@modular-prompt/driver 0.13.5 → 0.15.0

package/dist/mlx-ml/mlx-driver.d.ts

@@ -1,22 +1,23 @@
-import type { AIDriver, QueryOptions, QueryResult, StreamResult } from '../types.js';
-import type { FormatterOptions } from '../formatter/types.js';
-import type { MlxMlModelOptions, MlxModelCapabilities } from './types.js';
-import type { CompiledPrompt } from '@modular-prompt/core';
+import type { MlxBackendMode } from '../driver-registry/types.js';
+import { LocalInferenceDriver } from '../local-inference/driver.js';
+import { hasMessageElement } from '../local-inference/prompt-utils.js';
 import type { PromptCacheController } from '../cache-controller.js';
-/**
- * Check if the prompt contains MessageElement
- */
-export declare function hasMessageElement(prompt: CompiledPrompt): boolean;
+import type { MlxModelCapabilities } from './types.js';
+import type { MlxQueryOptions } from './mlx-options.js';
+import type { FormatterOptions } from '../formatter/types.js';
+export { hasMessageElement };
 /**
  * MLX ML driver configuration
  */
 export interface MlxDriverConfig {
     model: string;
-    defaultOptions?: Partial<MlxMlModelOptions>;
+    defaultOptions?: Partial<MlxQueryOptions>;
     formatterOptions?: FormatterOptions;
     /** VLM画像の最大辺ピクセル数（デフォルト: 768） */
     maxImageSize?: number;
-    /** VLMモデルをtext-onlyモードで使用する（VLM判定を抑制） */
+    /** MLX Python バックエンド（デフォルト: auto） */
+    backend?: MlxBackendMode;
+    /** VLMモデルをtext-onlyモードで使用する（backend: lm と同等。後方互換） */
     textOnly?: boolean;
     /** Speculative decoding用のdrafter model名 */
     drafterModel?: string;
@@ -26,63 +27,23 @@ export interface MlxDriverConfig {
     cacheController?: PromptCacheController;
 }
 /**
- * MLX ML driver using Python subprocess
+ * MLX ML driver using Python subprocess.
+ * 共通ロジックは LocalInferenceDriver に委譲する。
  */
-export declare class MlxDriver implements AIDriver {
-    private process;
-    private model;
-    private _defaultOptions;
-    private runtimeInfo;
-    private modelProcessor;
-    private formatterOptions;
-    private maxImageSize;
-    private queryLogger;
-    private cacheController?;
-    private cacheControllerBound;
-    get defaultOptions(): Partial<MlxMlModelOptions>;
-    set defaultOptions(value: Partial<MlxMlModelOptions>);
+export declare class MlxDriver extends LocalInferenceDriver {
+    private cacheControllerRaw?;
+    private cacheDisabledForVlm;
+    private readonly cacheBindingState;
     constructor(config: MlxDriverConfig);
-    /**
-     * Initialize process and cache runtime info
-     */
-    private ensureInitialized;
-    /**
-     * VLMモデルかどうかを判定
-     */
-    private isVLM;
-    /**
-     * Determine which API to use (chat or completion)
-     * Simple logic based on runtime info only
-     */
-    private determineApi;
-    /**
-     * モデルがnativeツール対応かを判定
-     * tool_call_format（Python側検出結果）を唯一の判断基準とする
-     */
-    private hasNativeToolSupport;
-    /**
-     * Execute query and return stream
-     * Common logic for query and streamQuery
-     */
-    private executeQuery;
-    /**
-     * Query the AI model with a compiled prompt
-     */
-    query(prompt: CompiledPrompt, options?: QueryOptions): Promise<QueryResult>;
-    /**
-     * Stream query implementation
-     */
-    streamQuery(prompt: CompiledPrompt, options?: QueryOptions): Promise<StreamResult>;
+    get defaultOptions(): Partial<MlxQueryOptions>;
+    set defaultOptions(value: Partial<MlxQueryOptions>);
     /**
      * Get model capabilities (public API)
      *
      * Returns runtime information converted to camelCase
      */
     getCapabilities(): Promise<MlxModelCapabilities>;
-    private logCacheStats;
-    /**
-     * Close the process
-     */
     close(): Promise<void>;
+    private logCacheStats;
 }
 //# sourceMappingURL=mlx-driver.d.ts.map

package/dist/mlx-ml/mlx-driver.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"mlx-driver.d.ts","sourceRoot":"","sources":["../../src/mlx-ml/mlx-driver.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,QAAQ,EAAE,YAAY,EAAE,WAAW,EAAE,YAAY,EAAgB,MAAM,aAAa,CAAC;AAEnG,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAI9D,OAAO,KAAK,EAAE,iBAAiB,EAAE,oBAAoB,EAAE,MAAM,YAAY,CAAC;AAI1E,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAK3D,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAQpE;;GAEG;AACH,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,cAAc,GAAG,OAAO,CAYjE;AAMD;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,KAAK,EAAE,MAAM,CAAC;IACd,cAAc,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC,CAAC;IAC5C,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,iCAAiC;IACjC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,yCAAyC;IACzC,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,2CAA2C;IAC3C,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,2DAA2D;IAC3D,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,0BAA0B;IAC1B,eAAe,CAAC,EAAE,qBAAqB,CAAC;CACzC;AAyED;;GAEG;AACH,qBAAa,SAAU,YAAW,QAAQ;IACxC,OAAO,CAAC,OAAO,CAAa;IAC5B,OAAO,CAAC,KAAK,CAAS;IACtB,OAAO,CAAC,eAAe,CAA6B;IACpD,OAAO,CAAC,WAAW,CAA+B;IAClD,OAAO,CAAC,cAAc,CAAC;IACvB,OAAO,CAAC,gBAAgB,CAAmB;IAC3C,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,WAAW,CAA0B;IAC7C,OAAO,CAAC,eAAe,CAAC,CAAwB;IAChD,OAAO,CAAC,oBAAoB,CAAS;IAErC,IAAI,cAAc,IAAI,OAAO,CAAC,iBAAiB,CAAC,CAE/C;IAED,IAAI,cAAc,CAAC,KAAK,EAAE,OAAO,CAAC,iBAAiB,CAAC,EAEnD;gBAEW,MAAM,EAAE,eAAe;IAiBnC;;OAEG;YACW,iBAAiB;IA2C/B;;OAEG;IACH,OAAO,CAAC,KAAK;IAIb;;;OAGG;IACH,OAAO,CAAC,YAAY;IASpB;;;OAGG;IACH,OAAO,CAAC,oBAAoB;IAI5B;;;OAGG;YACW,YAAY;IA4F1B;;OAEG;IACG,KAAK,CAAC,MAAM,EAAE,cAAc,EAAE,OAAO,CAAC,EAAE,YAAY,GAAG,OAAO,CAAC,WAAW,CAAC;IAajF;;OAEG;IACG,WAAW,CACf,MAAM,EAAE,cAAc,EACtB,OAAO,CAAC,EAAE,YAAY,GACrB,OAAO,CAAC,YAAY,CAAC;IA2HxB;;;;OAIG;IACG,eAAe,IAAI,OAAO,CAAC,oBAAoB,CAAC;IAsCtD,OAAO,CAAC,aAAa;IAqBrB;;OAEG;IACG,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;CAK7B"}
+{"version":3,"file":"mlx-driver.d.ts","sourceRoot":"","sources":["../../src/mlx-ml/mlx-driver.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,6BAA6B,CAAC;AAClE,OAAO,EAAE,oBAAoB,EAAE,MAAM,8BAA8B,CAAC;AACpE,OAAO,EAAE,iBAAiB,EAAE,MAAM,oCAAoC,CAAC;AACvE,OAAO,KAAK,EAAE,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AACpE,OAAO,KAAK,EAAE,oBAAoB,EAAE,MAAM,YAAY,CAAC;AACvD,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,kBAAkB,CAAC;AACxD,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAS9D,OAAO,EAAE,iBAAiB,EAAE,CAAC;AAE7B;;GAEG;AACH,MAAM,WAAW,eAAe;IAC9B,KAAK,EAAE,MAAM,CAAC;IACd,cAAc,CAAC,EAAE,OAAO,CAAC,eAAe,CAAC,CAAC;IAC1C,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,iCAAiC;IACjC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,qCAAqC;IACrC,OAAO,CAAC,EAAE,cAAc,CAAC;IACzB,qDAAqD;IACrD,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,2CAA2C;IAC3C,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,2DAA2D;IAC3D,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,0BAA0B;IAC1B,eAAe,CAAC,EAAE,qBAAqB,CAAC;CACzC;AAED;;;GAGG;AACH,qBAAa,SAAU,SAAQ,oBAAoB;IACjD,OAAO,CAAC,kBAAkB,CAAC,CAAwB;IACnD,OAAO,CAAC,mBAAmB,CAAS;IACpC,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAoB;gBAE1C,MAAM,EAAE,eAAe;IA6CnC,IAAI,cAAc,IAAI,OAAO,CAAC,eAAe,CAAC,CAE7C;IAED,IAAI,cAAc,CAAC,KAAK,EAAE,OAAO,CAAC,eAAe,CAAC,EAEjD;IAED;;;;OAIG;IACG,eAAe,IAAI,OAAO,CAAC,oBAAoB,CAAC;IA8CvC,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;IAQrC,OAAO,CAAC,aAAa;CAoBtB"}

package/dist/mlx-ml/mlx-driver.js

@@ -1,389 +1,57 @@
-import { isToolResult } from '../types.js';
-import { formatPromptAsMessages } from '../formatter/converter.js';
-import { formatCompletionPrompt } from '../formatter/completion-formatter.js';
+import { LocalInferenceDriver } from '../local-inference/driver.js';
+import { hasMessageElement } from '../local-inference/prompt-utils.js';
 import { MlxProcess } from './process/index.js';
-import { createModelSpecificProcessor, selectApi } from './process/model-specific.js';
-import { selectResponseProcessor } from './process/model-handlers.js';
-import { extractJSON } from '@modular-prompt/utils';
-import { formatToolDefinitionsAsText } from './tool-call-parser/index.js';
-import { convertMessages, convertToolDefinitions, extractImagePaths } from './mlx-message-utils.js';
-import { QueryLogger } from '../query-logger.js';
-import { extractCacheablePrefix } from '../cache-utils.js';
+import { mlxLocalInferenceAdapters } from './mlx-local-inference-adapters.js';
+import { bindMlxCacheOnCapabilitiesLoaded, createMlxCacheSupport, } from './mlx-cache-support.js';
 import { MlxCacheController } from './mlx-cache-controller.js';
-// ========================================================================
-// Utility Functions (exported for testing)
-// ========================================================================
+export { hasMessageElement };
 /**
- * Check if the prompt contains MessageElement
+ * MLX ML driver using Python subprocess.
+ * 共通ロジックは LocalInferenceDriver に委譲する。
  */
-export function hasMessageElement(prompt) {
-    const checkElements = (elements) => {
-        if (!elements)
-            return false;
-        return elements.some(element => {
-            const el = element;
-            return el?.type === 'message';
-        });
-    };
-    return checkElements(prompt.instructions) ||
-        checkElements(prompt.data) ||
-        checkElements(prompt.output);
-}
-const META_MARKER = '\x1e__META__:';
-function extractStreamMeta(content) {
-    const idx = content.lastIndexOf(META_MARKER);
-    if (idx === -1)
-        return { content, meta: {} };
-    const jsonStr = content.slice(idx + META_MARKER.length);
-    try {
-        return { content: content.slice(0, idx), meta: JSON.parse(jsonStr) };
-    }
-    catch {
-        return { content: content.slice(0, idx), meta: {} };
-    }
-}
-function createStreamIterable(stream) {
-    const chunks = [];
-    let resolveCompletion;
-    const completion = new Promise((resolve) => {
-        resolveCompletion = resolve;
-    });
-    const iterable = {
-        async *[Symbol.asyncIterator]() {
-            try {
-                let buffer = '';
-                let markerFound = false;
-                for await (const chunk of stream) {
-                    const str = chunk.toString();
-                    chunks.push(str);
-                    if (markerFound)
-                        continue;
-                    buffer += str;
-                    const markerIdx = buffer.indexOf(META_MARKER);
-                    if (markerIdx !== -1) {
-                        const text = buffer.slice(0, markerIdx);
-                        if (text)
-                            yield text;
-                        markerFound = true;
-                    }
-                    else {
-                        const safeLen = buffer.length - (META_MARKER.length - 1);
-                        if (safeLen > 0) {
-                            yield buffer.slice(0, safeLen);
-                            buffer = buffer.slice(safeLen);
-                        }
-                    }
-                }
-                if (!markerFound && buffer)
-                    yield buffer;
-                const raw = chunks.join('');
-                const { content, meta } = extractStreamMeta(raw);
-                resolveCompletion({ content, meta, error: null });
-            }
-            catch (error) {
-                const raw = chunks.join('');
-                const { content, meta } = extractStreamMeta(raw);
-                resolveCompletion({ content, meta, error: error });
-                throw error;
-            }
-        }
-    };
-    return { iterable, completion };
-}
-/**
- * MLX ML driver using Python subprocess
- */
-export class MlxDriver {
-    process;
-    model;
-    _defaultOptions;
-    runtimeInfo = null;
-    modelProcessor;
-    formatterOptions;
-    maxImageSize;
-    queryLogger = new QueryLogger('MLX');
-    cacheController;
-    cacheControllerBound = false;
-    get defaultOptions() {
-        return this._defaultOptions;
-    }
-    set defaultOptions(value) {
-        this._defaultOptions = value;
-    }
+export class MlxDriver extends LocalInferenceDriver {
+    cacheControllerRaw;
+    cacheDisabledForVlm = false;
+    cacheBindingState = { bound: false };
     constructor(config) {
-        this.model = config.model;
-        this._defaultOptions = config.defaultOptions || {};
-        this.formatterOptions = config.formatterOptions || {};
-        this.maxImageSize = config.maxImageSize ?? 768;
-        this.process = new MlxProcess(config.model, {
+        const mlxProcess = new MlxProcess(config.model, {
+            backend: config.backend,
             textOnly: config.textOnly,
             drafterModel: config.drafterModel,
-            draftBlockSize: config.draftBlockSize
+            draftBlockSize: config.draftBlockSize,
+        });
+        const cacheSupport = config.cacheController
+            ? createMlxCacheSupport(config.cacheController)
+            : undefined;
+        super({
+            model: config.model,
+            process: mlxProcess,
+            adapters: mlxLocalInferenceAdapters,
+            formatterOptions: config.formatterOptions,
+            maxImageSize: config.maxImageSize,
+            defaultOptions: config.defaultOptions,
+            loggerPrefix: 'MLX',
+            cache: cacheSupport,
+            onCapabilitiesLoaded: async (runtimeInfo, ctx) => {
+                if (!cacheSupport)
+                    return;
+                await bindMlxCacheOnCapabilitiesLoaded(cacheSupport, this.cacheBindingState, runtimeInfo, ctx, () => {
+                    this.queryLogger.log.info('VLM models do not support prompt caching — cacheController disabled');
+                    this.cacheDisabledForVlm = true;
+                    this.disableCacheSupport();
+                });
+            },
         });
-        this.modelProcessor = createModelSpecificProcessor(config.model);
-        this.cacheController = config.cacheController;
+        this.cacheControllerRaw = config.cacheController;
         if (config.drafterModel) {
             this.queryLogger.log.info(`Drafter model: ${config.drafterModel}`);
         }
     }
-    /**
-     * Initialize process and cache runtime info
-     */
-    async ensureInitialized() {
-        // Ensure process is initialized
-        await this.process.ensureInitialized();
-        // Cache runtime info if not already cached
-        if (!this.runtimeInfo) {
-            try {
-                this.runtimeInfo = await this.process.getCapabilities();
-                // Update formatterOptions with special tokens from runtime info
-                if (this.runtimeInfo.special_tokens) {
-                    this.formatterOptions.specialTokens = this.runtimeInfo.special_tokens;
-                }
-                // Update model processor with runtime context
-                this.modelProcessor.setRuntimeContext({
-                    chatRestrictions: this.runtimeInfo.chat_restrictions,
-                    modelKind: this.runtimeInfo.model_kind,
-                });
-                // Bind cache controller if provided and not yet bound
-                // NOTE: instanceof guard means VLM check only covers MlxCacheController.
-                // A custom PromptCacheController on a VLM model would bypass this — add a
-                // model-kind guard here if another implementation is introduced.
-                if (this.cacheController instanceof MlxCacheController && !this.cacheControllerBound) {
-                    if (this.runtimeInfo.model_kind === 'vlm') {
-                        this.queryLogger.log.info('VLM models do not support prompt caching — cacheController disabled');
-                        this.cacheController = undefined;
-                    }
-                    else {
-                        await this.cacheController.bind(this.process, this.formatterOptions, (msgs) => this.modelProcessor.applyChatSpecificProcessing(msgs));
-                        this.cacheControllerBound = true;
-                    }
-                }
-            }
-            catch (error) {
-                this.queryLogger.log.error('Failed to get MLX runtime info:', error instanceof Error ? error.message : String(error));
-            }
-        }
-    }
-    /**
-     * VLMモデルかどうかを判定
-     */
-    isVLM() {
-        return this.runtimeInfo?.model_kind === 'vlm';
-    }
-    /**
-     * Determine which API to use (chat or completion)
-     * Simple logic based on runtime info only
-     */
-    determineApi(options) {
-        return selectApi(options?.apiStrategy || 'auto', options?.mode, !!this.runtimeInfo?.features.apply_chat_template, this.modelProcessor.hasCompletionProcessor());
-    }
-    /**
-     * モデルがnativeツール対応かを判定
-     * tool_call_format（Python側検出結果）を唯一の判断基準とする
-     */
-    hasNativeToolSupport() {
-        return !!this.runtimeInfo?.features?.chat_template?.tool_call_format?.call_start;
-    }
-    /**
-     * Execute query and return stream
-     * Common logic for query and streamQuery
-     */
-    async executeQuery(prompt, mlxOptions, options) {
-        // APIを選択
-        const api = this.determineApi(options);
-        // tools変換
-        const tools = options?.tools ? convertToolDefinitions(options.tools) : undefined;
-        // completion API または nativeツール非対応の場合、tool定義をテキストとしてプロンプトに注入
-        let augmentedPrompt = prompt;
-        if (options?.tools && options.tools.length > 0 && (api === 'completion' || !this.hasNativeToolSupport())) {
-            const toolsText = formatToolDefinitionsAsText(options.tools, this.runtimeInfo?.special_tokens, this.runtimeInfo?.features?.chat_template?.tool_call_format);
-            augmentedPrompt = {
-                ...prompt,
-                instructions: [
-                    ...prompt.instructions,
-                    { type: 'text', content: toolsText }
-                ]
-            };
-        }
-        // Record all queries, regardless of API mode or cache usage
-        this.cacheController?.recordQuery?.();
-        let stream;
-        if (api === 'completion') {
-            let formattedPrompt = formatCompletionPrompt(augmentedPrompt, this.formatterOptions);
-            formattedPrompt = this.modelProcessor.applyCompletionSpecificProcessing(formattedPrompt);
-            stream = await this.process.completion(formattedPrompt, mlxOptions);
-        }
-        else {
-            const messages = formatPromptAsMessages(augmentedPrompt, this.formatterOptions);
-            const vlm = this.isVLM();
-            let mlxMessages = convertMessages(messages, vlm);
-            mlxMessages = this.modelProcessor.applyChatSpecificProcessing(mlxMessages);
-            const nativeTools = this.hasNativeToolSupport() && tools?.length ? tools : undefined;
-            const images = vlm
-                ? messages.flatMap(m => 'content' in m && !isToolResult(m) ? extractImagePaths(m.content) : [])
-                : [];
-            // Cache: chat APIのみ、以下の条件を全て満たす場合にキャッシュを使用
-            // - options.cache !== false（呼び出し側が明示的に無効化していない）
-            // - trustRemoteCode未指定（明示的なtrue/falseどちらもapply_chat_template kwargsに影響）
-            let cachePath;
-            let cacheTrimTokens;
-            const trustRemoteCode = mlxOptions.trustRemoteCode;
-            if (this.cacheController && options?.cache !== false && trustRemoteCode === undefined) {
-                const prefix = extractCacheablePrefix(augmentedPrompt);
-                const hasCacheableContent = prefix.instructions.length > 0 ||
-                    prefix.data.length > 0;
-                if (hasCacheableContent) {
-                    const cacheStart = performance.now();
-                    const handle = await this.cacheController.prepare({
-                        model: this.model,
-                        instructions: prefix.instructions,
-                        data: prefix.data,
-                        tools: nativeTools ? options.tools : undefined,
-                        reasoningEffort: options?.reasoningEffort,
-                        readOnly: options?.cache === 'read-only',
-                    });
-                    cachePath = handle.ref || undefined;
-                    cacheTrimTokens = handle.trimTokens;
-                    if (cachePath) {
-                        this.queryLogger.log.debug(`cache prepare ${(performance.now() - cacheStart).toFixed(0)}ms`, `(${prefix.instructions.length}i+${prefix.data.length}d)`, cacheTrimTokens != null ? `trim=${cacheTrimTokens}` : '');
-                    }
-                }
-            }
-            stream = await this.process.chat(mlxMessages, undefined, mlxOptions, nativeTools, images.length > 0 ? images : undefined, images.length > 0 ? this.maxImageSize : undefined, options?.reasoningEffort, cachePath, cacheTrimTokens);
-            const cacheTokensUsed = cachePath
-                ? (cacheTrimTokens ?? (this.cacheController instanceof MlxCacheController
-                    ? this.cacheController.readCacheTokenCount(cachePath) : 0))
-                : 0;
-            return { stream, cacheTokensUsed };
-        }
-        return { stream, cacheTokensUsed: 0 };
-    }
-    /**
-     * Query the AI model with a compiled prompt
-     */
-    async query(prompt, options) {
-        // Use streamQuery for consistency with other drivers
-        const { stream, result } = await this.streamQuery(prompt, options);
-        // Consume the stream to trigger completion
-        // This is necessary because the result promise only resolves when the stream is fully consumed
-        // eslint-disable-next-line @typescript-eslint/no-unused-vars
-        for await (const _chunk of stream) {
-            // Just consume the stream, don't need to do anything with the chunks
-        }
-        return result;
+    get defaultOptions() {
+        return super.defaultOptions;
     }
-    /**
-     * Stream query implementation
-     */
-    async streamQuery(prompt, options) {
-        await this.ensureInitialized();
-        // Merge options (only override if explicitly provided)
-        const mlxOptions = {
-            ...this.defaultOptions,
-            ...(options?.maxTokens !== undefined && { maxTokens: options.maxTokens }),
-            ...(options?.temperature !== undefined && { temperature: options.temperature }),
-            ...(options?.topP !== undefined && { topP: options.topP }),
-            ...(options?.topK !== undefined && { topK: options.topK }),
-        };
-        this.queryLogger.mark(mlxOptions);
-        // Use executeQuery for the actual stream generation
-        const queryStart = performance.now();
-        const { stream, cacheTokensUsed } = await this.executeQuery(prompt, mlxOptions, options);
-        const streamStart = performance.now();
-        this.queryLogger.log.debug(`setup ${(streamStart - queryStart).toFixed(0)}ms`);
-        // Convert stream to async iterable with collection
-        const { iterable, completion } = createStreamIterable(stream);
-        // Wrap iterable with phase-separated timing:
-        //   TTFT: streamStart → first chunk (inference latency, shows cache benefit)
-        //   generation: first chunk → last chunk (output speed, tok/s)
-        //   query total: queryStart → last chunk (end-to-end)
-        const queryLogger = this.queryLogger;
-        let firstChunkTime = 0;
-        const wrappedIterable = {
-            [Symbol.asyncIterator]() {
-                const inner = iterable[Symbol.asyncIterator]();
-                let firstChunk = true;
-                return {
-                    async next() {
-                        const result = await inner.next();
-                        if (!result.done) {
-                            if (firstChunk) {
-                                firstChunk = false;
-                                firstChunkTime = performance.now();
-                                queryLogger.log.debug(`TTFT ${(firstChunkTime - streamStart).toFixed(0)}ms`);
-                            }
-                        }
-                        if (result.done) {
-                            const now = performance.now();
-                            if (firstChunkTime > 0) {
-                                const genMs = now - firstChunkTime;
-                                queryLogger.log.debug(`generation ${genMs.toFixed(0)}ms (query total ${(now - queryStart).toFixed(0)}ms)`);
-                            }
-                            else {
-                                queryLogger.log.debug(`query total ${(performance.now() - queryStart).toFixed(0)}ms`);
-                            }
-                        }
-                        return result;
-                    },
-                    async return(value) {
-                        return inner.return?.(value) ?? { done: true, value: undefined };
-                    },
-                    async throw(e) {
-                        return inner.throw?.(e) ?? { done: true, value: undefined };
-                    },
-                };
-            },
-        };
-        // Create result promise that waits for stream completion
-        const cacheController = this.cacheController;
-        const resultPromise = completion.then(({ content, meta, error }) => {
-            // If there was an error, log and throw it
-            if (error) {
-                this.queryLogger.log.error('Stream error:', error.message);
-                throw error;
-            }
-            if (cacheController instanceof MlxCacheController && meta.prompt_tokens != null) {
-                cacheController.recordPromptTokens(meta.prompt_tokens, cacheTokensUsed);
-            }
-            // Log accurate generation stats if available
-            if (meta.generation_tokens != null && firstChunkTime > 0) {
-                const genMs = performance.now() - firstChunkTime;
-                const actualTps = (meta.generation_tokens / genMs * 1000).toFixed(1);
-                this.queryLogger.log.debug(`${meta.generation_tokens} tokens, ${actualTps} tok/s`);
-            }
-            // Response post-processing: thinking抽出 + tool call解析
-            const hasTools = options?.tools && options.tools.length > 0;
-            const responseProcessor = selectResponseProcessor(this.model, this.runtimeInfo, { enableToolParsing: !!hasTools });
-            const parsed = responseProcessor(content);
-            let finalContent = parsed.content;
-            const thinkingContent = parsed.thinkingContent;
-            const toolCalls = parsed.toolCalls;
-            if (thinkingContent) {
-                this.queryLogger.log.verbose('Thinking content:', thinkingContent);
-            }
-            // Handle structured output if schema is provided
-            let structuredOutput;
-            if (prompt.metadata?.outputSchema && finalContent) {
-                const extracted = extractJSON(finalContent, { multiple: false });
-                if (extracted.source !== 'none' && extracted.data !== null) {
-                    structuredOutput = extracted.data;
-                }
-            }
-            const finishReason = toolCalls ? 'tool_calls' : 'stop';
-            return {
-                content: finalContent,
-                thinkingContent,
-                structuredOutput,
-                toolCalls,
-                finishReason,
-                ...this.queryLogger.collect()
-            };
-        });
-        return {
-            stream: wrappedIterable,
-            result: resultPromise
-        };
+    set defaultOptions(value) {
+        super.defaultOptions = value ?? {};
     }
     /**
      * Get model capabilities (public API)
@@ -392,51 +60,62 @@ export class MlxDriver {
      */
     async getCapabilities() {
         await this.ensureInitialized();
-        if (!this.runtimeInfo) {
+        const runtimeInfo = this.getRuntimeInfo();
+        if (!runtimeInfo) {
             throw new Error('Failed to retrieve model capabilities');
         }
-        // Convert snake_case to camelCase
         return {
-            methods: this.runtimeInfo.methods,
-            specialTokens: this.runtimeInfo.special_tokens,
+            methods: runtimeInfo.methods,
+            specialTokens: runtimeInfo.special_tokens,
             features: {
-                hasChatTemplate: this.runtimeInfo.features.apply_chat_template,
-                vocabSize: this.runtimeInfo.features.vocab_size,
-                modelMaxLength: this.runtimeInfo.features.model_max_length,
-                chatTemplate: this.runtimeInfo.features.chat_template ? {
-                    supportedRoles: this.runtimeInfo.features.chat_template.supported_roles,
-                    preview: this.runtimeInfo.features.chat_template.preview,
-                    constraints: this.runtimeInfo.features.chat_template.constraints,
-                    toolCallFormat: this.runtimeInfo.features.chat_template.tool_call_format ? {
-                        toolParserType: this.runtimeInfo.features.chat_template.tool_call_format.tool_parser_type,
-                        callStart: this.runtimeInfo.features.chat_template.tool_call_format.call_start,
-                        callEnd: this.runtimeInfo.features.chat_template.tool_call_format.call_end,
-                        responseStart: this.runtimeInfo.features.chat_template.tool_call_format.response_start,
-                        responseEnd: this.runtimeInfo.features.chat_template.tool_call_format.response_end,
-                    } : undefined
-                } : undefined
+                hasChatTemplate: runtimeInfo.features.apply_chat_template,
+                vocabSize: runtimeInfo.features.vocab_size,
+                modelMaxLength: runtimeInfo.features.model_max_length,
+                chatTemplate: runtimeInfo.features.chat_template
+                    ? {
+                        supportedRoles: runtimeInfo.features.chat_template.supported_roles,
+                        preview: runtimeInfo.features.chat_template.preview,
+                        constraints: runtimeInfo.features.chat_template.constraints,
+                        toolCallFormat: runtimeInfo.features.chat_template.tool_call_format
+                            ? {
+                                toolParserType: runtimeInfo.features.chat_template.tool_call_format.tool_parser_type,
+                                callStart: runtimeInfo.features.chat_template.tool_call_format.call_start,
+                                callEnd: runtimeInfo.features.chat_template.tool_call_format.call_end,
+                                responseStart: runtimeInfo.features.chat_template.tool_call_format.response_start,
+                                responseEnd: runtimeInfo.features.chat_template.tool_call_format.response_end,
+                            }
+                            : undefined,
+                    }
+                    : undefined,
             },
-            chatRestrictions: this.runtimeInfo.chat_restrictions ? {
-                singleSystemAtStart: this.runtimeInfo.chat_restrictions.single_system_at_start,
-                maxSystemMessages: this.runtimeInfo.chat_restrictions.max_system_messages,
-                alternatingTurns: this.runtimeInfo.chat_restrictions.alternating_turns,
-                requiresUserLast: this.runtimeInfo.chat_restrictions.requires_user_last,
-                allowEmptyMessages: this.runtimeInfo.chat_restrictions.allow_empty_messages,
-            } : undefined
+            chatRestrictions: runtimeInfo.chat_restrictions
+                ? {
+                    singleSystemAtStart: runtimeInfo.chat_restrictions.single_system_at_start,
+                    maxSystemMessages: runtimeInfo.chat_restrictions.max_system_messages,
+                    alternatingTurns: runtimeInfo.chat_restrictions.alternating_turns,
+                    requiresUserLast: runtimeInfo.chat_restrictions.requires_user_last,
+                    allowEmptyMessages: runtimeInfo.chat_restrictions.allow_empty_messages,
+                }
+                : undefined,
         };
     }
+    async close() {
+        this.logCacheStats();
+        if (this.cacheDisabledForVlm) {
+            await this.cacheControllerRaw?.close();
+        }
+        await super.close();
+    }
     logCacheStats() {
-        if (!(this.cacheController instanceof MlxCacheController))
+        if (this.cacheDisabledForVlm)
             return;
-        const s = this.cacheController.getStats();
+        if (!(this.cacheControllerRaw instanceof MlxCacheController))
+            return;
+        const s = this.cacheControllerRaw.getStats();
         if (s.totalQueries === 0)
             return;
-        const queryBreakdown = s.incremental + s.fresh > 0
-            ? ` (incremental ${s.incremental}, fresh ${s.fresh})`
-            : '';
-        const parts = [
-            `cache stats: ${s.totalQueries} queries${queryBreakdown}`,
-        ];
+        const queryBreakdown = s.incremental + s.fresh > 0 ? ` (incremental ${s.incremental}, fresh ${s.fresh})` : '';
+        const parts = [`cache stats: ${s.totalQueries} queries${queryBreakdown}`];
         if (s.totalPromptTokens > 0) {
             const reusedRate = ((s.prefillReusedTokens / s.totalPromptTokens) * 100).toFixed(0);
             parts.push(`prompt ${s.totalPromptTokens} tokens, ${s.prefillReusedTokens} reused (${reusedRate}%)`);
@@ -446,13 +125,5 @@ export class MlxDriver {
         }
         this.queryLogger.log.verbose(parts.join(' | '));
     }
-    /**
-     * Close the process
-     */
-    async close() {
-        this.logCacheStats();
-        await this.cacheController?.close();
-        await this.process.exit();
-    }
 }
 //# sourceMappingURL=mlx-driver.js.map