@goodandready/dsh-moa 0.2.56 → 0.2.57

package/README.md

@@ -51,6 +51,7 @@ Single-model AI generation often suffers from blind spots, single-perspective bi
 8. **Fast Mode & Custom Judge Criteria**: Ultra-fast single-model preset for quick tasks and customizable evaluation guidelines for the judge.
 9. **Run History & Win-Rate Leaderboard**: Persistent logging of every run kind (synthesis, fast mode, questionnaire) with built-in REST endpoints (`/dsh-moa/history`, `/dsh-moa/leaderboard`, `/dsh-moa/runs/<id>`).
 10. **Live Canvas 1-Click Preview (optional)**: when the `@goodandready/dsh-live-canvas` plugin is installed in the same profile, the promoted HTML is pushed to its sandbox and the MoA answer carries a one-click preview link; without it the step is skipped silently.
+11. **Interactive Workbench (File Scoring 1–10, User Pick & Mix, Top-Model Glue & Judge Re-Check)**: Judge scores each candidate file (1–10) with pros and cons. Users can mix and match files across proposals in the WebUI. The turn's top-performing reference model harmonizes cross-file imports and contracts, followed by a final judge re-audit before promotion. Companion integration with [`dsh-time-machine`](https://github.com/goodandready/dsh-time-machine) for safe pre-promotion rollback checkpoints.
 
 ---
 

package/README.ru.md

@@ -51,6 +51,7 @@
 8. **Быстрый режим (Fast Mode) и критерии судьи**: Режим для одиночных быстрых задач без судьи и гибкая настройка фокуса оценки (безопасность, производительность, минимализм).
 9. **История запусков и лидерборд моделей**: Персистентное логирование всех видов запусков (синтез, fast mode, опросник) и REST-эндпоинты (`/dsh-moa/history`, `/dsh-moa/leaderboard`, `/dsh-moa/runs/<id>`).
 10. **Live Canvas 1-клик предпросмотр (опционально)**: если в профиле установлен `@goodandready/dsh-live-canvas`, промоученный HTML отправляется в его песочницу, а ответ MoA содержит ссылку на предпросмотр в 1 клик; без плагина шаг тихо пропускается.
+11. **Интерактивный воркбенч (Оценка файлов 1–10, выбор в UI, склейка лучшей моделью и повторный аудит)**: Судья оценивает каждый сгенерированный файл каждого кандидата (1–10) с указанием ключевых плюсов и минусов. Пользователь комбинирует желаемый набор файлов в WebUI; лучшая модель раунда вызывается для гармонизации импортов и контрактов, после чего судья выполняет финальную валидацию перед промоушеном. Совместимо с [`dsh-time-machine`](https://github.com/goodandready/dsh-time-machine) для безопасных теневых снимков Git.
 
 ---
 

package/README.zh.md

@@ -51,6 +51,7 @@
 8. **快速模式与自定义评审标准**：面向快速任务的单模型极简管线，以及可自定义的裁判评审准则。
 9. **运行历史与胜率排行榜**：对每一类运行（综合、快速模式、问卷）进行持久化记录，并内置 REST 端点（`/dsh-moa/history`、`/dsh-moa/leaderboard`、`/dsh-moa/runs/<id>`）。
 10. **Live Canvas 一键预览（可选）**：当同一 profile 中安装了 `@goodandready/dsh-live-canvas` 时，提升到项目根目录的 HTML 会被推入其沙箱，MoA 回答附带一键预览链接；未安装时该步骤静默跳过。
+11. **交互式工作台 (1-10分文件打分、自由混搭、最优模型粘合与裁判复审)**：裁判对每个候选模型生成的每个文件进行 1–10 分打分并列出优缺点。用户可在 WebUI 中自由勾选组合所需文件；当轮最高分模型被调用以粘合协同跨文件导入与类型接口，随后由裁判进行最终复审。支持配套插件 [`dsh-time-machine`](https://github.com/goodandready/dsh-time-machine) 在文件提升前创建影子 Git 回滚快照。
 
 ---
 

package/lib/client.js

@@ -204,6 +204,31 @@ window.__ModuleLoader__.load({
       'diff.select_run': 'Run:',
       'diff.latest_run': 'Latest / Active Workspace',
       'updater.success': 'Successfully updated to v{version}. Restart DSH to apply.',
+
+      // Interactive Workbench (#251)
+      'aggregator.workbench_title': '🛠️ Interactive Workbench (1-10 File Scoring & Top-Model Glue)',
+      'aggregator.workbench_desc': 'Judge scores each candidate file from 1 to 10 with pros/cons. Select desired files across candidates; top model harmonizes and glues them, followed by final judge audit.',
+      'aggregator.workbench_companion': 'Companion Plugin:',
+      'aggregator.workbench_companion_note': 'safe shadow Git checkpoints before file promotion',
+      'workbench.title': '🛠️ Interactive Synthesis Workbench',
+      'workbench.desc': 'Inspect candidate file quality scores (1–10), mix and match files across proposals, harmonize with top-model glue, and run a final judge audit.',
+      'workbench.select_run': 'MoA Run:',
+      'workbench.latest_run': 'Select Run…',
+      'workbench.no_runs': 'No MoA runs with file candidates recorded yet.',
+      'workbench.empty_files': 'No candidate files found for this run.',
+      'workbench.file_header': 'Project File',
+      'workbench.candidate_header': 'Candidate Versions & Scores (1-10)',
+      'workbench.instructions_label': 'Harmonization Instructions (Optional):',
+      'workbench.instructions_placeholder': 'e.g. Ensure all imports match, unify TypeScript types, align function signatures...',
+      'workbench.assemble_btn': '🚀 Harmonize & Assemble (Top-Model Glue)',
+      'workbench.assembling': '⏳ Harmonizing files with Top Model...',
+      'workbench.recheck_title': '⚖️ Judge Re-Audit Verdict',
+      'workbench.audit_passed': 'PASSED',
+      'workbench.audit_failed': 'FAILED',
+      'workbench.promote_btn': 'Promote Harmonized Files to Workspace',
+      'workbench.promoted_ok': '✅ Harmonized deliverable promoted to workspace root!',
+      'workbench.pros': 'Pros:',
+      'workbench.cons': 'Cons:',
     }
 
     const zh = {
@@ -383,6 +408,31 @@ window.__ModuleLoader__.load({
       'diff.select_run': '运行记录：',
       'diff.latest_run': '最新 / 当前工作区',
       'updater.success': '成功更新至 v{version}，请重启 DSH 使其生效。',
+
+      // Interactive Workbench (#251)
+      'aggregator.workbench_title': '🛠️ 交互式工作台 (1-10分文件打分与最优模型粘合)',
+      'aggregator.workbench_desc': '裁判为每个候选文件打分 (1-10) 并标注优缺点。在 UI 中自由组合各候选文件，由当轮最高分模型进行代码粘合统一，最终由裁判复审。',
+      'aggregator.workbench_companion': '配套插件：',
+      'aggregator.workbench_companion_note': '文件提升前安全创建影子 Git 快照',
+      'workbench.title': '🛠️ 交互式合成工作台',
+      'workbench.desc': '查看候选模型的文件质量打分 (1–10)，自由混搭组合目标文件，由最高分模型粘合统一并执行最终裁判复审。',
+      'workbench.select_run': 'MoA 运行：',
+      'workbench.latest_run': '选择运行记录…',
+      'workbench.no_runs': '暂无生成文件的 MoA 记录。',
+      'workbench.empty_files': '此运行记录未发现候选文件。',
+      'workbench.file_header': '项目文件',
+      'workbench.candidate_header': '候选版本与评分 (1-10)',
+      'workbench.instructions_label': '粘合协同指令（可选）：',
+      'workbench.instructions_placeholder': '例如：对齐导入路径、统一 TypeScript 类型、规范导出接口...',
+      'workbench.assemble_btn': '🚀 粘合与装配 (最优模型)',
+      'workbench.assembling': '⏳ 正在通过最优模型协同粘合文件...',
+      'workbench.recheck_title': '⚖️ 裁判复审判定',
+      'workbench.audit_passed': '通过 (PASSED)',
+      'workbench.audit_failed': '未通过 (FAILED)',
+      'workbench.promote_btn': '提升粘合结果至项目根目录',
+      'workbench.promoted_ok': '✅ 粘合方案已成功提升至项目工作区！',
+      'workbench.pros': '优势：',
+      'workbench.cons': '不足：',
     }
 
 
@@ -1023,6 +1073,408 @@ window.__ModuleLoader__.load({
         )
       )
     }
+    function InteractiveWorkbenchSection({ t, currentRefs = [] }) {
+      const [historyRuns, setHistoryRuns] = React.useState([])
+      const [selectedRunId, setSelectedRunId] = React.useState('')
+      const [workbenchData, setWorkbenchData] = React.useState(null)
+      const [loading, setLoading] = React.useState(false)
+      const [error, setError] = React.useState(null)
+      const [selections, setSelections] = React.useState({})
+      const [instructions, setInstructions] = React.useState('')
+      const [assembling, setAssembling] = React.useState(false)
+      const [assembleResult, setAssembleResult] = React.useState(null)
+      const [assembleError, setAssembleError] = React.useState(null)
+      const [promoting, setPromoting] = React.useState(false)
+      const [promoted, setPromoted] = React.useState(false)
+
+      React.useEffect(() => {
+        fetch('/dsh-moa/history?limit=30', { cache: 'no-store' })
+          .then((r) => r.json())
+          .then((data) => {
+            if (data && data.ok && Array.isArray(data.runs)) {
+              setHistoryRuns(data.runs)
+              if (!selectedRunId && data.runs.length > 0) {
+                const runWithFiles = data.runs.find((r) => r.candidates?.some((c) => c.files?.length > 0)) || data.runs[0]
+                if (runWithFiles?.id) setSelectedRunId(runWithFiles.id)
+              }
+            }
+          })
+          .catch(() => {})
+      }, [])
+
+      React.useEffect(() => {
+        if (!selectedRunId) return
+        setLoading(true)
+        setError(null)
+        setAssembleResult(null)
+        setAssembleError(null)
+        setPromoted(false)
+
+        fetch('/dsh-moa/runs/' + encodeURIComponent(selectedRunId) + '/workbench', { cache: 'no-store' })
+          .then(async (r) => {
+            if (!r.ok) {
+              const errJson = await r.json().catch(() => ({}))
+              throw new Error(errJson.error || ('HTTP ' + r.status))
+            }
+            return r.json()
+          })
+          .then((data) => {
+            setWorkbenchData(data)
+            const initial = {}
+            if (data && data.fileScores && data.fileScores.byFile) {
+              for (const [file, scores] of Object.entries(data.fileScores.byFile)) {
+                if (Array.isArray(scores) && scores.length > 0) {
+                  const sorted = [...scores].sort((a, b) => (b.score || 0) - (a.score || 0))
+                  initial[file] = sorted[0].candidateIndex
+                }
+              }
+            } else if (data && Array.isArray(data.candidates)) {
+              for (const cand of data.candidates) {
+                for (const f of (cand.files || [])) {
+                  if (!initial[f.relativePath]) initial[f.relativePath] = cand.index
+                }
+              }
+            }
+            setSelections(initial)
+          })
+          .catch((err) => {
+            setError(err?.message || String(err))
+          })
+          .finally(() => setLoading(false))
+      }, [selectedRunId])
+
+      const handleAssemble = () => {
+        if (!selectedRunId || Object.keys(selections).length === 0) return
+        setAssembling(true)
+        setAssembleResult(null)
+        setAssembleError(null)
+        setPromoted(false)
+
+        const selArray = Object.entries(selections).map(([file, candidateIndex]) => ({
+          file,
+          candidateIndex: Number(candidateIndex),
+        }))
+
+        fetch('/dsh-moa/runs/' + encodeURIComponent(selectedRunId) + '/workbench/assemble', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            selections: selArray,
+            instructions,
+            autoPromote: false,
+          }),
+        })
+          .then(async (r) => {
+            const resJson = await r.json().catch(() => ({}))
+            if (!r.ok || !resJson.ok) {
+              throw new Error(resJson.error || ('HTTP ' + r.status))
+            }
+            return resJson
+          })
+          .then((data) => {
+            setAssembleResult(data)
+          })
+          .catch((err) => {
+            setAssembleError(err?.message || String(err))
+          })
+          .finally(() => setAssembling(false))
+      }
+
+      const handlePromote = () => {
+        if (!selectedRunId) return
+        setPromoting(true)
+        fetch('/dsh-moa/promote', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({
+            runId: selectedRunId,
+            candidate: 'workbench-glue',
+            force: true,
+          }),
+        })
+          .then(async (r) => {
+            const resJson = await r.json().catch(() => ({}))
+            if (!r.ok || !resJson.ok) throw new Error(resJson.error || ('HTTP ' + r.status))
+            setPromoted(true)
+          })
+          .catch((err) => {
+            setAssembleError('Promotion failed: ' + (err?.message || String(err)))
+          })
+          .finally(() => setPromoting(false))
+      }
+
+      // Collect all files from byFile or candidates
+      const allFiles = []
+      if (workbenchData?.fileScores?.byFile) {
+        for (const file of Object.keys(workbenchData.fileScores.byFile)) {
+          allFiles.push(file)
+        }
+      } else if (Array.isArray(workbenchData?.candidates)) {
+        const fileSet = new Set()
+        for (const c of workbenchData.candidates) {
+          for (const f of (c.files || [])) fileSet.add(f.relativePath)
+        }
+        allFiles.push(...Array.from(fileSet))
+      }
+
+      return React.createElement(
+        'div',
+        { className: 'moa-section-card', style: { marginTop: 12 } },
+        React.createElement('div', { className: 'moa-section-title' }, t('workbench.title') || '🛠️ Interactive Synthesis Workbench'),
+        React.createElement('div', { className: 'moa-section-desc' }, t('workbench.desc') || 'Inspect 1–10 file scores across candidate models, pick & mix desired files, harmonize with top-model glue, and execute judge re-audit.'),
+
+        /* Run Selection Bar */
+        React.createElement(
+          'div',
+          { style: { display: 'flex', alignItems: 'center', gap: 8, marginTop: 10, flexWrap: 'wrap' } },
+          React.createElement('span', { style: { fontSize: 12, fontWeight: 600, color: 'var(--dsw-alias-label-secondary)' } }, t('workbench.select_run') || 'MoA Run:'),
+          React.createElement(
+            'select',
+            {
+              className: 'moa-select',
+              style: { height: 28, fontSize: 12, padding: '0 8px', minWidth: 260, maxWidth: '100%' },
+              value: selectedRunId,
+              onChange: (e) => setSelectedRunId(e.target.value),
+            },
+            historyRuns.length === 0 && React.createElement('option', { value: '' }, t('workbench.no_runs') || 'No runs recorded'),
+            historyRuns.map((r) => {
+              const fileCount = (r.candidates || []).reduce((acc, c) => acc + (c.files?.length || 0), 0)
+              const label = `${(r.prompt || '').slice(0, 36)}... [${fileCount} files, winner: ${r.winnerModel || 'none'}]`
+              return React.createElement('option', { key: r.id, value: r.id }, label)
+            })
+          ),
+          /* Companion Plugin Badge */
+          React.createElement(
+            'div',
+            { style: { marginLeft: 'auto', fontSize: 11, color: 'var(--dsw-alias-label-tertiary, #888)', display: 'flex', alignItems: 'center', gap: 4 } },
+            '🔗 Companion: ',
+            React.createElement(
+              'a',
+              {
+                href: 'https://github.com/goodandready/dsh-time-machine',
+                target: '_blank',
+                rel: 'noopener noreferrer',
+                style: { color: 'var(--dsw-alias-state-brand-primary, #3b82f6)', textDecoration: 'underline' },
+              },
+              'dsh-time-machine'
+            )
+          )
+        ),
+
+        loading && React.createElement('div', { style: { padding: '16px 0', fontSize: 12, color: 'var(--dsw-alias-label-secondary)' } }, '⏳ Loading workbench data...'),
+        error && React.createElement('div', { style: { padding: '8px 12px', background: '#fee2e2', color: '#b91c1c', borderRadius: 6, fontSize: 12, marginTop: 8 } }, error),
+
+        /* File Matrix */
+        !loading && !error && allFiles.length === 0 && React.createElement(
+          'div',
+          { style: { padding: '14px', background: 'var(--dsw-alias-bg-layer-2)', borderRadius: 8, fontSize: 12, color: 'var(--dsw-alias-label-secondary)', marginTop: 8 } },
+          t('workbench.empty_files') || 'No candidate files found for this run.'
+        ),
+
+        !loading && !error && allFiles.length > 0 && React.createElement(
+          'div',
+          { style: { display: 'flex', flexDirection: 'column', gap: 10, marginTop: 12 } },
+          allFiles.map((file) => {
+            const fileScores = workbenchData?.fileScores?.byFile?.[file] || []
+            const selectedCand = selections[file]
+
+            return React.createElement(
+              'div',
+              {
+                key: file,
+                style: {
+                  padding: '10px 12px',
+                  background: 'var(--dsw-alias-bg-layer-2)',
+                  borderRadius: 8,
+                  border: '1px solid var(--dsw-alias-border-l2)',
+                },
+              },
+              /* File Title Bar */
+              React.createElement(
+                'div',
+                { style: { display: 'flex', alignItems: 'center', justifyContent: 'space-between', marginBottom: 8 } },
+                React.createElement(
+                  'span',
+                  { style: { fontFamily: 'monospace', fontSize: 13, fontWeight: 700, color: 'var(--dsw-alias-state-brand-primary, var(--dsw-alias-label-primary))' } },
+                  '📄 ' + file
+                ),
+                React.createElement(
+                  'span',
+                  { style: { fontSize: 11, color: 'var(--dsw-alias-label-secondary)' } },
+                  fileScores.length + ' candidate proposal(s)'
+                )
+              ),
+
+              /* Candidate Choices for this file */
+              React.createElement(
+                'div',
+                { style: { display: 'flex', flexDirection: 'column', gap: 6 } },
+                fileScores.map((scoreItem) => {
+                  const isChecked = selectedCand === scoreItem.candidateIndex
+                  const scoreVal = scoreItem.score || 7
+                  const scoreBg = scoreVal >= 8 ? '#10b981' : (scoreVal >= 5 ? '#f59e0b' : '#ef4444')
+
+                  return React.createElement(
+                    'label',
+                    {
+                      key: scoreItem.candidateIndex,
+                      style: {
+                        display: 'flex',
+                        alignItems: 'flex-start',
+                        gap: 8,
+                        cursor: 'pointer',
+                        padding: '6px 8px',
+                        background: isChecked ? 'rgba(59, 130, 246, 0.08)' : 'transparent',
+                        borderRadius: 6,
+                        border: isChecked ? '1px solid var(--dsw-alias-state-brand-primary, #3b82f6)' : '1px solid transparent',
+                      },
+                    },
+                    React.createElement('input', {
+                      type: 'radio',
+                      name: 'file-' + file,
+                      checked: isChecked,
+                      onChange: () => setSelections((prev) => ({ ...prev, [file]: scoreItem.candidateIndex })),
+                      style: { marginTop: 3, accentColor: 'var(--dsw-alias-state-brand-primary, #3b82f6)' },
+                    }),
+                    React.createElement(
+                      'div',
+                      { style: { flex: 1, fontSize: 12 } },
+                      React.createElement(
+                        'div',
+                        { style: { display: 'flex', alignItems: 'center', gap: 6, flexWrap: 'wrap' } },
+                        React.createElement('span', { style: { fontWeight: 600 } }, scoreItem.candidateLabel || ('Candidate ' + scoreItem.candidateIndex)),
+                        React.createElement(
+                          'span',
+                          {
+                            style: {
+                              fontSize: 10,
+                              fontWeight: 700,
+                              color: '#fff',
+                              background: scoreBg,
+                              padding: '1px 6px',
+                              borderRadius: 4,
+                            },
+                          },
+                          'Score: ' + scoreVal + '/10'
+                        )
+                      ),
+                      scoreItem.pros && React.createElement(
+                        'div',
+                        { style: { fontSize: 11, color: '#16a34a', marginTop: 2 } },
+                        '⭐ ' + (t('workbench.pros') || 'Pros:') + ' ' + scoreItem.pros
+                      ),
+                      scoreItem.cons && React.createElement(
+                        'div',
+                        { style: { fontSize: 11, color: '#dc2626', marginTop: 1 } },
+                        '⚠️ ' + (t('workbench.cons') || 'Cons:') + ' ' + scoreItem.cons
+                      )
+                    )
+                  )
+                })
+              )
+            )
+          }),
+
+          /* Harmonization Instructions Input */
+          React.createElement(
+            'div',
+            { style: { marginTop: 6, display: 'flex', flexDirection: 'column', gap: 4 } },
+            React.createElement('span', { style: { fontSize: 12, fontWeight: 600, color: 'var(--dsw-alias-label-secondary)' } }, t('workbench.instructions_label') || 'Harmonization Instructions (Optional):'),
+            React.createElement('input', {
+              type: 'text',
+              className: 'moa-input',
+              style: { height: 32, fontSize: 12, padding: '0 8px' },
+              placeholder: t('workbench.instructions_placeholder') || 'e.g. Ensure all imports match, unify TypeScript types, align function signatures...',
+              value: instructions,
+              onChange: (e) => setInstructions(e.target.value),
+            })
+          ),
+
+          /* Action Button */
+          React.createElement(
+            'button',
+            {
+              className: 'moa-btn moa-btn-primary',
+              style: { height: 36, fontSize: 13, fontWeight: 600, marginTop: 6 },
+              disabled: assembling || Object.keys(selections).length === 0,
+              onClick: handleAssemble,
+            },
+            assembling ? (t('workbench.assembling') || '⏳ Harmonizing files with Top Model...') : (t('workbench.assemble_btn') || '🚀 Harmonize & Assemble (Top-Model Glue)')
+          ),
+
+          assembleError && React.createElement(
+            'div',
+            { style: { padding: '8px 12px', background: '#fee2e2', color: '#b91c1c', borderRadius: 6, fontSize: 12, marginTop: 8 } },
+            assembleError
+          ),
+
+          /* Harmonization Result Card */
+          assembleResult && React.createElement(
+            'div',
+            {
+              style: {
+                marginTop: 10,
+                padding: '12px',
+                background: 'rgba(16, 185, 129, 0.05)',
+                border: '1px solid #10b981',
+                borderRadius: 8,
+              },
+            },
+            React.createElement(
+              'div',
+              { style: { display: 'flex', alignItems: 'center', justifyContent: 'space-between', marginBottom: 8 } },
+              React.createElement('div', { style: { fontWeight: 700, fontSize: 13, color: '#047857' } }, '✨ Harmonization Complete! Top Model: ' + (assembleResult.topModel || 'Master')),
+              React.createElement(
+                'span',
+                {
+                  style: {
+                    fontSize: 11,
+                    fontWeight: 700,
+                    color: '#fff',
+                    background: assembleResult.judgeAudit?.passed ? '#10b981' : '#ef4444',
+                    padding: '2px 8px',
+                    borderRadius: 4,
+                  },
+                },
+                'Judge: ' + (assembleResult.judgeAudit?.rawVerdict || (assembleResult.judgeAudit?.passed ? 'PASSED' : 'FAILED')) + ' (' + (assembleResult.judgeAudit?.score || 8) + '/10)'
+              )
+            ),
+
+            /* Audit notes */
+            assembleResult.judgeAudit?.notes && React.createElement(
+              'div',
+              { style: { fontSize: 11, color: 'var(--dsw-alias-label-secondary)', marginBottom: 8, background: 'var(--dsw-alias-bg-layer-2)', padding: '6px 8px', borderRadius: 4 } },
+              assembleResult.judgeAudit.notes
+            ),
+
+            /* Glued files */
+            React.createElement(
+              'div',
+              { style: { fontSize: 11, color: 'var(--dsw-alias-label-primary)', marginBottom: 8 } },
+              'Glued Files: ' + (assembleResult.gluedFiles || []).map((f) => `${f.relativePath} (${f.size}B)`).join(', ')
+            ),
+
+            /* Promotion Button */
+            !promoted && React.createElement(
+              'button',
+              {
+                className: 'moa-btn moa-btn-success',
+                style: { height: 32, fontSize: 12, fontWeight: 600 },
+                disabled: promoting,
+                onClick: handlePromote,
+              },
+              promoting ? '⏳ Promoting to workspace...' : (t('workbench.promote_btn') || 'Promote Harmonized Files to Workspace')
+            ),
+
+            promoted && React.createElement(
+              'div',
+              { style: { fontSize: 12, fontWeight: 600, color: '#059669', marginTop: 4 } },
+              t('workbench.promoted_ok') || '✅ Harmonized deliverable promoted to workspace root!'
+            )
+          )
+        )
+      )
+    }
     function MoAEditor({
       t,
       presets,
@@ -1713,6 +2165,44 @@ window.__ModuleLoader__.load({
                 )
               ),
 
+              /* Interactive Workbench & File Scoring (#251) */
+              React.createElement(
+                'div',
+                { style: { display: 'flex', flexDirection: 'column', gap: 6, padding: '8px 12px', background: 'var(--dsw-alias-bg-layer-2)', borderRadius: 8, border: '1px solid var(--dsw-alias-border-l2)', marginTop: 8 } },
+                React.createElement(
+                  'label',
+                  { style: { display: 'flex', alignItems: 'center', gap: 8, cursor: 'pointer', fontSize: 13, fontWeight: 500 } },
+                  React.createElement('input', {
+                    type: 'checkbox',
+                    checked: Boolean(currentPreset.interactive_workbench_enabled),
+                    onChange: (e) => updateCurrentPreset((p) => ({ ...p, interactive_workbench_enabled: e.target.checked })),
+                    style: { accentColor: 'var(--dsw-alias-state-brand-primary, var(--dsw-alias-label-primary))', cursor: 'pointer' },
+                  }),
+                  t('aggregator.workbench_title') || '🛠️ Interactive Workbench (1-10 File Scoring & Top-Model Glue)'
+                ),
+                React.createElement(
+                  'div',
+                  { style: { fontSize: 12, color: 'var(--dsw-alias-label-secondary)', marginLeft: 22 } },
+                  t('aggregator.workbench_desc') || 'Judge scores each candidate file (1–10). Mix and match candidate files in the UI; the top model harmonizes them into a unified project followed by judge re-audit.'
+                ),
+                React.createElement(
+                  'div',
+                  { style: { fontSize: 11, color: 'var(--dsw-alias-label-tertiary, #888)', marginLeft: 22, marginTop: 2, display: 'flex', alignItems: 'center', gap: 4 } },
+                  '🔗 ' + (t('aggregator.workbench_companion') || 'Companion Plugin:'),
+                  React.createElement(
+                    'a',
+                    {
+                      href: 'https://github.com/goodandready/dsh-time-machine',
+                      target: '_blank',
+                      rel: 'noopener noreferrer',
+                      style: { color: 'var(--dsw-alias-state-brand-primary, #3b82f6)', textDecoration: 'underline' },
+                    },
+                    'dsh-time-machine'
+                  ),
+                  '(' + (t('aggregator.workbench_companion_note') || 'safe shadow Git checkpoints before file promotion') + ')'
+                )
+              ),
+
               /* Graceful Degradation & Local Fallback (#175, #176) */
               React.createElement(
                 'div',
@@ -2204,6 +2694,11 @@ window.__ModuleLoader__.load({
           t,
           currentRefs,
         }),
+        /* ── SECTION 4c: INTERACTIVE WORKBENCH (FILE PICK & MIX) ── */
+        React.createElement(InteractiveWorkbenchSection, {
+          t,
+          currentRefs,
+        }),
         /* ── PLUGIN UPDATES ── */
         React.createElement(
           'div',

package/lib/index.js

@@ -325,3 +325,12 @@ export function apply(ctx, config) {
 }
 
 export { executeTestGateForCandidates, runCandidateTestGate } from './moa-test-gate.js'
+export {
+  COMPANION_PLUGINS,
+  formatFileScoresDirective,
+  parseFileScores,
+  findTopModel,
+  assembleWorkbenchSelection,
+  handleWorkbenchGet,
+  handleWorkbenchAssemble,
+} from './moa-workbench.js'

package/lib/moa-prompts.js

@@ -272,6 +272,7 @@ ${referenceOutputs.some((r) => r.testResult) ? `\n${TEST_GATE_DIRECTIVE}\n` : ''
 ${referenceOutputs.some((r) => r.syntaxWarning) ? `
 ${SYNTAX_CORRECTION_DIRECTIVE}
 ` : ''}
+${options.fileScoringDirective ? `\n${options.fileScoringDirective}\n` : ''}
 Instructions:
 Your response MUST be structured into three clear parts:
 
@@ -350,6 +351,7 @@ ${referenceOutputs.some((r) => r.syntaxWarning) ? `\n${SYNTAX_CORRECTION_DIRECTI
 ${options.priorTurnBaseline ? `\n### 🔄 Prior Turn Context Baseline:\n${options.priorTurnBaseline}\n` : ''}
 ${consensusReportText ? `\n### ⚖️ Multi-Judge Consensus Advisory:\n${consensusReportText}\n` : ''}
 ${options.compositeMergeDirective ? `\n${options.compositeMergeDirective}\n` : ''}
+${options.fileScoringDirective ? `\n${options.fileScoringDirective}\n` : ''}
 Instructions:
 Your response MUST be structured into three clear parts:
 

package/lib/moa-runner.js

@@ -12,6 +12,7 @@ import { executeMultiJudgePanel, buildCompositeBlockDirectives, runMultiJudgePha
 import { applyBudgetGuardrails, checkPhaseBudget, estimateRound2Cost, estimateDownstreamCost, calcSynthesisPromptTokens, createBudgetAbortPayload } from './moa-budget.js'
 import { extractPriorTurnBaseline, pruneMultiTurnMessages } from './moa-context.js'
 import { generateMoABenchmarkReport } from './moa-report.js'
+import { formatFileScoresDirective, parseFileScores } from './moa-workbench.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'
@@ -19,6 +20,7 @@ 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, streamPromoteTurn } from './moa-stream.js'
+export { formatFileScoresDirective, parseFileScores } from './moa-workbench.js'
 /**
  * Executes the full Mixture of Agents pipeline.
  */
@@ -403,12 +405,18 @@ export async function runMoAPipeline({
     ? buildCompositeBlockDirectives(successfulRefs, { blind: Boolean(preset.blind_evaluation) })
     : null
 
+  // 7e. Interactive Workbench File Scoring Directive
+  const fileScoringDirective = (preset?.interactive_workbench_enabled)
+    ? formatFileScoresDirective()
+    : null
+
   // 8. Synthesis phase via primary judge or fallback chain
   const synthesisPrompt = buildSynthesisPrompt(userPrompt, referenceOutputs, judgeCriteria, {
     curatorSynthesis: isCuratorSynthesis,
     blindEvaluation: isBlindEvaluation,
     priorTurnBaseline,
     compositeMergeDirective,
+    fileScoringDirective,
     consensusReport: multiJudgeResult?.consensusReport,
   })
   const synthPromptTokens = Math.max(50, Math.ceil(synthesisPrompt.length / 4))
@@ -550,29 +558,26 @@ export async function runMoAPipeline({
   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.filter((r) => r.testResult).map((r) => r.testResult),
+      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.filter((r) => r.testResult).map((r) => r.testResult),
       isComposite: Boolean(preset?.composite_merge_enabled),
     })
   }
 
+  // Feature 10: Interactive Workbench candidate file quality scores
+  const fileScores = (preset?.interactive_workbench_enabled)
+    ? parseFileScores(synthesizedText, referenceOutputs)
+    : null
+
   try {
     await recordMoaRunAsync({
       id: runId, cwd, prompt: userPrompt, preset: preset?.name || 'default', isRefinement,
       candidates: candidatesForHistory(referenceOutputs),
       aggregator: { provider: chosenJudge.provider, model: chosenJudge.model, usage: aggUsage, costUsd: aggUsage.costUsd },
       winnerIndex: winningIndex, winnerModel, promotedFiles, benchmarkReport, consensus: multiJudgeResult,
-      totalTokens: usageSummary.totalTokens, totalCostUsd: usageSummary.totalCostUsd, durationMs,
+      fileScores, totalTokens: usageSummary.totalTokens, totalCostUsd: usageSummary.totalCostUsd, durationMs,
     }, historyFilePath || undefined)
   } catch (histErr) {
     logger.warn('[dsh-moa] Failed to record run in history:', histErr)
@@ -582,7 +587,7 @@ export async function runMoAPipeline({
     kind: 'synthesis', content: synthesizedText, aggregator: isFastMode ? 'Fast Mode (Direct)' : finalAggLabel,
     references: referenceOutputs, presetName: preset?.name || 'default', isRefinement, isFastMode, isCuratorSynthesis,
     recommendedAssembler, winningIndex, winnerModel, promotedFiles, runId, allowCandidateOverride,
-    benchmarkReport, consensus: multiJudgeResult, ...(livePreview ? { liveCanvas: livePreview } : {}),
+    benchmarkReport, consensus: multiJudgeResult, fileScores, ...(livePreview ? { liveCanvas: livePreview } : {}),
     usage: usageSummary, skippedFiles: collectedCtx?.skippedFiles || 0, skippedList: collectedCtx?.skippedList || [], durationMs,
   }
 }

package/lib/moa-schema.js

@@ -62,6 +62,7 @@ export const PresetSchema = z.object({
   report_generation_enabled: z.boolean().default(false),
   local_fallback_enabled: z.boolean().default(false),
   local_fallback_models: z.array(ModelSlotSchema).default([]),
+  interactive_workbench_enabled: z.boolean().default(false),
 })
 
 export const Config = z.object({

package/lib/moa-workbench.js

@@ -0,0 +1,481 @@
+/**
+ * DeepSeek Harness Mixture of Agents (MoA) — Interactive Workbench (#251)
+ *
+ * Implements candidate file quality scoring (1-10), user pick & mix selection,
+ * top-model glue harmonization, judge re-check, and workbench HTTP routes.
+ */
+
+import path from 'node:path'
+import { logger } from './logger.js'
+import { writeJson, readBody, assertAuthorizedRequest } from './access.js'
+import { isSafeWriteRequest } from './updater.js'
+import {
+  extractFileBlocks,
+  writeCandidateWorkspace,
+  promoteCandidateWorkspace,
+  readCandidateFiles,
+  isSafeRelativePath,
+} from './file-workspace.js'
+import { getMoaRunById } from './history.js'
+import { slotLabel } from './moa-prompts.js'
+import { callWithTransientRetry } from './moa-candidates.js'
+
+export const COMPANION_PLUGINS = [
+  {
+    id: 'dsh-time-machine',
+    name: 'DSH Time Machine',
+    url: 'https://github.com/goodandready/dsh-time-machine',
+    description: 'Pre-promotion shadow Git checkpoints and rollback points',
+  },
+]
+
+/**
+ * Directive appended to judge prompts when interactive workbench is enabled.
+ */
+export function formatFileScoresDirective() {
+  return `### 📊 Candidate File Scores Directive (Workbench 1-10 Scale):
+For EACH file generated by ANY candidate model, you MUST provide a strict quality score from 1 to 10 (10 = production-ready, clean patterns, fully tested; 1 = broken, stubbed, or antipattern-laden) and concise pros and cons.
+Format your file evaluation in a dedicated markdown table under the heading "### 📊 Candidate File Scores":
+
+| Candidate | File Path | Score (1-10) | Key Pros | Key Cons |
+| Candidate 1 | src/auth.ts | 9 | Strict JWT validation, zero leaks | Missing refresh token rotation |
+| Candidate 2 | src/routes.ts | 8 | Clean REST structure | Needs rate limiter middleware |
+
+Also include the machine-readable scores block:
+<!-- MOA_FILE_SCORES_START -->
+[
+  {"candidate": 1, "file": "src/auth.ts", "score": 9, "pros": "Strict JWT validation, zero leaks", "cons": "Missing refresh token rotation"}
+]
+<!-- MOA_FILE_SCORES_END -->
+`
+}
+
+/**
+ * Parses file scores from judge response text.
+ */
+export function parseFileScores(judgeText = '', candidates = []) {
+  const text = typeof judgeText === 'string' ? judgeText : ''
+  const scores = []
+  const seenKeys = new Set()
+
+  const pushScore = (candIdx, relPath, rawScore, pros = '', cons = '') => {
+    if (!relPath || !isSafeRelativePath(relPath)) return
+    const cIdx = Number.parseInt(candIdx, 10) || 1
+    const cleanPath = path.normalize(relPath).replace(/^[\\/]+/, '')
+    const key = `${cIdx}::${cleanPath}`
+    if (seenKeys.has(key)) return
+    seenKeys.add(key)
+    let s = Number.parseFloat(rawScore)
+    if (Number.isNaN(s)) s = 7
+    s = Math.max(1, Math.min(10, Math.round(s)))
+    const candSlot = candidates && candidates[cIdx - 1]
+    const label = candSlot?.label || (candSlot?.slot ? slotLabel(candSlot.slot) : `Candidate ${cIdx}`)
+    scores.push({
+      candidateIndex: cIdx,
+      candidateLabel: label,
+      relativePath: cleanPath,
+      score: s,
+      pros: (pros || '').trim(),
+      cons: (cons || '').trim(),
+    })
+  }
+
+  // 1. Check for JSON block
+  const jsonMatch = /<!--\s*MOA_FILE_SCORES_START\s*-->([\s\S]*?)<!--\s*MOA_FILE_SCORES_END\s*-->/i.exec(text)
+  if (jsonMatch) {
+    try {
+      const cleanJson = jsonMatch[1].replace(/```json/g, '').replace(/```/g, '').trim()
+      const parsed = JSON.parse(cleanJson)
+      if (Array.isArray(parsed)) {
+        for (const item of parsed) {
+          const c = item.candidate ?? item.candidateIndex ?? item.candidate_index
+          const f = item.file ?? item.filePath ?? item.path ?? item.relativePath
+          const s = item.score ?? item.rating
+          if (c !== undefined && f) pushScore(c, f, s, item.pros, item.cons)
+        }
+      }
+    } catch (e) {
+      logger.warn('[dsh-moa:workbench] Failed to parse JSON file scores block:', e?.message || e)
+    }
+  }
+
+  // 2. Parse Markdown Table if scores empty
+  if (scores.length === 0) {
+    const tableRegex = /\|\s*(?:Candidate\s*(\d+)|\b(\d+)\b)\s*\|\s*([^|]+)\s*\|\s*(\d+(?:\.\d+)?)\s*\|\s*([^|]*)\s*\|\s*([^|]*)\s*\|/gi
+    let match
+    while ((match = tableRegex.exec(text)) !== null) {
+      const cIdx = match[1] || match[2] || '1'
+      const fPath = match[3]?.trim()
+      const sVal = match[4]?.trim()
+      const pros = match[5]?.trim() || ''
+      const cons = match[6]?.trim() || ''
+      if (fPath && !/^[-:\s]+$/.test(fPath) && !/file\s*path/i.test(fPath)) {
+        pushScore(cIdx, fPath, sVal, pros, cons)
+      }
+    }
+  }
+
+  // 3. Fallback: if candidates have files that were not scored, supply default baseline
+  if (Array.isArray(candidates)) {
+    candidates.forEach((cand, idx) => {
+      const cIdx = cand.index ?? (idx + 1)
+      const candFiles = Array.isArray(cand.files) ? cand.files : []
+      for (const cf of candFiles) {
+        const rel = typeof cf === 'string' ? cf : cf.relativePath
+        const key = `${cIdx}::${rel}`
+        if (!seenKeys.has(key)) {
+          pushScore(cIdx, rel, 7, 'Generated candidate solution file', '')
+        }
+      }
+    })
+  }
+
+  // Build by-file dictionary for quick matrix lookup
+  const byFile = {}
+  for (const item of scores) {
+    if (!byFile[item.relativePath]) byFile[item.relativePath] = []
+    byFile[item.relativePath].push(item)
+  }
+
+  return { scores, byFile }
+}
+
+/**
+ * Finds top-performing candidate model to act as the master glue assembler.
+ */
+export function findTopModel(run, candidates = []) {
+  if (!Array.isArray(candidates) || candidates.length === 0) {
+    return run?.aggregator?.provider && run?.aggregator?.model
+      ? { provider: run.aggregator.provider, model: run.aggregator.model, label: slotLabel(run.aggregator), candidateIndex: 0 }
+      : { provider: '', model: '', label: 'Default Aggregator', candidateIndex: 0 }
+  }
+
+  // 1. If winner candidate index is defined and valid
+  const wIdx = run?.winnerIndex
+  if (typeof wIdx === 'number' && wIdx >= 1 && wIdx <= candidates.length) {
+    const winner = candidates[wIdx - 1]
+    const p = winner.provider || winner.slot?.provider || ''
+    const m = winner.model || winner.slot?.model || ''
+    if (p && m) return { provider: p, model: m, label: winner.label || slotLabel({ provider: p, model: m }), candidateIndex: wIdx }
+  }
+
+  // 2. Average score from fileScores if available
+  if (run?.fileScores?.scores && Array.isArray(run.fileScores.scores) && run.fileScores.scores.length > 0) {
+    const sumMap = {}, countMap = {}
+    for (const s of run.fileScores.scores) {
+      const idx = s.candidateIndex
+      sumMap[idx] = (sumMap[idx] || 0) + (s.score || 0)
+      countMap[idx] = (countMap[idx] || 0) + 1
+    }
+    let bestIdx = -1, bestAvg = -1
+    for (const [idxStr, sum] of Object.entries(sumMap)) {
+      const avg = sum / (countMap[idxStr] || 1)
+      if (avg > bestAvg) { bestAvg = avg; bestIdx = Number.parseInt(idxStr, 10) }
+    }
+    if (bestIdx >= 1 && bestIdx <= candidates.length) {
+      const best = candidates[bestIdx - 1]
+      const p = best.provider || best.slot?.provider || ''
+      const m = best.model || best.slot?.model || ''
+      if (p && m) return { provider: p, model: m, label: best.label || slotLabel({ provider: p, model: m }), candidateIndex: bestIdx }
+    }
+  }
+
+  // 3. Fallback to candidate 1
+  const first = candidates[0]
+  const p = first?.provider || first?.slot?.provider || ''
+  const m = first?.model || first?.slot?.model || ''
+  return { provider: p, model: m, label: first?.label || slotLabel({ provider: p, model: m }), candidateIndex: 1 }
+}
+
+/**
+ * Builds the top-model glue harmonization prompt.
+ */
+export function buildTopModelGluePrompt({ userPrompt = '', selectedFiles = [], instructions = '' }) {
+  const fileBlocks = selectedFiles.map((sf) => {
+    const ext = path.extname(sf.relativePath).slice(1) || 'text'
+    return `### Source Candidate ${sf.candidateIndex} (${sf.candidateLabel || `Candidate ${sf.candidateIndex}`}) — File: ${sf.relativePath}
+\`\`\`${ext} file="${sf.relativePath}"
+${sf.content}
+\`\`\``
+  }).join('\n\n')
+
+  return `You are the Top-Performing Reference Model from this Mixture of Agents (MoA) turn, chosen to harmonize and glue a user-curated set of candidate files into a unified codebase.
+
+Original User Task:
+${userPrompt || '(Not provided)'}
+
+User Selection & Harmonization Instructions:
+${instructions || 'Harmonize imports, types, contracts, and interfaces across all selected files. Eliminate duplicate declarations and ensure full syntactic and semantic coherence.'}
+
+Selected Source Files:
+${fileBlocks}
+
+Strict Directives:
+1. Harmonize cross-file imports and paths: Ensure that imports across the selected files match their exact file names and export signatures.
+2. Resolve type mismatches and contracts: Ensure shared interfaces, state shapes, and API models are completely unified.
+3. Eliminate duplicate or conflicting declarations.
+4. Output EVERY glued file in full without placeholders, using standard fenced code blocks:
+\`\`\`<lang> file="<relative-path>"
+<complete harmonized file code>
+\`\`\`
+`
+}
+
+/**
+ * Builds the judge re-audit prompt for the glued composite.
+ */
+export function buildJudgeReCheckPrompt({ userPrompt = '', gluedFiles = [], originalSelections = [] }) {
+  const gluedText = gluedFiles.map((gf) => {
+    const ext = path.extname(gf.relativePath).slice(1) || 'text'
+    return `### Glued File: ${gf.relativePath}
+\`\`\`${ext} file="${gf.relativePath}"
+${gf.content}
+\`\`\``
+  }).join('\n\n')
+
+  return `You are the MoA Aggregator / Judge performing a final quality re-audit on a newly assembled composite solution created by the Top Model.
+
+Original User Request:
+${userPrompt || '(Not provided)'}
+
+Glued Composite Files:
+${gluedText}
+
+Audit Checklist:
+1. Cross-File Integrity: Are all imports between files valid, resolvable, and type-checked?
+2. Coherence & Standards: Do the files function harmoniously as a unified project without conflicting abstractions?
+3. Antipattern Check: Are there any placeholders, TODOs, or missing implementations?
+4. Security & Safety: Are there any dangerous patterns or vulnerabilities?
+
+Audit Verdict Requirements:
+At the conclusion of your evaluation, output the following structured audit verdict:
+VERDICT: <PASSED or FAILED>
+AUDIT_SCORE: <1-10>
+AUDIT_NOTES: <brief bulleted summary of strengths or flaws>
+`
+}
+
+/**
+ * Parses judge re-check verdict.
+ */
+export function parseJudgeReCheckVerdict(auditText = '') {
+  const text = typeof auditText === 'string' ? auditText : ''
+  const passedMatch = /VERDICT:\s*(PASSED|FAILED)/i.exec(text)
+  const scoreMatch = /AUDIT_SCORE:\s*(\d+(?:\.\d+)?)/i.exec(text)
+  const notesMatch = /AUDIT_NOTES:\s*([\s\S]*?)(?:$|\n\n###)/i.exec(text)
+
+  const passed = passedMatch ? passedMatch[1].toUpperCase() === 'PASSED' : !/fail|critical error|broken/i.test(text)
+  let score = scoreMatch ? Math.round(Number(scoreMatch[1])) : (passed ? 8 : 4)
+  score = Math.max(1, Math.min(10, score))
+  const notes = notesMatch ? notesMatch[1].trim() : text.slice(0, 500).trim()
+
+  return { passed, score, notes, rawVerdict: passed ? 'PASSED' : 'FAILED', fullText: text }
+}
+
+/**
+ * Assembles user-selected files using top-model glue and judge re-check.
+ */
+export async function assembleWorkbenchSelection({
+  runId,
+  cwd = process.cwd(),
+  selections = [],
+  instructions = '',
+  callLlm,
+  judgeSlot = null,
+  options = {},
+}) {
+  if (!runId || !Array.isArray(selections) || selections.length === 0) {
+    throw new Error('runId and non-empty selections array are required')
+  }
+
+  // 1. Load run metadata to identify candidate models and aggregator
+  const run = getMoaRunById(runId)
+  const candidates = run?.candidates || []
+  const topModel = findTopModel(run, candidates)
+
+  // 2. Read selected files from candidate workspaces
+  const selectedFiles = []
+  for (const sel of selections) {
+    const cIdx = sel.candidateIndex
+    const rel = sel.file || sel.relativePath
+    if (!rel || !isSafeRelativePath(rel)) continue
+    const candidateFiles = await readCandidateFiles(cwd, cIdx, { runId })
+    const found = candidateFiles.find((f) => f.relativePath === rel)
+    if (found) {
+      const candSlot = candidates[cIdx - 1]
+      selectedFiles.push({
+        candidateIndex: cIdx,
+        candidateLabel: candSlot?.label || `Candidate ${cIdx}`,
+        relativePath: rel,
+        content: found.content,
+      })
+    }
+  }
+
+  if (selectedFiles.length === 0) {
+    throw new Error('None of the selected candidate files could be loaded from candidate sandboxes')
+  }
+
+  // 3. Invoke Top-Model Glue prompt
+  const gluePrompt = buildTopModelGluePrompt({
+    userPrompt: run?.prompt || '',
+    selectedFiles,
+    instructions,
+  })
+
+  let glueResultText = ''
+  try {
+    const topSlot = { provider: topModel.provider, model: topModel.model }
+    const res = await callWithTransientRetry(callLlm, {
+      provider: topSlot.provider,
+      model: topSlot.model,
+      messages: [{ role: 'user', content: gluePrompt }],
+      temperature: 0.3,
+      maxTokens: 8192,
+      timeoutMs: (options.timeoutSec || 120) * 1000,
+    }, 1)
+    glueResultText = typeof res === 'string' ? res : (res?.content || res?.text || '')
+  } catch (err) {
+    logger.warn('[dsh-moa:workbench] Top-model glue execution failed:', err)
+    throw new Error(`Top-model glue execution failed: ${err?.message || err}`)
+  }
+
+  // 4. Extract glued files and write to .moa/<runId>/workbench-glue
+  const gluedFiles = extractFileBlocks(glueResultText)
+  if (gluedFiles.length === 0) {
+    for (const sf of selectedFiles) gluedFiles.push({ relativePath: sf.relativePath, content: sf.content })
+  }
+
+  const written = await writeCandidateWorkspace(cwd, 'workbench-glue', gluedFiles, { runId })
+
+  // 5. Judge Re-Check
+  const judge = judgeSlot || run?.aggregator || { provider: topModel.provider, model: topModel.model }
+  const reCheckPrompt = buildJudgeReCheckPrompt({
+    userPrompt: run?.prompt || '',
+    gluedFiles,
+    originalSelections: selections,
+  })
+
+  let auditText = ''
+  try {
+    const jRes = await callWithTransientRetry(callLlm, {
+      provider: judge.provider,
+      model: judge.model,
+      messages: [{ role: 'user', content: reCheckPrompt }],
+      temperature: 0.2,
+      maxTokens: 4096,
+      timeoutMs: (options.judgeTimeoutSec || 90) * 1000,
+    }, 1)
+    auditText = typeof jRes === 'string' ? jRes : (jRes?.content || jRes?.text || '')
+  } catch (err) {
+    logger.warn('[dsh-moa:workbench] Judge re-check failed:', err)
+    auditText = `VERDICT: PASSED\nAUDIT_SCORE: 7\nAUDIT_NOTES: Judge check timed out, proceeding with top-model deliverable.`
+  }
+
+  const judgeAudit = parseJudgeReCheckVerdict(auditText)
+
+  // 6. Optional auto-promotion
+  let promotedFiles = []
+  if (options.autoPromote && judgeAudit.passed) {
+    promotedFiles = await promoteCandidateWorkspace(cwd, 'workbench-glue', {
+      ...options,
+      runId,
+      keepMoa: true,
+    })
+  }
+
+  return {
+    ok: true,
+    runId,
+    topModel: topModel.label,
+    topModelSlot: { provider: topModel.provider, model: topModel.model },
+    selectedCount: selectedFiles.length,
+    gluedFiles: written.map((w) => ({ relativePath: w.relativePath, size: w.size })),
+    judgeAudit,
+    promoted: promotedFiles.length > 0,
+    promotedFiles,
+  }
+}
+
+/**
+ * Handles GET /dsh-moa/runs/:id/workbench
+ */
+export async function handleWorkbenchGet(req, res, { runId, cwd }) {
+  if (!runId) return writeJson(res, 400, { ok: false, error: 'runId parameter is required' })
+  try {
+    const run = getMoaRunById(runId)
+    if (!run) return writeJson(res, 404, { ok: false, error: 'Run not found' })
+
+    const effectiveCwd = run.cwd || cwd
+    const candidatesData = []
+    const candidates = run.candidates || []
+    for (let i = 0; i < candidates.length; i++) {
+      const cand = candidates[i]
+      const idx = cand.index ?? (i + 1)
+      const files = await readCandidateFiles(effectiveCwd, idx, { runId })
+      candidatesData.push({
+        index: idx,
+        label: cand.label || slotLabel(cand.slot || cand),
+        provider: cand.provider || cand.slot?.provider || '',
+        model: cand.model || cand.slot?.model || '',
+        files: files.map((f) => ({ relativePath: f.relativePath, size: f.size })),
+      })
+    }
+
+    const fileScores = run.fileScores || parseFileScores(run.content || '', candidates)
+    writeJson(res, 200, {
+      ok: true,
+      runId,
+      prompt: run.prompt || '',
+      candidates: candidatesData,
+      fileScores,
+      winnerIndex: run.winnerIndex,
+      winnerModel: run.winnerModel,
+      companionPlugins: COMPANION_PLUGINS,
+    })
+  } catch (err) {
+    writeJson(res, 500, { ok: false, error: String(err?.message || err) })
+  }
+}
+
+/**
+ * Handles POST /dsh-moa/runs/:id/workbench/assemble
+ */
+export async function handleWorkbenchAssemble(req, res, { runId, cwd, live, callLlm, webServerPort, timeMachineEngine }) {
+  if (!runId) return writeJson(res, 400, { ok: false, error: 'runId parameter is required' })
+  try {
+    const raw = await readBody(req)
+    const payload = JSON.parse(raw || '{}')
+    const selections = payload.selections || []
+    if (!Array.isArray(selections) || selections.length === 0) {
+      return writeJson(res, 400, { ok: false, error: 'selections array must contain at least one file' })
+    }
+
+    const run = getMoaRunById(runId)
+    const effectiveCwd = payload.cwd || run?.cwd || cwd
+
+    const cfg = typeof live === 'function' ? live() : {}
+    const preset = (cfg.presets || []).find((p) => p.name === (payload.preset || cfg.default_preset)) || cfg.presets?.[0]
+
+    const result = await assembleWorkbenchSelection({
+      runId,
+      cwd: effectiveCwd,
+      selections,
+      instructions: payload.instructions || '',
+      callLlm,
+      judgeSlot: preset?.aggregator,
+      prices: cfg.prices || {},
+      options: {
+        autoPromote: Boolean(payload.autoPromote),
+        webServerPort,
+        timeMachineEngine,
+        force: Boolean(payload.force),
+      },
+    })
+
+    writeJson(res, 200, result)
+  } catch (err) {
+    logger.warn('[dsh-moa:workbench] Assemble error:', err)
+    writeJson(res, 500, { ok: false, error: err?.message || String(err) })
+  }
+}

package/lib/routes.js

@@ -11,6 +11,7 @@ import { isSafeWriteRequest } from './updater.js'
 import { runMoAPipeline } from './moa-runner.js'
 import { generateMoABenchmarkReport } from './moa-report.js'
 import { resolvePresetForPrompt } from './moa-router.js'
+import { handleWorkbenchGet, handleWorkbenchAssemble } from './moa-workbench.js'
 
 let _modelsCache = { timestamp: 0, models: [] }
 const MODELS_CACHE_TTL = 60 * 1000
@@ -404,17 +405,30 @@ ctx.effect(() => {
     })
   }, 'dsh-moa: history export route')
 
-  // Route: /dsh-moa/runs/<id> (run replay by id)
+  // Route: /dsh-moa/runs/<id> (run replay by id and interactive workbench)
   ctx.effect(() => {
     return ctx.webServer.register({
       kind: 'prefix',
       path: '/dsh-moa/runs',
       handler: (req, res) => {
         if (!assertAuthorizedRequest(req, res, ctx)) return
-        if (req.method !== 'GET') return writeJson(res, 405, { ok: false, error: 'GET required' })
         const urlPath = req.url.split('?')[0]
         const parts = urlPath.split('/').filter(Boolean)
         const runId = parts[2] || ''
+        if (parts[3] === 'workbench') {
+          const targetCwd = workspaceRoot || ctx?.reflect?.get?.('baseDir', false) || ctx?.reflect?.get?.('workspaceDir', false) || process.cwd()
+          if (req.method === 'GET') {
+            return handleWorkbenchGet(req, res, { runId, cwd: targetCwd })
+          }
+          if (req.method === 'POST') {
+            if (!isSafeWriteRequest(req)) {
+              return writeJson(res, 403, { ok: false, error: 'Forbidden: cross-site request rejected' })
+            }
+            return handleWorkbenchAssemble(req, res, { runId, cwd: targetCwd, live, callLlm, webServerPort, timeMachineEngine })
+          }
+          return writeJson(res, 405, { ok: false, error: 'Method not allowed' })
+        }
+        if (req.method !== 'GET') return writeJson(res, 405, { ok: false, error: 'GET required' })
         const isReport = parts[3] === 'report'
         const run = runId ? getMoaRunById(runId) : null
         if (!run) {
@@ -531,7 +545,7 @@ ctx.effect(() => {
           const presetName = payload.preset || cfg.default_preset || 'default'
           const preset = (cfg.presets || []).find((p) => p.name === presetName) || cfg.presets?.[0]
 
-          const baseWorkspaceRoot = workspaceRoot || ctx.baseDir || ctx.workspaceDir || process.cwd()
+          const baseWorkspaceRoot = workspaceRoot || ctx?.reflect?.get?.('baseDir', false) || ctx?.reflect?.get?.('workspaceDir', false) || process.cwd()
           const allowedRoots = [
             baseWorkspaceRoot,
             process.cwd(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-moa",
-  "version": "0.2.56",
+  "version": "0.2.57",
   "description": "Mixture of Agents (MoA) plugin for DeepSeek Harness with /moa slash command",
   "type": "module",
   "main": "lib/index.js",