elizaos--eliza/plugins/plugin-google-genai/models/text.ts

/**
 * Text-generation handlers backing every text `ModelType` tier
 * (nano/small/medium/large/mega, response-handler, action-planner). Each
 * handler resolves its concrete Gemini model name via `../utils/config`, builds
 * a `generateContent` request, and runs it through `recordLlmCall` so the call
 * lands in the trajectory log before returning.
 *
 * This module owns the translation between elizaOS's generic call shape and
 * Google's native `@google/genai` protocol: `normalizeToolsForGoogle` /
 * `normalizeToolConfigForGoogle` convert generic or native tool definitions to
 * `functionDeclarations` + `functionCallingConfig`, `resolveResponseJsonSchema`
 * routes structured-output schemas into `responseJsonSchema`, and
 * `buildPromptParts` inlines attachments (data URLs, remote URIs, raw bytes).
 * On the way back, `buildGoogleNativeTextResult` folds text, tool calls, finish
 * reason, and token usage into a single object that is returned as a
 * string-with-attached-fields (`GoogleTextModelResult`) whenever the caller
 * passed messages/tools/toolChoice/responseSchema, else as plain text.
 */
import type {
  GenerateTextParams,
  IAgentRuntime,
  JsonValue,
  RecordLlmCallDetails,
  TokenUsage,
  ToolCall,
} from "@elizaos/core";
import {
  buildCanonicalSystemPrompt,
  ElizaError,
  logger,
  ModelType,
  recordLlmCall,
  renderChatMessagesForPrompt,
  resolveEffectiveSystemPrompt,
} from "@elizaos/core";
import {
  createGoogleGenAI,
  getActionPlannerModel,
  getLargeModel,
  getMediumModel,
  getMegaModel,
  getNanoModel,
  getResponseHandlerModel,
  getSafetySettings,
  getSmallModel,
} from "../utils/config";
import { emitModelUsageEvent } from "../utils/events";
import { countTokens } from "../utils/tokenization";

const TEXT_NANO_MODEL_TYPE = ModelType.TEXT_NANO as string;
const TEXT_MEDIUM_MODEL_TYPE = ModelType.TEXT_MEDIUM as string;
const TEXT_SMALL_MODEL_TYPE = ModelType.TEXT_SMALL as string;
const TEXT_LARGE_MODEL_TYPE = ModelType.TEXT_LARGE as string;
const TEXT_MEGA_MODEL_TYPE = ModelType.TEXT_MEGA as string;
const RESPONSE_HANDLER_MODEL_TYPE = ModelType.RESPONSE_HANDLER as string;
const ACTION_PLANNER_MODEL_TYPE = ModelType.ACTION_PLANNER as string;

type ChatAttachment = {
  data: string | Uint8Array | URL;
  mediaType: string;
  filename?: string;
};

/**
 * Native Google GenAI tool input. Each function declaration carries a name,
 * description, and JSON Schema parameters object that the model can choose to
 * invoke.
 */
type GoogleFunctionDeclaration = {
  name: string;
  description?: string;
  parameters?: Record<string, unknown>;
};

type GoogleToolDeclaration = {
  functionDeclarations?: GoogleFunctionDeclaration[];
};

type GoogleFunctionCall = {
  id?: string;
  name?: string;
  args?: Record<string, unknown>;
};

type GoogleContentPart = {
  text?: string;
  thought?: boolean;
  functionCall?: GoogleFunctionCall;
};

type GoogleGenerateContentResponse = {
  text?: string;
  functionCalls?: GoogleFunctionCall[];
  candidates?: Array<{
    content?: {
      parts?: GoogleContentPart[];
    };
    finishReason?: string;
  }>;
  usageMetadata?: {
    promptTokenCount?: number;
    candidatesTokenCount?: number;
    totalTokenCount?: number;
    cachedContentTokenCount?: number;
  };
  modelVersion?: string;
  responseId?: string;
  createTime?: string;
};

type GoogleNativeTextResult = {
  text: string;
  toolCalls: ToolCall[];
  finishReason?: string;
  usage: TokenUsage;
  providerMetadata: Record<string, JsonValue | object | undefined>;
};

type GoogleTextModelResult = string & GoogleNativeTextResult;

type GenericToolDescriptor = {
  name?: string;
  description?: string;
  parameters?: unknown;
  inputSchema?: unknown;
  function?: { name?: string; description?: string; parameters?: unknown };
};

type LocalToolChoice =
  | "auto"
  | "required"
  | "none"
  | { type: "tool"; toolName?: string; name?: string }
  | { type: "function"; function: { name: string } }
  | { name: string };

type GenerateTextParamsWithAttachments = Omit<
  GenerateTextParams,
  "tools" | "toolChoice" | "responseSchema"
> & {
  attachments?: ChatAttachment[];
  /** Native or generic tool definitions; converted to Google functionDeclarations. */
  tools?:
    | GenericToolDescriptor[]
    | GoogleToolDeclaration[]
    | Record<string, GenericToolDescriptor>;
  /** Tool selection hint: "auto" | "required" | "none" | { type: "tool"; toolName } | { type: "function"; function }. */
  toolChoice?: LocalToolChoice;
  /** JSON Schema for structured output; routes through responseJsonSchema. */
  responseSchema?:
    | Record<string, unknown>
    | { schema: Record<string, unknown> };
};
type GoogleGenAIClient = NonNullable<ReturnType<typeof createGoogleGenAI>>;
type GenerateContentParams = Parameters<
  GoogleGenAIClient["models"]["generateContent"]
>[0];

function normalizeToolsForGoogle(
  tools: GenerateTextParamsWithAttachments["tools"],
): GoogleToolDeclaration[] | undefined {
  if (!tools) return undefined;

  // Already-shaped Google tools: array of { functionDeclarations: [...] }.
  if (
    Array.isArray(tools) &&
    tools.length > 0 &&
    typeof tools[0] === "object" &&
    tools[0] !== null &&
    "functionDeclarations" in (tools[0] as object)
  ) {
    return tools as GoogleToolDeclaration[];
  }

  const flat: GenericToolDescriptor[] = Array.isArray(tools)
    ? (tools as GenericToolDescriptor[])
    : Object.entries(tools).map(([name, value]) => ({ name, ...value }));

  const declarations: GoogleFunctionDeclaration[] = [];
  for (const tool of flat) {
    const name = tool.name ?? tool.function?.name;
    if (!name) {
      throw new Error("[GoogleGenAI] Tool definition is missing a name.");
    }
    const description = tool.description ?? tool.function?.description;
    const parameters = (tool.parameters ??
      tool.inputSchema ??
      tool.function?.parameters ?? {
        type: "object",
        properties: {},
      }) as Record<string, unknown>;
    declarations.push({
      name,
      ...(description ? { description } : {}),
      parameters,
    });
  }

  return declarations.length > 0
    ? [{ functionDeclarations: declarations }]
    : undefined;
}

function normalizeToolConfigForGoogle(
  toolChoice: GenerateTextParamsWithAttachments["toolChoice"],
):
  | {
      functionCallingConfig: {
        mode: "AUTO" | "ANY" | "NONE";
        allowedFunctionNames?: string[];
      };
    }
  | undefined {
  if (!toolChoice) return undefined;
  if (toolChoice === "auto") {
    return { functionCallingConfig: { mode: "AUTO" } };
  }
  if (toolChoice === "required") {
    return { functionCallingConfig: { mode: "ANY" } };
  }
  if (toolChoice === "none") {
    return { functionCallingConfig: { mode: "NONE" } };
  }
  let toolName: string | undefined;
  if ("type" in toolChoice) {
    toolName =
      toolChoice.type === "function"
        ? toolChoice.function.name
        : (toolChoice.toolName ?? toolChoice.name);
  } else {
    toolName = toolChoice.name;
  }
  if (toolName) {
    return {
      functionCallingConfig: {
        mode: "ANY",
        allowedFunctionNames: [toolName],
      },
    };
  }
  return undefined;
}

function resolveResponseJsonSchema(
  responseSchema: GenerateTextParamsWithAttachments["responseSchema"],
): Record<string, unknown> | undefined {
  if (!responseSchema) return undefined;
  if ("schema" in responseSchema && responseSchema.schema) {
    return responseSchema.schema as Record<string, unknown>;
  }
  return responseSchema as Record<string, unknown>;
}

function buildPromptParts(prompt: string, attachments?: ChatAttachment[]) {
  const parts: Array<
    | { text: string }
    | { fileData: { mimeType: string; fileUri: string } }
    | { inlineData: { mimeType: string; data: string } }
  > = [{ text: prompt }];

  for (const attachment of attachments ?? []) {
    if (attachment.data instanceof URL) {
      parts.push({
        fileData: {
          mimeType: attachment.mediaType,
          fileUri: attachment.data.toString(),
        },
      });
      continue;
    }

    if (
      typeof attachment.data === "string" &&
      /^https?:\/\//i.test(attachment.data)
    ) {
      parts.push({
        fileData: {
          mimeType: attachment.mediaType,
          fileUri: attachment.data,
        },
      });
      continue;
    }

    if (typeof attachment.data === "string") {
      const dataUrlMatch = attachment.data.match(
        /^data:([^;,]+);base64,(.+)$/i,
      );
      parts.push({
        inlineData: {
          mimeType: dataUrlMatch?.[1] ?? attachment.mediaType,
          data: dataUrlMatch?.[2] ?? attachment.data,
        },
      });
      continue;
    }

    parts.push({
      inlineData: {
        mimeType: attachment.mediaType,
        data: Buffer.from(attachment.data).toString("base64"),
      },
    });
  }

  return parts;
}

function resolveGoogleSystemInstruction(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): string | undefined {
  return resolveEffectiveSystemPrompt({
    params,
    fallback: buildCanonicalSystemPrompt({ character: runtime.character }),
  });
}

function resolveGooglePrompt(
  params: GenerateTextParamsWithAttachments,
  systemInstruction: string | undefined,
): string {
  return (
    renderChatMessagesForPrompt(params.messages, {
      omitDuplicateSystem: systemInstruction,
    }) ??
    params.prompt ??
    ""
  );
}

function getModelNameForType(
  runtime: IAgentRuntime,
  modelType: string,
): string {
  switch (modelType) {
    case TEXT_NANO_MODEL_TYPE:
      return getNanoModel(runtime);
    case TEXT_MEDIUM_MODEL_TYPE:
      return getMediumModel(runtime);
    case TEXT_SMALL_MODEL_TYPE:
      return getSmallModel(runtime);
    case TEXT_LARGE_MODEL_TYPE:
      return getLargeModel(runtime);
    case TEXT_MEGA_MODEL_TYPE:
      return getMegaModel(runtime);
    case RESPONSE_HANDLER_MODEL_TYPE:
      return getResponseHandlerModel(runtime);
    case ACTION_PLANNER_MODEL_TYPE:
      return getActionPlannerModel(runtime);
    default:
      return getLargeModel(runtime);
  }
}

function isRecord(value: unknown): value is Record<string, unknown> {
  return Boolean(value && typeof value === "object" && !Array.isArray(value));
}

function toJsonValue(value: unknown): JsonValue {
  if (
    value === null ||
    typeof value === "string" ||
    typeof value === "boolean"
  ) {
    return value;
  }
  if (typeof value === "number") {
    return Number.isFinite(value) ? value : null;
  }
  if (Array.isArray(value)) {
    return value.map((entry) => toJsonValue(entry));
  }
  if (isRecord(value)) {
    const record: Record<string, JsonValue> = {};
    for (const [key, entry] of Object.entries(value)) {
      if (entry !== undefined) {
        record[key] = toJsonValue(entry);
      }
    }
    return record;
  }
  return String(value);
}

function toToolArguments(value: unknown): Record<string, JsonValue> {
  if (!isRecord(value)) {
    return {};
  }
  const jsonValue = toJsonValue(value);
  return isRecord(jsonValue) ? (jsonValue as Record<string, JsonValue>) : {};
}

function readGoogleText(response: GoogleGenerateContentResponse): string {
  if (typeof response.text === "string") {
    return response.text;
  }
  return (
    response.candidates?.[0]?.content?.parts
      ?.filter((part) => !part.thought && typeof part.text === "string")
      .map((part) => part.text)
      .join("") ?? ""
  );
}

function readGoogleFunctionCalls(
  response: GoogleGenerateContentResponse,
): GoogleFunctionCall[] {
  if (
    Array.isArray(response.functionCalls) &&
    response.functionCalls.length > 0
  ) {
    return response.functionCalls;
  }
  const calls: GoogleFunctionCall[] = [];
  for (const candidate of response.candidates ?? []) {
    for (const part of candidate.content?.parts ?? []) {
      if (part.functionCall) {
        calls.push(part.functionCall);
      }
    }
  }
  return calls;
}

function normalizeGoogleToolCalls(
  response: GoogleGenerateContentResponse,
): ToolCall[] {
  return readGoogleFunctionCalls(response)
    .map((call, index): ToolCall | undefined => {
      const name = typeof call.name === "string" ? call.name : "";
      if (!name) {
        return undefined;
      }
      const id =
        typeof call.id === "string" && call.id.length > 0
          ? call.id
          : `google-genai-tool-call-${index + 1}`;
      const args = toToolArguments(call.args);
      return {
        id,
        name,
        arguments: args,
        toolName: name,
        toolCallId: id,
        type: "function",
        args,
        input: args,
      };
    })
    .filter((call): call is ToolCall => Boolean(call));
}

function normalizeGoogleFinishReason(
  response: GoogleGenerateContentResponse,
  toolCalls: ToolCall[],
): string | undefined {
  if (toolCalls.length > 0) {
    return "tool-calls";
  }
  return response.candidates?.find((candidate) => candidate.finishReason)
    ?.finishReason;
}

function firstNumber(...values: unknown[]): number | undefined {
  for (const value of values) {
    if (typeof value === "number" && Number.isFinite(value)) {
      return value;
    }
  }
  return undefined;
}

async function normalizeGoogleUsage(
  response: GoogleGenerateContentResponse,
  prompt: string,
  text: string,
): Promise<TokenUsage> {
  const metadata = response.usageMetadata;
  const promptTokens =
    firstNumber(metadata?.promptTokenCount) ?? (await countTokens(prompt));
  const completionTokens =
    firstNumber(metadata?.candidatesTokenCount) ?? (await countTokens(text));
  const totalTokens =
    firstNumber(metadata?.totalTokenCount) ?? promptTokens + completionTokens;

  return {
    promptTokens,
    completionTokens,
    totalTokens,
    cacheReadInputTokens: firstNumber(metadata?.cachedContentTokenCount),
  };
}

async function buildGoogleNativeTextResult(
  response: GoogleGenerateContentResponse,
  prompt: string,
  modelName: string,
): Promise<GoogleNativeTextResult> {
  const text = readGoogleText(response);
  const toolCalls = normalizeGoogleToolCalls(response);
  const usage = await normalizeGoogleUsage(response, prompt, text);

  return {
    text,
    toolCalls,
    finishReason: normalizeGoogleFinishReason(response, toolCalls),
    usage,
    providerMetadata: {
      provider: "google-genai",
      modelName,
      modelVersion: response.modelVersion,
      responseId: response.responseId,
      createTime: response.createTime,
      usageMetadata: response.usageMetadata,
    },
  };
}

function usesNativeTextResult(
  params: GenerateTextParamsWithAttachments,
): boolean {
  return Boolean(
    params.messages ||
      params.tools ||
      params.toolChoice ||
      params.responseSchema,
  );
}

function buildGoogleGenerationConfig(
  params: GenerateTextParamsWithAttachments,
  systemInstruction: string | undefined,
  temperature: number,
  maxTokens: number | undefined,
  stopSequences: string[],
): NonNullable<GenerateContentParams["config"]> {
  const tools = normalizeToolsForGoogle(params.tools);
  const toolConfig = normalizeToolConfigForGoogle(params.toolChoice);
  const responseJsonSchema = resolveResponseJsonSchema(params.responseSchema);

  const baseConfig: Record<string, unknown> = {
    temperature,
    topK: 40,
    topP: 0.95,
    stopSequences,
    safetySettings: getSafetySettings(),
    ...(typeof maxTokens === "number" ? { maxOutputTokens: maxTokens } : {}),
    ...(systemInstruction && { systemInstruction }),
    ...(tools ? { tools } : {}),
    ...(toolConfig ? { toolConfig } : {}),
    ...(responseJsonSchema
      ? {
          responseMimeType: "application/json",
          responseJsonSchema,
        }
      : {}),
  };

  return baseConfig as NonNullable<GenerateContentParams["config"]>;
}

function createLlmCallDetails(
  modelName: string,
  modelType: string,
  prompt: string,
  systemInstruction: string | undefined,
  temperature: number,
  maxTokens: number | undefined,
  maxTokensOmitted?: boolean,
): RecordLlmCallDetails {
  return {
    model: modelName,
    systemPrompt: systemInstruction ?? "",
    userPrompt: prompt,
    temperature,
    maxTokens: maxTokens ?? 0,
    maxTokensOmitted: maxTokensOmitted ? true : undefined,
    purpose: "external_llm",
    actionType: `google-genai.${modelType}.generateContent`,
  };
}

async function generateContentWithTrajectory(
  runtime: IAgentRuntime,
  genAI: GoogleGenAIClient,
  modelName: string,
  modelType: string,
  prompt: string,
  systemInstruction: string | undefined,
  temperature: number,
  maxTokens: number | undefined,
  maxTokensOmitted: boolean | undefined,
  request: GenerateContentParams,
  shouldReturnNativeResult: boolean,
): Promise<string> {
  const details = createLlmCallDetails(
    modelName,
    modelType,
    prompt,
    systemInstruction,
    temperature,
    maxTokens,
    maxTokensOmitted,
  );
  const response = await recordLlmCall(runtime, details, async () => {
    const result = (await genAI.models.generateContent(
      request,
    )) as GoogleGenerateContentResponse;
    const normalized = await buildGoogleNativeTextResult(
      result,
      prompt,
      modelName,
    );
    details.response = normalized.text;
    details.toolCalls = normalized.toolCalls;
    details.finishReason = normalized.finishReason;
    details.providerMetadata = normalized.providerMetadata;
    details.promptTokens = normalized.usage.promptTokens;
    details.completionTokens = normalized.usage.completionTokens;
    details.cacheReadInputTokens = normalized.usage.cacheReadInputTokens;
    return normalized;
  });

  // A completion with no text and no tool calls is a provider failure (safety
  // block, empty candidates, truncation) — never a legitimate result. Returning
  // "" here would fabricate a healthy-empty completion the planner cannot
  // distinguish from a real answer (#9324: throw, never fabricate).
  if (response.text.length === 0 && response.toolCalls.length === 0) {
    throw new ElizaError(
      `[Google GenAI] ${modelType} returned an empty completion${
        response.finishReason ? ` (finishReason: ${response.finishReason})` : ""
      }`,
      {
        code: "MODEL_EMPTY_COMPLETION",
        context: { modelType, finishReason: response.finishReason },
      },
    );
  }

  emitModelUsageEvent(runtime, modelType, prompt, response.usage);

  if (shouldReturnNativeResult) {
    return response as GoogleTextModelResult;
  }

  return response.text;
}

export async function handleTextSmall(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  const { stopSequences = [], temperature = 0.7, attachments } = params;
  const maxTokens = params.omitMaxTokens
    ? undefined
    : (params.maxTokens ?? 8192);
  const genAI = createGoogleGenAI(runtime);
  if (!genAI) {
    throw new Error("Google Generative AI client not initialized");
  }

  const modelName = getModelNameForType(runtime, TEXT_SMALL_MODEL_TYPE);

  logger.log(`[TEXT_SMALL] Using model: ${modelName}`);

  try {
    const systemInstruction = resolveGoogleSystemInstruction(runtime, params);
    const promptText = resolveGooglePrompt(params, systemInstruction);
    return await generateContentWithTrajectory(
      runtime,
      genAI,
      modelName,
      TEXT_SMALL_MODEL_TYPE,
      promptText,
      systemInstruction,
      temperature,
      maxTokens,
      params.omitMaxTokens,
      {
        model: modelName,
        contents:
          (attachments?.length ?? 0) > 0
            ? [
                {
                  role: "user",
                  parts: buildPromptParts(promptText, attachments),
                },
              ]
            : promptText,
        config: buildGoogleGenerationConfig(
          params,
          systemInstruction,
          temperature,
          maxTokens,
          stopSequences,
        ),
      },
      usesNativeTextResult(params),
    );
  } catch (error) {
    // error-policy:J2 context-adding rethrow — log the model label for
    // operators, then rethrow the original provider error unchanged so the
    // planner/runtime sees the real failure.
    logger.error(
      `[TEXT_SMALL] Error: ${error instanceof Error ? error.message : String(error)}`,
    );
    throw error;
  }
}

export async function handleTextLarge(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  const { stopSequences = [], temperature = 0.7, attachments } = params;
  const maxTokens = params.omitMaxTokens
    ? undefined
    : (params.maxTokens ?? 8192);
  const genAI = createGoogleGenAI(runtime);
  if (!genAI) {
    throw new Error("Google Generative AI client not initialized");
  }

  const modelName = getModelNameForType(runtime, TEXT_LARGE_MODEL_TYPE);

  logger.log(`[TEXT_LARGE] Using model: ${modelName}`);

  try {
    const systemInstruction = resolveGoogleSystemInstruction(runtime, params);
    const promptText = resolveGooglePrompt(params, systemInstruction);
    return await generateContentWithTrajectory(
      runtime,
      genAI,
      modelName,
      TEXT_LARGE_MODEL_TYPE,
      promptText,
      systemInstruction,
      temperature,
      maxTokens,
      params.omitMaxTokens,
      {
        model: modelName,
        contents:
          (attachments?.length ?? 0) > 0
            ? [
                {
                  role: "user",
                  parts: buildPromptParts(promptText, attachments),
                },
              ]
            : promptText,
        config: buildGoogleGenerationConfig(
          params,
          systemInstruction,
          temperature,
          maxTokens,
          stopSequences,
        ),
      },
      usesNativeTextResult(params),
    );
  } catch (error) {
    // error-policy:J2 context-adding rethrow — log the model label for
    // operators, then rethrow the original provider error unchanged.
    logger.error(
      `[TEXT_LARGE] Error: ${error instanceof Error ? error.message : String(error)}`,
    );
    throw error;
  }
}

export async function handleTextNano(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  return handleTextWithType(runtime, TEXT_NANO_MODEL_TYPE, params);
}

export async function handleTextMedium(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  return handleTextWithType(runtime, TEXT_MEDIUM_MODEL_TYPE, params);
}

export async function handleTextMega(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  return handleTextWithType(runtime, TEXT_MEGA_MODEL_TYPE, params);
}

export async function handleResponseHandler(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  return handleTextWithType(runtime, RESPONSE_HANDLER_MODEL_TYPE, params);
}

export async function handleActionPlanner(
  runtime: IAgentRuntime,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  return handleTextWithType(runtime, ACTION_PLANNER_MODEL_TYPE, params);
}

async function handleTextWithType(
  runtime: IAgentRuntime,
  modelType: string,
  params: GenerateTextParamsWithAttachments,
): Promise<string> {
  const { stopSequences = [], temperature = 0.7, attachments } = params;
  const maxTokens = params.omitMaxTokens
    ? undefined
    : (params.maxTokens ?? 8192);
  const genAI = createGoogleGenAI(runtime);
  if (!genAI) {
    throw new Error("Google Generative AI client not initialized");
  }

  const modelName = getModelNameForType(runtime, modelType);

  logger.log(`[${modelType}] Using model: ${modelName}`);

  try {
    const systemInstruction = resolveGoogleSystemInstruction(runtime, params);
    const promptText = resolveGooglePrompt(params, systemInstruction);
    return await generateContentWithTrajectory(
      runtime,
      genAI,
      modelName,
      modelType,
      promptText,
      systemInstruction,
      temperature,
      maxTokens,
      params.omitMaxTokens,
      {
        model: modelName,
        contents:
          (attachments?.length ?? 0) > 0
            ? [
                {
                  role: "user",
                  parts: buildPromptParts(promptText, attachments),
                },
              ]
            : promptText,
        config: buildGoogleGenerationConfig(
          params,
          systemInstruction,
          temperature,
          maxTokens,
          stopSequences,
        ),
      },
      usesNativeTextResult(params),
    );
  } catch (error) {
    // error-policy:J2 context-adding rethrow — log the model label for
    // operators, then rethrow the original provider error unchanged.
    logger.error(
      `[${modelType}] Error: ${error instanceof Error ? error.message : String(error)}`,
    );
    throw error;
  }
}