@dsh-suite/plugin-deus 0.2.0 → 0.4.0

package/README.md

@@ -26,6 +26,9 @@
    - 安装两个锚定 agent presets 到 `~/.dsh/.agent-presets/`（**窄锚** 2 工具 ~90% 触发 / **宽锚** ~8 工具 ~65% 换可用性），设置 > Agent presets 或会话 preset 选择器里直接选用，构成全程恒定；
    - 对 deus/minimal preset 会话**逐轮判定指纹**，面板「锚定维持」区实时显示每轮指纹与漂移状态；
    - 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。
 
 ## 安装 / Install
 
@@ -50,7 +53,7 @@ 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

@@ -42,6 +42,23 @@ window.__ModuleLoader__.load({
       colSession: '会话', colPreset: 'preset', colTurns: '轮次', colGodRate: '神版率', colLastFp: '最新指纹', colAnchorState: '状态',
       anchoredOn: '锚定维持中', drifted: '漂移→非神版', reanchor: '⚓ 重锚', autoReanchor: '漂移自动重锚',
       reanchorSent: '已发送重锚提示', dockAnchored: '锚定维持中',
+      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',
@@ -66,6 +83,23 @@ window.__ModuleLoader__.load({
       colSession: 'Session', colPreset: 'preset', colTurns: 'turns', colGodRate: 'god rate', colLastFp: 'latest', colAnchorState: 'state',
       anchoredOn: 'anchored', drifted: 'drifted', reanchor: '⚓ Re-anchor', autoReanchor: 'Auto re-anchor on drift',
       reanchorSent: 'Re-anchor nudge sent', dockAnchored: 'anchored',
+      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.',
     }
 
     const S = {
@@ -84,6 +118,11 @@ window.__ModuleLoader__.load({
       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' },
+      benchWrap: { width: '100%', background: '#0d1117', border: '1px solid #30363d', borderRadius: '10px', padding: '10px 12px', marginBottom: '6px' },
+      benchTa: { width: '100%', minHeight: '52px', background: '#161b22', border: '1px solid #30363d', borderRadius: '8px', color: '#c9d1d9', fontSize: '12px', padding: '6px 8px', boxSizing: 'border-box', resize: 'vertical' },
+      benchLabel: { fontSize: '11px', color: '#8b949e', margin: '6px 0 3px' },
+      benchTable: { width: '100%', borderCollapse: 'collapse', fontSize: '11px', marginTop: '6px' },
+      benchCell: { border: '1px solid #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' },
     }
@@ -91,6 +130,56 @@ window.__ModuleLoader__.load({
     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 = { god: '#d2a8ff', med: '#3fb950', pure: '#79c0ff', unknown: '#8b949e' }
+
+    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] || '#8b949e' } }, 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: '#d29922' } }, 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([])
@@ -101,6 +190,12 @@ 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')))
@@ -108,6 +203,7 @@ 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() }, [])
 
@@ -193,6 +289,8 @@ window.__ModuleLoader__.load({
         h('div', { style: S.h2 }, t('stats')),
         h('div', { style: S.status }, t('statsHint')),
         h('div', { style: { ...S.status, color: '#7d8590' } }, t('measured')),
+        h(Sparkline, { entries }),
+        entries.length >= 4 ? h('div', { style: { ...S.status, fontSize: '10px' } }, t('trend')) : null,
         stats && stats.modes.length > 0
           ? h('table', { style: S.table },
               h('thead', null, h('tr', null,
@@ -209,6 +307,15 @@ window.__ModuleLoader__.load({
           : h('div', { style: S.status }, t('noData')),
         h('div', { style: S.row },
           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,
+            onClick: async () => {
+              try {
+                await navigator.clipboard.writeText(buildMarkdownSummary(stats))
+                setCopied('__md'); setTimeout(() => setCopied(''), 2500)
+              } catch {}
+            },
+          }, '📋 ' + (copied === '__md' ? t('mdCopied') : t('copyMd'))) : null,
           h('button', { style: { ...S.btn, color: '#f85149', borderColor: '#f85149' }, onClick: resetLog }, '🗑 ' + t('resetLog')),
         ),
 
@@ -239,6 +346,23 @@ window.__ModuleLoader__.load({
             )
           : 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', onClick: () => refreshBench(benchRunId) }, '🔄 ' + t('benchRefresh')),
+          benchData && benchData.runs && benchData.runs.length > 0 ? h('select', {
+            style: { ...S.input, width: 'auto' }, 'data-deus': 'bench-runs',
+            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'))
@@ -276,16 +400,148 @@ window.__ModuleLoader__.load({
         try { return window.localStorage.getItem('deus.autoReanchor') === '1' } catch { return false }
       })
       const [reanchorMsg, setReanchorMsg] = useState('')
+      // v0.3 未锚定引导：当前会话的 agent preset（宿主读会话日志头行）
+      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(() => {
         let dead = false
-        const poll = () => fetch('/deus/anchor').then((r) => r.json()).then((d) => { if (!dead) setAnchor(d) }).catch(() => {})
+        const poll = () => {
+          fetch('/deus/anchor').then((r) => r.json()).then((d) => { if (!dead) setAnchor(d) }).catch(() => {})
+          if (sessionId) {
+            fetch('/deus/session-preset?sessionId=' + encodeURIComponent(sessionId))
+              .then((r) => r.json()).then((d) => { if (!dead) setSessPreset(d.preset || null) }).catch(() => {})
+          }
+        }
         poll()
         const timer = setInterval(poll, 4000)
         return () => { dead = true; clearInterval(timer) }
-      }, [])
+      }, [sessionId])
 
       const st = anchor && Array.isArray(anchor.sessions)
         ? anchor.sessions.find((s) => String(s.sessionId) === String(sessionId))
@@ -344,7 +600,41 @@ window.__ModuleLoader__.load({
 
       if (presets.length === 0) return null // 所有 hooks 之上不可早退（React hooks 顺序）
 
-      return h('div', { style: S.dock, title: t('dockHint') },
+      // v0.3 未锚定引导：非锚定会话 + preset 已知且非 deus/minimal 系 → 灰 chip
+      const UNANCHORED_PRESET_RE = /^(deus-|minimal)/
+      const unanchored = !st && sessPreset && !UNANCHORED_PRESET_RE.test(sessPreset)
+
+      const benchPanel = benchOpen ? h('div', { style: S.benchWrap, 'data-deus': 'bench-panel' },
+        h('div', { style: { fontSize: '12px', fontWeight: 600, color: '#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', 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', 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: '#161b22', border: '1px solid #30363d', borderRadius: '8px', color: '#c9d1d9', fontSize: '12px', padding: '4px 6px', width: '60px' },
+            onChange: (e) => setBenchN(e.target.value),
+          }),
+          benchRunning
+            ? h('button', { style: { ...S.btn, color: '#f85149', borderColor: '#f85149' }, 'data-deus': 'bench-abort', onClick: abortBench }, t('benchAbort'))
+            : h('button', { style: S.btnActive, 'data-deus': 'bench-run', 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: '#d29922' }, '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: '#8b949e' } }, '⚗'),
         presets.map((p) => h('button', {
           key: p.id, style: S.dockChip, title: t('dockHint'),
@@ -355,18 +645,81 @@ window.__ModuleLoader__.load({
           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' },
+          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',
+          },
+        }, 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 },
           (autoSend ? '☑ ' : '☐ ') + t('autoSend')),
         st ? h('button', { style: { ...S.dockChip, borderStyle: 'dashed' }, title: t('autoReanchor'), onClick: toggleAutoReanchor },
           (autoReanchor ? '☑ ' : '☐ ') + t('autoReanchor')) : null,
+        h('button', {
+          style: { ...S.dockChip, borderStyle: 'dashed', color: benchOpen ? '#d2a8ff' : '#8b949e', borderColor: benchOpen ? '#d2a8ff' : '#30363d' },
+          title: t('bench'), 'data-deus': 'bench-chip',
+          onClick: () => setBenchOpen((v) => !v),
+        }, '🧪 ' + t('bench')),
       )
+      return h('div', { style: { width: '100%' } }, 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）
+    function Sparkline({ entries }) {
+      const seq = entries.slice(0, 40).reverse() // /deus/log 最新在前 → 转时间序
+      if (seq.length < 4) return null
+      const B = 10
+      const pts = []
+      for (let i = 0; i < B; i++) {
+        const chunk = seq.slice(Math.floor(i * seq.length / B), Math.floor((i + 1) * seq.length / B))
+        if (chunk.length === 0) continue
+        const god = chunk.filter((e) => e.detected === 'god').length / chunk.length
+        pts.push([i, god])
+      }
+      if (pts.length < 2) return null
+      const W = 260, H = 48, P = 5
+      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' })),
+      )
+    }
+
+    // v0.3: 一键复制 Markdown 摘要（传播物料：表格 + 诚实声明 + 包链接）
+    function buildMarkdownSummary(stats) {
+      const pct1 = (x) => (x * 100).toFixed(1) + '%'
+      const lines = [
+        '## 神版触发率实测摘要 / God-mode trigger summary', '',
+        '| 模式 mode | 样本 n | 神版率 god | 95% CI | 中版 med | 纯区 pure |',
+        '|---|---|---|---|---|---|',
+      ]
+      for (const m of stats.modes || []) {
+        if (!m || !m.n) continue
+        lines.push(`| ${m.mode} | ${m.n} | ${pct1(m.godCI.rate)} | ${pct1(m.godCI.low)}–${pct1(m.godCI.high)} | ${m.med} | ${m.pure} |`)
+      }
+      if (stats.total && stats.total.n) {
+        lines.push(`| **ALL** | **${stats.total.n}** | **${pct1(stats.total.godCI.rate)}** | **${pct1(stats.total.godCI.low)}–${pct1(stats.total.godCI.high)}** | ${stats.total.med} | ${stats.total.pure} |`)
+      }
+      lines.push('',
+        `> ${stats.total ? stats.total.n : 0} 条本地日志，由 @dsh-suite/plugin-deus 记录。「神版/中版/纯区」为社区观察（X @NFT_Chen），未经 DeepSeek 官方证实；判定 = 推理流起手句式启发式，存在误判率。`,
+        '> https://www.npmjs.com/package/@dsh-suite/plugin-deus')
+      return lines.join('\n')
     }
 
     return {
       inject: ['slots', 'locale'],
       apply(ctx) {
         try {
+          try { window.__deusClient = '0.4.0' } 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

@@ -10,11 +10,11 @@
 //   2. JSONL experiment log at $DSH_HOME/deus-mode/log.jsonl
 //   3. Trigger preset library (5 built-in modes, user-editable via presets.json)
 //   4. /deus/* routes for the browser half (presets, log, stats+Wilson CI, version)
-import { existsSync, mkdirSync, readFileSync, writeFileSync, appendFileSync } from 'node:fs'
+import { existsSync, mkdirSync, readFileSync, writeFileSync, appendFileSync, readdirSync } from 'node:fs'
 import { createRequire } from 'node:module'
 import { join } from 'node:path'
 import { homedir } from 'node:os'
-import { gzipSync } from 'node:zlib'
+import { gzipSync, zstdDecompressSync } from 'node:zlib'
 
 export const name = 'deus-mode'
 export const inject = ['webServer', 'sessions']
@@ -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 = '0.2.0'
+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。
@@ -216,6 +235,40 @@ function ensureDir() {
   mkdirSync(dataDir(), { recursive: true })
 }
 
+// ── v0.3: session preset 探测（未锚定引导的数据源）─────────────────────────
+// session/event 流里 user/message 不带 agentPreset；权威来源是磁盘会话日志的
+// SessionHeader（jsonl.zstd 第一帧第一行，session-persistence-jsonl 格式）。
+// 读文件 → 找第二个 zstd 帧魔数切出第一帧 → 解压取 header.agentPreset。
+// 60s 缓存（blank 会话可在首轮前 recompose preset，不能永久缓存）。
+const ZSTD_MAGIC = Buffer.from([0x28, 0xb5, 0x2f, 0xfd])
+const presetCache = new Map() // sessionId -> { preset, ts }
+function readSessionPreset(sessionId) {
+  const hit = presetCache.get(sessionId)
+  if (hit && Date.now() - hit.ts < 60_000) return hit.preset
+  let preset = null
+  try {
+    const root = join(dshHome(), 'sessions')
+    if (existsSync(root)) {
+      for (const ws of readdirSync(root)) {
+        const f = join(root, ws, sessionId, 'session.jsonl.zstd')
+        if (!existsSync(f)) continue
+        const buf = readFileSync(f)
+        const second = buf.indexOf(ZSTD_MAGIC, 4)
+        const firstFrame = second > 4 ? buf.subarray(0, second) : buf
+        try {
+          const line0 = zstdDecompressSync(firstFrame).toString('utf8').split('\n')[0]
+          const header = JSON.parse(line0)
+          if (header && header.type === 'session' && typeof header.agentPreset === 'string') preset = header.agentPreset
+        } catch { /* 帧截断/格式异常 → preset 保持 null */ }
+        break
+      }
+    }
+  } catch { /* 磁盘问题 → null */ }
+  presetCache.set(sessionId, { preset, ts: Date.now() })
+  if (presetCache.size > 300) presetCache.delete(presetCache.keys().next().value)
+  return preset
+}
+
 // ── preset library: built-ins until the user edits, then presets.json wins ──
 function loadPresets() {
   try {
@@ -266,6 +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
@@ -286,13 +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
@@ -352,6 +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()
@@ -432,6 +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,
     })
   }
 
@@ -452,6 +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
@@ -625,6 +768,49 @@ export function apply(ctx) {
       ctx.webServer.register({ kind: 'exact', path: '/deus/stats', handler: (req, res) => {
         json(statsFrom(readLog(10000)))(req, res)
       } }),
+      // v0.3: 当前会话的 agent preset（dock 未锚定引导用；来源=会话日志头行）
+      ctx.webServer.register({ kind: 'exact', path: '/deus/session-preset', handler: (req, res) => {
+        const u = new URL(String(req.url || ''), 'http://localhost')
+        const sessionId = String(u.searchParams.get('sessionId') || '')
+        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": "0.2.0",
-  "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": "0.4.0",
+  "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": {