claude-brain 0.30.2 → 0.30.3

package/src/cli/commands/models.ts

@@ -1,979 +1,979 @@
-/**
- * Models Command — SLM Upgrade Phase 5
- * Manage SLM models: list, download, enable, disable, benchmark, stats.
- *
- * Usage:
- *   claude-brain models list
- *   claude-brain models download [--task <task>|all]
- *   claude-brain models enable <task>
- *   claude-brain models disable <task>
- *   claude-brain models benchmark <task>
- *   claude-brain models stats
- */
-
-import { readFileSync, existsSync, mkdirSync, writeFileSync, statSync, copyFileSync } from 'node:fs'
-import { join } from 'node:path'
-import { homedir } from 'node:os'
-import { parseArgs } from 'citty'
-import { renderLogo, theme, heading, dimText, successText, warningText, errorText, box, summaryPanel } from '@/cli/ui/index.js'
-import { progressBar } from '@/cli/ui/components.js'
-import { getHomePaths, getClaudeBrainHome } from '@/config/home'
-import { getTrainingStats, type TrainingTask } from '@/training/data-store'
-import type { ModelManifest, ModelManifestEntry, ModelTask } from '@/intelligence/model-manager'
-import { shouldRetrain, retrainTask, retrainAll, type RetrainConfig } from '@/training/retrain-pipeline'
-import { downloadFromHuggingFace, type HfManifest } from '@/intelligence/hf-downloader'
-import hfManifestData from '@/intelligence/hf-manifest.json'
-
-const ALL_TASKS: ModelTask[] = ['intent', 'entity', 'query', 'knowledge', 'compress', 'pattern']
-
-/** Default mode when disabling a task */
-const DISABLE_MODE: Record<ModelTask, string> = {
-  intent: 'regex',
-  entity: 'regex',
-  query: 'regex',
-  knowledge: 'regex',
-  compress: 'api',
-  pattern: 'regex',
-}
-
-export async function runModels() {
-  const args = parseArgs(process.argv.slice(3), {
-    subcommand: { type: 'positional', required: false, description: 'Subcommand: list, status, download, enable, disable, benchmark, stats, retrain' },
-    taskArg: { type: 'positional', required: false, description: 'Task name or "all" (for enable/disable/benchmark/retrain)' },
-    task: { type: 'string', description: 'Target task (for download --task)' },
-    source: { type: 'string', description: 'Source: local (default) or hf (Hugging Face Hub)' },
-    force: { type: 'boolean', description: 'Force retrain even if checks say not needed' },
-  })
-
-  const subcommand = args.subcommand || ''
-  const taskArg = (args.task || args.taskArg || '') as string
-
-  switch (subcommand) {
-    case 'list':
-      return listModels()
-    case 'status':
-      return showStatus()
-    case 'download':
-      return downloadModels(taskArg || 'all', (args.source as string) || 'local')
-    case 'enable':
-      return enableTask(taskArg)
-    case 'disable':
-      return disableTask(taskArg)
-    case 'benchmark':
-      return benchmarkTask(taskArg)
-    case 'stats':
-      return showStats()
-    case 'retrain':
-      return retrainModels(taskArg || 'all', !!args.force)
-    default:
-      return printModelsHelp()
-  }
-}
-
-// ─── list ─────────────────────────────────────────────────────────
-
-function listModels() {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('SLM Models'))
-  console.log()
-
-  const paths = getHomePaths()
-  const manifestPath = join(paths.models, 'manifest.json')
-
-  const items: Array<{ label: string; value: string; status?: 'success' | 'warning' | 'error' | 'info' }> = []
-
-  if (!existsSync(manifestPath)) {
-    items.push({ label: 'Manifest', value: 'Not found — run "models download" first', status: 'warning' })
-    console.log(summaryPanel('Models', items))
-    console.log()
-
-    // Still show task status even without manifest
-    for (const task of ALL_TASKS) {
-      console.log(`  ${theme.primary(task.padEnd(12))} ${warningText('not installed')}`)
-    }
-    console.log()
-    return
-  }
-
-  let manifest: ModelManifest
-  try {
-    manifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
-  } catch {
-    console.log(errorText('Failed to parse manifest.json'))
-    return
-  }
-
-  let totalSize = 0
-  const rows: string[] = []
-
-  for (const task of ALL_TASKS) {
-    const entry = manifest.models?.[task]
-    if (!entry) {
-      rows.push(`  ${theme.primary(task.padEnd(12))} ${warningText('not installed')}`)
-      continue
-    }
-
-    const modelPath = join(paths.models, entry.file)
-    const installed = existsSync(modelPath)
-
-    let sizeStr = '—'
-    if (installed) {
-      try {
-        const size = statSync(modelPath).size
-        totalSize += size
-        sizeStr = formatBytes(size)
-      } catch {
-        sizeStr = '?'
-      }
-    }
-
-    const status = installed ? successText('installed') : warningText('missing')
-    const version = entry.version ? dimText(`v${entry.version}`) : ''
-    const accuracy = entry.accuracy != null ? dimText(`${(entry.accuracy * 100).toFixed(1)}%`) : ''
-
-    rows.push(
-      `  ${theme.primary(task.padEnd(12))} ${status.padEnd(24)} ${version.padEnd(14)} ${accuracy.padEnd(10)} ${dimText(sizeStr)}`
-    )
-  }
-
-  // Header
-  console.log(
-    `  ${dimText('Task'.padEnd(12))} ${dimText('Status'.padEnd(14))} ${dimText('Version'.padEnd(14))} ${dimText('Accuracy'.padEnd(10))} ${dimText('Size')}`
-  )
-  for (const row of rows) {
-    console.log(row)
-  }
-  console.log()
-  console.log(`  ${dimText('Total size:')} ${dimText(formatBytes(totalSize))}`)
-  console.log()
-}
-
-// ─── status ──────────────────────────────────────────────────────
-
-async function showStatus() {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('SLM Inference Status'))
-  console.log()
-
-  // Check ONNX runtime availability
-  let onnxAvailable = false
-  let onnxBackend = 'not installed'
-  try {
-    await import('onnxruntime-node')
-    onnxAvailable = true
-    onnxBackend = 'onnxruntime-node (native)'
-  } catch {
-    try {
-      await import('onnxruntime-web')
-      onnxAvailable = true
-      onnxBackend = 'onnxruntime-web (WASM)'
-    } catch {
-      // Neither available
-    }
-  }
-
-  const config = loadConfigFile()
-  const slmEnabled = config.slm?.enabled ?? false
-  const taskConfig = config.slm?.tasks ?? {}
-
-  const paths = getHomePaths()
-  const manifest = loadManifest()
-
-  const headerItems: Array<{ label: string; value: string; status: 'success' | 'warning' | 'error' | 'info' }> = [
-    {
-      label: 'ONNX Runtime',
-      value: onnxBackend,
-      status: onnxAvailable ? 'success' : 'warning',
-    },
-    {
-      label: 'SLM Enabled',
-      value: slmEnabled ? 'yes' : 'no',
-      status: slmEnabled ? 'success' : 'info',
-    },
-    {
-      label: 'Confidence Threshold',
-      value: `${config.slm?.confidenceThreshold ?? 0.7}`,
-      status: 'info',
-    },
-    {
-      label: 'Models Dir',
-      value: paths.models,
-      status: existsSync(paths.models) ? 'success' : 'warning',
-    },
-  ]
-
-  console.log(summaryPanel('Configuration', headerItems))
-  console.log()
-
-  // Per-task status
-  const taskItems: Array<{ label: string; value: string; status: 'success' | 'warning' | 'error' | 'info' }> = []
-
-  for (const task of ALL_TASKS) {
-    const mode = taskConfig[task] || DISABLE_MODE[task]
-    const entry = manifest?.models?.[task]
-    const fileExists = entry ? existsSync(join(paths.models, entry.file)) : false
-
-    let statusStr = `mode: ${mode}`
-    if (entry) {
-      statusStr += fileExists ? ', model: available' : ', model: MISSING'
-    } else {
-      statusStr += ', no manifest entry'
-    }
-
-    const isActive = (mode === 'model' || mode === 'both') && fileExists && onnxAvailable
-    taskItems.push({
-      label: task,
-      value: statusStr,
-      status: isActive ? 'success' : mode === 'model' || mode === 'both' ? 'error' : 'info',
-    })
-  }
-
-  console.log(summaryPanel('Task Routing', taskItems))
-  console.log()
-
-  if (!onnxAvailable) {
-    console.log(warningText('  ONNX Runtime is not installed. Models cannot be loaded.'))
-    console.log(dimText('    Install with: npm install onnxruntime-node'))
-    console.log()
-  }
-
-  if (!slmEnabled && manifest) {
-    console.log(dimText('  SLM is disabled. Enable with: claude-brain models enable all'))
-    console.log()
-  }
-}
-
-function loadManifest(): ModelManifest | null {
-  const manifestPath = join(getHomePaths().models, 'manifest.json')
-  if (!existsSync(manifestPath)) return null
-  try {
-    return JSON.parse(readFileSync(manifestPath, 'utf-8')) as ModelManifest
-  } catch {
-    return null
-  }
-}
-
-// ─── download ─────────────────────────────────────────────────────
-
-async function downloadModels(taskFilter: string, source: string) {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('Download Models'))
-  console.log()
-
-  const paths = getHomePaths()
-
-  // Validate task filter
-  if (taskFilter !== 'all' && !ALL_TASKS.includes(taskFilter as ModelTask)) {
-    console.log(errorText(`Invalid task: ${taskFilter}`))
-    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}`))
-    return
-  }
-
-  const tasks = taskFilter === 'all' ? ALL_TASKS : [taskFilter as ModelTask]
-
-  // Ensure models directory exists
-  if (!existsSync(paths.models)) {
-    mkdirSync(paths.models, { recursive: true })
-    console.log(successText(`Created models directory: ${paths.models}`))
-  }
-
-  // Hugging Face Hub source
-  if (source === 'hf' || source === 'release') {
-    return downloadFromHF(tasks, paths.models)
-  }
-
-  // Local source — copy from ~/slm-training/models/
-  const sourceDir = join(homedir(), 'slm-training', 'models')
-
-  console.log(`  ${dimText('Source:')} ${sourceDir}`)
-  console.log(`  ${dimText('Target:')} ${paths.models}`)
-  console.log()
-
-  if (!existsSync(sourceDir)) {
-    console.log(errorText(`Source directory not found: ${sourceDir}`))
-    console.log(dimText('Train models first, then place .onnx and .json files in the source directory.'))
-    return
-  }
-
-  let installed = 0
-  let totalBytes = 0
-  const installedTasks: ModelTask[] = []
-  const manifestModels: Partial<Record<ModelTask, ModelManifestEntry>> = {}
-
-  // Load existing manifest to preserve entries for tasks we're not updating
-  const manifestPath = join(paths.models, 'manifest.json')
-  if (existsSync(manifestPath)) {
-    try {
-      const existing: ModelManifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
-      if (existing.models) {
-        Object.assign(manifestModels, existing.models)
-      }
-    } catch {
-      // Ignore corrupt manifest, we'll overwrite it
-    }
-  }
-
-  for (const task of tasks) {
-    const onnxFile = `${task}.onnx`
-    const metaFile = `${task}.json`
-    const srcOnnx = join(sourceDir, onnxFile)
-    const srcMeta = join(sourceDir, metaFile)
-
-    if (!existsSync(srcOnnx)) {
-      console.log(`  ${warningText(`${onnxFile} not found in source — skipping`)}`)
-      continue
-    }
-
-    // Copy ONNX model
-    const dstOnnx = join(paths.models, onnxFile)
-    copyFileSync(srcOnnx, dstOnnx)
-
-    // Verify copied file is valid (non-empty and size matches source)
-    const srcSize = statSync(srcOnnx).size
-    const dstSize = statSync(dstOnnx).size
-    if (dstSize === 0 || dstSize !== srcSize) {
-      console.log(`  ${errorText(`${onnxFile} copy verification failed (src: ${formatBytes(srcSize)}, dst: ${formatBytes(dstSize)}) — skipping`)}`)
-      continue
-    }
-
-    totalBytes += dstSize
-    console.log(`  Copying ${onnxFile}... ${successText('done')} ${dimText(`(${formatBytes(dstSize)})`)}`)
-
-    // Copy metadata if present
-    let meta: Partial<ModelManifestEntry> = {}
-    if (existsSync(srcMeta)) {
-      const dstMeta = join(paths.models, metaFile)
-      copyFileSync(srcMeta, dstMeta)
-      try {
-        meta = JSON.parse(readFileSync(dstMeta, 'utf-8'))
-      } catch {
-        console.log(`  ${warningText(`Failed to parse ${metaFile} — using defaults`)}`)
-      }
-    }
-
-    // Build manifest entry for this task
-    // Map training metadata fields → manifest fields:
-    //   val_acc → accuracy, block_size → maxSeqLen, model_name → params
-    const metaAny = meta as Record<string, any>
-    manifestModels[task] = {
-      version: meta.version ?? '0.1.0',
-      file: onnxFile,
-      sha256: meta.sha256,
-      params: meta.params ?? metaAny.model_name,
-      accuracy: meta.accuracy ?? metaAny.val_acc,
-      labels: meta.labels,
-      maxSeqLen: meta.maxSeqLen ?? metaAny.block_size,
-    }
-
-    installedTasks.push(task)
-    installed++
-  }
-
-  console.log()
-
-  if (installed === 0) {
-    console.log(warningText('No models were installed.'))
-    console.log()
-    return
-  }
-
-  // Write manifest
-  const manifest: ModelManifest = { models: manifestModels }
-  writeFileSync(manifestPath, JSON.stringify(manifest, null, 2))
-
-  // Auto-enable successfully installed models in config
-  const config = loadConfigFile()
-  if (!config.slm) config.slm = {}
-  config.slm.enabled = true
-  if (!config.slm.tasks) config.slm.tasks = {}
-  for (const task of installedTasks) {
-    config.slm.tasks[task] = 'model'
-  }
-  saveConfigFile(config)
-  updateConfigYml(installedTasks, 'model')
-
-  console.log(successText(`Installed ${installed} model${installed !== 1 ? 's' : ''} (total: ${formatBytes(totalBytes)})`))
-  console.log(successText(`Auto-enabled ${installedTasks.join(', ')} in config`))
-  console.log(dimText(`Manifest written to ${manifestPath}`))
-  console.log()
-}
-
-// ─── download from HF ─────────────────────────────────────────────
-
-const isTTY = process.stdout.isTTY === true
-
-async function downloadFromHF(tasks: ModelTask[], modelsDir: string) {
-  const manifest = hfManifestData as HfManifest
-
-  // Compute total download size
-  let totalSize = 0
-  for (const task of tasks) {
-    const entry = manifest.models[task]
-    if (entry) totalSize += entry.size
-  }
-
-  console.log(`  ${dimText('Source:')} Hugging Face Hub (${manifest.hfRepo})`)
-  console.log(`  ${dimText('Target:')} ${modelsDir}`)
-  console.log(`  ${dimText('Models:')} ${tasks.join(', ')}`)
-  console.log(`  ${dimText('Total:')}  ~${formatBytes(totalSize)}`)
-  console.log()
-
-  const results = await downloadFromHuggingFace(manifest, {
-    destDir: modelsDir,
-    tasks,
-    onProgress(task, downloaded, total) {
-      if (isTTY && total > 0) {
-        const pct = (downloaded / total) * 100
-        const bar = progressBar(pct, 25)
-        process.stdout.write(`\r  ${task.padEnd(12)} ${bar}  ${formatBytes(downloaded)} / ${formatBytes(total)}`)
-      }
-    },
-    onComplete(task, bytes) {
-      if (isTTY) process.stdout.write('\r' + ' '.repeat(80) + '\r')
-      console.log(`  ${successText(task.padEnd(12))} ${formatBytes(bytes)} ${dimText('SHA256 verified')}`)
-    },
-    onError(task, error) {
-      if (isTTY) process.stdout.write('\r' + ' '.repeat(80) + '\r')
-      console.log(`  ${errorText(task.padEnd(12))} ${error}`)
-    },
-  })
-
-  console.log()
-
-  const succeeded = results.filter(r => r.success)
-  if (succeeded.length === 0) {
-    console.log(warningText('No models were downloaded.'))
-    console.log()
-    return
-  }
-
-  // Build manifest from downloaded results + existing
-  const manifestPath = join(modelsDir, 'manifest.json')
-  const manifestModels: Partial<Record<ModelTask, ModelManifestEntry>> = {}
-
-  // Preserve existing manifest entries
-  if (existsSync(manifestPath)) {
-    try {
-      const existing: ModelManifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
-      if (existing.models) Object.assign(manifestModels, existing.models)
-    } catch { /* overwrite corrupt */ }
-  }
-
-  const installedTasks: ModelTask[] = []
-  let totalBytes = 0
-
-  for (const result of succeeded) {
-    const task = result.task as ModelTask
-    const entry = manifest.models[task]
-    if (!entry) continue
-
-    manifestModels[task] = {
-      version: entry.version,
-      file: entry.file,
-      sha256: entry.sha256,
-      params: entry.params,
-      accuracy: entry.accuracy ?? undefined,
-      labels: entry.labels,
-      maxSeqLen: entry.maxSeqLen,
-    }
-    installedTasks.push(task)
-    totalBytes += result.bytes
-  }
-
-  // Write manifest
-  const localManifest: ModelManifest = { models: manifestModels }
-  writeFileSync(manifestPath, JSON.stringify(localManifest, null, 2))
-
-  // Auto-enable models in config
-  const config = loadConfigFile()
-  if (!config.slm) config.slm = {}
-  config.slm.enabled = true
-  if (!config.slm.tasks) config.slm.tasks = {}
-  for (const task of installedTasks) {
-    config.slm.tasks[task] = 'model'
-  }
-  saveConfigFile(config)
-  updateConfigYml(installedTasks, 'model')
-
-  console.log(successText(`Downloaded ${succeeded.length} model${succeeded.length !== 1 ? 's' : ''} (${formatBytes(totalBytes)})`))
-  console.log(successText(`Auto-enabled ${installedTasks.join(', ')} in config`))
-  console.log(dimText(`Manifest written to ${manifestPath}`))
-  console.log()
-}
-
-// ─── enable ───────────────────────────────────────────────────────
-
-function enableTask(taskArg: string) {
-  const tasks = resolveTaskArg(taskArg, 'enable')
-  if (!tasks) return
-
-  console.log()
-  console.log(heading('Enable SLM Models'))
-  console.log()
-
-  const config = loadConfigFile()
-
-  // Ensure slm section exists
-  if (!config.slm) config.slm = {}
-  config.slm.enabled = true
-  if (!config.slm.tasks) config.slm.tasks = {}
-
-  for (const task of tasks) {
-    config.slm.tasks[task] = 'model'
-  }
-
-  saveConfigFile(config)
-  updateConfigYml(tasks, 'model')
-
-  for (const task of tasks) {
-    console.log(successText(`${task} -> model`))
-  }
-  console.log()
-  console.log(successText('SLM enabled'))
-  console.log(dimText('  Changes take effect on next server restart'))
-  console.log()
-}
-
-// ─── disable ──────────────────────────────────────────────────────
-
-function disableTask(taskArg: string) {
-  const tasks = resolveTaskArg(taskArg, 'disable')
-  if (!tasks) return
-
-  console.log()
-  console.log(heading('Disable SLM Models'))
-  console.log()
-
-  const config = loadConfigFile()
-  if (!config.slm) config.slm = {}
-  if (!config.slm.tasks) config.slm.tasks = {}
-
-  for (const task of tasks) {
-    config.slm.tasks[task] = DISABLE_MODE[task]
-  }
-
-  // If all tasks are now regex/api, disable SLM globally
-  const allDisabled = ALL_TASKS.every(t => {
-    const mode = config.slm.tasks[t] || DISABLE_MODE[t]
-    return mode === 'regex' || mode === 'api'
-  })
-  if (allDisabled) {
-    config.slm.enabled = false
-  }
-
-  saveConfigFile(config)
-  updateConfigYml(tasks, 'disable')
-
-  for (const task of tasks) {
-    console.log(warningText(`${task} -> ${DISABLE_MODE[task]}`))
-  }
-  console.log()
-  if (allDisabled) {
-    console.log(warningText('All tasks disabled — SLM globally disabled'))
-  }
-  console.log(dimText('  Changes take effect on next server restart'))
-  console.log()
-}
-
-/** Resolve a task argument to a list of tasks, or null if invalid */
-function resolveTaskArg(taskArg: string, verb: string): ModelTask[] | null {
-  if (!taskArg) {
-    console.log(errorText('Missing task argument'))
-    console.log(dimText(`Usage: claude-brain models ${verb} <task|all>`))
-    console.log(dimText(`Tasks: ${ALL_TASKS.join(', ')}, all`))
-    return null
-  }
-
-  if (taskArg === 'all') return [...ALL_TASKS]
-
-  if (!ALL_TASKS.includes(taskArg as ModelTask)) {
-    console.log(errorText(`Invalid task: ${taskArg}`))
-    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}, all`))
-    return null
-  }
-
-  return [taskArg as ModelTask]
-}
-
-// ─── benchmark ────────────────────────────────────────────────────
-
-async function benchmarkTask(task: string) {
-  if (!task) {
-    console.log(errorText('Missing task argument'))
-    console.log(dimText(`Usage: claude-brain models benchmark <task>`))
-    console.log(dimText(`Tasks: ${ALL_TASKS.join(', ')}`))
-    return
-  }
-
-  if (!ALL_TASKS.includes(task as ModelTask)) {
-    console.log(errorText(`Invalid task: ${task}`))
-    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}`))
-    return
-  }
-
-  console.log()
-  console.log(heading(`Benchmark: ${task}`))
-  console.log()
-
-  // Look for test data
-  const home = getClaudeBrainHome()
-  const testDataPath = join(home, 'training', 'benchmarks', 'baseline', `${task}_test.jsonl`)
-
-  if (!existsSync(testDataPath)) {
-    console.log(warningText(`Test data not found: ${testDataPath}`))
-    console.log(dimText('Generate test data first with: claude-brain export-training ' + task))
-    return
-  }
-
-  // Load test examples
-  const lines = readFileSync(testDataPath, 'utf-8').trim().split('\n').filter(Boolean)
-  if (lines.length === 0) {
-    console.log(warningText('Test file is empty'))
-    return
-  }
-
-  console.log(dimText(`Loaded ${lines.length} test examples from ${testDataPath}`))
-  console.log()
-
-  // Run through regex classifier for intent task
-  if (task === 'intent') {
-    await benchmarkIntent(lines)
-  } else {
-    // For non-intent tasks, just report data availability
-    console.log(dimText(`Benchmark for "${task}" requires model inference (ONNX).`))
-    console.log(dimText(`Regex-only benchmarking is available for the intent task.`))
-    console.log()
-    console.log(dimText(`Test data: ${lines.length} examples ready`))
-  }
-}
-
-async function benchmarkIntent(lines: string[]) {
-  // Dynamic import to avoid circular deps
-  const { IntentClassifier } = await import('@/routing/intent-classifier')
-  const classifier = new IntentClassifier()
-
-  let correct = 0
-  const labelCounts: Record<string, { tp: number; fp: number; fn: number }> = {}
-
-  for (const line of lines) {
-    let example: { input: string; output: { label: string } }
-    try {
-      example = JSON.parse(line)
-    } catch {
-      continue
-    }
-
-    const expected = example.output?.label
-    if (!expected) continue
-
-    const result = classifier.classify(example.input)
-    const predicted = result.primary
-
-    // Initialize counters
-    if (!labelCounts[expected]) labelCounts[expected] = { tp: 0, fp: 0, fn: 0 }
-    if (!labelCounts[predicted]) labelCounts[predicted] = { tp: 0, fp: 0, fn: 0 }
-
-    if (predicted === expected) {
-      correct++
-      labelCounts[expected].tp++
-    } else {
-      labelCounts[expected].fn++
-      labelCounts[predicted].fp++
-    }
-  }
-
-  const total = lines.length
-  const accuracy = total > 0 ? correct / total : 0
-
-  console.log(`  ${theme.bold('Overall Accuracy:')} ${accuracy >= 0.7 ? successText(`${(accuracy * 100).toFixed(1)}%`) : warningText(`${(accuracy * 100).toFixed(1)}%`)} (${correct}/${total})`)
-  console.log()
-
-  // Per-class metrics
-  console.log(`  ${dimText('Label'.padEnd(20))} ${dimText('Prec'.padEnd(8))} ${dimText('Recall'.padEnd(8))} ${dimText('F1'.padEnd(8))} ${dimText('Support')}`)
-  console.log(`  ${dimText('-'.repeat(52))}`)
-
-  const sortedLabels = Object.keys(labelCounts).sort()
-  for (const label of sortedLabels) {
-    const { tp, fp, fn } = labelCounts[label]
-    const precision = tp + fp > 0 ? tp / (tp + fp) : 0
-    const recall = tp + fn > 0 ? tp / (tp + fn) : 0
-    const f1 = precision + recall > 0 ? (2 * precision * recall) / (precision + recall) : 0
-    const support = tp + fn
-
-    console.log(
-      `  ${theme.primary(label.padEnd(20))} ${(precision * 100).toFixed(1).padStart(5)}%  ${(recall * 100).toFixed(1).padStart(5)}%  ${(f1 * 100).toFixed(1).padStart(5)}%  ${String(support).padStart(5)}`
-    )
-  }
-  console.log()
-}
-
-// ─── stats ────────────────────────────────────────────────────────
-
-async function showStats() {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('Training Data Statistics'))
-  console.log()
-
-  const stats = getTrainingStats()
-  const items: Array<{ label: string; value: string; status?: 'success' | 'warning' | 'error' | 'info' }> = []
-
-  let grandTotal = 0
-  let grandVerified = 0
-
-  for (const task of ALL_TASKS) {
-    const { total, verified } = stats[task]
-    grandTotal += total
-    grandVerified += verified
-    items.push({
-      label: task,
-      value: `${total} total, ${verified} verified`,
-      status: total > 0 ? 'success' : 'warning'
-    })
-  }
-
-  items.push({
-    label: 'Total',
-    value: `${grandTotal} total, ${grandVerified} verified`,
-    status: grandTotal > 0 ? 'success' : 'warning'
-  })
-
-  console.log(summaryPanel('Training Data', items))
-  console.log()
-
-  // Check for model_feedback table
-  try {
-    const { Database } = await import('bun:sqlite') as any
-    const dbPath = join(getClaudeBrainHome(), 'data', 'memory.db')
-    if (existsSync(dbPath)) {
-      const db = new Database(dbPath, { readonly: true })
-      const hasTable = db.prepare("SELECT name FROM sqlite_master WHERE type='table' AND name='model_feedback'").get()
-      if (hasTable) {
-        const disagreements = (db.prepare('SELECT COUNT(*) as c FROM model_feedback WHERE model_label != regex_label').get() as any)?.c ?? 0
-        const totalFeedback = (db.prepare('SELECT COUNT(*) as c FROM model_feedback').get() as any)?.c ?? 0
-        console.log(`  ${dimText('Model Feedback:')} ${totalFeedback} entries, ${disagreements} disagreements`)
-        console.log()
-      }
-      db.close()
-    }
-  } catch {
-    // model_feedback table doesn't exist yet — that's fine
-  }
-}
-
-// ─── retrain ──────────────────────────────────────────────────────
-
-async function retrainModels(taskFilter: string, force: boolean) {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('Retrain Models'))
-  console.log()
-
-  // Validate task filter
-  if (taskFilter !== 'all' && !ALL_TASKS.includes(taskFilter as ModelTask)) {
-    console.log(errorText(`Invalid task: ${taskFilter}`))
-    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}, all`))
-    return
-  }
-
-  // Build config from schema defaults + .claudebrainrc.json overrides
-  const userConfig = loadConfigFile()
-  const retrainCfg = userConfig?.slm?.retrain ?? {}
-
-  const config: RetrainConfig = {
-    minFeedbackCount: retrainCfg.minFeedbackCount ?? 100,
-    maxDisagreementRate: retrainCfg.maxDisagreementRate ?? 0.15,
-    pythonPath: retrainCfg.pythonPath ?? 'python3',
-    trainingDir: retrainCfg.trainingDir ?? '~/slm-training',
-    force,
-  }
-
-  console.log(`  ${dimText('Training dir:')} ${config.trainingDir}`)
-  console.log(`  ${dimText('Python path:')} ${config.pythonPath}`)
-  console.log(`  ${dimText('Min feedback:')} ${config.minFeedbackCount}`)
-  console.log(`  ${dimText('Max disagreement:')} ${(config.maxDisagreementRate * 100).toFixed(0)}%`)
-  if (force) console.log(`  ${warningText('Force mode enabled — skipping checks')}`)
-  console.log()
-
-  if (taskFilter === 'all') {
-    // Retrain all tasks that need it
-    const results = await retrainAll(config)
-
-    if (results.size === 0) {
-      console.log(dimText('  No tasks needed retraining.'))
-      console.log()
-      return
-    }
-
-    // Summary
-    console.log()
-    console.log(heading('Retrain Summary'))
-    console.log()
-    for (const [task, result] of results) {
-      if (result.success) {
-        const accStr = result.newAccuracy != null
-          ? `${(result.newAccuracy * 100).toFixed(1)}%`
-          : 'n/a'
-        const oldStr = result.oldAccuracy != null
-          ? ` (was ${(result.oldAccuracy * 100).toFixed(1)}%)`
-          : ''
-        console.log(`  ${successText(task.padEnd(12))} accuracy: ${accStr}${oldStr}  ${dimText(`${result.trainingDataCount} examples, ${(result.duration / 1000).toFixed(1)}s`)}`)
-      } else {
-        console.log(`  ${errorText(task.padEnd(12))} ${result.error}`)
-      }
-    }
-  } else {
-    // Single task
-    const task = taskFilter as ModelTask
-
-    if (!force) {
-      const check = shouldRetrain(task, config)
-      console.log(`  ${dimText('Feedback count:')} ${check.feedbackCount}`)
-      console.log(`  ${dimText('Disagreement rate:')} ${(check.disagreementRate * 100).toFixed(1)}%`)
-      console.log(`  ${dimText('Last retrain:')} ${check.lastRetrainDate ?? 'never'}`)
-      console.log(`  ${dimText('Needs retrain:')} ${check.needed ? 'yes' : 'no'} — ${check.reason}`)
-      console.log()
-
-      if (!check.needed) {
-        console.log(dimText('  Retrain not needed. Use --force to override.'))
-        console.log()
-        return
-      }
-    }
-
-    const result = await retrainTask(task, config)
-
-    console.log()
-    if (result.success) {
-      const accStr = result.newAccuracy != null
-        ? `${(result.newAccuracy * 100).toFixed(1)}%`
-        : 'n/a'
-      const oldStr = result.oldAccuracy != null
-        ? ` (was ${(result.oldAccuracy * 100).toFixed(1)}%)`
-        : ''
-      console.log(successText(`Retrain complete: accuracy ${accStr}${oldStr}`))
-      console.log(dimText(`  ${result.trainingDataCount} examples, ${(result.duration / 1000).toFixed(1)}s`))
-    } else {
-      console.log(errorText(`Retrain failed: ${result.error}`))
-    }
-  }
-
-  console.log()
-}
-
-// ─── help ─────────────────────────────────────────────────────────
-
-function printModelsHelp() {
-  console.log()
-  console.log(renderLogo())
-  console.log()
-  console.log(heading('SLM Model Management'))
-  console.log()
-
-  const subcommands = [
-    ['list', 'Show installed models and their status'],
-    ['status', 'Show inference routing and ONNX runtime status'],
-    ['download', 'Download models (--source local|hf, --task <task>|all)'],
-    ['enable <task|all>', 'Enable model inference for task(s)'],
-    ['disable <task|all>', 'Disable model inference for task(s)'],
-    ['benchmark <task>', 'Run accuracy benchmark on test data'],
-    ['stats', 'Show training data statistics'],
-    ['retrain [<task>|all]', 'Retrain models from feedback (--force)'],
-  ]
-
-  const lines = subcommands
-    .map(([cmd, desc]) => `  ${theme.primary(cmd!.padEnd(20))} ${dimText(desc!)}`)
-    .join('\n')
-
-  console.log(theme.bold('Usage:') + ' ' + dimText('claude-brain models <subcommand>'))
-  console.log()
-  console.log(theme.bold('Subcommands:'))
-  console.log(lines)
-  console.log()
-  console.log(theme.bold('Tasks:') + ' ' + dimText(ALL_TASKS.join(', ')))
-  console.log()
-  console.log(theme.bold('Examples:'))
-  console.log(`  ${dimText('claude-brain models list')}`)
-  console.log(`  ${dimText('claude-brain models status')}`)
-  console.log(`  ${dimText('claude-brain models download --source hf')}`)
-  console.log(`  ${dimText('claude-brain models download --source hf --task intent')}`)
-  console.log(`  ${dimText('claude-brain models download --source local')}`)
-  console.log(`  ${dimText('claude-brain models enable all')}`)
-  console.log(`  ${dimText('claude-brain models enable intent')}`)
-  console.log(`  ${dimText('claude-brain models disable pattern')}`)
-  console.log(`  ${dimText('claude-brain models benchmark intent')}`)
-  console.log(`  ${dimText('claude-brain models stats')}`)
-  console.log(`  ${dimText('claude-brain models retrain intent')}`)
-  console.log(`  ${dimText('claude-brain models retrain all --force')}`)
-  console.log()
-}
-
-// ─── helpers ──────────────────────────────────────────────────────
-
-const RC_FILE = '.claudebrainrc.json'
-
-function loadConfigFile(): Record<string, any> {
-  const rcPath = join(getClaudeBrainHome(), RC_FILE)
-  if (!existsSync(rcPath)) return {}
-  try {
-    return JSON.parse(readFileSync(rcPath, 'utf-8'))
-  } catch {
-    return {}
-  }
-}
-
-function saveConfigFile(config: Record<string, any>): void {
-  const rcPath = join(getClaudeBrainHome(), RC_FILE)
-  writeFileSync(rcPath, JSON.stringify(config, null, 2) + '\n', 'utf-8')
-}
-
-/** Best-effort update of config.yml task lines for user visibility */
-function updateConfigYml(tasks: ModelTask[], mode: string): void {
-  const ymlPath = join(getClaudeBrainHome(), 'config.yml')
-  if (!existsSync(ymlPath)) return
-
-  try {
-    let content = readFileSync(ymlPath, 'utf-8')
-
-    // Update each task line (e.g. "    intent: regex" -> "    intent: model")
-    for (const task of tasks) {
-      const newMode = mode === 'disable' ? DISABLE_MODE[task] : mode
-      const regex = new RegExp(`^(\\s*${task}:\\s*)\\S+`, 'm')
-      content = content.replace(regex, `$1${newMode}`)
-    }
-
-    // Update slm.enabled line if present
-    // Use a heuristic: find "enabled:" that appears after "slm:"
-    const slmIdx = content.indexOf('slm:')
-    if (slmIdx !== -1) {
-      const afterSlm = content.slice(slmIdx)
-      const enabledMatch = afterSlm.match(/^(\s+enabled:\s*)\S+/m)
-      if (enabledMatch) {
-        const config = loadConfigFile()
-        const slmEnabled = config.slm?.enabled ?? false
-        const newLine = `${enabledMatch[1]}${slmEnabled}`
-        content = content.slice(0, slmIdx) + afterSlm.replace(enabledMatch[0], newLine)
-      }
-    }
-
-    writeFileSync(ymlPath, content, 'utf-8')
-  } catch {
-    // Non-critical — .claudebrainrc.json is the authoritative config
-  }
-}
-
-function formatBytes(bytes: number): string {
-  if (bytes === 0) return '0 B'
-  const units = ['B', 'KB', 'MB', 'GB']
-  const i = Math.floor(Math.log(bytes) / Math.log(1024))
-  const val = bytes / Math.pow(1024, i)
-  return `${val.toFixed(i > 0 ? 1 : 0)} ${units[i]}`
-}
+/**
+ * Models Command — SLM Upgrade Phase 5
+ * Manage SLM models: list, download, enable, disable, benchmark, stats.
+ *
+ * Usage:
+ *   claude-brain models list
+ *   claude-brain models download [--task <task>|all]
+ *   claude-brain models enable <task>
+ *   claude-brain models disable <task>
+ *   claude-brain models benchmark <task>
+ *   claude-brain models stats
+ */
+
+import { readFileSync, existsSync, mkdirSync, writeFileSync, statSync, copyFileSync } from 'node:fs'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+import { parseArgs } from 'citty'
+import { renderLogo, theme, heading, dimText, successText, warningText, errorText, box, summaryPanel } from '@/cli/ui/index.js'
+import { progressBar } from '@/cli/ui/components.js'
+import { getHomePaths, getClaudeBrainHome } from '@/config/home'
+import { getTrainingStats, type TrainingTask } from '@/training/data-store'
+import type { ModelManifest, ModelManifestEntry, ModelTask } from '@/intelligence/model-manager'
+import { shouldRetrain, retrainTask, retrainAll, type RetrainConfig } from '@/training/retrain-pipeline'
+import { downloadFromHuggingFace, type HfManifest } from '@/intelligence/hf-downloader'
+import hfManifestData from '@/intelligence/hf-manifest.json'
+
+const ALL_TASKS: ModelTask[] = ['intent', 'entity', 'query', 'knowledge', 'compress', 'pattern']
+
+/** Default mode when disabling a task */
+const DISABLE_MODE: Record<ModelTask, string> = {
+  intent: 'regex',
+  entity: 'regex',
+  query: 'regex',
+  knowledge: 'regex',
+  compress: 'api',
+  pattern: 'regex',
+}
+
+export async function runModels() {
+  const args = parseArgs(process.argv.slice(3), {
+    subcommand: { type: 'positional', required: false, description: 'Subcommand: list, status, download, enable, disable, benchmark, stats, retrain' },
+    taskArg: { type: 'positional', required: false, description: 'Task name or "all" (for enable/disable/benchmark/retrain)' },
+    task: { type: 'string', description: 'Target task (for download --task)' },
+    source: { type: 'string', description: 'Source: local (default) or hf (Hugging Face Hub)' },
+    force: { type: 'boolean', description: 'Force retrain even if checks say not needed' },
+  })
+
+  const subcommand = args.subcommand || ''
+  const taskArg = (args.task || args.taskArg || '') as string
+
+  switch (subcommand) {
+    case 'list':
+      return listModels()
+    case 'status':
+      return showStatus()
+    case 'download':
+      return downloadModels(taskArg || 'all', (args.source as string) || 'local')
+    case 'enable':
+      return enableTask(taskArg)
+    case 'disable':
+      return disableTask(taskArg)
+    case 'benchmark':
+      return benchmarkTask(taskArg)
+    case 'stats':
+      return showStats()
+    case 'retrain':
+      return retrainModels(taskArg || 'all', !!args.force)
+    default:
+      return printModelsHelp()
+  }
+}
+
+// ─── list ─────────────────────────────────────────────────────────
+
+function listModels() {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('SLM Models'))
+  console.log()
+
+  const paths = getHomePaths()
+  const manifestPath = join(paths.models, 'manifest.json')
+
+  const items: Array<{ label: string; value: string; status?: 'success' | 'warning' | 'error' | 'info' }> = []
+
+  if (!existsSync(manifestPath)) {
+    items.push({ label: 'Manifest', value: 'Not found — run "models download" first', status: 'warning' })
+    console.log(summaryPanel('Models', items))
+    console.log()
+
+    // Still show task status even without manifest
+    for (const task of ALL_TASKS) {
+      console.log(`  ${theme.primary(task.padEnd(12))} ${warningText('not installed')}`)
+    }
+    console.log()
+    return
+  }
+
+  let manifest: ModelManifest
+  try {
+    manifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
+  } catch {
+    console.log(errorText('Failed to parse manifest.json'))
+    return
+  }
+
+  let totalSize = 0
+  const rows: string[] = []
+
+  for (const task of ALL_TASKS) {
+    const entry = manifest.models?.[task]
+    if (!entry) {
+      rows.push(`  ${theme.primary(task.padEnd(12))} ${warningText('not installed')}`)
+      continue
+    }
+
+    const modelPath = join(paths.models, entry.file)
+    const installed = existsSync(modelPath)
+
+    let sizeStr = '—'
+    if (installed) {
+      try {
+        const size = statSync(modelPath).size
+        totalSize += size
+        sizeStr = formatBytes(size)
+      } catch {
+        sizeStr = '?'
+      }
+    }
+
+    const status = installed ? successText('installed') : warningText('missing')
+    const version = entry.version ? dimText(`v${entry.version}`) : ''
+    const accuracy = entry.accuracy != null ? dimText(`${(entry.accuracy * 100).toFixed(1)}%`) : ''
+
+    rows.push(
+      `  ${theme.primary(task.padEnd(12))} ${status.padEnd(24)} ${version.padEnd(14)} ${accuracy.padEnd(10)} ${dimText(sizeStr)}`
+    )
+  }
+
+  // Header
+  console.log(
+    `  ${dimText('Task'.padEnd(12))} ${dimText('Status'.padEnd(14))} ${dimText('Version'.padEnd(14))} ${dimText('Accuracy'.padEnd(10))} ${dimText('Size')}`
+  )
+  for (const row of rows) {
+    console.log(row)
+  }
+  console.log()
+  console.log(`  ${dimText('Total size:')} ${dimText(formatBytes(totalSize))}`)
+  console.log()
+}
+
+// ─── status ──────────────────────────────────────────────────────
+
+async function showStatus() {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('SLM Inference Status'))
+  console.log()
+
+  // Check ONNX runtime availability
+  let onnxAvailable = false
+  let onnxBackend = 'not installed'
+  try {
+    await import('onnxruntime-node')
+    onnxAvailable = true
+    onnxBackend = 'onnxruntime-node (native)'
+  } catch {
+    try {
+      await import('onnxruntime-web')
+      onnxAvailable = true
+      onnxBackend = 'onnxruntime-web (WASM)'
+    } catch {
+      // Neither available
+    }
+  }
+
+  const config = loadConfigFile()
+  const slmEnabled = config.slm?.enabled ?? false
+  const taskConfig = config.slm?.tasks ?? {}
+
+  const paths = getHomePaths()
+  const manifest = loadManifest()
+
+  const headerItems: Array<{ label: string; value: string; status: 'success' | 'warning' | 'error' | 'info' }> = [
+    {
+      label: 'ONNX Runtime',
+      value: onnxBackend,
+      status: onnxAvailable ? 'success' : 'warning',
+    },
+    {
+      label: 'SLM Enabled',
+      value: slmEnabled ? 'yes' : 'no',
+      status: slmEnabled ? 'success' : 'info',
+    },
+    {
+      label: 'Confidence Threshold',
+      value: `${config.slm?.confidenceThreshold ?? 0.7}`,
+      status: 'info',
+    },
+    {
+      label: 'Models Dir',
+      value: paths.models,
+      status: existsSync(paths.models) ? 'success' : 'warning',
+    },
+  ]
+
+  console.log(summaryPanel('Configuration', headerItems))
+  console.log()
+
+  // Per-task status
+  const taskItems: Array<{ label: string; value: string; status: 'success' | 'warning' | 'error' | 'info' }> = []
+
+  for (const task of ALL_TASKS) {
+    const mode = taskConfig[task] || DISABLE_MODE[task]
+    const entry = manifest?.models?.[task]
+    const fileExists = entry ? existsSync(join(paths.models, entry.file)) : false
+
+    let statusStr = `mode: ${mode}`
+    if (entry) {
+      statusStr += fileExists ? ', model: available' : ', model: MISSING'
+    } else {
+      statusStr += ', no manifest entry'
+    }
+
+    const isActive = (mode === 'model' || mode === 'both') && fileExists && onnxAvailable
+    taskItems.push({
+      label: task,
+      value: statusStr,
+      status: isActive ? 'success' : mode === 'model' || mode === 'both' ? 'error' : 'info',
+    })
+  }
+
+  console.log(summaryPanel('Task Routing', taskItems))
+  console.log()
+
+  if (!onnxAvailable) {
+    console.log(warningText('  ONNX Runtime is not installed. Models cannot be loaded.'))
+    console.log(dimText('    Install with: npm install onnxruntime-node'))
+    console.log()
+  }
+
+  if (!slmEnabled && manifest) {
+    console.log(dimText('  SLM is disabled. Enable with: claude-brain models enable all'))
+    console.log()
+  }
+}
+
+function loadManifest(): ModelManifest | null {
+  const manifestPath = join(getHomePaths().models, 'manifest.json')
+  if (!existsSync(manifestPath)) return null
+  try {
+    return JSON.parse(readFileSync(manifestPath, 'utf-8')) as ModelManifest
+  } catch {
+    return null
+  }
+}
+
+// ─── download ─────────────────────────────────────────────────────
+
+async function downloadModels(taskFilter: string, source: string) {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('Download Models'))
+  console.log()
+
+  const paths = getHomePaths()
+
+  // Validate task filter
+  if (taskFilter !== 'all' && !ALL_TASKS.includes(taskFilter as ModelTask)) {
+    console.log(errorText(`Invalid task: ${taskFilter}`))
+    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}`))
+    return
+  }
+
+  const tasks = taskFilter === 'all' ? ALL_TASKS : [taskFilter as ModelTask]
+
+  // Ensure models directory exists
+  if (!existsSync(paths.models)) {
+    mkdirSync(paths.models, { recursive: true })
+    console.log(successText(`Created models directory: ${paths.models}`))
+  }
+
+  // Hugging Face Hub source
+  if (source === 'hf' || source === 'release') {
+    return downloadFromHF(tasks, paths.models)
+  }
+
+  // Local source — copy from ~/slm-training/models/
+  const sourceDir = join(homedir(), 'slm-training', 'models')
+
+  console.log(`  ${dimText('Source:')} ${sourceDir}`)
+  console.log(`  ${dimText('Target:')} ${paths.models}`)
+  console.log()
+
+  if (!existsSync(sourceDir)) {
+    console.log(errorText(`Source directory not found: ${sourceDir}`))
+    console.log(dimText('Train models first, then place .onnx and .json files in the source directory.'))
+    return
+  }
+
+  let installed = 0
+  let totalBytes = 0
+  const installedTasks: ModelTask[] = []
+  const manifestModels: Partial<Record<ModelTask, ModelManifestEntry>> = {}
+
+  // Load existing manifest to preserve entries for tasks we're not updating
+  const manifestPath = join(paths.models, 'manifest.json')
+  if (existsSync(manifestPath)) {
+    try {
+      const existing: ModelManifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
+      if (existing.models) {
+        Object.assign(manifestModels, existing.models)
+      }
+    } catch {
+      // Ignore corrupt manifest, we'll overwrite it
+    }
+  }
+
+  for (const task of tasks) {
+    const onnxFile = `${task}.onnx`
+    const metaFile = `${task}.json`
+    const srcOnnx = join(sourceDir, onnxFile)
+    const srcMeta = join(sourceDir, metaFile)
+
+    if (!existsSync(srcOnnx)) {
+      console.log(`  ${warningText(`${onnxFile} not found in source — skipping`)}`)
+      continue
+    }
+
+    // Copy ONNX model
+    const dstOnnx = join(paths.models, onnxFile)
+    copyFileSync(srcOnnx, dstOnnx)
+
+    // Verify copied file is valid (non-empty and size matches source)
+    const srcSize = statSync(srcOnnx).size
+    const dstSize = statSync(dstOnnx).size
+    if (dstSize === 0 || dstSize !== srcSize) {
+      console.log(`  ${errorText(`${onnxFile} copy verification failed (src: ${formatBytes(srcSize)}, dst: ${formatBytes(dstSize)}) — skipping`)}`)
+      continue
+    }
+
+    totalBytes += dstSize
+    console.log(`  Copying ${onnxFile}... ${successText('done')} ${dimText(`(${formatBytes(dstSize)})`)}`)
+
+    // Copy metadata if present
+    let meta: Partial<ModelManifestEntry> = {}
+    if (existsSync(srcMeta)) {
+      const dstMeta = join(paths.models, metaFile)
+      copyFileSync(srcMeta, dstMeta)
+      try {
+        meta = JSON.parse(readFileSync(dstMeta, 'utf-8'))
+      } catch {
+        console.log(`  ${warningText(`Failed to parse ${metaFile} — using defaults`)}`)
+      }
+    }
+
+    // Build manifest entry for this task
+    // Map training metadata fields → manifest fields:
+    //   val_acc → accuracy, block_size → maxSeqLen, model_name → params
+    const metaAny = meta as Record<string, any>
+    manifestModels[task] = {
+      version: meta.version ?? '0.1.0',
+      file: onnxFile,
+      sha256: meta.sha256,
+      params: meta.params ?? metaAny.model_name,
+      accuracy: meta.accuracy ?? metaAny.val_acc,
+      labels: meta.labels,
+      maxSeqLen: meta.maxSeqLen ?? metaAny.block_size,
+    }
+
+    installedTasks.push(task)
+    installed++
+  }
+
+  console.log()
+
+  if (installed === 0) {
+    console.log(warningText('No models were installed.'))
+    console.log()
+    return
+  }
+
+  // Write manifest
+  const manifest: ModelManifest = { models: manifestModels }
+  writeFileSync(manifestPath, JSON.stringify(manifest, null, 2))
+
+  // Auto-enable successfully installed models in config
+  const config = loadConfigFile()
+  if (!config.slm) config.slm = {}
+  config.slm.enabled = true
+  if (!config.slm.tasks) config.slm.tasks = {}
+  for (const task of installedTasks) {
+    config.slm.tasks[task] = 'model'
+  }
+  saveConfigFile(config)
+  updateConfigYml(installedTasks, 'model')
+
+  console.log(successText(`Installed ${installed} model${installed !== 1 ? 's' : ''} (total: ${formatBytes(totalBytes)})`))
+  console.log(successText(`Auto-enabled ${installedTasks.join(', ')} in config`))
+  console.log(dimText(`Manifest written to ${manifestPath}`))
+  console.log()
+}
+
+// ─── download from HF ─────────────────────────────────────────────
+
+const isTTY = process.stdout.isTTY === true
+
+async function downloadFromHF(tasks: ModelTask[], modelsDir: string) {
+  const manifest = hfManifestData as HfManifest
+
+  // Compute total download size
+  let totalSize = 0
+  for (const task of tasks) {
+    const entry = manifest.models[task]
+    if (entry) totalSize += entry.size
+  }
+
+  console.log(`  ${dimText('Source:')} Hugging Face Hub (${manifest.hfRepo})`)
+  console.log(`  ${dimText('Target:')} ${modelsDir}`)
+  console.log(`  ${dimText('Models:')} ${tasks.join(', ')}`)
+  console.log(`  ${dimText('Total:')}  ~${formatBytes(totalSize)}`)
+  console.log()
+
+  const results = await downloadFromHuggingFace(manifest, {
+    destDir: modelsDir,
+    tasks,
+    onProgress(task, downloaded, total) {
+      if (isTTY && total > 0) {
+        const pct = (downloaded / total) * 100
+        const bar = progressBar(pct, 25)
+        process.stdout.write(`\r  ${task.padEnd(12)} ${bar}  ${formatBytes(downloaded)} / ${formatBytes(total)}`)
+      }
+    },
+    onComplete(task, bytes) {
+      if (isTTY) process.stdout.write('\r' + ' '.repeat(80) + '\r')
+      console.log(`  ${successText(task.padEnd(12))} ${formatBytes(bytes)} ${dimText('SHA256 verified')}`)
+    },
+    onError(task, error) {
+      if (isTTY) process.stdout.write('\r' + ' '.repeat(80) + '\r')
+      console.log(`  ${errorText(task.padEnd(12))} ${error}`)
+    },
+  })
+
+  console.log()
+
+  const succeeded = results.filter(r => r.success)
+  if (succeeded.length === 0) {
+    console.log(warningText('No models were downloaded.'))
+    console.log()
+    return
+  }
+
+  // Build manifest from downloaded results + existing
+  const manifestPath = join(modelsDir, 'manifest.json')
+  const manifestModels: Partial<Record<ModelTask, ModelManifestEntry>> = {}
+
+  // Preserve existing manifest entries
+  if (existsSync(manifestPath)) {
+    try {
+      const existing: ModelManifest = JSON.parse(readFileSync(manifestPath, 'utf-8'))
+      if (existing.models) Object.assign(manifestModels, existing.models)
+    } catch { /* overwrite corrupt */ }
+  }
+
+  const installedTasks: ModelTask[] = []
+  let totalBytes = 0
+
+  for (const result of succeeded) {
+    const task = result.task as ModelTask
+    const entry = manifest.models[task]
+    if (!entry) continue
+
+    manifestModels[task] = {
+      version: entry.version,
+      file: entry.file,
+      sha256: entry.sha256,
+      params: entry.params,
+      accuracy: entry.accuracy ?? undefined,
+      labels: entry.labels,
+      maxSeqLen: entry.maxSeqLen,
+    }
+    installedTasks.push(task)
+    totalBytes += result.bytes
+  }
+
+  // Write manifest
+  const localManifest: ModelManifest = { models: manifestModels }
+  writeFileSync(manifestPath, JSON.stringify(localManifest, null, 2))
+
+  // Auto-enable models in config
+  const config = loadConfigFile()
+  if (!config.slm) config.slm = {}
+  config.slm.enabled = true
+  if (!config.slm.tasks) config.slm.tasks = {}
+  for (const task of installedTasks) {
+    config.slm.tasks[task] = 'model'
+  }
+  saveConfigFile(config)
+  updateConfigYml(installedTasks, 'model')
+
+  console.log(successText(`Downloaded ${succeeded.length} model${succeeded.length !== 1 ? 's' : ''} (${formatBytes(totalBytes)})`))
+  console.log(successText(`Auto-enabled ${installedTasks.join(', ')} in config`))
+  console.log(dimText(`Manifest written to ${manifestPath}`))
+  console.log()
+}
+
+// ─── enable ───────────────────────────────────────────────────────
+
+function enableTask(taskArg: string) {
+  const tasks = resolveTaskArg(taskArg, 'enable')
+  if (!tasks) return
+
+  console.log()
+  console.log(heading('Enable SLM Models'))
+  console.log()
+
+  const config = loadConfigFile()
+
+  // Ensure slm section exists
+  if (!config.slm) config.slm = {}
+  config.slm.enabled = true
+  if (!config.slm.tasks) config.slm.tasks = {}
+
+  for (const task of tasks) {
+    config.slm.tasks[task] = 'model'
+  }
+
+  saveConfigFile(config)
+  updateConfigYml(tasks, 'model')
+
+  for (const task of tasks) {
+    console.log(successText(`${task} -> model`))
+  }
+  console.log()
+  console.log(successText('SLM enabled'))
+  console.log(dimText('  Changes take effect on next server restart'))
+  console.log()
+}
+
+// ─── disable ──────────────────────────────────────────────────────
+
+function disableTask(taskArg: string) {
+  const tasks = resolveTaskArg(taskArg, 'disable')
+  if (!tasks) return
+
+  console.log()
+  console.log(heading('Disable SLM Models'))
+  console.log()
+
+  const config = loadConfigFile()
+  if (!config.slm) config.slm = {}
+  if (!config.slm.tasks) config.slm.tasks = {}
+
+  for (const task of tasks) {
+    config.slm.tasks[task] = DISABLE_MODE[task]
+  }
+
+  // If all tasks are now regex/api, disable SLM globally
+  const allDisabled = ALL_TASKS.every(t => {
+    const mode = config.slm.tasks[t] || DISABLE_MODE[t]
+    return mode === 'regex' || mode === 'api'
+  })
+  if (allDisabled) {
+    config.slm.enabled = false
+  }
+
+  saveConfigFile(config)
+  updateConfigYml(tasks, 'disable')
+
+  for (const task of tasks) {
+    console.log(warningText(`${task} -> ${DISABLE_MODE[task]}`))
+  }
+  console.log()
+  if (allDisabled) {
+    console.log(warningText('All tasks disabled — SLM globally disabled'))
+  }
+  console.log(dimText('  Changes take effect on next server restart'))
+  console.log()
+}
+
+/** Resolve a task argument to a list of tasks, or null if invalid */
+function resolveTaskArg(taskArg: string, verb: string): ModelTask[] | null {
+  if (!taskArg) {
+    console.log(errorText('Missing task argument'))
+    console.log(dimText(`Usage: claude-brain models ${verb} <task|all>`))
+    console.log(dimText(`Tasks: ${ALL_TASKS.join(', ')}, all`))
+    return null
+  }
+
+  if (taskArg === 'all') return [...ALL_TASKS]
+
+  if (!ALL_TASKS.includes(taskArg as ModelTask)) {
+    console.log(errorText(`Invalid task: ${taskArg}`))
+    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}, all`))
+    return null
+  }
+
+  return [taskArg as ModelTask]
+}
+
+// ─── benchmark ────────────────────────────────────────────────────
+
+async function benchmarkTask(task: string) {
+  if (!task) {
+    console.log(errorText('Missing task argument'))
+    console.log(dimText(`Usage: claude-brain models benchmark <task>`))
+    console.log(dimText(`Tasks: ${ALL_TASKS.join(', ')}`))
+    return
+  }
+
+  if (!ALL_TASKS.includes(task as ModelTask)) {
+    console.log(errorText(`Invalid task: ${task}`))
+    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}`))
+    return
+  }
+
+  console.log()
+  console.log(heading(`Benchmark: ${task}`))
+  console.log()
+
+  // Look for test data
+  const home = getClaudeBrainHome()
+  const testDataPath = join(home, 'training', 'benchmarks', 'baseline', `${task}_test.jsonl`)
+
+  if (!existsSync(testDataPath)) {
+    console.log(warningText(`Test data not found: ${testDataPath}`))
+    console.log(dimText('Generate test data first with: claude-brain export-training ' + task))
+    return
+  }
+
+  // Load test examples
+  const lines = readFileSync(testDataPath, 'utf-8').trim().split('\n').filter(Boolean)
+  if (lines.length === 0) {
+    console.log(warningText('Test file is empty'))
+    return
+  }
+
+  console.log(dimText(`Loaded ${lines.length} test examples from ${testDataPath}`))
+  console.log()
+
+  // Run through regex classifier for intent task
+  if (task === 'intent') {
+    await benchmarkIntent(lines)
+  } else {
+    // For non-intent tasks, just report data availability
+    console.log(dimText(`Benchmark for "${task}" requires model inference (ONNX).`))
+    console.log(dimText(`Regex-only benchmarking is available for the intent task.`))
+    console.log()
+    console.log(dimText(`Test data: ${lines.length} examples ready`))
+  }
+}
+
+async function benchmarkIntent(lines: string[]) {
+  // Dynamic import to avoid circular deps
+  const { IntentClassifier } = await import('@/routing/intent-classifier')
+  const classifier = new IntentClassifier()
+
+  let correct = 0
+  const labelCounts: Record<string, { tp: number; fp: number; fn: number }> = {}
+
+  for (const line of lines) {
+    let example: { input: string; output: { label: string } }
+    try {
+      example = JSON.parse(line)
+    } catch {
+      continue
+    }
+
+    const expected = example.output?.label
+    if (!expected) continue
+
+    const result = classifier.classify(example.input)
+    const predicted = result.primary
+
+    // Initialize counters
+    if (!labelCounts[expected]) labelCounts[expected] = { tp: 0, fp: 0, fn: 0 }
+    if (!labelCounts[predicted]) labelCounts[predicted] = { tp: 0, fp: 0, fn: 0 }
+
+    if (predicted === expected) {
+      correct++
+      labelCounts[expected].tp++
+    } else {
+      labelCounts[expected].fn++
+      labelCounts[predicted].fp++
+    }
+  }
+
+  const total = lines.length
+  const accuracy = total > 0 ? correct / total : 0
+
+  console.log(`  ${theme.bold('Overall Accuracy:')} ${accuracy >= 0.7 ? successText(`${(accuracy * 100).toFixed(1)}%`) : warningText(`${(accuracy * 100).toFixed(1)}%`)} (${correct}/${total})`)
+  console.log()
+
+  // Per-class metrics
+  console.log(`  ${dimText('Label'.padEnd(20))} ${dimText('Prec'.padEnd(8))} ${dimText('Recall'.padEnd(8))} ${dimText('F1'.padEnd(8))} ${dimText('Support')}`)
+  console.log(`  ${dimText('-'.repeat(52))}`)
+
+  const sortedLabels = Object.keys(labelCounts).sort()
+  for (const label of sortedLabels) {
+    const { tp, fp, fn } = labelCounts[label]
+    const precision = tp + fp > 0 ? tp / (tp + fp) : 0
+    const recall = tp + fn > 0 ? tp / (tp + fn) : 0
+    const f1 = precision + recall > 0 ? (2 * precision * recall) / (precision + recall) : 0
+    const support = tp + fn
+
+    console.log(
+      `  ${theme.primary(label.padEnd(20))} ${(precision * 100).toFixed(1).padStart(5)}%  ${(recall * 100).toFixed(1).padStart(5)}%  ${(f1 * 100).toFixed(1).padStart(5)}%  ${String(support).padStart(5)}`
+    )
+  }
+  console.log()
+}
+
+// ─── stats ────────────────────────────────────────────────────────
+
+async function showStats() {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('Training Data Statistics'))
+  console.log()
+
+  const stats = getTrainingStats()
+  const items: Array<{ label: string; value: string; status?: 'success' | 'warning' | 'error' | 'info' }> = []
+
+  let grandTotal = 0
+  let grandVerified = 0
+
+  for (const task of ALL_TASKS) {
+    const { total, verified } = stats[task]
+    grandTotal += total
+    grandVerified += verified
+    items.push({
+      label: task,
+      value: `${total} total, ${verified} verified`,
+      status: total > 0 ? 'success' : 'warning'
+    })
+  }
+
+  items.push({
+    label: 'Total',
+    value: `${grandTotal} total, ${grandVerified} verified`,
+    status: grandTotal > 0 ? 'success' : 'warning'
+  })
+
+  console.log(summaryPanel('Training Data', items))
+  console.log()
+
+  // Check for model_feedback table
+  try {
+    const { Database } = await import('bun:sqlite') as any
+    const dbPath = join(getClaudeBrainHome(), 'data', 'memory.db')
+    if (existsSync(dbPath)) {
+      const db = new Database(dbPath, { readonly: true })
+      const hasTable = db.prepare("SELECT name FROM sqlite_master WHERE type='table' AND name='model_feedback'").get()
+      if (hasTable) {
+        const disagreements = (db.prepare('SELECT COUNT(*) as c FROM model_feedback WHERE model_label != regex_label').get() as any)?.c ?? 0
+        const totalFeedback = (db.prepare('SELECT COUNT(*) as c FROM model_feedback').get() as any)?.c ?? 0
+        console.log(`  ${dimText('Model Feedback:')} ${totalFeedback} entries, ${disagreements} disagreements`)
+        console.log()
+      }
+      db.close()
+    }
+  } catch {
+    // model_feedback table doesn't exist yet — that's fine
+  }
+}
+
+// ─── retrain ──────────────────────────────────────────────────────
+
+async function retrainModels(taskFilter: string, force: boolean) {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('Retrain Models'))
+  console.log()
+
+  // Validate task filter
+  if (taskFilter !== 'all' && !ALL_TASKS.includes(taskFilter as ModelTask)) {
+    console.log(errorText(`Invalid task: ${taskFilter}`))
+    console.log(dimText(`Valid tasks: ${ALL_TASKS.join(', ')}, all`))
+    return
+  }
+
+  // Build config from schema defaults + .claudebrainrc.json overrides
+  const userConfig = loadConfigFile()
+  const retrainCfg = userConfig?.slm?.retrain ?? {}
+
+  const config: RetrainConfig = {
+    minFeedbackCount: retrainCfg.minFeedbackCount ?? 100,
+    maxDisagreementRate: retrainCfg.maxDisagreementRate ?? 0.15,
+    pythonPath: retrainCfg.pythonPath ?? 'python3',
+    trainingDir: retrainCfg.trainingDir ?? '~/slm-training',
+    force,
+  }
+
+  console.log(`  ${dimText('Training dir:')} ${config.trainingDir}`)
+  console.log(`  ${dimText('Python path:')} ${config.pythonPath}`)
+  console.log(`  ${dimText('Min feedback:')} ${config.minFeedbackCount}`)
+  console.log(`  ${dimText('Max disagreement:')} ${(config.maxDisagreementRate * 100).toFixed(0)}%`)
+  if (force) console.log(`  ${warningText('Force mode enabled — skipping checks')}`)
+  console.log()
+
+  if (taskFilter === 'all') {
+    // Retrain all tasks that need it
+    const results = await retrainAll(config)
+
+    if (results.size === 0) {
+      console.log(dimText('  No tasks needed retraining.'))
+      console.log()
+      return
+    }
+
+    // Summary
+    console.log()
+    console.log(heading('Retrain Summary'))
+    console.log()
+    for (const [task, result] of results) {
+      if (result.success) {
+        const accStr = result.newAccuracy != null
+          ? `${(result.newAccuracy * 100).toFixed(1)}%`
+          : 'n/a'
+        const oldStr = result.oldAccuracy != null
+          ? ` (was ${(result.oldAccuracy * 100).toFixed(1)}%)`
+          : ''
+        console.log(`  ${successText(task.padEnd(12))} accuracy: ${accStr}${oldStr}  ${dimText(`${result.trainingDataCount} examples, ${(result.duration / 1000).toFixed(1)}s`)}`)
+      } else {
+        console.log(`  ${errorText(task.padEnd(12))} ${result.error}`)
+      }
+    }
+  } else {
+    // Single task
+    const task = taskFilter as ModelTask
+
+    if (!force) {
+      const check = shouldRetrain(task, config)
+      console.log(`  ${dimText('Feedback count:')} ${check.feedbackCount}`)
+      console.log(`  ${dimText('Disagreement rate:')} ${(check.disagreementRate * 100).toFixed(1)}%`)
+      console.log(`  ${dimText('Last retrain:')} ${check.lastRetrainDate ?? 'never'}`)
+      console.log(`  ${dimText('Needs retrain:')} ${check.needed ? 'yes' : 'no'} — ${check.reason}`)
+      console.log()
+
+      if (!check.needed) {
+        console.log(dimText('  Retrain not needed. Use --force to override.'))
+        console.log()
+        return
+      }
+    }
+
+    const result = await retrainTask(task, config)
+
+    console.log()
+    if (result.success) {
+      const accStr = result.newAccuracy != null
+        ? `${(result.newAccuracy * 100).toFixed(1)}%`
+        : 'n/a'
+      const oldStr = result.oldAccuracy != null
+        ? ` (was ${(result.oldAccuracy * 100).toFixed(1)}%)`
+        : ''
+      console.log(successText(`Retrain complete: accuracy ${accStr}${oldStr}`))
+      console.log(dimText(`  ${result.trainingDataCount} examples, ${(result.duration / 1000).toFixed(1)}s`))
+    } else {
+      console.log(errorText(`Retrain failed: ${result.error}`))
+    }
+  }
+
+  console.log()
+}
+
+// ─── help ─────────────────────────────────────────────────────────
+
+function printModelsHelp() {
+  console.log()
+  console.log(renderLogo())
+  console.log()
+  console.log(heading('SLM Model Management'))
+  console.log()
+
+  const subcommands = [
+    ['list', 'Show installed models and their status'],
+    ['status', 'Show inference routing and ONNX runtime status'],
+    ['download', 'Download models (--source local|hf, --task <task>|all)'],
+    ['enable <task|all>', 'Enable model inference for task(s)'],
+    ['disable <task|all>', 'Disable model inference for task(s)'],
+    ['benchmark <task>', 'Run accuracy benchmark on test data'],
+    ['stats', 'Show training data statistics'],
+    ['retrain [<task>|all]', 'Retrain models from feedback (--force)'],
+  ]
+
+  const lines = subcommands
+    .map(([cmd, desc]) => `  ${theme.primary(cmd!.padEnd(20))} ${dimText(desc!)}`)
+    .join('\n')
+
+  console.log(theme.bold('Usage:') + ' ' + dimText('claude-brain models <subcommand>'))
+  console.log()
+  console.log(theme.bold('Subcommands:'))
+  console.log(lines)
+  console.log()
+  console.log(theme.bold('Tasks:') + ' ' + dimText(ALL_TASKS.join(', ')))
+  console.log()
+  console.log(theme.bold('Examples:'))
+  console.log(`  ${dimText('claude-brain models list')}`)
+  console.log(`  ${dimText('claude-brain models status')}`)
+  console.log(`  ${dimText('claude-brain models download --source hf')}`)
+  console.log(`  ${dimText('claude-brain models download --source hf --task intent')}`)
+  console.log(`  ${dimText('claude-brain models download --source local')}`)
+  console.log(`  ${dimText('claude-brain models enable all')}`)
+  console.log(`  ${dimText('claude-brain models enable intent')}`)
+  console.log(`  ${dimText('claude-brain models disable pattern')}`)
+  console.log(`  ${dimText('claude-brain models benchmark intent')}`)
+  console.log(`  ${dimText('claude-brain models stats')}`)
+  console.log(`  ${dimText('claude-brain models retrain intent')}`)
+  console.log(`  ${dimText('claude-brain models retrain all --force')}`)
+  console.log()
+}
+
+// ─── helpers ──────────────────────────────────────────────────────
+
+const RC_FILE = '.claudebrainrc.json'
+
+function loadConfigFile(): Record<string, any> {
+  const rcPath = join(getClaudeBrainHome(), RC_FILE)
+  if (!existsSync(rcPath)) return {}
+  try {
+    return JSON.parse(readFileSync(rcPath, 'utf-8'))
+  } catch {
+    return {}
+  }
+}
+
+function saveConfigFile(config: Record<string, any>): void {
+  const rcPath = join(getClaudeBrainHome(), RC_FILE)
+  writeFileSync(rcPath, JSON.stringify(config, null, 2) + '\n', 'utf-8')
+}
+
+/** Best-effort update of config.yml task lines for user visibility */
+function updateConfigYml(tasks: ModelTask[], mode: string): void {
+  const ymlPath = join(getClaudeBrainHome(), 'config.yml')
+  if (!existsSync(ymlPath)) return
+
+  try {
+    let content = readFileSync(ymlPath, 'utf-8')
+
+    // Update each task line (e.g. "    intent: regex" -> "    intent: model")
+    for (const task of tasks) {
+      const newMode = mode === 'disable' ? DISABLE_MODE[task] : mode
+      const regex = new RegExp(`^(\\s*${task}:\\s*)\\S+`, 'm')
+      content = content.replace(regex, `$1${newMode}`)
+    }
+
+    // Update slm.enabled line if present
+    // Use a heuristic: find "enabled:" that appears after "slm:"
+    const slmIdx = content.indexOf('slm:')
+    if (slmIdx !== -1) {
+      const afterSlm = content.slice(slmIdx)
+      const enabledMatch = afterSlm.match(/^(\s+enabled:\s*)\S+/m)
+      if (enabledMatch) {
+        const config = loadConfigFile()
+        const slmEnabled = config.slm?.enabled ?? false
+        const newLine = `${enabledMatch[1]}${slmEnabled}`
+        content = content.slice(0, slmIdx) + afterSlm.replace(enabledMatch[0], newLine)
+      }
+    }
+
+    writeFileSync(ymlPath, content, 'utf-8')
+  } catch {
+    // Non-critical — .claudebrainrc.json is the authoritative config
+  }
+}
+
+function formatBytes(bytes: number): string {
+  if (bytes === 0) return '0 B'
+  const units = ['B', 'KB', 'MB', 'GB']
+  const i = Math.floor(Math.log(bytes) / Math.log(1024))
+  const val = bytes / Math.pow(1024, i)
+  return `${val.toFixed(i > 0 ? 1 : 0)} ${units[i]}`
+}