claude-brain 0.30.2 → 0.30.3

package/src/training/retrain-pipeline.ts

@@ -1,394 +1,399 @@
-/**
- * Retrain Pipeline — SLM Upgrade Phase 6B
- * Automated retraining when enough feedback accumulates.
- *
- * Flow:
- *   1. Check if retraining is needed (feedback count + disagreement rate)
- *   2. Export merged training data (existing + feedback corrections) as JSONL
- *   3. Shell out to Python: prepare_data → train → evaluate → export_onnx
- *   4. Compare new accuracy to old, promote model if improved
- *   5. Update retrain-state.json with timestamp + stats
- */
-
-import { execSync } from 'node:child_process'
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'
-import { join } from 'node:path'
-import { homedir } from 'node:os'
-import { getClaudeBrainHome, getHomePaths } from '@/config/home'
-import {
-  exportTrainingData,
-  exportModelFeedback,
-  getModelFeedbackStats,
-  type TrainingTask,
-} from '@/training/data-store'
-import type { ModelManifest, ModelTask } from '@/intelligence/model-manager'
-
-// ── Types ────────────────────────────────────────────────────────────
-
-export interface RetrainConfig {
-  minFeedbackCount: number
-  maxDisagreementRate: number
-  pythonPath: string
-  trainingDir: string
-  force: boolean
-}
-
-export interface RetrainCheck {
-  needed: boolean
-  reason: string
-  feedbackCount: number
-  disagreementRate: number
-  lastRetrainDate: string | null
-}
-
-export interface RetrainResult {
-  task: string
-  success: boolean
-  error?: string
-  oldAccuracy?: number
-  newAccuracy?: number
-  trainingDataCount: number
-  duration: number
-}
-
-interface RetrainStateEntry {
-  lastRetrain: string
-  lastAccuracy: number
-  feedbackAtRetrain: number
-}
-
-type RetrainState = Record<string, RetrainStateEntry>
-
-// ── Helpers ──────────────────────────────────────────────────────────
-
-function resolveTrainingDir(trainingDir: string): string {
-  if (trainingDir.startsWith('~')) {
-    return join(homedir(), trainingDir.slice(1))
-  }
-  return trainingDir
-}
-
-function getRetrainStatePath(): string {
-  return join(getClaudeBrainHome(), 'data', 'retrain-state.json')
-}
-
-function loadRetrainState(): RetrainState {
-  const statePath = getRetrainStatePath()
-  if (!existsSync(statePath)) return {}
-  try {
-    return JSON.parse(readFileSync(statePath, 'utf-8'))
-  } catch {
-    return {}
-  }
-}
-
-function saveRetrainState(state: RetrainState): void {
-  const statePath = getRetrainStatePath()
-  const dir = join(getClaudeBrainHome(), 'data')
-  if (!existsSync(dir)) {
-    mkdirSync(dir, { recursive: true })
-  }
-  writeFileSync(statePath, JSON.stringify(state, null, 2))
-}
-
-function loadManifest(): ModelManifest | null {
-  const manifestPath = join(getHomePaths().models, 'manifest.json')
-  if (!existsSync(manifestPath)) return null
-  try {
-    return JSON.parse(readFileSync(manifestPath, 'utf-8'))
-  } catch {
-    return null
-  }
-}
-
-function saveManifest(manifest: ModelManifest): void {
-  const modelsDir = getHomePaths().models
-  if (!existsSync(modelsDir)) {
-    mkdirSync(modelsDir, { recursive: true })
-  }
-  writeFileSync(join(modelsDir, 'manifest.json'), JSON.stringify(manifest, null, 2))
-}
-
-// ── Public API ───────────────────────────────────────────────────────
-
-/**
- * Check if retraining is needed for a task based on feedback accumulation.
- */
-export function shouldRetrain(task: ModelTask, config: RetrainConfig): RetrainCheck {
-  const state = loadRetrainState()
-  const taskState = state[task]
-  const lastRetrainDate = taskState?.lastRetrain ?? null
-  const feedbackAtLastRetrain = taskState?.feedbackAtRetrain ?? 0
-
-  const feedbackStats = getModelFeedbackStats()
-  const taskStats = feedbackStats[task]
-
-  if (!taskStats) {
-    return {
-      needed: false,
-      reason: 'No feedback data available',
-      feedbackCount: 0,
-      disagreementRate: 0,
-      lastRetrainDate,
-    }
-  }
-
-  const newFeedbackCount = taskStats.total - feedbackAtLastRetrain
-  const disagreementRate = taskStats.disagreementRate
-
-  // Check if enough new feedback has accumulated
-  if (newFeedbackCount < config.minFeedbackCount) {
-    return {
-      needed: false,
-      reason: `Only ${newFeedbackCount} new feedback entries (need ${config.minFeedbackCount})`,
-      feedbackCount: taskStats.total,
-      disagreementRate,
-      lastRetrainDate,
-    }
-  }
-
-  // Check if disagreement rate exceeds threshold
-  if (disagreementRate > config.maxDisagreementRate) {
-    return {
-      needed: true,
-      reason: `Disagreement rate ${(disagreementRate * 100).toFixed(1)}% exceeds threshold ${(config.maxDisagreementRate * 100).toFixed(1)}%`,
-      feedbackCount: taskStats.total,
-      disagreementRate,
-      lastRetrainDate,
-    }
-  }
-
-  // Enough feedback accumulated even if disagreement rate is below threshold
-  return {
-    needed: true,
-    reason: `${newFeedbackCount} new feedback entries since last retrain`,
-    feedbackCount: taskStats.total,
-    disagreementRate,
-    lastRetrainDate,
-  }
-}
-
-/**
- * Export merged training data (training_data + feedback corrections) as JSONL.
- */
-export function exportMergedTrainingData(
-  task: ModelTask,
-  trainingDir: string
-): { path: string; count: number } {
-  const resolvedDir = resolveTrainingDir(trainingDir)
-  const dataDir = join(resolvedDir, 'data', task)
-  if (!existsSync(dataDir)) {
-    mkdirSync(dataDir, { recursive: true })
-  }
-
-  const outputPath = join(dataDir, 'train.jsonl')
-
-  // Get existing training data
-  const trainingLines = exportTrainingData(task as TrainingTask)
-
-  // Get feedback data (disagreements are useful training signal)
-  const feedbackLines = exportModelFeedback(task)
-
-  // Merge: training data first, then feedback
-  const allLines = [...trainingLines, ...feedbackLines]
-
-  writeFileSync(outputPath, allLines.join('\n') + '\n')
-
-  return { path: outputPath, count: allLines.length }
-}
-
-/**
- * Run the full retrain pipeline for a single task.
- */
-export async function retrainTask(
-  task: ModelTask,
-  config: RetrainConfig
-): Promise<RetrainResult> {
-  const startTime = Date.now()
-  const resolvedDir = resolveTrainingDir(config.trainingDir)
-
-  // Step 1: Export merged training data
-  console.log(`  [${task}] Exporting merged training data...`)
-  let dataCount: number
-  try {
-    const { count } = exportMergedTrainingData(task, config.trainingDir)
-    dataCount = count
-    console.log(`  [${task}] Exported ${count} training examples`)
-  } catch (err) {
-    return {
-      task,
-      success: false,
-      error: `Failed to export training data: ${err instanceof Error ? err.message : String(err)}`,
-      trainingDataCount: 0,
-      duration: Date.now() - startTime,
-    }
-  }
-
-  if (dataCount === 0) {
-    return {
-      task,
-      success: false,
-      error: 'No training data available',
-      trainingDataCount: 0,
-      duration: Date.now() - startTime,
-    }
-  }
-
-  // Step 2: Run Python training pipeline
-  const pythonSteps = [
-    { script: 'scripts/prepare_data.py', args: `--task ${task}`, label: 'Preparing data' },
-    { script: 'scripts/train.py', args: `--task ${task}`, label: 'Training model' },
-    { script: 'scripts/evaluate.py', args: `--task ${task} --save`, label: 'Evaluating model' },
-    { script: 'scripts/export_onnx.py', args: `--task ${task} --output-dir models/`, label: 'Exporting ONNX' },
-  ]
-
-  for (const step of pythonSteps) {
-    console.log(`  [${task}] ${step.label}...`)
-    const cmd = `${config.pythonPath} ${step.script} ${step.args}`
-    try {
-      execSync(cmd, {
-        cwd: resolvedDir,
-        stdio: 'pipe',
-        timeout: 600_000, // 10 minute timeout per step
-      })
-    } catch (err) {
-      const stderr = err instanceof Error && 'stderr' in err
-        ? (err as any).stderr?.toString().slice(0, 500)
-        : String(err)
-      return {
-        task,
-        success: false,
-        error: `Python step "${step.label}" failed: ${stderr}`,
-        trainingDataCount: dataCount,
-        duration: Date.now() - startTime,
-      }
-    }
-  }
-
-  // Step 3: Read evaluation results
-  const benchmarkPath = join(resolvedDir, 'benchmarks', `${task}.json`)
-  let newAccuracy: number | undefined
-  try {
-    if (existsSync(benchmarkPath)) {
-      const benchData = JSON.parse(readFileSync(benchmarkPath, 'utf-8'))
-      newAccuracy = benchData.accuracy ?? benchData.overall_accuracy
-    }
-  } catch {
-    console.log(`  [${task}] Warning: Could not read benchmark results`)
-  }
-
-  // Step 4: Compare with existing model accuracy
-  const manifest = loadManifest()
-  const oldAccuracy = manifest?.models[task]?.accuracy
-
-  if (newAccuracy != null && oldAccuracy != null && newAccuracy < oldAccuracy && !config.force) {
-    console.log(`  [${task}] New accuracy ${(newAccuracy * 100).toFixed(1)}% < old ${(oldAccuracy * 100).toFixed(1)}% — skipping model replacement`)
-    // Still record the retrain attempt
-    const state = loadRetrainState()
-    const feedbackStats = getModelFeedbackStats()
-    state[task] = {
-      lastRetrain: new Date().toISOString(),
-      lastAccuracy: oldAccuracy,
-      feedbackAtRetrain: feedbackStats[task]?.total ?? 0,
-    }
-    saveRetrainState(state)
-
-    return {
-      task,
-      success: false,
-      error: `New accuracy (${(newAccuracy * 100).toFixed(1)}%) lower than current (${(oldAccuracy * 100).toFixed(1)}%). Use --force to override.`,
-      oldAccuracy,
-      newAccuracy,
-      trainingDataCount: dataCount,
-      duration: Date.now() - startTime,
-    }
-  }
-
-  // Step 5: Copy new ONNX model to ~/.claude-brain/models/
-  const modelsDir = getHomePaths().models
-  if (!existsSync(modelsDir)) {
-    mkdirSync(modelsDir, { recursive: true })
-  }
-
-  const onnxSource = join(resolvedDir, 'models', `${task}.onnx`)
-  if (existsSync(onnxSource)) {
-    const onnxDest = join(modelsDir, `${task}.onnx`)
-    const onnxData = readFileSync(onnxSource)
-    writeFileSync(onnxDest, onnxData)
-    console.log(`  [${task}] Model copied to ${onnxDest}`)
-
-    // Update manifest
-    const currentManifest = loadManifest() ?? { models: {} }
-    const existing = currentManifest.models[task]
-    currentManifest.models[task] = {
-      version: bumpVersion(existing?.version),
-      file: `${task}.onnx`,
-      accuracy: newAccuracy,
-      labels: existing?.labels,
-      maxSeqLen: existing?.maxSeqLen,
-    }
-    saveManifest(currentManifest)
-    console.log(`  [${task}] Manifest updated`)
-  } else {
-    console.log(`  [${task}] Warning: ONNX file not found at ${onnxSource}`)
-  }
-
-  // Step 6: Update retrain state
-  const state = loadRetrainState()
-  const feedbackStats = getModelFeedbackStats()
-  state[task] = {
-    lastRetrain: new Date().toISOString(),
-    lastAccuracy: newAccuracy ?? oldAccuracy ?? 0,
-    feedbackAtRetrain: feedbackStats[task]?.total ?? 0,
-  }
-  saveRetrainState(state)
-
-  console.log(`  [${task}] Retrain complete in ${((Date.now() - startTime) / 1000).toFixed(1)}s`)
-
-  return {
-    task,
-    success: true,
-    oldAccuracy,
-    newAccuracy,
-    trainingDataCount: dataCount,
-    duration: Date.now() - startTime,
-  }
-}
-
-/**
- * Run retrain for all tasks that need it.
- */
-export async function retrainAll(
-  config: RetrainConfig
-): Promise<Map<string, RetrainResult>> {
-  const ALL_TASKS: ModelTask[] = ['intent', 'entity', 'query', 'knowledge', 'compress', 'pattern']
-  const results = new Map<string, RetrainResult>()
-
-  for (const task of ALL_TASKS) {
-    const check = shouldRetrain(task, config)
-
-    if (!check.needed && !config.force) {
-      console.log(`  [${task}] Skipping — ${check.reason}`)
-      continue
-    }
-
-    if (config.force && !check.needed) {
-      console.log(`  [${task}] Forcing retrain (${check.reason})`)
-    }
-
-    const result = await retrainTask(task, config)
-    results.set(task, result)
-  }
-
-  return results
-}
-
-// ── Internal Helpers ─────────────────────────────────────────────────
-
-function bumpVersion(current?: string): string {
-  if (!current) return '1.0.0'
-  const parts = current.split('.').map(Number)
-  // Bump patch version
-  parts[2] = (parts[2] ?? 0) + 1
-  return parts.join('.')
-}
+/**
+ * Retrain Pipeline — SLM Upgrade Phase 6B
+ * Automated retraining when enough feedback accumulates.
+ *
+ * Flow:
+ *   1. Check if retraining is needed (feedback count + disagreement rate)
+ *   2. Export merged training data (existing + feedback corrections) as JSONL
+ *   3. Shell out to Python: prepare_data → train → evaluate → export_onnx
+ *   4. Compare new accuracy to old, promote model if improved
+ *   5. Update retrain-state.json with timestamp + stats
+ */
+
+import { spawnSync } from 'node:child_process'
+import { existsSync, mkdirSync, readFileSync, writeFileSync, realpathSync } from 'node:fs'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+import { getClaudeBrainHome, getHomePaths } from '@/config/home'
+import {
+  exportTrainingData,
+  exportModelFeedback,
+  getModelFeedbackStats,
+  type TrainingTask,
+} from '@/training/data-store'
+import type { ModelManifest, ModelTask } from '@/intelligence/model-manager'
+
+// ── Types ────────────────────────────────────────────────────────────
+
+export interface RetrainConfig {
+  minFeedbackCount: number
+  maxDisagreementRate: number
+  pythonPath: string
+  trainingDir: string
+  force: boolean
+}
+
+export interface RetrainCheck {
+  needed: boolean
+  reason: string
+  feedbackCount: number
+  disagreementRate: number
+  lastRetrainDate: string | null
+}
+
+export interface RetrainResult {
+  task: string
+  success: boolean
+  error?: string
+  oldAccuracy?: number
+  newAccuracy?: number
+  trainingDataCount: number
+  duration: number
+}
+
+interface RetrainStateEntry {
+  lastRetrain: string
+  lastAccuracy: number
+  feedbackAtRetrain: number
+}
+
+type RetrainState = Record<string, RetrainStateEntry>
+
+// ── Helpers ──────────────────────────────────────────────────────────
+
+function resolveTrainingDir(trainingDir: string): string {
+  if (trainingDir.startsWith('~')) {
+    return join(homedir(), trainingDir.slice(1))
+  }
+  return trainingDir
+}
+
+function getRetrainStatePath(): string {
+  return join(getClaudeBrainHome(), 'data', 'retrain-state.json')
+}
+
+function loadRetrainState(): RetrainState {
+  const statePath = getRetrainStatePath()
+  if (!existsSync(statePath)) return {}
+  try {
+    return JSON.parse(readFileSync(statePath, 'utf-8'))
+  } catch {
+    return {}
+  }
+}
+
+function saveRetrainState(state: RetrainState): void {
+  const statePath = getRetrainStatePath()
+  const dir = join(getClaudeBrainHome(), 'data')
+  if (!existsSync(dir)) {
+    mkdirSync(dir, { recursive: true })
+  }
+  writeFileSync(statePath, JSON.stringify(state, null, 2))
+}
+
+function loadManifest(): ModelManifest | null {
+  const manifestPath = join(getHomePaths().models, 'manifest.json')
+  if (!existsSync(manifestPath)) return null
+  try {
+    return JSON.parse(readFileSync(manifestPath, 'utf-8'))
+  } catch {
+    return null
+  }
+}
+
+function saveManifest(manifest: ModelManifest): void {
+  const modelsDir = getHomePaths().models
+  if (!existsSync(modelsDir)) {
+    mkdirSync(modelsDir, { recursive: true })
+  }
+  writeFileSync(join(modelsDir, 'manifest.json'), JSON.stringify(manifest, null, 2))
+}
+
+// ── Public API ───────────────────────────────────────────────────────
+
+/**
+ * Check if retraining is needed for a task based on feedback accumulation.
+ */
+export function shouldRetrain(task: ModelTask, config: RetrainConfig): RetrainCheck {
+  const state = loadRetrainState()
+  const taskState = state[task]
+  const lastRetrainDate = taskState?.lastRetrain ?? null
+  const feedbackAtLastRetrain = taskState?.feedbackAtRetrain ?? 0
+
+  const feedbackStats = getModelFeedbackStats()
+  const taskStats = feedbackStats[task]
+
+  if (!taskStats) {
+    return {
+      needed: false,
+      reason: 'No feedback data available',
+      feedbackCount: 0,
+      disagreementRate: 0,
+      lastRetrainDate,
+    }
+  }
+
+  const newFeedbackCount = taskStats.total - feedbackAtLastRetrain
+  const disagreementRate = taskStats.disagreementRate
+
+  // Check if enough new feedback has accumulated
+  if (newFeedbackCount < config.minFeedbackCount) {
+    return {
+      needed: false,
+      reason: `Only ${newFeedbackCount} new feedback entries (need ${config.minFeedbackCount})`,
+      feedbackCount: taskStats.total,
+      disagreementRate,
+      lastRetrainDate,
+    }
+  }
+
+  // Check if disagreement rate exceeds threshold
+  if (disagreementRate > config.maxDisagreementRate) {
+    return {
+      needed: true,
+      reason: `Disagreement rate ${(disagreementRate * 100).toFixed(1)}% exceeds threshold ${(config.maxDisagreementRate * 100).toFixed(1)}%`,
+      feedbackCount: taskStats.total,
+      disagreementRate,
+      lastRetrainDate,
+    }
+  }
+
+  // Enough feedback accumulated even if disagreement rate is below threshold
+  return {
+    needed: true,
+    reason: `${newFeedbackCount} new feedback entries since last retrain`,
+    feedbackCount: taskStats.total,
+    disagreementRate,
+    lastRetrainDate,
+  }
+}
+
+/**
+ * Export merged training data (training_data + feedback corrections) as JSONL.
+ */
+export function exportMergedTrainingData(
+  task: ModelTask,
+  trainingDir: string
+): { path: string; count: number } {
+  const resolvedDir = resolveTrainingDir(trainingDir)
+  const dataDir = join(resolvedDir, 'data', task)
+  if (!existsSync(dataDir)) {
+    mkdirSync(dataDir, { recursive: true })
+  }
+
+  const outputPath = join(dataDir, 'train.jsonl')
+
+  // Get existing training data
+  const trainingLines = exportTrainingData(task as TrainingTask)
+
+  // Get feedback data (disagreements are useful training signal)
+  const feedbackLines = exportModelFeedback(task)
+
+  // Merge: training data first, then feedback
+  const allLines = [...trainingLines, ...feedbackLines]
+
+  writeFileSync(outputPath, allLines.join('\n') + '\n')
+
+  return { path: outputPath, count: allLines.length }
+}
+
+/**
+ * Run the full retrain pipeline for a single task.
+ */
+export async function retrainTask(
+  task: ModelTask,
+  config: RetrainConfig
+): Promise<RetrainResult> {
+  const startTime = Date.now()
+  const resolvedDir = resolveTrainingDir(config.trainingDir)
+
+  // Step 1: Export merged training data
+  console.log(`  [${task}] Exporting merged training data...`)
+  let dataCount: number
+  try {
+    const { count } = exportMergedTrainingData(task, config.trainingDir)
+    dataCount = count
+    console.log(`  [${task}] Exported ${count} training examples`)
+  } catch (err) {
+    return {
+      task,
+      success: false,
+      error: `Failed to export training data: ${err instanceof Error ? err.message : String(err)}`,
+      trainingDataCount: 0,
+      duration: Date.now() - startTime,
+    }
+  }
+
+  if (dataCount === 0) {
+    return {
+      task,
+      success: false,
+      error: 'No training data available',
+      trainingDataCount: 0,
+      duration: Date.now() - startTime,
+    }
+  }
+
+  // Step 2: Run Python training pipeline
+  const pythonSteps = [
+    { script: 'scripts/prepare_data.py', args: `--task ${task}`, label: 'Preparing data' },
+    { script: 'scripts/train.py', args: `--task ${task}`, label: 'Training model' },
+    { script: 'scripts/evaluate.py', args: `--task ${task} --save`, label: 'Evaluating model' },
+    { script: 'scripts/export_onnx.py', args: `--task ${task} --output-dir models/`, label: 'Exporting ONNX' },
+  ]
+
+  let resolvedCwd: string
+  try {
+    resolvedCwd = realpathSync(resolvedDir)
+  } catch {
+    resolvedCwd = resolvedDir
+  }
+
+  for (const step of pythonSteps) {
+    console.log(`  [${task}] ${step.label}...`)
+    const args = [step.script, ...step.args.split(' ')]
+    const result = spawnSync(config.pythonPath, args, {
+      cwd: resolvedCwd,
+      encoding: 'utf-8',
+      stdio: 'pipe',
+      timeout: 600_000, // 10 minute timeout per step
+    })
+    if (result.error || result.status !== 0) {
+      const stderr = result.stderr?.slice(0, 500) || result.error?.message || 'Unknown error'
+      return {
+        task,
+        success: false,
+        error: `Python step "${step.label}" failed: ${stderr}`,
+        trainingDataCount: dataCount,
+        duration: Date.now() - startTime,
+      }
+    }
+  }
+
+  // Step 3: Read evaluation results
+  const benchmarkPath = join(resolvedDir, 'benchmarks', `${task}.json`)
+  let newAccuracy: number | undefined
+  try {
+    if (existsSync(benchmarkPath)) {
+      const benchData = JSON.parse(readFileSync(benchmarkPath, 'utf-8'))
+      newAccuracy = benchData.accuracy ?? benchData.overall_accuracy
+    }
+  } catch {
+    console.log(`  [${task}] Warning: Could not read benchmark results`)
+  }
+
+  // Step 4: Compare with existing model accuracy
+  const manifest = loadManifest()
+  const oldAccuracy = manifest?.models[task]?.accuracy
+
+  if (newAccuracy != null && oldAccuracy != null && newAccuracy < oldAccuracy && !config.force) {
+    console.log(`  [${task}] New accuracy ${(newAccuracy * 100).toFixed(1)}% < old ${(oldAccuracy * 100).toFixed(1)}% — skipping model replacement`)
+    // Still record the retrain attempt
+    const state = loadRetrainState()
+    const feedbackStats = getModelFeedbackStats()
+    state[task] = {
+      lastRetrain: new Date().toISOString(),
+      lastAccuracy: oldAccuracy,
+      feedbackAtRetrain: feedbackStats[task]?.total ?? 0,
+    }
+    saveRetrainState(state)
+
+    return {
+      task,
+      success: false,
+      error: `New accuracy (${(newAccuracy * 100).toFixed(1)}%) lower than current (${(oldAccuracy * 100).toFixed(1)}%). Use --force to override.`,
+      oldAccuracy,
+      newAccuracy,
+      trainingDataCount: dataCount,
+      duration: Date.now() - startTime,
+    }
+  }
+
+  // Step 5: Copy new ONNX model to ~/.claude-brain/models/
+  const modelsDir = getHomePaths().models
+  if (!existsSync(modelsDir)) {
+    mkdirSync(modelsDir, { recursive: true })
+  }
+
+  const onnxSource = join(resolvedDir, 'models', `${task}.onnx`)
+  if (existsSync(onnxSource)) {
+    const onnxDest = join(modelsDir, `${task}.onnx`)
+    const onnxData = readFileSync(onnxSource)
+    writeFileSync(onnxDest, onnxData)
+    console.log(`  [${task}] Model copied to ${onnxDest}`)
+
+    // Update manifest
+    const currentManifest = loadManifest() ?? { models: {} }
+    const existing = currentManifest.models[task]
+    currentManifest.models[task] = {
+      version: bumpVersion(existing?.version),
+      file: `${task}.onnx`,
+      accuracy: newAccuracy,
+      labels: existing?.labels,
+      maxSeqLen: existing?.maxSeqLen,
+    }
+    saveManifest(currentManifest)
+    console.log(`  [${task}] Manifest updated`)
+  } else {
+    console.log(`  [${task}] Warning: ONNX file not found at ${onnxSource}`)
+  }
+
+  // Step 6: Update retrain state
+  const state = loadRetrainState()
+  const feedbackStats = getModelFeedbackStats()
+  state[task] = {
+    lastRetrain: new Date().toISOString(),
+    lastAccuracy: newAccuracy ?? oldAccuracy ?? 0,
+    feedbackAtRetrain: feedbackStats[task]?.total ?? 0,
+  }
+  saveRetrainState(state)
+
+  console.log(`  [${task}] Retrain complete in ${((Date.now() - startTime) / 1000).toFixed(1)}s`)
+
+  return {
+    task,
+    success: true,
+    oldAccuracy,
+    newAccuracy,
+    trainingDataCount: dataCount,
+    duration: Date.now() - startTime,
+  }
+}
+
+/**
+ * Run retrain for all tasks that need it.
+ */
+export async function retrainAll(
+  config: RetrainConfig
+): Promise<Map<string, RetrainResult>> {
+  const ALL_TASKS: ModelTask[] = ['intent', 'entity', 'query', 'knowledge', 'compress', 'pattern']
+  const results = new Map<string, RetrainResult>()
+
+  for (const task of ALL_TASKS) {
+    const check = shouldRetrain(task, config)
+
+    if (!check.needed && !config.force) {
+      console.log(`  [${task}] Skipping — ${check.reason}`)
+      continue
+    }
+
+    if (config.force && !check.needed) {
+      console.log(`  [${task}] Forcing retrain (${check.reason})`)
+    }
+
+    const result = await retrainTask(task, config)
+    results.set(task, result)
+  }
+
+  return results
+}
+
+// ── Internal Helpers ─────────────────────────────────────────────────
+
+function bumpVersion(current?: string): string {
+  if (!current) return '1.0.0'
+  const parts = current.split('.').map(Number)
+  // Bump patch version
+  parts[2] = (parts[2] ?? 0) + 1
+  return parts.join('.')
+}