@goodandready/dsh-moa 0.2.26 → 0.2.29

package/lib/moa-router.js

@@ -0,0 +1,102 @@
+/**
+ * Smart Preset Router for Mixture of Agents (MoA).
+ * Supports heuristic keyword classification and optional JEV/LLM zero-shot routing.
+ */
+
+const INTENT_RULES = [
+  { preset: 'security-audit', regex: /\b(security|vulnerability|exploit|cve|xss|injection|sanitize|threat|auth|permission|secret)\b/i },
+  { preset: 'bug-hunter', regex: /\b(bug|fix|error|crash|exception|regression|reproduce|broken|fail|issue)\b/i },
+  { preset: 'refactor-cleanup', regex: /\b(refactor|clean|simplify|ponytail|yagni|prune|dedup|modular|dead\s*code)\b/i },
+  { preset: 'code-review', regex: /\b(review|critique|audit|pr|pull\s*request|diff|inspect|check\s*code)\b/i },
+  { preset: 'frontend-ui', regex: /\b(frontend|ui|css|html|react|vue|tailwind|styling|styles?|stylesheet|component|button|layout|page|landing|modal)\b/i },
+  { preset: 'deep-architect', regex: /\b(architect|distributed|system|microservice|database|scalab|pipeline|infra|schema)\b/i },
+  { preset: 'math-logic', regex: /\b(math|algorithm|proof|matrix|calc|complexity|combinatorics|graph|tree|dynamic\s*programming)\b/i },
+  { preset: 'creative-brainstorm', regex: /\b(brainstorm|idea|concept|creative|alternative|options|feature\s*idea)\b/i },
+  { preset: 'fast-audit', regex: /\b(quick|fast|brief|summary|one-line|short|tldr)\b/i },
+]
+
+/**
+ * Fast keyword-based intent classification.
+ */
+export function classifyPromptIntent(prompt = '') {
+  if (!prompt || typeof prompt !== 'string') return null
+  const text = prompt.trim()
+  for (const rule of INTENT_RULES) {
+    if (rule.regex.test(text)) {
+      return rule.preset
+    }
+  }
+  return null
+}
+
+/**
+ * Resolves the optimal preset for a given user prompt.
+ */
+export async function resolvePresetForPrompt({
+  prompt = '',
+  presets = [],
+  defaultPreset = 'default',
+  routingModel = null,
+  callLlm = null,
+  enabled = false,
+  timeoutMs = 2000,
+}) {
+  const availableNames = (presets || []).map((p) => p.name).filter(Boolean)
+  const safeDefault = availableNames.includes(defaultPreset) ? defaultPreset : (availableNames[0] || 'default')
+
+  if (!enabled || !prompt || typeof prompt !== 'string') {
+    return { presetName: safeDefault, reason: 'default', isAutoRouted: false }
+  }
+
+  // 1. If LLM routing model (e.g. JEV) is configured and callable, run fast zero-shot classifier
+  if (routingModel?.provider && routingModel?.model && typeof callLlm === 'function') {
+    try {
+      const systemPrompt = `You are the MoA Preset Router. Given user prompt, classify which preset is best suited.
+Available presets: ${availableNames.join(', ')}
+Respond with ONLY the chosen preset name and nothing else.`
+
+      const abortCtrl = new AbortController()
+      const timer = setTimeout(() => abortCtrl.abort(), timeoutMs)
+      timer.unref?.()
+
+      const res = await callLlm({
+        provider: routingModel.provider,
+        model: routingModel.model,
+        messages: [
+          { role: 'system', content: systemPrompt },
+          { role: 'user', content: prompt.slice(0, 500) },
+        ],
+        temperature: 0.1,
+        maxTokens: 30,
+        signal: abortCtrl.signal,
+      })
+      clearTimeout(timer)
+
+      const rawChoice = typeof res === 'string' ? res : (res?.content || res?.text || '')
+      const match = availableNames.find((name) => new RegExp(`\\b${name}\\b`, 'i').test(rawChoice))
+      if (match) {
+        return {
+          presetName: match,
+          reason: 'llm_classifier',
+          routerModel: `${routingModel.provider}:${routingModel.model}`,
+          isAutoRouted: true,
+        }
+      }
+    } catch {
+      // Fallback silently to heuristic rules on timeout or network error
+    }
+  }
+
+  // 2. Keyword heuristic classifier
+  const heuristic = classifyPromptIntent(prompt)
+  if (heuristic && availableNames.includes(heuristic)) {
+    return {
+      presetName: heuristic,
+      reason: 'keyword_heuristic',
+      isAutoRouted: true,
+    }
+  }
+
+  // 3. Fallback to default
+  return { presetName: safeDefault, reason: 'default_fallback', isAutoRouted: false }
+}

package/lib/moa-runner.js

@@ -1,14 +1,19 @@
+import { logger } from './logger.js'
 import path from 'node:path'
 import crypto from 'node:crypto'
 import { bestEffort } from './best-effort.js'
 import { extractFileBlocks, collectProjectContext, formatProjectContext, isRefinementTask, writeCandidateWorkspace, promoteCandidateWorkspace, cleanMoaWorkspaces, verifyFileSyntax } from './file-workspace.js'
 import { estimateTokenCost, summarizeMoAUsage } from './pricing.js'
-import { recordMoaRun, recordMoaRunAsync, candidatesForHistory } from './history.js'
+import { recordMoaRunAsync, candidatesForHistory } from './history.js'
 import { createPromotedPreview } from './live-canvas.js'
 import { slotLabel, cleanAdvisoryMessages, isBroadPromptRequiringQuestions, buildQuestionSynthesisPrompt, buildCuratorSynthesisPrompt, buildSynthesisPrompt, buildPeerCritiquePrompt, ROLE_PERSONA_PROMPTS, SYSTEM_ROLE_PROPOSER, ANTIPATTERNS_RUBRIC } from './moa-prompts.js'
 import { parseWinnerIndex, parseRecommendedAssembler, parseMoACommand, stripOrSummarizeCode, formatMoAResponse } from './moa-parser.js'
 import { REFERENCE_SYSTEM_PROMPT, callWithTransientRetry, runReferencesParallel } from './moa-candidates.js'
 import { executeTestGateForCandidates, runCandidateTestGate } from './moa-test-gate.js'
+import { executeMultiJudgePanel, buildCompositeBlockDirectives } from './moa-multi-judge.js'
+import { applyBudgetGuardrails } from './moa-budget.js'
+import { extractPriorTurnBaseline, pruneMultiTurnMessages } from './moa-context.js'
+import { generateMoABenchmarkReport } from './moa-report.js'
 
 // Re-exports for consumers & backward compatibility
 export { slotLabel, cleanAdvisoryMessages, isBroadPromptRequiringQuestions, buildQuestionSynthesisPrompt, buildCuratorSynthesisPrompt, buildSynthesisPrompt, ANTIPATTERNS_RUBRIC } from './moa-prompts.js'
@@ -16,6 +21,7 @@ export { parseWinnerIndex, parseRecommendedAssembler, parseMoACommand, stripOrSu
 export { estimateTokenCost, summarizeMoAUsage } from './pricing.js'
 export { REFERENCE_SYSTEM_PROMPT, callWithTransientRetry, runReferencesParallel } from './moa-candidates.js'
 export { executeTestGateForCandidates, runCandidateTestGate } from './moa-test-gate.js'
+export { streamMoATurn } from './moa-stream.js'
 
 /**
  * Executes the full Mixture of Agents pipeline.
@@ -60,6 +66,38 @@ export async function runMoAPipeline({
   const isBlindEvaluation = Boolean(preset?.blind_evaluation), isPeerCritiqueEnabled = Boolean(preset?.peer_critique_enabled)
   const allowCandidateOverride = Boolean(preset?.allow_candidate_override), runId = crypto.randomUUID()
 
+  // Feature 6: Budget Guardrail check
+  const budgetCheck = applyBudgetGuardrails({
+    references: referenceModels,
+    enabled: preset?.budget_guard_enabled,
+    maxBudgetUsd: preset?.max_budget_usd,
+    action: preset?.budget_action || 'trim',
+    prices,
+    promptLength: userPrompt?.length || 1000,
+  })
+  if (!budgetCheck.allowed) {
+    return {
+      kind: 'failure',
+      content: `🛑 Budget Guardrail Abort: ${budgetCheck.reason}`,
+      aggregator: slotLabel(primaryJudge),
+      references: [],
+      presetName: preset?.name || 'default',
+      isRefinement: false,
+      isFastMode,
+      winningIndex: 0,
+      winnerModel: 'none',
+      promotedFiles: [],
+      usage: { totalTokens: 0, totalCostUsd: 0, candidates: [], aggregator: { totalTokens: 0, costUsd: 0 } },
+      skippedFiles: 0,
+      skippedList: [],
+      durationMs: Date.now() - startTime,
+    }
+  }
+  const effectiveReferences = budgetCheck.references
+  if (budgetCheck.action === 'trim' && typeof onProgress === 'function') {
+    onProgress(`✂️ *${budgetCheck.reason}*\n`)
+  }
+
   // 2. Collect project context for refinement tasks
   const collectedCtx = await collectProjectContext(cwd, 16000)
   const isRefinement = Boolean(collectedCtx?.files?.length > 0 && isRefinementTask(userPrompt, collectedCtx.files))
@@ -79,23 +117,28 @@ export async function runMoAPipeline({
     })
   }
 
-  // 3. Build prompts & evaluate broad questionnaire needs
+  // 3. Build prompts & Feature 8 multi-turn context
+  const priorTurnBaseline = preset?.multi_turn_enabled !== false ? extractPriorTurnBaseline(messages) : null
+  const prunedHistory = preset?.multi_turn_enabled !== false ? pruneMultiTurnMessages(messages) : messages
+
   const candidateSystemPrompt = projectContext
     ? `${REFERENCE_SYSTEM_PROMPT}\n\n### Current Project Files & Context:\n${projectContext}`
     : REFERENCE_SYSTEM_PROMPT
 
   const askClarifyingQuestions = preset?.ask_clarifying_questions !== false
   const needsQuestions = askClarifyingQuestions && isBroadPromptRequiringQuestions(userPrompt, messages)
+  const enrichedMessages = [...prunedHistory, { role: 'user', content: userPrompt }]
 
-  const enrichedMessages = [...messages, { role: 'user', content: userPrompt }]
-
-  // 4. Parallel fan-out to candidate models with quorum & transient retry
+  // 4. Parallel fan-out to candidate models with temperature gradient & local fallback
   const referenceOutputs = await runReferencesParallel(
-    referenceModels,
+    effectiveReferences,
     enrichedMessages,
     {
       systemPrompt: candidateSystemPrompt,
       temperature: refTemp,
+      temperature_gradient_enabled: Boolean(preset?.temperature_gradient_enabled),
+      local_fallback_enabled: Boolean(preset?.local_fallback_enabled),
+      local_fallback_models: preset?.local_fallback_models || [],
       maxTokens,
       prices,
       timeoutSec: refTimeoutSec,
@@ -185,7 +228,7 @@ export async function runMoAPipeline({
       }
       qUsage = estimateTokenCost(primaryJudge, qFallbackUsage, prices)
     } catch (err) {
-      console.warn('[dsh-moa] Questionnaire synthesis failed, proceeding with fallback questions:', err)
+      logger.warn('[dsh-moa] Questionnaire synthesis failed, proceeding with fallback questions:', err)
       questionsContent = `### Clarification of Requirements: "${userPrompt}"\n\n1. **Architecture & Scope**: Single-file deliverable or multi-module project structure?\n2. **Design & Style**: Minimalist, dark mode, or clean neutral theme?\n3. **Functional Priorities**: Core MVP or comprehensive extended implementation?\n\n*Reply with your preferences (e.g. "1, 2") or proceed with defaults.*`
     }
 
@@ -195,7 +238,7 @@ export async function runMoAPipeline({
     const durationMs = Date.now() - startTime
 
     try {
-      recordMoaRun({
+      await recordMoaRunAsync({
         prompt: userPrompt,
         preset: preset?.name || 'default',
         isRefinement,
@@ -209,7 +252,7 @@ export async function runMoAPipeline({
         durationMs,
       }, historyFilePath || undefined)
     } catch (histErr) {
-      console.warn('[dsh-moa] Failed to record questionnaire run in history:', histErr?.message || histErr)
+      logger.warn('[dsh-moa] Failed to record questionnaire run in history:', histErr?.message || histErr)
     }
 
     return {
@@ -251,7 +294,7 @@ export async function runMoAPipeline({
     const durationMs = Date.now() - startTime
 
     try {
-      recordMoaRun({
+      await recordMoaRunAsync({
         prompt: userPrompt,
         preset: preset?.name || 'default',
         isRefinement,
@@ -266,7 +309,7 @@ export async function runMoAPipeline({
         durationMs,
       }, historyFilePath || undefined)
     } catch (histErr) {
-      console.warn('[dsh-moa] Failed to record fast synthesis run in history:', histErr?.message || histErr)
+      logger.warn('[dsh-moa] Failed to record fast synthesis run in history:', histErr?.message || histErr)
     }
 
     return {
@@ -334,7 +377,7 @@ export async function runMoAPipeline({
           cand.costUsd = Number(((cand.costUsd || 0) + extraCost.costUsd).toFixed(5))
         }
       } catch (r2CostErr) {
-        console.warn('[dsh-moa] Failed to estimate Round 2 cost:', r2CostErr?.message || r2CostErr)
+        logger.warn('[dsh-moa] Failed to estimate Round 2 cost:', r2CostErr?.message || r2CostErr)
       }
     })
     await Promise.allSettled(r2Promises)
@@ -347,10 +390,32 @@ export async function runMoAPipeline({
     })
   }
 
+  // 7c. Feature 2: Multi-Judge Panel & Consensus Voting
+  let multiJudgeResult = null
+  if (preset?.multi_judge_enabled && successfulRefs.length > 1) {
+    multiJudgeResult = await executeMultiJudgePanel({
+      judges: (preset.judge_models && preset.judge_models.length > 0) ? preset.judge_models : [primaryJudge],
+      candidates: successfulRefs,
+      userPrompt,
+      callLlm,
+      strategy: preset.judge_voting_strategy || 'majority',
+      signal,
+      onProgress,
+    })
+  }
+
+  // 7d. Feature 3: Composite Block Merge Directive
+  const compositeMergeDirective = (preset?.composite_merge_enabled && successfulRefs.length > 1)
+    ? buildCompositeBlockDirectives(successfulRefs)
+    : null
+
   // 8. Synthesis phase via primary judge or fallback chain
   const synthesisPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, {
     curatorSynthesis: isCuratorSynthesis,
     blindEvaluation: isBlindEvaluation,
+    priorTurnBaseline,
+    compositeMergeDirective,
+    consensusReport: multiJudgeResult?.consensusReport,
   })
   let synthesizedText = ''
   let aggUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 }
@@ -396,7 +461,7 @@ export async function runMoAPipeline({
       break
     } catch (err) {
       lastJudgeError = err
-      console.warn(`[dsh-moa] Judge ${currentLabel} failed:`, err)
+      logger.warn(`[dsh-moa] Judge ${currentLabel} failed:`, err)
       if (typeof onProgress === 'function') {
         const nextJudge = judgesChain[jIdx + 1]
         const nextHint = nextJudge ? ` Trying fallback ${slotLabel(nextJudge)}...` : ''
@@ -412,7 +477,10 @@ export async function runMoAPipeline({
   }
 
   // 9. Evaluate winner & promote files
-  const winningIndex = parseWinnerIndex(synthesizedText, 1, referenceOutputs.length)
+  const winningIndex = (multiJudgeResult?.consensus && multiJudgeResult?.winningCandidateIndex)
+    ? multiJudgeResult.winningCandidateIndex
+    : parseWinnerIndex(synthesizedText, 1, referenceOutputs.length)
+
   const recommendedAssembler = isCuratorSynthesis
     ? parseRecommendedAssembler(synthesizedText, winningIndex, referenceOutputs.length)
     : null
@@ -435,15 +503,34 @@ export async function runMoAPipeline({
   }
 
   const livePreview = await createPromotedPreview(liveCanvas, cwd, promotedFiles, referenceOutputs)
-
   const durationMs = Date.now() - startTime
   const usageSummary = summarizeMoAUsage(referenceOutputs, aggUsage)
   const winningRef = referenceOutputs[winningIndex - 1]
   const winnerModel = winningRef?.label || slotLabel(referenceModels[0])
   const finalAggLabel = slotLabel(chosenJudge)
 
+  // Feature 9: Benchmark & Post-Mortem PR report
+  let benchmarkReport = null
+  if (preset?.report_generation_enabled) {
+    benchmarkReport = generateMoABenchmarkReport({
+      runId,
+      timestamp: new Date().toISOString(),
+      preset: preset?.name || 'default',
+      prompt: userPrompt,
+      durationMs,
+      usage: usageSummary,
+      costUsd: usageSummary.totalCostUsd,
+      candidates: referenceOutputs,
+      winningIndex,
+      winningLabel: winnerModel,
+      consensus: multiJudgeResult,
+      testGate: referenceOutputs.find((r) => r.testResult)?.testResult || null,
+      isComposite: Boolean(preset?.composite_merge_enabled),
+    })
+  }
+
   try {
-    recordMoaRun({
+    await recordMoaRunAsync({
       id: runId,
       prompt: userPrompt,
       preset: preset?.name || 'default',
@@ -453,12 +540,14 @@ export async function runMoAPipeline({
       winnerIndex: winningIndex,
       winnerModel,
       promotedFiles,
+      benchmarkReport,
+      consensus: multiJudgeResult,
       totalTokens: usageSummary.totalTokens,
       totalCostUsd: usageSummary.totalCostUsd,
       durationMs,
     }, historyFilePath || undefined)
   } catch (histErr) {
-    console.warn('[dsh-moa] Failed to record run in history:', histErr)
+    logger.warn('[dsh-moa] Failed to record run in history:', histErr)
   }
 
   return {
@@ -476,6 +565,8 @@ export async function runMoAPipeline({
     promotedFiles,
     runId,
     allowCandidateOverride,
+    benchmarkReport,
+    consensus: multiJudgeResult,
     ...(livePreview ? { liveCanvas: livePreview } : {}),
     usage: usageSummary,
     skippedFiles: collectedCtx?.skippedFiles || 0,
@@ -483,113 +574,3 @@ export async function runMoAPipeline({
     durationMs,
   }
 }
-
-/**
- * Streams a full MoA turn into chat with live aggregator tokens and progress feedback.
- */
-export async function* streamMoATurn({ targetPreset, userPrompt, messages, callLlm, cwd, prices, historyFilePath, liveCanvas }, options = {}) {
-  const signal = options?.signal
-  if (signal?.aborted) return
-
-  yield { type: 'block-start', index: 0, blockType: 'text' }
-  yield { type: 'text-delta', index: 0, text: '🧠 *Mixture of Agents started...*\n\n' }
-
-  // Async push-queue for zero-latency live delta streaming
-  const queue = []
-  let notify = null
-
-  const pushUpdate = (text) => {
-    queue.push({ type: 'text-delta', index: 0, text })
-    if (notify) {
-      notify()
-      notify = null
-    }
-  }
-
-  let done = false
-  const pipelinePromise = runMoAPipeline({
-    userPrompt,
-    messages,
-    preset: targetPreset,
-    callLlm,
-    cwd,
-    onProgress: pushUpdate,
-    onStreamDelta: (delta) => pushUpdate(delta),
-    prices,
-    historyFilePath,
-    liveCanvas,
-    signal,
-  })
-    .catch((err) => ({ error: err }))
-    .finally(() => {
-      done = true
-      if (notify) {
-        notify()
-        notify = null
-      }
-    })
-
-  const startTime = Date.now()
-  let lastYieldTime = Date.now()
-
-  try {
-    while (!done || queue.length > 0) {
-      if (signal?.aborted) {
-        await cleanMoaWorkspaces(cwd)
-        return
-      }
-
-      while (queue.length > 0) {
-        const item = queue.shift()
-        yield item
-        lastYieldTime = Date.now()
-      }
-
-      if (done) break
-
-      // Wait for next push item or max 2.5s heartbeat
-      await Promise.race([
-        new Promise((resolve) => { notify = resolve }),
-        new Promise((resolve) => { const t = setTimeout(resolve, 2500); t.unref?.(); }),
-      ])
-
-      const elapsedSec = Math.floor((Date.now() - startTime) / 1000)
-      if (!done && Date.now() - lastYieldTime >= 3000) {
-        yield { type: 'text-delta', index: 0, text: `⏳ *[${elapsedSec}s] Still processing...*\n` }
-        lastYieldTime = Date.now()
-      }
-    }
-
-    const result = await pipelinePromise
-    if (signal?.aborted) {
-      await cleanMoaWorkspaces(cwd)
-      return
-    }
-
-    if (result?.error) {
-      const errText = '\n\n⚠️ **Mixture of Agents error**: ' + (result.error?.message || String(result.error))
-      yield { type: 'text-delta', index: 0, text: errText }
-      yield { type: 'block-end', index: 0, block: { type: 'text', text: errText } }
-      yield { type: 'finish', reason: { kind: 'stop' } }
-      return
-    }
-
-    const formatted = formatMoAResponse({
-      moaResult: result,
-      presetName: targetPreset?.name || 'default',
-    })
-
-    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: result?.usage?.totalTokens || 0,
-        outputTokens: Math.round((formatted.length || 0) / 4),
-      },
-    }
-    yield { type: 'finish', reason: { kind: 'stop' } }
-  } finally {
-    // cleanup
-  }
-}

package/lib/moa-stream.js

@@ -0,0 +1,117 @@
+/**
+ * Zero-latency Live Streaming Generator for MoA Turns
+ */
+
+import { cleanMoaWorkspaces } from './file-workspace.js'
+import { formatMoAResponse } from './moa-parser.js'
+import { runMoAPipeline } from './moa-runner.js'
+
+/**
+ * Streams a full MoA turn into chat with live aggregator tokens and progress feedback.
+ */
+export async function* streamMoATurn({ targetPreset, userPrompt, messages, callLlm, cwd, prices, historyFilePath, liveCanvas }, options = {}) {
+  const signal = options?.signal
+  if (signal?.aborted) return
+
+  yield { type: 'block-start', index: 0, blockType: 'text' }
+  yield { type: 'text-delta', index: 0, text: '🧠 *Mixture of Agents started...*\n\n' }
+
+  // Async push-queue for zero-latency live delta streaming
+  const queue = []
+  let notify = null
+
+  const pushUpdate = (text) => {
+    queue.push({ type: 'text-delta', index: 0, text })
+    if (notify) {
+      notify()
+      notify = null
+    }
+  }
+
+  let done = false
+  const pipelinePromise = runMoAPipeline({
+    userPrompt,
+    messages,
+    preset: targetPreset,
+    callLlm,
+    cwd,
+    onProgress: pushUpdate,
+    onStreamDelta: (delta) => pushUpdate(delta),
+    prices,
+    historyFilePath,
+    liveCanvas,
+    signal,
+  })
+    .catch((err) => ({ error: err }))
+    .finally(() => {
+      done = true
+      if (notify) {
+        notify()
+        notify = null
+      }
+    })
+
+  const startTime = Date.now()
+  let lastYieldTime = Date.now()
+
+  try {
+    while (!done || queue.length > 0) {
+      if (signal?.aborted) {
+        await cleanMoaWorkspaces(cwd)
+        return
+      }
+
+      while (queue.length > 0) {
+        const item = queue.shift()
+        yield item
+        lastYieldTime = Date.now()
+      }
+
+      if (done) break
+
+      // Wait for next push item or max 2.5s heartbeat
+      await Promise.race([
+        new Promise((resolve) => { notify = resolve }),
+        new Promise((resolve) => { const t = setTimeout(resolve, 2500); t.unref?.(); }),
+      ])
+
+      const elapsedSec = Math.floor((Date.now() - startTime) / 1000)
+      if (!done && Date.now() - lastYieldTime >= 3000) {
+        yield { type: 'text-delta', index: 0, text: `⏳ *[${elapsedSec}s] Still processing...*\n` }
+        lastYieldTime = Date.now()
+      }
+    }
+
+    const result = await pipelinePromise
+    if (signal?.aborted) {
+      await cleanMoaWorkspaces(cwd)
+      return
+    }
+
+    if (result?.error) {
+      const errText = '\n\n⚠️ **Mixture of Agents error**: ' + (result.error?.message || String(result.error))
+      yield { type: 'text-delta', index: 0, text: errText }
+      yield { type: 'block-end', index: 0, block: { type: 'text', text: errText } }
+      yield { type: 'finish', reason: { kind: 'stop' } }
+      return
+    }
+
+    const formatted = formatMoAResponse({
+      moaResult: result,
+      presetName: targetPreset?.name || 'default',
+    })
+
+    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: result?.usage?.totalTokens || 0,
+        outputTokens: Math.round((formatted.length || 0) / 4),
+      },
+    }
+    yield { type: 'finish', reason: { kind: 'stop' } }
+  } finally {
+    // cleanup
+  }
+}

package/lib/moa-test-gate.js

@@ -4,6 +4,7 @@ import fsSync from 'node:fs'
 import path from 'node:path'
 import { spawn } from 'node:child_process'
 import { assertPathContained } from './file-workspace.js'
+import { bestEffort } from './best-effort.js'
 
 /**
  * Resolves the directory for candidate sandbox.
@@ -179,9 +180,9 @@ export async function runCandidateTestGate({
     }
   } finally {
     if (stageDir) {
-      try {
-        await fs.rm(stageDir, { recursive: true, force: true })
-      } catch {}
+      await bestEffort('moa-test-gate cleanup stageDir', () =>
+        fs.rm(stageDir, { recursive: true, force: true })
+      )
     }
   }
 }