@goodandready/dsh-moa 0.2.10 → 0.2.11

package/lib/moa-runner.js

@@ -5,7 +5,7 @@
  * - Parallel fan-out to reference models with transient retry & quorum straggler mitigation
  * - Prompt caching aligned message structures
  * - Curator synthesis & antipatterns evaluation
- * - Aggregator fallback chain for resilience
+ * - Aggregator fallback chain for resilience & malformed output recovery
  * - Live token streaming for aggregator
  * - File promotion & Live Canvas sandbox preview
  */
@@ -20,311 +20,48 @@ import {
   cleanMoaWorkspaces,
 } from './file-workspace.js'
 import { estimateTokenCost } from './pricing.js'
-import { recordMoaRun } from './history.js'
+import { recordMoaRun, recordMoaRunAsync } from './history.js'
+import {
+  slotLabel,
+  cleanAdvisoryMessages,
+  isBroadPromptRequiringQuestions,
+  buildQuestionSynthesisPrompt,
+  buildCuratorSynthesisPrompt,
+  buildSynthesisPrompt,
+  SYSTEM_ROLE_PROPOSER,
+  ANTIPATTERNS_RUBRIC,
+} from './moa-prompts.js'
+import {
+  parseWinnerIndex,
+  parseRecommendedAssembler,
+  parseMoACommand,
+  stripOrSummarizeCode,
+  formatMoAResponse,
+} from './moa-parser.js'
+
+// Re-export prompt and parser functions for external consumers / backwards compatibility
+export {
+  slotLabel,
+  cleanAdvisoryMessages,
+  isBroadPromptRequiringQuestions,
+  buildQuestionSynthesisPrompt,
+  buildCuratorSynthesisPrompt,
+  buildSynthesisPrompt,
+  ANTIPATTERNS_RUBRIC,
+} from './moa-prompts.js'
+
+export {
+  parseWinnerIndex,
+  parseRecommendedAssembler,
+  parseMoACommand,
+  stripOrSummarizeCode,
+  formatMoAResponse,
+} from './moa-parser.js'
 
 export { estimateTokenCost } from './pricing.js'
 
 export const DEFAULT_PROMPT = 'Please propose an optimal, well-structured, production-ready solution with full code and explanations.'
-
-export const REFERENCE_SYSTEM_PROMPT = `You are an expert AI software architect and senior engineer acting as a candidate proposer in a Mixture of Agents (MoA) ensemble.
-Your task is to provide the highest-quality, robust, complete, production-ready solution to the user request.
-Write clean, modern, fully functional code without placeholders or shortcuts.
-When generating files for a project, explicitly specify file paths using fenced code blocks with file annotations, e.g.:
-\`\`\`html file="index.html"
-\`\`\`
-\`\`\`javascript file="script.js"
-\`\`\`
-\`\`\`css file="style.css"
-\`\`\``
-
-export const ANTIPATTERNS_RUBRIC = `### 🚫 Strict Antipatterns Evaluation Checklist
-Penalize and strictly downgrade candidates exhibiting any of the following flaws:
-1. 🚫 Lazy Code & Placeholders:
-   - Phrases like "// ... rest of code unchanged", "/* TODO: implement */", incomplete functions or stubs returning null/mock without notice.
-2. 🚫 Blind Mocking:
-   - Hardcoded dummy arrays instead of real dynamic logic, user input handling, or real API integration.
-3. 🚫 Silent Failures & Missing Error Handling:
-   - Missing try/catch around async calls, fetch, JSON.parse; lack of user-facing fallback or retry states.
-4. 🚫 AI Slop UI & Poor Ergonomics:
-   - Generic purple/cyan gradients on pure black backgrounds, blurry drop-shadows without borders, lack of typographic hierarchy, low contrast (e.g. light gray text on white).
-5. 🚫 Missing UI States:
-   - Lack of loading state (spinner/skeleton), error display with retry action, or empty state when no data exists.
-6. 🚫 Broken Layout & Mobile Incompatibility:
-   - Fixed pixel widths (e.g. width: 800px) overflowing small viewports; unscrollable modal dialogs.
-7. 🚫 Monolithic God Objects & Overengineering:
-   - Dumping 1000+ lines into a single unmaintainable file, or building 10+ abstraction layers for a 2-function task.
-8. 🚫 Context Amnesia & Regressions:
-   - Dropping or breaking previously functioning project features while adding new code.`
-
-export function slotLabel(slot) {
-  if (!slot) return 'unknown'
-  if (typeof slot === 'string') return slot
-  if (slot.provider && slot.model) return `${slot.provider}:${slot.model}`
-  return slot.model || slot.provider || 'unknown'
-}
-
-/**
- * Extracts and sanitizes conversation history for candidate models.
- * Robust against complex DSH message content types (strings, part arrays, objects, tool calls).
- */
-export function cleanAdvisoryMessages(messages = [], maxCharBudget = 24000) {
-  if (!Array.isArray(messages)) return []
-
-  const trimmed = []
-  for (const msg of messages) {
-    if (!msg || typeof msg !== 'object') continue
-    const role = msg.role
-    if (role !== 'user' && role !== 'assistant') continue
-
-    let text = ''
-    if (typeof msg.content === 'string') {
-      text = msg.content
-    } else if (Array.isArray(msg.content)) {
-      text = msg.content
-        .filter((part) => part && typeof part === 'object' && part.type === 'text' && typeof part.text === 'string')
-        .map((part) => part.text)
-        .join('\n')
-    } else if (msg.content && typeof msg.content === 'object') {
-      if (typeof msg.content.text === 'string') {
-        text = msg.content.text
-      } else if (typeof msg.content.content === 'string') {
-        text = msg.content.content
-      }
-    } else if (typeof msg.text === 'string') {
-      text = msg.text
-    }
-
-    text = text.trim()
-    if (!text) continue
-
-    if (text.length > maxCharBudget) {
-      const head = text.slice(0, Math.floor(maxCharBudget * 0.7))
-      const tail = text.slice(-Math.floor(maxCharBudget * 0.3))
-      text = `${head}\n... [trimmed ${text.length - maxCharBudget} characters] ...\n${tail}`
-    }
-
-    trimmed.push({ role, content: text })
-  }
-  return trimmed
-}
-
-export function isBroadPromptRequiringQuestions(userPrompt = '', messages = []) {
-  if (!userPrompt || typeof userPrompt !== 'string') return false
-  const p = userPrompt.trim()
-  const wordCount = p.split(/\s+/).length
-
-  if (/^(да|нет|1|2|3|4|ок|погнали|давай|yes|no)\b/i.test(p) && wordCount <= 5) {
-    return false
-  }
-
-  const hasRecentQuestion = messages.some((m) => {
-    const text = typeof m?.content === 'string' ? m.content : JSON.stringify(m?.content || '')
-    return text.includes('Уточнение требований') || text.includes('опросник') || text.includes('Вариант 1')
-  })
-  if (hasRecentQuestion) {
-    return false
-  }
-
-  const creationTriggers = [
-    'сделай', 'создай', 'напиши', 'разработай', 'придумай', 'реализуй',
-    'make', 'build', 'create', 'generate', 'develop',
-  ]
-  const startsWithCreation = creationTriggers.some((t) => p.toLowerCase().startsWith(t))
-
-  if (startsWithCreation && wordCount <= 18) {
-    return true
-  }
-
-  const vagueNouns = ['приложение', 'игру', 'сервис', 'сайт', 'лендинг', 'калькулятор', 'виджет', 'дашборд', 'app', 'game', 'tool', 'website']
-  if (vagueNouns.some((n) => p.toLowerCase().includes(n))) {
-    if (wordCount <= 12) return true
-  }
-
-  return false
-}
-
-export function buildQuestionSynthesisPrompt(userPrompt, referenceOutputs = []) {
-  const joined = referenceOutputs
-    .map((r, i) => `Advisor ${i + 1} (${r.label}):\n${r.text}`)
-    .join('\n\n')
-
-  return `You are the lead architect and judge in a Mixture of Agents (MoA) ensemble.
-The user gave the task:
-"${userPrompt}"
-
-The advisors proposed the following decision points and clarifications:
-${joined}
-
-Your task is to synthesize a single, compact, friendly and structured questionnaire (2-4 questions) in the same language as the user's prompt.
-Each question must offer 2-3 concrete recommended answer options (e.g.: 1. Format: single-file HTML/JS or React? 2. Style: minimalism, iOS or neubrutalism?).
-At the end, add a note that the user can answer briefly (e.g.: "1, 2, dark theme") or trust the defaults.`
-}
-
-export function buildCuratorSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '') {
-  const joined = referenceOutputs
-    .map((r, i) => {
-      const fileSummary = (r.files && r.files.length > 0)
-        ? ` [Files created: ${r.files.map((f) => f.relativePath).join(', ')}]`
-        : ''
-      let textContent = r.text
-      if (referenceOutputs.length >= 3 && textContent.length > 3000) {
-        textContent = stripOrSummarizeCode(textContent)
-      }
-      return `Candidate ${i + 1} — ${r.label}:${fileSummary}\n${textContent}`
-    })
-    .join('\n\n')
-
-  const criteriaBlock = judgeCriteria && judgeCriteria.trim()
-    ? `\n### 🎯 Additional evaluation criteria:\n${judgeCriteria.trim()}\n`
-    : ''
-
-  return `You are the expert Lead Technical Curator and Solution Architect in a Mixture of Agents (MoA) ensemble.
-Your mission is not merely to select one candidate, but to synthesize the optimal solution by extracting the finest components from each candidate's response, identifying potential flaws using the strict antipatterns rubric, and selecting/advising which single agent model is best suited to assemble the final unified deliverable.
-
-User Request:
-${userPrompt}
-${criteriaBlock}
-Candidate proposals:
-${joined}
-
-${ANTIPATTERNS_RUBRIC}
-
-Instructions:
-Your response MUST be structured into three clear parts (respond in the same language as the user's prompt, e.g. Russian):
-
-### 1. 🔍 Curator Analysis & Component Breakdown
-- For EACH candidate, provide:
-  - ⭐ **Strongest aspects** (e.g. robust architecture, superior UI/CSS design, clean data validation).
-  - ⚠️ **Defects or Antipatterns** found from the checklist above.
-- Highlight which candidate provides the best foundation for each component (e.g. Candidate 1 for core logic, Candidate 2 for visual UI).
-
-### 2. 🧩 Assembly Recipe & Recommended Master Assembler
-- Recommend the best single agent model to assemble and finalize the solution:
-  RECOMMENDED_ASSEMBLER: <number from 1 to N> (<provider:model>)
-- State the machine winner index marker for file promotion:
-  WINNER_CANDIDATE_INDEX: <number from 1 to N>
-- Provide the exact blueprint / instructions for combining the best pieces into a unified deliverable.
-
-### 3. 🚀 Unified Solution & Execution Guide
-- Present the final synthesized code or complete instructions combining the best candidate features.
-- How to run, verify, and use the deliverable.`
-}
-
-export function buildSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '', options = {}) {
-  if (options.curatorSynthesis) {
-    return buildCuratorSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria)
-  }
-
-  const joined = referenceOutputs
-    .map((r, i) => {
-      const fileSummary = (r.files && r.files.length > 0)
-        ? ` [Files created: ${r.files.map((f) => f.relativePath).join(', ')}]`
-        : ''
-      let textContent = r.text
-      if (referenceOutputs.length >= 3 && textContent.length > 3000) {
-        textContent = stripOrSummarizeCode(textContent)
-      }
-      return `Reference ${i + 1} — ${r.label}:${fileSummary}\n${textContent}`
-    })
-    .join('\n\n')
-
-  const criteriaBlock = judgeCriteria && judgeCriteria.trim()
-    ? `\n### 🎯 Additional evaluation criteria from the user:\n${judgeCriteria.trim()}\n`
-    : ''
-
-  return `You are the expert aggregator/judge in a Mixture of Agents (MoA) process. You evaluate solutions from multiple candidate models, judge which one is best (or how to combine their best parts), and deliver the final authoritative verdict and solution.
-
-Original user prompt:
-${userPrompt}
-${criteriaBlock}
-Reference responses from candidate models:
-${joined}
-
-${ANTIPATTERNS_RUBRIC}
-
-Instructions:
-Your response MUST be structured into three clear parts (respond in the same language as the user's prompt, e.g. Russian):
-
-### 1. ⚖️ Judge verdict and comparative analysis
-- **Winner**: clearly name the model and candidate number (e.g. "Winner: Candidate 1 (opencode-go:deepseek-v4-flash)" or "Reference 1 (label) is chosen").
-- Always add the machine winner-selection marker:
-  WINNER_CANDIDATE_INDEX: <number from 1 to N>
-- **Why this choice**: compare code, architecture, strengths, weaknesses and reliability of all candidates in detail.
-
-### 2. 📁 Project files created
-- List the winner files promoted to the project root and their purpose.
-
-### 3. 🚀 How to run and use
-- Describe how to open and run the created project.`
-}
-
-export function parseWinnerIndex(judgeText, defaultIndex = 1, candidateCount = Infinity) {
-  if (!judgeText || typeof judgeText !== 'string') return defaultIndex
-  const inRange = (idx) => !isNaN(idx) && idx >= 1 && idx <= candidateCount
-  const match = /WINNER_CANDIDATE_INDEX:\s*(\d+)/i.exec(judgeText)
-  if (match) {
-    const idx = parseInt(match[1], 10)
-    if (inRange(idx)) return idx
-  }
-  const candMatch = /(?:Кандидат|Candidate|Reference)\s*(\d+)\b/i.exec(judgeText)
-  if (candMatch) {
-    const idx = parseInt(candMatch[1], 10)
-    if (inRange(idx)) return idx
-  }
-  return defaultIndex
-}
-
-export function parseRecommendedAssembler(judgeText, defaultIndex = 1, candidateCount = Infinity) {
-  if (!judgeText || typeof judgeText !== 'string') return { index: defaultIndex, label: '' }
-  const inRange = (idx) => !isNaN(idx) && idx >= 1 && idx <= candidateCount
-  const match = /RECOMMENDED_ASSEMBLER:\s*(\d+)(?:\s*\(([^)]+)\))?/i.exec(judgeText)
-  if (match) {
-    const idx = parseInt(match[1], 10)
-    if (inRange(idx)) {
-      return { index: idx, label: match[2]?.trim() || '' }
-    }
-  }
-  return { index: defaultIndex, label: '' }
-}
-
-export function parseMoACommand(text, presets = []) {
-  if (typeof text !== 'string' || !text.startsWith('/moa')) {
-    return null
-  }
-
-  const remainder = text.slice(4).trim()
-  if (!remainder) {
-    return {
-      presetName: 'default',
-      prompt: '',
-    }
-  }
-
-  const matchPreset = /^--preset(?:=|\s+)([a-zA-Z0-9_-]+)\s*(.*)/s.exec(remainder)
-  if (matchPreset) {
-    return {
-      presetName: matchPreset[1],
-      prompt: matchPreset[2] || '',
-    }
-  }
-
-  const parts = remainder.split(/\s+/)
-  const firstWord = parts[0]
-  if (Array.isArray(presets)) {
-    const matched = presets.find((p) => p.name === firstWord)
-    if (matched) {
-      return {
-        presetName: matched.name,
-        prompt: parts.slice(1).join(' ').trim(),
-      }
-    }
-  }
-
-  return {
-    presetName: 'default',
-    prompt: remainder,
-  }
-}
+export const REFERENCE_SYSTEM_PROMPT = SYSTEM_ROLE_PROPOSER
 
 /**
  * Invokes LLM call with transient retry for recoverable network/rate-limit errors.
@@ -357,9 +94,9 @@ export async function runReferencesParallel(references, messages, options = {},
 
   const advisoryMessages = cleanAdvisoryMessages(messages)
   const systemPrompt = options.systemPrompt || REFERENCE_SYSTEM_PROMPT
-  // Deterministic prefix for optimal prompt caching hit rate (#2.3)
+  // Deterministic prefix for optimal prompt caching hit rate
   const fullMessages = [{ role: 'system', content: systemPrompt }, ...advisoryMessages]
-  const timeoutMs = options.timeoutMs ?? 75000
+  const timeoutMs = options.timeoutMs ?? ((options.timeoutSec ?? 60) * 1000)
   const maxRetries = options.maxRetries ?? 0
   const quorumEnabled = Boolean(options.quorumEnabled)
   const gracePeriodMs = (options.gracePeriodSec ?? 10) * 1000
@@ -391,14 +128,16 @@ export async function runReferencesParallel(references, messages, options = {},
             messages: fullMessages,
             temperature: options.temperature ?? 0.6,
             maxTokens: options.maxTokens ?? 4096,
+            timeoutMs,
             signal: abortControllers[i].signal,
           },
           maxRetries
         )
 
         const timeoutPromise = new Promise((_, reject) => {
-          const timer = setTimeout(() => reject(new Error(`Timeout after ${timeoutMs / 1000}s`)), timeoutMs)
+          const timer = setTimeout(() => reject(new Error(`Timeout after ${Math.round(timeoutMs / 1000)}s`)), timeoutMs)
           timer.unref?.()
+          abortControllers[i].signal.addEventListener('abort', () => clearTimeout(timer), { once: true })
         })
 
         const res = await Promise.race([callPromise, timeoutPromise])
@@ -412,7 +151,7 @@ export async function runReferencesParallel(references, messages, options = {},
 
         finishedCount++
         if (typeof onProgress === 'function') {
-          const costStr = costInfo.costUsd > 0 ? ` (~\${costInfo.costUsd.toFixed(4)})` : ''
+          const costStr = costInfo.costUsd > 0 ? ` (~\$${costInfo.costUsd.toFixed(4)})` : ''
           onProgress(`✅ *Candidate ${i + 1}/${total} (${label}) finished${costStr}*\n`)
         }
 
@@ -450,22 +189,22 @@ export async function runReferencesParallel(references, messages, options = {},
     runOne()
   })
 
-  // Quorum straggler mitigation: exit as soon as grace period expires
+  // Straggler mitigation via Quorum + Grace Period
   if (quorumEnabled && total >= 3) {
     const quorumTarget = Math.max(2, Math.min(total - 1, Math.ceil(total * 0.6)))
     let graceTimer = null
-    let quorumTriggered = false
 
     await new Promise((resolve) => {
       const checkQuorum = () => {
         if (finishedCount >= total) {
           if (graceTimer) clearTimeout(graceTimer)
-          return resolve()
+          resolve()
+          return
         }
-        if (finishedCount >= quorumTarget && !quorumTriggered) {
-          quorumTriggered = true
+
+        if (finishedCount >= quorumTarget && !graceTimer) {
           if (typeof onProgress === 'function') {
-            onProgress(`⏳ *Quorum reached (${finishedCount}/${total}). Grace period ${gracePeriodMs / 1000}s for remaining models...*\n`)
+            onProgress(`⏳ *Quorum reached (${finishedCount}/${total}). Starting ${Math.round(gracePeriodMs / 1000)}s grace period for stragglers...*\n`)
           }
           graceTimer = setTimeout(() => {
             if (typeof onProgress === 'function' && finishedCount < total) {
@@ -478,7 +217,7 @@ export async function runReferencesParallel(references, messages, options = {},
             }
             resolve()
           }, gracePeriodMs)
-          // active timer keeps event loop alive
+          graceTimer.unref?.()
         }
       }
 
@@ -548,7 +287,7 @@ async function createPromotedPreview(liveCanvas, cwd, promotedFiles, referenceOu
 }
 
 /**
- * Main MoA pipeline runner with Curator Synthesis, Fallback Chains, and Live Token Streaming.
+ * Executes the full Mixture of Agents pipeline.
  */
 export async function runMoAPipeline({
   userPrompt,
@@ -556,57 +295,58 @@ export async function runMoAPipeline({
   preset,
   callLlm,
   cwd = process.cwd(),
-  onProgress,
-  onStreamDelta,
   prices = {},
-  historyFilePath,
+  onProgress = null,
+  historyFilePath = null,
   liveCanvas = null,
+  onStreamDelta = null,
 }) {
   const startTime = Date.now()
-  const referenceModels = preset?.reference_models || [
-    { provider: 'opencode-go', model: 'deepseek-v4-flash' },
-    { provider: 'grok', model: 'grok-build-0.1' },
-  ]
-  const primaryJudge = preset?.aggregator || { provider: 'codex', model: 'gpt-5.6-sol' }
+
+  // 1. Resolve configurations
+  const referenceModels = Array.isArray(preset?.reference_models) && preset.reference_models.length > 0
+    ? preset.reference_models
+    : [{ provider: 'opencode-go', model: 'deepseek-v4-flash' }]
+
+  const primaryJudge = preset?.aggregator?.provider && preset?.aggregator?.model
+    ? preset.aggregator
+    : { provider: 'codex', model: 'gpt-5.6-sol' }
+
   const fallbackJudges = Array.isArray(preset?.aggregator_fallbacks) ? preset.aggregator_fallbacks : []
   const judgesChain = [primaryJudge, ...fallbackJudges]
 
-  const refTemp = preset?.reference_temperature ?? 0.6
-  const aggTemp = preset?.aggregator_temperature ?? 0.4
-  const maxTokens = preset?.max_tokens ?? 4096
-  const judgeCriteria = preset?.judge_criteria ?? ''
-  const isFastMode = referenceModels.length === 1
+  const refTemp = typeof preset?.reference_temperature === 'number' ? preset.reference_temperature : 0.6
+  const aggTemp = typeof preset?.aggregator_temperature === 'number' ? preset.aggregator_temperature : 0.4
+  const maxTokens = typeof preset?.max_tokens === 'number' ? preset.max_tokens : 4096
+  const judgeCriteria = preset?.judge_criteria || ''
+  const isFastMode = referenceModels.length === 1 && !preset?.curator_synthesis
   const isCuratorSynthesis = Boolean(preset?.curator_synthesis)
   const isStreamAggregator = preset?.stream_aggregator !== false
   const isQuorumEnabled = Boolean(preset?.quorum_enabled)
-  const gracePeriodSec = preset?.grace_period_sec ?? 10
-  const candidateRetries = preset?.retry_count ?? 1
-
-  // 1. Collect current project workspace context
-  if (typeof onProgress === 'function') {
-    onProgress('🔍 *Scanning project context...*\n')
-  }
-  const projectContext = await collectProjectContext(cwd)
-  const isRefinement = isRefinementTask(userPrompt, projectContext.files)
-
-  // 2. Check if broad prompt requires user questionnaire
-  const askQuestions = preset?.ask_clarifying_questions !== false
-  const needsQuestions = !isFastMode && askQuestions && !isRefinement && isBroadPromptRequiringQuestions(userPrompt, messages)
-
-  // 3. Build enriched prompt for candidates
-  let candidateSystemPrompt = REFERENCE_SYSTEM_PROMPT
-  if (isRefinement && projectContext.files.length > 0) {
-    const fileList = projectContext.files.map((f) => `- \`${f.relativePath}\` (${f.content.length} chars)`).join('\n')
-    const fileContents = projectContext.files.map((f) => `### File: ${f.relativePath}\n\`\`\`\n${f.content}\n\`\`\``).join('\n\n')
-    candidateSystemPrompt += `\n\nExisting project structure:\n${fileList}\n\nProject files:\n${fileContents}\n\nYou are modifying an existing project. Output modified or new files with explicit file paths.`
+  const gracePeriodSec = typeof preset?.grace_period_sec === 'number' ? preset.grace_period_sec : 10
+  const candidateRetries = 1
+  const refTimeoutSec = typeof preset?.reference_timeout_sec === 'number' ? preset.reference_timeout_sec : 60
+  const aggTimeoutSec = typeof preset?.aggregator_timeout_sec === 'number' ? preset.aggregator_timeout_sec : 180
+
+  // 2. Collect project context for refinement tasks
+  const isRefinement = isRefinementTask(userPrompt)
+  let projectContext = ''
+  if (isRefinement) {
+    if (typeof onProgress === 'function') {
+      onProgress('🔍 *Reading project files for refinement context...*\n')
+    }
+    projectContext = await collectProjectContext(cwd, 16000)
   }
 
-  let promptForCandidates = userPrompt
-  if (needsQuestions) {
-    promptForCandidates += "\n(Note: the task is broad. Propose the key architectural and functional decision points to clarify the user's requirements.)"
-  }
+  // 3. Build prompts & evaluate broad questionnaire needs
+  const candidateSystemPrompt = projectContext
+    ? `${REFERENCE_SYSTEM_PROMPT}\n\n### Current Project Files & Context:\n${projectContext}`
+    : REFERENCE_SYSTEM_PROMPT
 
-  const enrichedMessages = [...messages, { role: 'user', content: promptForCandidates }]
+  const askClarifyingQuestions = preset?.ask_clarifying_questions !== false
+  const needsQuestions = askClarifyingQuestions && isBroadPromptRequiringQuestions(userPrompt, messages)
+
+  const enrichedMessages = [...messages, { role: 'user', content: userPrompt }]
 
   // 4. Parallel fan-out to candidate models with quorum & transient retry
   const referenceOutputs = await runReferencesParallel(
@@ -617,6 +357,7 @@ export async function runMoAPipeline({
       temperature: refTemp,
       maxTokens,
       prices,
+      timeoutSec: refTimeoutSec,
       quorumEnabled: isQuorumEnabled,
       gracePeriodSec,
       maxRetries: candidateRetries,
@@ -676,21 +417,27 @@ export async function runMoAPipeline({
         messages: [{ role: 'user', content: questionPrompt }],
         temperature: 0.3,
         maxTokens: 2048,
+        timeoutMs: aggTimeoutSec * 1000,
       }, 1)
       questionsContent = typeof qRes === 'string' ? qRes : (qRes?.content || qRes?.text || '')
-      const qFallback = (typeof qRes === 'object' && qRes?.usage) ? qRes.usage : {
+      const qFallbackUsage = (typeof qRes === 'object' && qRes?.usage) ? qRes.usage : {
         inputTokens: Math.max(1, Math.round(questionPrompt.length / 4)),
         outputTokens: Math.max(1, Math.round(questionsContent.length / 4)),
       }
-      qUsage = estimateTokenCost(primaryJudge, qFallback, prices)
-    } catch {
-      questionsContent = '### Project requirements clarification\nPlease specify the implementation details and the desired stack.'
+      qUsage = estimateTokenCost(primaryJudge, qFallbackUsage, prices)
+    } catch (err) {
+      console.warn('[dsh-moa] Questionnaire synthesis failed, proceeding with fallback questions:', err)
+      questionsContent = `### Уточнение требований по задаче: "${userPrompt}"\n\n` +
+        '1. **Формат решения**: однофайловый HTML/JS или многомодульный проект?\n' +
+        '2. **Стиль и визуальное оформление**: минимализм, темная тема или нейтральный интерфейс?\n' +
+        '3. **Функциональные приоритеты**: базовый MVP или расширенная реализация?\n\n' +
+        '*Ответьте кратко (например, "1, 2") или доверьтесь выбору по умолчанию.*'
     }
 
     await cleanMoaWorkspaces(cwd)
-
-    const qTokens = referenceOutputs.reduce((acc, r) => acc + (r.usage?.totalTokens || 0), 0) + qUsage.totalTokens
-    const qCostUsd = Number((referenceOutputs.reduce((acc, r) => acc + (r.costUsd || 0), 0) + qUsage.costUsd).toFixed(5))
+    const totalTokens = referenceOutputs.reduce((acc, r) => acc + (r.usage?.totalTokens || 0), 0) + qUsage.totalTokens
+    const totalCostUsd = Number((referenceOutputs.reduce((acc, r) => acc + (r.costUsd || 0), 0) + qUsage.costUsd).toFixed(5))
+    const durationMs = Date.now() - startTime
 
     try {
       recordMoaRun({
@@ -698,41 +445,51 @@ export async function runMoAPipeline({
         preset: preset?.name || 'default',
         isRefinement,
         candidates: candidatesForHistory(referenceOutputs),
-        aggregator: null,
+        aggregator: { provider: primaryJudge.provider, model: primaryJudge.model, usage: qUsage, costUsd: qUsage.costUsd },
         winnerIndex: -1,
-        winnerModel: '',
+        winnerModel: 'none',
         promotedFiles: [],
-        totalTokens: qTokens,
-        totalCostUsd: qCostUsd,
-        durationMs: Date.now() - startTime,
-      }, historyFilePath)
-    } catch (histErr) {
-      console.warn('[dsh-moa] Failed to record questionnaire run in history:', histErr)
-    }
+        totalTokens,
+        totalCostUsd,
+        durationMs,
+      }, historyFilePath || undefined)
+    } catch {}
 
     return {
       kind: 'questions',
       content: questionsContent,
+      aggregator: slotLabel(primaryJudge),
       references: referenceOutputs,
       presetName: preset?.name || 'default',
-      isRefinement: false,
+      isRefinement,
+      isFastMode,
+      winningIndex: 0,
+      winnerModel: 'none',
+      promotedFiles: [],
+      usage: {
+        totalTokens,
+        totalCostUsd,
+        candidates: referenceOutputs.map((r) => ({ label: r.label, usage: r.usage, costUsd: r.costUsd })),
+        aggregator: qUsage,
+      },
+      durationMs,
     }
   }
 
-  // 7. Fast mode bypass for single candidate
+  // 7. Fast Mode (1 candidate, bypass judge)
   if (isFastMode) {
-    const single = referenceOutputs[0]
+    const single = successfulRefs[0]
     let promotedFiles = []
-    if (single.files && single.files.length > 0) {
-      promotedFiles = await promoteCandidateWorkspace(cwd, 1)
+    if (single.files?.length > 0) {
+      promotedFiles = await promoteCandidateWorkspace(cwd, single.index)
     } else {
       await cleanMoaWorkspaces(cwd)
     }
 
+    const livePreview = await createPromotedPreview(liveCanvas, cwd, promotedFiles, referenceOutputs)
     const totalTokens = single.usage?.totalTokens || 0
-    const totalCostUsd = single.costUsd || 0
+    const totalCostUsd = Number((single.costUsd || 0).toFixed(5))
     const durationMs = Date.now() - startTime
-    const livePreview = await createPromotedPreview(liveCanvas, cwd, promotedFiles, referenceOutputs)
 
     try {
       recordMoaRun({
@@ -742,16 +499,14 @@ export async function runMoAPipeline({
         isFastMode: true,
         candidates: candidatesForHistory(referenceOutputs),
         aggregator: null,
-        winnerIndex: 1,
+        winnerIndex: single.index,
         winnerModel: single.label,
         promotedFiles,
         totalTokens,
         totalCostUsd,
         durationMs,
-      }, historyFilePath)
-    } catch (histErr) {
-      console.warn('[dsh-moa] Failed to record fast-mode run in history:', histErr)
-    }
+      }, historyFilePath || undefined)
+    } catch {}
 
     return {
       kind: 'synthesis',
@@ -761,28 +516,22 @@ export async function runMoAPipeline({
       presetName: preset?.name || 'default',
       isRefinement,
       isFastMode: true,
-      winningIndex: 1,
+      winningIndex: single.index,
       winnerModel: single.label,
       promotedFiles,
       ...(livePreview ? { liveCanvas: livePreview } : {}),
       usage: {
         totalTokens,
         totalCostUsd,
-        candidates: [{ label: single.label, usage: single.usage, costUsd: single.costUsd }],
+        candidates: referenceOutputs.map((r) => ({ label: r.label, usage: r.usage, costUsd: r.costUsd })),
         aggregator: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 },
       },
       durationMs,
     }
   }
 
-  // 8. Aggregator / Judge synthesis with fallback chain (#3.3) & streaming (#2.1)
-  const synthesisPrompt = buildSynthesisPrompt(
-    userPrompt,
-    referenceOutputs,
-    judgeCriteria,
-    { curatorSynthesis: isCuratorSynthesis }
-  )
-
+  // 8. Synthesis phase via primary judge or fallback chain
+  const synthesisPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, { curatorSynthesis: isCuratorSynthesis })
   let synthesizedText = ''
   let aggUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 }
   let chosenJudge = primaryJudge
@@ -794,9 +543,9 @@ export async function runMoAPipeline({
     const currentLabel = slotLabel(currentJudge)
 
     if (typeof onProgress === 'function') {
-      const judgeTitle = isCuratorSynthesis ? 'Lead Curator' : 'Judge'
+      const modeTitle = isCuratorSynthesis ? 'Curator' : 'Judge'
       const fallbackBadge = jIdx > 0 ? ` (Fallback #${jIdx})` : ''
-      onProgress(`⚖️ *${judgeTitle} (${currentLabel})${fallbackBadge} evaluates candidates and synthesizes the solution...*\n`)
+      onProgress(`⚖️ *${modeTitle} (${currentLabel}${fallbackBadge}) synthesizing solution...*\n`)
     }
 
     try {
@@ -806,6 +555,7 @@ export async function runMoAPipeline({
         messages: [{ role: 'user', content: synthesisPrompt }],
         temperature: aggTemp,
         maxTokens,
+        timeoutMs: aggTimeoutSec * 1000,
         onStreamDelta: (delta) => {
           if (isStreamAggregator && typeof onStreamDelta === 'function') {
             onStreamDelta(delta)
@@ -814,6 +564,7 @@ export async function runMoAPipeline({
       }, 0)
 
       synthesizedText = typeof aggRes === 'string' ? aggRes : (aggRes?.content || aggRes?.text || '')
+
       const aggFallbackUsage = (typeof aggRes === 'object' && aggRes?.usage) ? aggRes.usage : {
         inputTokens: Math.max(1, Math.round(synthesisPrompt.length / 4)),
         outputTokens: Math.max(1, Math.round(synthesizedText.length / 4)),
@@ -885,7 +636,7 @@ export async function runMoAPipeline({
       totalTokens,
       totalCostUsd,
       durationMs,
-    }, historyFilePath)
+    }, historyFilePath || undefined)
   } catch (histErr) {
     console.warn('[dsh-moa] Failed to record run in history:', histErr)
   }
@@ -915,103 +666,8 @@ export async function runMoAPipeline({
 }
 
 /**
- * Replaces large code blocks with concise file/code summaries to avoid token waste and huge chat dumps.
+ * Streams a full MoA turn into chat with live aggregator tokens and progress feedback.
  */
-export function stripOrSummarizeCode(text) {
-  if (!text || typeof text !== 'string') return ''
-  return text.replace(/```([a-zA-Z0-9_\-\.\/]*)\s*([\w\.\/\-]+\.[a-zA-Z0-9]+)?\n([\s\S]*?)```/g, (match, lang, fileTag, code) => {
-    const lines = code.trim().split('\n')
-    if (lines.length <= 3 && !/html|jsx|tsx|vue|svelte|css|js|ts/i.test(lang)) {
-      return match
-    }
-    const fileHint = fileTag || (code.match(/^\s*(?:\/\/|#|<!--|\/\*)\s*(?:file|filepath|path):\s*([^\s*]+)/im)?.[1])
-    const label = fileHint ? `file \`${fileHint}\`` : (lang ? `code \`${lang}\`` : 'code')
-    return `\n> 📄 *[${label} - ${lines.length} lines saved to disk]*\n`
-  })
-}
-
-export function formatMoAResponse({ moaResult, presetName }) {
-  const parts = []
-  const pName = presetName || moaResult?.presetName || 'default'
-  const judge = moaResult?.aggregator || 'unknown'
-  const refs = moaResult?.references || []
-
-  if (moaResult?.kind === 'questions') {
-    parts.push('## 🧠 Mixture of Agents — Requirements Clarification')
-    parts.push(`*Judge (${judge}) and the advisors analyzed the task:*\n`)
-    parts.push(moaResult.content)
-    return parts.join('\n')
-  }
-
-  if (moaResult?.kind === 'failure') {
-    parts.push('## ⚠️ Mixture of Agents — Execution Failed')
-    parts.push(moaResult.content)
-    return parts.join('\n')
-  }
-
-  const modeBadge = moaResult?.isFastMode
-    ? '⚡ Fast Mode'
-    : (moaResult?.isCuratorSynthesis ? `🧠 Curator: ${judge}` : `Judge: ${judge}`)
-  parts.push(`## 🧠 Mixture of Agents (Preset: ${pName} | ${modeBadge})`)
-  parts.push('')
-
-  if (moaResult?.isRefinement) {
-    parts.push('> 🔄 **Mode**: Iterative project refinement (Refinement)')
-  }
-
-  const hasPromoted = moaResult?.promotedFiles && moaResult.promotedFiles.length > 0
-  if (hasPromoted) {
-    parts.push(`> 📦 **Files created in the project**: \`${moaResult.promotedFiles.join('`, `')}\``)
-  }
-
-  if (moaResult?.recommendedAssembler?.label) {
-    parts.push(`> 🎯 **Recommended Master Assembler**: Candidate ${moaResult.recommendedAssembler.index} (\`${moaResult.recommendedAssembler.label}\`)`)
-  }
-
-  if (moaResult?.liveCanvas?.previewUrl) {
-    parts.push(`> 🎨 **Live Canvas**: [🚀 Открыть ${moaResult.liveCanvas.title || 'превью'} в Live Canvas](${moaResult.liveCanvas.previewUrl}) | [↗ Открыть в новой вкладке](${moaResult.liveCanvas.previewUrl})`)
-  }
-
-  // Cost tracking card
-  if (moaResult?.usage) {
-    const u = moaResult.usage
-    const costStr = u.totalCostUsd > 0 ? `~\$${u.totalCostUsd.toFixed(4)}` : 'Free'
-    const tokStr = u.totalTokens >= 1000 ? `${(u.totalTokens / 1000).toFixed(1)}k` : `${u.totalTokens}`
-    parts.push(`> 💰 **Run cost**: ${costStr} (${tokStr} tokens total)`)
-  }
-
-  parts.push('')
-
-  if (!moaResult?.isFastMode) {
-    parts.push(`### ⚖️ Judge verdict and final synthesis (Synthesis: ${judge})`)
-    parts.push('')
-    const cleanJudgeContent = hasPromoted
-      ? stripOrSummarizeCode(moaResult?.content || '')
-      : (moaResult?.content || '(нет ответа)')
-    parts.push(cleanJudgeContent)
-    parts.push('')
-  }
-
-  if (Array.isArray(refs) && refs.length > 0) {
-    const title = moaResult?.isFastMode ? '### 🚀 Candidate generation result:' : `### 👥 Advisor responses (${refs.length}):`
-    parts.push(title)
-    parts.push('')
-    refs.forEach((ref, i) => {
-      const statusIcon = ref.ok ? '✅' : '⚠️'
-      const fileBadge = ref.files?.length ? ` (${ref.files.length} файл(ов))` : ''
-      const costBadge = ref.costUsd > 0 ? ` [~\$${ref.costUsd.toFixed(4)}]` : ''
-      parts.push(`#### ${statusIcon} Model ${i + 1}: ${ref.label}${fileBadge}${costBadge}`)
-      parts.push('')
-      parts.push(stripOrSummarizeCode(ref.text))
-      parts.push('')
-      parts.push('---')
-      parts.push('')
-    })
-  }
-
-  return parts.join('\n')
-}
-
 export async function* streamMoATurn({ targetPreset, userPrompt, messages, callLlm, cwd, prices, historyFilePath, liveCanvas }, options = {}) {
   const signal = options?.signal
   if (signal?.aborted) return
@@ -1116,9 +772,6 @@ export async function* streamMoATurn({ targetPreset, userPrompt, messages, callL
     }
     yield { type: 'finish', reason: { kind: 'stop' } }
   } finally {
-    if (signal?.aborted) {
-      await cleanMoaWorkspaces(cwd)
-    }
+    // cleanup
   }
 }
-