genai-lite 0.4.2 → 0.4.3

package/README.md

@@ -151,6 +151,8 @@ const llmService = new LLMService(myKeyProvider);
 
 Run models locally via [llama.cpp](https://github.com/ggml-org/llama.cpp) server. Model IDs can be any name—they're not validated since you load your own GGUF models.
 
+**Automatic Capability Detection:** genai-lite automatically detects capabilities (reasoning support, context windows, token limits) for known open-weights models (Qwen3, etc.) by matching the GGUF filename from the server. No configuration needed.
+
 **Example models:**
 - `llama-3-8b-instruct` - Llama 3 8B Instruct
 - `llama-3-70b-instruct` - Llama 3 70B Instruct
@@ -212,6 +214,7 @@ Some models include advanced reasoning/thinking capabilities that enhance their
 - **Anthropic**: Claude Sonnet 4, Claude Opus 4, Claude 3.7 Sonnet
 - **Google Gemini**: Gemini 2.5 Pro (always on), Gemini 2.5 Flash, Gemini 2.5 Flash-Lite Preview
 - **OpenAI**: o4-mini (always on)
+- **llama.cpp**: Qwen3, DeepSeek-R1, GPT-OSS (requires `--reasoning-format deepseek` server flag)
 
 See the [Reasoning Mode](#reasoning-mode) section for usage details.
 
@@ -794,12 +797,20 @@ Get GGUF models from Hugging Face, for example:
 # Basic usage
 llama-server -m /path/to/model.gguf --port 8080
 
-# With more options
+# With reasoning support (for Qwen3, DeepSeek-R1, etc.)
+llama-server -m /path/to/model.gguf \
+  --port 8080 \
+  --jinja \
+  --reasoning-format deepseek
+
+# Full options
 llama-server -m /path/to/model.gguf \
   --port 8080 \
-  -c 4096 \           # Context size
-  -np 4 \             # Parallel requests
-  --threads 8         # CPU threads
+  --jinja \                    # Required for reasoning
+  --reasoning-format deepseek \ # Extract reasoning from <think> tags
+  -c 4096 \                    # Context size
+  -np 4 \                      # Parallel requests
+  --threads 8                  # CPU threads
 ```
 
 ### Basic Usage
@@ -1077,7 +1088,9 @@ import type {
 import {
   LlamaCppClientAdapter,
   LlamaCppServerClient,
-  createFallbackModelInfo
+  createFallbackModelInfo,
+  detectGgufCapabilities,
+  KNOWN_GGUF_MODELS
 } from 'genai-lite';
 
 import type {
@@ -1090,7 +1103,10 @@ import type {
   LlamaCppPropsResponse,
   LlamaCppMetricsResponse,
   LlamaCppSlot,
-  LlamaCppSlotsResponse
+  LlamaCppSlotsResponse,
+  LlamaCppModel,
+  LlamaCppModelsResponse,
+  GgufModelPattern
 } from 'genai-lite';
 ```
 

package/dist/index.d.ts

@@ -8,9 +8,10 @@ export { fromEnvironment } from "./providers/fromEnvironment";
 export { LlamaCppClientAdapter } from "./llm/clients/LlamaCppClientAdapter";
 export { LlamaCppServerClient } from "./llm/clients/LlamaCppServerClient";
 export type { LlamaCppClientConfig, } from "./llm/clients/LlamaCppClientAdapter";
-export type { LlamaCppHealthResponse, LlamaCppTokenizeResponse, LlamaCppDetokenizeResponse, LlamaCppEmbeddingResponse, LlamaCppInfillResponse, LlamaCppPropsResponse, LlamaCppMetricsResponse, LlamaCppSlot, LlamaCppSlotsResponse, } from "./llm/clients/LlamaCppServerClient";
+export type { LlamaCppHealthResponse, LlamaCppTokenizeResponse, LlamaCppDetokenizeResponse, LlamaCppEmbeddingResponse, LlamaCppInfillResponse, LlamaCppPropsResponse, LlamaCppMetricsResponse, LlamaCppSlot, LlamaCppSlotsResponse, LlamaCppModel, LlamaCppModelsResponse, } from "./llm/clients/LlamaCppServerClient";
 export { renderTemplate } from "./prompting/template";
 export { countTokens, getSmartPreview, extractRandomVariables } from "./prompting/content";
 export { parseStructuredContent, parseRoleTags, extractInitialTaggedContent, parseTemplateWithMetadata } from "./prompting/parser";
 export type { TemplateMetadata } from "./prompting/parser";
-export { createFallbackModelInfo } from "./llm/config";
+export { createFallbackModelInfo, detectGgufCapabilities, KNOWN_GGUF_MODELS } from "./llm/config";
+export type { GgufModelPattern } from "./llm/config";

package/dist/index.js

@@ -14,7 +14,7 @@ var __exportStar = (this && this.__exportStar) || function(m, exports) {
     for (var p in m) if (p !== "default" && !Object.prototype.hasOwnProperty.call(exports, p)) __createBinding(exports, m, p);
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.createFallbackModelInfo = exports.parseTemplateWithMetadata = exports.extractInitialTaggedContent = exports.parseRoleTags = exports.parseStructuredContent = exports.extractRandomVariables = exports.getSmartPreview = exports.countTokens = exports.renderTemplate = exports.LlamaCppServerClient = exports.LlamaCppClientAdapter = exports.fromEnvironment = exports.LLMService = void 0;
+exports.KNOWN_GGUF_MODELS = exports.detectGgufCapabilities = exports.createFallbackModelInfo = exports.parseTemplateWithMetadata = exports.extractInitialTaggedContent = exports.parseRoleTags = exports.parseStructuredContent = exports.extractRandomVariables = exports.getSmartPreview = exports.countTokens = exports.renderTemplate = exports.LlamaCppServerClient = exports.LlamaCppClientAdapter = exports.fromEnvironment = exports.LLMService = void 0;
 // --- LLM Service ---
 var LLMService_1 = require("./llm/LLMService");
 Object.defineProperty(exports, "LLMService", { enumerable: true, get: function () { return LLMService_1.LLMService; } });
@@ -44,3 +44,5 @@ Object.defineProperty(exports, "extractInitialTaggedContent", { enumerable: true
 Object.defineProperty(exports, "parseTemplateWithMetadata", { enumerable: true, get: function () { return parser_1.parseTemplateWithMetadata; } });
 var config_1 = require("./llm/config");
 Object.defineProperty(exports, "createFallbackModelInfo", { enumerable: true, get: function () { return config_1.createFallbackModelInfo; } });
+Object.defineProperty(exports, "detectGgufCapabilities", { enumerable: true, get: function () { return config_1.detectGgufCapabilities; } });
+Object.defineProperty(exports, "KNOWN_GGUF_MODELS", { enumerable: true, get: function () { return config_1.KNOWN_GGUF_MODELS; } });

package/dist/llm/LLMService.js

@@ -31,7 +31,7 @@ class LLMService {
         this.adapterRegistry = new AdapterRegistry_1.AdapterRegistry();
         this.requestValidator = new RequestValidator_1.RequestValidator();
         this.settingsManager = new SettingsManager_1.SettingsManager();
-        this.modelResolver = new ModelResolver_1.ModelResolver(this.presetManager);
+        this.modelResolver = new ModelResolver_1.ModelResolver(this.presetManager, this.adapterRegistry);
     }
     /**
      * Gets list of supported LLM providers
@@ -69,7 +69,7 @@ class LLMService {
         console.log(`LLMService.sendMessage called with presetId: ${request.presetId}, provider: ${request.providerId}, model: ${request.modelId}`);
         try {
             // Resolve model information from preset or direct IDs
-            const resolved = this.modelResolver.resolve(request);
+            const resolved = await this.modelResolver.resolve(request);
             if (resolved.error) {
                 return resolved.error;
             }
@@ -327,7 +327,7 @@ class LLMService {
         let modelContext = null;
         if (options.presetId || (options.providerId && options.modelId)) {
             // Resolve model information
-            const resolved = this.modelResolver.resolve({
+            const resolved = await this.modelResolver.resolve({
                 presetId: options.presetId,
                 providerId: options.providerId,
                 modelId: options.modelId,

package/dist/llm/clients/LlamaCppClientAdapter.d.ts

@@ -1,4 +1,4 @@
-import type { LLMResponse, LLMFailureResponse } from "../types";
+import type { LLMResponse, LLMFailureResponse, ModelInfo } from "../types";
 import type { ILLMClientAdapter, InternalLLMChatRequest } from "./types";
 import { LlamaCppServerClient } from "./LlamaCppServerClient";
 /**
@@ -49,12 +49,30 @@ export declare class LlamaCppClientAdapter implements ILLMClientAdapter {
     private baseURL;
     private checkHealth;
     private serverClient;
+    private cachedModelCapabilities;
+    private detectionAttempted;
     /**
      * Creates a new llama.cpp client adapter
      *
      * @param config Optional configuration for the adapter
      */
     constructor(config?: LlamaCppClientConfig);
+    /**
+     * Gets model capabilities by detecting the loaded GGUF model
+     *
+     * This method caches the result to avoid repeated HTTP calls.
+     * Cache is automatically cleared on connection errors in sendMessage().
+     *
+     * @returns Detected model capabilities or null if detection fails
+     */
+    getModelCapabilities(): Promise<Partial<ModelInfo> | null>;
+    /**
+     * Clears the cached model capabilities
+     *
+     * Called automatically on connection errors, or can be called manually
+     * if the server has been restarted with a different model.
+     */
+    clearModelCache(): void;
     /**
      * Sends a chat message to llama.cpp server
      *

package/dist/llm/clients/LlamaCppClientAdapter.js

@@ -10,6 +10,7 @@ const openai_1 = __importDefault(require("openai"));
 const types_1 = require("./types");
 const adapterErrorUtils_1 = require("./adapterErrorUtils");
 const LlamaCppServerClient_1 = require("./LlamaCppServerClient");
+const config_1 = require("../config");
 /**
  * Client adapter for llama.cpp server integration
  *
@@ -52,10 +53,68 @@ class LlamaCppClientAdapter {
      * @param config Optional configuration for the adapter
      */
     constructor(config) {
+        this.cachedModelCapabilities = null;
+        this.detectionAttempted = false;
         this.baseURL = config?.baseURL || 'http://localhost:8080';
         this.checkHealth = config?.checkHealth || false;
         this.serverClient = new LlamaCppServerClient_1.LlamaCppServerClient(this.baseURL);
     }
+    /**
+     * Gets model capabilities by detecting the loaded GGUF model
+     *
+     * This method caches the result to avoid repeated HTTP calls.
+     * Cache is automatically cleared on connection errors in sendMessage().
+     *
+     * @returns Detected model capabilities or null if detection fails
+     */
+    async getModelCapabilities() {
+        // Return cached result if available
+        if (this.cachedModelCapabilities !== null) {
+            return this.cachedModelCapabilities;
+        }
+        // Return null if we already tried and failed
+        if (this.detectionAttempted) {
+            return null;
+        }
+        // Attempt detection
+        try {
+            console.log(`Detecting model capabilities from llama.cpp server at ${this.baseURL}`);
+            const { data } = await this.serverClient.getModels();
+            if (!data || data.length === 0) {
+                console.warn('No models loaded in llama.cpp server');
+                this.detectionAttempted = true;
+                return null;
+            }
+            const ggufFilename = data[0].id;
+            const capabilities = (0, config_1.detectGgufCapabilities)(ggufFilename);
+            // Cache the result (even if null)
+            this.cachedModelCapabilities = capabilities;
+            this.detectionAttempted = true;
+            if (capabilities) {
+                console.log(`Cached model capabilities for: ${ggufFilename}`);
+            }
+            else {
+                console.log(`No known pattern matched for: ${ggufFilename}`);
+            }
+            return capabilities;
+        }
+        catch (error) {
+            console.warn('Failed to detect model capabilities:', error);
+            this.detectionAttempted = true;
+            return null;
+        }
+    }
+    /**
+     * Clears the cached model capabilities
+     *
+     * Called automatically on connection errors, or can be called manually
+     * if the server has been restarted with a different model.
+     */
+    clearModelCache() {
+        this.cachedModelCapabilities = null;
+        this.detectionAttempted = false;
+        console.log('Cleared model capabilities cache');
+    }
     /**
      * Sends a chat message to llama.cpp server
      *
@@ -132,6 +191,13 @@ class LlamaCppClientAdapter {
         }
         catch (error) {
             console.error("llama.cpp API error:", error);
+            // Clear cache on connection errors so we re-detect on next request
+            const errorMessage = error?.message || String(error);
+            if (errorMessage.includes("ECONNREFUSED") ||
+                errorMessage.includes("fetch failed") ||
+                errorMessage.includes("connect")) {
+                this.clearModelCache();
+            }
             return this.createErrorResponse(error, request);
         }
     }
@@ -219,19 +285,33 @@ class LlamaCppClientAdapter {
         if (!choice || !choice.message) {
             throw new Error("No valid choices in llama.cpp completion response");
         }
+        // Extract reasoning content if available
+        // llama.cpp returns reasoning in reasoning_content field when using --reasoning-format
+        let reasoning;
+        if (choice.message.reasoning_content) {
+            reasoning = choice.message.reasoning_content;
+        }
         return {
             id: completion.id,
             provider: request.providerId,
             model: completion.model || request.modelId,
             created: completion.created,
-            choices: completion.choices.map((c) => ({
-                message: {
-                    role: "assistant",
-                    content: c.message.content || "",
-                },
-                finish_reason: c.finish_reason,
-                index: c.index,
-            })),
+            choices: completion.choices.map((c) => {
+                const mappedChoice = {
+                    message: {
+                        role: "assistant",
+                        content: c.message.content || "",
+                    },
+                    finish_reason: c.finish_reason,
+                    index: c.index,
+                };
+                // Include reasoning if available and not excluded
+                const messageReasoning = c.message.reasoning_content;
+                if (messageReasoning && request.settings.reasoning && !request.settings.reasoning.exclude) {
+                    mappedChoice.reasoning = messageReasoning;
+                }
+                return mappedChoice;
+            }),
             usage: completion.usage
                 ? {
                     prompt_tokens: completion.usage.prompt_tokens,

package/dist/llm/clients/LlamaCppServerClient.d.ts

@@ -62,6 +62,23 @@ export interface LlamaCppSlot {
 export interface LlamaCppSlotsResponse {
     slots: LlamaCppSlot[];
 }
+/**
+ * Individual model information from /v1/models endpoint
+ */
+export interface LlamaCppModel {
+    id: string;
+    object?: string;
+    created?: number;
+    owned_by?: string;
+    [key: string]: any;
+}
+/**
+ * Response from the /v1/models endpoint
+ */
+export interface LlamaCppModelsResponse {
+    object: string;
+    data: LlamaCppModel[];
+}
 /**
  * Client for interacting with llama.cpp server's management and utility endpoints
  *
@@ -158,4 +175,23 @@ export declare class LlamaCppServerClient {
      * @throws Error if the request fails or endpoint is not enabled
      */
     getSlots(): Promise<LlamaCppSlotsResponse>;
+    /**
+     * Retrieves the list of models loaded in the llama.cpp server
+     *
+     * This uses the OpenAI-compatible /v1/models endpoint to get information
+     * about the currently loaded model(s). Typically llama.cpp serves one model
+     * at a time, but this returns an array for API compatibility.
+     *
+     * @returns Promise resolving to models response with array of loaded models
+     * @throws Error if the request fails
+     *
+     * @example
+     * ```typescript
+     * const client = new LlamaCppServerClient('http://localhost:8080');
+     * const { data } = await client.getModels();
+     * console.log('Loaded model:', data[0].id);
+     * // Output: "Qwen2.5-7B-Instruct-Q4_K_M.gguf"
+     * ```
+     */
+    getModels(): Promise<LlamaCppModelsResponse>;
 }

package/dist/llm/clients/LlamaCppServerClient.js

@@ -188,5 +188,30 @@ class LlamaCppServerClient {
         }
         return await response.json();
     }
+    /**
+     * Retrieves the list of models loaded in the llama.cpp server
+     *
+     * This uses the OpenAI-compatible /v1/models endpoint to get information
+     * about the currently loaded model(s). Typically llama.cpp serves one model
+     * at a time, but this returns an array for API compatibility.
+     *
+     * @returns Promise resolving to models response with array of loaded models
+     * @throws Error if the request fails
+     *
+     * @example
+     * ```typescript
+     * const client = new LlamaCppServerClient('http://localhost:8080');
+     * const { data } = await client.getModels();
+     * console.log('Loaded model:', data[0].id);
+     * // Output: "Qwen2.5-7B-Instruct-Q4_K_M.gguf"
+     * ```
+     */
+    async getModels() {
+        const response = await fetch(`${this.baseURL}/v1/models`);
+        if (!response.ok) {
+            throw new Error(`Get models failed: ${response.status} ${response.statusText}`);
+        }
+        return await response.json();
+    }
 }
 exports.LlamaCppServerClient = LlamaCppServerClient;

package/dist/llm/config.d.ts

@@ -31,6 +31,46 @@ export declare const MODEL_DEFAULT_SETTINGS: Record<string, Partial<LLMSettings>
  * Supported LLM providers
  */
 export declare const SUPPORTED_PROVIDERS: ProviderInfo[];
+/**
+ * Pattern definition for detecting GGUF model capabilities
+ */
+export interface GgufModelPattern {
+    /** Pattern to match in the GGUF filename (case-insensitive substring match) */
+    pattern: string;
+    /** Human-readable name for the model */
+    name: string;
+    /** Optional description */
+    description?: string;
+    /** Model capabilities (reasoning config, context window, etc.) */
+    capabilities: Partial<ModelInfo>;
+}
+/**
+ * Known GGUF model patterns for capability detection
+ *
+ * Order matters: more specific patterns should come before generic ones.
+ * First matching pattern wins.
+ *
+ * Example: "Qwen3-0.6B-0522" should be before "Qwen3-0.6B"
+ */
+export declare const KNOWN_GGUF_MODELS: GgufModelPattern[];
+/**
+ * Detects model capabilities from GGUF filename
+ *
+ * Performs case-insensitive substring matching against known model patterns.
+ * Returns the first matching pattern's capabilities (array order determines priority).
+ *
+ * @param ggufFilename - The GGUF model filename (e.g., "Qwen3-8B-Instruct-Q4_K_M.gguf")
+ * @returns Partial ModelInfo with detected capabilities, or null if no match
+ *
+ * @example
+ * ```typescript
+ * const caps = detectGgufCapabilities("Qwen3-8B-Instruct-Q4_K_M.gguf");
+ * if (caps?.reasoning?.supported) {
+ *   console.log("This model supports thinking!");
+ * }
+ * ```
+ */
+export declare function detectGgufCapabilities(ggufFilename: string): Partial<ModelInfo> | null;
 /**
  * Supported LLM models with their configurations
  * ModelInfo is similar to Cline model info
@@ -82,9 +122,10 @@ export declare function isModelSupported(modelId: string, providerId: string): b
  *
  * @param modelId - The model ID to create info for
  * @param providerId - The provider ID
- * @returns ModelInfo with default/placeholder values
+ * @param capabilities - Optional detected capabilities to merge (e.g., from GGUF detection)
+ * @returns ModelInfo with default/placeholder values, enhanced with detected capabilities
  */
-export declare function createFallbackModelInfo(modelId: string, providerId: string): ModelInfo;
+export declare function createFallbackModelInfo(modelId: string, providerId: string, capabilities?: Partial<ModelInfo>): ModelInfo;
 /**
  * Gets merged default settings for a specific model and provider
  *

package/dist/llm/config.js

@@ -2,7 +2,8 @@
 // AI Summary: Configuration for LLM module including default settings, supported providers, and models.
 // Defines operational parameters and available LLM options for the application.
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.SUPPORTED_MODELS = exports.SUPPORTED_PROVIDERS = exports.MODEL_DEFAULT_SETTINGS = exports.PROVIDER_DEFAULT_SETTINGS = exports.DEFAULT_LLM_SETTINGS = exports.ADAPTER_CONFIGS = exports.ADAPTER_CONSTRUCTORS = void 0;
+exports.SUPPORTED_MODELS = exports.KNOWN_GGUF_MODELS = exports.SUPPORTED_PROVIDERS = exports.MODEL_DEFAULT_SETTINGS = exports.PROVIDER_DEFAULT_SETTINGS = exports.DEFAULT_LLM_SETTINGS = exports.ADAPTER_CONFIGS = exports.ADAPTER_CONSTRUCTORS = void 0;
+exports.detectGgufCapabilities = detectGgufCapabilities;
 exports.getProviderById = getProviderById;
 exports.getModelById = getModelById;
 exports.getModelsByProvider = getModelsByProvider;
@@ -133,6 +134,150 @@ exports.SUPPORTED_PROVIDERS = [
         allowUnknownModels: true, // Test provider accepts any model
     },
 ];
+/**
+ * Known GGUF model patterns for capability detection
+ *
+ * Order matters: more specific patterns should come before generic ones.
+ * First matching pattern wins.
+ *
+ * Example: "Qwen3-0.6B-0522" should be before "Qwen3-0.6B"
+ */
+exports.KNOWN_GGUF_MODELS = [
+    // Qwen 3 Series - All support thinking/reasoning
+    {
+        pattern: "qwen3-30b",
+        name: "Qwen 3 30B",
+        description: "Qwen 3 30B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 16384,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 38912,
+            },
+        },
+    },
+    {
+        pattern: "qwen3-14b",
+        name: "Qwen 3 14B",
+        description: "Qwen 3 14B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 38912,
+            },
+        },
+    },
+    {
+        pattern: "qwen3-8b",
+        name: "Qwen 3 8B",
+        description: "Qwen 3 8B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 38912,
+            },
+        },
+    },
+    {
+        pattern: "qwen3-4b",
+        name: "Qwen 3 4B",
+        description: "Qwen 3 4B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 131072,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 38912,
+            },
+        },
+    },
+    {
+        pattern: "qwen3-1.7b",
+        name: "Qwen 3 1.7B",
+        description: "Qwen 3 1.7B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 30720,
+            },
+        },
+    },
+    {
+        pattern: "qwen3-0.6b",
+        name: "Qwen 3 0.6B",
+        description: "Qwen 3 0.6B model with thinking capabilities",
+        capabilities: {
+            maxTokens: 8192,
+            contextWindow: 32768,
+            supportsImages: false,
+            supportsPromptCache: false,
+            reasoning: {
+                supported: true,
+                enabledByDefault: false,
+                canDisable: true,
+                maxBudget: 30720,
+            },
+        },
+    },
+    // Add more model patterns here as needed
+    // DeepSeek, Llama, etc.
+];
+/**
+ * Detects model capabilities from GGUF filename
+ *
+ * Performs case-insensitive substring matching against known model patterns.
+ * Returns the first matching pattern's capabilities (array order determines priority).
+ *
+ * @param ggufFilename - The GGUF model filename (e.g., "Qwen3-8B-Instruct-Q4_K_M.gguf")
+ * @returns Partial ModelInfo with detected capabilities, or null if no match
+ *
+ * @example
+ * ```typescript
+ * const caps = detectGgufCapabilities("Qwen3-8B-Instruct-Q4_K_M.gguf");
+ * if (caps?.reasoning?.supported) {
+ *   console.log("This model supports thinking!");
+ * }
+ * ```
+ */
+function detectGgufCapabilities(ggufFilename) {
+    const lowerFilename = ggufFilename.toLowerCase();
+    // First match wins (array is pre-ordered from specific to generic)
+    for (const model of exports.KNOWN_GGUF_MODELS) {
+        if (lowerFilename.includes(model.pattern.toLowerCase())) {
+            console.log(`Detected GGUF model: ${model.name} (pattern: ${model.pattern})`);
+            return model.capabilities;
+        }
+    }
+    // No match found
+    return null;
+}
 /**
  * Supported LLM models with their configurations
  * ModelInfo is similar to Cline model info
@@ -521,10 +666,11 @@ function isModelSupported(modelId, providerId) {
  *
  * @param modelId - The model ID to create info for
  * @param providerId - The provider ID
- * @returns ModelInfo with default/placeholder values
+ * @param capabilities - Optional detected capabilities to merge (e.g., from GGUF detection)
+ * @returns ModelInfo with default/placeholder values, enhanced with detected capabilities
  */
-function createFallbackModelInfo(modelId, providerId) {
-    return {
+function createFallbackModelInfo(modelId, providerId, capabilities) {
+    const defaults = {
         id: modelId,
         name: modelId,
         providerId: providerId,
@@ -536,6 +682,23 @@ function createFallbackModelInfo(modelId, providerId) {
         supportsImages: false,
         supportsPromptCache: false,
     };
+    // Merge detected capabilities if provided
+    if (capabilities) {
+        return {
+            ...defaults,
+            ...capabilities,
+            // Always preserve these from defaults/params
+            id: modelId,
+            name: capabilities.name || modelId,
+            providerId: providerId,
+            // For local models, pricing is always 0
+            inputPrice: 0,
+            outputPrice: 0,
+            cacheWritesPrice: undefined,
+            cacheReadsPrice: undefined,
+        };
+    }
+    return defaults;
 }
 /**
  * Gets merged default settings for a specific model and provider

package/dist/llm/services/ModelResolver.d.ts

@@ -1,5 +1,6 @@
 import type { LLMFailureResponse, LLMSettings, ModelInfo } from "../types";
 import { PresetManager } from "./PresetManager";
+import { AdapterRegistry } from "./AdapterRegistry";
 /**
  * Options for model selection
  */
@@ -24,12 +25,13 @@ export interface ModelResolution {
  */
 export declare class ModelResolver {
     private presetManager;
-    constructor(presetManager: PresetManager);
+    private adapterRegistry;
+    constructor(presetManager: PresetManager, adapterRegistry: AdapterRegistry);
     /**
      * Resolves model information from either a preset ID or provider/model IDs
      *
      * @param options Options containing either presetId or providerId/modelId
      * @returns Resolved model info and settings or error response
      */
-    resolve(options: ModelSelectionOptions): ModelResolution;
+    resolve(options: ModelSelectionOptions): Promise<ModelResolution>;
 }

package/dist/llm/services/ModelResolver.js

@@ -6,8 +6,9 @@ const config_1 = require("../config");
  * Resolves model information from presets or direct provider/model IDs
  */
 class ModelResolver {
-    constructor(presetManager) {
+    constructor(presetManager, adapterRegistry) {
         this.presetManager = presetManager;
+        this.adapterRegistry = adapterRegistry;
     }
     /**
      * Resolves model information from either a preset ID or provider/model IDs
@@ -15,7 +16,7 @@ class ModelResolver {
      * @param options Options containing either presetId or providerId/modelId
      * @returns Resolved model info and settings or error response
      */
-    resolve(options) {
+    async resolve(options) {
         // If presetId is provided, use it
         if (options.presetId) {
             const preset = this.presetManager.resolvePreset(options.presetId);
@@ -94,15 +95,31 @@ class ModelResolver {
         if (!modelInfo) {
             // Check if provider allows unknown models
             const provider = (0, config_1.getProviderById)(options.providerId);
+            // For llamacpp, try to detect capabilities from the adapter's cache
+            let detectedCapabilities;
+            if (options.providerId === 'llamacpp') {
+                try {
+                    const adapter = this.adapterRegistry.getAdapter('llamacpp');
+                    // Check if adapter has the getModelCapabilities method
+                    if (adapter && typeof adapter.getModelCapabilities === 'function') {
+                        const capabilities = await adapter.getModelCapabilities();
+                        detectedCapabilities = capabilities || undefined;
+                    }
+                }
+                catch (error) {
+                    console.warn('Failed to detect GGUF model capabilities:', error);
+                    // Continue with fallback
+                }
+            }
             if (provider?.allowUnknownModels) {
-                // Flexible provider (e.g., llamacpp) - silent fallback
-                modelInfo = (0, config_1.createFallbackModelInfo)(options.modelId, options.providerId);
+                // Flexible provider (e.g., llamacpp) - silent fallback with detected capabilities
+                modelInfo = (0, config_1.createFallbackModelInfo)(options.modelId, options.providerId, detectedCapabilities);
             }
             else {
                 // Strict provider - warn but allow
                 console.warn(`⚠️  Unknown model "${options.modelId}" for provider "${options.providerId}". ` +
                     `Using default settings. This may fail at the provider API if the model doesn't exist.`);
-                modelInfo = (0, config_1.createFallbackModelInfo)(options.modelId, options.providerId);
+                modelInfo = (0, config_1.createFallbackModelInfo)(options.modelId, options.providerId, detectedCapabilities);
             }
         }
         return {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "genai-lite",
-  "version": "0.4.2",
+  "version": "0.4.3",
   "description": "A lightweight, portable toolkit for interacting with various Generative AI APIs.",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",