@falai/agent 2.2.3 → 2.4.1

package/src/providers/OpenRouterProvider.ts

@@ -4,24 +4,13 @@
  */
 
 import OpenAI from "openai";
-import type { ChatCompletionCreateParamsNonStreaming } from "openai/resources/chat/completions";
-import { FunctionParameters } from "openai/resources/shared.mjs";
 
-import type {
-  AiProvider,
-  AgentStructuredResponse,
-  GenerateMessageInput,
-  GenerateMessageOutput,
-  GenerateMessageStreamChunk,
-  StructuredSchema,
-} from "../types";
-import type { HistoryItem } from "../types/history";
-import { withTimeoutAndRetry, logger } from "../utils";
-
-const DEFAULT_RETRY_CONFIG = {
-  timeout: 60000,
-  retries: 3,
-};
+import type { ProviderCapabilities } from "../types/ai";
+import type { ErrorClassificationOptions } from "./errorClassification";
+import {
+  OpenAICompatibleProvider,
+  type OpenAICompatibleRequestConfig,
+} from "./OpenAICompatibleProvider";
 
 /**
  * Configuration options for OpenRouter provider
@@ -39,9 +28,7 @@ export interface OpenRouterProviderOptions {
   /** Optional app name for OpenRouter rankings */
   siteName?: string;
   /** Default parameters - uses ChatCompletionCreateParamsNonStreaming from openai package */
-  config?: Partial<
-    Omit<ChatCompletionCreateParamsNonStreaming, "model" | "messages">
-  >;
+  config?: OpenAICompatibleRequestConfig;
   /** Retry configuration */
   retryConfig?: {
     timeout?: number;
@@ -49,96 +36,35 @@ export interface OpenRouterProviderOptions {
   };
 }
 
-/**
- * Type guard for errors with status/code properties
- */
-interface ErrorWithStatus {
-  status?: number;
-  code?: string;
-  message?: string;
-  type?: string;
-}
-
-/**
- * Type guard to check if error is ErrorWithStatus
- */
-function isErrorWithStatus(error: unknown): error is ErrorWithStatus {
-  return (
-    typeof error === "object" &&
-    error !== null &&
-    ("status" in error || "code" in error || "message" in error)
-  );
-}
-
-/**
- * Safely extract error message
- */
-function getErrorMessage(error: unknown): string {
-  if (error instanceof Error) {
-    return error.message;
-  }
-  if (isErrorWithStatus(error) && error.message) {
-    return error.message;
-  }
-  return String(error);
-}
-
-/**
- * Determines if an error should trigger backup model usage
- */
-const shouldUseBackupModel = (error: unknown): boolean => {
-  if (!isErrorWithStatus(error)) {
-    return false;
-  }
-
-  // Server errors
-  if (error.status === 500 || error.status === 503) {
-    return true;
-  }
-
-  // Rate limiting
-  if (error.status === 429) {
-    return true;
-  }
-
-  // Model not available or overloaded
-  if (error.code === "model_not_found" || error.code === "model_overloaded") {
-    return true;
-  }
-
-  const message = getErrorMessage(error);
-  if (
-    message.includes("overloaded") ||
-    message.includes("unavailable") ||
-    message.includes("internal error") ||
-    message.includes("Internal error") ||
-    message.includes("capacity")
-  ) {
-    return true;
-  }
-
-  return false;
-};
-
 /**
  * OpenRouter provider implementation
  * Provides access to multiple AI models through OpenRouter's unified API
  */
-export class OpenRouterProvider implements AiProvider {
+export class OpenRouterProvider extends OpenAICompatibleProvider {
   public readonly name = "openrouter";
-  private client: OpenAI;
-  private primaryModel: string;
-  private backupModels: string[];
-  private config?: Partial<
-    Omit<ChatCompletionCreateParamsNonStreaming, "model" | "messages">
-  >;
-  private retryConfig: { timeout: number; retries: number };
+  public readonly capabilities: ProviderCapabilities = {
+    supportsTools: true,
+    supportsNativeJsonSchema: true,
+    supportsStreaming: true,
+    supportsStreamingToolCalls: true,
+    supportsPromptCaching: false,
+  };
+
+  protected readonly logLabel = "OPENROUTER";
+  protected readonly displayName = "OpenRouter";
+
+  /** OpenRouter additionally reports capacity issues in error messages */
+  protected override readonly classificationOptions: ErrorClassificationOptions =
+    {
+      overloadedCodes: ["model_not_found", "model_overloaded"],
+      overloadedMessages: ["capacity"],
+    };
 
   constructor(options: OpenRouterProviderOptions) {
     const {
       apiKey,
       model,
-      backupModels = [],
+      backupModels,
       siteUrl,
       siteName,
       config,
@@ -154,503 +80,19 @@ export class OpenRouterProvider implements AiProvider {
     }
 
     // Initialize OpenAI client with OpenRouter base URL
-    this.client = new OpenAI({
-      apiKey,
-      baseURL: "https://openrouter.ai/api/v1",
-      defaultHeaders: {
-        "HTTP-Referer": siteUrl || "",
-        "X-Title": siteName || "",
-      },
-    });
-
-    this.primaryModel = model;
-    this.backupModels = backupModels;
-    this.config = config;
-    this.retryConfig = {
-      timeout: retryConfig?.timeout || DEFAULT_RETRY_CONFIG.timeout,
-      retries: retryConfig?.retries || DEFAULT_RETRY_CONFIG.retries,
-    };
-  }
-
-  /**
-   * Build OpenRouter-formatted messages from HistoryItem[] array.
-   * OpenRouter uses OpenAI-compatible format.
-   */
-  private buildOpenRouterMessages(history: HistoryItem[]): Array<unknown> {
-    const messages: Array<unknown> = [];
-
-    for (const item of history) {
-      switch (item.role) {
-        case "system":
-          messages.push({ role: "system", content: item.content });
-          break;
-        case "user":
-          messages.push({ role: "user", content: item.content });
-          break;
-        case "assistant":
-          if (item.tool_calls && item.tool_calls.length > 0) {
-            messages.push({
-              role: "assistant",
-              content: item.content || null,
-              tool_calls: item.tool_calls.map(tc => ({
-                id: tc.id,
-                type: "function",
-                function: {
-                  name: tc.name,
-                  arguments: JSON.stringify(tc.arguments),
-                },
-              })),
-            });
-          } else {
-            messages.push({ role: "assistant", content: item.content || "" });
-          }
-          break;
-        case "tool":
-          messages.push({
-            role: "tool",
-            tool_call_id: item.tool_call_id,
-            content: typeof item.content === "string" ? item.content : JSON.stringify(item.content),
-          });
-          break;
-      }
-    }
-
-    return messages;
-  }
-
-  /**
-   * Adapt common schema format to OpenRouter's format.
-   * OpenRouter is OpenAI-compatible and uses standard JSON Schema.
-   *
-   * @private
-   */
-  private adaptSchemaForOpenRouter(
-    schema: StructuredSchema
-  ): Record<string, unknown> {
-    // OpenRouter is OpenAI-compatible and uses standard JSON Schema
-    // Our StructuredSchema is already JSON Schema compatible
-    return schema as Record<string, unknown>;
-  }
-
-  async generateMessage<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): Promise<GenerateMessageOutput<TStructured>> {
-    return this.generateWithBackup<TContext, TStructured>(input);
-  }
-
-  async *generateMessageStream<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): AsyncGenerator<GenerateMessageStreamChunk<TStructured>> {
-    yield* this.generateStreamWithBackup<TContext, TStructured>(input);
-  }
-
-  private async generateWithBackup<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): Promise<GenerateMessageOutput<TStructured>> {
-    // Try primary model first
-    try {
-      return await this.generateWithModel(this.primaryModel, input);
-    } catch (primaryError: unknown) {
-      const primaryErrMsg = getErrorMessage(primaryError);
-      logger.warn(
-        `[OPENROUTER] Primary model ${this.primaryModel} failed: ${primaryErrMsg}`
-      );
-
-      if (!shouldUseBackupModel(primaryError)) {
-        throw primaryError;
-      }
-
-      logger.debug(`[OPENROUTER] Trying backup models`);
-
-      let lastBackupError: unknown = primaryError;
-
-      for (let i = 0; i < this.backupModels.length; i++) {
-        const backupModel = this.backupModels[i];
-        logger.debug(
-          `[OPENROUTER] Trying backup model ${i + 1}/${this.backupModels.length
-          }: ${backupModel}`
-        );
-
-        try {
-          const result = await this.generateWithModel(backupModel, input);
-          logger.debug(`[OPENROUTER] Backup model ${backupModel} succeeded`);
-          return result as GenerateMessageOutput<TStructured>;
-        } catch (backupError: unknown) {
-          const backupErrMsg = getErrorMessage(backupError);
-          logger.warn(
-            `[OPENROUTER] Backup model ${backupModel} failed: ${backupErrMsg}`
-          );
-          lastBackupError = backupError;
-
-          if (
-            !shouldUseBackupModel(backupError) &&
-            i < this.backupModels.length - 1
-          ) {
-            logger.debug(
-              `[OPENROUTER] Backup model error doesn't qualify for further attempts`
-            );
-            break;
-          }
-        }
-      }
-
-      const lastBackupErrMsg = getErrorMessage(lastBackupError);
-      logger.error(
-        `[OPENROUTER] All models failed. Primary: ${primaryErrMsg}, Last backup: ${lastBackupErrMsg}`
-      );
-      throw lastBackupError;
-    }
-  }
-
-  private async generateWithModel<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    model: string,
-    input: GenerateMessageInput<TContext>
-  ): Promise<GenerateMessageOutput<TStructured>> {
-    const operation = async (): Promise<GenerateMessageOutput> => {
-      const historyMessages = this.buildOpenRouterMessages(input.history);
-      historyMessages.push({ role: "user", content: input.prompt });
-
-      const params: ChatCompletionCreateParamsNonStreaming = {
-        model,
-        messages: historyMessages as ChatCompletionCreateParamsNonStreaming["messages"],
-        ...this.config,
-      };
-
-      // Override with input parameters if provided
-      if (input.parameters?.maxOutputTokens !== undefined) {
-        params.max_tokens = input.parameters.maxOutputTokens;
-      }
-
-      // Add tools if provided
-      if (input.tools && input.tools.length > 0) {
-        params.tools = input.tools.map((tool) => ({
-          type: "function" as const,
-          function: {
-            name: tool.name || tool.id,
-            description: tool.description,
-            parameters: tool.parameters as FunctionParameters, // JSON schema
-          },
-        }));
-        params.tool_choice = "auto";
-      }
-
-      // Use structured output API if JSON schema is provided
-      if (input.parameters?.jsonSchema) {
-        const response = await this.client.responses.parse({
-          model,
-          instructions: input.prompt,
-          input: "",
-          reasoning: {
-            effort: input.parameters?.reasoning?.effort || "low",
-          },
-          text: {
-            format: {
-              type: "json_schema",
-              name: input.parameters?.schemaName || "structured_output",
-              // Adapt common schema format to OpenRouter's format
-              schema: this.adaptSchemaForOpenRouter(
-                input.parameters.jsonSchema
-              ),
-            },
-          },
-        });
-
-        if (!response.output_parsed) {
-          throw new Error("No parsed output returned from OpenRouter");
-        }
-
-        const structured = response.output_parsed as AgentStructuredResponse;
-        const message = structured.message;
-
-        return {
-          message,
-          metadata: {
-            model: response.model,
-            tokensUsed: response.usage?.total_tokens,
-            promptTokens: response.usage?.input_tokens,
-            completionTokens: response.usage?.output_tokens,
-          },
-          structured,
-        };
-      }
-
-      // Fall back to regular chat completions API if no schema provided
-      const response = await this.client.chat.completions.create(params);
-
-      const message = response.choices[0]?.message?.content || "";
-
-      let toolCalls: Array<{
-        toolName: string;
-        arguments: Record<string, unknown>;
-      }> = [];
-      if (response.choices?.[0]?.message?.tool_calls) {
-        toolCalls = response.choices[0].message.tool_calls
-          .filter((toolCall) => toolCall.type === "function")
-          .map((toolCall) => {
-            let toolCallArguments: Record<string, unknown> = {};
-            try {
-              toolCallArguments = JSON.parse(
-                toolCall.function.arguments
-              ) as Record<string, unknown>;
-            } catch (error) {
-              logger.warn(
-                `[OPENROUTER] Failed to parse tool call arguments: ${getErrorMessage(
-                  error
-                )}`
-              );
-              toolCallArguments = {};
-            }
-            return {
-              toolName: toolCall.function.name,
-              arguments: toolCallArguments,
-            };
-          });
-      }
-
-      // Only throw error if we have no text AND no function calls
-      if (!message && toolCalls.length === 0) {
-        throw new Error("No response from OpenRouter");
-      }
-
-      return {
-        message,
-        metadata: {
-          model: response.model,
-          finishReason: response.choices[0]?.finish_reason,
-          tokensUsed: response.usage?.total_tokens,
-          promptTokens: response.usage?.prompt_tokens,
-          completionTokens: response.usage?.completion_tokens,
+    super({
+      client: new OpenAI({
+        apiKey,
+        baseURL: "https://openrouter.ai/api/v1",
+        defaultHeaders: {
+          "HTTP-Referer": siteUrl || "",
+          "X-Title": siteName || "",
         },
-        structured:
-          toolCalls.length > 0
-            ? ({ message, toolCalls } as AgentStructuredResponse)
-            : undefined,
-      };
-    };
-
-    return withTimeoutAndRetry(
-      operation,
-      this.retryConfig.timeout,
-      this.retryConfig.retries,
-      `OpenRouter ${model}`
-    ) as Promise<GenerateMessageOutput<TStructured>>;
-  }
-
-  private async *generateStreamWithBackup<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): AsyncGenerator<GenerateMessageStreamChunk<TStructured>> {
-    // Try primary model first
-    try {
-      yield* this.generateStreamWithModel(this.primaryModel, input);
-    } catch (primaryError: unknown) {
-      const primaryErrMsg = getErrorMessage(primaryError);
-      logger.warn(
-        `[OPENROUTER] Primary model ${this.primaryModel} failed: ${primaryErrMsg}`
-      );
-
-      if (!shouldUseBackupModel(primaryError)) {
-        throw primaryError;
-      }
-
-      logger.debug(`[OPENROUTER] Trying backup models for streaming`);
-
-      let lastBackupError: unknown = primaryError;
-
-      for (let i = 0; i < this.backupModels.length; i++) {
-        const backupModel = this.backupModels[i];
-        logger.debug(
-          `[OPENROUTER] Trying backup model ${i + 1}/${this.backupModels.length
-          }: ${backupModel}`
-        );
-
-        try {
-          yield* this.generateStreamWithModel(backupModel, input);
-          logger.debug(`[OPENROUTER] Backup model ${backupModel} succeeded`);
-          return;
-        } catch (backupError: unknown) {
-          const backupErrMsg = getErrorMessage(backupError);
-          logger.warn(
-            `[OPENROUTER] Backup model ${backupModel} failed: ${backupErrMsg}`
-          );
-          lastBackupError = backupError;
-
-          if (
-            !shouldUseBackupModel(backupError) &&
-            i < this.backupModels.length - 1
-          ) {
-            logger.debug(
-              `[OPENROUTER] Backup model error doesn't qualify for further attempts`
-            );
-            break;
-          }
-        }
-      }
-
-      const lastBackupErrMsg = getErrorMessage(lastBackupError);
-      logger.error(
-        `[OPENROUTER] All models failed. Primary: ${primaryErrMsg}, Last backup: ${lastBackupErrMsg}`
-      );
-      throw lastBackupError;
-    }
-  }
-
-  private async *generateStreamWithModel<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    model: string,
-    input: GenerateMessageInput<TContext>
-  ): AsyncGenerator<GenerateMessageStreamChunk<TStructured>> {
-    const historyMessages = this.buildOpenRouterMessages(input.history);
-    historyMessages.push({ role: "user" as const, content: input.prompt });
-
-    const params = {
-      ...this.config,
+      }),
       model,
-      messages: historyMessages as ChatCompletionCreateParamsNonStreaming["messages"],
-      stream: true as const,
-    };
-
-    // Override with input parameters if provided
-    if (input.parameters?.maxOutputTokens !== undefined) {
-      params.max_tokens = input.parameters.maxOutputTokens;
-    }
-
-    // Add tools if provided
-    if (input.tools && input.tools.length > 0) {
-      params.tools = input.tools.map((tool) => ({
-        type: "function" as const,
-        function: {
-          name: tool.name || tool.id,
-          description: tool.description,
-          parameters: tool.parameters as FunctionParameters, // JSON schema
-        },
-      }));
-      params.tool_choice = "auto";
-    }
-
-    // Streaming path does not support responses.parse; if schema present,
-    // request JSON object and parse at the end.
-    if (input.parameters?.jsonSchema) {
-      params.response_format = { type: "json_object" };
-    }
-
-    const stream = await this.client.chat.completions.create(params);
-
-    let accumulated = "";
-    let currentModel = model;
-    let finishReason: string | undefined;
-    let promptTokens: number | undefined;
-    let completionTokens: number | undefined;
-    let totalTokens: number | undefined;
-    const toolCalls: Array<{
-      toolName: string;
-      arguments: Record<string, unknown>;
-    }> = [];
-
-    for await (const chunk of stream) {
-      currentModel = chunk.model;
-      const delta = chunk.choices[0]?.delta?.content || "";
-
-      // Extract tool calls from delta
-      if (chunk.choices[0]?.delta?.tool_calls) {
-        for (const toolCall of chunk.choices[0].delta.tool_calls) {
-          if (toolCall.function) {
-            let toolCallArguments: Record<string, unknown> = {};
-            try {
-              toolCallArguments = toolCall.function.arguments
-                ? (JSON.parse(toolCall.function.arguments) as Record<
-                  string,
-                  unknown
-                >)
-                : {};
-            } catch (error) {
-              logger.warn(
-                `[OPENROUTER] Failed to parse tool call arguments in stream: ${getErrorMessage(
-                  error
-                )}`
-              );
-              toolCallArguments = {};
-            }
-            toolCalls.push({
-              toolName: toolCall.function.name || "",
-              arguments: toolCallArguments,
-            });
-          }
-        }
-      }
-
-      if (delta) {
-        accumulated += delta;
-        yield {
-          delta,
-          accumulated,
-          done: false,
-        };
-      }
-
-      if (chunk.choices[0]?.finish_reason) {
-        finishReason = chunk.choices[0].finish_reason;
-      }
-
-      // OpenRouter may include usage in the final chunk
-      if (chunk.usage) {
-        promptTokens = chunk.usage.prompt_tokens;
-        completionTokens = chunk.usage.completion_tokens;
-        totalTokens = chunk.usage.total_tokens;
-      }
-    }
-
-    // Parse JSON response if schema was provided
-    let structured: TStructured | undefined;
-    if (input.parameters?.jsonSchema && accumulated) {
-      try {
-        structured = JSON.parse(accumulated) as TStructured;
-      } catch (error) {
-        logger.warn(
-          "[OPENROUTER] Failed to parse JSON response in stream:",
-          error
-        );
-      }
-    }
-
-    // If tools were used, include them in structured response
-    if (toolCalls.length > 0) {
-      structured = {
-        ...(structured || {}),
-        message: (structured as AgentStructuredResponse | undefined)?.message || accumulated,
-        toolCalls,
-      } as TStructured;
-    }
-
-    // Yield final chunk
-    yield {
-      delta: "",
-      accumulated,
-      done: true,
-      metadata: {
-        model: currentModel,
-        finishReason,
-        tokensUsed: totalTokens,
-        promptTokens,
-        completionTokens,
-      },
-      structured,
-    };
+      backupModels,
+      config,
+      retryConfig,
+    });
   }
 }