oh-my-knowledge 0.22.0 → 0.24.0

package/dist/src/renderer/summary.js

@@ -1,4 +1,5 @@
 import { e, fmtNum, fmtCost, fmtDuration, COLORS, t } from './layout.js';
+import { generateAnalysisSummary } from '../analysis/report-diagnostics.js';
 import { pValueCategory } from '../eval-core/statistics.js';
 import { computeVerdict } from '../eval-core/verdict.js';
 /**
@@ -10,19 +11,6 @@ import { computeVerdict } from '../eval-core/verdict.js';
  * UNDERPOWERED / NOISE / CAUTIOUS — none of which are "ship" but none of which
  * are "regress" either. SOLO is grey because it's an info pill, not a verdict.
  */
-// v0.21 B.4 — verdict level → status icon. PROGRESS/REGRESS 用实心点强信号,
-// CAUTIOUS 三角(警示),NOISE 空心圆(有信号但无效果),UNDERPOWERED 部分填充
-// (不够数据),SOLO 描边圆(单变体非对比).
-function levelIcon(level) {
-    switch (level) {
-        case 'PROGRESS': return '●';
-        case 'REGRESS': return '●';
-        case 'CAUTIOUS': return '▲';
-        case 'NOISE': return '◌';
-        case 'UNDERPOWERED': return '◔';
-        case 'SOLO': return '○';
-    }
-}
 // v0.21 B.4 — verdict 一句话总结. 把 level 标签 + Δ 数字 + action 推荐 + 原因
 // 压成一个完整的中文/英文句子, 用 variant 名直接说"X 比 Y 怎么样, 怎么办".
 // 用户读完一句话就懂结论, 不需要先学 jargon (Δ / CI / NOISE / SHIP).
@@ -36,7 +24,7 @@ function verdictOneLine(level, lang, treatment, control) {
             case 'CAUTIOUS': return `${t} 比 ${c} 略好 — 但建议再仔细看,差距很小或某层未达标`;
             case 'REGRESS': return `${t} 比 ${c} 明显更差 — 不要发布`;
             case 'NOISE': return `${t} 和 ${c} 没看出明显差别 — 可以多加几条用例再试`;
-            case 'UNDERPOWERED': return `用例数太少,看不出 ${t} 和 ${c} 的差别 — 多跑几个再看`;
+            case 'UNDERPOWERED': return `评测用例数太少,看不出 ${t} 和 ${c} 的差别 — 多跑几个再看`;
             case 'SOLO': return `只跑了一组,需要加对照组才能对比`;
         }
     }
@@ -75,7 +63,7 @@ export function levelTooltip(level, lang) {
             case 'REGRESS': return '实验组分数显著劣于对照组';
             case 'CAUTIOUS': return '实验组略优于对照组,但差距小或某层未达 gate';
             case 'NOISE': return '两组分数差距置信区间跨过 0,统计上分辨不出效果';
-            case 'UNDERPOWERED': return '用例数太少,需要多加几条再看';
+            case 'UNDERPOWERED': return '评测用例数太少,需要多加几条再看';
             case 'SOLO': return '只跑了一组,没做对比';
         }
     }
@@ -95,7 +83,7 @@ export function levelTooltip(level, lang) {
 // 去掉旧版 headline 和 significance 的内容重复 (perPair[0].headline 是干净的
 // Δ+CI 核心, headline 拼装版只在 SOLO mode 用).
 //
-// computeVerdict 在 each mode + 历史脏 report 上有 NPE 风险 (顶层 summary
+// computeVerdict 在历史混合批量 report 上有 NPE 风险 (顶层 summary
 // 缺 variant 数据时 evaluateLayerGates 访问 .avgFactScore 炸). 加 try/catch
 // 让 renderer 不 crash, 改为静默 skip pill — 这是 v0.21 B.4 的 scope 范围,
 // verdict.ts/layer-gates.ts 的 defensive 修复留作单独 task.
@@ -318,11 +306,11 @@ export function renderSaturationCurve(saturation, variants, lang) {
         : `<p style="font-size:13px;color:var(--text-muted);margin:8px 0 0">${lang === 'zh' ? '提示:repeat ≥ 5 时才会输出饱和判定。当前数据只够画曲线,不足以下结论。' : 'Note: saturation verdict needs repeat ≥ 5. Current data plots the curve only.'}</p>`;
     return `
     <h2 style="margin-top:24px">${lang === 'zh' ? '饱和曲线 (Saturation curve)' : 'Saturation curve'}</h2>
-    <p style="font-size:13px;color:var(--text-secondary);margin:4px 0 12px">${lang === 'zh' ? '随累积用例数 N 增长的均值与 95% CI。CI 宽度衰减放缓即饱和——再多用例对结论无实质收益。' : 'Mean and 95% CI as cumulative N grows. When CI shrink rate flattens, the evidence has saturated — more samples buy little.'}</p>
+    <p style="font-size:13px;color:var(--text-secondary);margin:4px 0 12px">${lang === 'zh' ? '随累积评测用例数 N 增长的均值与 95% CI。CI 宽度衰减放缓即饱和——再多评测用例对结论无实质收益。' : 'Mean and 95% CI as cumulative N grows. When CI shrink rate flattens, the evidence has saturated — more samples buy little.'}</p>
     <svg viewBox="0 0 ${width} ${height}" xmlns="http://www.w3.org/2000/svg" style="width:100%;max-width:${width}px;height:auto;display:block">
       ${yTicks}
       <text x="${padL - 36}" y="${padT + plotH / 2}" font-size="11" text-anchor="middle" fill="var(--text-muted)" transform="rotate(-90 ${padL - 36} ${padT + plotH / 2})">${lang === 'zh' ? '均值' : 'mean'}</text>
-      <text x="${padL + plotW / 2}" y="${height - 4}" font-size="11" text-anchor="middle" fill="var(--text-muted)">${lang === 'zh' ? '累积用例数 N' : 'cumulative N'}</text>
+      <text x="${padL + plotW / 2}" y="${height - 4}" font-size="11" text-anchor="middle" fill="var(--text-muted)">${lang === 'zh' ? '累积评测用例数 N' : 'cumulative N'}</text>
       ${xTicks}
       ${seriesParts.join('\n')}
     </svg>
@@ -374,10 +362,18 @@ export function renderSummaryCards(variants, summary, lang, variance) {
         // LLM 评价层 cell
         const judgeCell = renderLayerCell(vd?.byLayer?.judge?.mean, s.avgJudgeScore);
         // Cost — only show execution cost (judge cost is tool overhead, not skill cost)
+        // execCostReported === false 时(如 codex executor)显示「—」并 tooltip 解释,
+        // 跟"真的花了 $0(全 cached / 短 prompt)"区分开。
         const execCost = s.totalExecCostUSD || 0;
+        const costReported = s.execCostReported !== false;
         const hasCost = execCost > 0 || (s.avgTotalTokens || 0) > 0;
+        const costUnreportedTooltip = lang === 'zh'
+            ? 'executor 不报 USD 成本(如 codex CLI),无法估算'
+            : 'executor does not report USD cost (e.g. codex CLI); not measurable';
         const costCell = hasCost
-            ? `<td class="summary-cell"><div class="summary-value">${fmtCost(execCost)}</div><div class="summary-detail">${fmtNum(s.avgTotalTokens)} tokens/${t('tokPerReq', lang).replace('tokens/', '')}</div></td>`
+            ? (costReported
+                ? `<td class="summary-cell"><div class="summary-value">${fmtCost(execCost)}</div><div class="summary-detail">${fmtNum(s.avgTotalTokens)} tokens/${t('tokPerReq', lang).replace('tokens/', '')}</div></td>`
+                : `<td class="summary-cell" title="${e(costUnreportedTooltip)}"><div class="summary-value" style="color:var(--text-muted)">${fmtCost(0, false)}</div><div class="summary-detail">${fmtNum(s.avgTotalTokens)} tokens/${t('tokPerReq', lang).replace('tokens/', '')}</div></td>`)
             : `<td class="summary-cell"><span style="color:var(--text-muted)">N/A</span></td>`;
         // Efficiency
         const effDetails = [];
@@ -456,10 +452,10 @@ export function renderSummaryCards(variants, summary, lang, variance) {
         return `<tr><td style="border-left:3px solid ${color};padding-left:12px"><strong>${e(v)}</strong></td>${factCell}${behaviorCell}${judgeCell}${costCell}${effCell}${stabCell}</tr>`;
     }).join('');
     const guideModalId = 'guide-six-dims';
-    const guideTitle = lang === 'zh' ? '如何阅读六维对比？' : 'How to read this 6-dim comparison?';
+    const guideTitle = lang === 'zh' ? '如何阅读六维对比？' : 'How to read this six-dimension comparison?';
     const guideIntro = lang === 'zh'
         ? '每行是一个实验分组（Variant），六列分别衡量不同维度：'
-        : 'Each row is a Variant. Six columns measure independent dimensions:';
+        : 'Each row is an experiment variant. Six columns measure independent dimensions:';
     const icon = (emoji) => `<span aria-hidden="true">${emoji}</span>`;
     // 维度分隔加粗(border-top 2px),让六维的视觉边界更明显。sub 缩进从 28 收到 22。
     const dim = 'style="padding:12px 0 4px;border-top:2px solid var(--border);color:var(--text-primary);font-weight:600"';
@@ -469,10 +465,10 @@ export function renderSummaryCards(variants, summary, lang, variance) {
     const sub = 'style="padding:2px 0 2px 22px;font-size:12px;color:var(--text-secondary);font-weight:500"';
     const subDesc = 'style="padding:2px 0;font-size:12px;color:var(--text-muted)"';
     const guideRows = lang === 'zh' ? `
-    <tr><td ${dimFirst}>${icon('📋')} <strong>事实</strong></td><td ${dimFirstDesc}>模型的输出说得对不对（事实声明层面）。靠规则断言判：关键词是否出现、JSON 格式是否合法等，答错了直接不给分。</td></tr>
-    <tr><td ${dim}>${icon('🛠️')} <strong>行为</strong></td><td ${dimDesc}>模型做事的过程有没有走对路。靠规则断言判：该调的工具有没有调、有没有超过轮次/成本上限。</td></tr>
-    <tr><td ${dim}>${icon('💬')} <strong>LLM 评价</strong></td><td ${dimDesc}>请一个 LLM 当评委，让它读被测模型的输出内容，按预先写好的评分规则（英文叫 rubric）打个 1-5 分。主观但能抓到规则断言判不了的"整体好不好"——比如回答是否清晰、有没有答非所问。</td></tr>
-    <tr><td ${dim}>${icon('💰')} <strong>成本</strong></td><td ${dimDesc}>跑这次评测花了多少 API 调用费（只算执行成本，评委那个 LLM 的钱不算进来）。</td></tr>
+    <tr><td ${dimFirst}>${icon('📋')} <strong>事实</strong></td><td ${dimFirstDesc}>任务执行模型的输出说得对不对（事实声明层面）。靠规则断言判：关键词是否出现、JSON 格式是否合法等，答错了直接不给分。</td></tr>
+    <tr><td ${dim}>${icon('🛠️')} <strong>行为</strong></td><td ${dimDesc}>任务执行模型做事的过程有没有走对路。靠规则断言判：该调的工具有没有调、有没有超过轮次/成本上限。</td></tr>
+    <tr><td ${dim}>${icon('💬')} <strong>LLM 评价</strong></td><td ${dimDesc}>请一个 LLM 当评委，让它读任务执行模型的输出内容，按预先写好的评分规则（英文叫 rubric）打个 1-5 分。主观但能抓到规则断言判不了的"整体好不好"——比如回答是否清晰、有没有答非所问。</td></tr>
+    <tr><td ${dim}>${icon('💰')} <strong>执行成本</strong></td><td ${dimDesc}>任务执行模型跑这次评测花了多少 API 调用费。评委模型成本单独作为评测开销记录，不算进 skill 自身成本。</td></tr>
     <tr><td ${dim}>${icon('⚡')} <strong>效率</strong></td><td ${dimDesc}>一次评测平均跑多久；附带轮次数和工具调用次数。</td></tr>
     <tr><td ${dim}>${icon('🛡️')} <strong>稳定性</strong></td><td ${dimDesc}>同一份测试跑很多次，分数抖不抖。抖得越少越稳定。<strong>跑一次看不出稳定性</strong>——至少要 <code>--repeat ≥ 2</code>，不然显示"—"。</td></tr>
     <tr><td ${sub}>稳定 / 较稳 / 波动大</td><td ${subDesc}>分数波动比例 &lt;5% = 稳定 · 5~15% = 一般 · &gt;15% = 波动大</td></tr>
@@ -480,10 +476,10 @@ export function renderSummaryCards(variants, summary, lang, variance) {
     <tr><td ${sub}>CV</td><td ${subDesc}>分数抖动幅度占平均分的比例（例：CV 2% = 分数波动大约是平均分的 2%）</td></tr>
     <tr><td ${sub}>95% CI</td><td ${subDesc}>如果跑无数次求平均，真实平均分有 95% 概率落在这个范围里——范围越窄，这次测出的均值越可信</td></tr>
   ` : `
-    <tr><td ${dimFirst}>${icon('📋')} <strong>Fact</strong></td><td ${dimFirstDesc}>Whether the model's output is factually correct. Checked by rule-based assertions — keyword matches, JSON schema validity, etc. Wrong = zero.</td></tr>
-    <tr><td ${dim}>${icon('🛠️')} <strong>Behavior</strong></td><td ${dimDesc}>Whether the model followed the right process. Checked by rule-based assertions — did it call the expected tools, stay within turn/cost limits.</td></tr>
-    <tr><td ${dim}>${icon('💬')} <strong>LLM judge</strong></td><td ${dimDesc}>A separate LLM acts as judge: it reads the tested model's output and scores it 1-5 against a predefined rubric. Subjective, but catches "overall feel" that rule-based assertions miss — e.g., whether the answer is clear, whether it's on-topic.</td></tr>
-    <tr><td ${dim}>${icon('💰')} <strong>Cost</strong></td><td ${dimDesc}>API cost of this run (execution only — the judge LLM's cost isn't included here).</td></tr>
+    <tr><td ${dimFirst}>${icon('📋')} <strong>Fact</strong></td><td ${dimFirstDesc}>Whether the task execution model's output is factually correct. Checked by rule-based assertions — keyword matches, JSON schema validity, etc. Wrong = zero.</td></tr>
+    <tr><td ${dim}>${icon('🛠️')} <strong>Behavior</strong></td><td ${dimDesc}>Whether the task execution model followed the right process. Checked by rule-based assertions — did it call the expected tools, stay within turn/cost limits.</td></tr>
+    <tr><td ${dim}>${icon('💬')} <strong>LLM judge</strong></td><td ${dimDesc}>A separate LLM acts as judge: it reads the task execution model's output and scores it 1-5 against a predefined rubric. Subjective, but catches "overall feel" that rule-based assertions miss — e.g., whether the answer is clear, whether it's on-topic.</td></tr>
+    <tr><td ${dim}>${icon('💰')} <strong>Exec cost</strong></td><td ${dimDesc}>API cost for the task execution model. Judge model cost is tracked as evaluation overhead, not skill cost.</td></tr>
     <tr><td ${dim}>${icon('⚡')} <strong>Efficiency</strong></td><td ${dimDesc}>Average time per evaluation, plus turn counts and tool call stats.</td></tr>
     <tr><td ${dim}>${icon('🛡️')} <strong>Stability</strong></td><td ${dimDesc}>How much the score swings when you repeat the same test. Less swing = more stable. <strong>You can't measure stability from a single run</strong> — need <code>--repeat ≥ 2</code>, otherwise shows "—".</td></tr>
     <tr><td ${sub}>Stable / Moderate / Variable</td><td ${subDesc}>Score swing as % of mean: &lt;5% = Stable · 5~15% = Moderate · &gt;15% = Variable</td></tr>
@@ -492,7 +488,7 @@ export function renderSummaryCards(variants, summary, lang, variance) {
     <tr><td ${sub}>95% CI</td><td ${subDesc}>If you ran infinitely many times, the true mean has a 95% chance of falling in this range — narrower = you can trust the measured mean more</td></tr>
   `;
     return `
-    <h2 style="display:flex;align-items:center;gap:4px">${lang === 'zh' ? '六维对比' : '6-Dim Comparison'} <button type="button" class="hint-btn" onclick="openModal('${guideModalId}')" aria-label="${e(guideTitle)}" aria-haspopup="dialog">?</button></h2>
+    <h2 style="display:flex;align-items:center;gap:4px">${lang === 'zh' ? '六维对比' : 'Six-Dimension Comparison'} <button type="button" class="hint-btn" onclick="openModal('${guideModalId}')" aria-label="${e(guideTitle)}" aria-haspopup="dialog">?</button></h2>
     <div id="${guideModalId}" class="modal-overlay" role="dialog" aria-modal="true" aria-labelledby="${guideModalId}-title" onclick="if(event.target===this)closeModal('${guideModalId}')">
       <div class="modal-content">
         <div class="modal-header">
@@ -548,7 +544,7 @@ function renderJudgeAgreementBlock(variants, summary, lang) {
         <th>${t('judgeModelsLabel', lang)}</th>
         <th title="${t('pearsonDesc', lang)}">${t('pearsonLabel', lang)}</th>
         <th title="${t('madDesc', lang)}">${t('madLabel', lang)}</th>
-        <th>${lang === 'zh' ? '用例数' : 'Samples'}</th>
+        <th>${lang === 'zh' ? '评测用例数' : 'Samples'}</th>
         <th>${lang === 'zh' ? 'Judge 对数' : 'Pairs'}</th>
       </tr></thead>
       <tbody>${rows}</tbody>
@@ -682,9 +678,12 @@ function getVariantMetricMean(variance, variant, key) {
         return v.mean;
     return v.byMetric?.[key]?.mean ?? null;
 }
-export function renderVarianceComparisons(variance, lang, layeredStatsOpen = false) {
+export function renderVarianceComparisons(variance, lang, layeredStatsOpen = false, summary) {
     if (!variance || !variance.comparisons || variance.comparisons.length === 0)
         return '';
+    // 任一 variant 的 exec 或 judge cost 未报告 → cost 行跳过(显示「—」会跟"持平"难区分,
+    // 索性直接不渲染该指标行)。判 cost 列前看 summary,缺 summary 当全 reported 走旧路径。
+    const allCostReported = !summary || Object.values(summary).every((v) => v.execCostReported !== false && v.judgeCostReported !== false);
     const modalId = 'guide-variance-comparisons';
     const title = lang === 'zh' ? '方差与显著性' : 'Variance & Significance';
     const guideTitle = lang === 'zh' ? '如何阅读这张表？' : 'How to read this table?';
@@ -830,6 +829,9 @@ export function renderVarianceComparisons(variance, lang, layeredStatsOpen = fal
         // Collect available metric rows for this comparison
         const availableMetrics = [];
         for (const cfg of METRIC_CONFIGS) {
+            // cost 行:任一 variant 不报 cost 时跳过(数据是占位 0,显示"持平"误导)
+            if (cfg.key === 'cost' && !allCostReported)
+                continue;
             const m = pickMetricFromComparison(comp, cfg.key);
             if (m)
                 availableMetrics.push({ cfg, metric: m });
@@ -864,7 +866,7 @@ export function renderVarianceComparisons(variance, lang, layeredStatsOpen = fal
         { label: '显著性', desc: 't 检验结论，基于 p<0.05 阈值。回答"差异真不真"，和效应量"差多大"互补' },
         { label: 'p 值', desc: '假设真的没差异时，观察到当前差距的概率。越小越可信。0.05 只是约定阈值', sub: true },
         { label: 't 值', desc: '均值差 ÷ 估计误差，需配合 df 和效应量解读，不能单独看', sub: true },
-        { label: 'df 自由度', desc: '≈"有效用例数"。--repeat 3 时通常 2~4；想达到 20+ 需 --repeat 10+', sub: true },
+        { label: 'df 自由度', desc: '≈"有效评测用例数"。--repeat 3 时通常 2~4；想达到 20+ 需 --repeat 10+', sub: true },
     ];
     const glossaryEn = [
         { label: 'Gap', desc: 'Cross-run mean winner + absolute difference (raw units)' },
@@ -947,7 +949,7 @@ function severityDot(severity) {
     };
     return `<span style="display:inline-block;width:6px;height:6px;border-radius:50%;background:${color[severity] || 'var(--text-muted)'};margin-right:8px;flex-shrink:0;margin-top:6px"></span>`;
 }
-function renderSummaryStructured(summary) {
+function renderSummaryStructured(summary, lang) {
     const markerRegex = /【([^】]+)】/g;
     const markers = [];
     let match;
@@ -962,9 +964,17 @@ function renderSummaryStructured(summary) {
     if (sections.length === 0) {
         return `<div style="padding:14px 18px;font-size:13px;line-height:1.8;color:var(--text-secondary);background:var(--bg-surface);border-radius:var(--radius);border:1px solid var(--border)">${e(summary)}</div>`;
     }
+    const labelMap = {
+        结论: { zh: '结论', en: 'Conclusion' },
+        Conclusion: { zh: '结论', en: 'Conclusion' },
+        '关键差异': { zh: '关键差异', en: 'Key differences' },
+        'Key differences': { zh: '关键差异', en: 'Key differences' },
+        综合洞察: { zh: '综合洞察', en: 'Synthesis' },
+        Synthesis: { zh: '综合洞察', en: 'Synthesis' },
+    };
     const sectionHtml = sections.map((section) => `
       <div style="display:flex;gap:12px;align-items:baseline">
-        <span style="flex-shrink:0;font-size:11px;font-weight:600;color:var(--text-muted);letter-spacing:0.03em;min-width:56px">${e(section.label)}</span>
+        <span style="flex-shrink:0;font-size:11px;font-weight:600;color:var(--text-muted);letter-spacing:0.03em;min-width:56px">${e(labelMap[section.label]?.[lang] || section.label)}</span>
         <span style="color:var(--text-secondary);font-size:13px;line-height:1.7">${e(section.content)}</span>
       </div>`).join('');
     return `
@@ -972,24 +982,188 @@ function renderSummaryStructured(summary) {
       ${sectionHtml}
     </div>`;
 }
-export function renderAnalysis(analysis, lang) {
+function asRecord(value) {
+    return value && typeof value === 'object' && !Array.isArray(value) ? value : {};
+}
+function asArray(value) {
+    return Array.isArray(value) ? value : [];
+}
+function num(value) {
+    return typeof value === 'number' && Number.isFinite(value) ? value : null;
+}
+function localizedInsightMessage(insight, report, lang) {
+    const details = asRecord(insight.details);
+    const detailItems = asArray(insight.details);
+    const totalSamples = report?.results?.length || detailItems.length || num(details.total) || 0;
+    const type = insight.type;
+    switch (type) {
+        case 'low_discrimination_all_passed': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `${count} 个断言所有变体均通过，区分度低`
+                : `${count} assertions passed for every variant, so they have low discrimination power`;
+        }
+        case 'low_discrimination_all_failed': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `${count} 个断言所有变体均失败，断言可能过严或配置有误`
+                : `${count} assertions failed for every variant; they may be too strict or misconfigured`;
+        }
+        case 'uniform_scores': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `${count}/${totalSamples} 个评测用例在变体间分差 < 0.5，区分度较低`
+                : `${count}/${totalSamples} samples differ by less than 0.5 across variants, so they have low discrimination power`;
+        }
+        case 'all_pass':
+            return lang === 'zh'
+                ? '所有断言在所有变体上全部通过，断言可能过于宽松'
+                : 'All assertions passed for all variants; the assertions may be too loose';
+        case 'all_fail':
+            return lang === 'zh'
+                ? '所有断言在所有变体上全部失败，请检查断言配置'
+                : 'All assertions failed for all variants; check the assertion configuration';
+        case 'suggest_repeat': {
+            const variant = String(details.variant || 'variant');
+            const min = num(details.min);
+            const max = num(details.max);
+            const range = min != null && max != null ? `${min}~${max}` : '';
+            return lang === 'zh'
+                ? `${variant} 的分数跨度较大${range ? `（${range}）` : ''}，建议使用 --repeat 多轮评测`
+                : `${variant} has a wide score range${range ? ` (${range})` : ''}; run with --repeat to measure variance`;
+        }
+        case 'low_tool_success_rate': {
+            const variant = String(details.variant || 'variant');
+            const rate = num(details.toolSuccessRate);
+            const pct = rate != null ? `${(rate * 100).toFixed(0)}%` : 'low';
+            return lang === 'zh'
+                ? `${variant} 的工具调用成功率仅 ${pct}，可能存在工具选择或参数问题`
+                : `${variant} has only ${pct} tool-call success, which may indicate tool-selection or parameter issues`;
+        }
+        case 'tool_permission_error': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `检测到 ${count} 次工具权限错误，实验结论可能被环境问题污染`
+                : `${count} tool permission errors were detected; the conclusion may be contaminated by environment issues`;
+        }
+        case 'trace_integrity_gap': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `${count} 个 variant 的 trace 覆盖率低于 75%，报告可能不足以解释 agent 行为差异`
+                : `${count} variants have trace coverage below 75%, so the report may not fully explain agent behavior differences`;
+        }
+        case 'agent_assertion_discrimination_low': {
+            const total = num(details.total) || 0;
+            const discriminative = num(details.discriminative) || 0;
+            const pct = total > 0 ? Math.round((discriminative / total) * 100) : 0;
+            return lang === 'zh'
+                ? `agent 断言区分度偏低，只有 ${pct}% 的断言真正拉开了变体差异`
+                : `Agent assertion discrimination is low; only ${pct}% of assertions separate variants`;
+        }
+        case 'agent_assertion_discrimination_ok': {
+            const total = num(details.total) || 0;
+            const discriminative = num(details.discriminative) || 0;
+            const pct = total > 0 ? Math.round((discriminative / total) * 100) : 0;
+            return lang === 'zh'
+                ? `agent 断言区分度达标，${pct}% 的断言能区分变体差异`
+                : `Agent assertion discrimination is healthy; ${pct}% of assertions separate variants`;
+        }
+        case 'efficiency_gap': {
+            const variant = String(details.variant || 'variant');
+            const baseline = String(details.baseline || 'baseline');
+            return lang === 'zh'
+                ? `${variant} 与 ${baseline} 在效率维度存在明显差异`
+                : `${variant} differs materially from ${baseline} on efficiency`;
+        }
+        case 'tool_count_gap': {
+            const variant = String(details.variant || 'variant');
+            const baseline = String(details.baseline || 'baseline');
+            const baseTools = num(details.baseTools);
+            const otherTools = num(details.otherTools);
+            const delta = baseTools != null && otherTools != null ? Math.abs(otherTools - baseTools).toFixed(1) : '';
+            return lang === 'zh'
+                ? `${variant} 与 ${baseline} 的工具调用次数差异明显${delta ? `（差 ${delta} 次）` : ''}`
+                : `${variant} differs materially from ${baseline} in tool-call count${delta ? ` (${delta} calls apart)` : ''}`;
+        }
+        case 'high_cost_sample': {
+            const count = detailItems.length;
+            return lang === 'zh'
+                ? `${count} 个评测用例成本显著高于平均值`
+                : `${count} samples cost materially more than average`;
+        }
+        default:
+            return lang === 'zh'
+                ? `结构化诊断：${type}`
+                : `Structured diagnostic: ${type}`;
+    }
+}
+function localizedSuggestion(insight, lang) {
+    switch (insight.type) {
+        case 'low_discrimination_all_passed':
+            return lang === 'zh'
+                ? '把全通过断言替换为能检测 skill 独有细节的断言，例如特定参数名、配置值或流程要求'
+                : 'Replace always-passing assertions with checks for artifact-specific details, such as parameter names, config values, or workflow requirements';
+        case 'low_discrimination_all_failed':
+            return lang === 'zh'
+                ? '检查断言条件是否正确，或降低匹配要求，避免 broken 用例污染结论'
+                : 'Check whether the assertion condition is correct, or relax the matcher to avoid broken samples contaminating the conclusion';
+        case 'uniform_scores':
+            return lang === 'zh'
+                ? '增加更有挑战性的评测用例，或把 rubric 写得更能区分优劣'
+                : 'Add more challenging samples, or make the rubric more discriminative';
+        case 'all_pass':
+            return lang === 'zh'
+                ? '增加更严格的断言来区分不同变体的质量'
+                : 'Add stricter assertions that can distinguish variant quality';
+        case 'all_fail':
+            return lang === 'zh'
+                ? '优先检查评测配置和断言目标，确认用例不是整体不可达'
+                : 'Check the evaluation config and assertion target first; make sure the sample is reachable';
+        case 'suggest_repeat':
+            return lang === 'zh'
+                ? '使用 --repeat 3 或更多轮次获取方差、置信区间和显著性检验'
+                : 'Run with --repeat 3 or more to obtain variance, confidence intervals, and significance tests';
+        case 'low_tool_success_rate':
+            return lang === 'zh'
+                ? '检查失败工具调用，必要时在 skill 中补充工具选择和参数约束'
+                : 'Inspect failed tool calls and, if needed, add tool-selection and parameter guidance to the artifact';
+        case 'tool_permission_error':
+            return lang === 'zh'
+                ? '先解决工具权限或工作目录问题，再解读分数差异'
+                : 'Fix tool permission or working-directory issues before interpreting score differences';
+        case 'trace_integrity_gap':
+            return lang === 'zh'
+                ? '补齐 turns、toolCalls、timing 和完整输出采集，确保报告能解释行为差异'
+                : 'Capture turns, tool calls, timing, and full output so the report can explain behavior differences';
+        case 'agent_assertion_discrimination_low':
+            return lang === 'zh'
+                ? '重写 agent 断言时优先约束工具路径、关键文件读取和 turns 上限'
+                : 'When rewriting agent assertions, prioritize tool path, key file reads, and turn-limit constraints';
+        default:
+            return lang === 'zh'
+                ? '查看结构化 details 字段定位原因'
+                : 'Inspect the structured details field for the underlying evidence';
+    }
+}
+export function renderAnalysis(report, lang) {
+    const analysis = report?.analysis;
     if (!analysis)
         return '';
-    const { insights, suggestions } = analysis;
-    if ((!insights || insights.length === 0) && (!suggestions || suggestions.length === 0))
+    const { insights } = analysis;
+    if ((!insights || insights.length === 0) && !report)
         return '';
     const issues = (insights || []).filter((insight) => !isConclusion(insight));
     const issueLabel = lang === 'zh' ? '问题与建议' : 'Issues & Suggestions';
-    const safeSuggestions = suggestions || [];
+    const safeSuggestions = issues.map((issue) => localizedSuggestion(issue, lang));
     let issuesHtml = '';
-    if (issues.length > 0 || safeSuggestions.length > 0) {
-        const maxRows = Math.max(issues.length, safeSuggestions.length);
+    if (issues.length > 0) {
+        const maxRows = issues.length;
         const rows = [];
         for (let i = 0; i < maxRows; i++) {
             const issue = issues[i];
             const suggestion = safeSuggestions[i];
             const issueContent = issue
-                ? `${severityDot(issue.severity)}<span>${e(issue.message)}</span>`
+                ? `${severityDot(issue.severity)}<span>${e(localizedInsightMessage(issue, report, lang))}</span>`
                 : '';
             const suggestionContent = suggestion
                 ? e(suggestion)
@@ -1013,9 +1187,12 @@ export function renderAnalysis(analysis, lang) {
       </div>
     `;
     }
-    const summaryHtml = analysis.summary
-        ? renderSummaryStructured(analysis.summary)
+    const generatedSummary = report ? generateAnalysisSummary(report, lang) : undefined;
+    const summaryHtml = generatedSummary
+        ? renderSummaryStructured(generatedSummary, lang)
         : '';
+    if (!summaryHtml && !issuesHtml)
+        return '';
     return `
     <h2 data-i18n="autoAnalysis">${t('autoAnalysis', lang)}</h2>
     ${summaryHtml}
@@ -1168,8 +1345,8 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
         return '';
     const title = lang === 'zh' ? '本次测评：测试用例 × 知识库' : 'This Evaluation: Test Set × Knowledge Base';
     const desc = lang === 'zh'
-        ? '展示本次测评用例和知识库的交互画像——用到哪些知识（使用情况）· 哪些知识想找但没找到或模型表达不确定（盲区）。'
-        : 'How this test set interacts with the KB — which knowledge was exercised (usage) · which was missed or flagged as uncertain (gaps).';
+        ? '展示本次测评用例和知识库的交互画像——用到哪些知识（使用情况）· 哪些知识想找但没找到或任务执行模型表达不确定（盲区）。'
+        : 'How this test set interacts with the KB — which knowledge was exercised (usage) · which knowledge the task execution model missed or flagged as uncertain (gaps).';
     const readHint = lang === 'zh'
         ? '💡 读表：两者同时高 → 知识库内容有问题（有文件但答不出）· 同时低 → 测评用例太浅（没触到复杂场景）· 使用高 + 盲区低 → 理想但警惕用例驯化'
         : '💡 Read together: both high → KB content issues (files exist but can\'t answer) · both low → test set too shallow · high use + low gap → ideal, but beware sample-set overfitting';
@@ -1180,7 +1357,7 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
     ]));
     const signalTypeLabels = {
         failed_search: { zh: '搜索未命中', en: 'Search miss' },
-        explicit_marker: { zh: '模型标记缺口', en: 'Model-flagged gap' },
+        explicit_marker: { zh: '任务执行模型标记缺口', en: 'Execution model flag' },
         hedging: { zh: '表达不确定', en: 'Hedging' },
         repeated_failure: { zh: '反复未命中', en: 'Repeated miss' },
     };
@@ -1274,7 +1451,7 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
             const softShare = pct - weightedPct;
             const weightedHint = lang === 'zh'
                 ? (softShare >= 10
-                    ? `<strong>实际盲区 ${weightedPct}%</strong> · 另 ${softShare}% 为模型表达不确定(软信号,建议对照清单复核)`
+                    ? `<strong>实际盲区 ${weightedPct}%</strong> · 另 ${softShare}% 为任务执行模型表达不确定(软信号,建议对照清单复核)`
                     : `<strong>实际盲区 ${weightedPct}%</strong> · 主要来自确定的搜索未命中`)
                 : (softShare >= 10
                     ? `<strong>real gaps ${weightedPct}%</strong> · another ${softShare}% is hedging (review list below)`
@@ -1297,7 +1474,7 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
                 ? `<div style="font-size:var(--fs-micro);color:var(--text-muted);margin-top:6px">${lang === 'zh' ? `另 ${gap.signals.length - INVENTORY_CAP} 条未展示` : `+${gap.signals.length - INVENTORY_CAP} more not shown`}</div>`
                 : '';
             const detailsLabel = lang === 'zh'
-                ? `展开 ${gap.signals.length} 条证据（按严重度上色: 红=确定 / 黄=模型自述 / 灰=犹豫）`
+                ? `展开 ${gap.signals.length} 条证据（按严重度上色: 红=确定 / 黄=执行模型自述 / 灰=犹豫）`
                 : `Show all ${gap.signals.length} evidence items (red=confirmed · yellow=self-flagged · gray=hedging)`;
             gapInner = `
         <div class="ki-col-header">
@@ -1307,7 +1484,7 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
         <div class="ki-bar" role="progressbar" aria-valuenow="${pct}" aria-valuemin="0" aria-valuemax="100" aria-label="${e(barLabel)}">
           <div class="ki-bar-fill" style="width:${barW}%;background:${pctColor}"></div>
         </div>
-        <div style="font-size:11px;color:var(--text-muted);margin-bottom:4px">${gap.samplesWithGap}/${gap.sampleCount} ${lang === 'zh' ? '个用例出现搜索未命中或表达不确定' : 'samples with search miss / hedging'}</div>
+        <div style="font-size:11px;color:var(--text-muted);margin-bottom:4px">${gap.samplesWithGap}/${gap.sampleCount} ${lang === 'zh' ? '个评测用例出现搜索未命中或表达不确定' : 'samples with search miss / hedging'}</div>
         <div style="font-size:var(--fs-detail);color:var(--text-secondary);margin-bottom:8px">${weightedHint}</div>
         ${typeBadges ? `<div>${typeBadges}</div>` : ''}
         ${inventory ? `<details class="ki-details"><summary>${detailsLabel}</summary>
@@ -1356,11 +1533,11 @@ export function renderGapSection(gapReports, lang) {
         return '';
     const title = lang === 'zh' ? '本次测评的知识盲区' : 'Knowledge gaps in this evaluation';
     const desc = lang === 'zh'
-        ? '本次测评中，哪些知识想找但没找到、或模型表达不确定。数字高不一定代表知识库不全——也可能是测试用例问的领域知识库未覆盖。'
-        : 'Which knowledge the model tried to find but missed, or expressed uncertainty about. High numbers do not necessarily mean the KB is incomplete — the test set may be asking about areas the KB never covered.';
+        ? '本次测评中，哪些知识想找但没找到、或任务执行模型表达不确定。数字高不一定代表知识库不全——也可能是测试用例问的领域知识库未覆盖。'
+        : 'Which knowledge the task execution model tried to find but missed, or expressed uncertainty about. High numbers do not necessarily mean the KB is incomplete — the test set may be asking about areas the KB never covered.';
     const signalTypeLabels = {
         failed_search: { zh: '搜索未命中', en: 'Search miss' },
-        explicit_marker: { zh: '模型标记缺口', en: 'Model-flagged gap' },
+        explicit_marker: { zh: '任务执行模型标记缺口', en: 'Execution model flag' },
         hedging: { zh: '表达不确定', en: 'Hedging' },
         repeated_failure: { zh: '反复未命中', en: 'Repeated miss' },
     };
@@ -1377,7 +1554,7 @@ export function renderGapSection(gapReports, lang) {
         const softSignalShare = pct - weightedPct;
         const weightedHint = lang === 'zh'
             ? (softSignalShare >= 10
-                ? `<strong>实际盲区 ${weightedPct}%</strong> · 另外 ${softSignalShare}% 为模型表达不确定(软信号,建议对照右侧清单复核)`
+                ? `<strong>实际盲区 ${weightedPct}%</strong> · 另外 ${softSignalShare}% 为任务执行模型表达不确定(软信号,建议对照右侧清单复核)`
                 : `<strong>实际盲区 ${weightedPct}%</strong> · 主要来自确定的搜索未命中`)
             : (softSignalShare >= 10
                 ? `<strong>real gaps ${weightedPct}%</strong> · another ${softSignalShare}% is hedging (soft signals — review the list on the right)`
@@ -1423,7 +1600,7 @@ export function renderGapSection(gapReports, lang) {
         <div style="width:${barW}%;height:100%;background:${pctColor};border-radius:4px"></div>
       </div>
       <div style="font-size:var(--fs-detail);color:var(--text-muted);margin-bottom:4px">
-        ${report.samplesWithGap} / ${report.sampleCount} ${lang === 'zh' ? '个用例出现搜索未命中或表达不确定' : 'samples with search miss or hedging'}
+        ${report.samplesWithGap} / ${report.sampleCount} ${lang === 'zh' ? '个评测用例出现搜索未命中或表达不确定' : 'samples with search miss or hedging'}
       </div>
       <div style="font-size:var(--fs-detail);color:var(--text-secondary);margin-bottom:10px">${weightedHint}</div>
       ${typeBadges ? `<div style="margin-bottom:10px">${typeBadges}</div>` : ''}