@vybestack/llxprt-code-core 0.6.1-nightly.251202.1e208436b → 0.6.1-nightly.251204.b119e390d

package/dist/src/providers/openai-vercel/OpenAIVercelProvider.js

@@ -0,0 +1,943 @@
+/**
+ * Copyright 2025 Vybestack LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+/**
+ * @plan PLAN-20250218-STATELESSPROVIDER.P04
+ * @requirement REQ-SP-001
+ *
+ * OpenAI provider implemented on top of Vercel AI SDK v5, using the
+ * OpenAI chat completions API via @ai-sdk/openai + ai.
+ */
+import crypto from 'node:crypto';
+import * as Ai from 'ai';
+import { createOpenAI } from '@ai-sdk/openai';
+import { BaseProvider, } from '../BaseProvider.js';
+import { DebugLogger } from '../../debug/index.js';
+import { ToolFormatter } from '../../tools/ToolFormatter.js';
+import { processToolParameters } from '../../tools/doubleEscapeUtils.js';
+import { getCoreSystemPromptAsync } from '../../core/prompts.js';
+import { resolveUserMemory } from '../utils/userMemory.js';
+import { convertToVercelMessages } from './messageConversion.js';
+import { resolveRuntimeAuthToken } from '../utils/authToken.js';
+import { filterOpenAIRequestParams } from '../openai/openaiRequestParams.js';
+import { isLocalEndpoint } from '../utils/localEndpoint.js';
+import { AuthenticationError, wrapError } from './errors.js';
+const streamText = Ai.streamText;
+const generateText = Ai.generateText;
+/**
+ * Vercel OpenAI-based provider using AI SDK v5.
+ *
+ * NOTE:
+ * - No dependency on the official `openai` SDK.
+ * - Uses `openai.chat(modelId)` to talk to the Chat Completions API.
+ * - Tools are configured via AI SDK `tool()` with JSON schema input.
+ */
+export class OpenAIVercelProvider extends BaseProvider {
+    getLogger() {
+        return new DebugLogger('llxprt:provider:openaivercel');
+    }
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Constructor reduced to minimal initialization - no state captured.
+     */
+    constructor(apiKey, baseURL, config, oauthManager) {
+        // Normalize empty string to undefined for proper precedence handling
+        const normalizedApiKey = apiKey && apiKey.trim() !== '' ? apiKey : undefined;
+        super({
+            name: 'openaivercel',
+            apiKey: normalizedApiKey,
+            baseURL,
+            envKeyNames: ['OPENAI_API_KEY'],
+            // AI SDK-based provider does not use OAuth directly here.
+            isOAuthEnabled: false,
+            oauthProvider: undefined,
+            oauthManager,
+        }, config);
+    }
+    supportsOAuth() {
+        return false;
+    }
+    /**
+     * Create an OpenAI provider instance for this call using AI SDK v5.
+     *
+     * Uses the resolved runtime auth token and baseURL, and still allows
+     * local endpoints without authentication (for Ollama-style servers).
+     */
+    async createOpenAIClient(options) {
+        const authToken = (await resolveRuntimeAuthToken(options.resolved.authToken)) ?? '';
+        const baseURL = options.resolved.baseURL ?? this.baseProviderConfig.baseURL;
+        // Allow local endpoints without authentication
+        if (!authToken && !isLocalEndpoint(baseURL)) {
+            throw new AuthenticationError(`Auth token unavailable for runtimeId=${options.runtime?.runtimeId} (REQ-SP4-003).`, this.name);
+        }
+        const headers = this.getCustomHeaders();
+        return createOpenAI({
+            apiKey: authToken || undefined,
+            baseURL: baseURL || undefined,
+            headers: headers || undefined,
+        });
+    }
+    /**
+     * Extract model parameters from normalized options instead of settings service.
+     * This mirrors OpenAIProvider but feeds AI SDK call options instead.
+     */
+    extractModelParamsFromOptions(options) {
+        const providerSettings = options.settings?.getProviderSettings(this.name) ?? {};
+        const configEphemerals = options.invocation?.ephemerals ?? {};
+        const filteredProviderParams = filterOpenAIRequestParams(providerSettings);
+        const filteredEphemeralParams = filterOpenAIRequestParams(configEphemerals);
+        if (!filteredProviderParams && !filteredEphemeralParams) {
+            return undefined;
+        }
+        return {
+            ...(filteredProviderParams ?? {}),
+            ...(filteredEphemeralParams ?? {}),
+        };
+    }
+    /**
+     * Tool formatter instances cannot be shared between stateless calls,
+     * so construct a fresh one for every invocation.
+     */
+    createToolFormatter() {
+        return new ToolFormatter();
+    }
+    getAiJsonSchema() {
+        try {
+            const candidate = Ai.jsonSchema;
+            return typeof candidate === 'function'
+                ? candidate
+                : undefined;
+        }
+        catch {
+            return undefined;
+        }
+    }
+    getAiTool() {
+        try {
+            const candidate = Ai.tool;
+            return typeof candidate === 'function'
+                ? candidate
+                : undefined;
+        }
+        catch {
+            return undefined;
+        }
+    }
+    /**
+     * Normalize tool IDs from various formats to OpenAI-style format.
+     * Kept for compatibility with existing history/tool logic.
+     */
+    normalizeToOpenAIToolId(id) {
+        const sanitize = (value) => value.replace(/[^a-zA-Z0-9_]/g, '') ||
+            'call_' + crypto.randomUUID().replace(/-/g, '');
+        // If already in OpenAI format, return as-is
+        if (id.startsWith('call_')) {
+            return sanitize(id);
+        }
+        // For history format, extract the UUID and add OpenAI prefix
+        if (id.startsWith('hist_tool_')) {
+            const uuid = id.substring('hist_tool_'.length);
+            return sanitize('call_' + uuid);
+        }
+        // For Anthropic format, extract the UUID and add OpenAI prefix
+        if (id.startsWith('toolu_')) {
+            const uuid = id.substring('toolu_'.length);
+            return sanitize('call_' + uuid);
+        }
+        // Unknown format - assume it's a raw UUID
+        return sanitize('call_' + id);
+    }
+    /**
+     * Normalize tool IDs from OpenAI-style format to history format.
+     */
+    normalizeToHistoryToolId(id) {
+        // If already in history format, return as-is
+        if (id.startsWith('hist_tool_')) {
+            return id;
+        }
+        // For OpenAI format, extract the UUID and add history prefix
+        if (id.startsWith('call_')) {
+            const uuid = id.substring('call_'.length);
+            return 'hist_tool_' + uuid;
+        }
+        // For Anthropic format, extract the UUID and add history prefix
+        if (id.startsWith('toolu_')) {
+            const uuid = id.substring('toolu_'.length);
+            return 'hist_tool_' + uuid;
+        }
+        // Unknown format - assume it's a raw UUID
+        return 'hist_tool_' + id;
+    }
+    /**
+     * Convert internal history IContent[] to AI SDK ModelMessage[].
+     *
+     * This implementation uses textual tool replay for past tool calls/results.
+     * New tool calls in the current response still use structured ToolCallBlocks.
+     */
+    convertToModelMessages(contents) {
+        return convertToVercelMessages(contents);
+    }
+    /**
+     * Build an AI SDK ToolSet from already-normalized OpenAI-style tool definitions.
+     *
+     * Input is the same array produced by ToolFormatter.convertGeminiToFormat(…, 'openai'|'qwen').
+     */
+    buildVercelTools(formattedTools) {
+        if (!formattedTools || formattedTools.length === 0) {
+            return undefined;
+        }
+        const jsonSchemaFn = this.getAiJsonSchema() ??
+            ((schema) => schema);
+        const toolFn = this.getAiTool() ??
+            ((config) => config);
+        const toolsRecord = {};
+        for (const t of formattedTools) {
+            if (!t || t.type !== 'function')
+                continue;
+            const fn = t.function;
+            if (!fn?.name)
+                continue;
+            if (toolsRecord[fn.name])
+                continue;
+            const inputSchema = fn.parameters
+                ? jsonSchemaFn(fn.parameters)
+                : jsonSchemaFn({
+                    type: 'object',
+                    properties: {},
+                    additionalProperties: false,
+                });
+            toolsRecord[fn.name] = toolFn({
+                description: fn.description,
+                inputSchema,
+                // No execute() – we only surface tool calls back to the caller,
+                // execution is handled by the existing external tool pipeline.
+            });
+        }
+        return Object.keys(toolsRecord).length > 0 ? toolsRecord : undefined;
+    }
+    mapUsageToMetadata(usage) {
+        if (!usage)
+            return undefined;
+        const promptTokens = usage.inputTokens ??
+            usage.promptTokens ??
+            0;
+        const completionTokens = usage.outputTokens ??
+            usage.completionTokens ??
+            0;
+        const totalTokens = usage.totalTokens ??
+            (typeof promptTokens === 'number' && typeof completionTokens === 'number'
+                ? promptTokens + completionTokens
+                : 0);
+        return {
+            promptTokens,
+            completionTokens,
+            totalTokens,
+        };
+    }
+    /**
+     * Get a short preview of a message's content for debug logging.
+     */
+    getContentPreview(content, maxLength = 200) {
+        if (content === null || content === undefined) {
+            return undefined;
+        }
+        if (typeof content === 'string') {
+            if (content.length <= maxLength) {
+                return content;
+            }
+            return `${content.slice(0, maxLength)}…`;
+        }
+        if (Array.isArray(content)) {
+            // text parts, tool-call parts, etc.
+            const textParts = content.map((part) => {
+                if (typeof part === 'object' &&
+                    part !== null &&
+                    'type' in part &&
+                    part.type === 'text') {
+                    return part.text ?? '';
+                }
+                try {
+                    return JSON.stringify(part);
+                }
+                catch {
+                    return '[unserializable part]';
+                }
+            });
+            const joined = textParts.join('\n');
+            if (joined.length <= maxLength) {
+                return joined;
+            }
+            return `${joined.slice(0, maxLength)}…`;
+        }
+        try {
+            const serialized = JSON.stringify(content);
+            if (serialized.length <= maxLength) {
+                return serialized;
+            }
+            return `${serialized.slice(0, maxLength)}…`;
+        }
+        catch {
+            return '[unserializable content]';
+        }
+    }
+    /**
+     * Core chat completion implementation using AI SDK v5.
+     *
+     * This replaces the original OpenAI SDK v5 client usage with:
+     *   - createOpenAI({ apiKey, baseURL })
+     *   - openai.chat(modelId)
+     *   - generateText / streamText
+     */
+    async *generateChatCompletionWithOptions(options) {
+        const logger = this.getLogger();
+        const { contents, tools, metadata } = options;
+        const modelId = options.resolved.model || this.getDefaultModel();
+        const abortSignal = metadata?.abortSignal;
+        const ephemerals = options.invocation?.ephemerals ?? {};
+        const toolFormatter = this.createToolFormatter();
+        const resolved = options.resolved;
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIVercelProvider] Resolved request context`, {
+                provider: this.name,
+                model: modelId,
+                resolvedModel: resolved.model,
+                resolvedBaseUrl: resolved.baseURL,
+                authTokenPresent: Boolean(resolved.authToken),
+                messageCount: contents.length,
+                toolCount: tools?.length ?? 0,
+                metadataKeys: Object.keys(metadata ?? {}),
+            });
+        }
+        // Determine streaming vs non-streaming mode (default: enabled)
+        const streamingSetting = ephemerals['streaming'];
+        const streamingResolved = options.resolved?.streaming;
+        const streamingEnabled = streamingResolved === false
+            ? false
+            : streamingResolved === true
+                ? true
+                : streamingSetting !== 'disabled';
+        // System prompt (same core-prompt mechanism as OpenAIProvider)
+        const flattenedToolNames = tools?.flatMap((group) => group.functionDeclarations
+            .map((decl) => decl.name)
+            .filter((name) => !!name)) ?? [];
+        const toolNamesArg = tools === undefined ? undefined : Array.from(new Set(flattenedToolNames));
+        const userMemory = await resolveUserMemory(options.userMemory, () => options.invocation?.userMemory);
+        const systemPrompt = await getCoreSystemPromptAsync(userMemory, modelId, toolNamesArg);
+        // Convert internal history to AI SDK ModelMessages with structured tool replay.
+        const messages = this.convertToModelMessages(contents);
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIVercelProvider] Chat payload snapshot`, {
+                messageCount: messages.length,
+                messages: messages.map((msg) => ({
+                    role: msg.role,
+                    contentPreview: this.getContentPreview(msg.content),
+                })),
+            });
+        }
+        // Detect tool format ('openai' or 'qwen') and convert Gemini tools to OpenAI-style definitions
+        const detectedFormat = this.detectToolFormat();
+        const formattedTools = toolFormatter.convertGeminiToFormat(tools, detectedFormat);
+        if (logger.enabled && formattedTools) {
+            logger.debug(() => `[OpenAIVercelProvider] Tool conversion summary`, {
+                detectedFormat,
+                hasTools: !!formattedTools,
+                toolCount: formattedTools.length,
+                toolNames: formattedTools.map((t) => t.function.name),
+            });
+        }
+        // Build AI SDK ToolSet
+        const aiTools = this.buildVercelTools(formattedTools);
+        // Model parameters (temperature, top_p, etc.)
+        const modelParams = this.extractModelParamsFromOptions(options) ?? {};
+        const maxTokensMeta = metadata?.maxTokens ??
+            ephemerals['max-tokens'];
+        const maxTokensOverride = modelParams['max_tokens'] ?? undefined;
+        const maxOutputTokens = typeof maxTokensMeta === 'number' && Number.isFinite(maxTokensMeta)
+            ? maxTokensMeta
+            : typeof maxTokensOverride === 'number' &&
+                Number.isFinite(maxTokensOverride)
+                ? maxTokensOverride
+                : undefined;
+        const temperature = modelParams['temperature'];
+        const topP = modelParams['top_p'];
+        const presencePenalty = modelParams['presence_penalty'];
+        const frequencyPenalty = modelParams['frequency_penalty'];
+        const stopSetting = modelParams['stop'];
+        const stopSequences = typeof stopSetting === 'string'
+            ? [stopSetting]
+            : Array.isArray(stopSetting)
+                ? stopSetting
+                : undefined;
+        const seed = modelParams['seed'];
+        const maxRetries = ephemerals['retries'] ?? 2; // AI SDK default is 2
+        // Instantiate AI SDK OpenAI provider + chat model
+        const openaiProvider = await this.createOpenAIClient(options);
+        const providerWithChat = openaiProvider;
+        const model = (providerWithChat.chat
+            ? providerWithChat.chat(modelId)
+            : providerWithChat(modelId));
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIVercelProvider] Sending chat request`, {
+                model: modelId,
+                baseURL: resolved.baseURL ?? this.getBaseURL(),
+                streamingEnabled,
+                hasTools: !!aiTools,
+                toolCount: aiTools ? Object.keys(aiTools).length : 0,
+                maxOutputTokens,
+            });
+        }
+        if (streamingEnabled) {
+            // Streaming mode via streamText()
+            const streamOptions = {
+                model,
+                system: systemPrompt,
+                messages,
+                tools: aiTools,
+                maxOutputTokens,
+                temperature,
+                topP,
+                presencePenalty,
+                frequencyPenalty,
+                stopSequences,
+                seed,
+                maxRetries,
+                abortSignal,
+            };
+            if (maxOutputTokens !== undefined) {
+                streamOptions['maxTokens'] = maxOutputTokens;
+            }
+            let result;
+            try {
+                result = await streamText(streamOptions);
+            }
+            catch (error) {
+                logger.error(() => `[OpenAIVercelProvider] streamText failed: ${error instanceof Error ? error.message : String(error)}`, { error });
+                throw wrapError(error, this.name);
+            }
+            const collectedToolCalls = [];
+            let totalUsage;
+            let finishReason;
+            const hasFullStream = result &&
+                typeof result === 'object' &&
+                'fullStream' in result;
+            if (hasFullStream && result.fullStream) {
+                try {
+                    for await (const part of result.fullStream) {
+                        if (abortSignal?.aborted) {
+                            break;
+                        }
+                        switch (part.type) {
+                            case 'text-delta': {
+                                const text = typeof part.text === 'string' ? part.text : '';
+                                if (text) {
+                                    const content = {
+                                        speaker: 'ai',
+                                        blocks: [
+                                            {
+                                                type: 'text',
+                                                text,
+                                            },
+                                        ],
+                                    };
+                                    yield content;
+                                }
+                                break;
+                            }
+                            case 'tool-call': {
+                                // Single completed tool call with already-parsed input
+                                if (part.toolCallId && part.toolName) {
+                                    collectedToolCalls.push({
+                                        toolCallId: String(part.toolCallId),
+                                        toolName: String(part.toolName),
+                                        input: part.input,
+                                    });
+                                }
+                                break;
+                            }
+                            case 'finish': {
+                                totalUsage = part.totalUsage;
+                                finishReason = part.finishReason;
+                                if (logger.enabled) {
+                                    logger.debug(() => `[OpenAIVercelProvider] streamText finished with reason: ${part.finishReason}`, {
+                                        finishReason: part.finishReason,
+                                        hasUsage: !!totalUsage,
+                                        toolCallCount: collectedToolCalls.length,
+                                    });
+                                }
+                                break;
+                            }
+                            case 'error': {
+                                throw part.error ?? new Error('Streaming error from AI SDK');
+                            }
+                            default:
+                                // Ignore other parts: reasoning, source, start-step, finish-step, etc.
+                                break;
+                        }
+                    }
+                }
+                catch (error) {
+                    if (abortSignal?.aborted ||
+                        (error &&
+                            typeof error === 'object' &&
+                            'name' in error &&
+                            error.name === 'AbortError')) {
+                        logger.debug(() => `[OpenAIVercelProvider] Streaming response cancelled by AbortSignal`);
+                        throw error;
+                    }
+                    logger.error(() => `[OpenAIVercelProvider] Error processing streaming response: ${error instanceof Error ? error.message : String(error)}`, { error });
+                    throw wrapError(error, this.name);
+                }
+            }
+            else {
+                const legacyStream = result;
+                try {
+                    if (legacyStream.textStream) {
+                        for await (const textChunk of legacyStream.textStream) {
+                            if (!textChunk) {
+                                continue;
+                            }
+                            yield {
+                                speaker: 'ai',
+                                blocks: [
+                                    {
+                                        type: 'text',
+                                        text: textChunk,
+                                    },
+                                ],
+                            };
+                        }
+                    }
+                }
+                catch (error) {
+                    if (abortSignal?.aborted ||
+                        (error &&
+                            typeof error === 'object' &&
+                            'name' in error &&
+                            error.name === 'AbortError')) {
+                        throw error;
+                    }
+                    logger.error(() => `[OpenAIVercelProvider] Legacy streaming response failed: ${error instanceof Error ? error.message : String(error)}`, { error });
+                    throw wrapError(error, this.name);
+                }
+                const legacyToolCalls = (legacyStream.toolCalls
+                    ? await legacyStream.toolCalls.catch(() => [])
+                    : []) ?? [];
+                for (const call of legacyToolCalls) {
+                    collectedToolCalls.push({
+                        toolCallId: String(call.toolCallId ?? crypto.randomUUID()),
+                        toolName: String(call.toolName ?? 'unknown_tool'),
+                        input: call.input,
+                    });
+                }
+                totalUsage = legacyStream.usage
+                    ? await legacyStream.usage.catch(() => undefined)
+                    : undefined;
+                finishReason = legacyStream.finishReason
+                    ? await legacyStream.finishReason.catch(() => undefined)
+                    : undefined;
+            }
+            // Emit accumulated tool calls as a single IContent, with usage metadata if available
+            if (collectedToolCalls.length > 0) {
+                const blocks = collectedToolCalls.map((call) => {
+                    let argsString = '{}';
+                    try {
+                        argsString =
+                            typeof call.input === 'string'
+                                ? call.input
+                                : JSON.stringify(call.input ?? {});
+                    }
+                    catch {
+                        argsString = '{}';
+                    }
+                    const processedParameters = processToolParameters(argsString, call.toolName);
+                    return {
+                        type: 'tool_call',
+                        id: this.normalizeToHistoryToolId(this.normalizeToOpenAIToolId(call.toolCallId)),
+                        name: call.toolName,
+                        parameters: processedParameters,
+                    };
+                });
+                const usageMeta = this.mapUsageToMetadata(totalUsage);
+                const metadata = usageMeta || finishReason
+                    ? {
+                        ...(usageMeta ? { usage: usageMeta } : {}),
+                        ...(finishReason ? { finishReason } : {}),
+                    }
+                    : undefined;
+                const toolContent = {
+                    speaker: 'ai',
+                    blocks,
+                    ...(metadata ? { metadata } : {}),
+                };
+                yield toolContent;
+            }
+            else {
+                // Emit metadata-only message so callers can see usage/finish reason
+                const usageMeta = this.mapUsageToMetadata(totalUsage);
+                const metadata = usageMeta || finishReason
+                    ? {
+                        ...(usageMeta ? { usage: usageMeta } : {}),
+                        ...(finishReason ? { finishReason } : {}),
+                    }
+                    : undefined;
+                if (metadata) {
+                    yield {
+                        speaker: 'ai',
+                        blocks: [],
+                        metadata,
+                    };
+                }
+            }
+        }
+        else {
+            // Non-streaming mode via generateText()
+            let result;
+            try {
+                const aiToolFn = this.getAiTool();
+                const toolsForGenerate = (!aiToolFn && formattedTools ? formattedTools : aiTools) ?? undefined;
+                const generateOptions = {
+                    model,
+                    system: systemPrompt,
+                    messages,
+                    tools: toolsForGenerate,
+                    maxOutputTokens,
+                    temperature,
+                    topP,
+                    presencePenalty,
+                    frequencyPenalty,
+                    stopSequences,
+                    seed,
+                    maxRetries,
+                    abortSignal,
+                };
+                if (maxOutputTokens !== undefined) {
+                    generateOptions['maxTokens'] = maxOutputTokens;
+                }
+                result = await generateText(generateOptions);
+            }
+            catch (error) {
+                logger.error(() => `[OpenAIVercelProvider] Non-streaming chat completion failed: ${error instanceof Error ? error.message : String(error)}`, { error });
+                throw wrapError(error, this.name);
+            }
+            const blocks = [];
+            if (result.text) {
+                blocks.push({
+                    type: 'text',
+                    text: result.text,
+                });
+            }
+            // Typed tool calls from AI SDK; execution is not automatic because we did not provide execute().
+            const toolCalls = 'toolCalls' in result && result.toolCalls ? await result.toolCalls : [];
+            for (const call of toolCalls) {
+                const toolName = call.toolName ?? 'unknown_tool';
+                const id = call.toolCallId ?? crypto.randomUUID();
+                const rawInput = call.input ??
+                    call.args ??
+                    call.arguments;
+                let argsString = '{}';
+                try {
+                    argsString =
+                        typeof rawInput === 'string'
+                            ? rawInput
+                            : JSON.stringify(rawInput ?? {});
+                }
+                catch {
+                    argsString = '{}';
+                }
+                const processedParameters = processToolParameters(argsString, toolName);
+                blocks.push({
+                    type: 'tool_call',
+                    id: this.normalizeToHistoryToolId(this.normalizeToOpenAIToolId(id)),
+                    name: toolName,
+                    parameters: processedParameters,
+                });
+            }
+            if (blocks.length > 0 || result.usage) {
+                const usageMeta = this.mapUsageToMetadata(result.usage);
+                const content = {
+                    speaker: 'ai',
+                    blocks,
+                    ...(usageMeta
+                        ? {
+                            metadata: {
+                                usage: usageMeta,
+                            },
+                        }
+                        : {}),
+                };
+                yield content;
+            }
+        }
+    }
+    /**
+     * Models listing – uses HTTP GET /models via fetch instead of the OpenAI SDK.
+     * Falls back to a small static list if the request fails.
+     */
+    async getModels() {
+        const logger = this.getLogger();
+        try {
+            const authToken = await this.getAuthToken();
+            const baseURL = this.getBaseURL() ?? 'https://api.openai.com/v1';
+            const url = baseURL.endsWith('/') || baseURL.endsWith('\\')
+                ? `${baseURL}models`
+                : `${baseURL}/models`;
+            const headers = {
+                ...(this.getCustomHeaders() ?? {}),
+            };
+            if (authToken) {
+                headers['Authorization'] = `Bearer ${authToken}`;
+            }
+            const res = await fetch(url, {
+                headers,
+            });
+            if (!res.ok) {
+                throw new Error(`HTTP ${res.status}`);
+            }
+            const data = (await res.json());
+            const models = [];
+            for (const model of data.data ?? []) {
+                // Filter out non-chat models (embeddings, audio, image, etc.)
+                if (!/embedding|whisper|audio|tts|image|vision|dall[- ]?e|moderation/i.test(model.id)) {
+                    const contextWindow = model.context_window ??
+                        model.contextWindow;
+                    models.push({
+                        id: model.id,
+                        name: model.name ?? model.id,
+                        provider: this.name,
+                        supportedToolFormats: ['openai'],
+                        ...(typeof contextWindow === 'number'
+                            ? { contextWindow }
+                            : undefined),
+                    });
+                }
+            }
+            const sortedModels = models.length > 0
+                ? models.sort((a, b) => a.name.localeCompare(b.name))
+                : this.getFallbackModels();
+            return sortedModels;
+        }
+        catch (error) {
+            logger.debug(() => `Error fetching models from OpenAI via Vercel provider: ${error}`);
+            return this.getFallbackModels();
+        }
+    }
+    getFallbackModels() {
+        const providerName = this.name;
+        const models = [
+            {
+                id: 'gpt-3.5-turbo',
+                name: 'GPT-3.5 Turbo',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 16385,
+            },
+            {
+                id: 'gpt-4',
+                name: 'GPT-4',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 8192,
+            },
+            {
+                id: 'gpt-4-turbo',
+                name: 'GPT-4 Turbo',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 128000,
+            },
+            {
+                id: 'gpt-4o',
+                name: 'GPT-4o',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 128000,
+            },
+            {
+                id: 'gpt-4o-mini',
+                name: 'GPT-4o Mini',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 128000,
+            },
+            {
+                id: 'o1-mini',
+                name: 'o1-mini',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 128000,
+            },
+            {
+                id: 'o1-preview',
+                name: 'o1-preview',
+                provider: providerName,
+                supportedToolFormats: ['openai'],
+                contextWindow: 128000,
+            },
+        ];
+        return models.sort((a, b) => a.name.localeCompare(b.name));
+    }
+    getDefaultModel() {
+        const baseURL = this.getBaseURL();
+        if (baseURL &&
+            (baseURL.includes('qwen') || baseURL.includes('dashscope'))) {
+            return process.env.LLXPRT_DEFAULT_MODEL || 'qwen3-coder-plus';
+        }
+        return process.env.LLXPRT_DEFAULT_MODEL || 'gpt-4o';
+    }
+    getCurrentModel() {
+        return this.getModel();
+    }
+    // No client caching for AI SDK provider – kept as no-op for compatibility.
+    clearClientCache(runtimeKey) {
+        void runtimeKey;
+    }
+    clearState() {
+        this.clearClientCache();
+        this.clearAuthCache();
+    }
+    getServerTools() {
+        return [];
+    }
+    async invokeServerTool(toolName, _params, _config, _signal) {
+        throw new Error(`Server tool '${toolName}' not supported by OpenAIVercelProvider`);
+    }
+    getToolFormat() {
+        const format = this.detectToolFormat();
+        const logger = new DebugLogger('llxprt:provider:openaivercel');
+        logger.debug(() => `getToolFormat() called, returning: ${format}`, {
+            provider: this.name,
+            model: this.getModel(),
+            format,
+        });
+        return format;
+    }
+    /**
+     * Detects the tool call format based on the model being used.
+     * Mirrors OpenAIProvider behavior so existing ToolFormatter logic works.
+     */
+    detectToolFormat() {
+        const modelName = (this.getModel() || this.getDefaultModel()).toLowerCase();
+        const logger = new DebugLogger('llxprt:provider:openaivercel');
+        if (modelName.includes('glm-4')) {
+            logger.debug(() => `Auto-detected 'qwen' format for GLM-4.x model: ${modelName}`);
+            return 'qwen';
+        }
+        if (modelName.includes('qwen')) {
+            logger.debug(() => `Auto-detected 'qwen' format for Qwen model: ${modelName}`);
+            return 'qwen';
+        }
+        logger.debug(() => `Using default 'openai' format for model: ${modelName}`);
+        return 'openai';
+    }
+    parseToolResponse(response) {
+        return response;
+    }
+    /**
+     * Disallow memoization of model params to preserve stateless behavior.
+     */
+    setModelParams(_params) {
+        throw new Error('ProviderCacheError("Attempted to memoize model parameters for openaivercel")');
+    }
+    /**
+     * Gets model parameters from SettingsService per call (stateless).
+     * Mirrors OpenAIProvider.getModelParams for compatibility.
+     */
+    getModelParams() {
+        try {
+            const settingsService = this.resolveSettingsService();
+            const providerSettings = settingsService.getProviderSettings(this.name);
+            const reservedKeys = new Set([
+                'enabled',
+                'apiKey',
+                'api-key',
+                'apiKeyfile',
+                'api-keyfile',
+                'baseUrl',
+                'base-url',
+                'model',
+                'toolFormat',
+                'tool-format',
+                'toolFormatOverride',
+                'tool-format-override',
+                'defaultModel',
+            ]);
+            const params = {};
+            if (providerSettings) {
+                for (const [key, value] of Object.entries(providerSettings)) {
+                    if (reservedKeys.has(key) || value === undefined || value === null) {
+                        continue;
+                    }
+                    params[key] = value;
+                }
+            }
+            return Object.keys(params).length > 0 ? params : undefined;
+        }
+        catch (error) {
+            this.getLogger().debug(() => `Failed to get OpenAIVercel provider settings from SettingsService: ${error}`);
+            return undefined;
+        }
+    }
+    /**
+     * Determines whether a response should be retried based on error codes.
+     *
+     * This is retained for compatibility with existing retryWithBackoff
+     * callers, even though AI SDK's generateText/streamText have their
+     * own built-in retry logic.
+     */
+    shouldRetryResponse(error) {
+        const logger = new DebugLogger('llxprt:provider:openaivercel');
+        // Don't retry if it's a "successful" 200 error wrapper
+        if (error &&
+            typeof error === 'object' &&
+            'status' in error &&
+            error.status === 200) {
+            return false;
+        }
+        let status;
+        if (error && typeof error === 'object' && 'status' in error) {
+            status = error.status;
+        }
+        if (!status && error && typeof error === 'object' && 'response' in error) {
+            const response = error.response;
+            if (response && typeof response === 'object' && 'status' in response) {
+                status = response.status;
+            }
+        }
+        if (!status && error instanceof Error) {
+            if (error.message.includes('429')) {
+                status = 429;
+            }
+        }
+        logger.debug(() => `shouldRetryResponse checking error:`, {
+            hasError: !!error,
+            errorType: error && typeof error === 'object'
+                ? error.constructor?.name
+                : undefined,
+            status,
+            errorMessage: error instanceof Error ? error.message : String(error),
+            errorKeys: error && typeof error === 'object' ? Object.keys(error) : [],
+            errorData: error && typeof error === 'object' && 'error' in error
+                ? error.error
+                : undefined,
+        });
+        const shouldRetry = Boolean(status === 429 || status === 503 || status === 504);
+        if (shouldRetry) {
+            logger.debug(() => `Will retry request due to status ${status}`);
+        }
+        return shouldRetry;
+    }
+}
+//# sourceMappingURL=OpenAIVercelProvider.js.map