cawdex 1.35.63 → 1.35.65

package/dist/benchmark-trace.d.ts

@@ -819,6 +819,12 @@ export interface SourceResearchCoverage {
     resultSources: string[];
     topUrls: string[];
     recentDays: number[];
+    datedHitCount: number;
+    freshHitCount: number;
+    staleHitCount: number;
+    unknownDateHitCount: number;
+    newestDate: string | null;
+    oldestDate: string | null;
     freshTargetedCoverage: boolean;
     kaggleCompetitionsSkipped: boolean;
     coverageNotes: string[];

package/dist/benchmark-trace.js

@@ -2032,9 +2032,17 @@ export function buildBenchmarkTrajectoryQuality(events, usage = emptyBenchmarkUs
     if (sourceResearchUsed && !sourceResearchCoverage.completeTargetedCoverage) {
         warnings.push('source research was partial; targeted benchmark research should cover arXiv, GitHub github_kind:"all", Hugging Face kind:"all", and Kaggle kaggle_kind:"both" when external research is relevant.');
     }
-    if (sourceResearchUsed && sourceResearchCoverage.completeTargetedCoverage && !sourceResearchCoverage.freshTargetedCoverage) {
+    if (sourceResearchUsed && sourceResearchCoverage.completeTargetedCoverage && sourceResearchCoverage.recentDays.length === 0) {
         warnings.push('targeted source research omitted recent_days; newest-science and leaderboard work should bound arXiv, GitHub, and Hugging Face recency before relying on external evidence.');
     }
+    if (sourceResearchUsed &&
+        sourceResearchCoverage.completeTargetedCoverage &&
+        sourceResearchCoverage.recentDays.length > 0 &&
+        sourceResearchCoverage.sourceHitCount > 0 &&
+        sourceResearchHasFreshnessAccounting(sourceResearchCoverage) &&
+        sourceResearchCoverage.freshHitCount === 0) {
+        warnings.push(`targeted source research requested recent_days but produced no dated fresh hits (${formatSourceCoverage(sourceResearchCoverage)}). Inspect stale or undated source evidence before relying on newest-science claims.`);
+    }
     if (sourceResearchUsed && sourceResearchCoverage.sourceHitCount === 0) {
         warnings.push('source research produced no parsed source hits; broaden the query or verify endpoint/auth failures before relying on it.');
     }
@@ -2859,9 +2867,17 @@ function buildBenchmarkProcessDefects(input) {
     if (input.sourceResearchUsed && !input.sourceResearchCoverage.completeTargetedCoverage) {
         add('partial_source_research', 'source_research', 'medium', null, 'Source research was used but did not satisfy targeted arXiv/GitHub/Hugging Face/Kaggle coverage.', formatSourceCoverage(input.sourceResearchCoverage));
     }
-    if (input.sourceResearchUsed && input.sourceResearchCoverage.completeTargetedCoverage && !input.sourceResearchCoverage.freshTargetedCoverage) {
+    if (input.sourceResearchUsed && input.sourceResearchCoverage.completeTargetedCoverage && input.sourceResearchCoverage.recentDays.length === 0) {
         add('source_research_missing_recency', 'source_research', 'medium', null, 'Targeted source research did not include a recency window for newest-science evidence.', formatSourceCoverage(input.sourceResearchCoverage));
     }
+    if (input.sourceResearchUsed &&
+        input.sourceResearchCoverage.completeTargetedCoverage &&
+        input.sourceResearchCoverage.recentDays.length > 0 &&
+        input.sourceResearchCoverage.sourceHitCount > 0 &&
+        sourceResearchHasFreshnessAccounting(input.sourceResearchCoverage) &&
+        input.sourceResearchCoverage.freshHitCount === 0) {
+        add('source_research_no_fresh_hits', 'source_research', 'medium', null, 'Targeted source research requested a recency window but produced no dated fresh hits.', formatSourceCoverage(input.sourceResearchCoverage));
+    }
     if (input.sourceResearchUsed && input.sourceResearchCoverage.sourceHitCount === 0) {
         add('source_research_no_hits', 'source_research', 'medium', null, 'Source research produced no parsed source hits.', formatSourceCoverage(input.sourceResearchCoverage));
     }
@@ -3135,6 +3151,12 @@ export function buildSourceResearchCoverage(events) {
         resultSources: [],
         topUrls: [],
         recentDays: [],
+        datedHitCount: 0,
+        freshHitCount: 0,
+        staleHitCount: 0,
+        unknownDateHitCount: 0,
+        newestDate: null,
+        oldestDate: null,
         freshTargetedCoverage: false,
         kaggleCompetitionsSkipped: false,
         coverageNotes: [],
@@ -3194,9 +3216,32 @@ export function buildSourceResearchCoverage(events) {
             coverage.huggingFaceKinds.includes('all') &&
             coverage.kaggleKinds.includes('both') &&
             !coverage.kaggleCompetitionsSkipped;
-    coverage.freshTargetedCoverage = coverage.completeTargetedCoverage && coverage.recentDays.length > 0;
+    coverage.freshTargetedCoverage =
+        coverage.completeTargetedCoverage &&
+            coverage.recentDays.length > 0 &&
+            sourceResearchHasFreshEvidence(coverage);
     return coverage;
 }
+function sourceResearchFreshnessAccountingCount(coverage) {
+    return coverage.datedHitCount +
+        coverage.freshHitCount +
+        coverage.staleHitCount +
+        coverage.unknownDateHitCount;
+}
+function sourceResearchHasFreshnessAccounting(coverage) {
+    return sourceResearchFreshnessAccountingCount(coverage) > 0;
+}
+function sourceResearchHasFreshEvidence(coverage) {
+    if (coverage.sourceHitCount <= 0)
+        return false;
+    if (!sourceResearchHasFreshnessAccounting(coverage)) {
+        // Older traces predate per-hit date accounting. Preserve compatibility:
+        // a recency-bounded targeted query remains "fresh" unless newer evidence
+        // explicitly proves every hit is stale or undated.
+        return true;
+    }
+    return coverage.freshHitCount > 0;
+}
 export function buildBenchmarkSourceMiningCoverage(events) {
     const coverage = {
         catalogCallCount: 0,
@@ -6622,6 +6667,19 @@ function collectSourceResearchJsonEvidence(coverage, packet) {
         coverage.sourceHitCount += digestHitCount;
     if (digestErrorCount != null)
         coverage.sourceErrorCount += digestErrorCount;
+    const digestDatedCount = numberFromUnknown(digest?.datedHitCount);
+    const digestFreshCount = numberFromUnknown(digest?.freshHitCount);
+    const digestStaleCount = numberFromUnknown(digest?.staleHitCount);
+    const digestUnknownDateCount = numberFromUnknown(digest?.unknownDateHitCount);
+    if (digestDatedCount != null)
+        coverage.datedHitCount += digestDatedCount;
+    if (digestFreshCount != null)
+        coverage.freshHitCount += digestFreshCount;
+    if (digestStaleCount != null)
+        coverage.staleHitCount += digestStaleCount;
+    if (digestUnknownDateCount != null)
+        coverage.unknownDateHitCount += digestUnknownDateCount;
+    mergeSourceResearchDateRange(coverage, stringFromUnknown(digest?.oldestDate), stringFromUnknown(digest?.newestDate));
     const digestSources = objectFromUnknown(digest?.sources);
     if (digestSources) {
         for (const source of Object.keys(digestSources)) {
@@ -6639,14 +6697,54 @@ function collectSourceResearchJsonEvidence(coverage, packet) {
         const record = objectFromUnknown(hit);
         const source = typeof record?.source === 'string' ? record.source : '';
         const url = typeof record?.url === 'string' ? record.url : '';
+        const date = typeof record?.date === 'string' ? record.date : '';
         if (source)
             pushUnique(coverage.resultSources, normalizeResearchSourceLabel(source));
         if (url && coverage.topUrls.length < 12)
             pushUnique(coverage.topUrls, url.replace(/[),.;]+$/, ''));
+        if (digestDatedCount == null && digestFreshCount == null && digestStaleCount == null && digestUnknownDateCount == null) {
+            collectSourceResearchHitFreshness(coverage, date, recentDays);
+        }
     }
     if (digestErrorCount == null)
         coverage.sourceErrorCount += stringArrayFromUnknown(packet.errors).length;
 }
+function collectSourceResearchHitFreshness(coverage, date, recentDays) {
+    const normalized = normalizeTraceIsoDate(date);
+    if (!normalized) {
+        coverage.unknownDateHitCount++;
+        return;
+    }
+    coverage.datedHitCount++;
+    mergeSourceResearchDateRange(coverage, normalized, normalized);
+    if (Number.isFinite(recentDays) && recentDays > 0) {
+        const ms = Date.parse(normalized);
+        if (Number.isFinite(ms) && ms >= Date.now() - Math.floor(recentDays) * 24 * 60 * 60 * 1000) {
+            coverage.freshHitCount++;
+        }
+        else {
+            coverage.staleHitCount++;
+        }
+    }
+}
+function mergeSourceResearchDateRange(coverage, oldest, newest) {
+    const normalizedOldest = normalizeTraceIsoDate(oldest);
+    const normalizedNewest = normalizeTraceIsoDate(newest);
+    if (normalizedOldest && (!coverage.oldestDate || normalizedOldest < coverage.oldestDate)) {
+        coverage.oldestDate = normalizedOldest;
+    }
+    if (normalizedNewest && (!coverage.newestDate || normalizedNewest > coverage.newestDate)) {
+        coverage.newestDate = normalizedNewest;
+    }
+}
+function normalizeTraceIsoDate(value) {
+    if (!value)
+        return undefined;
+    const ms = Date.parse(value);
+    if (!Number.isFinite(ms))
+        return undefined;
+    return new Date(ms).toISOString().slice(0, 10);
+}
 function applySourceCoverageNote(coverage, note) {
     const normalized = note.replace(/\s+/g, ' ').trim();
     if (!normalized)
@@ -6762,6 +6860,22 @@ function collectSourceResearchEvidence(coverage, output) {
         if (authLine) {
             applySourceAuthPreview(coverage, authLine[1]);
         }
+        const datedHits = line.match(/^-\s+dated_hits:\s+(\d+)/i);
+        if (datedHits)
+            coverage.datedHitCount += Number(datedHits[1]);
+        const freshHits = line.match(/^-\s+fresh_hits:\s+(\d+)/i);
+        if (freshHits)
+            coverage.freshHitCount += Number(freshHits[1]);
+        const staleHits = line.match(/^-\s+stale_hits:\s+(\d+)/i);
+        if (staleHits)
+            coverage.staleHitCount += Number(staleHits[1]);
+        const unknownDateHits = line.match(/^-\s+unknown_date_hits:\s+(\d+)/i);
+        if (unknownDateHits)
+            coverage.unknownDateHitCount += Number(unknownDateHits[1]);
+        const dateRange = line.match(/^-\s+date_range:\s+(\d{4}-\d{2}-\d{2})\.\.(\d{4}-\d{2}-\d{2})/i);
+        if (dateRange) {
+            mergeSourceResearchDateRange(coverage, dateRange[1], dateRange[2]);
+        }
         const url = line.match(/^https?:\/\/\S+/i)?.[0];
         if (url && coverage.topUrls.length < 12) {
             pushUnique(coverage.topUrls, url.replace(/[),.;]+$/, ''));
@@ -6814,6 +6928,7 @@ export function buildBenchmarkCompletionReminder(events, usageEvents = [], messa
         || warning.includes('multiple full skill prompts')
         || warning.includes('source research was partial')
         || warning.includes('targeted source research omitted recent_days')
+        || warning.includes('targeted source research requested recent_days but produced no dated fresh hits')
         || warning.includes('source research produced no parsed source hits')
         || warning.includes('source research reported')
         || warning.includes('Kaggle competition research')
@@ -7438,6 +7553,9 @@ function numberFromUnknown(value) {
     const n = Number(value);
     return Number.isFinite(n) ? n : null;
 }
+function stringFromUnknown(value) {
+    return typeof value === 'string' ? value : undefined;
+}
 function booleanFromUnknown(value) {
     return typeof value === 'boolean' ? value : null;
 }
@@ -7489,6 +7607,11 @@ function formatSourceCoverage(coverage) {
         huggingFace,
         kaggle,
         coverage.recentDays.length ? `recent_days:${coverage.recentDays.join('|')}` : 'recent_days:none',
+        `dated_hits:${coverage.datedHitCount}`,
+        `fresh_hits:${coverage.freshHitCount}`,
+        `stale_hits:${coverage.staleHitCount}`,
+        `unknown_date_hits:${coverage.unknownDateHitCount}`,
+        coverage.oldestDate && coverage.newestDate ? `date_range:${coverage.oldestDate}..${coverage.newestDate}` : null,
         coverage.resultSources.length ? `result_sources:${coverage.resultSources.slice(0, 8).join('|')}` : null,
         authSignals ? `auth:${authSignals}` : null,
         coverage.kaggleCompetitionsSkipped ? 'kaggle_competitions:skipped' : null,
@@ -8092,6 +8215,11 @@ function formatResearchSourcesJsonPreviewLines(packet) {
         `- sources: ${sourceSummary}`,
         authPreview ? `- auth: ${authPreview}` : null,
         `- top_urls: ${topUrls.length ? topUrls.join(' | ') : 'none'}`,
+        `- dated_hits: ${numberFromUnknown(digest?.datedHitCount) ?? 0}`,
+        `- fresh_hits: ${numberFromUnknown(digest?.freshHitCount) ?? 0}`,
+        `- stale_hits: ${numberFromUnknown(digest?.staleHitCount) ?? 0}`,
+        `- unknown_date_hits: ${numberFromUnknown(digest?.unknownDateHitCount) ?? 0}`,
+        `- date_range: ${stringFromUnknown(digest?.oldestDate) && stringFromUnknown(digest?.newestDate) ? `${stringFromUnknown(digest?.oldestDate)}..${stringFromUnknown(digest?.newestDate)}` : 'none'}`,
     ].filter((line) => line != null);
     lines.push('', ...digestLines);
     for (const hit of arrayFromUnknown(packet.hits)) {
@@ -8106,6 +8234,9 @@ function formatResearchSourcesJsonPreviewLines(packet) {
         lines.push('', `## ${source}: ${title}`);
         if (url)
             lines.push(url);
+        if (typeof record.date === 'string' && record.date.trim()) {
+            lines.push(`date ${record.date.trim()}`);
+        }
     }
     const errors = stringArrayFromUnknown(packet.errors);
     if (errors.length > 0) {