@dsh-suite/plugin-deus 0.3.0 → 0.4.1

package/README.md

@@ -28,6 +28,25 @@
    - 对话页 dock 出现 `⚓ 锚定维持中` 状态 chip；**漂移时变红**，点击发送重锚提示词（可开「漂移自动重锚」）；重锚轮日志以 `reanchor` 模式独立记录，便于对比重锚前后维持率。
 6. **未锚定引导（v0.3）** — 当前会话不在 deus/minimal 系 preset 时，dock 显示灰态 `◌ 非神版会话` chip，点击展开切换指引（把「25 工具会话注入≈0%」的实测结论落地为产品引导）。
 7. **统计导出与分享（v0.3）** — 面板一键导出 CSV、**复制 Markdown 摘要**（Wilson CI 表格 + 诚实声明 + 包链接，可直接发帖），并内嵌近期神版率迷你趋势图（最近 40 条日志分桶折线）。
+8. **提示词 A/B 实验台（v0.4）** — 设计文档的「纯噪音退路」转正：对话页 dock 的 🧪 **实验台**里输入**任意**提示词变体 A/B，各跑 N 次**真实会话注入**（每次试验自动点「新建会话」隔离上下文），宿主按提示词文本队列配对判定，产出**指纹分布对比表**——每类的 Wilson 95% CI、Δ(A−B)、两比例 z 检验与显著性标记。Settings 面板可回看/切换历史实验。deus 由此从「神模工具」升级为「提示词实验台」，神模只是第一个被测假设。
+
+## Theme & Skin Compatibility / 主题与皮肤兼容
+
+This plugin follows the DSH Web styling contract
+([`docs/web-styling.md`](https://github.com/deepseek-ai/deepseek-harness/blob/main/docs/web-styling.md)):
+every color routes through `--dsw-alias-*` design tokens (dark literals are kept
+only as fallbacks for hosts predating the alias table), with no per-theme CSS
+selectors.
+
+It also implements the [dsh-web Skin Center compatibility contract](https://github.com/zhu1090093659/dsh-web/tree/main/packages/skins/skin-center/contracts)
+at **L1 (token coverage)** and **L2 (semantic attributes)**: the settings panel
+root carries `data-dsh-plugin="plugin-deus"` + `data-dsh-surface="settings-modal"`,
+the composer dock root carries `data-dsh-surface="chat-input"`, and skin-relevant
+nodes are tagged with `data-dsh-part` (`banner` / `card` / `button-ghost` /
+`input` / `textarea` / `select` / `chip` / `panel`), so third-party skins get
+full restyle coverage without relying on unstable class names.
+
+本插件遵循 DSH Web 样式契约：全部颜色经 `--dsw-alias-*` 设计令牌引用（暗色字面量仅作旧宿主兜底），无按主题分支的 CSS 选择器；并实现 dsh-web 皮肤中心兼容契约 **L1（令牌覆盖）+ L2（语义属性）**——设置面板根节点输出 `data-dsh-plugin="plugin-deus"` + `data-dsh-surface="settings-modal"`，输入区 dock 根节点输出 `data-dsh-surface="chat-input"`，换肤相关节点标注 `data-dsh-part`，第三方皮肤无需依赖不稳定类名即可完整换肤。
 
 ## 安装 / Install
 
@@ -52,7 +71,7 @@ npx @deepseek-ai/dsh plugin --profile web add ./plugin-deus   # 在其父目录
 
 - 日志只写本地 `$DSH_HOME/deus-mode/log.jsonl`，不上传、可一键清空、可导出 CSV。
 - 全部是用户正常交互（prompt 注入 + 回复观测），无越权、无绕过。
-- 若实验证明无差异：本插件的身份即**通用极简提示词 A/B 测试台**，"神模"只是第一个被测假设。
+- 若实验证明无差异：本插件的身份即**通用极简提示词 A/B 测试台**，"神模"只是第一个被测假设。（v0.4 起实验台已内建：🧪 dock 芯片直接开跑。）
 
 ## 实测依据 / Measured basis
 

package/lib/client.js

@@ -45,6 +45,20 @@ window.__ModuleLoader__.load({
       unanchored: '◌ 非神版会话', unanchoredTitle: '当前会话不是锚定 preset——实测 25 工具会话注入触发率≈0%。点击查看怎么切换',
       guideText: '实测结论：standard（25 工具）会话里文本注入几乎无法触发神版（0%）。想用神版：新建会话时在 Agent preset 选择器选「神模扳机 · 窄锚 / 宽锚」（实测触发率 ~90% / ~65%）。',
       copyMd: '复制 Markdown 摘要', mdCopied: '已复制 ✓ 去发帖吧', trend: '近期神版率趋势',
+      bench: '实验台',
+      benchHint: '提示词 A/B 实验台：任意两个提示词变体各跑 N 次真实会话注入（每次自动开新会话隔离上下文），对比起手式指纹分布。Wilson 95% CI + 两比例 z 检验；小样本只作倾向信号。神模只是第一个被测假设。',
+      benchA: '变体 A 提示词', benchB: '变体 B 提示词', benchTimes: '每变体次数 N',
+      benchRun: '▶ 开始实验', benchAbort: '⏹ 中止', benchRunning: '实验进行中',
+      benchProgress: '进度', benchWaiting: '等待回复…', benchDone: '实验完成',
+      benchIsolationWarn: '⚠ 未找到「新建会话」按钮，本轮在同会话内连跑，上下文未隔离',
+      benchTimeout: '等待回复超时（240s），实验中止',
+      benchEmpty: '请先填写变体 A 与变体 B 的提示词',
+      benchReportTitle: '指纹分布对比', benchRunId: '实验编号', benchStarted: '开始', benchFinished: '结束',
+      colFp: '指纹', colDelta: 'Δ(A−B)', colZ: 'z', colP: 'p', colSig: '显著',
+      benchNone: '还没有实验记录。在对话页输入框上方的 ⚗ 芯片行打开「实验台」跑一轮。',
+      benchPickRun: '选择实验', benchRefresh: '刷新',
+      benchDefA: '用三句话向聪明的初中生解释「熵」。',
+      benchDefB: '你是一位世界顶尖科普作家，回答前先在脑内推演三遍。用三句话向聪明的初中生解释「熵」。',
     }
     const en = {
       nav: 'Deus Trigger', sub: 'Minimal-prompt trigger bench · inject → classify → stats',
@@ -72,31 +86,114 @@ window.__ModuleLoader__.load({
       unanchored: '◌ Not a deus session', unanchoredTitle: 'This session is not on an anchored preset — injection measured ≈0% effective on 25-tool sessions. Click for how to switch',
       guideText: 'Measured: text injection barely triggers god-mode in standard (25-tool) sessions (0%). To get it: start a new session and pick "Deus Trigger · narrow / wide anchor" in the Agent preset picker (~90% / ~65% measured trigger rate).',
       copyMd: 'Copy Markdown summary', mdCopied: 'Copied ✓ go post it', trend: 'Recent god-rate trend',
+      bench: 'A/B Bench',
+      benchHint: 'Prompt A/B bench: run any two prompt variants N times each via real session injection (a fresh session per trial isolates context), then compare opener-fingerprint distributions. Wilson 95% CI + two-proportion z-test; small samples are directional only. God-mode is just the first hypothesis under test.',
+      benchA: 'Variant A prompt', benchB: 'Variant B prompt', benchTimes: 'Trials per variant N',
+      benchRun: '▶ Run experiment', benchAbort: '⏹ Abort', benchRunning: 'Running',
+      benchProgress: 'Progress', benchWaiting: 'waiting for reply…', benchDone: 'Experiment done',
+      benchIsolationWarn: '⚠ "New session" button not found; trials ran in the same session without context isolation',
+      benchTimeout: 'Timed out waiting for reply (240s); aborted',
+      benchEmpty: 'Fill in both variant A and variant B prompts first',
+      benchReportTitle: 'Fingerprint distribution comparison', benchRunId: 'Run', benchStarted: 'Started', benchFinished: 'Finished',
+      colFp: 'fingerprint', colDelta: 'Δ(A−B)', colZ: 'z', colP: 'p', colSig: 'sig',
+      benchNone: 'No experiments yet. Open the "A/B Bench" from the ⚗ chip row above the composer on a conversation page.',
+      benchPickRun: 'Pick run', benchRefresh: 'Refresh',
+      benchDefA: 'Explain "entropy" to a bright middle-schooler in three sentences.',
+      benchDefB: 'You are a world-class science writer; rehearse the answer three times in your head before speaking. Explain "entropy" to a bright middle-schooler in three sentences.',
     }
 
+    // DSH Web 样式契约（docs/web-styling.md）：全部颜色经 --dsw-alias-* 令牌引用；
+    // 暗色字面量仅作宿主早于令牌表时的兜底。无按主题分支的选择器，皮肤可自由覆写。
     const S = {
-      banner: { background: '#2d2406', border: '1px solid #9e6a03', borderRadius: '8px', padding: '10px 12px', fontSize: '12px', color: '#e3b341', lineHeight: '1.6', marginBottom: '14px' },
-      status: { fontSize: '11px', color: '#8b949e', marginBottom: '10px' },
-      card: { background: '#161b22', border: '1px solid #30363d', borderRadius: '10px', padding: '14px', marginBottom: '12px' },
+      banner: { background: 'var(--dsw-alias-state-warn-tertiary, #2d2406)', border: '1px solid var(--dsw-alias-state-warn-primary, #9e6a03)', borderRadius: '8px', padding: '10px 12px', fontSize: '12px', color: 'var(--dsw-alias-state-warn-label, #e3b341)', lineHeight: '1.6', marginBottom: '14px' },
+      status: { fontSize: '11px', color: 'var(--dsw-alias-label-secondary, #8b949e)', marginBottom: '10px' },
+      card: { background: 'var(--dsw-alias-bg-layer-1, #161b22)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '10px', padding: '14px', marginBottom: '12px' },
       row: { display: 'flex', gap: '8px', alignItems: 'center', flexWrap: 'wrap', marginTop: '10px' },
-      btn: { background: 'transparent', border: '1px solid #30363d', borderRadius: '6px', color: '#8b949e', padding: '6px 12px', fontSize: '13px', cursor: 'pointer' },
-      btnActive: { background: '#21262d', border: '1px solid #8b949e', borderRadius: '6px', color: '#e6edf3', padding: '6px 12px', fontSize: '13px', cursor: 'pointer' },
-      chip: { background: '#0d1117', border: '1px solid #30363d', borderRadius: '20px', color: '#8b949e', padding: '2px 10px', fontSize: '11px' },
-      name: { fontSize: '14px', fontWeight: '700', color: '#e6edf3' },
-      desc: { fontSize: '12px', color: '#8b949e', marginTop: '4px', lineHeight: '1.5' },
-      h2: { fontSize: '13px', fontWeight: '600', color: '#e6edf3', margin: '18px 0 8px' },
-      table: { width: '100%', borderCollapse: 'collapse', fontSize: '12px', color: '#c9d1d9' },
-      th: { textAlign: 'left', padding: '6px 8px', borderBottom: '1px solid #30363d', color: '#8b949e', fontWeight: '600' },
-      td: { padding: '6px 8px', borderBottom: '1px solid #21262d' },
-      input: { background: '#0d1117', border: '1px solid #30363d', borderRadius: '6px', color: '#e6edf3', padding: '6px 8px', fontSize: '12px', width: '100%', boxSizing: 'border-box' },
-      mono: { fontFamily: 'monospace', fontSize: '12px', color: '#c9d1d9' },
+      btn: { background: 'transparent', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '6px', color: 'var(--dsw-alias-label-secondary, #8b949e)', padding: '6px 12px', fontSize: '13px', cursor: 'pointer' },
+      btnActive: { background: 'var(--dsw-alias-bg-layer-3, #21262d)', border: '1px solid var(--dsw-alias-label-secondary, #8b949e)', borderRadius: '6px', color: 'var(--dsw-alias-label-primary, #e6edf3)', padding: '6px 12px', fontSize: '13px', cursor: 'pointer' },
+      chip: { background: 'var(--dsw-alias-bg-layer-2, #0d1117)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '20px', color: 'var(--dsw-alias-label-secondary, #8b949e)', padding: '2px 10px', fontSize: '11px' },
+      name: { fontSize: '14px', fontWeight: '700', color: 'var(--dsw-alias-label-primary, #e6edf3)' },
+      desc: { fontSize: '12px', color: 'var(--dsw-alias-label-secondary, #8b949e)', marginTop: '4px', lineHeight: '1.5' },
+      h2: { fontSize: '13px', fontWeight: '600', color: 'var(--dsw-alias-label-primary, #e6edf3)', margin: '18px 0 8px' },
+      table: { width: '100%', borderCollapse: 'collapse', fontSize: '12px', color: 'var(--dsw-alias-label-primary, #c9d1d9)' },
+      th: { textAlign: 'left', padding: '6px 8px', borderBottom: '1px solid var(--dsw-alias-border-l1, #30363d)', color: 'var(--dsw-alias-label-secondary, #8b949e)', fontWeight: '600' },
+      td: { padding: '6px 8px', borderBottom: '1px solid var(--dsw-alias-border-l1, #21262d)' },
+      input: { background: 'var(--dsw-alias-bg-layer-2, #0d1117)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '6px', color: 'var(--dsw-alias-label-primary, #e6edf3)', padding: '6px 8px', fontSize: '12px', width: '100%', boxSizing: 'border-box' },
+      mono: { fontFamily: 'monospace', fontSize: '12px', color: 'var(--dsw-alias-label-primary, #c9d1d9)' },
+      benchWrap: { width: '100%', background: 'var(--dsw-alias-bg-layer-2, #0d1117)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '10px', padding: '10px 12px', marginBottom: '6px' },
+      benchTa: { width: '100%', minHeight: '52px', background: 'var(--dsw-alias-bg-layer-1, #161b22)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '8px', color: 'var(--dsw-alias-label-primary, #c9d1d9)', fontSize: '12px', padding: '6px 8px', boxSizing: 'border-box', resize: 'vertical' },
+      benchLabel: { fontSize: '11px', color: 'var(--dsw-alias-label-secondary, #8b949e)', margin: '6px 0 3px' },
+      benchTable: { width: '100%', borderCollapse: 'collapse', fontSize: '11px', marginTop: '6px' },
+      benchCell: { border: '1px solid var(--dsw-alias-border-l1, #21262d)', padding: '3px 6px', textAlign: 'right', whiteSpace: 'nowrap' },
       dock: { display: 'flex', gap: '6px', alignItems: 'center', flexWrap: 'wrap', padding: '4px 0' },
-      dockChip: { background: '#161b22', border: '1px solid #30363d', borderRadius: '16px', color: '#8b949e', padding: '3px 12px', fontSize: '12px', cursor: 'pointer' },
+      dockChip: { background: 'var(--dsw-alias-bg-layer-1, #161b22)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '16px', color: 'var(--dsw-alias-label-secondary, #8b949e)', padding: '3px 12px', fontSize: '12px', cursor: 'pointer' },
     }
 
+    // 指纹分类强调色：god=紫 / med=绿 / pure=蓝 / unknown=灰。dsw 无图表系列令牌，
+    // 取语义最近的别名 + 字面量兜底。状态色统一走 state-* 令牌。
+    const CLS_ACCENT = {
+      god: 'var(--dsw-alias-state-business-primary, #d2a8ff)',
+      med: 'var(--dsw-alias-state-success-primary, #3fb950)',
+      pure: 'var(--dsw-alias-label-primary-bluish, #79c0ff)',
+      unknown: 'var(--dsw-alias-label-tertiary, #8b949e)',
+    }
+    const STATE_OK = 'var(--dsw-alias-state-success-primary, #3fb950)'
+    const STATE_ERR = 'var(--dsw-alias-state-error-primary, #f85149)'
+    const STATE_WARN = 'var(--dsw-alias-state-warn-primary, #d29922)'
+
     const pct = (ci) => `${(ci.rate * 100).toFixed(0)}% [${(ci.low * 100).toFixed(0)}–${(ci.high * 100).toFixed(0)}]`
 
     // ── tier A: settings section panel ──────────────────────────────────────
+    // ── v0.4 实验台：报告拉取与 A/B 对比表（Dock 控制台与设置面板共用）─────
+    function fetchBenchReport(runId) {
+      const url = runId ? '/deus/bench/report?runId=' + encodeURIComponent(runId) : '/deus/bench/report'
+      return fetch(url).then((r) => r.json())
+    }
+
+    const BENCH_CLS_COLORS = CLS_ACCENT
+
+    function benchCellFmt(g, n) {
+      if (!g || !n) return '—'
+      return g.n + '/' + n + ' · ' + Math.round(100 * g.rate) + '% [' + Math.round(100 * g.low) + '–' + Math.round(100 * g.high) + ']'
+    }
+
+    function BenchTable(props) {
+      const report = props.report
+      const t = props.t
+      if (!report) return null
+      const A = report.order[0]
+      const B = report.order[1] || 'B'
+      const vA = report.variants[A]
+      const vB = report.variants[B]
+      const sig = (p) => (p == null ? '' : p < 0.01 ? '✱✱' : p < 0.05 ? '✱' : '')
+      return h('div', { 'data-deus': 'bench-report', style: { marginTop: '8px' } },
+        h('div', { style: S.status },
+          t('benchRunId') + ': ' + report.runId + ' · n=' + report.n +
+          ' · ' + t('benchStarted') + ' ' + String(report.startedAt).slice(11, 19) +
+          ' · ' + t('benchFinished') + ' ' + String(report.finishedAt).slice(11, 19)),
+        h('table', { style: S.benchTable },
+          h('thead', null, h('tr', null,
+            h('th', { style: { ...S.benchCell, textAlign: 'left' } }, t('colFp')),
+            h('th', { style: S.benchCell }, 'A (n=' + (vA ? vA.n : 0) + ')'),
+            h('th', { style: S.benchCell }, 'B (n=' + (vB ? vB.n : 0) + ')'),
+            h('th', { style: S.benchCell }, t('colDelta')),
+            h('th', { style: S.benchCell }, t('colZ')),
+            h('th', { style: S.benchCell }, t('colP')),
+            h('th', { style: S.benchCell }, t('colSig')))),
+          h('tbody', null, (report.comparison || []).map((row) => h('tr', { key: row.cls },
+            h('td', { style: { ...S.benchCell, textAlign: 'left', color: BENCH_CLS_COLORS[row.cls] || CLS_ACCENT.unknown } }, t('detected_' + row.cls)),
+            h('td', { style: S.benchCell }, benchCellFmt(row.a, vA && vA.n)),
+            h('td', { style: S.benchCell }, benchCellFmt(row.b, vB && vB.n)),
+            h('td', { style: S.benchCell }, row.delta == null ? '—' : (row.delta >= 0 ? '+' : '') + Math.round(row.delta * 100) + 'pp'),
+            h('td', { style: S.benchCell }, row.z == null ? '—' : Number(row.z).toFixed(2)),
+            h('td', { style: S.benchCell }, row.p == null ? '—' : row.p < 0.001 ? '<0.001' : Number(row.p).toFixed(3)),
+            h('td', { style: { ...S.benchCell, color: STATE_WARN } }, sig(row.p)),
+          )))),
+        vA && vA.prompt ? h('div', { style: { ...S.status, marginTop: '6px' } }, 'A: ' + JSON.stringify(vA.prompt)) : null,
+        vB && vB.prompt ? h('div', { style: S.status }, 'B: ' + JSON.stringify(vB.prompt)) : null,
+      )
+    }
+
     function Panel(props) {
       const t = props.t
       const [presets, setPresets] = useState([])
@@ -107,6 +204,12 @@ window.__ModuleLoader__.load({
       const [err, setErr] = useState('')
       const [copied, setCopied] = useState('')
       const [editing, setEditing] = useState(null) // {id, label_zh, prompt} draft
+      const [benchData, setBenchData] = useState(null) // { ok, report, runs }
+      const [benchRunId, setBenchRunId] = useState('')
+
+      function refreshBench(runId) {
+        fetchBenchReport(runId || undefined).then((d) => setBenchData(d)).catch(() => {})
+      }
 
       function refresh() {
         fetch('/deus/presets').then((r) => r.json()).then((d) => setPresets(d.presets || [])).catch(() => setErr(t('loadFail')))
@@ -114,6 +217,7 @@ window.__ModuleLoader__.load({
         fetch('/deus/log').then((r) => r.json()).then((d) => setEntries(d.entries || [])).catch(() => {})
         fetch('/deus/version').then((r) => r.json()).then(setVer).catch(() => {})
         fetch('/deus/anchor').then((r) => r.json()).then(setAnchor).catch(() => {})
+        refreshBench(benchRunId)
       }
       useEffect(() => { refresh() }, [])
 
@@ -156,49 +260,49 @@ window.__ModuleLoader__.load({
         setEditing(null)
       }
 
-      return h('div', { style: { maxWidth: '860px' } },
-        h('div', { style: { fontSize: '14px', fontWeight: '600', color: '#e6edf3', marginBottom: '4px' } }, '⚗ ' + t('nav')),
+      return h('div', { style: { maxWidth: '860px' }, 'data-dsh-plugin': 'plugin-deus', 'data-dsh-surface': 'settings-modal' },
+        h('div', { style: { fontSize: '14px', fontWeight: '600', color: 'var(--dsw-alias-label-primary, #e6edf3)', marginBottom: '4px' } }, '⚗ ' + t('nav')),
         h('div', { style: S.status }, t('sub')),
-        h('div', { style: S.banner }, '⚠️ ' + t('honesty')),
-        err ? h('div', { style: { color: '#f85149', fontSize: '12px', marginBottom: '10px' } }, err) : null,
+        h('div', { style: S.banner, 'data-dsh-part': 'banner' }, '⚠️ ' + t('honesty')),
+        err ? h('div', { style: { color: STATE_ERR, fontSize: '12px', marginBottom: '10px' } }, err) : null,
         ver ? h('div', { style: { ...S.status, fontFamily: 'monospace' } },
           `${t('dshVer')}: ${ver.dsh} · ${t('pluginVer')}: ${ver.plugin} · ${t('logPath')}: ${ver.logPath}`) : null,
 
         h('div', { style: S.h2 }, t('presets')),
         h('div', { style: S.status }, t('presetsHint')),
-        presets.map((p) => h('div', { key: p.id, style: S.card },
+        presets.map((p) => h('div', { key: p.id, style: S.card, 'data-dsh-part': 'card' },
           h('div', { style: { display: 'flex', alignItems: 'baseline', gap: '10px', flexWrap: 'wrap' } },
             h('span', { style: S.name }, p.label_zh),
-            h('span', { style: { fontSize: '11px', color: '#8b949e' } }, p.id),
-            p.builtin ? h('span', { style: S.chip }, 'built-in') : null,
+            h('span', { style: { fontSize: '11px', color: 'var(--dsw-alias-label-secondary, #8b949e)' } }, p.id),
+            p.builtin ? h('span', { style: S.chip, 'data-dsh-part': 'chip' }, 'built-in') : null,
           ),
           h('div', { style: { ...S.mono, marginTop: '6px' } }, p.prompt === null ? t('empty') : JSON.stringify(p.prompt)),
           h('div', { style: S.row },
-            h('button', { style: copied === p.id ? S.btnActive : S.btn, onClick: () => copyPrompt(p) },
+            h('button', { style: copied === p.id ? S.btnActive : S.btn, 'data-dsh-part': 'button-ghost', onClick: () => copyPrompt(p) },
               copied === p.id ? '📋 ' + t('copied') : '📋 ' + t('copy')),
-            h('button', { style: S.btn, onClick: () => setEditing({ id: p.id, label_zh: p.label_zh, prompt: p.prompt === null ? '' : p.prompt }) }, '✏️ ' + t('edit')),
-            h('button', { style: S.btn, onClick: () => saveAll(presets.filter((x) => x.id !== p.id)) }, '🗑 ' + t('del')),
+            h('button', { style: S.btn, 'data-dsh-part': 'button-ghost', onClick: () => setEditing({ id: p.id, label_zh: p.label_zh, prompt: p.prompt === null ? '' : p.prompt }) }, '✏️ ' + t('edit')),
+            h('button', { style: S.btn, 'data-dsh-part': 'button-ghost', onClick: () => saveAll(presets.filter((x) => x.id !== p.id)) }, '🗑 ' + t('del')),
           ),
         )),
-        editing ? h('div', { style: S.card },
+        editing ? h('div', { style: S.card, 'data-dsh-part': 'card' },
           h('div', { style: S.status }, t('editId')),
-          h('input', { style: S.input, value: editing.id, onChange: (e) => setEditing({ ...editing, id: e.target.value }) }),
+          h('input', { style: S.input, 'data-dsh-part': 'input', value: editing.id, onChange: (e) => setEditing({ ...editing, id: e.target.value }) }),
           h('div', { style: { ...S.status, marginTop: '8px' } }, t('editLabel')),
-          h('input', { style: S.input, value: editing.label_zh, onChange: (e) => setEditing({ ...editing, label_zh: e.target.value }) }),
+          h('input', { style: S.input, 'data-dsh-part': 'input', value: editing.label_zh, onChange: (e) => setEditing({ ...editing, label_zh: e.target.value }) }),
           h('div', { style: { ...S.status, marginTop: '8px' } }, t('editPrompt')),
-          h('textarea', { style: { ...S.input, minHeight: '60px' }, value: editing.prompt, onChange: (e) => setEditing({ ...editing, prompt: e.target.value }) }),
+          h('textarea', { style: { ...S.input, minHeight: '60px' }, 'data-dsh-part': 'textarea', value: editing.prompt, onChange: (e) => setEditing({ ...editing, prompt: e.target.value }) }),
           h('div', { style: S.row },
-            h('button', { style: S.btnActive, onClick: commitEdit }, '💾 ' + t('save')),
-            h('button', { style: S.btn, onClick: () => setEditing(null) }, t('cancel')),
+            h('button', { style: S.btnActive, 'data-dsh-part': 'button-ghost', onClick: commitEdit }, '💾 ' + t('save')),
+            h('button', { style: S.btn, 'data-dsh-part': 'button-ghost', onClick: () => setEditing(null) }, t('cancel')),
           ),
         ) : h('div', { style: S.row },
-          h('button', { style: S.btn, onClick: () => setEditing({ id: '', label_zh: '', prompt: '' }) }, '➕ ' + t('add')),
-          h('button', { style: S.btn, onClick: resetPresets }, '↩ ' + t('resetPresets')),
+          h('button', { style: S.btn, 'data-dsh-part': 'button-ghost', onClick: () => setEditing({ id: '', label_zh: '', prompt: '' }) }, '➕ ' + t('add')),
+          h('button', { style: S.btn, 'data-dsh-part': 'button-ghost', onClick: resetPresets }, '↩ ' + t('resetPresets')),
         ),
 
         h('div', { style: S.h2 }, t('stats')),
         h('div', { style: S.status }, t('statsHint')),
-        h('div', { style: { ...S.status, color: '#7d8590' } }, t('measured')),
+        h('div', { style: { ...S.status, color: 'var(--dsw-alias-label-tertiary, #7d8590)' } }, t('measured')),
         h(Sparkline, { entries }),
         entries.length >= 4 ? h('div', { style: { ...S.status, fontSize: '10px' } }, t('trend')) : null,
         stats && stats.modes.length > 0
@@ -209,7 +313,7 @@ window.__ModuleLoader__.load({
               h('tbody', null, stats.modes.map((m) => h('tr', { key: m.mode },
                 h('td', { style: S.td }, m.mode),
                 h('td', { style: S.td }, m.n),
-                h('td', { style: { ...S.td, color: '#d2a8ff' } }, pct(m.godCI)),
+                h('td', { style: { ...S.td, color: CLS_ACCENT.god } }, pct(m.godCI)),
                 h('td', { style: S.td }, pct(m.medCI)),
                 h('td', { style: S.td }, pct(m.pureCI)),
               ))),
@@ -219,6 +323,7 @@ window.__ModuleLoader__.load({
           h('a', { href: '/deus/log.csv', download: 'deus-log.csv', style: { ...S.btn, textDecoration: 'none', display: 'inline-block' } }, '⬇ ' + t('exportCsv')),
           stats ? h('button', {
             style: copied === '__md' ? S.btnActive : S.btn,
+            'data-dsh-part': 'button-ghost',
             onClick: async () => {
               try {
                 await navigator.clipboard.writeText(buildMarkdownSummary(stats))
@@ -226,7 +331,7 @@ window.__ModuleLoader__.load({
               } catch {}
             },
           }, '📋 ' + (copied === '__md' ? t('mdCopied') : t('copyMd'))) : null,
-          h('button', { style: { ...S.btn, color: '#f85149', borderColor: '#f85149' }, onClick: resetLog }, '🗑 ' + t('resetLog')),
+          h('button', { style: { ...S.btn, color: STATE_ERR, borderColor: STATE_ERR }, 'data-dsh-part': 'button-ghost', onClick: resetLog }, '🗑 ' + t('resetLog')),
         ),
 
         h('div', { style: S.h2 }, t('anchor')),
@@ -234,7 +339,7 @@ window.__ModuleLoader__.load({
         anchor && anchor.agentPresets && h('div', { style: S.card },
           h('div', { style: S.status }, t('anchorPresetInstalled')),
           anchor.agentPresets.map((p) => h('div', { key: p.id, style: { display: 'flex', gap: '8px', alignItems: 'center', marginTop: '6px' } },
-            h('span', { style: { ...S.chip, color: p.installed ? '#3fb950' : '#f85149', borderColor: p.installed ? '#3fb950' : '#f85149' } }, p.installed ? '✓' : '✗'),
+            h('span', { style: { ...S.chip, color: p.installed ? STATE_OK : STATE_ERR, borderColor: p.installed ? STATE_OK : STATE_ERR }, 'data-dsh-part': 'chip' }, p.installed ? '✓' : '✗'),
             h('span', { style: S.name }, p.name || p.id),
             p.reason ? h('span', { style: S.status }, '(' + p.reason + ')') : null,
           )),
@@ -249,22 +354,39 @@ window.__ModuleLoader__.load({
                 h('td', { style: { ...S.td, ...S.mono } }, s.sessionId.slice(0, 8)),
                 h('td', { style: S.td }, s.preset),
                 h('td', { style: S.td }, s.total),
-                h('td', { style: { ...S.td, color: '#d2a8ff' } }, s.total ? Math.round(100 * s.god / s.total) + '%' : '—'),
+                h('td', { style: { ...S.td, color: CLS_ACCENT.god } }, s.total ? Math.round(100 * s.god / s.total) + '%' : '—'),
                 h('td', { style: S.td }, s.lastFp ? t('detected_' + s.lastFp) : '—'),
-                h('td', { style: { ...S.td, color: s.drifted ? '#f85149' : '#3fb950' } }, s.drifted ? '⚠ ' + t('drifted') : '⚓ ' + t('anchoredOn')),
+                h('td', { style: { ...S.td, color: s.drifted ? STATE_ERR : STATE_OK } }, s.drifted ? '⚠ ' + t('drifted') : '⚓ ' + t('anchoredOn')),
               ))),
             )
           : h('div', { style: S.status }, t('anchorNone')),
 
+        h('div', { style: S.h2 }, '🧪 ' + t('bench')),
+        h('div', { style: S.status }, t('benchHint')),
+        h('div', { style: S.row },
+          h('button', { style: S.btn, 'data-deus': 'bench-refresh', 'data-dsh-part': 'button-ghost', onClick: () => refreshBench(benchRunId) }, '🔄 ' + t('benchRefresh')),
+          benchData && benchData.runs && benchData.runs.length > 0 ? h('select', {
+            style: { ...S.input, width: 'auto' }, 'data-deus': 'bench-runs', 'data-dsh-part': 'select',
+            value: benchRunId,
+            onChange: (e) => { setBenchRunId(e.target.value); refreshBench(e.target.value) },
+          },
+            h('option', { value: '' }, t('benchPickRun')),
+            benchData.runs.map((r) => h('option', { key: r.runId, value: r.runId }, r.runId + ' · n=' + r.n)),
+          ) : null,
+        ),
+        benchData && benchData.report
+          ? h(BenchTable, { report: benchData.report, t })
+          : h('div', { style: S.status, 'data-deus': 'bench-empty' }, t('benchNone')),
+
         h('div', { style: S.h2 }, t('recent')),
         h('div', { style: S.status }, t('recentHint')),
         entries.length === 0 ? h('div', { style: S.status }, t('noData'))
           : entries.slice(0, 10).map((e, i) => h('div', { key: i, style: { ...S.card, padding: '10px 12px' } },
               h('div', { style: { display: 'flex', gap: '8px', alignItems: 'baseline', flexWrap: 'wrap' } },
-                h('span', { style: S.chip }, e.prompt_mode),
-                h('span', { style: { ...S.chip, color: e.detected === 'god' ? '#d2a8ff' : e.detected === 'pure' ? '#79c0ff' : e.detected === 'med' ? '#3fb950' : '#8b949e' } },
+                h('span', { style: S.chip, 'data-dsh-part': 'chip' }, e.prompt_mode),
+                h('span', { style: { ...S.chip, color: e.detected === 'god' ? CLS_ACCENT.god : e.detected === 'pure' ? CLS_ACCENT.pure : e.detected === 'med' ? CLS_ACCENT.med : CLS_ACCENT.unknown }, 'data-dsh-part': 'chip' },
                   t('detected_' + (e.detected || 'unknown'))),
-                h('span', { style: { fontSize: '11px', color: '#8b949e' } }, e.ts),
+                h('span', { style: { fontSize: '11px', color: 'var(--dsw-alias-label-tertiary, #8b949e)' } }, e.ts),
               ),
               h('div', { style: { ...S.mono, marginTop: '6px', whiteSpace: 'pre-wrap' } }, e.first_sentence),
             )),
@@ -297,6 +419,129 @@ window.__ModuleLoader__.load({
       const [sessPreset, setSessPreset] = useState(null)
       const [showGuide, setShowGuide] = useState(false)
       const lastReanchorKey = useRef('')
+
+      // ── v0.4 实验台（A/B bench）──────────────────────────────────────────
+      const [benchOpen, setBenchOpen] = useState(false)
+      const [benchA, setBenchA] = useState(() => {
+        try { return window.localStorage.getItem('deus.benchA') || t('benchDefA') } catch { return t('benchDefA') }
+      })
+      const [benchB, setBenchB] = useState(() => {
+        try { return window.localStorage.getItem('deus.benchB') || t('benchDefB') } catch { return t('benchDefB') }
+      })
+      const [benchN, setBenchN] = useState(3)
+      const [benchRunning, setBenchRunning] = useState(false)
+      const [benchProg, setBenchProg] = useState(null) // { a, b, n, label, waiting }
+      const [benchReport, setBenchReport] = useState(null)
+      const [benchNote, setBenchNote] = useState('')
+      const benchAbort = useRef(false)
+      const benchRunIdRef = useRef('')
+      const sidRef = useRef(sessionId)
+      const actionsRef = useRef(actions)
+      useEffect(() => { sidRef.current = sessionId; actionsRef.current = actions })
+
+      function sleep(ms) { return new Promise((r) => setTimeout(r, ms)) }
+
+      // 每次试验开一个新会话以隔离上下文：点侧栏「新建会话」按钮，
+      // 然后等 dock 的 sessionId prop 变化（运行时重渲染本会带回新 id）。
+      async function benchNewSession(prevSid) {
+        const sels = [
+          'button[aria-label="新建会话"]', 'button[aria-label="New session"]',
+          '[role="button"][aria-label="新建会话"]', '[role="button"][aria-label="New session"]',
+        ]
+        let btn = null
+        for (const sel of sels) { btn = document.querySelector(sel); if (btn) break }
+        if (!btn) {
+          btn = Array.from(document.querySelectorAll('button, [role="button"]'))
+            .find((b) => /新建会话|New session/i.test((b.getAttribute('aria-label') || '') + ' ' + (b.title || '')))
+        }
+        if (!btn) return false
+        btn.click()
+        const t0 = Date.now()
+        while (Date.now() - t0 < 10000) {
+          await sleep(300)
+          if (String(sidRef.current || '') !== String(prevSid || '')) return true
+        }
+        return false
+      }
+
+      async function abortBench() {
+        benchAbort.current = true
+        try {
+          await fetch('/deus/bench/clear', {
+            method: 'POST', headers: { 'content-type': 'application/json' },
+            body: JSON.stringify({ runId: benchRunIdRef.current }),
+          })
+        } catch { /* best-effort */ }
+      }
+
+      async function runBench() {
+        const aText = benchA.trim()
+        const bText = benchB.trim()
+        if (!aText || !bText) { setBenchNote(t('benchEmpty')); return }
+        setBenchRunning(true); setBenchNote(''); setBenchReport(null)
+        benchAbort.current = false
+        const runId = 'bench-' + Date.now().toString(36)
+        benchRunIdRef.current = runId
+        const n = Math.max(1, Math.min(20, Math.floor(Number(benchN) || 1)))
+        const trials = []
+        for (let i = 0; i < n; i++) { trials.push(['A', aText, i + 1]); trials.push(['B', bText, i + 1]) }
+        let doneA = 0
+        let doneB = 0
+        const total = trials.length
+        try { window.__deusBench = { running: true, runId, done: 0, total } } catch { /* ignore */ }
+        try {
+          for (const [variant, text, idx] of trials) {
+            if (benchAbort.current) break
+            setBenchProg({ a: doneA, b: doneB, n, label: variant + '#' + idx, waiting: false })
+            // 1) 新会话隔离（失败则降级为同会话连跑并亮警告）
+            const prevSid = sidRef.current
+            const switched = await benchNewSession(prevSid)
+            if (!switched) { setBenchNote(t('benchIsolationWarn')); await sleep(400) }
+            if (benchAbort.current) break
+            // 2) 等 composer actions 就绪
+            const w0 = Date.now()
+            while (!(actionsRef.current && typeof actionsRef.current.setDraft === 'function')) {
+              if (Date.now() - w0 > 8000) break
+              await sleep(250)
+            }
+            // 3) 宿主标记（文本队列配对）→ 注入并发送
+            try {
+              await fetch('/deus/bench/mark', {
+                method: 'POST', headers: { 'content-type': 'application/json' },
+                body: JSON.stringify({ runId, variant, prompt: text }),
+              })
+            } catch { /* marking best-effort */ }
+            const acts = actionsRef.current
+            if (!acts || typeof acts.setDraft !== 'function' || typeof acts.submit !== 'function') {
+              setBenchNote(t('benchTimeout')); break
+            }
+            acts.setDraft(text)
+            await sleep(200)
+            acts.submit()
+            // 4) 轮询报告计数，等宿主把这轮的指纹判定写进日志
+            setBenchProg({ a: doneA, b: doneB, n, label: variant + '#' + idx, waiting: true })
+            const doneSoFar = doneA + doneB
+            const t0 = Date.now()
+            let arrived = false
+            while (Date.now() - t0 < 240000) {
+              if (benchAbort.current) break
+              await sleep(2000)
+              const rep = await fetchBenchReport(runId).catch(() => null)
+              if (rep && rep.report && rep.report.n >= doneSoFar + 1) { arrived = true; break }
+            }
+            if (!arrived && !benchAbort.current) { setBenchNote(t('benchTimeout')); break }
+            if (variant === 'A') doneA++; else doneB++
+            try { window.__deusBench = { running: true, runId, done: doneA + doneB, total } } catch { /* ignore */ }
+            await sleep(600)
+          }
+        } finally {
+          setBenchRunning(false)
+          setBenchProg(null)
+          try { window.__deusBench = { running: false, runId, done: doneA + doneB, total } } catch { /* ignore */ }
+        }
+        const fin = await fetchBenchReport(runId).catch(() => null)
+        if (fin && fin.report) setBenchReport(fin.report)
+      }
       const REANCHOR = 'We need to continue working on this together. Let us pick up where we left off. 我们继续协作，接着上一步往下做。'
 
       useEffect(() => {
@@ -374,34 +619,72 @@ window.__ModuleLoader__.load({
       const UNANCHORED_PRESET_RE = /^(deus-|minimal)/
       const unanchored = !st && sessPreset && !UNANCHORED_PRESET_RE.test(sessPreset)
 
-      return h('div', { style: S.dock, title: t('dockHint') },
-        h('span', { style: { fontSize: '11px', color: '#8b949e' } }, '⚗'),
+      const benchPanel = benchOpen ? h('div', { style: S.benchWrap, 'data-deus': 'bench-panel', 'data-dsh-part': 'panel' },
+        h('div', { style: { fontSize: '12px', fontWeight: 600, color: 'var(--dsw-alias-label-primary, #e6edf3)' } }, '🧪 ' + t('bench')),
+        h('div', { style: S.status }, t('benchHint')),
+        h('div', { style: S.benchLabel }, t('benchA')),
+        h('textarea', {
+          style: S.benchTa, 'data-deus': 'bench-a', 'data-dsh-part': 'textarea', value: benchA, disabled: benchRunning,
+          onChange: (e) => { setBenchA(e.target.value); try { window.localStorage.setItem('deus.benchA', e.target.value) } catch { /* ignore */ } },
+        }),
+        h('div', { style: S.benchLabel }, t('benchB')),
+        h('textarea', {
+          style: S.benchTa, 'data-deus': 'bench-b', 'data-dsh-part': 'textarea', value: benchB, disabled: benchRunning,
+          onChange: (e) => { setBenchB(e.target.value); try { window.localStorage.setItem('deus.benchB', e.target.value) } catch { /* ignore */ } },
+        }),
+        h('div', { style: { ...S.row, marginTop: '8px', alignItems: 'center' } },
+          h('span', { style: { ...S.benchLabel, margin: 0 } }, t('benchTimes')),
+          h('input', {
+            type: 'number', min: 1, max: 20, 'data-deus': 'bench-n', value: benchN, disabled: benchRunning,
+            style: { background: 'var(--dsw-alias-bg-layer-1, #161b22)', border: '1px solid var(--dsw-alias-border-l1, #30363d)', borderRadius: '8px', color: 'var(--dsw-alias-label-primary, #c9d1d9)', fontSize: '12px', padding: '4px 6px', width: '60px' },
+            onChange: (e) => setBenchN(e.target.value),
+          }),
+          benchRunning
+            ? h('button', { style: { ...S.btn, color: STATE_ERR, borderColor: STATE_ERR }, 'data-deus': 'bench-abort', 'data-dsh-part': 'button-ghost', onClick: abortBench }, t('benchAbort'))
+            : h('button', { style: S.btnActive, 'data-deus': 'bench-run', 'data-dsh-part': 'button-ghost', onClick: runBench }, t('benchRun')),
+        ),
+        benchProg ? h('div', { style: S.status, 'data-deus': 'bench-progress' },
+          t('benchRunning') + ' · ' + t('benchProgress') + ': A ' + benchProg.a + '/' + benchProg.n + ' · B ' + benchProg.b + '/' + benchProg.n + ' · ' + benchProg.label + (benchProg.waiting ? ' ' + t('benchWaiting') : '')) : null,
+        benchNote ? h('div', { style: { ...S.status, color: STATE_WARN }, 'data-deus': 'bench-note' }, benchNote) : null,
+        benchReport ? h(BenchTable, { report: benchReport, t }) : null,
+      ) : null
+
+      const dockRow = h('div', { style: S.dock, title: t('dockHint') },
+        h('span', { style: { fontSize: '11px', color: 'var(--dsw-alias-label-secondary, #8b949e)' } }, '⚗'),
         presets.map((p) => h('button', {
-          key: p.id, style: S.dockChip, title: t('dockHint'),
+          key: p.id, style: S.dockChip, 'data-dsh-part': 'chip', title: t('dockHint'),
           onClick: () => fire(p),
         }, (autoSend ? '🚀 ' : '⚡ ') + p.label_zh)),
         st ? h('button', {
-          style: { ...S.dockChip, color: st.drifted ? '#f85149' : '#3fb950', borderColor: st.drifted ? '#f85149' : '#3fb950' },
+          style: { ...S.dockChip, color: st.drifted ? STATE_ERR : STATE_OK, borderColor: st.drifted ? STATE_ERR : STATE_OK },
+          'data-dsh-part': 'chip',
           title: st.drifted ? t('reanchor') : t('dockAnchored'),
           onClick: st.drifted ? fireReanchor : undefined,
         }, st.drifted ? '⚠ ' + t('reanchor') : `⚓ ${t('dockAnchored')} ${st.god}/${st.total}`) : null,
         unanchored ? h('button', {
-          style: { ...S.dockChip, color: '#8b949e', borderColor: '#8b949e', borderStyle: 'dotted' },
+          style: { ...S.dockChip, color: 'var(--dsw-alias-label-secondary, #8b949e)', borderColor: 'var(--dsw-alias-border-l1, #8b949e)', borderStyle: 'dotted' },
+          'data-dsh-part': 'chip',
           title: t('unanchoredTitle'),
           onClick: () => setShowGuide((v) => !v),
         }, t('unanchored')) : null,
         unanchored && showGuide ? h('div', {
           style: {
-            fontSize: '11px', color: '#8b949e', lineHeight: 1.6, maxWidth: '520px',
-            borderLeft: '2px solid #8b949e', paddingLeft: '8px', margin: '2px 0',
+            fontSize: '11px', color: 'var(--dsw-alias-label-secondary, #8b949e)', lineHeight: 1.6, maxWidth: '520px',
+            borderLeft: '2px solid var(--dsw-alias-border-l1, #8b949e)', paddingLeft: '8px', margin: '2px 0',
           },
         }, t('guideText')) : null,
-        reanchorMsg ? h('span', { style: { fontSize: '11px', color: '#3fb950' } }, reanchorMsg) : null,
-        h('button', { style: { ...S.dockChip, borderStyle: 'dashed' }, title: t('autoSend'), onClick: toggleAuto },
+        reanchorMsg ? h('span', { style: { fontSize: '11px', color: STATE_OK } }, reanchorMsg) : null,
+        h('button', { style: { ...S.dockChip, borderStyle: 'dashed' }, 'data-dsh-part': 'chip', title: t('autoSend'), onClick: toggleAuto },
           (autoSend ? '☑ ' : '☐ ') + t('autoSend')),
-        st ? h('button', { style: { ...S.dockChip, borderStyle: 'dashed' }, title: t('autoReanchor'), onClick: toggleAutoReanchor },
+        st ? h('button', { style: { ...S.dockChip, borderStyle: 'dashed' }, 'data-dsh-part': 'chip', title: t('autoReanchor'), onClick: toggleAutoReanchor },
           (autoReanchor ? '☑ ' : '☐ ') + t('autoReanchor')) : null,
+        h('button', {
+          style: { ...S.dockChip, borderStyle: 'dashed', color: benchOpen ? CLS_ACCENT.god : 'var(--dsw-alias-label-secondary, #8b949e)', borderColor: benchOpen ? CLS_ACCENT.god : 'var(--dsw-alias-border-l1, #30363d)' },
+          title: t('bench'), 'data-deus': 'bench-chip', 'data-dsh-part': 'chip',
+          onClick: () => setBenchOpen((v) => !v),
+        }, '🧪 ' + t('bench')),
       )
+      return h('div', { style: { width: '100%' }, 'data-dsh-plugin': 'plugin-deus', 'data-dsh-surface': 'chat-input' }, benchPanel, dockRow)
     }
 
     // v0.3: 触发率迷你趋势（最近 40 条日志分 10 桶的 god 率折线）
@@ -421,10 +704,11 @@ window.__ModuleLoader__.load({
       const xy = pts.map(([i, v]) => [P + (W - 2 * P) * i / (B - 1), H - P - (H - 2 * P) * v])
       const dAttr = xy.map((p, i) => (i ? 'L' : 'M') + p[0].toFixed(1) + ',' + p[1].toFixed(1)).join(' ')
       return h('svg', { width: W, height: H, viewBox: `0 0 ${W} ${H}`, style: { display: 'block', margin: '4px 0 2px' } },
-        h('path', { d: `M${P},${H - P} L${W - P},${H - P}`, stroke: '#30363d', strokeWidth: 1, fill: 'none' }),
-        h('path', { d: `M${P},${P} L${W - P},${P}`, stroke: '#30363d', strokeWidth: 1, strokeDasharray: '3 3', fill: 'none' }),
-        h('path', { d: dAttr, stroke: '#d2a8ff', strokeWidth: 2, fill: 'none' }),
-        xy.map((p, i) => h('circle', { key: i, cx: p[0], cy: p[1], r: 2.5, fill: '#d2a8ff' })),
+        // var() 在 SVG 表现属性里不解析，必须走 style（CSS 层才能吃自定义属性）
+        h('path', { d: `M${P},${H - P} L${W - P},${H - P}`, style: { stroke: 'var(--dsw-alias-border-l1, #30363d)' }, strokeWidth: 1, fill: 'none' }),
+        h('path', { d: `M${P},${P} L${W - P},${P}`, style: { stroke: 'var(--dsw-alias-border-l1, #30363d)' }, strokeWidth: 1, strokeDasharray: '3 3', fill: 'none' }),
+        h('path', { d: dAttr, style: { stroke: CLS_ACCENT.god }, strokeWidth: 2, fill: 'none' }),
+        xy.map((p, i) => h('circle', { key: i, cx: p[0], cy: p[1], r: 2.5, style: { fill: CLS_ACCENT.god } })),
       )
     }
 
@@ -453,6 +737,7 @@ window.__ModuleLoader__.load({
       inject: ['slots', 'locale'],
       apply(ctx) {
         try {
+          try { window.__deusClient = '0.4.1' } catch { /* non-browser */ }
           ctx.effect(() => ctx.locale.register(NS, { zh, en }), 'deus-mode: dicts')
           const t = ctx.locale.bind(NS)
           ctx.slots.inject('settings.section', () => ctx.slots.register({

package/lib/index.js

@@ -78,8 +78,27 @@ function wilson(k, n) {
   return { rate: p, low: Math.max(0, center - half), high: Math.min(1, center + half) }
 }
 
+// ── 两比例 z 检验（v0.4 实验台 A/B 对比用）─────────────────────────────────
+// 标准正态 CDF（Abramowitz–Stegun 7.1.26 近似，|ε| < 7.5e-8）
+function normalCdf(x) {
+  const t = 1 / (1 + 0.2316419 * Math.abs(x))
+  const d = 0.3989423 * Math.exp((-x * x) / 2)
+  const p = d * t * (0.3193815 + t * (-0.3565638 + t * (1.781478 + t * (-1.821256 + t * 1.330274))))
+  return x > 0 ? 1 - p : p
+}
+
+/** 两比例 z 检验（双侧）。任一样本为空 → null。 */
+function twoPropZ(kA, nA, kB, nB) {
+  if (!nA || !nB) return null
+  const pooled = (kA + kB) / (nA + nB)
+  const se = Math.sqrt(pooled * (1 - pooled) * (1 / nA + 1 / nB))
+  if (!se) return { z: 0, p: 1 }
+  const z = (kA / nA - kB / nB) / se
+  return { z, p: 2 * (1 - normalCdf(Math.abs(z))) }
+}
+
 // ── shipped agent presets (v0.2 锚定维持的静态层) ───────────────────────────
-const PLUGIN_VERSION = '0.3.0'
+const PLUGIN_VERSION = '0.4.0'
 // 实测依据（research/deus-mode-matrix.md §8/§9）：神版触发需要 Minimal persona
 // ∧ 小工具目录；且本 harness 没有竞品的 promoteOn 机制——preset 构成全程恒定，
 // 所以只要会话跑在锚定 preset 上，注入剥离/工具裁剪自动延续到每一轮。
@@ -300,6 +319,8 @@ function statsFrom(entries) {
   const byMode = new Map()
   for (const e of entries) {
     const mode = e.prompt_mode || 'unknown'
+    if (mode === 'bench') continue // v0.4: 实验台样本有独立的 A/B 报告，不混入神模统计
+
     if (!byMode.has(mode)) byMode.set(mode, { mode, n: 0, pure: 0, med: 0, god: 0, unknown: 0 })
     const row = byMode.get(mode)
     row.n += 1
@@ -320,13 +341,85 @@ function statsFrom(entries) {
 }
 
 function toCsv(entries) {
-  const head = 'ts,sessionId,prompt_mode,prompt_text,detected,source,turn,first_sentence,model'
+  const head = 'ts,sessionId,prompt_mode,prompt_text,detected,source,turn,first_sentence,model,bench_run,bench_variant'
   const esc = (v) => `"${String(v ?? '').replace(/"/g, '""')}"`
   const rows = entries.map((e) =>
-    [e.ts, e.sessionId, e.prompt_mode, e.prompt_text, e.detected, e.source, e.turn, e.first_sentence, e.model].map(esc).join(','))
+    [e.ts, e.sessionId, e.prompt_mode, e.prompt_text, e.detected, e.source, e.turn, e.first_sentence, e.model, e.bench_run, e.bench_variant].map(esc).join(','))
   return [head, ...rows].join('\n') + '\n'
 }
 
+// ── v0.4 实验台：A/B 报告（指纹分布对比 + Wilson CI + 两比例 z 检验）────────
+// 设计文档「纯噪音退路」转正：面板输入任意提示词变体 A/B，各跑 N 次真实会话
+// 注入，这里把 log 里 prompt_mode='bench' 的样本按 bench_run/bench_variant
+// 分组，输出每类的 Wilson 95% CI 与 A−B 差异检验。神模只是第一个被测假设。
+const BENCH_CLASSES = ['god', 'med', 'pure', 'unknown']
+
+export function benchReportFrom(entries, runId) {
+  const bench = entries.filter((e) => e.prompt_mode === 'bench' && e.bench_run)
+  const runMap = new Map()
+  for (const e of bench) {
+    let r = runMap.get(e.bench_run)
+    if (!r) {
+      r = { runId: e.bench_run, n: 0, firstTs: e.ts, lastTs: e.ts, variants: new Set() }
+      runMap.set(e.bench_run, r)
+    }
+    r.n += 1
+    if (e.ts < r.firstTs) r.firstTs = e.ts
+    if (e.ts > r.lastTs) r.lastTs = e.ts
+    if (e.bench_variant) r.variants.add(e.bench_variant)
+  }
+  const runs = [...runMap.values()]
+    .map((r) => ({ runId: r.runId, n: r.n, firstTs: r.firstTs, lastTs: r.lastTs, variants: [...r.variants].sort() }))
+    .sort((a, b) => (a.lastTs < b.lastTs ? 1 : -1))
+  const picked = runId ? runs.find((r) => r.runId === runId) : runs[0]
+  if (!picked) return { report: null, runs }
+  const rows = bench.filter((e) => e.bench_run === picked.runId)
+  const byVariant = {}
+  for (const e of rows) {
+    const v = String(e.bench_variant || '?')
+    const g = (byVariant[v] ||= { variant: v, n: 0, counts: { god: 0, med: 0, pure: 0, unknown: 0 }, prompt: null })
+    g.n += 1
+    if (g.counts[e.detected] !== undefined) g.counts[e.detected] += 1
+    else g.counts.unknown += 1
+    if (!g.prompt && e.prompt_text) g.prompt = e.prompt_text
+  }
+  for (const g of Object.values(byVariant)) {
+    g.classes = {}
+    for (const c of BENCH_CLASSES) {
+      g.classes[c] = { n: g.counts[c], ...wilson(g.counts[c], g.n) }
+    }
+  }
+  const order = Object.keys(byVariant).sort()
+  const comparison = []
+  if (order.length >= 2) {
+    const A = byVariant[order[0]]
+    const B = byVariant[order[1]]
+    for (const c of BENCH_CLASSES) {
+      const t = twoPropZ(A.counts[c], A.n, B.counts[c], B.n)
+      comparison.push({
+        cls: c,
+        a: A.classes[c],
+        b: B.classes[c],
+        delta: A.n && B.n ? A.counts[c] / A.n - B.counts[c] / B.n : null,
+        z: t ? t.z : null,
+        p: t ? t.p : null,
+      })
+    }
+  }
+  return {
+    report: {
+      runId: picked.runId,
+      n: rows.length,
+      variants: byVariant,
+      order,
+      comparison,
+      startedAt: picked.firstTs,
+      finishedAt: picked.lastTs,
+    },
+    runs,
+  }
+}
+
 // ── version auto-detect (differentiator: no manual checking) ────────────────
 function detectDshVersion() {
   if (process.env.DSH_VERSION) return process.env.DSH_VERSION
@@ -386,6 +479,10 @@ export function apply(ctx) {
 
   // Pending trigger marks: sessionId -> { mode, prompt_text, ts, auto }
   const pending = new Map()
+  // v0.4 实验台：待配对的 bench 样本队列 [{ runId, variant, prompt, ts }]
+  // （文本精确匹配配对，见 /deus/bench/mark 路由注释）
+  const benchQueue = []
+  const BENCH_QUEUE_TTL = 10 * 60 * 1000
   // Chunk accumulation while a trigger is pending: sessionId -> { turn, step, reasoning, text }
   // （实测校准：指纹在 reasoning-delta 流里，text-delta 兜底；见 detectMode 注释）
   const buffers = new Map()
@@ -466,6 +563,9 @@ export function apply(ctx) {
       source,
       first_sentence: firstSentenceOf(classifyOn),
       model: mark.model || undefined,
+      // v0.4 实验台：A/B 样本归属（仅 bench 标记的注入轮带这两个字段）
+      bench_run: mark.benchRun || undefined,
+      bench_variant: mark.benchVariant || undefined,
     })
   }
 
@@ -486,6 +586,15 @@ export function apply(ctx) {
     if (event.type === 'user/message') {
       const text = textOf(event.data && event.data.content).trim()
       if (text !== '' && !pending.has(sessionId)) {
+        // v0.4 实验台：bench 队列优先于 preset exact-match（显式实验意图优先）
+        const now = Date.now()
+        const bi = benchQueue.findIndex((q) => q.prompt === text && now - q.ts < BENCH_QUEUE_TTL)
+        if (bi >= 0) {
+          const q = benchQueue.splice(bi, 1)[0]
+          buffers.delete(sessionId)
+          pending.set(sessionId, { mode: 'bench', prompt_text: q.prompt, benchRun: q.runId, benchVariant: q.variant, ts: now, auto: false })
+          return
+        }
         const hit = loadPresets().find((p) => p.prompt !== null && p.prompt.trim() === text)
         if (hit) {
           buffers.delete(sessionId) // 防止上轮残留 buffer 污染注入轮判定
@@ -666,6 +775,42 @@ export function apply(ctx) {
         if (!sessionId) { json({ ok: false, error: 'missing sessionId' }, 400)(req, res); return }
         json({ sessionId, preset: readSessionPreset(sessionId) })(req, res)
       } }),
+      // v0.4 实验台：A/B 变体注入标记。与 /deus/trigger 的 sessionId 预标记不同，
+      // 这里走「提示词文本队列」配对（与档 A 的 exact-match 同思路）：客户端每次
+      // 试验前把 (runId, variant, prompt) 入队，user/message 到达时按文本精确
+      // 匹配出队并标记该会话注入轮。这样 bench 不依赖「新会话 id 何时出现」
+      // 的时序——新建空白会话在首条消息前可能没有稳定 id。
+      ctx.webServer.register({ kind: 'exact', path: '/deus/bench/mark', handler: (req, res) => {
+        readBody(req, (body) => {
+          const runId = String(body.runId || '').slice(0, 64)
+          const variant = String(body.variant || '').slice(0, 16)
+          const prompt = String(body.prompt || '').slice(0, 4000)
+          if (!runId || !variant || !prompt) {
+            json({ ok: false, error: 'missing runId / variant / prompt' }, 400)(req, res)
+            return
+          }
+          benchQueue.push({ runId, variant, prompt, ts: Date.now() })
+          if (benchQueue.length > 200) benchQueue.splice(0, benchQueue.length - 200)
+          json({ ok: true, queued: benchQueue.length })(req, res)
+        })
+      } }),
+      // 中止/清理：丢弃某个 run（或全部）未消耗的队列项
+      ctx.webServer.register({ kind: 'exact', path: '/deus/bench/clear', handler: (req, res) => {
+        readBody(req, (body) => {
+          const runId = String(body.runId || '')
+          const before = benchQueue.length
+          for (let i = benchQueue.length - 1; i >= 0; i--) {
+            if (!runId || benchQueue[i].runId === runId) benchQueue.splice(i, 1)
+          }
+          json({ ok: true, dropped: before - benchQueue.length })(req, res)
+        })
+      } }),
+      // v0.4 实验台：A/B 报告（?runId= 指定，缺省取最近一次 run；附全部 run 列表）
+      ctx.webServer.register({ kind: 'exact', path: '/deus/bench/report', handler: (req, res) => {
+        const u = new URL(String(req.url || ''), 'http://localhost')
+        const runId = String(u.searchParams.get('runId') || '')
+        json({ ok: true, ...benchReportFrom(readLog(10000), runId || undefined) })(req, res)
+      } }),
       ctx.webServer.register({ kind: 'exact', path: '/deus/version', handler: (req, res) => {
         let dsh = 'unknown'
         try { dsh = detectDshVersion() } catch (e) { ctx.logger.warn('[plugin-deus] version detect failed:', e) }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@dsh-suite/plugin-deus",
-  "version": "0.3.0",
-  "description": "神模扳机 Deus Trigger — 极简提示词触发实验台：预设库 + 一键注入 + 起手识别器 + 触发率统计（Wilson CI）。社区观察未获官方证实，本插件不承诺解锁任何隐藏能力。",
+  "version": "0.4.1",
+  "description": "神模扳机 Deus Trigger — 提示词 A/B 实验台：任意提示词变体各跑 N 次真实会话注入，指纹分布对比（Wilson CI + 两比例 z 检验）；预设库 + 一键注入 + 起手识别器 + 触发率统计。社区观察未获官方证实，本插件不承诺解锁任何隐藏能力。",
   "type": "module",
   "main": "lib/index.js",
   "exports": {