@vybestack/llxprt-code-core 0.4.7 → 0.5.0-nightly.251102.6bb3db7a

package/dist/src/providers/openai/OpenAIProvider.js

@@ -21,334 +21,212 @@ import OpenAI from 'openai';
 import * as http from 'http';
 import * as https from 'https';
 import * as net from 'net';
-import { BaseProvider } from '../BaseProvider.js';
+import { BaseProvider, } from '../BaseProvider.js';
 import { DebugLogger } from '../../debug/index.js';
-import { getSettingsService } from '../../settings/settingsServiceInstance.js';
 import { ToolFormatter } from '../../tools/ToolFormatter.js';
 import { processToolParameters } from '../../tools/doubleEscapeUtils.js';
 import { getCoreSystemPromptAsync } from '../../core/prompts.js';
-import { retryWithBackoff, isNetworkTransientError, } from '../../utils/retry.js';
+import { retryWithBackoff } from '../../utils/retry.js';
+import { resolveUserMemory } from '../utils/userMemory.js';
+import { resolveRuntimeAuthToken } from '../utils/authToken.js';
+import { filterOpenAIRequestParams } from './openaiRequestParams.js';
 export class OpenAIProvider extends BaseProvider {
     name = 'openai';
-    logger;
-    toolFormatter;
-    _cachedClient;
-    _cachedClientKey;
-    modelParams;
+    getLogger() {
+        return new DebugLogger('llxprt:provider:openai');
+    }
+    /**
+     * @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;
         // Detect if this is a Qwen endpoint
         // CRITICAL FIX: For now, only use base URL check in constructor since `this.name` isn't available yet
         // The name-based check will be handled in the supportsOAuth() method after construction
-        const isQwenEndpoint = !!(baseURL &&
-            (baseURL.includes('dashscope.aliyuncs.com') ||
-                baseURL.includes('api.qwen.com') ||
-                baseURL.includes('qwen')));
+        let isQwenEndpoint = false;
+        if (baseURL) {
+            try {
+                const hostname = new URL(baseURL).hostname.toLowerCase();
+                isQwenEndpoint =
+                    hostname === 'dashscope.aliyuncs.com' ||
+                        hostname.endsWith('.dashscope.aliyuncs.com') ||
+                        hostname === 'api.qwen.com' ||
+                        hostname.endsWith('.qwen.com');
+            }
+            catch {
+                const lowered = baseURL.toLowerCase();
+                isQwenEndpoint =
+                    lowered.includes('dashscope.aliyuncs.com') ||
+                        lowered.includes('api.qwen.com') ||
+                        lowered.includes('qwen.com');
+            }
+        }
+        const forceQwenOAuth = Boolean(config?.forceQwenOAuth);
         // Initialize base provider with auth configuration
         super({
             name: 'openai',
             apiKey: normalizedApiKey,
             baseURL,
             envKeyNames: ['OPENAI_API_KEY'], // Support environment variable fallback
-            isOAuthEnabled: isQwenEndpoint && !!oauthManager,
-            oauthProvider: isQwenEndpoint ? 'qwen' : undefined,
+            isOAuthEnabled: (isQwenEndpoint || forceQwenOAuth) && !!oauthManager,
+            oauthProvider: isQwenEndpoint || forceQwenOAuth ? 'qwen' : undefined,
             oauthManager,
         }, config);
-        this.toolFormatter = new ToolFormatter();
-        // new DebugLogger('llxprt:core:toolformatter'), // TODO: Fix ToolFormatter constructor
-        // Setup debug logger
-        this.logger = new DebugLogger('llxprt:provider:openai');
-        this.loadModelParamsFromSettings().catch((error) => {
-            this.logger.debug(() => `Failed to initialize model params from SettingsService: ${error}`);
-        });
+        // @plan:PLAN-20251023-STATELESS-HARDENING.P08
+        // @requirement:REQ-SP4-002
+        // No constructor-captured state - all values sourced from normalized options per call
     }
-    getSocketSettings() {
-        const settings = this.providerConfig?.getEphemeralSettings?.() || {};
-        const timeoutSetting = settings['socket-timeout'];
-        const keepAliveSetting = settings['socket-keepalive'];
-        const noDelaySetting = settings['socket-nodelay'];
-        const hasExplicitValue = (setting) => setting !== undefined && setting !== null;
-        if (!hasExplicitValue(timeoutSetting) &&
-            !hasExplicitValue(keepAliveSetting) &&
-            !hasExplicitValue(noDelaySetting)) {
+    /**
+     * Create HTTP/HTTPS agents with socket configuration for local AI servers
+     * Returns undefined if no socket settings are configured
+     *
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Now sources ephemeral settings from call options instead of provider config
+     */
+    createHttpAgents(options) {
+        // Get socket configuration from call options or fallback to provider config
+        const settingsFromInvocation = options?.invocation?.ephemerals;
+        const settings = settingsFromInvocation ??
+            this.providerConfig?.getEphemeralSettings?.() ??
+            {};
+        // Check if any socket settings are explicitly configured
+        const hasSocketSettings = 'socket-timeout' in settings ||
+            'socket-keepalive' in settings ||
+            'socket-nodelay' in settings;
+        // Only create custom agents if socket settings are configured
+        if (!hasSocketSettings) {
             return undefined;
         }
-        const timeout = typeof timeoutSetting === 'number' && Number.isFinite(timeoutSetting)
-            ? timeoutSetting
-            : Number(timeoutSetting) > 0
-                ? Number(timeoutSetting)
-                : 60000;
-        const keepAlive = keepAliveSetting === undefined || keepAliveSetting === null
-            ? true
-            : keepAliveSetting !== false;
-        const noDelay = noDelaySetting === undefined || noDelaySetting === null
-            ? true
-            : noDelaySetting !== false;
-        return {
-            timeout,
-            keepAlive,
-            noDelay,
-        };
-    }
-    createSocketAwareFetch(config) {
-        const { timeout, keepAlive, noDelay } = config;
-        const maxRetries = 2;
-        const retryDelay = 1000;
-        const partialResponseThreshold = 2;
-        const buildHeaders = (init) => {
-            const baseHeaders = {
-                Accept: 'text/event-stream',
-                Connection: keepAlive ? 'keep-alive' : 'close',
-                'Cache-Control': 'no-cache',
-            };
-            if (!init?.headers) {
-                return baseHeaders;
-            }
-            const appendHeader = (key, value) => {
-                baseHeaders[key] = value;
-            };
-            const headers = init.headers;
-            if (headers instanceof Headers) {
-                headers.forEach((value, key) => {
-                    appendHeader(key, value);
-                });
-            }
-            else if (Array.isArray(headers)) {
-                headers.forEach(([key, value]) => {
-                    if (typeof value === 'string') {
-                        appendHeader(key, value);
-                    }
-                });
-            }
-            else if (typeof headers === 'object') {
-                Object.entries(headers).forEach(([key, value]) => {
-                    if (typeof value === 'string') {
-                        appendHeader(key, value);
-                    }
-                    else if (Array.isArray(value)) {
-                        appendHeader(key, value.join(', '));
-                    }
-                    else if (value !== undefined && value !== null) {
-                        appendHeader(key, String(value));
-                    }
-                });
-            }
-            return baseHeaders;
-        };
-        const collectResponseHeaders = (rawHeaders) => {
-            const headers = new Headers();
-            for (const [key, value] of Object.entries(rawHeaders)) {
-                if (!key)
-                    continue;
-                if (Array.isArray(value)) {
-                    headers.append(key, value.join(', '));
-                }
-                else if (value !== undefined) {
-                    headers.append(key, value);
-                }
-            }
-            return headers;
-        };
-        const writeRequestBody = (req, body) => {
-            if (!body) {
-                req.end();
-                return;
-            }
-            if (typeof body === 'string' || body instanceof Buffer) {
-                req.write(body);
-                req.end();
-                return;
-            }
-            if (body instanceof ArrayBuffer) {
-                req.write(Buffer.from(body));
-                req.end();
-                return;
-            }
-            if (ArrayBuffer.isView(body)) {
-                req.write(Buffer.from(body.buffer, body.byteOffset, body.byteLength));
-                req.end();
-                return;
-            }
-            try {
-                req.write(body);
-            }
-            catch {
-                req.write(String(body));
-            }
-            req.end();
-        };
-        const delay = (ms) => new Promise((resolve) => {
-            setTimeout(resolve, ms);
+        // Socket configuration with defaults for when settings ARE configured
+        const socketTimeout = settings['socket-timeout'] || 60000; // 60 seconds default
+        const socketKeepAlive = settings['socket-keepalive'] !== false; // true by default
+        const socketNoDelay = settings['socket-nodelay'] !== false; // true by default
+        // Create HTTP agent with socket options
+        const httpAgent = new http.Agent({
+            keepAlive: socketKeepAlive,
+            keepAliveMsecs: 1000,
+            timeout: socketTimeout,
         });
-        const makeRequest = async (url, init, attempt = 0) => new Promise((resolve, reject) => {
-            let parsedUrl;
-            try {
-                parsedUrl = new URL(url);
-            }
-            catch (error) {
-                reject(new Error(`Invalid URL provided to socket-aware fetch: ${url} (${String(error)})`));
-                return;
-            }
-            const isHttps = parsedUrl.protocol === 'https:';
-            const httpModule = isHttps ? https : http;
-            const options = {
-                hostname: parsedUrl.hostname,
-                port: parsedUrl.port ? Number(parsedUrl.port) : isHttps ? 443 : 80,
-                path: `${parsedUrl.pathname}${parsedUrl.search}`,
-                method: init?.method?.toUpperCase() || 'GET',
-                headers: buildHeaders(init),
-            };
-            const req = httpModule.request(options, (res) => {
-                const chunks = [];
-                let chunkCount = 0;
-                res.on('data', (chunk) => {
-                    chunkCount += 1;
-                    if (typeof chunk === 'string') {
-                        chunks.push(Buffer.from(chunk));
-                    }
-                    else {
-                        chunks.push(chunk);
-                    }
-                });
-                res.on('end', () => {
-                    const bodyBuffer = Buffer.concat(chunks);
-                    resolve(new Response(bodyBuffer, {
-                        status: res.statusCode ?? 0,
-                        statusText: res.statusMessage ?? '',
-                        headers: collectResponseHeaders(res.headers),
-                    }));
-                });
-                res.on('error', async (error) => {
-                    if (chunkCount >= partialResponseThreshold &&
-                        attempt < maxRetries) {
-                        await delay(retryDelay);
-                        try {
-                            const retryResponse = await makeRequest(url, init, attempt + 1);
-                            resolve(retryResponse);
-                            return;
-                        }
-                        catch (retryError) {
-                            reject(retryError);
-                            return;
-                        }
-                    }
-                    reject(new Error(`Response stream error: ${String(error)}`));
-                });
-            });
-            req.on('socket', (socket) => {
+        // Create HTTPS agent with socket options
+        const httpsAgent = new https.Agent({
+            keepAlive: socketKeepAlive,
+            keepAliveMsecs: 1000,
+            timeout: socketTimeout,
+        });
+        // Apply TCP_NODELAY if enabled (reduces latency for local servers)
+        if (socketNoDelay) {
+            const originalCreateConnection = httpAgent.createConnection;
+            httpAgent.createConnection = function (options, callback) {
+                const socket = originalCreateConnection.call(this, options, callback);
                 if (socket instanceof net.Socket) {
-                    socket.setTimeout(timeout);
-                    socket.setKeepAlive(keepAlive, 1000);
-                    socket.setNoDelay(noDelay);
+                    socket.setNoDelay(true);
                 }
-            });
-            req.setTimeout(timeout, () => {
-                req.destroy(new Error(`Request timed out after ${timeout}ms`));
-            });
-            if (init?.signal) {
-                const abortHandler = () => {
-                    const abortError = new Error('Request aborted');
-                    abortError.name = 'AbortError';
-                    req.destroy(abortError);
-                };
-                if (init.signal.aborted) {
-                    abortHandler();
-                    return;
-                }
-                init.signal.addEventListener('abort', abortHandler);
-                req.on('close', () => {
-                    init.signal?.removeEventListener('abort', abortHandler);
-                });
-            }
-            req.on('error', async (error) => {
-                if (attempt < maxRetries) {
-                    await delay(retryDelay);
-                    try {
-                        const retryResponse = await makeRequest(url, init, attempt + 1);
-                        resolve(retryResponse);
-                        return;
-                    }
-                    catch (retryError) {
-                        reject(retryError);
-                        return;
-                    }
+                return socket;
+            };
+            const originalHttpsCreateConnection = httpsAgent.createConnection;
+            httpsAgent.createConnection = function (options, callback) {
+                const socket = originalHttpsCreateConnection.call(this, options, callback);
+                if (socket instanceof net.Socket) {
+                    socket.setNoDelay(true);
                 }
-                reject(new Error(`Request failed: ${String(error)}`));
-            });
-            writeRequestBody(req, init?.body ?? null);
-        });
-        return async (input, init) => {
-            const url = typeof input === 'string'
-                ? input
-                : input instanceof URL
-                    ? input.href
-                    : input.url;
-            if (typeof url !== 'string') {
-                return fetch(input, init);
-            }
-            return makeRequest(url, init);
-        };
-    }
-    async loadModelParamsFromSettings() {
-        const params = await this.getModelParamsFromSettings();
-        this.modelParams = params;
-    }
-    async resolveModelParams() {
-        if (this.modelParams) {
-            return this.modelParams;
+                return socket;
+            };
         }
-        const params = await this.getModelParamsFromSettings();
-        if (params) {
-            this.modelParams = params;
+        return { httpAgent, httpsAgent };
+    }
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-002
+     * Extract model parameters from normalized options instead of settings service
+     */
+    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 params;
+        return {
+            ...(filteredProviderParams ?? {}),
+            ...(filteredEphemeralParams ?? {}),
+        };
     }
     /**
-     * Get or create OpenAI client instance
-     * Will use the API key from resolved auth
-     * @returns OpenAI client instance
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Resolve runtime key from normalized options for client scoping
      */
-    async getClient() {
-        const resolvedKey = await this.getAuthToken();
-        // Use the unified getBaseURL() method from BaseProvider
-        const baseURL = this.getBaseURL();
-        const socketSettings = this.getSocketSettings();
-        const socketKey = socketSettings
-            ? JSON.stringify(socketSettings)
-            : 'default';
-        const clientKey = `${baseURL}-${resolvedKey}-${socketKey}`;
-        // Clear cache if we have no valid auth (e.g., after logout)
-        if (!resolvedKey && this._cachedClient) {
-            this._cachedClient = undefined;
-            this._cachedClientKey = undefined;
+    resolveRuntimeKey(options) {
+        if (options.runtime?.runtimeId) {
+            return options.runtime.runtimeId;
         }
-        // Return cached client if available and auth hasn't changed
-        if (this._cachedClient && this._cachedClientKey === clientKey) {
-            return this._cachedClient;
+        const metadataRuntimeId = options.metadata?.runtimeId;
+        if (typeof metadataRuntimeId === 'string' && metadataRuntimeId.trim()) {
+            return metadataRuntimeId.trim();
         }
-        const baseOptions = {
-            apiKey: resolvedKey || '',
-            baseURL,
-            // CRITICAL: Disable OpenAI SDK's built-in retries so our retry logic can handle them
-            // This allows us to track throttle wait times properly
+        const callId = options.settings.get('call-id');
+        if (typeof callId === 'string' && callId.trim()) {
+            return `call:${callId.trim()}`;
+        }
+        return 'openai.runtime.unscoped';
+    }
+    /**
+     * Tool formatter instances cannot be shared between stateless calls,
+     * so construct a fresh one for every invocation.
+     *
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     */
+    createToolFormatter() {
+        return new ToolFormatter();
+    }
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P09
+     * @requirement:REQ-SP4-002
+     * Instantiates a fresh OpenAI client per call to preserve stateless behaviour.
+     */
+    instantiateClient(authToken, baseURL, agents) {
+        const clientOptions = {
+            apiKey: authToken || '',
             maxRetries: 0,
         };
-        if (socketSettings) {
-            baseOptions.timeout = socketSettings.timeout;
-            baseOptions.fetch = this.createSocketAwareFetch(socketSettings);
+        if (baseURL && baseURL.trim() !== '') {
+            clientOptions.baseURL = baseURL;
         }
-        // Create new client with current auth and optional socket configuration
-        // Cast to unknown then to the expected type to bypass TypeScript's structural checking
-        this._cachedClient = new OpenAI(baseOptions);
-        this._cachedClientKey = clientKey;
-        return this._cachedClient;
+        if (agents) {
+            clientOptions.httpAgent = agents.httpAgent;
+            clientOptions.httpsAgent = agents.httpsAgent;
+        }
+        return new OpenAI(clientOptions);
+    }
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P09
+     * @requirement:REQ-SP4-002
+     * Creates a client scoped to the active runtime metadata without caching.
+     */
+    async getClient(options) {
+        const authToken = (await resolveRuntimeAuthToken(options.resolved.authToken)) ?? '';
+        if (!authToken) {
+            throw new Error(`ProviderCacheError("Auth token unavailable for runtimeId=${options.runtime?.runtimeId} (REQ-SP4-003).")`);
+        }
+        const baseURL = options.resolved.baseURL ?? this.baseProviderConfig.baseURL;
+        const agents = this.createHttpAgents(options);
+        return this.instantiateClient(authToken, baseURL, agents);
     }
     /**
      * Check if OAuth is supported for this provider
      * Qwen endpoints support OAuth, standard OpenAI does not
      */
     supportsOAuth() {
+        const providerConfig = this.providerConfig;
+        if (providerConfig?.forceQwenOAuth) {
+            return true;
+        }
         // CRITICAL FIX: Check provider name first for cases where base URL is changed by profiles
         // This handles the cerebrasqwen3 profile case where base-url is changed to cerebras.ai
         // but the provider name is still 'qwen' due to Object.defineProperty override
@@ -370,7 +248,10 @@ export class OpenAIProvider extends BaseProvider {
         try {
             // Always try to fetch models, regardless of auth status
             // Local endpoints often work without authentication
-            const client = await this.getClient();
+            const authToken = await this.getAuthToken();
+            const baseURL = this.getBaseURL();
+            const agents = this.createHttpAgents();
+            const client = this.instantiateClient(authToken, baseURL, agents);
             const response = await client.models.list();
             const models = [];
             for await (const model of response) {
@@ -387,7 +268,7 @@ export class OpenAIProvider extends BaseProvider {
             return models;
         }
         catch (error) {
-            this.logger.debug(() => `Error fetching models from OpenAI: ${error}`);
+            this.getLogger().debug(() => `Error fetching models from OpenAI: ${error}`);
             // Return a hardcoded list as fallback
             return this.getFallbackModels();
         }
@@ -397,9 +278,6 @@ export class OpenAIProvider extends BaseProvider {
     }
     getDefaultModel() {
         // Return hardcoded default - do NOT call getModel() to avoid circular dependency
-        if (this.providerConfig?.defaultModel) {
-            return this.providerConfig.defaultModel;
-        }
         // Check if this is a Qwen provider instance based on baseURL
         const baseURL = this.getBaseURL();
         if (baseURL &&
@@ -408,15 +286,6 @@ export class OpenAIProvider extends BaseProvider {
         }
         return process.env.LLXPRT_DEFAULT_MODEL || 'gpt-5';
     }
-    /**
-     * Set the model to use for this provider
-     * This updates the model in ephemeral settings so it's immediately available
-     */
-    setModel(modelId) {
-        const settingsService = getSettingsService();
-        settingsService.set('model', modelId);
-        this.logger.debug(() => `Model set to: ${modelId}`);
-    }
     /**
      * Get the currently selected model
      */
@@ -424,32 +293,51 @@ export class OpenAIProvider extends BaseProvider {
         return this.getModel();
     }
     /**
-     * Clear the cached OpenAI client
-     * Should be called when authentication state changes (e.g., after logout)
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P09
+     * @requirement:REQ-SP4-002
+     * No-op retained for compatibility because clients are no longer cached.
      */
     // eslint-disable-next-line @typescript-eslint/explicit-member-accessibility
-    clearClientCache() {
-        this._cachedClient = undefined;
-        this._cachedClientKey = undefined;
+    clearClientCache(runtimeKey) {
+        void runtimeKey;
     }
     /**
      * Override isAuthenticated for qwen provider to check OAuth directly
      */
     async isAuthenticated() {
         const config = this.providerConfig;
+        const directApiKey = this.baseProviderConfig.apiKey;
+        if (typeof directApiKey === 'string' && directApiKey.trim() !== '') {
+            return true;
+        }
+        try {
+            const nonOAuthToken = await this.authResolver.resolveAuthentication({
+                settingsService: this.resolveSettingsService(),
+                includeOAuth: false,
+            });
+            if (typeof nonOAuthToken === 'string' && nonOAuthToken.trim() !== '') {
+                return true;
+            }
+        }
+        catch (error) {
+            if (process.env.DEBUG) {
+                this.getLogger().debug(() => `[openai] non-OAuth authentication resolution failed: ${error instanceof Error ? error.message : String(error)}`);
+            }
+        }
         if (this.name === 'qwen' && config?.forceQwenOAuth) {
-            // For qwen with forceQwenOAuth, check OAuth directly
-            if (this.baseProviderConfig.oauthManager) {
-                try {
-                    const oauthProviderName = this.baseProviderConfig.oauthProvider || 'qwen';
-                    const token = await this.baseProviderConfig.oauthManager.getToken(oauthProviderName);
-                    return token !== null;
-                }
-                catch {
-                    return false;
+            try {
+                const token = await this.authResolver.resolveAuthentication({
+                    settingsService: this.resolveSettingsService(),
+                    includeOAuth: true,
+                });
+                return typeof token === 'string' && token.trim() !== '';
+            }
+            catch (error) {
+                if (process.env.DEBUG) {
+                    this.getLogger().debug(() => `[openai] forced OAuth authentication failed: ${error instanceof Error ? error.message : String(error)}`);
                 }
+                return false;
             }
-            return false;
         }
         // For non-qwen providers, use the normal check
         return super.isAuthenticated();
@@ -462,15 +350,6 @@ export class OpenAIProvider extends BaseProvider {
         // If this is the qwen provider and we have forceQwenOAuth, skip SettingsService checks
         const config = this.providerConfig;
         if (this.name === 'qwen' && config?.forceQwenOAuth) {
-            // Check cache first (short-lived cache to avoid repeated OAuth calls)
-            if (this.cachedAuthToken &&
-                this.authCacheTimestamp &&
-                Date.now() - this.authCacheTimestamp < this.AUTH_CACHE_DURATION) {
-                return this.cachedAuthToken;
-            }
-            // Clear stale cache
-            this.cachedAuthToken = undefined;
-            this.authCacheTimestamp = undefined;
             // For qwen, skip directly to OAuth without checking SettingsService
             // Use 'qwen' as the provider name even if baseProviderConfig.oauthProvider is not set
             const oauthProviderName = this.baseProviderConfig.oauthProvider || 'qwen';
@@ -478,9 +357,6 @@ export class OpenAIProvider extends BaseProvider {
                 try {
                     const token = await this.baseProviderConfig.oauthManager.getToken(oauthProviderName);
                     if (token) {
-                        // Cache the token briefly
-                        this.cachedAuthToken = token;
-                        this.authCacheTimestamp = Date.now();
                         return token;
                     }
                 }
@@ -558,25 +434,36 @@ export class OpenAIProvider extends BaseProvider {
         return 'hist_tool_' + id;
     }
     /**
-     * Generate chat completion with IContent interface
-     * Internally converts to OpenAI API format, but only yields IContent
-     * @param contents Array of content blocks (text and tool_call)
-     * @param tools Array of available tools
+     * @plan PLAN-20250218-STATELESSPROVIDER.P04
+     * @requirement REQ-SP-001
+     * @pseudocode base-provider.md lines 7-15
+     * @pseudocode provider-invocation.md lines 8-12
      */
-    async *generateChatCompletion(contents, tools) {
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P09
+     * @requirement:REQ-SP4-002
+     * Generate chat completion with per-call client instantiation.
+     */
+    async *generateChatCompletionWithOptions(options) {
+        const callFormatter = this.createToolFormatter();
+        const client = await this.getClient(options);
+        const runtimeKey = this.resolveRuntimeKey(options);
+        const { tools } = options;
+        const logger = new DebugLogger('llxprt:provider:openai');
         // Debug log what we receive
-        if (this.logger.enabled) {
-            this.logger.debug(() => `[OpenAIProvider] generateChatCompletion received tools:`, {
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIProvider] generateChatCompletion received tools:`, {
                 hasTools: !!tools,
                 toolsLength: tools?.length,
                 toolsType: typeof tools,
                 isArray: Array.isArray(tools),
                 firstToolName: tools?.[0]?.functionDeclarations?.[0]?.name,
                 toolsStructure: tools ? 'available' : 'undefined',
+                runtimeKey,
             });
         }
-        // Pass tools directly in Gemini format - they'll be converted in generateChatCompletionImpl
-        const generator = this.generateChatCompletionImpl(contents, tools, undefined, undefined, undefined);
+        // Pass tools directly in Gemini format - they'll be converted per call
+        const generator = this.generateChatCompletionImpl(options, callFormatter, client, logger);
         for await (const item of generator) {
             yield item;
         }
@@ -646,27 +533,44 @@ export class OpenAIProvider extends BaseProvider {
         return messages;
     }
     /**
-     * Internal implementation for chat completion
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Internal implementation for chat completion using normalized options
      */
-    async *generateChatCompletionImpl(contents, tools, maxTokens, abortSignal, modelName) {
-        // Always look up model from SettingsService
-        const model = modelName || this.getModel() || this.getDefaultModel();
+    async *generateChatCompletionImpl(options, toolFormatter, client, logger) {
+        const { contents, tools, metadata } = options;
+        const model = options.resolved.model || this.getDefaultModel();
+        const abortSignal = metadata?.abortSignal;
+        const ephemeralSettings = options.invocation?.ephemerals ?? {};
+        if (logger.enabled) {
+            const resolved = options.resolved;
+            logger.debug(() => `[OpenAIProvider] Resolved request context`, {
+                provider: this.name,
+                model,
+                resolvedModel: resolved.model,
+                resolvedBaseUrl: resolved.baseURL,
+                authTokenPresent: Boolean(resolved.authToken),
+                messageCount: contents.length,
+                toolCount: tools?.length ?? 0,
+                metadataKeys: Object.keys(metadata ?? {}),
+            });
+        }
         // Convert IContent to OpenAI messages format
         const messages = this.convertToOpenAIMessages(contents);
         // Detect the tool format to use (once at the start of the method)
         const detectedFormat = this.detectToolFormat();
         // Log the detected format for debugging
-        this.logger.debug(() => `[OpenAIProvider] Using tool format '${detectedFormat}' for model '${model}'`, {
+        logger.debug(() => `[OpenAIProvider] Using tool format '${detectedFormat}' for model '${model}'`, {
             model,
             detectedFormat,
             provider: this.name,
         });
         // Convert Gemini format tools to the detected format
-        let formattedTools = this.toolFormatter.convertGeminiToFormat(tools, detectedFormat);
+        let formattedTools = toolFormatter.convertGeminiToFormat(tools, detectedFormat);
         // CRITICAL FIX: Ensure we never pass an empty tools array
         // The OpenAI API errors when tools=[] but a tool call is attempted
         if (Array.isArray(formattedTools) && formattedTools.length === 0) {
-            this.logger.warn(() => `[OpenAIProvider] CRITICAL: Formatted tools is empty array! Setting to undefined to prevent API errors.`, {
+            logger.warn(() => `[OpenAIProvider] CRITICAL: Formatted tools is empty array! Setting to undefined to prevent API errors.`, {
                 model,
                 inputTools: tools,
                 inputToolsLength: tools?.length,
@@ -676,8 +580,8 @@ export class OpenAIProvider extends BaseProvider {
             formattedTools = undefined;
         }
         // Debug log the conversion result - enhanced logging for intermittent issues
-        if (this.logger.enabled && formattedTools) {
-            this.logger.debug(() => `[OpenAIProvider] Tool conversion summary:`, {
+        if (logger.enabled && formattedTools) {
+            logger.debug(() => `[OpenAIProvider] Tool conversion summary:`, {
                 detectedFormat,
                 inputHadTools: !!tools,
                 inputToolsLength: tools?.length,
@@ -689,48 +593,61 @@ export class OpenAIProvider extends BaseProvider {
             });
         }
         // Get streaming setting from ephemeral settings (default: enabled)
-        const streamingSetting = this.providerConfig?.getEphemeralSettings?.()?.['streaming'];
+        const streamingSetting = ephemeralSettings['streaming'];
         const streamingEnabled = streamingSetting !== 'disabled';
         // Get the system prompt
         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 = this.globalConfig?.getUserMemory
-            ? this.globalConfig.getUserMemory()
-            : '';
+        /**
+         * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+         * @requirement:REQ-SP4-003
+         * Source user memory from normalized options instead of global config
+         */
+        const userMemory = await resolveUserMemory(options.userMemory, () => options.invocation?.userMemory);
         const systemPrompt = await getCoreSystemPromptAsync(userMemory, model, toolNamesArg);
         // Add system prompt as the first message in the array
         const messagesWithSystem = [
             { role: 'system', content: systemPrompt },
             ...messages,
         ];
+        const maxTokens = metadata?.maxTokens ??
+            ephemeralSettings['max-tokens'];
         // Build request - only include tools if they exist and are not empty
         // IMPORTANT: Create a deep copy of tools to prevent mutation issues
         const requestBody = {
             model,
             messages: messagesWithSystem,
-            ...(formattedTools && formattedTools.length > 0
-                ? {
-                    // Deep clone the tools array to prevent any mutation issues
-                    tools: JSON.parse(JSON.stringify(formattedTools)),
-                    // Add tool_choice for Qwen/Cerebras to ensure proper tool calling
-                    tool_choice: 'auto',
-                }
-                : {}),
-            max_tokens: maxTokens,
             stream: streamingEnabled,
         };
-        const modelParams = await this.resolveModelParams();
-        if (modelParams) {
-            Object.assign(requestBody, modelParams);
+        if (formattedTools && formattedTools.length > 0) {
+            requestBody.tools = JSON.parse(JSON.stringify(formattedTools));
+            requestBody.tool_choice = 'auto';
+        }
+        /**
+         * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+         * @requirement:REQ-SP4-002
+         * Extract per-call request overrides from normalized options instead of cached state
+         */
+        const requestOverrides = this.extractModelParamsFromOptions(options);
+        if (requestOverrides) {
+            if (logger.enabled) {
+                logger.debug(() => `[OpenAIProvider] Applying request overrides`, {
+                    overrideKeys: Object.keys(requestOverrides),
+                });
+            }
+            Object.assign(requestBody, requestOverrides);
+        }
+        if (typeof maxTokens === 'number' && Number.isFinite(maxTokens)) {
+            requestBody.max_tokens = maxTokens;
         }
         // Debug log request summary for Cerebras/Qwen
-        if (this.logger.enabled &&
-            (model.toLowerCase().includes('qwen') ||
-                this.getBaseURL()?.includes('cerebras'))) {
-            this.logger.debug(() => `Request to ${this.getBaseURL()} for model ${model}:`, {
-                baseURL: this.getBaseURL(),
+        const baseURL = options.resolved.baseURL ?? this.getBaseURL();
+        if (logger.enabled &&
+            (model.toLowerCase().includes('qwen') || baseURL?.includes('cerebras'))) {
+            logger.debug(() => `Request to ${baseURL} for model ${model}:`, {
+                baseURL,
                 model,
                 streamingEnabled,
                 hasTools: 'tools' in requestBody,
@@ -739,10 +656,7 @@ export class OpenAIProvider extends BaseProvider {
                 toolsInRequest: 'tools' in requestBody ? requestBody.tools?.length : 'not included',
             });
         }
-        // Get OpenAI client
-        const client = await this.getClient();
         // Get retry settings from ephemeral settings
-        const ephemeralSettings = this.providerConfig?.getEphemeralSettings?.() || {};
         const maxRetries = ephemeralSettings['retries'] ?? 6; // Default for OpenAI
         const initialDelayMs = ephemeralSettings['retrywait'] ?? 4000; // Default for OpenAI
         // Get stream options from ephemeral settings (default: include usage for token tracking)
@@ -752,23 +666,43 @@ export class OpenAIProvider extends BaseProvider {
             Object.assign(requestBody, { stream_options: streamOptions });
         }
         // Log the exact tools being sent for debugging
-        if (this.logger.enabled && 'tools' in requestBody) {
-            this.logger.debug(() => `[OpenAIProvider] Exact tools being sent to API:`, {
+        if (logger.enabled && 'tools' in requestBody) {
+            logger.debug(() => `[OpenAIProvider] Exact tools being sent to API:`, {
                 toolCount: requestBody.tools?.length,
                 toolNames: requestBody.tools?.map((t) => 'function' in t ? t.function?.name : undefined),
                 firstTool: requestBody.tools?.[0],
             });
         }
         // Wrap the API call with retry logic using centralized retry utility
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIProvider] Sending chat request`, {
+                model,
+                baseURL: baseURL ?? this.getBaseURL(),
+                streamingEnabled,
+                toolCount: formattedTools?.length ?? 0,
+                hasAuthToken: Boolean(options.resolved.authToken),
+                requestHasSystemPrompt: Boolean(systemPrompt?.length),
+                messageCount: messagesWithSystem.length,
+            });
+        }
         let response;
         // Debug log throttle tracker status
-        this.logger.debug(() => `Retry configuration:`, {
+        logger.debug(() => `Retry configuration:`, {
             hasThrottleTracker: !!this.throttleTracker,
             throttleTrackerType: typeof this.throttleTracker,
             maxRetries,
             initialDelayMs,
         });
         const customHeaders = this.getCustomHeaders();
+        if (logger.enabled) {
+            logger.debug(() => `[OpenAIProvider] Request body preview`, {
+                model: requestBody.model,
+                hasStop: 'stop' in requestBody,
+                hasMaxTokens: 'max_tokens' in requestBody,
+                hasResponseFormat: 'response_format' in requestBody,
+                overrideKeys: requestOverrides ? Object.keys(requestOverrides) : [],
+            });
+        }
         try {
             response = await retryWithBackoff(() => client.chat.completions.create(requestBody, {
                 ...(abortSignal ? { signal: abortSignal } : {}),
@@ -786,7 +720,7 @@ export class OpenAIProvider extends BaseProvider {
             if (errorMessage.includes('Tool is not present in the tools list') &&
                 (model.toLowerCase().includes('qwen') ||
                     this.getBaseURL()?.includes('cerebras'))) {
-                this.logger.error('Cerebras/Qwen API error: Tool not found despite being in request. This is a known API issue.', {
+                logger.error('Cerebras/Qwen API error: Tool not found despite being in request. This is a known API issue.', {
                     error,
                     model,
                     toolsProvided: formattedTools?.length || 0,
@@ -800,6 +734,21 @@ export class OpenAIProvider extends BaseProvider {
                 throw enhancedError;
             }
             // Re-throw other errors as-is
+            const capturedErrorMessage = error instanceof Error ? error.message : String(error);
+            const status = typeof error === 'object' &&
+                error !== null &&
+                'status' in error &&
+                typeof error.status === 'number'
+                ? error.status
+                : undefined;
+            logger.error(() => `[OpenAIProvider] Chat completion failed for model '${model}' at '${baseURL ?? this.getBaseURL() ?? 'default'}': ${capturedErrorMessage}`, {
+                model,
+                baseURL: baseURL ?? this.getBaseURL(),
+                streamingEnabled,
+                hasTools: formattedTools?.length ?? 0,
+                requestHasSystemPrompt: !!systemPrompt,
+                status,
+            });
             throw error;
         }
         // Check if response is streaming or not
@@ -829,7 +778,7 @@ export class OpenAIProvider extends BaseProvider {
                         continue;
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
-                        this.logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
+                        logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
                             model,
                             finishReason: choice.finish_reason,
                             hasAccumulatedText: _accumulatedText.length > 0,
@@ -838,7 +787,7 @@ export class OpenAIProvider extends BaseProvider {
                         });
                         // If finish_reason is 'length', the response was cut off
                         if (choice.finish_reason === 'length') {
-                            this.logger.debug(() => `Response truncated due to length limit for model ${model}`);
+                            logger.debug(() => `Response truncated due to length limit for model ${model}`);
                         }
                         // Flush any buffered text when stream finishes
                         if (textBuffer.length > 0) {
@@ -860,7 +809,7 @@ export class OpenAIProvider extends BaseProvider {
                         _accumulatedText += deltaContent;
                         // Debug log for providers that need buffering
                         if (shouldBufferText) {
-                            this.logger.debug(() => `[Streaming] Chunk content for ${detectedFormat} format:`, {
+                            logger.debug(() => `[Streaming] Chunk content for ${detectedFormat} format:`, {
                                 deltaContent,
                                 length: deltaContent.length,
                                 hasNewline: deltaContent.includes('\n'),
@@ -871,7 +820,8 @@ export class OpenAIProvider extends BaseProvider {
                             textBuffer += deltaContent;
                             // Emit buffered text when we have a complete sentence or paragraph
                             // Look for natural break points
-                            if (textBuffer.endsWith('. ') ||
+                            if (textBuffer.includes('\n') ||
+                                textBuffer.endsWith('. ') ||
                                 textBuffer.endsWith('! ') ||
                                 textBuffer.endsWith('? ') ||
                                 textBuffer.length > 100) {
@@ -941,7 +891,7 @@ export class OpenAIProvider extends BaseProvider {
                     if (errorMessage.includes('Tool is not present in the tools list') &&
                         (model.toLowerCase().includes('qwen') ||
                             this.getBaseURL()?.includes('cerebras'))) {
-                        this.logger.error('Cerebras/Qwen API error: Tool not found despite being in request. This is a known API issue.', {
+                        logger.error('Cerebras/Qwen API error: Tool not found despite being in request. This is a known API issue.', {
                             error,
                             model,
                             toolsProvided: formattedTools?.length || 0,
@@ -952,7 +902,7 @@ export class OpenAIProvider extends BaseProvider {
                         enhancedError.originalError = error;
                         throw enhancedError;
                     }
-                    this.logger.error('Error processing streaming response:', error);
+                    logger.error('Error processing streaming response:', error);
                     throw error;
                 }
             }
@@ -1031,7 +981,7 @@ export class OpenAIProvider extends BaseProvider {
             }
             // Log finish reason for debugging Qwen issues
             if (choice.finish_reason) {
-                this.logger.debug(() => `[Non-streaming] Response finish_reason: ${choice.finish_reason}`, {
+                logger.debug(() => `[Non-streaming] Response finish_reason: ${choice.finish_reason}`, {
                     model,
                     finishReason: choice.finish_reason,
                     hasContent: !!choice.message?.content,
@@ -1042,7 +992,7 @@ export class OpenAIProvider extends BaseProvider {
                 });
                 // Warn if the response was truncated
                 if (choice.finish_reason === 'length') {
-                    this.logger.warn(() => `Response truncated due to max_tokens limit for model ${model}. Consider increasing max_tokens.`);
+                    logger.warn(() => `Response truncated due to max_tokens limit for model ${model}. Consider increasing max_tokens.`);
                 }
             }
             const blocks = [];
@@ -1108,103 +1058,89 @@ export class OpenAIProvider extends BaseProvider {
         }
     }
     /**
-     * Update model parameters and persist them in the SettingsService.
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-002
+     * Memoization of model parameters disabled for stateless provider
      */
-    setModelParams(params) {
-        if (params === undefined) {
-            this.modelParams = undefined;
-            this.setModelParamsInSettings(undefined).catch((error) => {
-                this.logger.debug(() => `Failed to clear model params in SettingsService: ${error}`);
-            });
-            return;
-        }
-        const updated = { ...(this.modelParams ?? {}) };
-        for (const [key, value] of Object.entries(params)) {
-            if (value === undefined || value === null) {
-                delete updated[key];
-            }
-            else {
-                updated[key] = value;
-            }
-        }
-        this.modelParams = Object.keys(updated).length > 0 ? updated : undefined;
-        this.setModelParamsInSettings(this.modelParams).catch((error) => {
-            this.logger.debug(() => `Failed to persist model params to SettingsService: ${error}`);
-        });
+    setModelParams(_params) {
+        throw new Error('ProviderCacheError("Attempted to memoize model parameters for openai")');
     }
+    /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Gets model parameters from SettingsService per call (stateless)
+     */
     getModelParams() {
-        return this.modelParams;
+        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 OpenAI provider settings from SettingsService: ${error}`);
+            return undefined;
+        }
     }
     /**
-     * Get the tool format for this provider
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
+     * Get the tool format for this provider using normalized options
      * @returns The tool format to use
      */
     getToolFormat() {
         const format = this.detectToolFormat();
-        this.logger.debug(() => `getToolFormat() called, returning: ${format}`, {
+        const logger = new DebugLogger('llxprt:provider:openai');
+        logger.debug(() => `getToolFormat() called, returning: ${format}`, {
             provider: this.name,
             model: this.getModel(),
             format,
         });
         return format;
     }
-    /**
-     * Set tool format override for this provider
-     * @param format The format to use, or null to clear override
-     */
-    setToolFormatOverride(format) {
-        const settingsService = getSettingsService();
-        if (format === null) {
-            settingsService.setProviderSetting(this.name, 'toolFormat', 'auto');
-            this.logger.debug(() => `Tool format override cleared for ${this.name}`);
-        }
-        else {
-            settingsService.setProviderSetting(this.name, 'toolFormat', format);
-            this.logger.debug(() => `Tool format override set to '${format}' for ${this.name}`);
-        }
-        // Clear cached client to ensure new format takes effect
-        this._cachedClient = undefined;
-        this._cachedClientKey = undefined;
-    }
     /**
      * Detects the tool call format based on the model being used
      * @returns The detected tool format ('openai' or 'qwen')
      */
     detectToolFormat() {
-        try {
-            // Check for toolFormat override in provider settings
-            const settingsService = getSettingsService();
-            const currentSettings = settingsService['settings'];
-            const providerSettings = currentSettings?.providers?.[this.name];
-            const toolFormatOverride = providerSettings?.toolFormat;
-            // If explicitly set to a specific format (not 'auto'), use it
-            if (toolFormatOverride && toolFormatOverride !== 'auto') {
-                this.logger.debug(() => `Using tool format override '${toolFormatOverride}' for ${this.name}`);
-                return toolFormatOverride;
-            }
-        }
-        catch (error) {
-            this.logger.debug(() => `Failed to detect tool format from SettingsService: ${error}`);
-        }
         // Auto-detect based on model name if set to 'auto' or not set
         const modelName = (this.getModel() || this.getDefaultModel()).toLowerCase();
-        // Check for GLM models (glm-4.5, glm-4-6, etc.) which require Qwen handling
-        if (modelName.includes('glm-')) {
-            this.logger.debug(() => `Auto-detected 'qwen' format for GLM model: ${modelName}`);
-            return 'qwen';
-        }
-        // Check for MiniMax models (minimax, mini-max, etc.) which require Qwen handling
-        if (modelName.includes('minimax') || modelName.includes('mini-max')) {
-            this.logger.debug(() => `Auto-detected 'qwen' format for MiniMax model: ${modelName}`);
+        const logger = new DebugLogger('llxprt:provider:openai');
+        // Check for GLM-4 models (glm-4, glm-4.5, glm-4.6, glm-4-5, etc.)
+        if (modelName.includes('glm-4')) {
+            logger.debug(() => `Auto-detected 'qwen' format for GLM-4.x model: ${modelName}`);
             return 'qwen';
         }
         // Check for qwen models
         if (modelName.includes('qwen')) {
-            this.logger.debug(() => `Auto-detected 'qwen' format for Qwen model: ${modelName}`);
+            logger.debug(() => `Auto-detected 'qwen' format for Qwen model: ${modelName}`);
             return 'qwen';
         }
         // Default to 'openai' format
-        this.logger.debug(() => `Using default 'openai' format for model: ${modelName}`);
+        logger.debug(() => `Using default 'openai' format for model: ${modelName}`);
         return 'openai';
     }
     /**
@@ -1218,11 +1154,14 @@ export class OpenAIProvider extends BaseProvider {
         return response;
     }
     /**
+     * @plan:PLAN-20251023-STATELESS-HARDENING.P08
+     * @requirement:REQ-SP4-003
      * Determines whether a response should be retried based on error codes
      * @param error The error object from the API response
      * @returns true if the request should be retried, false otherwise
      */
     shouldRetryResponse(error) {
+        const logger = new DebugLogger('llxprt:provider:openai');
         // Don't retry if we're streaming chunks - just continue processing
         if (error &&
             typeof error === 'object' &&
@@ -1250,7 +1189,7 @@ export class OpenAIProvider extends BaseProvider {
             }
         }
         // Log what we're seeing
-        this.logger.debug(() => `shouldRetryResponse checking error:`, {
+        logger.debug(() => `shouldRetryResponse checking error:`, {
             hasError: !!error,
             errorType: error?.constructor?.name,
             status,
@@ -1258,15 +1197,11 @@ export class OpenAIProvider extends BaseProvider {
             errorKeys: error && typeof error === 'object' ? Object.keys(error) : [],
         });
         // Retry on 429 rate limit errors or 5xx server errors
-        if (status === 429 || (status && status >= 500 && status < 600)) {
-            this.logger.debug(() => `Will retry request due to status ${status}`);
-            return true;
+        const shouldRetry = Boolean(status === 429 || (status && status >= 500 && status < 600));
+        if (shouldRetry) {
+            logger.debug(() => `Will retry request due to status ${status}`);
         }
-        if (isNetworkTransientError(error)) {
-            this.logger.debug(() => 'Will retry request due to transient network error signature (connection-level failure).');
-            return true;
-        }
-        return false;
+        return shouldRetry;
     }
 }
 //# sourceMappingURL=OpenAIProvider.js.map