oh-my-knowledge 0.48.0 → 0.49.0

package/dist/renderer/skill-health-renderer.js

@@ -11,7 +11,7 @@
  *   - layout.ts 的整体 HTML skeleton + CSS (ki-card / ki-col / ki-bar 全部现成)
  *   - 报告视觉和 eval HTML 保持一致，读者无学习成本切换
  */
-import { confidenceOf } from '../observability/skill-health-analyzer.js';
+import { confidenceOf, toolStabilityOf } from '../observability/skill-health-analyzer.js';
 import { COLORS, e, t } from './layout.js';
 import { icon } from './icons.js';
 import { reportShell } from './report-shell.js';
@@ -79,11 +79,32 @@ function renderSkillCard(skill, variantColor, lang) {
     // ─── 右栏: gap signals ────────────────────────
     const softShare = gapPct - weightedPct;
     const failureRatePct = Math.round(skill.toolFailureRate * 100);
-    const instabilityNote = skill.stability === 'very-unstable'
+    const resolvedToolCalls = skill.toolResolvedCount
+        ?? Math.max(0, skill.toolCallCount - (skill.toolUnknownCount ?? 0));
+    const cancelledToolCalls = skill.toolCancelledCount ?? 0;
+    const comparableToolCalls = Math.max(0, resolvedToolCalls - cancelledToolCalls);
+    const unknownToolCalls = skill.toolUnknownCount ?? Math.max(0, skill.toolCallCount - resolvedToolCalls);
+    const stability = toolStabilityOf(skill.toolFailureRate, comparableToolCalls, skill.toolCallCount);
+    const stabilityNote = stability === 'very-unstable'
         ? ` · <span style="color:var(--red)">${lang === 'zh' ? `失败率 ${failureRatePct}%,gap 可能是环境问题` : `failure rate ${failureRatePct}%, gap likely env issue`}</span>`
-        : skill.stability === 'unstable'
+        : stability === 'unstable'
             ? ` · <span style="color:var(--yellow)">${lang === 'zh' ? `失败率 ${failureRatePct}%,gap 可能含噪声` : `failure rate ${failureRatePct}%, gap may be noisy`}</span>`
-            : '';
+            : stability === 'unknown'
+                ? ` · <span style="color:var(--text-faint)">${cancelledToolCalls > 0
+                    ? (lang === 'zh' ? '没有可比较结果，工具调用已取消' : 'no comparable outcomes; calls cancelled')
+                    : (lang === 'zh' ? '工具结果状态不可测' : 'tool outcomes unavailable')}</span>`
+                : '';
+    const outcomeEvidenceNote = stability === 'unknown'
+        ? ''
+        : cancelledToolCalls > 0 || unknownToolCalls > 0
+            ? ` · <span style="color:var(--text-faint)">${lang === 'zh'
+                ? `${comparableToolCalls}/${skill.toolCallCount} 次结果可比较${cancelledToolCalls > 0 ? `，${cancelledToolCalls} 次取消` : ''}`
+                : `${comparableToolCalls}/${skill.toolCallCount} outcomes comparable${cancelledToolCalls > 0 ? `, ${cancelledToolCalls} cancelled` : ''}`}</span>`
+            : comparableToolCalls > 0 && comparableToolCalls < 5
+                ? ` · <span style="color:var(--text-faint)">${lang === 'zh'
+                    ? `仅 ${comparableToolCalls} 次结果，结论置信度低`
+                    : `only ${comparableToolCalls} outcomes, low confidence`}</span>`
+                : '';
     const weightedHintBase = softShare >= 10
         ? `<strong>${lang === 'zh' ? '加权盲区' : 'weighted gap'} ${weightedPct}%</strong> · ${softShare}% ${lang === 'zh' ? '为软信号(建议复核)' : 'soft signals (review)'}`
         : `<strong>${lang === 'zh' ? '加权盲区' : 'weighted gap'} ${weightedPct}%</strong> · ${lang === 'zh' ? '以硬证据为主' : 'mostly hard evidence'}`;
@@ -92,7 +113,7 @@ function renderSkillCard(skill, variantColor, lang) {
             ? `${skill.segmentCount} 段，可信度${confidence === 'underpowered' ? '不足，仅供参考' : '偏低'}`
             : `N=${skill.segmentCount}, ${confidence} confidence`}</span>`
         : '';
-    const weightedHint = weightedHintBase + instabilityNote + confNote;
+    const weightedHint = weightedHintBase + stabilityNote + outcomeEvidenceNote + confNote;
     const signalBadges = Object.entries(gap.byType)
         .filter(([, n]) => n > 0)
         .map(([k, n]) => {
@@ -118,19 +139,25 @@ function renderSkillCard(skill, variantColor, lang) {
     ${signalBadges ? `<div>${signalBadges}</div>` : ''}
   `;
     // 失败率 badge: 用 skill.stability 决定颜色
-    const stabilityColor = skill.stability === 'very-unstable'
+    const stabilityColor = stability === 'very-unstable'
         ? 'var(--red)'
-        : skill.stability === 'unstable'
+        : stability === 'unstable'
             ? 'var(--yellow)'
             : 'var(--text-muted)';
-    const failureLabel = skill.toolCallCount > 0
-        ? `${skill.toolFailureCount}/${skill.toolCallCount} ${lang === 'zh' ? '失败' : 'failed'} (${failureRatePct}%)`
-        : `0 ${lang === 'zh' ? '次工具调用' : 'tool calls'}`;
+    const failureLabel = skill.toolCallCount > 0 && comparableToolCalls === 0
+        ? cancelledToolCalls > 0 && unknownToolCalls === 0
+            ? `${cancelledToolCalls} ${lang === 'zh' ? '次取消' : 'cancelled'}`
+            : cancelledToolCalls > 0
+                ? `${cancelledToolCalls} ${lang === 'zh' ? '次取消' : 'cancelled'} · ${unknownToolCalls} ${lang === 'zh' ? '次状态未知' : 'unknown'}`
+                : `${skill.toolCallCount} ${lang === 'zh' ? '次状态未知' : 'unknown outcomes'}`
+        : comparableToolCalls > 0
+            ? `${skill.toolFailureCount}/${comparableToolCalls} ${lang === 'zh' ? '失败' : 'failed'} (${failureRatePct}%)${cancelledToolCalls > 0 ? ` · ${cancelledToolCalls} ${lang === 'zh' ? '取消' : 'cancelled'}` : ''}`
+            : `0 ${lang === 'zh' ? '次工具调用' : 'tool calls'}`;
     // ─── 成本/耗时行(第四轴,skill 维度聚合) ─────────
     // 旧 JSON (加 usage 字段前生成的) 缺此字段,降级为全 0
     const u = skill.usage ?? {
         inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0,
-        totalTokens: 0, durationMs: 0, numTurns: 0,
+        totalTokens: 0, tokenObservedSegmentCount: 0, tokenCoverage: 0, durationMs: 0, numTurns: 0,
         avgTokensPerSegment: 0, avgDurationMsPerSegment: 0,
     };
     const billableTokens = (u.inputTokens ?? 0) + (u.outputTokens ?? 0);
@@ -140,7 +167,11 @@ function renderSkillCard(skill, variantColor, lang) {
     const durSec = (u.durationMs / 1000).toFixed(1);
     const avgDurSec = (u.avgDurationMsPerSegment / 1000).toFixed(1);
     const cachedSeg = cachedTokens > 0 ? ` + ${cachedK}k ${lang === 'zh' ? '缓存' : 'cached'}` : '';
-    const usageLine = `${billK}k tokens${cachedSeg} · ${durSec}s (${lang === 'zh' ? '均' : 'avg'} ${avgDurSec}s/${lang === 'zh' ? '段' : 'seg'}) · ${u.numTurns} ${lang === 'zh' ? '轮次' : 'turns'}`;
+    const tokenCoverage = u.tokenCoverage ?? 0;
+    const tokenText = tokenCoverage > 0
+        ? `${billK}k tokens${cachedSeg}${tokenCoverage < 1 ? ` (${Math.round(tokenCoverage * 100)}% ${lang === 'zh' ? '覆盖' : 'coverage'})` : ''}`
+        : lang === 'zh' ? 'tokens —（未观测）' : 'tokens — (unobserved)';
+    const usageLine = `${tokenText} · ${durSec}s (${lang === 'zh' ? '均' : 'avg'} ${avgDurSec}s/${lang === 'zh' ? '段' : 'seg'}) · ${u.numTurns} ${lang === 'zh' ? '轮次' : 'turns'}`;
     // ─── Card 结构 ───────────────────────────────
     return `
   <div class="ki-card" style="border-left:3px solid ${variantColor}">
@@ -231,8 +262,34 @@ export function renderSkillHealthReport(report, lang = 'zh') {
     <div class="rs-stat rs-stat--total"><span class="rs-stat-num">${meta.toolCallCount}</span><span class="rs-stat-label">${zh ? '工具调用' : 'Tool calls'}</span></div>
     <div class="rs-stat ${overall.weightedGapRate >= 0.3 ? 'rs-stat--fail' : overall.weightedGapRate >= 0.1 ? 'rs-stat--warn' : 'rs-stat--pass'}"><span class="rs-stat-num">${Math.round(overall.weightedGapRate * 100)}%</span><span class="rs-stat-label">${zh ? '加权盲区' : 'Weighted gap'}</span></div>
   </div>`;
+    const ingestion = meta.ingestion;
+    const ingestionNotice = ingestion && (ingestion.malformedRecordCount > 0
+        || ingestion.ignoredValueCount > 0
+        || ingestion.unknownEventCount > 0)
+        ? `<div style="margin:12px 0;padding:10px 12px;border:1px solid var(--yellow);border-radius:8px;background:var(--yellow-bg);color:var(--text-secondary);font-size:13px">
+        <strong style="color:var(--yellow)">${zh ? '观测输入需要复核' : 'Observation input needs review'}</strong>
+        <span style="margin-left:8px">${zh
+            ? `${ingestion.malformedRecordCount} 条格式损坏记录，${ingestion.ignoredValueCount} 个非对象值，${ingestion.unknownEventCount} 个未识别事件。`
+            : `${ingestion.malformedRecordCount} malformed records, ${ingestion.ignoredValueCount} non-object values, ${ingestion.unknownEventCount} unrecognized events.`}</span>
+      </div>`
+        : '';
+    const timestampedSegmentCount = meta.timestampedSegmentCount
+        ?? (meta.timeRange.from && meta.timeRange.to ? meta.segmentCount : 0);
+    const timestampCoverage = meta.timestampCoverage
+        ?? (meta.segmentCount > 0 ? timestampedSegmentCount / meta.segmentCount : 1);
+    const excludedUntimestampedSegmentCount = meta.excludedUntimestampedSegmentCount ?? 0;
+    const timestampNotice = timestampCoverage < 1 || excludedUntimestampedSegmentCount > 0
+        ? `<div style="margin:12px 0;padding:10px 12px;border:1px solid var(--yellow);border-radius:8px;background:var(--yellow-bg);color:var(--text-secondary);font-size:13px">
+        <strong style="color:var(--yellow)">${zh ? '时间范围不完整' : 'Incomplete time coverage'}</strong>
+        <span style="margin-left:8px">${zh
+            ? `${meta.segmentCount} 个入选片段中，${timestampedSegmentCount} 个具有可观测时间；时间筛选另排除了 ${excludedUntimestampedSegmentCount} 个无时间戳片段。`
+            : `${timestampedSegmentCount} of ${meta.segmentCount} selected segments have observed timestamps; time filtering excluded ${excludedUntimestampedSegmentCount} additional untimestamped segments.`}</span>
+      </div>`
+        : '';
     const body = [
         statsBar,
+        ingestionNotice,
+        timestampNotice,
         renderWatermark(report, lang),
         `<div class="rs-panel">
       <div class="rs-panel-title">${zh ? '各 skill 健康度' : 'Per-skill health'} <small>${zh ? '按使用量降序' : 'sorted by usage'}</small></div>

package/dist/renderer/summary.js

@@ -4,6 +4,7 @@ import { generateAnalysisSummary } from '../analysis/report-diagnostics.js';
 import { pValueCategory } from '../eval-core/statistics.js';
 import { ciLevelLabel } from '../eval-core/bootstrap.js';
 import { computeVerdict, medianStabilityCV, STABILITY_UNSTABLE_CV, ENSEMBLE_STRONG_PEARSON, ENSEMBLE_DISSENT_PEARSON } from '../eval-core/verdict.js';
+import { ownRecordValue, setOwnRecordValue } from '../shared/record-count.js';
 /**
  * Verdict pill — sticky banner at the top of the HTML report giving the same
  * one-line conclusion as the `omk eval` CLI. Both surfaces share the
@@ -522,9 +523,13 @@ export function renderSummaryCards(variants, summary, lang, variance) {
         // Cost — 只算 execution cost(judge cost 是工具开销,不算 skill 成本)。
         const execCost = s.totalExecCostUSD || 0;
         const costReported = s.execCostReported !== false;
-        const hasCost = execCost > 0 || (s.avgTotalTokens || 0) > 0;
+        const tokenCoverage = s.tokenUsageCoverageRate ?? 1;
+        const hasMeasuredTokens = tokenCoverage > 0;
+        const hasCost = (s.totalSamples || 0) > 0;
         const costPerSample = (s.totalSamples || 0) > 0 ? execCost / s.totalSamples : execCost;
-        const tokenDetail = `${fmtNum(s.avgTotalTokens)} tokens/${t('tokPerReq', lang).replace('tokens/', '')}`;
+        const tokenDetail = hasMeasuredTokens
+            ? `${fmtNum(s.avgTotalTokens)} tokens/${t('tokPerReq', lang).replace('tokens/', '')}${tokenCoverage < 1 ? ` · ${Math.round(tokenCoverage * 100)}% ${lang === 'zh' ? '覆盖' : 'coverage'}` : ''}`
+            : lang === 'zh' ? 'tokens 未观测' : 'tokens unobserved';
         const costDim = hasCost
             ? { label: dimText('dimCost', lang), value: costReported ? fmtCost(execCost) : fmtCost(0, false), valueColor: costReported ? 'var(--text-primary)' : 'var(--text-muted)', primary: false, bg: '', detailLines: [tokenDetail], title: costReported ? undefined : costUnreportedTooltip, icon: 'cost' }
             : { label: dimText('dimCost', lang), value: 'N/A', valueColor: 'var(--text-muted)', primary: false, bg: '', detailLines: [], icon: 'cost' };
@@ -1671,8 +1676,22 @@ export function renderAgentOverview(variants, summary, lang) {
             return '';
         const color = COLORS[i % COLORS.length];
         const avgTools = stats.avgToolCalls ?? 0;
-        const successRate = stats.toolSuccessRate != null ? `${(stats.toolSuccessRate * 100).toFixed(0)}%` : '—';
-        const successRateColor = (stats.toolSuccessRate ?? 1) >= 0.8 ? 'var(--green)' : 'var(--red)';
+        const outcomeSuffix = [
+            stats.avgToolCancelled
+                ? lang === 'zh' ? `${stats.avgToolCancelled} 次取消` : `${stats.avgToolCancelled} cancelled`
+                : '',
+            stats.avgToolUnknown
+                ? lang === 'zh' ? `${stats.avgToolUnknown} 次状态未知` : `${stats.avgToolUnknown} unknown`
+                : '',
+        ].filter(Boolean).map((value) => ` · ${value}`).join('');
+        const successRate = stats.toolSuccessRate != null
+            ? `${(stats.toolSuccessRate * 100).toFixed(0)}% OK${outcomeSuffix}`
+            : stats.avgToolCancelled || stats.avgToolUnknown
+                ? outcomeSuffix.replace(/^ · /, '')
+                : lang === 'zh' ? '状态未知' : 'unknown';
+        const successRateColor = stats.toolSuccessRate == null
+            ? 'var(--text-muted)'
+            : stats.toolSuccessRate >= 0.8 ? 'var(--green)' : 'var(--red)';
         const turns = stats.avgFullNumTurns ?? stats.avgNumTurns ?? 0;
         const distributionEntries = Object.entries(stats.toolDistribution || {}).sort((a, b) => b[1] - a[1]);
         const distSummary = distributionEntries.length > 0
@@ -1686,7 +1705,7 @@ export function renderAgentOverview(variants, summary, lang) {
         <span class="ctx-sep">·</span>
         <span>${avgTools} ${lang === 'zh' ? '工具/次' : 'tools/req'}</span>
         <span class="ctx-sep">·</span>
-        <span style="color:${successRateColor}">${successRate} OK</span>
+        <span style="color:${successRateColor}">${successRate}</span>
       </span>
     </div>`;
     }).join('');
@@ -1726,7 +1745,8 @@ export function renderCoverageSection(coverage, lang) {
         const uncoveredByType = {};
         for (const entry of knowledgeCoverage.entries.filter((item) => !item.accessed)) {
             const category = entry.path.startsWith('repos/') ? 'code' : entry.type;
-            (uncoveredByType[category] = uncoveredByType[category] || []).push(entry.path);
+            const files = ownRecordValue(uncoveredByType, category) ?? setOwnRecordValue(uncoveredByType, category, []);
+            files.push(entry.path);
         }
         const hintLines = [];
         const typeLabels = lang === 'zh'
@@ -1850,7 +1870,8 @@ export function renderKnowledgeInteractionSection(coverage, gapReports, lang) {
             const uncoveredByType = {};
             for (const entry of cov.entries.filter((item) => !item.accessed)) {
                 const category = entry.path.startsWith('repos/') ? 'code' : entry.type;
-                (uncoveredByType[category] = uncoveredByType[category] || []).push(entry.path);
+                const files = ownRecordValue(uncoveredByType, category) ?? setOwnRecordValue(uncoveredByType, category, []);
+                files.push(entry.path);
             }
             const typeLabels = lang === 'zh'
                 ? { principle: '原则文件', semantic: '语义索引', design: '设计文档', code: '代码路径', script: '脚本工具', other: '其他知识' }

package/dist/renderer/table.js

@@ -113,11 +113,28 @@ function renderVariantBlock(d, variantName, comparedTo, lang) {
     const totalMs = d.timing?.totalMs || d.durationMs;
     const metaParts = [];
     if (d.numToolCalls != null && d.numToolCalls > 0) {
-        const srColor = (d.toolSuccessRate ?? 1) >= 0.8 ? 'var(--green)' : 'var(--red)';
-        const srText = d.toolSuccessRate != null ? `${(d.toolSuccessRate * 100).toFixed(0)}%` : '—';
-        metaParts.push(`<div class="sc-meta-row"><span class="sc-meta-name">${lang === 'zh' ? '工具调用' : 'Tools'}</span>${d.numToolCalls} <span style="color:${srColor}">(${srText} OK)</span></div>`);
+        const hasResolvedOutcome = d.toolSuccessRate != null;
+        const srColor = hasResolvedOutcome
+            ? d.toolSuccessRate >= 0.8 ? 'var(--green)' : 'var(--red)'
+            : 'var(--text-muted)';
+        const outcomeSuffix = [
+            (d.numToolCancelled ?? 0) > 0
+                ? lang === 'zh' ? `${d.numToolCancelled} 次取消` : `${d.numToolCancelled} cancelled`
+                : '',
+            (d.numToolUnknown ?? 0) > 0
+                ? lang === 'zh' ? `${d.numToolUnknown} 次状态未知` : `${d.numToolUnknown} unknown`
+                : '',
+        ].filter(Boolean).map((value) => ` · ${value}`).join('');
+        const outcomeText = hasResolvedOutcome
+            ? `${(d.toolSuccessRate * 100).toFixed(0)}% OK${outcomeSuffix}`
+            : outcomeSuffix
+                ? outcomeSuffix.replace(/^ · /, '')
+                : lang === 'zh'
+                    ? `${d.numToolCalls} 次状态未知`
+                    : `${d.numToolCalls} unknown`;
+        metaParts.push(`<div class="sc-meta-row"><span class="sc-meta-name">${lang === 'zh' ? '工具调用' : 'Tools'}</span>${d.numToolCalls} <span style="color:${srColor}">(${outcomeText})</span></div>`);
     }
-    metaParts.push(`<div class="sc-meta-row"><span class="sc-meta-name">${lang === 'zh' ? 'Tokens' : 'Tokens'}</span>${fmtNum(d.totalTokens)}</div>`);
+    metaParts.push(`<div class="sc-meta-row"><span class="sc-meta-name">${lang === 'zh' ? 'Tokens' : 'Tokens'}</span>${d.tokenUsageReportedByExecutor === false ? '—' : fmtNum(d.totalTokens)}</div>`);
     metaParts.push(`<div class="sc-meta-row"><span class="sc-meta-name">${lang === 'zh' ? '耗时' : 'Duration'}</span>${fmtDuration(totalMs)}</div>`);
     // Error message
     const errorHtml = !d.ok && d.error

package/dist/renderer/test-view.d.ts

@@ -49,6 +49,7 @@ export declare function sampleVerdict(details: AssertionDetail[] | undefined): '
  * 避免使用者误以为是要修的真 bug。
  */
 export declare function isTripwireSample(snapshot: SampleSnapshot | undefined, variant: VariantResult): boolean;
+export declare function renderTurnContent(turn: TurnInfo, lang: Lang): string;
 export declare function renderTestView(report: EvaluationReport, lang: Lang): string;
 /** 单测视角的 CSS。
  *

package/dist/renderer/test-view.js

@@ -15,6 +15,7 @@
  */
 import { e, t } from './layout.js';
 import { icon } from './icons.js';
+import { isToolCallFailure, toolCallStatus } from '../shared/tool-call-status.js';
 // ──────────────────────────────────────────────────────────────────
 // Assertion → Turn 绑定:扫执行轨迹找断言匹配的 turn 编号
 // ──────────────────────────────────────────────────────────────────
@@ -182,10 +183,30 @@ function renderDesignSection(snapshot, lang) {
        <div class="tv-mocks">${snapshot.mocks.map((m, i) => renderMock(m, i + 1)).join('')}</div>
        </details>`
         : '';
+    const executionContract = snapshot.cwd
+        || snapshot.environment
+        || snapshot.mocksStrict !== undefined
+        || (snapshot.allowedTools && snapshot.allowedTools.length > 0)
+        || (snapshot.expectedTools && snapshot.expectedTools.length > 0)
+        ? `<details class="tv-collapsible"><summary>${lang === 'zh' ? '执行契约' : 'Execution contract'}</summary>
+       <pre class="tv-pre">${e(JSON.stringify({
+            ...(snapshot.cwd ? { cwd: snapshot.cwd } : {}),
+            ...(snapshot.environment ? { environment: snapshot.environment } : {}),
+            ...(snapshot.mocksStrict !== undefined ? { mocksStrict: snapshot.mocksStrict } : {}),
+            ...(snapshot.allowedTools ? { allowedTools: snapshot.allowedTools } : {}),
+            ...(snapshot.expectedTools ? { expectedTools: snapshot.expectedTools } : {}),
+        }, null, 2))}</pre>
+       </details>`
+        : '';
+    const dimensionsBlock = snapshot.dimensions && Object.keys(snapshot.dimensions).length > 0
+        ? `<details class="tv-collapsible"><summary>${lang === 'zh' ? '评分维度' : 'Scoring dimensions'}</summary><pre class="tv-pre">${e(JSON.stringify(snapshot.dimensions, null, 2))}</pre></details>`
+        : '';
     return `<div class="tv-section--design-body">
     <div class="tv-prompt"><strong>${lang === 'zh' ? '提示词(Prompt):' : 'Prompt:'}</strong> <span>${e(snapshot.prompt)}</span></div>
     ${snapshot.rubric ? `<div class="tv-rubric"><strong>${lang === 'zh' ? '评分标准(Rubric):' : 'Rubric:'}</strong> <span>${e(snapshot.rubric)}</span></div>` : ''}
     ${snapshot.context ? `<details class="tv-collapsible"><summary>${lang === 'zh' ? '上下文(Context)' : 'Context'}</summary><pre class="tv-pre">${e(snapshot.context)}</pre></details>` : ''}
+    ${executionContract}
+    ${dimensionsBlock}
     ${meta.length > 0 ? `<div class="tv-meta-pills">${meta.join(' ')}</div>` : ''}
     ${expectedAssertions ? `<div class="tv-expected">
       <div class="tv-expected-h">${lang === 'zh' ? '期望（assertions）：' : 'Expected (assertions):'}</div>
@@ -270,17 +291,14 @@ function renderExecutionTrace(variant, detailsByTurn, matchedMockNumbersByTurn,
     }
     const stepsHtml = turns.map((turn, i) => {
         const turnNum = i + 1;
-        const isAssistant = turn.role === 'assistant';
-        const thinkingBlock = isAssistant && turn.content
-            ? `<details class="tv-thinking"><summary>${lang === 'zh' ? '展开 thinking' : 'show thinking'} (${turn.content.length} ${lang === 'zh' ? '字' : 'chars'})</summary><div class="tv-thinking-body">${e(turn.content)}</div></details>`
-            : '';
-        const calls = (turn.toolCalls ?? []).map((tc) => renderToolCall(tc, turnNum, matchedMockNumbersByTurn.get(turnNum) ?? [])).join('');
+        const contentBlock = renderTurnContent(turn, lang);
+        const calls = (turn.toolCalls ?? []).map((tc) => renderToolCall(tc, turnNum, matchedMockNumbersByTurn.get(turnNum) ?? [], lang)).join('');
         const assertsForTurn = (detailsByTurn.get(turnNum) ?? []);
         const turnAssertsHtml = assertsForTurn.map((d) => `<div class="tv-trace-assert ${d.passed ? 'tv-assert--pass' : 'tv-assert--fail'}">${d.passed ? '✅' : '❌'} <code>${e(d.type)}</code> <code class="tv-assert-value">${e(typeof d.value === 'object' ? JSON.stringify(d.value) : String(d.value))}</code></div>`).join('');
         const durationTag = turn.durationMs ? `<span class="tv-turn-dur">${(turn.durationMs / 1000).toFixed(1)}s</span>` : '';
         return `<div class="tv-turn">
       <div class="tv-turn-h">Turn ${turnNum} · <code>${e(turn.role)}</code> ${durationTag}</div>
-      ${thinkingBlock}
+      ${contentBlock}
       ${calls}
       ${turnAssertsHtml}
     </div>`;
@@ -294,14 +312,31 @@ function renderExecutionTrace(variant, detailsByTurn, matchedMockNumbersByTurn,
     ${finalOutput}
   </div>`;
 }
-function renderToolCall(tc, _turnNum, mockHitsHere) {
+export function renderTurnContent(turn, lang) {
+    if (!turn.content)
+        return '';
+    if (turn.role === 'assistant') {
+        return `<details class="tv-thinking"><summary>${lang === 'zh' ? '展开 thinking' : 'show thinking'} (${turn.content.length} ${lang === 'zh' ? '字' : 'chars'})</summary><div class="tv-thinking-body">${e(turn.content)}</div></details>`;
+    }
+    const label = turn.role === 'user'
+        ? lang === 'zh' ? '用户输入' : 'User input'
+        : lang === 'zh' ? '工具结果' : 'Tool result';
+    return `<div class="tv-turn-content"><div class="tv-turn-content-h">${label}</div><pre class="tv-pre">${e(turn.content)}</pre></div>`;
+}
+function renderToolCall(tc, _turnNum, mockHitsHere, lang) {
     const inputRepr = typeof tc.input === 'string' ? tc.input : JSON.stringify(tc.input ?? '');
     const outputRepr = typeof tc.output === 'string' ? tc.output : JSON.stringify(tc.output ?? '');
     const mockBadge = mockHitsHere.length > 0
         ? `<span class="tv-mock-badge" title="matched mock #${mockHitsHere.join(',#')}">✓ mock#${mockHitsHere.join(',#')}</span>`
         : '';
-    return `<div class="tv-call ${tc.success ? '' : 'tv-call--err'}">
-    <div class="tv-call-h">🔧 <code>${e(tc.tool)}</code> ${mockBadge}</div>
+    const status = toolCallStatus(tc);
+    const statusLabel = status === 'unknown'
+        ? `<span class="tv-status">${e(lang === 'zh' ? '状态未知' : 'unknown status')}</span>`
+        : status === 'cancelled'
+            ? `<span class="tv-status">${e(lang === 'zh' ? '已取消' : 'cancelled')}</span>`
+            : '';
+    return `<div class="tv-call ${isToolCallFailure(tc) ? 'tv-call--err' : ''}">
+    <div class="tv-call-h">🔧 <code>${e(tc.tool)}</code> ${statusLabel} ${mockBadge}</div>
     <details><summary>input</summary><pre class="tv-pre">${e(inputRepr.length > 1000 ? inputRepr.slice(0, 1000) + '…' : inputRepr)}</pre></details>
     <details><summary>output</summary><pre class="tv-pre">${e(outputRepr.length > 1000 ? outputRepr.slice(0, 1000) + '…' : outputRepr)}</pre></details>
   </div>`;

package/dist/server/indexed-report-store.js

@@ -4,7 +4,7 @@
  * 三数据源合并、按 id dedup(live 盖卡片,优先级 项目 > 全局 > 卡片):
  *   - live(当前项目 `.omk/reports`) + live(全局 `~/.oh-my-knowledge/reports`):完整真身,永远新鲜(无需 backfill);
  *   - 索引卡片(state/artifact-index/report/):覆盖**别的项目**的报告(当前 studio 的 cwd live 扫不到),
- *     卡片只含 meta+summary(无 results),详情按 `card.path` 读真身。
+ *     卡片只用于发现 `card.path`;列表与详情都回源并严格解析真身。
  *
  * 与 createOverlayReportStore(项目盖全局、记录优先)的区别:这里是**机器级 merge**(看全本机),不是项目优先;
  * 可比性红线在比较层(verdict / findByVariant 锁同口径)而非视图层,故跨项目 merge 列表合法。
@@ -13,7 +13,7 @@
 import { existsSync, statSync, readdirSync } from 'node:fs';
 import { dirname, join } from 'node:path';
 import { createFileStore } from './report-store.js';
-import { artifactIndexDir, listReportCards, listLiveReportCards, cardToReportDocument, removeReportCard, cardTargetSentinel } from '../eval-core/artifact-index.js';
+import { artifactIndexDir, listReportCards, listLiveReportCards, removeReportCard, cardTargetSentinel } from '../eval-core/artifact-index.js';
 function isEvaluation(r) {
     return r.kind === 'evaluation';
 }
@@ -47,19 +47,20 @@ export function createIndexedReportStore({ projectDir, globalDir }) {
             return 'none';
         }
     }
-    function cardDocs() {
+    async function cardDocs() {
         const fp = cardFingerprint();
         if (fp === cachedFp && cachedCards)
-            return cachedCards;
-        // 机器级 list 展示用 live 卡片:真身被带外删/移走的悬空卡片不进列表(把卡片当活指针)。
-        // get() 仍走 raw listReportCards 做 by-id best-effort(直链命中真身则给真身、悬空降级壳)。
-        const docs = listLiveReportCards().map(cardToReportDocument);
+            return structuredClone(cachedCards);
+        // 卡片只是发现指针。机器级 list 也必须回源并走 createFileStore 的 canonical parser，
+        // 不能把索引里的 meta/summary 壳伪装成完整 ReportDocument。
+        const loaded = await Promise.all(listLiveReportCards().map((card) => createFileStore(dirname(card.path)).get(card.id)));
+        const docs = loaded.filter((doc) => doc !== null);
         cachedFp = fp;
         cachedCards = docs;
-        return docs;
+        return structuredClone(docs);
     }
     async function list() {
-        const [proj, glob] = await Promise.all([project.list(), global.list()]);
+        const [proj, glob, indexed] = await Promise.all([project.list(), global.list(), cardDocs()]);
         const seen = new Set();
         const merged = [];
         for (const r of proj) {
@@ -74,7 +75,7 @@ export function createIndexedReportStore({ projectDir, globalDir }) {
                 merged.push(r);
             }
         }
-        for (const r of cardDocs()) {
+        for (const r of indexed) {
             if (!seen.has(r.id)) {
                 seen.add(r.id);
                 merged.push(r);
@@ -88,19 +89,14 @@ export function createIndexedReportStore({ projectDir, globalDir }) {
         const live = (await project.get(id)) ?? (await global.get(id));
         if (live)
             return live;
-        // 别项目:按卡片 path 读真身;悬空(项目被移动/删除)→ 返回卡片壳(results:[])不抛、不 500。
+        // 别项目:按卡片 path 读真身。悬空或损坏都 fail closed，不从 scratch 卡片伪造报告壳。
         const card = listReportCards().find((c) => c.id === id);
         if (!card)
             return null;
-        const fromFile = await createFileStore(dirname(card.path)).get(id);
-        return fromFile ?? cardToReportDocument(card);
+        return createFileStore(dirname(card.path)).get(id);
     }
     async function exists(id) {
-        if (await project.exists(id))
-            return true;
-        if (await global.exists(id))
-            return true;
-        return listReportCards().some((c) => c.id === id);
+        return (await get(id)) !== null;
     }
     async function save(id, report) {
         return project.save(id, report);

package/dist/server/job-store.js

@@ -1,11 +1,18 @@
 import { promises as fs } from 'node:fs';
 import { existsSync, mkdirSync } from 'node:fs';
 import { join } from 'node:path';
+import { writeJsonFileAtomicAsync } from '../shared/atomic-json.js';
+import { isSafeEvaluationJobId, isValidEvaluationJob, } from '../shared/evaluation-job.js';
+import { KeyedMutex } from '../shared/keyed-mutex.js';
 async function listJsonFiles(dir) {
     try {
         const entries = await fs.readdir(dir, { withFileTypes: true });
         return entries
-            .filter((entry) => entry.isFile() && entry.name.endsWith('.json'))
+            .filter((entry) => {
+            if (!entry.isFile() || !entry.name.endsWith('.json'))
+                return false;
+            return isSafeJobId(entry.name.slice(0, -'.json'.length));
+        })
             .map((entry) => entry.name);
     }
     catch {
@@ -13,53 +20,84 @@ async function listJsonFiles(dir) {
     }
 }
 export function createFileJobStore(dir) {
+    const mutations = new KeyedMutex();
     if (!existsSync(dir)) {
         mkdirSync(dir, { recursive: true });
     }
     async function list() {
         const files = await listJsonFiles(dir);
-        const jobs = await Promise.all(files.map(async (file) => JSON.parse(await fs.readFile(join(dir, file), 'utf-8'))));
+        const jobs = (await Promise.all(files.map(async (file) => {
+            const id = file.slice(0, -'.json'.length);
+            try {
+                return parseJob(await fs.readFile(join(dir, file), 'utf-8'), id);
+            }
+            catch {
+                return null;
+            }
+        }))).filter((job) => job !== null);
         return jobs.sort((a, b) => (b.createdAt || '').localeCompare(a.createdAt || ''));
     }
     async function get(id) {
+        const path = jobFilePath(dir, id);
+        if (!path)
+            return null;
         try {
-            return JSON.parse(await fs.readFile(join(dir, `${id}.json`), 'utf-8'));
+            return parseJob(await fs.readFile(path, 'utf-8'), id);
         }
         catch {
             return null;
         }
     }
+    async function saveUnlocked(id, job) {
+        const path = jobFilePath(dir, id);
+        if (!path || !isValidEvaluationJob(job, id))
+            throw new Error('invalid job');
+        await writeJsonFileAtomicAsync(path, job);
+    }
     async function save(id, job) {
-        await fs.writeFile(join(dir, `${id}.json`), JSON.stringify(job, null, 2));
+        await mutations.run(id, () => saveUnlocked(id, job));
     }
     async function update(id, mutator) {
-        const current = await get(id);
-        if (!current)
+        if (!isSafeJobId(id))
             return null;
-        const updated = mutator(current);
-        await save(id, updated);
-        return updated;
+        return mutations.run(id, async () => {
+            const current = await get(id);
+            if (!current)
+                return null;
+            const updated = mutator(current);
+            await saveUnlocked(id, updated);
+            return updated;
+        });
     }
     async function remove(id) {
-        try {
-            await fs.unlink(join(dir, `${id}.json`));
-            return true;
-        }
-        catch (err) {
-            const fsError = err;
-            if (fsError.code === 'ENOENT')
-                return false;
-            throw err;
-        }
+        const path = jobFilePath(dir, id);
+        if (!path)
+            return false;
+        return mutations.run(id, async () => {
+            try {
+                await fs.unlink(path);
+                return true;
+            }
+            catch (err) {
+                const fsError = err;
+                if (fsError.code === 'ENOENT')
+                    return false;
+                throw err;
+            }
+        });
     }
     async function exists(id) {
-        try {
-            await fs.access(join(dir, `${id}.json`));
-            return true;
-        }
-        catch {
-            return false;
-        }
+        return (await get(id)) !== null;
     }
     return { list, get, save, update, remove, exists };
 }
+function isSafeJobId(id) {
+    return isSafeEvaluationJobId(id);
+}
+function jobFilePath(dir, id) {
+    return isSafeJobId(id) ? join(dir, `${id}.json`) : null;
+}
+function parseJob(raw, expectedId) {
+    const value = JSON.parse(raw);
+    return isValidEvaluationJob(value, expectedId) ? value : null;
+}