@falai/agent 2.7.0 → 3.1.0

package/src/providers/OpenAICompatibleProvider.ts

@@ -1,785 +1,82 @@
 /**
- * Shared base class for OpenAI-compatible chat-completions providers
- * (OpenAI, DeepSeek, OpenRouter).
+ * Shared base for OpenAI-compatible chat-completions providers (OpenAI,
+ * DeepSeek, OpenRouter, and anything else speaking the dialect).
  *
- * Handles message/history building, tool-call parsing, streaming chunk
- * handling, backup-model fallback, retry wiring, schema passthrough and
- * normalized error wrapping. Subclasses supply the configured client,
- * naming, capabilities and any genuinely provider-specific behavior via
- * the protected hooks.
+ * Everything that used to live here — message building, tool-call assembly,
+ * streaming, retry, backup models, error wrapping — is now in
+ * `@providerkit/core` and `ProviderAdapter`. What is left is the one decision
+ * this layer still owns: which wire a structured request goes out on.
  */
 
-import OpenAI from "openai";
-import type {
-  ChatCompletionChunk,
-  ChatCompletionCreateParamsNonStreaming,
-  ChatCompletionCreateParamsStreaming,
-} from "openai/resources/chat/completions";
-import { FunctionParameters } from "openai/resources/shared.mjs";
+import { createOpenAIProvider, createResponsesProvider, type Provider } from "@providerkit/core";
 
-import type {
-  AiProvider,
-  GenerateMessageInput,
-  GenerateMessageOutput,
-  GenerateMessageStreamChunk,
-  AgentStructuredResponse,
-  StructuredSchema,
-} from "../types";
-import type { ProviderCapabilities } from "../types/ai";
-import type { HistoryItem } from "../types/history";
-import type { ResponseInput } from "openai/resources/responses/responses";
-import { withTimeoutAndRetry, withStreamRetry, withBackupFallback, streamWithBackupFallback, backupFallbackLogging, resolveRetryConfig, logger, assertUsableCompletion, combineAbortSignals } from "../utils";
-import {
-  classifyProviderError,
-  getErrorMessage,
-  isBackupEligible,
-  isRetriableProviderError,
-  toProviderError,
-  type ErrorClassificationOptions,
-} from "./errorClassification";
-
-/**
- * Default request parameters shared by OpenAI-compatible providers
- */
-export type OpenAICompatibleRequestConfig = Partial<
-  Omit<ChatCompletionCreateParamsNonStreaming, "model" | "messages">
->;
+import { ProviderAdapter, type ProviderAdapterInit, type RequestConfig } from "./ProviderAdapter.js";
 
 /**
  * How structured (JSON-schema) output is requested from the endpoint:
- * - `"responses_parse"` (default) — OpenAI's native `responses.parse` API
- *   (OpenAI, OpenRouter). Most compatible endpoints do not implement it.
- * - `"json_schema"` — chat completions with a `json_schema` response_format.
- *   The broadest enforced mode (DeepSeek, Groq, Together, Fireworks, vLLM, Azure…).
- * - `"json_object"` — chat completions with a `json_object` response_format,
- *   for servers that lack json_schema enforcement (parsed at the end).
+ * - `"responses_parse"` — OpenAI's Responses API, which enforces the schema
+ *   natively. Most compatible endpoints do not implement it.
+ * - `"json_schema"` — chat completions with a `json_schema` response format.
+ *   The broadest enforced mode (DeepSeek, Groq, Together, Fireworks, vLLM…).
+ * - `"json_object"` — chat completions with plain JSON mode, for servers with
+ *   no schema enforcement at all. The caller validates.
  */
 export type StructuredOutputMode = "responses_parse" | "json_schema" | "json_object";
 
-/**
- * Initialization values supplied by subclasses
- */
-export interface OpenAICompatibleProviderInit {
-  /** Configured OpenAI SDK client (subclass sets API key, base URL, headers) */
-  client: OpenAI;
-  /** Primary model */
-  model: string;
-  /** Backup models to try if primary fails (default: []) */
-  backupModels?: string[];
-  /** Default request parameters */
-  config?: OpenAICompatibleRequestConfig;
-  /** Retry configuration */
-  retryConfig?: {
-    timeout?: number;
-    retries?: number;
-  };
-  /**
-   * How structured output is requested (default `"responses_parse"`). Lifts the
-   * strategy that used to be a per-subclass override into shared config.
-   */
+export interface OpenAICompatibleProviderInit
+  extends Omit<ProviderAdapterInit, "provider"> {
+  apiKey: string;
+  baseUrl?: string;
+  /** Names the provider in errors, and picks core's effort dialect. */
+  id: string;
+  headers?: Record<string, string>;
+  config?: RequestConfig;
   structuredOutput?: StructuredOutputMode;
+  /** OpenRouter upstream-host pin, to keep a conversation's prompt cache warm. */
+  providerOrder?: string[];
+  /** Replacement `fetch`, for tests that script the wire. */
+  fetchImpl?: typeof fetch;
 }
 
-/**
- * Base implementation of AiProvider for OpenAI-compatible APIs with
- * backup models and retry logic.
- */
-export abstract class OpenAICompatibleProvider implements AiProvider {
-  public abstract readonly name: string;
-  public abstract readonly capabilities: ProviderCapabilities;
-
-  /** Uppercase tag used in log lines, e.g. "OPENAI" */
-  protected abstract readonly logLabel: string;
-  /** Human-readable name used in retry/error messages, e.g. "OpenAI" */
-  protected abstract readonly displayName: string;
-
-  protected readonly client: OpenAI;
-  protected readonly primaryModel: string;
-  protected readonly backupModels: string[];
-  protected readonly config?: OpenAICompatibleRequestConfig;
-  protected readonly retryConfig: { timeout: number; retries: number };
-  /** Structured-output strategy; see {@link StructuredOutputMode}. */
-  protected readonly structuredOutput: StructuredOutputMode;
-
-  /**
-   * Provider-specific error classification signals.
-   * model_not_found/model_overloaded historically triggered backup models.
-   */
-  protected readonly classificationOptions: ErrorClassificationOptions = {
-    overloadedCodes: ["model_not_found", "model_overloaded"],
-  };
+export abstract class OpenAICompatibleProvider extends ProviderAdapter {
+  protected readonly config?: RequestConfig;
 
   protected constructor(init: OpenAICompatibleProviderInit) {
-    this.client = init.client;
-    this.primaryModel = init.model;
-    this.backupModels = init.backupModels ?? [];
-    this.config = init.config;
-    this.retryConfig = resolveRetryConfig(init.retryConfig);
-    this.structuredOutput = init.structuredOutput ?? "responses_parse";
-  }
-
-  // ---------------------------------------------------------------------
-  // Public API
-  // ---------------------------------------------------------------------
-
-  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);
-  }
-
-  // ---------------------------------------------------------------------
-  // Provider-specific hooks
-  // ---------------------------------------------------------------------
-
-  /**
-   * Map interaction history onto Responses-API input items so structured
-   * generations carry the same context as chat completions: plain messages
-   * pass through as easy-input messages, and tool exchanges map to
-   * function_call / function_call_output pairs keyed by the original ids.
-   * The prompt is appended as the final user turn, mirroring how the
-   * chat-completions path builds its messages array.
-   */
-  private buildResponsesInput(history: HistoryItem[], prompt: string): ResponseInput {
-    const items: ResponseInput = [];
-
-    for (const item of history) {
-      switch (item.role) {
-        case "system":
-          items.push({ role: "system", content: item.content });
-          break;
-        case "user":
-          items.push({ role: "user", content: item.content });
-          break;
-        case "assistant": {
-          if (item.content) {
-            items.push({ role: "assistant", content: item.content });
-          }
-          for (const toolCall of item.tool_calls ?? []) {
-            items.push({
-              type: "function_call",
-              call_id: toolCall.id,
-              name: toolCall.name,
-              arguments: JSON.stringify(toolCall.arguments),
-            });
-          }
-          break;
-        }
-        case "tool":
-          items.push({
-            type: "function_call_output",
-            call_id: item.tool_call_id,
-            output:
-              typeof item.content === "string"
-                ? item.content
-                : JSON.stringify(item.content),
-          });
-          break;
-      }
-    }
-
-    items.push({ role: "user", content: prompt });
-    return items;
-  }
-
-  /**
-   * Generate a structured (JSON schema) response. Default uses the
-   * responses.parse API (OpenAI, OpenRouter). Subclasses whose API lacks
-   * responses.parse (e.g. DeepSeek) override this to use chat completions
-   * with a structured response_format instead.
-   */
-  protected async executeStructuredGenerate(
-    model: string,
-    input: GenerateMessageInput<unknown>,
-    jsonSchema: StructuredSchema
-  ): Promise<GenerateMessageOutput> {
-    // Non-native modes go through chat completions, where
-    // `structuredResponseFormat` selects the response_format.
-    if (this.structuredOutput !== "responses_parse") {
-      return this.executeChatCompletion(model, input);
-    }
-
-    const response = await this.client.responses.parse(
-      {
-        model,
-        // Full conversation context: mapped history plus the prompt as the
-        // final user turn. This used to send only `instructions` with an empty
-        // `input`, leaving every structured turn without conversation memory.
-        input: this.buildResponsesInput(input.history, input.prompt),
-        reasoning: {
-          effort: input.parameters?.reasoning?.effort || "low",
-        },
-        text: {
-          format: {
-            type: "json_schema",
-            name: input.parameters?.schemaName || "structured_output",
-            // Adapt common schema format to the provider's format
-            schema: this.adaptSchema(jsonSchema),
-          },
-        },
-      },
-      { signal: input.signal }
-    );
-
-    if (!response.output_parsed) {
-      throw new Error(`No parsed output returned from ${this.displayName}`);
-    }
-
-    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,
-    };
-  }
-
-  /**
-   * response_format applied to chat-completions requests when a JSON schema is
-   * present, selected by {@link StructuredOutputMode}: `"json_schema"` for native
-   * enforcement, else a generic JSON object parsed at the end. (`"responses_parse"`
-   * never reaches chat completions for non-streaming, but streaming has no parse
-   * API, so it falls back to `json_object` here.)
-   */
-  protected structuredResponseFormat(
-    jsonSchema: StructuredSchema,
-    schemaName: string | undefined
-  ): ChatCompletionCreateParamsNonStreaming["response_format"] {
-    if (this.structuredOutput === "json_schema") {
-      return {
-        type: "json_schema" as const,
-        json_schema: {
-          name: schemaName || "structured_output",
-          schema: this.adaptSchema(jsonSchema),
-        },
-      };
-    }
-    return { type: "json_object" };
-  }
-
-  /**
-   * Hook for provider-specific streaming request parameters
-   * (e.g. DeepSeek's stream_options.include_usage).
-   */
-  protected configureStreamParams(
-    _params: ChatCompletionCreateParamsStreaming
-  ): void {
-    // Default: no extra parameters
-  }
-
-  /**
-   * Hook for provider-specific stream delta handling
-   * (e.g. DeepSeek's reasoning_content).
-   */
-  protected onStreamDelta(_delta: ChatCompletionChunk.Choice.Delta): void {
-    // Default: no extra handling
-  }
-
-  // ---------------------------------------------------------------------
-  // Shared building blocks
-  // ---------------------------------------------------------------------
-
-  /**
-   * Adapt common schema format to the provider's format.
-   * OpenAI-compatible APIs use standard JSON Schema, so this is a
-   * passthrough.
-   */
-  protected adaptSchema(schema: StructuredSchema): Record<string, unknown> {
-    // Our StructuredSchema is already JSON Schema compatible
-    return schema;
-  }
-
-  /**
-   * Build OpenAI-formatted messages from HistoryItem[] array.
-   * Maps directly to ChatCompletionMessageParam format.
-   */
-  protected buildChatMessages(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;
-  }
-
-  /**
-   * Map framework tool definitions to OpenAI function tools.
-   */
-  private buildToolParams(
-    tools: NonNullable<GenerateMessageInput<unknown>["tools"]>
-  ): ChatCompletionCreateParamsNonStreaming["tools"] {
-    return tools.map((tool) => ({
-      type: "function" as const,
-      function: {
-        name: tool.name || tool.id,
-        description: tool.description,
-        parameters: tool.parameters as FunctionParameters, // JSON schema
-      },
-    }));
-  }
-
-  /**
-   * Parse tool calls from a (complete or delta) tool_calls array,
-   * tolerating malformed JSON arguments.
-   */
-  private parseToolCallArguments(
-    rawArguments: string | undefined,
-    streaming: boolean
-  ): Record<string, unknown> {
-    if (!rawArguments) {
-      return {};
-    }
-    try {
-      return JSON.parse(rawArguments) as Record<string, unknown>;
-    } catch (error) {
-      logger.warn(
-        `[${this.logLabel}] Failed to parse tool call arguments${
-          streaming ? " in stream" : ""
-        }: ${getErrorMessage(error)}`
-      );
-      return {};
-    }
-  }
-
-  /**
-   * Determines if an error should trigger backup model usage.
-   * Derived from the normalized error classification.
-   */
-  protected shouldUseBackupModel(error: unknown): boolean {
-    return isBackupEligible(
-      classifyProviderError(error, this.classificationOptions)
-    );
-  }
-
-  /**
-   * Wrap a terminal failure in a normalized ProviderError.
-   */
-  protected wrapTerminalError(error: unknown): Error {
-    return toProviderError(error, this.name, this.classificationOptions);
-  }
-
-  // ---------------------------------------------------------------------
-  // Generation with backup models
-  // ---------------------------------------------------------------------
-
-  private async generateWithBackup<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): Promise<GenerateMessageOutput<TStructured>> {
-    const observer = backupFallbackLogging(
-      `[${this.logLabel}]`,
-      (error) => this.shouldUseBackupModel(error)
-    );
-
-    try {
-      return await withBackupFallback({
-        models: [this.primaryModel, ...this.backupModels],
-        attempt: (model) =>
-          this.generateWithModel<TContext, TStructured>(model, input),
-        shouldTryBackup: (error) => this.shouldUseBackupModel(error),
-        ...observer.callbacks,
-      });
-    } catch (error: unknown) {
-      observer.logExhausted(error);
-      throw this.wrapTerminalError(error);
-    }
-  }
-
-  private async *generateStreamWithBackup<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    input: GenerateMessageInput<TContext>
-  ): AsyncGenerator<GenerateMessageStreamChunk<TStructured>> {
-    const observer = backupFallbackLogging(
-      `[${this.logLabel}]`,
-      (error) => this.shouldUseBackupModel(error),
-      { streaming: true }
-    );
-
-    try {
-      yield* streamWithBackupFallback({
-        models: [this.primaryModel, ...this.backupModels],
-        attempt: (model) =>
-          withStreamRetry(
-            (signal) => this.generateStreamWithModel<TContext, TStructured>(model, input, signal),
-            { maxRetries: this.retryConfig.retries, firstChunkTimeoutMs: this.retryConfig.timeout, operationName: `${this.logLabel} ${model} stream`, isRetriable: (error) => isRetriableProviderError(error, this.classificationOptions) }
-          ),
-        shouldTryBackup: (error) => this.shouldUseBackupModel(error),
-        ...observer.callbacks,
-      });
-    } catch (error: unknown) {
-      observer.logExhausted(error);
-      throw this.wrapTerminalError(error);
-    }
-  }
-
-  // ---------------------------------------------------------------------
-  // Non-streaming generation
-  // ---------------------------------------------------------------------
-
-  private async generateWithModel<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    model: string,
-    input: GenerateMessageInput<TContext>
-  ): Promise<GenerateMessageOutput<TStructured>> {
-    const operation = async (signal: AbortSignal): Promise<GenerateMessageOutput> => {
-      // Thread the per-attempt timeout signal through input.signal so whichever
-      // path runs — responses.parse (structured) or chat.completions — cancels
-      // its upstream call when the deadline fires.
-      const reqInput = { ...input, signal: combineAbortSignals(input.signal, signal) };
-      // Use structured output path if JSON schema is provided
-      if (reqInput.parameters?.jsonSchema) {
-        return this.executeStructuredGenerate(
-          model,
-          reqInput,
-          reqInput.parameters.jsonSchema
-        );
-      }
-      return this.executeChatCompletion(model, reqInput);
-    };
-
-    return withTimeoutAndRetry(
-      operation,
-      this.retryConfig.timeout,
-      this.retryConfig.retries,
-      `${this.displayName} ${model}`,
-      (error) => isRetriableProviderError(error, this.classificationOptions)
-    ) as Promise<GenerateMessageOutput<TStructured>>;
-  }
-
-  /**
-   * Non-streaming chat-completions request: builds messages/tools,
-   * extracts text and tool calls, and merges structured output.
-   */
-  protected async executeChatCompletion(
-    model: string,
-    input: GenerateMessageInput<unknown>
-  ): Promise<GenerateMessageOutput> {
-    const historyMessages = this.buildChatMessages(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 = this.buildToolParams(input.tools);
-      params.tool_choice = "auto";
-    }
-
-    // Use structured output format if JSON schema is provided
-    if (input.parameters?.jsonSchema) {
-      params.response_format = this.structuredResponseFormat(
-        input.parameters.jsonSchema,
-        input.parameters.schemaName
-      );
-    }
-
-    const response = await this.client.chat.completions.create(params, {
-      signal: input.signal,
+    const mode = init.structuredOutput ?? "responses_parse";
+    super({
+      provider: buildProvider(init, mode),
+      model: init.model,
+      ...(init.config ? { defaults: init.config } : {}),
+      ...(init.backupModels ? { backupModels: init.backupModels } : {}),
+      ...(init.retryConfig ? { retryConfig: init.retryConfig } : {}),
     });
-
-    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) => ({
-          toolName: toolCall.function.name,
-          arguments: this.parseToolCallArguments(
-            toolCall.function.arguments,
-            false
-          ),
-        }));
-    }
-
-    // Only throw error if we have no text AND no function calls
-    if (!message && toolCalls.length === 0) {
-      throw new Error(`No response from ${this.displayName}`);
-    }
-
-    // Parse structured output if schema was provided
-    let structured: AgentStructuredResponse | undefined;
-    if (input.parameters?.jsonSchema && message) {
-      try {
-        structured = JSON.parse(message) as AgentStructuredResponse;
-      } catch (error) {
-        logger.warn(`[${this.logLabel}] Failed to parse JSON response:`, error);
-      }
-    }
-
-    // If tools were used, include them in structured response
-    if (toolCalls.length > 0) {
-      structured = {
-        ...(structured || {}),
-        message: structured?.message || message,
-        toolCalls,
-      } as AgentStructuredResponse;
-    }
-
-    // A parsed-but-blank message with no tool calls is as empty as no text;
-    // the shared guard throws so withTimeoutAndRetry retries instead of
-    // returning {"message":""}.
-    assertUsableCompletion(structured, message, toolCalls.length, this.displayName);
-
-    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,
-      },
-      structured,
-    };
+    this.config = init.config;
   }
+}
 
-  // ---------------------------------------------------------------------
-  // Streaming generation
-  // ---------------------------------------------------------------------
-
-  private async *generateStreamWithModel<
-    TContext = unknown,
-    TStructured = AgentStructuredResponse
-  >(
-    model: string,
-    input: GenerateMessageInput<TContext>,
-    attemptSignal?: AbortSignal
-  ): AsyncGenerator<GenerateMessageStreamChunk<TStructured>> {
-    // Build messages from history and append prompt as final user message
-    const historyMessages = this.buildChatMessages(input.history);
-    historyMessages.push({ role: "user" as const, content: input.prompt });
-
-    const params: ChatCompletionCreateParamsStreaming = {
-      ...this.config,
-      model,
-      messages:
-        historyMessages as ChatCompletionCreateParamsNonStreaming["messages"],
-      stream: true as const,
-    };
-    this.configureStreamParams(params);
-
-    // 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 = this.buildToolParams(input.tools);
-      params.tool_choice = "auto";
-    }
-
-    // Streaming path does not support responses.parse; if schema present,
-    // request structured output and parse at the end.
-    if (input.parameters?.jsonSchema) {
-      params.response_format = this.structuredResponseFormat(
-        input.parameters.jsonSchema,
-        input.parameters.schemaName
-      );
-    }
-
-    const stream = await this.client.chat.completions.create(params, {
-      signal: combineAbortSignals(input.signal, attemptSignal),
+function buildProvider(init: OpenAICompatibleProviderInit, mode: StructuredOutputMode): Provider {
+  if (mode === "responses_parse") {
+    return createResponsesProvider({
+      apiKey: init.apiKey,
+      model: init.model,
+      id: init.id,
+      ...(init.baseUrl ? { baseUrl: `${init.baseUrl}` } : {}),
+      ...(init.headers ? { headers: init.headers } : {}),
+      ...(init.config?.maxTokens ? { maxTokens: init.config.maxTokens } : {}),
+      ...(init.config?.effort ? { effort: init.config.effort } : {}),
+      ...(init.fetchImpl ? { fetchImpl: init.fetchImpl } : {}),
     });
-
-    let accumulated = "";
-    let currentModel = model;
-    let finishReason: string | undefined;
-    let promptTokens: number | undefined;
-    let completionTokens: number | undefined;
-    let totalTokens: number | undefined;
-    // Tool calls arrive across MANY deltas: the first carries the id/name,
-    // later ones carry argument-string fragments that must be concatenated.
-    // Accumulate per call (keyed by `index`) and parse once at stream end —
-    // parsing per fragment produced N broken entries for one real call.
-    const pendingToolCalls = new Map<
-      number,
-      { toolName: string; argumentsBuffer: string }
-    >();
-    let lastSlot = 0;
-
-    for await (const chunk of stream) {
-      currentModel = chunk.model;
-
-      // Some providers include usage in (possibly choice-less) chunks
-      if (chunk.usage) {
-        promptTokens = chunk.usage.prompt_tokens;
-        completionTokens = chunk.usage.completion_tokens;
-        totalTokens = chunk.usage.total_tokens;
-      }
-
-      const choice = chunk.choices?.[0];
-      if (!choice) continue;
-
-      // Accumulate streamed tool-call fragments by index
-      if (choice.delta?.tool_calls) {
-        for (const toolCall of choice.delta.tool_calls) {
-          // Spec-compliant chunks carry `index`; servers that omit it send
-          // one call's fragments sequentially, so keep appending to lastSlot.
-          const slot =
-            typeof toolCall.index === "number" ? toolCall.index : lastSlot;
-          lastSlot = slot;
-          const entry = pendingToolCalls.get(slot) ?? {
-            toolName: "",
-            argumentsBuffer: "",
-          };
-          if (toolCall.function?.name) {
-            entry.toolName = toolCall.function.name;
-          }
-          if (toolCall.function?.arguments) {
-            entry.argumentsBuffer += toolCall.function.arguments;
-          }
-          pendingToolCalls.set(slot, entry);
-        }
-      }
-
-      // Provider-specific delta handling (e.g. DeepSeek reasoning_content)
-      this.onStreamDelta(choice.delta);
-
-      const delta = choice.delta?.content || "";
-      if (delta) {
-        accumulated += delta;
-        yield {
-          delta,
-          accumulated,
-          done: false,
-        };
-      }
-
-      if (choice.finish_reason) {
-        finishReason = choice.finish_reason;
-      }
-    }
-
-    // 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(
-          `[${this.logLabel}] Failed to parse JSON response in stream:`,
-          error
-        );
-      }
-    }
-
-    const toolCalls: Array<{
-      toolName: string;
-      arguments: Record<string, unknown>;
-    }> = [];
-    for (const [, entry] of [...pendingToolCalls.entries()].sort(
-      ([a], [b]) => a - b
-    )) {
-      if (!entry.toolName) continue;
-      toolCalls.push({
-        toolName: entry.toolName,
-        arguments: this.parseToolCallArguments(entry.argumentsBuffer, true),
-      });
-    }
-
-    // Include tool calls in structured response (even without JSON schema)
-    if (toolCalls.length > 0) {
-      structured = {
-        ...(structured || {}),
-        message:
-          (structured as AgentStructuredResponse | undefined)?.message ||
-          accumulated,
-        toolCalls,
-      } as TStructured;
-    }
-
-    // Empty-completion guard — same definition as the non-streaming path, so the
-    // stream retries / falls back to backup instead of emitting an empty message.
-    assertUsableCompletion(structured, accumulated, toolCalls.length, this.displayName);
-
-    // Yield final chunk
-    yield {
-      delta: "",
-      accumulated,
-      done: true,
-      metadata: {
-        model: currentModel,
-        finishReason,
-        tokensUsed: totalTokens,
-        promptTokens,
-        completionTokens,
-      },
-      structured,
-    };
   }
+  return createOpenAIProvider({
+    apiKey: init.apiKey,
+    model: init.model,
+    id: init.id,
+    jsonMode: mode === "json_schema" ? "schema" : "object",
+    ...(init.baseUrl ? { baseUrl: init.baseUrl } : {}),
+    ...(init.headers ? { headers: init.headers } : {}),
+    ...(init.config?.maxTokens ? { maxTokens: init.config.maxTokens } : {}),
+    ...(init.config?.effort ? { effort: init.config.effort } : {}),
+    ...(init.providerOrder ? { providerOrder: init.providerOrder } : {}),
+    ...(init.fetchImpl ? { fetchImpl: init.fetchImpl } : {}),
+  });
 }