@goodandready/dsh-moa 0.2.11 → 0.2.13

package/lib/index.js

@@ -23,7 +23,9 @@ import {
   getMoaHistory,
   getMoaLeaderboard,
   getMoaRunById,
+  exportMoaHistory,
 } from './history.js'
+import { promoteCandidateWorkspace } from './file-workspace.js'
 
 import { createLiveCanvasClient } from './live-canvas.js'
 
@@ -41,6 +43,7 @@ export const PriceRow = z.object({
 export const ModelSlotSchema = z.object({
   provider: z.string().default('opencode-go'),
   model: z.string().default('deepseek-v4-flash'),
+  role_persona: z.string().default(''),
 })
 
 export const PresetSchema = z.object({
@@ -58,6 +61,9 @@ export const PresetSchema = z.object({
   aggregator_temperature: z.number().default(0.4),
   reference_timeout_sec: z.number().default(60),
   aggregator_timeout_sec: z.number().default(180),
+  blind_evaluation: z.boolean().default(false),
+  peer_critique_enabled: z.boolean().default(false),
+  allow_candidate_override: z.boolean().default(false),
   max_tokens: z.number().default(4096),
   judge_criteria: z.string().default(''),
 })
@@ -420,7 +426,7 @@ export function apply(ctx, config) {
     })
   }, 'dsh-moa: history route')
 
-  // Route: /dsh-moa/leaderboard (#11)
+  // Route: /dsh-moa/leaderboard
   ctx.effect(() => {
     return ctx.webServer.register({
       kind: 'exact',
@@ -431,7 +437,21 @@ export function apply(ctx, config) {
           return
         }
         try {
-          const data = getMoaLeaderboard()
+          const url = new URL(req.url || '', 'http://127.0.0.1')
+          const preset = url.searchParams.get('preset') || null
+          const format = url.searchParams.get('format') || 'json'
+
+          if (format === 'csv') {
+            const csv = exportMoaHistory(undefined, { format: 'csv', preset })
+            res.writeHead(200, {
+              'Content-Type': 'text/csv; charset=utf-8',
+              'Content-Disposition': 'attachment; filename="moa-leaderboard.csv"',
+            })
+            res.end(csv)
+            return
+          }
+
+          const data = getMoaLeaderboard(undefined, { preset })
           writeJson(res, 200, { ok: true, ...data })
         } catch (err) {
           writeJson(res, 500, { ok: false, error: err?.message || String(err) })
@@ -440,6 +460,71 @@ export function apply(ctx, config) {
     })
   }, 'dsh-moa: leaderboard route')
 
+  // Route: /dsh-moa/promote (Candidate Override Action)
+  ctx.effect(() => {
+    return ctx.webServer.register({
+      kind: 'exact',
+      path: '/dsh-moa/promote',
+      handler: async (req, res) => {
+        if (req.method !== 'POST') {
+          writeJson(res, 405, { ok: false, error: 'POST required' })
+          return
+        }
+        try {
+          const raw = await readBody(req)
+          const body = JSON.parse(raw || '{}')
+          const candidateIndex = parseInt(body?.candidateIndex, 10)
+          if (isNaN(candidateIndex) || candidateIndex < 1) {
+            writeJson(res, 400, { ok: false, error: 'Valid candidateIndex required (>= 1)' })
+            return
+          }
+          const cwd = body?.cwd || process.cwd()
+          const promotedFiles = await promoteCandidateWorkspace(cwd, candidateIndex, { keepMoa: true })
+          writeJson(res, 200, { ok: true, candidateIndex, promotedFiles })
+        } catch (err) {
+          writeJson(res, 500, { ok: false, error: err?.message || String(err) })
+        }
+      },
+    })
+  }, 'dsh-moa: promote candidate route')
+
+  // Route: /dsh-moa/history/export
+  ctx.effect(() => {
+    return ctx.webServer.register({
+      kind: 'exact',
+      path: '/dsh-moa/history/export',
+      handler: async (req, res) => {
+        if (req.method !== 'GET') {
+          writeJson(res, 405, { ok: false, error: 'GET required' })
+          return
+        }
+        try {
+          const url = new URL(req.url || '', 'http://127.0.0.1')
+          const preset = url.searchParams.get('preset') || null
+          const format = url.searchParams.get('format') || 'json'
+          const data = exportMoaHistory(undefined, { format, preset })
+
+          if (format === 'csv') {
+            res.writeHead(200, {
+              'Content-Type': 'text/csv; charset=utf-8',
+              'Content-Disposition': 'attachment; filename="moa-history.csv"',
+            })
+            res.end(data)
+            return
+          }
+
+          res.writeHead(200, {
+            'Content-Type': 'application/json; charset=utf-8',
+            'Content-Disposition': 'attachment; filename="moa-history.json"',
+          })
+          res.end(data)
+        } catch (err) {
+          writeJson(res, 500, { ok: false, error: err?.message || String(err) })
+        }
+      },
+    })
+  }, 'dsh-moa: history export route')
+
   // Route: /dsh-moa/runs/<id> (run replay by id)
   ctx.effect(() => {
     return ctx.webServer.register({

package/lib/moa-parser.js

@@ -45,6 +45,17 @@ export function parseMoACommand(text, presets = []) {
     return null
   }
 
+  // Check for promote subcommand: /moa promote <runId> <candidateIndex> or /moa promote <candidateIndex>
+  const promoteMatch = /^\/moa\s+promote\s+([a-zA-Z0-9_\-]+)(?:\s+(\d+))?/i.exec(text.trim())
+  if (promoteMatch) {
+    const hasTwoArgs = Boolean(promoteMatch[2])
+    return {
+      isPromote: true,
+      runId: hasTwoArgs ? promoteMatch[1] : null,
+      candidateIndex: hasTwoArgs ? parseInt(promoteMatch[2], 10) : parseInt(promoteMatch[1], 10),
+    }
+  }
+
   const remainder = text.slice(4).trim()
   if (!remainder) {
     return {
@@ -177,5 +188,18 @@ export function formatMoAResponse({ moaResult, presetName }) {
     })
   }
 
+  if (moaResult?.allowCandidateOverride && refs.length > 1) {
+    const runId = moaResult?.runId || ''
+    parts.push('### 🔄 Candidate Override Actions')
+    parts.push('To apply an alternative candidate\'s files instead of the judge\'s pick:')
+    refs.forEach((ref, idx) => {
+      const isWinner = (idx + 1) === moaResult.winningIndex
+      const tag = isWinner ? ' *(Current Judge Pick)*' : ''
+      const promoteCmd = runId ? `/moa promote ${runId} ${idx + 1}` : `/moa promote ${idx + 1}`
+      parts.push(`- \`${promoteCmd}\` — Candidate ${idx + 1} (${ref.label})${tag}`)
+    })
+    parts.push('')
+  }
+
   return parts.join('\n')
 }

package/lib/moa-prompts.js

@@ -33,6 +33,38 @@ Penalize and strictly downgrade candidates exhibiting any of the following flaws
 8. 🚫 Context Amnesia & Regressions:
    - Dropping or breaking previously functioning project features while adding new code.`
 
+export const ROLE_PERSONA_PROMPTS = {
+  minimalist: `### 🎯 Specialized Engineering Persona: THE MINIMALIST (Ponytail)
+- Prioritize the standard library and native platform capabilities over third-party dependencies.
+- Zero boilerplate, no premature abstractions, interfaces, or unneeded layers.
+- Write the shortest, cleanest working solution with maximum readability.`,
+
+  robustness: `### 🎯 Specialized Engineering Persona: ROBUSTNESS & DEFENSIVE DESIGN
+- Emphasize paranoid input validation, boundary checking, and comprehensive error handling.
+- Anticipate network failures, null/undefined edge cases, malformed data, and race conditions.
+- Ensure grace under failure and informative, actionable error reporting.`,
+
+  performance: `### 🎯 Specialized Engineering Persona: HIGH PERFORMANCE & EFFICIENCY
+- Prioritize asymptotic time and space complexity, minimal memory allocations, and zero redundant work.
+- Optimize hot paths, streaming data processing, and efficient algorithmic structures.
+- Document the Big-O complexity and performance characteristics of your solution.`,
+
+  tester: `### 🎯 Specialized Engineering Persona: TESTABILITY & VERIFICATION
+- Structure code for maximum testability, modular decoupling, and clear side-effect boundaries.
+- Include comprehensive unit/integration test specifications verifying happy paths and edge cases.
+- Emphasize deterministic behavior and assertions.`,
+
+  general: '',
+}
+
+export const SYNTAX_CORRECTION_DIRECTIVE = `### ⚠️ Syntax Warning & Auto-Fix Directive:
+Some candidate proposals have detected syntax flaws (annotated with [⚠️ Syntax Warning]).
+CRITICAL JUDGE DIRECTIVE: If a candidate with a syntax flaw presents superior architecture, algorithm, or engineering logic compared to other candidates, DO NOT reject them solely for this syntax flaw!
+Instead, you MUST correct the syntax error directly during synthesis/assembly and declare that candidate the winner.`
+
+export const LANGUAGE_MIRRORING_DIRECTIVE = `### 🌐 Language Mirroring Requirement
+CRITICAL: You MUST write your entire analysis, reasoning, verdict, explanations and instructions in the EXACT SAME LANGUAGE as the user's prompt (e.g. Russian if the prompt is written in Russian, English if in English). Do NOT switch or translate to English unless explicitly requested by the user.`
+
 /**
  * Formats a provider + model slot into a readable string key.
  */
@@ -141,15 +173,50 @@ The user gave the task:
 The advisors proposed the following decision points and clarifications:
 ${joined}
 
+${LANGUAGE_MIRRORING_DIRECTIVE}
+
 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.`
 }
 
 /**
- * Builds the curator synthesis prompt with component analysis and model recommendation.
+ * Builds the curator synthesis prompt with component analysis, model recommendation, and optional blind review.
  */
-export function buildCuratorSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '') {
+/**
+ * Builds the Round 2 Consilium / Peer Critique prompt for candidate models.
+ */
+export function buildPeerCritiquePrompt(userPrompt, myProposal, otherProposals = [], isBlind = true) {
+  const othersText = otherProposals
+    .map((p, idx) => {
+      const label = isBlind ? `Candidate ${idx + 1}` : (p.label || `Candidate ${idx + 1}`)
+      return `### ${label} Alternative Proposal:
+${p.text}`
+    })
+    .join('\n\n---\n\n')
+
+  return `You are participating in Round 2 (Consilium / Peer Critique & Solution Refinement) of an MoA ensemble.
+
+Original User Request:
+${userPrompt}
+
+Your Initial Solution (Round 1):
+${myProposal}
+
+Alternative Solutions from Peer Candidates:
+${othersText}
+
+Instructions:
+1. 🔍 Peer Critique: Briefly evaluate the other candidates' solutions. Point out any hidden bugs, race conditions, antipatterns, or edge-case omissions in their logic.
+2. 🛡️ Defend & Refine: Adopt the strongest architectural ideas or optimizations from competitors to enhance your own solution. Fix any oversights in your own code.
+3. 🚀 Final Refined Deliverable: Present your complete, production-ready, improved code with full explanations.
+
+${LANGUAGE_MIRRORING_DIRECTIVE}`
+}
+
+export function buildCuratorSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '', options = {}) {
+  const isBlind = Boolean(options.blindEvaluation)
+
   const joined = referenceOutputs
     .map((r, i) => {
       const fileSummary = (r.files && r.files.length > 0)
@@ -159,7 +226,11 @@ export function buildCuratorSynthesisPrompt(userPrompt, referenceOutputs = [], j
       if (referenceOutputs.length >= 3 && textContent.length > 3000) {
         textContent = stripOrSummarizeCode(textContent)
       }
-      return `Candidate ${i + 1} — ${r.label}:${fileSummary}\n${textContent}`
+      const syntaxNote = r.syntaxWarning ? ` [⚠️ Syntax Warning: ${r.syntaxWarning}]` : ''
+      const header = isBlind
+        ? `Candidate ${i + 1}:${syntaxNote}${fileSummary}`
+        : `Candidate ${i + 1} — ${r.label}:${syntaxNote}${fileSummary}`
+      return `${header}\n${textContent}`
     })
     .join('\n\n')
 
@@ -178,8 +249,12 @@ ${joined}
 
 ${ANTIPATTERNS_RUBRIC}
 
+${LANGUAGE_MIRRORING_DIRECTIVE}
+${referenceOutputs.some((r) => r.syntaxWarning) ? `
+${SYNTAX_CORRECTION_DIRECTIVE}
+` : ''}
 Instructions:
-Your response MUST be structured into three clear parts (respond in the same language as the user's prompt, e.g. Russian):
+Your response MUST be structured into three clear parts:
 
 ### 1. 🔍 Curator Analysis & Component Breakdown
 - For EACH candidate, provide:
@@ -189,7 +264,7 @@ Your response MUST be structured into three clear parts (respond in the same lan
 
 ### 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>)
+  RECOMMENDED_ASSEMBLER: <number from 1 to N> ${isBlind ? '' : '(<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.
@@ -200,13 +275,15 @@ Your response MUST be structured into three clear parts (respond in the same lan
 }
 
 /**
- * Builds the judge synthesis prompt for standard or curator mode.
+ * Builds the judge synthesis prompt for standard or curator mode with optional blind evaluation.
  */
 export function buildSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCriteria = '', options = {}) {
   if (options.curatorSynthesis) {
-    return buildCuratorSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria)
+    return buildCuratorSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, options)
   }
 
+  const isBlind = Boolean(options.blindEvaluation)
+
   const joined = referenceOutputs
     .map((r, i) => {
       const fileSummary = (r.files && r.files.length > 0)
@@ -216,7 +293,11 @@ export function buildSynthesisPrompt(userPrompt, referenceOutputs = [], judgeCri
       if (referenceOutputs.length >= 3 && textContent.length > 3000) {
         textContent = stripOrSummarizeCode(textContent)
       }
-      return `Reference ${i + 1} — ${r.label}:${fileSummary}\n${textContent}`
+      const syntaxNote = r.syntaxWarning ? ` [⚠️ Syntax Warning: ${r.syntaxWarning}]` : ''
+      const header = isBlind
+        ? `Reference ${i + 1}:${syntaxNote}${fileSummary}`
+        : `Reference ${i + 1} — ${r.label}:${syntaxNote}${fileSummary}`
+      return `${header}\n${textContent}`
     })
     .join('\n\n')
 
@@ -234,11 +315,13 @@ ${joined}
 
 ${ANTIPATTERNS_RUBRIC}
 
+${LANGUAGE_MIRRORING_DIRECTIVE}
+${referenceOutputs.some((r) => r.syntaxWarning) ? `\n${SYNTAX_CORRECTION_DIRECTIVE}\n` : ''}
 Instructions:
-Your response MUST be structured into three clear parts (respond in the same language as the user's prompt, e.g. Russian):
+Your response MUST be structured into three clear parts:
 
 ### 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").
+- **Winner**: clearly name the chosen candidate (e.g. "Winner: Candidate 1" or "Reference 1 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.

package/lib/moa-runner.js

@@ -11,53 +11,16 @@
  */
 
 import path from 'node:path'
-import {
-  extractFileBlocks,
-  collectProjectContext,
-  isRefinementTask,
-  writeCandidateWorkspace,
-  promoteCandidateWorkspace,
-  cleanMoaWorkspaces,
-} from './file-workspace.js'
+import crypto from 'node:crypto'
+import { extractFileBlocks, collectProjectContext, isRefinementTask, writeCandidateWorkspace, promoteCandidateWorkspace, cleanMoaWorkspaces, verifyFileSyntax } from './file-workspace.js'
 import { estimateTokenCost } from './pricing.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'
+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'
 
+// Re-exports for consumers & backward 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.'
@@ -94,8 +57,6 @@ 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
-  const fullMessages = [{ role: 'system', content: systemPrompt }, ...advisoryMessages]
   const timeoutMs = options.timeoutMs ?? ((options.timeoutSec ?? 60) * 1000)
   const maxRetries = options.maxRetries ?? 0
   const quorumEnabled = Boolean(options.quorumEnabled)
@@ -117,6 +78,10 @@ export async function runReferencesParallel(references, messages, options = {},
 
   references.forEach((slot, i) => {
     const label = slotLabel(slot)
+    const persona = slot?.role_persona && ROLE_PERSONA_PROMPTS[slot.role_persona]
+      ? `\n\n${ROLE_PERSONA_PROMPTS[slot.role_persona]}`
+      : ''
+    const slotMessages = [{ role: 'system', content: systemPrompt + persona }, ...advisoryMessages]
 
     const runOne = async () => {
       try {
@@ -125,7 +90,7 @@ export async function runReferencesParallel(references, messages, options = {},
           {
             provider: slot.provider,
             model: slot.model,
-            messages: fullMessages,
+            messages: slotMessages,
             temperature: options.temperature ?? 0.6,
             maxTokens: options.maxTokens ?? 4096,
             timeoutMs,
@@ -144,7 +109,7 @@ export async function runReferencesParallel(references, messages, options = {},
         const text = typeof res === 'string' ? res : (res?.content || res?.text || '')
 
         const fallbackUsage = (typeof res === 'object' && res?.usage) ? res.usage : {
-          inputTokens: Math.max(1, Math.round(fullMessages.map((m) => m.content).join('').length / 4)),
+          inputTokens: Math.max(1, Math.round(slotMessages.map((m) => m.content).join('').length / 4)),
           outputTokens: Math.max(1, Math.round(text.length / 4)),
         }
         const costInfo = estimateTokenCost(slot, fallbackUsage, options.prices)
@@ -327,6 +292,10 @@ export async function runMoAPipeline({
   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
+  const isBlindEvaluation = Boolean(preset?.blind_evaluation)
+  const isPeerCritiqueEnabled = Boolean(preset?.peer_critique_enabled)
+  const allowCandidateOverride = Boolean(preset?.allow_candidate_override)
+  const runId = crypto.randomUUID()
 
   // 2. Collect project context for refinement tasks
   const isRefinement = isRefinementTask(userPrompt)
@@ -530,8 +499,50 @@ export async function runMoAPipeline({
     }
   }
 
+  // 7b. Consilium Round 2 (Peer Critique) if enabled
+  if (isPeerCritiqueEnabled && successfulRefs.length > 1) {
+    if (typeof onProgress === 'function') {
+      onProgress('🤝 *Consilium Round 2: Candidates reviewing peer proposals in parallel...*\n')
+    }
+    const r2Promises = successfulRefs.map(async (cand) => {
+      const opponents = successfulRefs
+        .filter((c) => c.index !== cand.index)
+        .map((c) => ({ label: isBlindEvaluation ? `Candidate ${c.index}` : c.label, text: c.text }))
+      const r2Prompt = buildPeerCritiquePrompt(userPrompt, cand.text, opponents, isBlindEvaluation)
+      try {
+        const slot = referenceModels[cand.index - 1] || { provider: cand.provider, model: cand.model }
+        const res = await callWithTransientRetry(callLlm, {
+          provider: slot.provider,
+          model: slot.model,
+          messages: [{ role: 'user', content: r2Prompt }],
+          temperature: refTemp,
+          maxTokens,
+          timeoutMs: refTimeoutSec * 1000,
+        }, candidateRetries)
+        const refinedText = typeof res === 'string' ? res : (res?.content || res?.text || cand.text)
+        cand.text = refinedText
+        const r2Files = extractFileBlocks(refinedText)
+        if (r2Files.length > 0) {
+          cand.files = r2Files
+          cand.syntaxWarning = (verifyFileSyntax(r2Files) || []).map((w) => `${w.file}: ${w.error}`).join('; ')
+        }
+        if (typeof res === 'object' && res?.usage) {
+          cand.usage.inputTokens = (cand.usage.inputTokens || 0) + (res.usage.inputTokens || 0)
+          cand.usage.outputTokens = (cand.usage.outputTokens || 0) + (res.usage.outputTokens || 0)
+          cand.usage.totalTokens = (cand.usage.totalTokens || 0) + (res.usage.totalTokens || 0)
+          const extraCost = estimateTokenCost(slot, res.usage, prices)
+          cand.costUsd = Number(((cand.costUsd || 0) + extraCost.costUsd).toFixed(5))
+        }
+      } catch {}
+    })
+    await Promise.allSettled(r2Promises)
+  }
+
   // 8. Synthesis phase via primary judge or fallback chain
-  const synthesisPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, { curatorSynthesis: isCuratorSynthesis })
+  const synthesisPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, {
+    curatorSynthesis: isCuratorSynthesis,
+    blindEvaluation: isBlindEvaluation,
+  })
   let synthesizedText = ''
   let aggUsage = { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUsd: 0 }
   let chosenJudge = primaryJudge
@@ -600,15 +611,16 @@ export async function runMoAPipeline({
   const synthesizedFiles = extractFileBlocks(synthesizedText)
   let promotedFiles = []
 
+  const keepWorkspaces = allowCandidateOverride
   if (synthesizedFiles.length > 0) {
     await writeCandidateWorkspace(cwd, 'curator-synthesis', synthesizedFiles)
-    promotedFiles = await promoteCandidateWorkspace(cwd, 'curator-synthesis')
+    promotedFiles = await promoteCandidateWorkspace(cwd, 'curator-synthesis', { keepMoa: keepWorkspaces })
   } else if (referenceOutputs[winningIndex - 1]?.files?.length > 0) {
-    promotedFiles = await promoteCandidateWorkspace(cwd, winningIndex)
+    promotedFiles = await promoteCandidateWorkspace(cwd, winningIndex, { keepMoa: keepWorkspaces })
   } else if (successfulRefs[0]?.files?.length > 0) {
     const fallbackIdx = successfulRefs[0].index
-    promotedFiles = await promoteCandidateWorkspace(cwd, fallbackIdx)
-  } else {
+    promotedFiles = await promoteCandidateWorkspace(cwd, fallbackIdx, { keepMoa: keepWorkspaces })
+  } else if (!keepWorkspaces) {
     await cleanMoaWorkspaces(cwd)
   }
 
@@ -625,6 +637,7 @@ export async function runMoAPipeline({
   // Record run to history
   try {
     recordMoaRun({
+      id: runId,
       prompt: userPrompt,
       preset: preset?.name || 'default',
       isRefinement,
@@ -654,6 +667,8 @@ export async function runMoAPipeline({
     winningIndex,
     winnerModel,
     promotedFiles,
+    runId,
+    allowCandidateOverride,
     ...(livePreview ? { liveCanvas: livePreview } : {}),
     usage: {
       totalTokens,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-moa",
-  "version": "0.2.11",
+  "version": "0.2.13",
   "description": "Mixture of Agents (MoA) plugin for DeepSeek Harness with /moa slash command",
   "type": "module",
   "main": "lib/index.js",