@goodandready/dsh-moa 0.2.2 → 0.2.4

package/lib/moa-runner.js

@@ -1,6 +1,7 @@
 /**
- * Mixture of Agents (MoA) execution engine with Interactive Questioning
- * and Isolated Multi-Candidate File Execution.
+ * Mixture of Agents (MoA) execution engine with Interactive Questioning,
+ * Isolated Multi-Candidate File Execution, Native Cost Tracking,
+ * Refinement Mode, and Multi-Candidate Scalability.
  */
 
 import {
@@ -8,13 +9,19 @@ import {
   writeCandidateWorkspace,
   promoteCandidateWorkspace,
   cleanMoaWorkspaces,
+  collectProjectContext,
+  isRefinementTask,
 } from './file-workspace.js'
 
+import { recordMoaRun } from './history.js'
+
 export {
   extractFileBlocks,
   writeCandidateWorkspace,
   promoteCandidateWorkspace,
   cleanMoaWorkspaces,
+  collectProjectContext,
+  isRefinementTask,
 }
 
 export const REFERENCE_SYSTEM_PROMPT = `You are an expert candidate engineer model in a Mixture of Agents (MoA) architecture.
@@ -34,12 +41,56 @@ RULES:
 4. If the user request is broad, make opinionated, high-quality technical decisions and deliver a complete working project.
 5. Never leave placeholders like "// TODO" or "... rest of code". Deliver exhaustive, working code.`
 
+export const REFINEMENT_SYSTEM_PROMPT = `You are an expert software engineer performing an iterative modification / refinement on an existing project.
+You are provided with the current codebase files and the user's delta request.
+
+RULES:
+1. Modify the existing files or create new files to fulfill the user's change request.
+2. Deliver complete, production-ready updated code for every modified file.
+3. Format each updated file in markdown code blocks with explicit file attribute:
+   \`\`\`javascript file="src/app.js"
+   ...
+   \`\`\`
+4. Keep the existing architecture, styling conventions and dependencies consistent.`
+
 export const ADVISOR_QUESTION_PROMPT = `You are a senior technical advisor in a Mixture of Agents (MoA) architecture.
 The user has provided a prompt that may have multiple design choices, architectural paths, or underspecified requirements.
 
 Review the user prompt and identify 1 to 3 critical, high-impact clarifying questions or architectural options that would define the implementation (e.g. framework/vanilla, features, design style, target environment).
 Keep questions very clear, structured, and actionable. Avoid trivial questions. Respond directly in Russian.`
 
+/**
+ * Built-in pricing table for major LLM model families ($ per 1M tokens) (#10).
+ */
+export const MODEL_PRICING_REGISTRY = [
+  { pattern: /deepseek.*flash|v4.*flash/i, inputPerM: 0.14, outputPerM: 0.28 },
+  { pattern: /deepseek.*reasoner|r1/i, inputPerM: 0.55, outputPerM: 2.19 },
+  { pattern: /deepseek/i, inputPerM: 0.27, outputPerM: 1.10 },
+  { pattern: /gpt-4o-mini|gpt-5.*mini/i, inputPerM: 0.15, outputPerM: 0.60 },
+  { pattern: /gpt-5|gpt-4o|sol/i, inputPerM: 2.50, outputPerM: 10.00 },
+  { pattern: /grok/i, inputPerM: 2.00, outputPerM: 10.00 },
+  { pattern: /claude-3-5-sonnet|claude-3-7-sonnet/i, inputPerM: 3.00, outputPerM: 15.00 },
+  { pattern: /claude-3-5-haiku/i, inputPerM: 0.80, outputPerM: 4.00 },
+  { pattern: /claude/i, inputPerM: 3.00, outputPerM: 15.00 },
+  { pattern: /qwen/i, inputPerM: 0.40, outputPerM: 1.20 },
+  { pattern: /commandcode/i, inputPerM: 0.50, outputPerM: 1.50 },
+  { pattern: /.*/, inputPerM: 0.50, outputPerM: 1.50 },
+]
+
+export function estimateTokenCost(slot, usage = {}) {
+  const modelStr = `${slot?.provider || ''} ${slot?.model || ''}`.trim()
+  const match = MODEL_PRICING_REGISTRY.find((p) => p.pattern.test(modelStr)) || MODEL_PRICING_REGISTRY[MODEL_PRICING_REGISTRY.length - 1]
+  const inTokens = usage.inputTokens || usage.promptTokens || usage.input_tokens || 0
+  const outTokens = usage.outputTokens || usage.completionTokens || usage.output_tokens || 0
+  const cost = (inTokens * match.inputPerM / 1_000_000) + (outTokens * match.outputPerM / 1_000_000)
+  return {
+    inputTokens: inTokens,
+    outputTokens: outTokens,
+    totalTokens: inTokens + outTokens,
+    costUsd: Number(cost.toFixed(5)),
+  }
+}
+
 export function slotLabel(slot) {
   if (!slot || typeof slot !== 'object') return 'unknown'
   const prov = String(slot.provider || '').trim()
@@ -79,44 +130,38 @@ export function cleanAdvisoryMessages(messages = [], maxCharBudget = 4000) {
 
     trimmed.push({ role, content: text })
   }
-
   return trimmed
 }
 
-/**
- * Checks if the prompt should trigger a clarifying questions phase
- * instead of immediate execution.
- */
 export function isBroadPromptRequiringQuestions(userPrompt = '', messages = []) {
-  const prompt = (userPrompt || '').trim()
-  if (!prompt) return false
-
-  // Find the last assistant message in history
-  const lastAssistant = [...messages].reverse().find((m) => m && m.role === 'assistant')
-  const lastAssistantText = typeof lastAssistant?.content === 'string'
-    ? lastAssistant.content
-    : (Array.isArray(lastAssistant?.content)
-        ? lastAssistant.content.map((c) => (typeof c === 'string' ? c : c.text || '')).join('')
-        : '')
-
-  // If the last assistant message was an MoA questionnaire, the user is answering it now!
-  const isAnsweringQuestions =
-    lastAssistantText.includes('Уточнение требований') ||
-    lastAssistantText.includes('Уточните, пожалуйста') ||
-    lastAssistantText.includes('опросник')
-
-  if (isAnsweringQuestions) {
+  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
   }
 
-  // Trigger on Russian broad verbs
-  const lower = prompt.toLowerCase()
-  const broadKeywords = ['создай', 'сделай', 'разработай', 'напиши', 'построй', 'сгенерируй']
-  const hasBroadKeyword = broadKeywords.some((k) => lower.includes(k))
+  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
+  }
 
-  // If it contains a broad keyword and is relatively brief (<= 12 words), ask questions!
-  if (hasBroadKeyword) {
-    const wordCount = prompt.split(/\s+/).filter(Boolean).length
+  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
   }
 
@@ -140,21 +185,30 @@ ${joined}
 В конце добавь примечание, что пользователь может ответить кратко (например: "1, 2, темная тема") или довериться выбору по умолчанию.`
 }
 
-export function buildSynthesisPrompt(userPrompt, referenceOutputs = []) {
+export function buildSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '') {
   const joined = referenceOutputs
     .map((r, i) => {
       const fileSummary = (r.files && r.files.length > 0)
         ? ` [Созданные файлы: ${r.files.map((f) => f.relativePath).join(', ')}]`
         : ''
-      return `Reference ${i + 1} — ${r.label}:${fileSummary ? ` ${fileSummary}` : ''}\n${r.text}`
+      // If 3+ candidates, summarize text to prevent judge context overflow (#13)
+      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### 🎯 Дополнительные критерии оценки от пользователя:\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}
 
@@ -181,9 +235,10 @@ export function parseWinnerIndex(judgeText, defaultIndex = 1) {
     const idx = parseInt(match[1], 10)
     if (!isNaN(idx) && idx >= 1) return idx
   }
-  // Fallback: look for "Кандидат 1", "Reference 1", "Кандидат 2", "Reference 2"
-  if (/(?:Кандидат|Reference)\s*2\b/i.test(judgeText) && !/(?:Кандидат|Reference)\s*1\b/i.test(judgeText)) {
-    return 2
+  const candMatch = /(?:Кандидат|Reference)\s*(\d+)\b/i.exec(judgeText)
+  if (candMatch) {
+    const idx = parseInt(candMatch[1], 10)
+    if (!isNaN(idx) && idx >= 1) return idx
   }
   return defaultIndex
 }
@@ -263,6 +318,13 @@ export async function runReferencesParallel(references, messages, options = {},
       const text = (typeof res === 'string' ? res : (res?.content || res?.text || '')).trim()
       const files = extractFileBlocks(text)
 
+      // Estimate tokens & cost
+      const rawUsage = res?.usage || {
+        inputTokens: Math.round(JSON.stringify(fullMessages).length / 4),
+        outputTokens: Math.round(text.length / 4),
+      }
+      const costInfo = estimateTokenCost(slot, rawUsage)
+
       if (typeof onProgress === 'function') {
         const fileMsg = files.length > 0 ? ` (создано файлов: ${files.length})` : ''
         onProgress(`✅ *Кандидат ${i + 1} (${label}) завершил ответ${fileMsg}.*\n`)
@@ -275,6 +337,8 @@ export async function runReferencesParallel(references, messages, options = {},
         model: slot.model,
         text: text || '(empty response)',
         files,
+        usage: costInfo,
+        costUsd: costInfo.costUsd,
         ok: true,
       }
     } catch (err) {
@@ -289,6 +353,8 @@ export async function runReferencesParallel(references, messages, options = {},
         model: slot.model,
         text: `[failed: ${err?.message || String(err)}]`,
         files: [],
+        usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 },
+        costUsd: 0,
         ok: false,
       }
     }
@@ -299,10 +365,10 @@ export async function runReferencesParallel(references, messages, options = {},
 
 /**
  * Runs the full Mixture of Agents pipeline:
- * 1. Checks if questions are needed
+ * 1. Checks if questions or refinement needed
  * 2. Parallel candidate proposers write to .moa/candidate-X/
- * 3. Aggregator judges and picks winner
- * 4. Promotes winner files and cleans up
+ * 3. Aggregator judges and picks winner (or bypassed in Fast Mode)
+ * 4. Promotes winner files, records history and calculates costs
  */
 export async function runMoAPipeline({
   userPrompt,
@@ -317,15 +383,31 @@ export async function runMoAPipeline({
     throw new Error('callLlm function is required for runMoAPipeline')
   }
 
+  const startTime = Date.now()
   const referenceModels = preset?.reference_models || preset?.referenceModels || []
   const aggregator = preset?.aggregator || { provider: 'default', model: 'default' }
   const refTemp = preset?.reference_temperature ?? preset?.referenceTemperature ?? 0.6
   const aggTemp = preset?.aggregator_temperature ?? preset?.aggregatorTemperature ?? 0.4
   const maxTokens = preset?.max_tokens ?? preset?.maxTokens ?? 4096
+  const judgeCriteria = preset?.judge_criteria || preset?.judgeCriteria || ''
   const workDir = cwd || process.cwd()
 
+  // Collect project context (#6) and check refinement task (#4)
+  const projectCtx = await collectProjectContext(workDir, 16000)
+  const isRefinement = isRefinementTask(userPrompt, projectCtx.files)
+
+  // System prompt selection
+  let candidateSystemPrompt = REFERENCE_SYSTEM_PROMPT
+  if (isRefinement && projectCtx.files.length > 0) {
+    const fileList = projectCtx.files.map((f) => `### File: ${f.relativePath}\n\`\`\`\n${f.content}\n\`\`\``).join('\n\n')
+    candidateSystemPrompt = `${REFINEMENT_SYSTEM_PROMPT}\n\n## Existing Project Files:\n${fileList}`
+    if (typeof onProgress === 'function') {
+      onProgress(`🔄 *[Refinement Mode]: Обнаружен существующий проект (${projectCtx.files.length} файлов). Кандидаты вносят точечные изменения...*\n\n`)
+    }
+  }
+
   // Phase 1: Check if clarifying questions should be asked
-  const needsQuestions = !skipQuestions && isBroadPromptRequiringQuestions(userPrompt, messages)
+  const needsQuestions = !skipQuestions && !isRefinement && isBroadPromptRequiringQuestions(userPrompt, messages)
   if (needsQuestions) {
     if (typeof onProgress === 'function') {
       onProgress('🔍 *Задача общего характера. Советники формируют ключевые развилки...*\n\n')
@@ -371,15 +453,19 @@ export async function runMoAPipeline({
     }
   }
 
+  // FAST MODE: single candidate model bypasses aggregator (#14)
+  const isFastMode = referenceModels.length === 1
+
   // Phase 2: Parallel execution & file generation
   if (typeof onProgress === 'function') {
-    onProgress(`🚀 *Запускаю параллельную реализацию советниками (${referenceModels.length})...*\n\n`)
+    const modeLabel = isFastMode ? '⚡ Fast Mode' : `советниками (${referenceModels.length})`
+    onProgress(`🚀 *Запускаю реализацию ${modeLabel}...*\n\n`)
   }
 
   const referenceOutputs = await runReferencesParallel(
     referenceModels,
     [...messages, { role: 'user', content: userPrompt }],
-    { referenceTemperature: refTemp, maxTokens },
+    { systemPrompt: candidateSystemPrompt, referenceTemperature: refTemp, maxTokens },
     callLlm,
     onProgress,
   )
@@ -399,35 +485,51 @@ export async function runMoAPipeline({
     }
   }
 
-  // Phase 3: Aggregator evaluation
+  let synthesizedText = ''
+  let winningIndex = 1
+  let aggUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 }
   const aggLabel = slotLabel(aggregator)
-  if (typeof onProgress === 'function') {
-    onProgress(`\n⚖️ *Все кандидаты завершили генерацию. Судья (${aggLabel}) оценивает код и файлы...*\n\n`)
-  }
 
-  const synthPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs)
+  if (isFastMode) {
+    // Fast mode: promote candidate 1 directly without judge
+    synthesizedText = referenceOutputs[0]?.text || '(empty fast response)'
+    winningIndex = 1
+  } else {
+    // Phase 3: Aggregator evaluation
+    if (typeof onProgress === 'function') {
+      onProgress(`\n⚖️ *Все кандидаты завершили генерацию. Судья (${aggLabel}) оценивает код и файлы...*\n\n`)
+    }
 
-  let synthesizedText = ''
-  try {
-    const aggPromise = callLlm({
-      provider: aggregator.provider,
-      model: aggregator.model,
-      messages: [{ role: 'user', content: synthPrompt }],
-      temperature: aggTemp,
-      maxTokens,
-    })
-    const aggTimeout = new Promise((_, reject) =>
-      setTimeout(() => reject(new Error(`Timeout after 90s waiting for aggregator ${aggLabel}`)), 90000).unref()
-    )
-    const res = await Promise.race([aggPromise, aggTimeout])
-    synthesizedText = (typeof res === 'string' ? res : (res?.content || res?.text || '')).trim()
-  } catch (err) {
-    synthesizedText = `[Aggregator error: ${err?.message || String(err)}]\n\nFallback candidate outputs:\n\n` +
-      referenceOutputs.map((r, i) => `### ${r.label}\n${r.text}`).join('\n\n')
+    const synthPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria)
+
+    try {
+      const aggPromise = callLlm({
+        provider: aggregator.provider,
+        model: aggregator.model,
+        messages: [{ role: 'user', content: synthPrompt }],
+        temperature: aggTemp,
+        maxTokens,
+      })
+      const aggTimeout = new Promise((_, reject) =>
+        setTimeout(() => reject(new Error(`Timeout after 90s waiting for aggregator ${aggLabel}`)), 90000).unref()
+      )
+      const res = await Promise.race([aggPromise, aggTimeout])
+      synthesizedText = (typeof res === 'string' ? res : (res?.content || res?.text || '')).trim()
+
+      const rawAggUsage = res?.usage || {
+        inputTokens: Math.round(synthPrompt.length / 4),
+        outputTokens: Math.round(synthesizedText.length / 4),
+      }
+      aggUsage = estimateTokenCost(aggregator, rawAggUsage)
+    } catch (err) {
+      synthesizedText = `[Aggregator error: ${err?.message || String(err)}]\n\nFallback candidate outputs:\n\n` +
+        referenceOutputs.map((r, i) => `### ${r.label}\n${r.text}`).join('\n\n')
+    }
+
+    winningIndex = parseWinnerIndex(synthesizedText, 1)
   }
 
   // Phase 4: Promote winner files and cleanup
-  const winningIndex = parseWinnerIndex(synthesizedText, 1)
   let promotedFiles = []
   try {
     promotedFiles = await promoteCandidateWorkspace(workDir, winningIndex)
@@ -448,7 +550,8 @@ export async function runMoAPipeline({
         const lcRes = await fetch(`http://127.0.0.1:${port}/dsh-live-canvas/api/open-file`, {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
-          body: JSON.stringify({ filePath: previewCandidate })
+          body: JSON.stringify({ filePath: previewCandidate }),
+          signal: AbortSignal.timeout(500),
         })
         if (lcRes.ok) {
           const lcData = await lcRes.json()
@@ -470,21 +573,55 @@ export async function runMoAPipeline({
     }
   }
 
+  // Calculate totals
+  const totalTokens = referenceOutputs.reduce((acc, r) => acc + (r.usage?.totalTokens || 0), 0) + aggUsage.totalTokens
+  const totalCostUsd = Number((referenceOutputs.reduce((acc, r) => acc + (r.costUsd || 0), 0) + aggUsage.costUsd).toFixed(5))
+  const durationMs = Date.now() - startTime
+
+  const winningRef = referenceOutputs[winningIndex - 1]
+  const winnerModel = winningRef?.label || slotLabel(referenceModels[0])
+
+  // Record run to history (#9, #10, #11)
+  try {
+    recordMoaRun({
+      prompt: userPrompt,
+      preset: preset?.name || 'default',
+      isRefinement,
+      candidates: referenceOutputs,
+      aggregator: isFastMode ? null : { provider: aggregator.provider, model: aggregator.model, usage: aggUsage, costUsd: aggUsage.costUsd },
+      winnerIndex: winningIndex,
+      winnerModel,
+      promotedFiles,
+      totalTokens,
+      totalCostUsd,
+      durationMs,
+    })
+  } catch (histErr) {
+    console.warn('[dsh-moa] Failed to record run in history:', histErr)
+  }
+
   return {
     kind: 'synthesis',
     content: synthesizedText,
-    aggregator: aggLabel,
+    aggregator: isFastMode ? 'Fast Mode (Direct)' : aggLabel,
     references: referenceOutputs,
     presetName: preset?.name || 'default',
+    isRefinement,
+    isFastMode,
     winningIndex,
+    winnerModel,
     promotedFiles,
     liveCanvas,
+    usage: {
+      totalTokens,
+      totalCostUsd,
+      candidates: referenceOutputs.map(r => ({ label: r.label, usage: r.usage, costUsd: r.costUsd })),
+      aggregator: aggUsage,
+    },
+    durationMs,
   }
 }
 
-/**
- * Formats full MoA output showing candidate outputs, judge synthesis, and promoted files.
- */
 /**
  * Replaces large code blocks with concise file/code summaries to avoid token waste and huge chat dumps.
  */
@@ -492,7 +629,6 @@ 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')
-    // Keep very short commands (<= 3 lines) that are not HTML/JSX/source files
     if (lines.length <= 3 && !/html|jsx|tsx|vue|svelte|css|js|ts/i.test(lang)) {
       return match
     }
@@ -515,9 +651,14 @@ export function formatMoAResponse({ moaResult, presetName }) {
     return parts.join('\n')
   }
 
-  parts.push(`## 🧠 Mixture of Agents (Пресет: ${pName} | Судья: ${judge})`)
+  const modeBadge = moaResult?.isFastMode ? '⚡ Fast Mode' : `Судья: ${judge}`
+  parts.push(`## 🧠 Mixture of Agents (Пресет: ${pName} | ${modeBadge})`)
   parts.push('')
 
+  if (moaResult?.isRefinement) {
+    parts.push('> 🔄 **Режим**: Итеративная доработка проекта (Refinement)')
+  }
+
   const hasPromoted = moaResult?.promotedFiles && moaResult.promotedFiles.length > 0
   if (hasPromoted) {
     parts.push(`> 📦 **Созданы файлы в проекте**: \`${moaResult.promotedFiles.join('`, `')}\``)
@@ -527,25 +668,35 @@ export function formatMoAResponse({ moaResult, presetName }) {
     parts.push(`> 🎨 **Live Canvas**: [🚀 Открыть ${moaResult.liveCanvas.title || 'превью'} в Live Canvas](${moaResult.liveCanvas.previewUrl}) | [↗ Открыть в новой вкладке](${moaResult.liveCanvas.previewUrl})`)
   }
 
-  if (hasPromoted || moaResult?.liveCanvas) {
-    parts.push('')
+  // Cost tracking card (#10)
+  if (moaResult?.usage) {
+    const u = moaResult.usage
+    const costStr = u.totalCostUsd > 0 ? `~\$${u.totalCostUsd.toFixed(4)}` : 'Бесплатно'
+    const tokStr = u.totalTokens >= 1000 ? `${(u.totalTokens / 1000).toFixed(1)}k` : `${u.totalTokens}`
+    parts.push(`> 💰 **Стоимость запуска**: ${costStr} (всего ${tokStr} токенов)`)
   }
 
-  parts.push(`### ⚖️ Вердикт судьи и итоговое решение (Синтез: ${judge})`)
-  parts.push('')
-  const cleanJudgeContent = hasPromoted
-    ? stripOrSummarizeCode(moaResult?.content || '')
-    : (moaResult?.content || '(нет ответа)')
-  parts.push(cleanJudgeContent)
   parts.push('')
 
+  if (!moaResult?.isFastMode) {
+    parts.push(`### ⚖️ Вердикт судьи и итоговое решение (Синтез: ${judge})`)
+    parts.push('')
+    const cleanJudgeContent = hasPromoted
+      ? stripOrSummarizeCode(moaResult?.content || '')
+      : (moaResult?.content || '(нет ответа)')
+    parts.push(cleanJudgeContent)
+    parts.push('')
+  }
+
   if (Array.isArray(refs) && refs.length > 0) {
-    parts.push(`### 👥 Ответы моделей-советников (${refs.length}):`)
+    const title = moaResult?.isFastMode ? '### 🚀 Результат генерации кандидата:' : `### 👥 Ответы моделей-советников (${refs.length}):`
+    parts.push(title)
     parts.push('')
     refs.forEach((ref, i) => {
       const statusIcon = ref.ok ? '✅' : '⚠️'
       const fileBadge = ref.files?.length ? ` (${ref.files.length} файл(ов))` : ''
-      parts.push(`#### ${statusIcon} Модель ${i + 1}: ${ref.label}${fileBadge}`)
+      const costBadge = ref.costUsd > 0 ? ` [~\$${ref.costUsd.toFixed(4)}]` : ''
+      parts.push(`#### ${statusIcon} Модель ${i + 1}: ${ref.label}${fileBadge}${costBadge}`)
       parts.push('')
       parts.push(stripOrSummarizeCode(ref.text))
       parts.push('')
@@ -618,7 +769,7 @@ export class MoaRunnerAdapter {
       const startTime = Date.now()
       let lastYieldTime = Date.now()
 
-      // Interval to yield progressive updates while pipeline runs
+      // Yield progressive updates while pipeline runs
       while (true) {
         if (signal?.aborted) return
 
@@ -659,7 +810,7 @@ export class MoaRunnerAdapter {
 
       yield { type: 'text-delta', index: 0, text: '\n---\n\n' + formatted }
       yield { type: 'block-end', index: 0, block: { type: 'text', text: formatted } }
-      yield { type: 'usage', usage: { inputTokens: 100, outputTokens: formatted.length } }
+      yield { type: 'usage', usage: { inputTokens: moaResult?.usage?.totalTokens || 100, outputTokens: formatted.length } }
       yield { type: 'finish', reason: { kind: 'stop' } }
     } catch (err) {
       if (signal?.aborted) return

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-moa",
-  "version": "0.2.2",
+  "version": "0.2.4",
   "description": "Mixture of Agents (MoA) plugin for DeepSeek Harness with /moa slash command",
   "type": "module",
   "main": "lib/index.js",