@fugood/bricks-project 2.24.0-beta.2 → 2.24.0-beta.21

package/tools/mcp-tools/huggingface.ts

@@ -0,0 +1,762 @@
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { z } from 'zod'
+import { JSON5 } from 'bun'
+import * as TOON from '@toon-format/toon'
+import { gguf } from '@huggingface/gguf'
+
+// Hugging Face API configuration
+const HF_API_URL = 'https://huggingface.co/api'
+const { HF_TOKEN } = process.env
+
+// Helper function to convert BigInt to number for JSON serialization
+const convertBigIntToNumber = (value: unknown): number | unknown => {
+  if (typeof value === 'bigint') {
+    return Number(value)
+  }
+  return value
+}
+
+// Extract GGUF metadata from a GGUF file URL with 10s timeout
+const extractGGUFMetadata = async (url: string) => {
+  const timeoutPromise = new Promise<never>((_, reject) =>
+    setTimeout(() => reject(new Error('GGUF metadata extraction timeout')), 10000),
+  )
+
+  try {
+    const { metadata } = (await Promise.race([gguf(url), timeoutPromise])) as { metadata: any }
+    const architecture = metadata['general.architecture']
+
+    const ggufSimplifiedMetadata: Record<string, unknown> = {
+      name: metadata['general.name'],
+      size_label: metadata['general.size_label'],
+      basename: metadata['general.basename'],
+      architecture: metadata['general.architecture'],
+      file_type: convertBigIntToNumber(metadata['general.file_type']),
+      quantization_version: convertBigIntToNumber(metadata['general.quantization_version']),
+    }
+
+    if (!architecture) return ggufSimplifiedMetadata
+
+    // Helper to add converted value if defined
+    const addIfDefined = (target: Record<string, unknown>, key: string, value: unknown) => {
+      if (value !== undefined) target[key] = convertBigIntToNumber(value)
+    }
+
+    // Architecture-specific transformer parameters
+    addIfDefined(ggufSimplifiedMetadata, 'n_ctx_train', metadata[`${architecture}.context_length`])
+    addIfDefined(ggufSimplifiedMetadata, 'n_layer', metadata[`${architecture}.block_count`])
+    addIfDefined(ggufSimplifiedMetadata, 'n_embd', metadata[`${architecture}.embedding_length`])
+    addIfDefined(ggufSimplifiedMetadata, 'n_head', metadata[`${architecture}.attention.head_count`])
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'n_head_kv',
+      metadata[`${architecture}.attention.head_count_kv`],
+    )
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'n_embd_head_k',
+      metadata[`${architecture}.attention.key_length`],
+    )
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'n_embd_head_v',
+      metadata[`${architecture}.attention.value_length`],
+    )
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'n_swa',
+      metadata[`${architecture}.attention.sliding_window`],
+    )
+
+    // SSM (Mamba) parameters for recurrent/hybrid models
+    addIfDefined(ggufSimplifiedMetadata, 'ssm_d_conv', metadata[`${architecture}.ssm.conv_kernel`])
+    addIfDefined(ggufSimplifiedMetadata, 'ssm_d_state', metadata[`${architecture}.ssm.state_size`])
+    addIfDefined(ggufSimplifiedMetadata, 'ssm_d_inner', metadata[`${architecture}.ssm.inner_size`])
+    addIfDefined(ggufSimplifiedMetadata, 'ssm_n_group', metadata[`${architecture}.ssm.group_count`])
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'ssm_dt_rank',
+      metadata[`${architecture}.ssm.time_step_rank`],
+    )
+
+    // RWKV parameters
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'rwkv_head_size',
+      metadata[`${architecture}.rwkv.head_size`],
+    )
+    addIfDefined(
+      ggufSimplifiedMetadata,
+      'rwkv_token_shift_count',
+      metadata[`${architecture}.rwkv.token_shift_count`],
+    )
+
+    return ggufSimplifiedMetadata
+  } catch (error) {
+    console.error('Failed to extract GGUF metadata:', error)
+    return null
+  }
+}
+
+type HFSibling = {
+  rfilename: string
+  size?: number
+  lfs?: { sha256?: string }
+  blobId?: string
+}
+
+type HFModel = {
+  id: string
+  author?: string
+  downloads?: number
+  likes?: number
+  tags?: string[]
+  pipeline_tag?: string
+  siblings?: HFSibling[]
+  config?: { model_type?: string }
+  cardData?: { model_type?: string }
+}
+
+const supportedLlmTasks = [
+  'text-generation',
+  'image-text-to-text',
+  'text2text-generation',
+  'conversational',
+]
+
+type GeneratorType =
+  | 'GeneratorLLM'
+  | 'GeneratorVectorStore'
+  | 'GeneratorReranker'
+  | 'GeneratorGGMLTTS'
+  | 'GeneratorGGMLTTSVocoder'
+  | 'GeneratorOnnxLLM'
+  | 'GeneratorOnnxSTT'
+  | 'GeneratorTTS'
+  | 'GeneratorMlxLLM'
+
+type ModelKind = 'gguf' | 'onnx' | 'mlx'
+
+interface GeneratorConfig {
+  modelKind: ModelKind
+  filter: string
+  taskFilter?: string[]
+  filePattern?: RegExp
+  hasValidStructure?: (siblings: HFSibling[]) => boolean
+}
+
+// Helper to check valid ONNX structure
+const hasValidOnnxStructure = (siblings: HFSibling[]): boolean => {
+  const hasConfigJson = siblings.some((file) => file.rfilename === 'config.json')
+  const hasOnnxModel = siblings.some(
+    (file) => file.rfilename.includes('onnx/') && file.rfilename.endsWith('.onnx'),
+  )
+  return hasConfigJson && hasOnnxModel
+}
+
+// Detect quantization types from ONNX files
+const detectOnnxQuantizationTypes = (siblings: HFSibling[]): string[] => {
+  const onnxFiles = siblings.filter((file) => file.rfilename.endsWith('.onnx'))
+  const quantTypes = new Set<string>()
+
+  onnxFiles.forEach((file) => {
+    const filename = file.rfilename
+    if (!filename.endsWith('.onnx')) return
+    const postfix = /_(q8|q4|q4f16|fp16|int8|int4|uint8|bnb4|quantized)\.onnx$/.exec(filename)?.[1]
+    if (!postfix) {
+      quantTypes.add('auto')
+      quantTypes.add('none')
+    } else {
+      quantTypes.add(postfix === 'quantized' ? 'q8' : postfix)
+    }
+  })
+
+  return Array.from(quantTypes)
+}
+
+// Find speaker embedding files for TTS models
+const findSpeakerEmbedFiles = (siblings: HFSibling[]): HFSibling[] =>
+  siblings.filter((file) => file.rfilename.startsWith('voices/') && file.rfilename.endsWith('.bin'))
+
+const generatorConfigs: Record<GeneratorType, GeneratorConfig> = {
+  GeneratorLLM: {
+    modelKind: 'gguf',
+    filter: 'gguf',
+    taskFilter: supportedLlmTasks,
+    filePattern: /\.gguf$/,
+  },
+  GeneratorVectorStore: {
+    modelKind: 'gguf',
+    filter: 'gguf',
+    filePattern: /\.gguf$/,
+  },
+  GeneratorReranker: {
+    modelKind: 'gguf',
+    filter: 'gguf,reranker',
+    filePattern: /\.gguf$/,
+  },
+  GeneratorGGMLTTS: {
+    modelKind: 'gguf',
+    filter: 'gguf,text-to-speech',
+    filePattern: /\.gguf$/,
+  },
+  GeneratorGGMLTTSVocoder: {
+    modelKind: 'gguf',
+    filter: 'gguf,feature-extraction',
+    filePattern: /\.gguf$/,
+  },
+  GeneratorOnnxLLM: {
+    modelKind: 'onnx',
+    filter: 'onnx',
+    taskFilter: supportedLlmTasks,
+    hasValidStructure: hasValidOnnxStructure,
+  },
+  GeneratorOnnxSTT: {
+    modelKind: 'onnx',
+    filter: 'onnx,automatic-speech-recognition',
+    hasValidStructure: hasValidOnnxStructure,
+  },
+  GeneratorTTS: {
+    modelKind: 'onnx',
+    filter: 'onnx,text-to-speech',
+    hasValidStructure: hasValidOnnxStructure,
+  },
+  GeneratorMlxLLM: {
+    modelKind: 'mlx',
+    filter: 'mlx',
+    taskFilter: supportedLlmTasks,
+  },
+}
+
+const searchHFModels = async (filter: string, search?: string, limit = 50): Promise<HFModel[]> => {
+  const params = new URLSearchParams({
+    limit: String(limit),
+    full: 'true',
+    config: 'true',
+    sort: 'likes',
+    direction: '-1',
+  })
+  if (filter) params.set('filter', filter)
+  if (search) params.set('search', search)
+
+  const headers: Record<string, string> = {}
+  if (HF_TOKEN) {
+    headers['Authorization'] = `Bearer ${HF_TOKEN}`
+  }
+
+  const response = await fetch(`${HF_API_URL}/models?${params.toString()}`, { headers })
+  if (!response.ok) {
+    throw new Error(`Hugging Face API error: ${response.status} ${response.statusText}`)
+  }
+  return response.json()
+}
+
+const fetchHFModelDetails = async (modelId: string): Promise<HFModel> => {
+  const params = new URLSearchParams({ blobs: 'true' })
+
+  const headers: Record<string, string> = {}
+  if (HF_TOKEN) {
+    headers['Authorization'] = `Bearer ${HF_TOKEN}`
+  }
+
+  const response = await fetch(`${HF_API_URL}/models/${modelId}?${params.toString()}`, { headers })
+  if (!response.ok) {
+    throw new Error(`Hugging Face API error: ${response.status} ${response.statusText}`)
+  }
+  return response.json()
+}
+
+// Example: Mixtral-8x22B-v0.1.IQ3_XS-00001-of-00005.gguf
+const ggufSplitPattern = /-(\d{5})-of-(\d{5})\.gguf$/
+
+export function register(server: McpServer) {
+  server.tool(
+    'huggingface_search',
+    {
+      generatorType: z
+        .enum([
+          'GeneratorLLM',
+          'GeneratorVectorStore',
+          'GeneratorReranker',
+          'GeneratorGGMLTTS',
+          'GeneratorGGMLTTSVocoder',
+          'GeneratorOnnxLLM',
+          'GeneratorOnnxSTT',
+          'GeneratorTTS',
+          'GeneratorMlxLLM',
+        ])
+        .describe('Generator type to search models for')
+        .default('GeneratorLLM'),
+      query: z.string().describe('Search keywords for models on Hugging Face').optional(),
+      limit: z.number().min(1).max(100).optional().default(20),
+      includeFiles: z
+        .boolean()
+        .optional()
+        .default(false)
+        .describe('Include list of model files (requires additional API calls)'),
+    },
+    async ({ generatorType, query, limit, includeFiles }) => {
+      try {
+        const config = generatorConfigs[generatorType]
+        const models = await searchHFModels(config.filter, query, limit)
+
+        // Filter models based on generator configuration
+        const filteredModels = models.filter((model) => {
+          const modelTags = model.tags || []
+
+          // Check task filter if configured
+          if (config.taskFilter && !config.taskFilter.some((t) => modelTags.includes(t))) {
+            return false
+          }
+
+          // Check structure validation for ONNX models
+          if (config.hasValidStructure && model.siblings) {
+            if (!config.hasValidStructure(model.siblings)) {
+              return false
+            }
+          }
+
+          return true
+        })
+
+        // Build result models
+        let results: Array<{
+          id: string
+          author?: string
+          downloads?: number
+          likes?: number
+          pipeline_tag?: string
+          model_type?: string
+          model_kind: ModelKind
+          files?: Array<{ filename: string; size?: number }>
+          quantization_types?: string[]
+          speaker_embed_files?: Array<{ filename: string; size?: number }>
+        }> = filteredModels.map((model) => ({
+          id: model.id,
+          author: model.author,
+          downloads: model.downloads,
+          likes: model.likes,
+          pipeline_tag: model.pipeline_tag,
+          model_type: model.config?.model_type || model.cardData?.model_type,
+          model_kind: config.modelKind,
+        }))
+
+        if (includeFiles) {
+          results = await Promise.all(
+            results.map(async (model) => {
+              try {
+                const details = await fetchHFModelDetails(model.id)
+                const siblings = details.siblings || []
+
+                if (config.modelKind === 'gguf') {
+                  const ggufFiles = siblings
+                    .filter((file) => config.filePattern?.test(file.rfilename))
+                    .map((file) => ({
+                      filename: file.rfilename,
+                      size: file.size,
+                    }))
+                  return { ...model, files: ggufFiles }
+                } else {
+                  // ONNX models
+                  const quantTypes = detectOnnxQuantizationTypes(siblings)
+                  const speakerFiles =
+                    generatorType === 'GeneratorTTS' ? findSpeakerEmbedFiles(siblings) : []
+
+                  return {
+                    ...model,
+                    quantization_types: quantTypes,
+                    ...(speakerFiles.length > 0 && {
+                      speaker_embed_files: speakerFiles.map((f) => ({
+                        filename: f.rfilename,
+                        size: f.size,
+                      })),
+                    }),
+                  }
+                }
+              } catch {
+                return model
+              }
+            }),
+          )
+        }
+
+        return {
+          content: [
+            {
+              type: 'text',
+              text: TOON.encode({
+                count: results.length,
+                generatorType,
+                modelKind: config.modelKind,
+                models: results,
+                hf_token_configured: !!HF_TOKEN,
+              }),
+            },
+          ],
+        }
+      } catch (err: any) {
+        return {
+          content: [{ type: 'text', text: `Failed to search models: ${err.message}` }],
+        }
+      }
+    },
+  )
+
+  server.tool(
+    'huggingface_select',
+    {
+      generatorType: z
+        .enum([
+          'GeneratorLLM',
+          'GeneratorVectorStore',
+          'GeneratorReranker',
+          'GeneratorGGMLTTS',
+          'GeneratorGGMLTTSVocoder',
+          'GeneratorOnnxLLM',
+          'GeneratorOnnxSTT',
+          'GeneratorTTS',
+          'GeneratorMlxLLM',
+        ])
+        .describe('Generator type for model selection')
+        .default('GeneratorLLM'),
+      // eslint-disable-next-line camelcase
+      model_id: z
+        .string()
+        .describe('Hugging Face model ID (e.g., "unsloth/Llama-3.2-1B-Instruct-GGUF")'),
+      filename: z
+        .string()
+        .describe('Model filename to select (required for GGUF models)')
+        .optional(),
+      quantize_type: z
+        .string()
+        .describe('Quantization type for ONNX models (e.g., "q8", "fp16", "auto")')
+        .optional()
+        .default('auto'),
+      speaker_embed_file: z.string().describe('Speaker embedding file for TTS models').optional(),
+    },
+    // eslint-disable-next-line camelcase
+    async ({
+      generatorType,
+      model_id: modelId,
+      filename,
+      quantize_type: quantizeType,
+      speaker_embed_file: speakerEmbedFile,
+    }) => {
+      try {
+        const config = generatorConfigs[generatorType]
+        const details = await fetchHFModelDetails(modelId)
+        const siblings = details.siblings || []
+
+        // Handle ONNX models
+        // ONNX generators expect: model (HF model ID), modelType, quantizeType
+        if (config.modelKind === 'onnx') {
+          const quantTypes = detectOnnxQuantizationTypes(siblings)
+          const speakerFiles =
+            generatorType === 'GeneratorTTS' ? findSpeakerEmbedFiles(siblings) : []
+          const selectedSpeakerFile = speakerEmbedFile
+            ? siblings.find((f) => f.rfilename === speakerEmbedFile)
+            : undefined
+
+          const selectedQuantType = quantTypes.includes(quantizeType || 'auto')
+            ? quantizeType
+            : 'auto'
+          const modelType = details.config?.model_type || details.cardData?.model_type
+
+          // Result format matches ONNX generator property names (camelCase)
+          const result = {
+            // Primary model fields for generator
+            model: modelId,
+            modelType,
+            quantizeType: selectedQuantType,
+            // Speaker embedding for TTS generators
+            ...(selectedSpeakerFile && {
+              speakerEmbedUrl: `https://huggingface.co/${modelId}/resolve/main/${selectedSpeakerFile.rfilename}?download=true`,
+              speakerEmbedHash: selectedSpeakerFile.lfs?.sha256,
+              speakerEmbedHashType: 'sha256',
+            }),
+            // Additional info
+            availableQuantizeTypes: quantTypes,
+            ...(speakerFiles.length > 0 && {
+              availableSpeakerEmbedFiles: speakerFiles.map((f) => f.rfilename),
+            }),
+            _hfRepoInfo: {
+              repo: modelId,
+              model: {
+                id: details.id,
+                downloads: details.downloads,
+                likes: details.likes,
+                author: details.author,
+              },
+              generatorType,
+              modelType,
+              quantizeType: selectedQuantType,
+              ...(selectedSpeakerFile && {
+                speakerEmbedFile: selectedSpeakerFile.rfilename,
+              }),
+            },
+          }
+
+          // Return JSON for huggingface_select to allow direct parsing by consumers
+          return {
+            content: [
+              {
+                type: 'text',
+                text: `${JSON5.stringify(result, null, 2)} // Please use _hfRepoInfo undocumented field with @ts-ignore`,
+              },
+            ],
+          }
+        }
+
+        // Handle MLX models (repo-level, like ONNX)
+        if (config.modelKind === 'mlx') {
+          // Fetch config.json for model architecture metadata
+          const headers: Record<string, string> = {}
+          if (HF_TOKEN) headers['Authorization'] = `Bearer ${HF_TOKEN}`
+          let mlxModelConfig: Record<string, any> | null = null
+          try {
+            const configRes = await fetch(
+              `https://huggingface.co/${modelId}/resolve/main/config.json`,
+              { headers },
+            )
+            if (configRes.ok) mlxModelConfig = await configRes.json()
+          } catch {
+            // Non-critical
+          }
+
+          const modelType =
+            mlxModelConfig?.model_type || details.config?.model_type || details.cardData?.model_type
+
+          // Build MLX model metadata from config.json (for hardware guardrails)
+          const cfg = mlxModelConfig || ({} as Record<string, any>)
+          const textCfg = cfg.text_config || cfg
+          const numHeads = textCfg.num_attention_heads || textCfg.n_heads || 0
+          const hiddenSize = textCfg.hidden_size || textCfg.dim || 0
+          const kvLoraRank = textCfg.kv_lora_rank || 0
+          const quant = cfg.quantization || cfg.quantization_config || null
+
+          // Sum safetensors/npz file sizes for model weight bytes
+          const modelBytes = siblings
+            .filter((f) => /\.(safetensors|npz)$/.test(f.rfilename))
+            .reduce((sum, f) => sum + (f.size ?? 0), 0)
+
+          // Build _mlxDownloadFiles list (safetensors, json, jinja, tokenizer.model)
+          const mlxDownloadFiles = siblings
+            .filter(
+              (f) =>
+                f.rfilename.endsWith('.safetensors') ||
+                f.rfilename.endsWith('.json') ||
+                f.rfilename.endsWith('.jinja') ||
+                f.rfilename === 'tokenizer.model',
+            )
+            .map((f) => ({
+              url: `https://huggingface.co/${modelId}/resolve/main/${f.rfilename}?download=true`,
+              filename: `${modelId.replace('/', '-')}/${f.rfilename}`,
+              hash_type: f.lfs ? 'sha256' : f.blobId ? 'sha1' : undefined,
+              sha256: f.lfs?.sha256,
+              sha1: f.lfs ? undefined : f.blobId,
+            }))
+
+          const result = {
+            modelId,
+            modelType,
+            _mlxDownloadFiles: mlxDownloadFiles,
+            _hfRepoInfo: {
+              repo: modelId,
+              model: {
+                id: details.id,
+                downloads: details.downloads,
+                likes: details.likes,
+                author: details.author,
+              },
+              generatorType,
+              modelType,
+              ...(modelBytes > 0 && { size: modelBytes }),
+              ...(mlxModelConfig && {
+                mlxModelMetadata: {
+                  hiddenSize,
+                  numLayers: textCfg.num_hidden_layers || textCfg.n_layers || 0,
+                  numHeads,
+                  numKvHeads: textCfg.num_key_value_heads ?? numHeads,
+                  headDim:
+                    textCfg.head_dim ||
+                    textCfg.v_head_dim ||
+                    (numHeads > 0 && hiddenSize > 0 ? hiddenSize / numHeads : 0),
+                  maxCtx: textCfg.max_position_embeddings || 0,
+                  kvLoraRank,
+                  qkRopeHeadDim: textCfg.qk_rope_head_dim || 0,
+                  isMLA: kvLoraRank > 0,
+                  quantBits: quant?.bits || null,
+                  quantGroupSize: quant?.group_size || null,
+                },
+              }),
+            },
+          }
+
+          return {
+            content: [
+              {
+                type: 'text',
+                text: `${JSON5.stringify(result, null, 2)} // IMPORTANT: Use _mlxDownloadFiles / _hfRepoInfo undocumented fields with @ts-ignore`,
+              },
+            ],
+          }
+        }
+
+        // Handle GGUF models
+        if (!filename) {
+          // List available GGUF files
+          const ggufFiles = siblings
+            .filter((file) => config.filePattern?.test(file.rfilename))
+            .map((file) => ({
+              filename: file.rfilename,
+              size: file.size,
+            }))
+
+          // Return JSON for huggingface_select to allow direct parsing by consumers
+          return {
+            content: [
+              {
+                type: 'text',
+                text: JSON.stringify(
+                  {
+                    error: 'filename is required for GGUF models',
+                    available_files: ggufFiles,
+                  },
+                  null,
+                  2,
+                ),
+              },
+            ],
+          }
+        }
+
+        // Find the selected file
+        const selectedFile = siblings.find((f) => f.rfilename === filename)
+        if (!selectedFile) {
+          return {
+            content: [{ type: 'text', text: `File "${filename}" not found in model ${modelId}` }],
+          }
+        }
+
+        // Check if it's a split file
+        const matched = filename.match(ggufSplitPattern)
+        const isSplit = !!matched
+
+        // Find mmproj file if available (only for LLM generators)
+        const mmprojFile =
+          generatorType === 'GeneratorLLM'
+            ? siblings.find((f) => /^mmproj-/i.test(f.rfilename))
+            : undefined
+
+        // Extract GGUF metadata (for split files, metadata is in the first split)
+        const metadataFilename = isSplit
+          ? filename.replace(ggufSplitPattern, '-00001-of-$2.gguf')
+          : filename
+        const ggufUrl = `https://huggingface.co/${modelId}/resolve/main/${metadataFilename}`
+        const ggufSimplifiedMetadata = await extractGGUFMetadata(ggufUrl)
+
+        if (isSplit) {
+          const [, , splitTotal] = matched!
+          const splitFiles = Array.from({ length: Number(splitTotal) }, (_, i) => {
+            const split = String(i + 1).padStart(5, '0')
+            const splitRFilename = filename.replace(
+              ggufSplitPattern,
+              `-${split}-of-${splitTotal}.gguf`,
+            )
+            const sibling = siblings.find((sb) => sb.rfilename === splitRFilename)
+            return {
+              rfilename: splitRFilename,
+              size: sibling?.size,
+              lfs: sibling?.lfs,
+            }
+          })
+
+          const first = splitFiles[0]
+          const rest = splitFiles.slice(1)
+
+          const result = {
+            url: `https://huggingface.co/${modelId}/resolve/main/${first.rfilename}?download=true`,
+            hash: first.lfs?.sha256,
+            hash_type: 'sha256',
+            _ggufSplitFiles: rest.map((split) => ({
+              url: `https://huggingface.co/${modelId}/resolve/main/${split.rfilename}?download=true`,
+              hash: split.lfs?.sha256,
+              hash_type: 'sha256',
+            })),
+            ...(mmprojFile && {
+              mmproj_url: `https://huggingface.co/${modelId}/resolve/main/${mmprojFile.rfilename}?download=true`,
+              mmproj_hash: mmprojFile.lfs?.sha256,
+              mmproj_hash_type: 'sha256',
+            }),
+            _hfRepoInfo: {
+              repo: modelId,
+              model: {
+                id: details.id,
+                downloads: details.downloads,
+                likes: details.likes,
+                author: details.author,
+              },
+              generatorType,
+              isSplit: true,
+              files: splitFiles.map((f) => f.rfilename),
+              sizes: splitFiles.map((f) => f.size),
+              size: splitFiles.reduce((acc, f) => acc + (f.size ?? 0), 0),
+              ...(mmprojFile && {
+                mmprojFile: mmprojFile.rfilename,
+                mmprojSize: mmprojFile.size,
+              }),
+              ...(ggufSimplifiedMetadata && { ggufSimplifiedMetadata }),
+            },
+          }
+
+          // Return JSON for huggingface_select to allow direct parsing by consumers
+          return {
+            content: [{ type: 'text', text: JSON5.stringify(result, null, 2) }],
+          }
+        } else {
+          const result = {
+            url: `https://huggingface.co/${modelId}/resolve/main/${filename}?download=true`,
+            hash: selectedFile.lfs?.sha256,
+            hash_type: 'sha256',
+            ...(mmprojFile && {
+              mmproj_url: `https://huggingface.co/${modelId}/resolve/main/${mmprojFile.rfilename}?download=true`,
+              mmproj_hash: mmprojFile.lfs?.sha256,
+              mmproj_hash_type: 'sha256',
+            }),
+            _hfRepoInfo: {
+              repo: modelId,
+              model: {
+                id: details.id,
+                downloads: details.downloads,
+                likes: details.likes,
+                author: details.author,
+              },
+              generatorType,
+              file: filename,
+              size: selectedFile.size,
+              ...(mmprojFile && {
+                mmprojFile: mmprojFile.rfilename,
+                mmprojSize: mmprojFile.size,
+              }),
+              ...(ggufSimplifiedMetadata && { ggufSimplifiedMetadata }),
+            },
+          }
+
+          // Return JSON for huggingface_select to allow direct parsing by consumers
+          return {
+            content: [
+              {
+                type: 'text',
+                text: `${JSON5.stringify(result, null, 2)} // IMPORTANT: Use _hfRepoInfo undocumented field with @ts-ignore`,
+              },
+            ],
+          }
+        }
+      } catch (err: any) {
+        return {
+          content: [{ type: 'text', text: `Failed to select model: ${err.message}` }],
+        }
+      }
+    },
+  )
+}