\n */\nexport function flattenModalityTokenCounts(\n  modalities?: Array<ModalityTokenCount>,\n): FlattenedModalityTokens {\n  if (!modalities || modalities.length === 0) {\n    return {}\n  }\n\n  const result: FlattenedModalityTokens = {}\n\n  for (const item of modalities) {\n    if (!item.modality || item.tokenCount === undefined) {\n      continue\n    }\n\n    const modality = item.modality.toUpperCase()\n    const count = item.tokenCount\n\n    switch (modality) {\n      case 'TEXT':\n        result.textTokens = (result.textTokens ?? 0) + count\n        break\n      case 'IMAGE':\n        result.imageTokens = (result.imageTokens ?? 0) + count\n        break\n      case 'AUDIO':\n        result.audioTokens = (result.audioTokens ?? 0) + count\n        break\n      case 'VIDEO':\n        result.videoTokens = (result.videoTokens ?? 0) + count\n        break\n      case 'DOCUMENT':\n        result.documentTokens = (result.documentTokens ?? 0) + count\n        break\n    }\n  }\n\n  return result\n}\n\n/**\n * Checks if a FlattenedModalityTokens object has any values set.\n */\nexport function hasModalityTokens(tokens: FlattenedModalityTokens): boolean {\n  return (\n    tokens.textTokens !== undefined ||\n    tokens.imageTokens !== undefined ||\n    tokens.audioTokens !== undefined ||\n    tokens.videoTokens !== undefined ||\n    tokens.documentTokens !== undefined\n  )\n}\n\n/**\n * Gemini-specific provider usage details.\n * These fields are unique to Gemini and placed in providerUsageDetails.\n */\nexport type GeminiProviderUsageDetails = {\n  /**\n   * The traffic type for this request.\n   * Can indicate whether request was handled by different service tiers.\n   */\n  trafficType?: string\n  /**\n   * Number of tokens in the results from tool executions,\n   * which are provided back to the model as input.\n   */\n  toolUsePromptTokenCount?: number\n  /**\n   * Detailed breakdown by modality of the token counts from\n   * the results of tool executions.\n   */\n  toolUsePromptTokensDetails?: Array<{\n    modality: string\n    tokenCount: number\n  }>\n  /**\n   * Detailed breakdown of cache tokens by modality.\n   * More granular than the normalized cachedTokens field.\n   */\n  cacheTokensDetails?: Array<{\n    modality: string\n    tokenCount: number\n  }>\n}\n\n/**\n * Build normalized TokenUsage from Gemini's usageMetadata.\n * Handles modality breakdowns and thinking tokens. Returns `undefined` when the\n * provider reported no usage metadata, so callers omit the field rather than\n * fabricating zeroed totals.\n */\nexport function buildGeminiUsage(\n  usageMetadata: GenerateContentResponseUsageMetadata | undefined | null,\n): TokenUsage<GeminiProviderUsageDetails> | undefined {\n  if (!usageMetadata) return undefined\n\n  const promptTokens = usageMetadata.promptTokenCount ?? 0\n  const completionTokens = usageMetadata.candidatesTokenCount ?? 0\n\n  const result = buildBaseUsage<GeminiProviderUsageDetails>({\n    promptTokens: promptTokens,\n    completionTokens: completionTokens,\n    totalTokens:\n      usageMetadata.totalTokenCount ?? promptTokens + completionTokens,\n  })\n\n  // Add prompt token details\n  // Flatten modality breakdown for prompt\n  const promptModalities = flattenModalityTokenCounts(\n    usageMetadata.promptTokensDetails,\n  )\n  const cachedTokens = usageMetadata.cachedContentTokenCount\n\n  const promptTokensDetails = {\n    ...(hasModalityTokens(promptModalities) ? promptModalities : {}),\n    ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n  }\n\n  // Add completion token details\n  // Flatten modality breakdown for candidates (output)\n  const completionModalities = flattenModalityTokenCounts(\n    usageMetadata.candidatesTokensDetails,\n  )\n  const thoughtsTokens = usageMetadata.thoughtsTokenCount\n\n  const completionTokensDetails = {\n    ...(hasModalityTokens(completionModalities) ? completionModalities : {}),\n    // Map thoughtsTokenCount to reasoningTokens for consistency with OpenAI\n    ...(thoughtsTokens !== undefined && thoughtsTokens > 0\n      ? { reasoningTokens: thoughtsTokens }\n      : {}),\n  }\n\n  // Add provider-specific details\n  const providerDetails: GeminiProviderUsageDetails = {\n    ...(usageMetadata.trafficType\n      ? { trafficType: usageMetadata.trafficType }\n      : {}),\n    ...(usageMetadata.toolUsePromptTokenCount !== undefined &&\n    usageMetadata.toolUsePromptTokenCount > 0\n      ? { toolUsePromptTokenCount: usageMetadata.toolUsePromptTokenCount }\n      : {}),\n    ...(usageMetadata.toolUsePromptTokensDetails &&\n    usageMetadata.toolUsePromptTokensDetails.length > 0\n      ? {\n          toolUsePromptTokensDetails:\n            usageMetadata.toolUsePromptTokensDetails.map((item) => ({\n              modality: item.modality || 'UNKNOWN',\n              tokenCount: item.tokenCount ?? 0,\n            })),\n        }\n      : {}),\n    ...(usageMetadata.cacheTokensDetails &&\n    usageMetadata.cacheTokensDetails.length > 0\n      ? {\n          cacheTokensDetails: usageMetadata.cacheTokensDetails.map((item) => ({\n            modality: item.modality || 'UNKNOWN',\n            tokenCount: item.tokenCount ?? 0,\n          })),\n        }\n      : {}),\n  }\n\n  // Add prompt token details if available\n  if (Object.keys(promptTokensDetails).length > 0) {\n    result.promptTokensDetails = promptTokensDetails\n  }\n  // Add provider details if available\n  if (Object.keys(providerDetails).length > 0) {\n    result.providerUsageDetails = providerDetails\n  }\n  // Add completion token details if available\n  if (Object.keys(completionTokensDetails).length > 0) {\n    result.completionTokensDetails = completionTokensDetails\n  }\n\n  return result\n}\n"],"names":[],"mappings":";AA6BO,SAAS,2BACd,YACyB;AACzB,MAAI,CAAC,cAAc,WAAW,WAAW,GAAG;AAC1C,WAAO,CAAA;AAAA,EACT;AAEA,QAAM,SAAkC,CAAA;AAExC,aAAW,QAAQ,YAAY;AAC7B,QAAI,CAAC,KAAK,YAAY,KAAK,eAAe,QAAW;AACnD;AAAA,IACF;AAEA,UAAM,WAAW,KAAK,SAAS,YAAA;AAC/B,UAAM,QAAQ,KAAK;AAEnB,YAAQ,UAAA;AAAA,MACN,KAAK;AACH,eAAO,cAAc,OAAO,cAAc,KAAK;AAC/C;AAAA,MACF,KAAK;AACH,eAAO,eAAe,OAAO,eAAe,KAAK;AACjD;AAAA,MACF,KAAK;AACH,eAAO,eAAe,OAAO,eAAe,KAAK;AACjD;AAAA,MACF,KAAK;AACH,eAAO,eAAe,OAAO,eAAe,KAAK;AACjD;AAAA,MACF,KAAK;AACH,eAAO,kBAAkB,OAAO,kBAAkB,KAAK;AACvD;AAAA,IAAA;AAAA,EAEN;AAEA,SAAO;AACT;AAKO,SAAS,kBAAkB,QAA0C;AAC1E,SACE,OAAO,eAAe,UACtB,OAAO,gBAAgB,UACvB,OAAO,gBAAgB,UACvB,OAAO,gBAAgB,UACvB,OAAO,mBAAmB;AAE9B;AAyCO,SAAS,iBACd,eACoD;AACpD,MAAI,CAAC,cAAe,QAAO;AAE3B,QAAM,eAAe,cAAc,oBAAoB;AACvD,QAAM,mBAAmB,cAAc,wBAAwB;AAE/D,QAAM,SAAS,eAA2C;AAAA,IACxD;AAAA,IACA;AAAA,IACA,aACE,cAAc,mBAAmB,eAAe;AAAA,EAAA,CACnD;AAID,QAAM,mBAAmB;AAAA,IACvB,cAAc;AAAA,EAAA;AAEhB,QAAM,eAAe,cAAc;AAEnC,QAAM,sBAAsB;AAAA,IAC1B,GAAI,kBAAkB,gBAAgB,IAAI,mBAAmB,CAAA;AAAA,IAC7D,GAAI,iBAAiB,UAAa,eAAe,IAAI,EAAE,aAAA,IAAiB,CAAA;AAAA,EAAC;AAK3E,QAAM,uBAAuB;AAAA,IAC3B,cAAc;AAAA,EAAA;AAEhB,QAAM,iBAAiB,cAAc;AAErC,QAAM,0BAA0B;AAAA,IAC9B,GAAI,kBAAkB,oBAAoB,IAAI,uBAAuB,CAAA;AAAA;AAAA,IAErE,GAAI,mBAAmB,UAAa,iBAAiB,IACjD,EAAE,iBAAiB,mBACnB,CAAA;AAAA,EAAC;AAIP,QAAM,kBAA8C;AAAA,IAClD,GAAI,cAAc,cACd,EAAE,aAAa,cAAc,YAAA,IAC7B,CAAA;AAAA,IACJ,GAAI,cAAc,4BAA4B,UAC9C,cAAc,0BAA0B,IACpC,EAAE,yBAAyB,cAAc,wBAAA,IACzC,CAAA;AAAA,IACJ,GAAI,cAAc,8BAClB,cAAc,2BAA2B,SAAS,IAC9C;AAAA,MACE,4BACE,cAAc,2BAA2B,IAAI,CAAC,UAAU;AAAA,QACtD,UAAU,KAAK,YAAY;AAAA,QAC3B,YAAY,KAAK,cAAc;AAAA,MAAA,EAC/B;AAAA,IAAA,IAEN,CAAA;AAAA,IACJ,GAAI,cAAc,sBAClB,cAAc,mBAAmB,SAAS,IACtC;AAAA,MACE,oBAAoB,cAAc,mBAAmB,IAAI,CAAC,UAAU;AAAA,QAClE,UAAU,KAAK,YAAY;AAAA,QAC3B,YAAY,KAAK,cAAc;AAAA,MAAA,EAC/B;AAAA,IAAA,IAEJ,CAAA;AAAA,EAAC;AAIP,MAAI,OAAO,KAAK,mBAAmB,EAAE,SAAS,GAAG;AAC/C,WAAO,sBAAsB;AAAA,EAC/B;AAEA,MAAI,OAAO,KAAK,eAAe,EAAE,SAAS,GAAG;AAC3C,WAAO,uBAAuB;AAAA,EAChC;AAEA,MAAI,OAAO,KAAK,uBAAuB,EAAE,SAAS,GAAG;AACnD,WAAO,0BAA0B;AAAA,EACnC;AAEA,SAAO;AACT;"}
+{"version":3,"file":"usage.js","names":[],"sources":["../../src/usage.ts"],"sourcesContent":["import { buildBaseUsage } from '@tanstack/ai'\nimport type { TokenUsage } from '@tanstack/ai'\nimport type {\n  GenerateContentResponseUsageMetadata,\n  ModalityTokenCount,\n} from '@google/genai'\n\n/**\n * Flattened modality token counts for normalized usage reporting.\n * Maps Gemini's ModalityTokenCount array to individual fields.\n */\nexport interface FlattenedModalityTokens {\n  /** Text tokens */\n  textTokens?: number\n  /** Image tokens */\n  imageTokens?: number\n  /** Audio tokens */\n  audioTokens?: number\n  /** Video tokens */\n  videoTokens?: number\n  /** Document tokens (e.g. PDF inputs) */\n  documentTokens?: number\n}\n\n/**\n * Flattens Gemini's ModalityTokenCount array into individual token fields.\n * Extracts TEXT, IMAGE, AUDIO, VIDEO, DOCUMENT modality counts into a\n * normalized structure.\n */\nexport function flattenModalityTokenCounts(\n  modalities?: Array<ModalityTokenCount>,\n): FlattenedModalityTokens {\n  if (!modalities || modalities.length === 0) {\n    return {}\n  }\n\n  const result: FlattenedModalityTokens = {}\n\n  for (const item of modalities) {\n    if (!item.modality || item.tokenCount === undefined) {\n      continue\n    }\n\n    const modality = item.modality.toUpperCase()\n    const count = item.tokenCount\n\n    switch (modality) {\n      case 'TEXT':\n        result.textTokens = (result.textTokens ?? 0) + count\n        break\n      case 'IMAGE':\n        result.imageTokens = (result.imageTokens ?? 0) + count\n        break\n      case 'AUDIO':\n        result.audioTokens = (result.audioTokens ?? 0) + count\n        break\n      case 'VIDEO':\n        result.videoTokens = (result.videoTokens ?? 0) + count\n        break\n      case 'DOCUMENT':\n        result.documentTokens = (result.documentTokens ?? 0) + count\n        break\n    }\n  }\n\n  return result\n}\n\n/**\n * Checks if a FlattenedModalityTokens object has any values set.\n */\nexport function hasModalityTokens(tokens: FlattenedModalityTokens): boolean {\n  return (\n    tokens.textTokens !== undefined ||\n    tokens.imageTokens !== undefined ||\n    tokens.audioTokens !== undefined ||\n    tokens.videoTokens !== undefined ||\n    tokens.documentTokens !== undefined\n  )\n}\n\n/**\n * Gemini-specific provider usage details.\n * These fields are unique to Gemini and placed in providerUsageDetails.\n */\nexport type GeminiProviderUsageDetails = {\n  /**\n   * The traffic type for this request.\n   * Can indicate whether request was handled by different service tiers.\n   */\n  trafficType?: string\n  /**\n   * Number of tokens in the results from tool executions,\n   * which are provided back to the model as input.\n   */\n  toolUsePromptTokenCount?: number\n  /**\n   * Detailed breakdown by modality of the token counts from\n   * the results of tool executions.\n   */\n  toolUsePromptTokensDetails?: Array<{\n    modality: string\n    tokenCount: number\n  }>\n  /**\n   * Detailed breakdown of cache tokens by modality.\n   * More granular than the normalized cachedTokens field.\n   */\n  cacheTokensDetails?: Array<{\n    modality: string\n    tokenCount: number\n  }>\n}\n\n/**\n * Build normalized TokenUsage from Gemini's usageMetadata.\n * Handles modality breakdowns and thinking tokens. Returns `undefined` when the\n * provider reported no usage metadata, so callers omit the field rather than\n * fabricating zeroed totals.\n */\nexport function buildGeminiUsage(\n  usageMetadata: GenerateContentResponseUsageMetadata | undefined | null,\n): TokenUsage<GeminiProviderUsageDetails> | undefined {\n  if (!usageMetadata) return undefined\n\n  const promptTokens = usageMetadata.promptTokenCount ?? 0\n  const completionTokens = usageMetadata.candidatesTokenCount ?? 0\n\n  const result = buildBaseUsage<GeminiProviderUsageDetails>({\n    promptTokens: promptTokens,\n    completionTokens: completionTokens,\n    totalTokens:\n      usageMetadata.totalTokenCount ?? promptTokens + completionTokens,\n  })\n\n  // Add prompt token details\n  // Flatten modality breakdown for prompt\n  const promptModalities = flattenModalityTokenCounts(\n    usageMetadata.promptTokensDetails,\n  )\n  const cachedTokens = usageMetadata.cachedContentTokenCount\n\n  const promptTokensDetails = {\n    ...(hasModalityTokens(promptModalities) ? promptModalities : {}),\n    ...(cachedTokens !== undefined && cachedTokens > 0 ? { cachedTokens } : {}),\n  }\n\n  // Add completion token details\n  // Flatten modality breakdown for candidates (output)\n  const completionModalities = flattenModalityTokenCounts(\n    usageMetadata.candidatesTokensDetails,\n  )\n  const thoughtsTokens = usageMetadata.thoughtsTokenCount\n\n  const completionTokensDetails = {\n    ...(hasModalityTokens(completionModalities) ? completionModalities : {}),\n    // Map thoughtsTokenCount to reasoningTokens for consistency with OpenAI\n    ...(thoughtsTokens !== undefined && thoughtsTokens > 0\n      ? { reasoningTokens: thoughtsTokens }\n      : {}),\n  }\n\n  // Add provider-specific details\n  const providerDetails: GeminiProviderUsageDetails = {\n    ...(usageMetadata.trafficType\n      ? { trafficType: usageMetadata.trafficType }\n      : {}),\n    ...(usageMetadata.toolUsePromptTokenCount !== undefined &&\n    usageMetadata.toolUsePromptTokenCount > 0\n      ? { toolUsePromptTokenCount: usageMetadata.toolUsePromptTokenCount }\n      : {}),\n    ...(usageMetadata.toolUsePromptTokensDetails &&\n    usageMetadata.toolUsePromptTokensDetails.length > 0\n      ? {\n          toolUsePromptTokensDetails:\n            usageMetadata.toolUsePromptTokensDetails.map((item) => ({\n              modality: item.modality || 'UNKNOWN',\n              tokenCount: item.tokenCount ?? 0,\n            })),\n        }\n      : {}),\n    ...(usageMetadata.cacheTokensDetails &&\n    usageMetadata.cacheTokensDetails.length > 0\n      ? {\n          cacheTokensDetails: usageMetadata.cacheTokensDetails.map((item) => ({\n            modality: item.modality || 'UNKNOWN',\n            tokenCount: item.tokenCount ?? 0,\n          })),\n        }\n      : {}),\n  }\n\n  // Add prompt token details if available\n  if (Object.keys(promptTokensDetails).length > 0) {\n    result.promptTokensDetails = promptTokensDetails\n  }\n  // Add provider details if available\n  if (Object.keys(providerDetails).length > 0) {\n    result.providerUsageDetails = providerDetails\n  }\n  // Add completion token details if available\n  if (Object.keys(completionTokensDetails).length > 0) {\n    result.completionTokensDetails = completionTokensDetails\n  }\n\n  return result\n}\n"],"mappings":";;;;;;;AA6BA,SAAgB,2BACd,YACyB;CACzB,IAAI,CAAC,cAAc,WAAW,WAAW,GACvC,OAAO,CAAC;CAGV,MAAM,SAAkC,CAAC;CAEzC,KAAK,MAAM,QAAQ,YAAY;EAC7B,IAAI,CAAC,KAAK,YAAY,KAAK,eAAe,KAAA,GACxC;EAGF,MAAM,WAAW,KAAK,SAAS,YAAY;EAC3C,MAAM,QAAQ,KAAK;EAEnB,QAAQ,UAAR;GACE,KAAK;IACH,OAAO,cAAc,OAAO,cAAc,KAAK;IAC/C;GACF,KAAK;IACH,OAAO,eAAe,OAAO,eAAe,KAAK;IACjD;GACF,KAAK;IACH,OAAO,eAAe,OAAO,eAAe,KAAK;IACjD;GACF,KAAK;IACH,OAAO,eAAe,OAAO,eAAe,KAAK;IACjD;GACF,KAAK;IACH,OAAO,kBAAkB,OAAO,kBAAkB,KAAK;IACvD;EACJ;CACF;CAEA,OAAO;AACT;;;;AAKA,SAAgB,kBAAkB,QAA0C;CAC1E,OACE,OAAO,eAAe,KAAA,KACtB,OAAO,gBAAgB,KAAA,KACvB,OAAO,gBAAgB,KAAA,KACvB,OAAO,gBAAgB,KAAA,KACvB,OAAO,mBAAmB,KAAA;AAE9B;;;;;;;AAyCA,SAAgB,iBACd,eACoD;CACpD,IAAI,CAAC,eAAe,OAAO,KAAA;CAE3B,MAAM,eAAe,cAAc,oBAAoB;CACvD,MAAM,mBAAmB,cAAc,wBAAwB;CAE/D,MAAM,SAAS,eAA2C;EAC1C;EACI;EAClB,aACE,cAAc,mBAAmB,eAAe;CACpD,CAAC;CAID,MAAM,mBAAmB,2BACvB,cAAc,mBAChB;CACA,MAAM,eAAe,cAAc;CAEnC,MAAM,sBAAsB;EAC1B,GAAI,kBAAkB,gBAAgB,IAAI,mBAAmB,CAAC;EAC9D,GAAI,iBAAiB,KAAA,KAAa,eAAe,IAAI,EAAE,aAAa,IAAI,CAAC;CAC3E;CAIA,MAAM,uBAAuB,2BAC3B,cAAc,uBAChB;CACA,MAAM,iBAAiB,cAAc;CAErC,MAAM,0BAA0B;EAC9B,GAAI,kBAAkB,oBAAoB,IAAI,uBAAuB,CAAC;EAEtE,GAAI,mBAAmB,KAAA,KAAa,iBAAiB,IACjD,EAAE,iBAAiB,eAAe,IAClC,CAAC;CACP;CAGA,MAAM,kBAA8C;EAClD,GAAI,cAAc,cACd,EAAE,aAAa,cAAc,YAAY,IACzC,CAAC;EACL,GAAI,cAAc,4BAA4B,KAAA,KAC9C,cAAc,0BAA0B,IACpC,EAAE,yBAAyB,cAAc,wBAAwB,IACjE,CAAC;EACL,GAAI,cAAc,8BAClB,cAAc,2BAA2B,SAAS,IAC9C,EACE,4BACE,cAAc,2BAA2B,KAAK,UAAU;GACtD,UAAU,KAAK,YAAY;GAC3B,YAAY,KAAK,cAAc;EACjC,EAAE,EACN,IACA,CAAC;EACL,GAAI,cAAc,sBAClB,cAAc,mBAAmB,SAAS,IACtC,EACE,oBAAoB,cAAc,mBAAmB,KAAK,UAAU;GAClE,UAAU,KAAK,YAAY;GAC3B,YAAY,KAAK,cAAc;EACjC,EAAE,EACJ,IACA,CAAC;CACP;CAGA,IAAI,OAAO,KAAK,mBAAmB,CAAC,CAAC,SAAS,GAC5C,OAAO,sBAAsB;CAG/B,IAAI,OAAO,KAAK,eAAe,CAAC,CAAC,SAAS,GACxC,OAAO,uBAAuB;CAGhC,IAAI,OAAO,KAAK,uBAAuB,CAAC,CAAC,SAAS,GAChD,OAAO,0BAA0B;CAGnC,OAAO;AACT"}

package/dist/esm/utils/client.js

@@ -1,30 +1,37 @@
 import { GoogleGenAI } from "@google/genai";
-import { generateId as generateId$1, getApiKeyFromEnv } from "@tanstack/ai-utils";
+import { generateId, getApiKeyFromEnv } from "@tanstack/ai-utils";
+//#region src/utils/client.ts
+/**
+* Creates a Google Generative AI client instance
+*/
 function createGeminiClient(config) {
-  return new GoogleGenAI({
-    ...config,
-    apiKey: config.apiKey
-  });
+	return new GoogleGenAI({
+		...config,
+		apiKey: config.apiKey
+	});
 }
+/**
+* Gets Google API key from environment variables
+* @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found
+*/
 function getGeminiApiKeyFromEnv() {
-  try {
-    return getApiKeyFromEnv("GOOGLE_API_KEY");
-  } catch {
-    try {
-      return getApiKeyFromEnv("GEMINI_API_KEY");
-    } catch {
-      throw new Error(
-        "GOOGLE_API_KEY or GEMINI_API_KEY is not set. Please set one of these environment variables or pass the API key directly."
-      );
-    }
-  }
+	try {
+		return getApiKeyFromEnv("GOOGLE_API_KEY");
+	} catch {
+		try {
+			return getApiKeyFromEnv("GEMINI_API_KEY");
+		} catch {
+			throw new Error("GOOGLE_API_KEY or GEMINI_API_KEY is not set. Please set one of these environment variables or pass the API key directly.");
+		}
+	}
 }
-function generateId(prefix) {
-  return generateId$1(prefix);
+/**
+* Generates a unique ID with a prefix
+*/
+function generateId$1(prefix) {
+	return generateId(prefix);
 }
-export {
-  createGeminiClient,
-  generateId,
-  getGeminiApiKeyFromEnv
-};
-//# sourceMappingURL=client.js.map
+//#endregion
+export { createGeminiClient, generateId$1 as generateId, getGeminiApiKeyFromEnv };
+
+//# sourceMappingURL=client.js.map

package/dist/esm/utils/client.js.map

@@ -1 +1 @@
-{"version":3,"file":"client.js","sources":["../../../src/utils/client.ts"],"sourcesContent":["import { GoogleGenAI } from '@google/genai'\nimport { generateId as _generateId, getApiKeyFromEnv } from '@tanstack/ai-utils'\nimport type { GoogleGenAIOptions } from '@google/genai'\n\nexport interface GeminiClientConfig extends GoogleGenAIOptions {\n  apiKey: string\n}\n\n/**\n * Creates a Google Generative AI client instance\n */\nexport function createGeminiClient(config: GeminiClientConfig): GoogleGenAI {\n  return new GoogleGenAI({\n    ...config,\n    apiKey: config.apiKey,\n  })\n}\n\n/**\n * Gets Google API key from environment variables\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found\n */\nexport function getGeminiApiKeyFromEnv(): string {\n  try {\n    return getApiKeyFromEnv('GOOGLE_API_KEY')\n  } catch {\n    try {\n      return getApiKeyFromEnv('GEMINI_API_KEY')\n    } catch {\n      throw new Error(\n        'GOOGLE_API_KEY or GEMINI_API_KEY is not set. Please set one of these environment variables or pass the API key directly.',\n      )\n    }\n  }\n}\n\n/**\n * Generates a unique ID with a prefix\n */\nexport function generateId(prefix: string): string {\n  return _generateId(prefix)\n}\n"],"names":["_generateId"],"mappings":";;AAWO,SAAS,mBAAmB,QAAyC;AAC1E,SAAO,IAAI,YAAY;AAAA,IACrB,GAAG;AAAA,IACH,QAAQ,OAAO;AAAA,EAAA,CAChB;AACH;AAMO,SAAS,yBAAiC;AAC/C,MAAI;AACF,WAAO,iBAAiB,gBAAgB;AAAA,EAC1C,QAAQ;AACN,QAAI;AACF,aAAO,iBAAiB,gBAAgB;AAAA,IAC1C,QAAQ;AACN,YAAM,IAAI;AAAA,QACR;AAAA,MAAA;AAAA,IAEJ;AAAA,EACF;AACF;AAKO,SAAS,WAAW,QAAwB;AACjD,SAAOA,aAAY,MAAM;AAC3B;"}
+{"version":3,"file":"client.js","names":[],"sources":["../../../src/utils/client.ts"],"sourcesContent":["import { GoogleGenAI } from '@google/genai'\nimport { generateId as _generateId, getApiKeyFromEnv } from '@tanstack/ai-utils'\nimport type { GoogleGenAIOptions } from '@google/genai'\n\nexport interface GeminiClientConfig extends GoogleGenAIOptions {\n  apiKey: string\n}\n\n/**\n * Creates a Google Generative AI client instance\n */\nexport function createGeminiClient(config: GeminiClientConfig): GoogleGenAI {\n  return new GoogleGenAI({\n    ...config,\n    apiKey: config.apiKey,\n  })\n}\n\n/**\n * Gets Google API key from environment variables\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found\n */\nexport function getGeminiApiKeyFromEnv(): string {\n  try {\n    return getApiKeyFromEnv('GOOGLE_API_KEY')\n  } catch {\n    try {\n      return getApiKeyFromEnv('GEMINI_API_KEY')\n    } catch {\n      throw new Error(\n        'GOOGLE_API_KEY or GEMINI_API_KEY is not set. Please set one of these environment variables or pass the API key directly.',\n      )\n    }\n  }\n}\n\n/**\n * Generates a unique ID with a prefix\n */\nexport function generateId(prefix: string): string {\n  return _generateId(prefix)\n}\n"],"mappings":";;;;;;AAWA,SAAgB,mBAAmB,QAAyC;CAC1E,OAAO,IAAI,YAAY;EACrB,GAAG;EACH,QAAQ,OAAO;CACjB,CAAC;AACH;;;;;AAMA,SAAgB,yBAAiC;CAC/C,IAAI;EACF,OAAO,iBAAiB,gBAAgB;CAC1C,QAAQ;EACN,IAAI;GACF,OAAO,iBAAiB,gBAAgB;EAC1C,QAAQ;GACN,MAAM,IAAI,MACR,0HACF;EACF;CACF;AACF;;;;AAKA,SAAgB,aAAW,QAAwB;CACjD,OAAO,WAAY,MAAM;AAC3B"}

package/dist/esm/utils/index.js

@@ -0,0 +1,2 @@
+import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "./client.js";
+export { createGeminiClient, generateId, getGeminiApiKeyFromEnv };

package/dist/esm/video/video-provider-options.js

@@ -1,26 +1,79 @@
 import { GEMINI_INTERACTIONS_VIDEO_MODELS } from "../model-meta.js";
+//#region src/video/video-provider-options.ts
+/**
+* Gemini Video Generation Provider Options
+*
+* Covers two request paths behind the one video adapter:
+* - Veo models — long-running operations via `:predictLongRunning`
+*   (https://ai.google.dev/gemini-api/docs/video)
+* - Gemini Omni Flash — background jobs via the Interactions API
+*   (https://ai.google.dev/gemini-api/docs/omni)
+*
+* @experimental Video generation is an experimental feature and may change.
+*/
+/**
+* Runtime guard for the Interactions-served video models.
+* @experimental Omni video generation is an experimental feature and may change.
+*/
 function isInteractionsVideoModel(model) {
-  return GEMINI_INTERACTIONS_VIDEO_MODELS.includes(
-    model
-  );
+	return GEMINI_INTERACTIONS_VIDEO_MODELS.includes(model);
 }
-const GEMINI_VIDEO_DURATIONS = {
-  "veo-3.1-generate-preview": { kind: "discrete", values: [4, 6, 8] },
-  "veo-3.1-fast-generate-preview": { kind: "discrete", values: [4, 6, 8] },
-  "veo-3.1-lite-generate-preview": { kind: "discrete", values: [4, 6, 8] },
-  "gemini-omni-flash-preview": {
-    kind: "range",
-    min: 3,
-    max: 10,
-    unit: "seconds"
-  }
+/**
+* Runtime duration table backing `availableDurations()` / `snapDuration()`.
+*
+* Veo values are curated from the official docs
+* (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec
+* types the `:predictLongRunning` request's `parameters` as unconstrained,
+* so it carries no per-model duration information to derive these from.
+* Omni Flash's 3–10s range was verified against the live API
+* (2026-07-02): `response_format.duration` takes a `"<seconds>s"` string,
+* fractional values are accepted, out-of-range values are rejected with
+* "minimum allowed 3s" / "maximum allowed 10s", and omitting it defaults
+* to a 10-second clip.
+*
+* @experimental Video generation is an experimental feature and may change.
+*/
+var GEMINI_VIDEO_DURATIONS = {
+	"veo-3.1-generate-preview": {
+		kind: "discrete",
+		values: [
+			4,
+			6,
+			8
+		]
+	},
+	"veo-3.1-fast-generate-preview": {
+		kind: "discrete",
+		values: [
+			4,
+			6,
+			8
+		]
+	},
+	"veo-3.1-lite-generate-preview": {
+		kind: "discrete",
+		values: [
+			4,
+			6,
+			8
+		]
+	},
+	"gemini-omni-flash-preview": {
+		kind: "range",
+		min: 3,
+		max: 10,
+		unit: "seconds"
+	}
 };
+/**
+* Look up the duration options for a Gemini video model.
+*
+* @experimental Video generation is an experimental feature and may change.
+*/
 function getGeminiVideoDurationOptions(model) {
-  return GEMINI_VIDEO_DURATIONS[model];
+	return GEMINI_VIDEO_DURATIONS[model];
 }
-export {
-  GEMINI_VIDEO_DURATIONS,
-  getGeminiVideoDurationOptions,
-  isInteractionsVideoModel
-};
-//# sourceMappingURL=video-provider-options.js.map
+//#endregion
+export { GEMINI_VIDEO_DURATIONS, getGeminiVideoDurationOptions, isInteractionsVideoModel };
+
+//# sourceMappingURL=video-provider-options.js.map

package/dist/esm/video/video-provider-options.js.map

@@ -1 +1 @@
-{"version":3,"file":"video-provider-options.js","sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Gemini Video Generation Provider Options\n *\n * Covers two request paths behind the one video adapter:\n * - Veo models — long-running operations via `:predictLongRunning`\n *   (https://ai.google.dev/gemini-api/docs/video)\n * - Gemini Omni Flash — background jobs via the Interactions API\n *   (https://ai.google.dev/gemini-api/docs/omni)\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nimport { GEMINI_INTERACTIONS_VIDEO_MODELS } from '../model-meta'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GenerateVideosConfig, Interactions } from '@google/genai'\nimport type { GEMINI_VIDEO_MODELS } from '../model-meta'\n\n/**\n * Model type for Gemini video generation (Veo + Omni Flash).\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModel = (typeof GEMINI_VIDEO_MODELS)[number]\n\n/**\n * Video models served by the Interactions API (Gemini Omni Flash) rather\n * than Veo's `:predictLongRunning` operations flow.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiInteractionsVideoModel =\n  (typeof GEMINI_INTERACTIONS_VIDEO_MODELS)[number]\n\n/**\n * Runtime guard for the Interactions-served video models.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport function isInteractionsVideoModel(\n  model: GeminiVideoModel,\n): model is GeminiInteractionsVideoModel {\n  return (GEMINI_INTERACTIONS_VIDEO_MODELS as ReadonlyArray<string>).includes(\n    model,\n  )\n}\n\n/**\n * Supported aspect ratios for Gemini video generation. This is the `size`\n * value for the Gemini video adapter — both Veo and Omni Flash express\n * output shape as an aspect ratio (plus an optional `resolution` in Veo's\n * `modelOptions`), not pixel dimensions.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoSize = '16:9' | '9:16'\n\n/**\n * Provider-specific options for Gemini Veo video generation.\n *\n * Derived from the SDK's `GenerateVideosConfig`, minus the fields the\n * adapter manages itself:\n * - `durationSeconds` — set via the typed top-level `duration` option\n *   (use `adapter.snapDuration(seconds)` to coerce raw seconds)\n * - `aspectRatio` — set via the top-level `size` option\n * - `lastFrame` / `referenceImages` — set via image parts in the `prompt`\n *   with `metadata.role: 'end_frame'` / `'reference'`\n * - `httpOptions` / `abortSignal` — client-level transport concerns\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoProviderOptions = Omit<\n  GenerateVideosConfig,\n  | 'durationSeconds'\n  | 'aspectRatio'\n  | 'lastFrame'\n  | 'referenceImages'\n  | 'httpOptions'\n  | 'abortSignal'\n>\n\n/**\n * Provider-specific options for Gemini Omni Flash video generation on the\n * Interactions API.\n *\n * Derived from the SDK's `Interactions.CreateModelInteractionParamsNonStreaming`,\n * minus the fields the adapter manages itself:\n * - `model` / `input` — set from the adapter's model and the `prompt`\n * - `stream` / `background` — the adapter always creates a background job\n *   and polls it through the `generateVideo` jobs API\n * - `response_modalities` / `response_format` — the adapter requests video\n *   output and maps the top-level `size` option onto\n *   `response_format.aspect_ratio`\n * - `tools` / `response_mime_type` — not applicable to video generation\n *\n * Notable passthroughs:\n * - `previous_interaction_id` — conversational video editing: chain a new\n *   prompt onto a prior Omni interaction to refine its video\n * - `generation_config.video_config.task` — pin the task mode\n *   (`'text_to_video' | 'image_to_video' | 'reference_to_video' | 'edit'`)\n *   instead of letting the model infer it\n *\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiOmniVideoProviderOptions = Omit<\n  Interactions.CreateModelInteractionParamsNonStreaming,\n  | 'model'\n  | 'input'\n  | 'stream'\n  | 'background'\n  | 'response_modalities'\n  | 'response_format'\n  | 'response_mime_type'\n  | 'tools'\n>\n\n/**\n * Model-specific provider options mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelProviderOptionsByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? GeminiOmniVideoProviderOptions\n    : GeminiVideoProviderOptions\n}\n\n/**\n * Model-specific size (aspect ratio) mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelSizeByName = {\n  [TModel in GeminiVideoModel]: GeminiVideoSize\n}\n\n/**\n * Per-model prompt input modalities. Every Veo model accepts image\n * conditioning inputs (first frame, last frame, reference images) alongside\n * the text prompt. Omni Flash additionally accepts video inputs (short\n * reference clips / videos to edit).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelInputModalitiesByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? readonly ['image', 'video']\n    : readonly ['image']\n}\n\n/**\n * Per-model duration unions (seconds, as numbers — Veo's\n * `parameters.durationSeconds` field is numeric; Omni Flash accepts a\n * continuous 3–10 second range, fractional seconds included, so it stays\n * `number` — the adapter rejects out-of-range values at job creation,\n * against the range entry below).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelDurationByName = {\n  'veo-3.1-generate-preview': 4 | 6 | 8\n  'veo-3.1-fast-generate-preview': 4 | 6 | 8\n  'veo-3.1-lite-generate-preview': 4 | 6 | 8\n  'gemini-omni-flash-preview': number\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n *\n * Veo values are curated from the official docs\n * (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec\n * types the `:predictLongRunning` request's `parameters` as unconstrained,\n * so it carries no per-model duration information to derive these from.\n * Omni Flash's 3–10s range was verified against the live API\n * (2026-07-02): `response_format.duration` takes a `\"<seconds>s\"` string,\n * fractional values are accepted, out-of-range values are rejected with\n * \"minimum allowed 3s\" / \"maximum allowed 10s\", and omitting it defaults\n * to a 10-second clip.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GEMINI_VIDEO_DURATIONS: {\n  readonly [TModel in GeminiVideoModel]: DurationOptions<\n    GeminiVideoModelDurationByName[TModel]\n  >\n} = {\n  'veo-3.1-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-fast-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-lite-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'gemini-omni-flash-preview': {\n    kind: 'range',\n    min: 3,\n    max: 10,\n    unit: 'seconds',\n  },\n}\n\n/**\n * Look up the duration options for a Gemini video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGeminiVideoDurationOptions<TModel extends GeminiVideoModel>(\n  model: TModel,\n): DurationOptions<GeminiVideoModelDurationByName[TModel]> {\n  return GEMINI_VIDEO_DURATIONS[model]\n}\n"],"names":[],"mappings":";AAkCO,SAAS,yBACd,OACuC;AACvC,SAAQ,iCAA2D;AAAA,IACjE;AAAA,EAAA;AAEJ;AAwIO,MAAM,yBAIT;AAAA,EACF,4BAA4B,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EAChE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EACrE,iCAAiC,EAAE,MAAM,YAAY,QAAQ,CAAC,GAAG,GAAG,CAAC,EAAA;AAAA,EACrE,6BAA6B;AAAA,IAC3B,MAAM;AAAA,IACN,KAAK;AAAA,IACL,KAAK;AAAA,IACL,MAAM;AAAA,EAAA;AAEV;AAOO,SAAS,8BACd,OACyD;AACzD,SAAO,uBAAuB,KAAK;AACrC;"}
+{"version":3,"file":"video-provider-options.js","names":[],"sources":["../../../src/video/video-provider-options.ts"],"sourcesContent":["/**\n * Gemini Video Generation Provider Options\n *\n * Covers two request paths behind the one video adapter:\n * - Veo models — long-running operations via `:predictLongRunning`\n *   (https://ai.google.dev/gemini-api/docs/video)\n * - Gemini Omni Flash — background jobs via the Interactions API\n *   (https://ai.google.dev/gemini-api/docs/omni)\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nimport { GEMINI_INTERACTIONS_VIDEO_MODELS } from '../model-meta'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type { GenerateVideosConfig, Interactions } from '@google/genai'\nimport type { GEMINI_VIDEO_MODELS } from '../model-meta'\n\n/**\n * Model type for Gemini video generation (Veo + Omni Flash).\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModel = (typeof GEMINI_VIDEO_MODELS)[number]\n\n/**\n * Video models served by the Interactions API (Gemini Omni Flash) rather\n * than Veo's `:predictLongRunning` operations flow.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiInteractionsVideoModel =\n  (typeof GEMINI_INTERACTIONS_VIDEO_MODELS)[number]\n\n/**\n * Runtime guard for the Interactions-served video models.\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport function isInteractionsVideoModel(\n  model: GeminiVideoModel,\n): model is GeminiInteractionsVideoModel {\n  return (GEMINI_INTERACTIONS_VIDEO_MODELS as ReadonlyArray<string>).includes(\n    model,\n  )\n}\n\n/**\n * Supported aspect ratios for Gemini video generation. This is the `size`\n * value for the Gemini video adapter — both Veo and Omni Flash express\n * output shape as an aspect ratio (plus an optional `resolution` in Veo's\n * `modelOptions`), not pixel dimensions.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoSize = '16:9' | '9:16'\n\n/**\n * Provider-specific options for Gemini Veo video generation.\n *\n * Derived from the SDK's `GenerateVideosConfig`, minus the fields the\n * adapter manages itself:\n * - `durationSeconds` — set via the typed top-level `duration` option\n *   (use `adapter.snapDuration(seconds)` to coerce raw seconds)\n * - `aspectRatio` — set via the top-level `size` option\n * - `lastFrame` / `referenceImages` — set via image parts in the `prompt`\n *   with `metadata.role: 'end_frame'` / `'reference'`\n * - `httpOptions` / `abortSignal` — client-level transport concerns\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoProviderOptions = Omit<\n  GenerateVideosConfig,\n  | 'durationSeconds'\n  | 'aspectRatio'\n  | 'lastFrame'\n  | 'referenceImages'\n  | 'httpOptions'\n  | 'abortSignal'\n>\n\n/**\n * Provider-specific options for Gemini Omni Flash video generation on the\n * Interactions API.\n *\n * Derived from the SDK's `Interactions.CreateModelInteractionParamsNonStreaming`,\n * minus the fields the adapter manages itself:\n * - `model` / `input` — set from the adapter's model and the `prompt`\n * - `stream` / `background` — the adapter always creates a background job\n *   and polls it through the `generateVideo` jobs API\n * - `response_modalities` / `response_format` — the adapter requests video\n *   output and maps the top-level `size` option onto\n *   `response_format.aspect_ratio`\n * - `tools` / `response_mime_type` — not applicable to video generation\n *\n * Notable passthroughs:\n * - `previous_interaction_id` — conversational video editing: chain a new\n *   prompt onto a prior Omni interaction to refine its video\n * - `generation_config.video_config.task` — pin the task mode\n *   (`'text_to_video' | 'image_to_video' | 'reference_to_video' | 'edit'`)\n *   instead of letting the model infer it\n *\n * @experimental Omni video generation is an experimental feature and may change.\n */\nexport type GeminiOmniVideoProviderOptions = Omit<\n  Interactions.CreateModelInteractionParamsNonStreaming,\n  | 'model'\n  | 'input'\n  | 'stream'\n  | 'background'\n  | 'response_modalities'\n  | 'response_format'\n  | 'response_mime_type'\n  | 'tools'\n>\n\n/**\n * Model-specific provider options mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelProviderOptionsByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? GeminiOmniVideoProviderOptions\n    : GeminiVideoProviderOptions\n}\n\n/**\n * Model-specific size (aspect ratio) mapping.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelSizeByName = {\n  [TModel in GeminiVideoModel]: GeminiVideoSize\n}\n\n/**\n * Per-model prompt input modalities. Every Veo model accepts image\n * conditioning inputs (first frame, last frame, reference images) alongside\n * the text prompt. Omni Flash additionally accepts video inputs (short\n * reference clips / videos to edit).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelInputModalitiesByName = {\n  [TModel in GeminiVideoModel]: TModel extends GeminiInteractionsVideoModel\n    ? readonly ['image', 'video']\n    : readonly ['image']\n}\n\n/**\n * Per-model duration unions (seconds, as numbers — Veo's\n * `parameters.durationSeconds` field is numeric; Omni Flash accepts a\n * continuous 3–10 second range, fractional seconds included, so it stays\n * `number` — the adapter rejects out-of-range values at job creation,\n * against the range entry below).\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport type GeminiVideoModelDurationByName = {\n  'veo-3.1-generate-preview': 4 | 6 | 8\n  'veo-3.1-fast-generate-preview': 4 | 6 | 8\n  'veo-3.1-lite-generate-preview': 4 | 6 | 8\n  'gemini-omni-flash-preview': number\n}\n\n/**\n * Runtime duration table backing `availableDurations()` / `snapDuration()`.\n *\n * Veo values are curated from the official docs\n * (https://ai.google.dev/gemini-api/docs/video) — the Gemini OpenAPI spec\n * types the `:predictLongRunning` request's `parameters` as unconstrained,\n * so it carries no per-model duration information to derive these from.\n * Omni Flash's 3–10s range was verified against the live API\n * (2026-07-02): `response_format.duration` takes a `\"<seconds>s\"` string,\n * fractional values are accepted, out-of-range values are rejected with\n * \"minimum allowed 3s\" / \"maximum allowed 10s\", and omitting it defaults\n * to a 10-second clip.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport const GEMINI_VIDEO_DURATIONS: {\n  readonly [TModel in GeminiVideoModel]: DurationOptions<\n    GeminiVideoModelDurationByName[TModel]\n  >\n} = {\n  'veo-3.1-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-fast-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'veo-3.1-lite-generate-preview': { kind: 'discrete', values: [4, 6, 8] },\n  'gemini-omni-flash-preview': {\n    kind: 'range',\n    min: 3,\n    max: 10,\n    unit: 'seconds',\n  },\n}\n\n/**\n * Look up the duration options for a Gemini video model.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport function getGeminiVideoDurationOptions<TModel extends GeminiVideoModel>(\n  model: TModel,\n): DurationOptions<GeminiVideoModelDurationByName[TModel]> {\n  return GEMINI_VIDEO_DURATIONS[model]\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAkCA,SAAgB,yBACd,OACuC;CACvC,OAAQ,iCAA2D,SACjE,KACF;AACF;;;;;;;;;;;;;;;;AAwIA,IAAa,yBAIT;CACF,4BAA4B;EAAE,MAAM;EAAY,QAAQ;GAAC;GAAG;GAAG;EAAC;CAAE;CAClE,iCAAiC;EAAE,MAAM;EAAY,QAAQ;GAAC;GAAG;GAAG;EAAC;CAAE;CACvE,iCAAiC;EAAE,MAAM;EAAY,QAAQ;GAAC;GAAG;GAAG;EAAC;CAAE;CACvE,6BAA6B;EAC3B,MAAM;EACN,KAAK;EACL,KAAK;EACL,MAAM;CACR;AACF;;;;;;AAOA,SAAgB,8BACd,OACyD;CACzD,OAAO,uBAAuB;AAChC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tanstack/ai-gemini",
-  "version": "0.20.0",
+  "version": "0.21.0",
   "description": "Google Gemini adapter for TanStack AI chat, images, speech, audio generation, and structured outputs.",
   "author": "Tanner Linsley",
   "license": "MIT",
@@ -56,23 +56,23 @@
   "dependencies": {
     "@google/genai": "^2.10.0",
     "partial-json": "^0.1.7",
-    "@tanstack/ai-utils": "0.3.1"
+    "@tanstack/ai-utils": "0.4.0"
   },
   "peerDependencies": {
-    "@tanstack/ai": "^0.41.0"
+    "@tanstack/ai": "^0.43.0"
   },
   "devDependencies": {
     "@vitest/coverage-v8": "4.0.14",
-    "vite": "^7.3.3",
+    "vite": "^8.1.4",
     "zod": "^4.2.0",
-    "@tanstack/ai": "0.41.0"
+    "@tanstack/ai": "0.43.0"
   },
   "scripts": {
     "build": "vite build",
     "clean": "premove ./build ./dist",
-    "lint:fix": "eslint ./src --fix",
+    "lint:fix": "oxlint src --type-aware --fix",
     "test:build": "publint --strict",
-    "test:eslint": "eslint ./src",
+    "test:oxlint": "oxlint src --type-aware",
     "test:lib": "vitest",
     "test:lib:dev": "pnpm test:lib --watch",
     "test:types": "tsc"