@remnic/core 1.1.12 → 1.1.13

package/dist/chunk-D46YSIYX.js

@@ -0,0 +1,892 @@
+import {
+  listJsonFiles,
+  listNamedFiles,
+  readJsonFile
+} from "./chunk-LPSF4OQH.js";
+
+// src/evals.ts
+import path from "path";
+import { cp, mkdir, rm, stat, writeFile } from "fs/promises";
+function isRecord(value) {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function assertString(value, field) {
+  if (typeof value !== "string" || value.trim().length === 0) {
+    throw new Error(`${field} must be a non-empty string`);
+  }
+  return value.trim();
+}
+function optionalStringArray(value, field) {
+  if (value === void 0) return void 0;
+  if (!Array.isArray(value)) {
+    throw new Error(`${field} must be an array of strings`);
+  }
+  const out = value.filter((item) => typeof item === "string").map((item) => item.trim()).filter((item) => item.length > 0);
+  if (out.length !== value.length) {
+    throw new Error(`${field} must be an array of non-empty strings`);
+  }
+  return out;
+}
+function resolveEvalStoreDir(memoryDir, overrideDir) {
+  if (typeof overrideDir === "string" && overrideDir.trim().length > 0) {
+    return overrideDir.trim();
+  }
+  return path.join(memoryDir, "state", "evals");
+}
+function assertSafePathSegment(value, field) {
+  if (value === "." || value === ".." || value.includes("/") || value.includes("\\")) {
+    throw new Error(`${field} must be a safe path segment`);
+  }
+  return value;
+}
+function assertSafeBenchmarkId(benchmarkId) {
+  return assertSafePathSegment(benchmarkId, "benchmarkId");
+}
+var ISO_UTC_TIMESTAMP_RE = /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/;
+function assertIsoTimestamp(value, field) {
+  const timestamp = assertString(value, field);
+  if (!ISO_UTC_TIMESTAMP_RE.test(timestamp)) {
+    throw new Error(`${field} must be a valid ISO timestamp`);
+  }
+  const parsed = Date.parse(timestamp);
+  if (!Number.isFinite(parsed)) {
+    throw new Error(`${field} must be a valid ISO timestamp`);
+  }
+  const normalized = new Date(parsed).toISOString();
+  if (normalized !== timestamp) {
+    throw new Error(`${field} must be a valid ISO timestamp`);
+  }
+  return normalized;
+}
+function assertPathWithin(rootDir, targetPath, field) {
+  const resolvedRoot = path.resolve(rootDir);
+  const resolvedTarget = path.resolve(targetPath);
+  const relative = path.relative(resolvedRoot, resolvedTarget);
+  if (relative === "" || !relative.startsWith("..") && !path.isAbsolute(relative)) {
+    return;
+  }
+  throw new Error(`${field} must stay within ${rootDir}`);
+}
+function validateEvalBenchmarkManifest(raw, options) {
+  if (!isRecord(raw)) throw new Error("benchmark manifest must be an object");
+  if (raw.schemaVersion !== 1) throw new Error("schemaVersion must be 1");
+  if (!Array.isArray(raw.cases)) throw new Error("cases must be an array");
+  const benchmarkTypeRaw = typeof raw.benchmarkType === "string" && raw.benchmarkType.trim().length > 0 ? raw.benchmarkType.trim() : "standard";
+  if (!["standard", "memory-red-team"].includes(benchmarkTypeRaw)) {
+    throw new Error("benchmarkType must be one of standard|memory-red-team");
+  }
+  const cases = raw.cases.map((item, index) => {
+    if (!isRecord(item)) throw new Error(`cases[${index}] must be an object`);
+    return {
+      id: assertString(item.id, `cases[${index}].id`),
+      prompt: assertString(item.prompt, `cases[${index}].prompt`),
+      expectedSignals: optionalStringArray(item.expectedSignals, `cases[${index}].expectedSignals`),
+      notes: typeof item.notes === "string" && item.notes.trim().length > 0 ? item.notes.trim() : void 0
+    };
+  });
+  const benchmarkType = benchmarkTypeRaw;
+  if (benchmarkType === "memory-red-team" && options?.memoryRedTeamBenchEnabled !== true) {
+    throw new Error("memory-red-team benchmark packs require memoryRedTeamBenchEnabled");
+  }
+  const attackClass = typeof raw.attackClass === "string" && raw.attackClass.trim().length > 0 ? raw.attackClass.trim() : void 0;
+  const targetSurface = typeof raw.targetSurface === "string" && raw.targetSurface.trim().length > 0 ? raw.targetSurface.trim() : void 0;
+  if (benchmarkType === "memory-red-team" && attackClass === void 0) {
+    throw new Error("attackClass must be a non-empty string");
+  }
+  if (benchmarkType === "memory-red-team" && targetSurface === void 0) {
+    throw new Error("targetSurface must be a non-empty string");
+  }
+  return {
+    schemaVersion: 1,
+    benchmarkId: assertString(raw.benchmarkId, "benchmarkId"),
+    benchmarkType,
+    title: assertString(raw.title, "title"),
+    description: typeof raw.description === "string" && raw.description.trim().length > 0 ? raw.description.trim() : void 0,
+    tags: optionalStringArray(raw.tags, "tags"),
+    sourceLinks: optionalStringArray(raw.sourceLinks, "sourceLinks"),
+    attackClass,
+    targetSurface,
+    cases
+  };
+}
+function validateEvalRunSummary(raw) {
+  if (!isRecord(raw)) throw new Error("eval run summary must be an object");
+  if (raw.schemaVersion !== 1) throw new Error("schemaVersion must be 1");
+  const status = assertString(raw.status, "status");
+  if (!["running", "completed", "failed", "partial"].includes(status)) {
+    throw new Error("status must be one of running|completed|failed|partial");
+  }
+  const totalCases = Number(raw.totalCases);
+  const passedCases = Number(raw.passedCases);
+  const failedCases = Number(raw.failedCases);
+  if (!Number.isFinite(totalCases) || totalCases < 0) throw new Error("totalCases must be a non-negative number");
+  if (!Number.isFinite(passedCases) || passedCases < 0) throw new Error("passedCases must be a non-negative number");
+  if (!Number.isFinite(failedCases) || failedCases < 0) throw new Error("failedCases must be a non-negative number");
+  const metrics = parseOptionalEvalRunMetrics(raw.metrics);
+  return {
+    schemaVersion: 1,
+    runId: assertString(raw.runId, "runId"),
+    benchmarkId: assertString(raw.benchmarkId, "benchmarkId"),
+    status,
+    startedAt: assertString(raw.startedAt, "startedAt"),
+    completedAt: typeof raw.completedAt === "string" && raw.completedAt.trim().length > 0 ? raw.completedAt.trim() : void 0,
+    totalCases,
+    passedCases,
+    failedCases,
+    metrics,
+    notes: typeof raw.notes === "string" && raw.notes.trim().length > 0 ? raw.notes.trim() : void 0,
+    gitRef: typeof raw.gitRef === "string" && raw.gitRef.trim().length > 0 ? raw.gitRef.trim() : void 0
+  };
+}
+function validateEvalBaselineSnapshot(raw) {
+  if (!isRecord(raw)) throw new Error("eval baseline snapshot must be an object");
+  if (raw.schemaVersion !== 1) throw new Error("schemaVersion must be 1");
+  if (!Array.isArray(raw.benchmarks)) throw new Error("benchmarks must be an array");
+  const benchmarks = raw.benchmarks.map((item, index) => {
+    if (!isRecord(item)) throw new Error(`benchmarks[${index}] must be an object`);
+    const passRate = Number(item.passRate);
+    if (!Number.isFinite(passRate) || passRate < 0 || passRate > 1) {
+      throw new Error(`benchmarks[${index}].passRate must be a number between 0 and 1`);
+    }
+    const metrics = parseOptionalEvalRunMetrics(item.metrics);
+    return {
+      benchmarkId: assertString(item.benchmarkId, `benchmarks[${index}].benchmarkId`),
+      runId: assertString(item.runId, `benchmarks[${index}].runId`),
+      completedAt: typeof item.completedAt === "string" && item.completedAt.trim().length > 0 ? item.completedAt.trim() : void 0,
+      gitRef: typeof item.gitRef === "string" && item.gitRef.trim().length > 0 ? item.gitRef.trim() : void 0,
+      passRate,
+      metrics
+    };
+  });
+  const benchmarkCount = Number(raw.benchmarkCount);
+  if (!Number.isFinite(benchmarkCount) || benchmarkCount < 0) {
+    throw new Error("benchmarkCount must be a non-negative number");
+  }
+  if (benchmarkCount !== benchmarks.length) {
+    throw new Error("benchmarkCount must match benchmarks.length");
+  }
+  return {
+    schemaVersion: 1,
+    snapshotId: assertString(raw.snapshotId, "snapshotId"),
+    createdAt: assertString(raw.createdAt, "createdAt"),
+    sourceRootDir: assertString(raw.sourceRootDir, "sourceRootDir"),
+    benchmarkCount,
+    benchmarks,
+    notes: typeof raw.notes === "string" && raw.notes.trim().length > 0 ? raw.notes.trim() : void 0,
+    gitRef: typeof raw.gitRef === "string" && raw.gitRef.trim().length > 0 ? raw.gitRef.trim() : void 0
+  };
+}
+function parseOptionalEvalRunMetrics(raw) {
+  if (!isRecord(raw)) return void 0;
+  return {
+    recallPrecisionAtK: typeof raw.recallPrecisionAtK === "number" ? raw.recallPrecisionAtK : void 0,
+    actionOutcomeScore: typeof raw.actionOutcomeScore === "number" ? raw.actionOutcomeScore : void 0,
+    objectiveStateCoverage: typeof raw.objectiveStateCoverage === "number" ? raw.objectiveStateCoverage : void 0,
+    causalPathRecall: typeof raw.causalPathRecall === "number" ? raw.causalPathRecall : void 0,
+    trustViolationRate: typeof raw.trustViolationRate === "number" ? raw.trustViolationRate : void 0,
+    creationRecoveryScore: typeof raw.creationRecoveryScore === "number" ? raw.creationRecoveryScore : void 0
+  };
+}
+function validateEvalShadowRecallRecord(raw) {
+  if (!isRecord(raw)) throw new Error("eval shadow recall record must be an object");
+  if (raw.schemaVersion !== 1) throw new Error("schemaVersion must be 1");
+  const recallMode = assertString(raw.recallMode, "recallMode");
+  if (!["no_recall", "minimal", "full", "graph_mode"].includes(recallMode)) {
+    throw new Error("recallMode must be one of no_recall|minimal|full|graph_mode");
+  }
+  const source = assertString(raw.source, "source");
+  if (!["none", "hot_qmd", "hot_embedding", "cold_fallback", "recent_scan"].includes(source)) {
+    throw new Error("source must be one of none|hot_qmd|hot_embedding|cold_fallback|recent_scan");
+  }
+  const promptLength = Number(raw.promptLength);
+  const retrievalQueryLength = Number(raw.retrievalQueryLength);
+  const recallResultLimit = Number(raw.recallResultLimit);
+  const recalledMemoryCount = Number(raw.recalledMemoryCount);
+  const contextChars = Number(raw.contextChars);
+  const durationMs = Number(raw.durationMs);
+  for (const [field, value] of [
+    ["promptLength", promptLength],
+    ["retrievalQueryLength", retrievalQueryLength],
+    ["recallResultLimit", recallResultLimit],
+    ["recalledMemoryCount", recalledMemoryCount],
+    ["contextChars", contextChars],
+    ["durationMs", durationMs]
+  ]) {
+    if (!Number.isFinite(value) || value < 0) {
+      throw new Error(`${field} must be a non-negative number`);
+    }
+  }
+  const memoryIds = optionalStringArray(raw.memoryIds, "memoryIds") ?? [];
+  if (typeof raw.injected !== "boolean") throw new Error("injected must be a boolean");
+  let timings;
+  if (raw.timings !== void 0) {
+    if (!isRecord(raw.timings)) throw new Error("timings must be an object of strings");
+    const out = {};
+    for (const [key, value] of Object.entries(raw.timings)) {
+      if (typeof value !== "string") throw new Error("timings must be an object of strings");
+      out[key] = value;
+    }
+    timings = out;
+  }
+  const identityInjectionModeRaw = typeof raw.identityInjectionMode === "string" && raw.identityInjectionMode.trim().length > 0 ? raw.identityInjectionMode.trim() : void 0;
+  if (identityInjectionModeRaw !== void 0 && !["recovery_only", "minimal", "full", "none"].includes(identityInjectionModeRaw)) {
+    throw new Error("identityInjectionMode must be one of recovery_only|minimal|full|none");
+  }
+  return {
+    schemaVersion: 1,
+    traceId: assertSafePathSegment(assertString(raw.traceId, "traceId"), "traceId"),
+    recordedAt: assertIsoTimestamp(raw.recordedAt, "recordedAt"),
+    sessionKey: assertString(raw.sessionKey, "sessionKey"),
+    promptHash: assertString(raw.promptHash, "promptHash"),
+    promptLength,
+    retrievalQueryHash: assertString(raw.retrievalQueryHash, "retrievalQueryHash"),
+    retrievalQueryLength,
+    recallMode,
+    recallResultLimit,
+    source,
+    recalledMemoryCount,
+    injected: raw.injected,
+    contextChars,
+    memoryIds,
+    policyVersion: typeof raw.policyVersion === "string" && raw.policyVersion.trim().length > 0 ? raw.policyVersion.trim() : void 0,
+    identityInjectionMode: identityInjectionModeRaw,
+    identityInjectedChars: typeof raw.identityInjectedChars === "number" && Number.isFinite(raw.identityInjectedChars) ? raw.identityInjectedChars : void 0,
+    identityInjectionTruncated: typeof raw.identityInjectionTruncated === "boolean" ? raw.identityInjectionTruncated : void 0,
+    durationMs,
+    timings
+  };
+}
+var LOWER_IS_BETTER_METRICS = /* @__PURE__ */ new Set(["trustViolationRate"]);
+function computePassRate(run) {
+  return run.totalCases > 0 ? run.passedCases / run.totalCases : 0;
+}
+function latestCompletedRunsByBenchmark(runs) {
+  const sorted = [...runs].filter((run) => run.status === "completed").sort((a, b) => {
+    const aTime = Date.parse(a.completedAt ?? a.startedAt);
+    const bTime = Date.parse(b.completedAt ?? b.startedAt);
+    return (Number.isNaN(bTime) ? 0 : bTime) - (Number.isNaN(aTime) ? 0 : aTime);
+  });
+  const out = /* @__PURE__ */ new Map();
+  for (const run of sorted) {
+    if (!out.has(run.benchmarkId)) {
+      out.set(run.benchmarkId, run);
+    }
+  }
+  return out;
+}
+function compareMetricDeltas(baseMetrics, candidateMetrics) {
+  const deltas = {};
+  const regressions = [];
+  const improvements = [];
+  if (!baseMetrics || !candidateMetrics) {
+    return { deltas, regressions, improvements };
+  }
+  for (const metric of Object.keys(baseMetrics)) {
+    const baseValue = baseMetrics[metric];
+    const candidateValue = candidateMetrics[metric];
+    if (typeof baseValue !== "number" || typeof candidateValue !== "number") continue;
+    const delta = candidateValue - baseValue;
+    deltas[metric] = delta;
+    if (delta === 0) continue;
+    const lowerIsBetter = LOWER_IS_BETTER_METRICS.has(metric);
+    const improved = lowerIsBetter ? delta < 0 : delta > 0;
+    const summary = `${metric} ${baseValue} -> ${candidateValue}`;
+    if (improved) {
+      improvements.push(summary);
+    } else {
+      regressions.push(summary);
+    }
+  }
+  return { deltas, regressions, improvements };
+}
+function formatEvalBaselineDeltaMarkdown(report) {
+  const lines = [
+    "# Eval Baseline Delta Report",
+    "",
+    `- Passed: ${report.passed ? "yes" : "no"}`,
+    `- Baseline snapshot: ${report.baselineSnapshotId}`,
+    `- Baseline created: ${report.baselineCreatedAt}`,
+    `- Baseline source root: ${report.baselineSourceRootDir}`,
+    `- Candidate root: ${report.candidateRootDir}`,
+    `- Benchmarks compared: ${report.comparedBenchmarks}`
+  ];
+  if (report.missingCandidateBenchmarks.length > 0) {
+    lines.push(`- Missing candidate benchmarks: ${report.missingCandidateBenchmarks.join(", ")}`);
+  }
+  lines.push(
+    `- Invalid candidate artifacts: benchmarks=${report.invalidArtifacts.candidate.benchmarks}, runs=${report.invalidArtifacts.candidate.runs}, shadows=${report.invalidArtifacts.candidate.shadows}, baselines=${report.invalidArtifacts.candidate.baselines}`,
+    "",
+    "## Regressions"
+  );
+  if (report.regressions.length === 0) {
+    lines.push("- none");
+  } else {
+    for (const regression of report.regressions) lines.push(`- ${regression}`);
+  }
+  lines.push("", "## Improvements");
+  if (report.improvements.length === 0) {
+    lines.push("- none");
+  } else {
+    for (const improvement of report.improvements) lines.push(`- ${improvement}`);
+  }
+  lines.push("", "## Benchmark Deltas");
+  if (report.deltas.length === 0) {
+    lines.push("- none");
+  } else {
+    for (const delta of report.deltas) {
+      lines.push(
+        `- ${delta.benchmarkId}: passRate ${delta.basePassRate} -> ${delta.candidatePassRate} (delta ${delta.passRateDelta})`
+      );
+    }
+  }
+  return lines.join("\n");
+}
+async function collectEvalStoreSnapshot(options) {
+  const rootDir = options.rootDir;
+  const benchmarkDir = path.join(rootDir, "benchmarks");
+  const runsDir = path.join(rootDir, "runs");
+  const shadowDir = path.join(rootDir, "shadow");
+  const baselineDir = path.join(rootDir, "baselines");
+  const benchmarkFiles = await listNamedFiles(benchmarkDir, "manifest.json");
+  const runFiles = await listJsonFiles(runsDir);
+  const shadowFiles = await listJsonFiles(shadowDir);
+  const baselineFiles = await listJsonFiles(baselineDir);
+  const invalidBenchmarks = [];
+  const invalidRuns = [];
+  const invalidShadows = [];
+  const invalidBaselines = [];
+  const manifests = [];
+  for (const filePath of benchmarkFiles) {
+    try {
+      manifests.push(
+        validateEvalBenchmarkManifest(await readJsonFile(filePath), {
+          memoryRedTeamBenchEnabled: options.memoryRedTeamBenchEnabled
+        })
+      );
+    } catch (error) {
+      invalidBenchmarks.push({
+        path: filePath,
+        error: error instanceof Error ? error.message : String(error)
+      });
+    }
+  }
+  const runs = [];
+  for (const filePath of runFiles) {
+    try {
+      runs.push(validateEvalRunSummary(await readJsonFile(filePath)));
+    } catch (error) {
+      invalidRuns.push({
+        path: filePath,
+        error: error instanceof Error ? error.message : String(error)
+      });
+    }
+  }
+  const shadows = [];
+  for (const filePath of shadowFiles) {
+    try {
+      shadows.push(validateEvalShadowRecallRecord(await readJsonFile(filePath)));
+    } catch (error) {
+      invalidShadows.push({
+        path: filePath,
+        error: error instanceof Error ? error.message : String(error)
+      });
+    }
+  }
+  const baselines = [];
+  for (const filePath of baselineFiles) {
+    try {
+      baselines.push(validateEvalBaselineSnapshot(await readJsonFile(filePath)));
+    } catch (error) {
+      invalidBaselines.push({
+        path: filePath,
+        error: error instanceof Error ? error.message : String(error)
+      });
+    }
+  }
+  runs.sort((a, b) => {
+    const aTime = Date.parse(a.completedAt ?? a.startedAt);
+    const bTime = Date.parse(b.completedAt ?? b.startedAt);
+    return (Number.isNaN(bTime) ? 0 : bTime) - (Number.isNaN(aTime) ? 0 : aTime);
+  });
+  shadows.sort((a, b) => b.recordedAt.localeCompare(a.recordedAt));
+  baselines.sort((a, b) => b.createdAt.localeCompare(a.createdAt));
+  const tags = /* @__PURE__ */ new Set();
+  const attackClasses = /* @__PURE__ */ new Set();
+  const sourceLinks = /* @__PURE__ */ new Set();
+  const targetSurfaces = /* @__PURE__ */ new Set();
+  let totalCases = 0;
+  let redTeam = 0;
+  for (const manifest of manifests) {
+    totalCases += manifest.cases.length;
+    if (manifest.benchmarkType === "memory-red-team") {
+      redTeam += 1;
+      if (manifest.attackClass) attackClasses.add(manifest.attackClass);
+      if (manifest.targetSurface) targetSurfaces.add(manifest.targetSurface);
+    }
+    for (const tag of manifest.tags ?? []) tags.add(tag);
+    for (const link of manifest.sourceLinks ?? []) sourceLinks.add(link);
+  }
+  return {
+    status: {
+      enabled: options.enabled,
+      shadowModeEnabled: options.shadowModeEnabled,
+      rootDir,
+      benchmarkDir,
+      runsDir,
+      benchmarks: {
+        total: benchmarkFiles.length,
+        valid: manifests.length,
+        invalid: invalidBenchmarks.length,
+        redTeam,
+        totalCases,
+        attackClasses: [...attackClasses].sort(),
+        tags: [...tags].sort(),
+        targetSurfaces: [...targetSurfaces].sort(),
+        sourceLinks: [...sourceLinks].sort()
+      },
+      runs: {
+        total: runFiles.length,
+        invalid: invalidRuns.length,
+        completed: runs.filter((run) => run.status === "completed").length,
+        failed: runs.filter((run) => run.status === "failed").length,
+        partial: runs.filter((run) => run.status === "partial").length,
+        running: runs.filter((run) => run.status === "running").length,
+        latestRunId: runs[0]?.runId,
+        latestBenchmarkId: runs[0]?.benchmarkId,
+        latestCompletedAt: runs[0]?.completedAt
+      },
+      shadows: {
+        total: shadowFiles.length,
+        invalid: invalidShadows.length,
+        latestTraceId: shadows[0]?.traceId,
+        latestRecordedAt: shadows[0]?.recordedAt,
+        latestSessionKey: shadows[0]?.sessionKey
+      },
+      baselines: {
+        enabled: options.baselineSnapshotsEnabled === true,
+        total: baselineFiles.length,
+        invalid: invalidBaselines.length,
+        latestSnapshotId: baselines[0]?.snapshotId,
+        latestCreatedAt: baselines[0]?.createdAt,
+        latestBenchmarkCount: baselines[0]?.benchmarkCount
+      },
+      latestRun: runs[0],
+      latestShadow: shadows[0],
+      latestBaseline: baselines[0],
+      invalidBenchmarks,
+      invalidRuns,
+      invalidShadows,
+      invalidBaselines
+    },
+    manifests,
+    runs,
+    shadows,
+    baselines
+  };
+}
+async function resolveBenchmarkManifestPath(sourcePath) {
+  const info = await stat(sourcePath);
+  if (info.isDirectory()) {
+    return {
+      sourceKind: "directory",
+      manifestPath: path.join(sourcePath, "manifest.json")
+    };
+  }
+  if (info.isFile()) {
+    return {
+      sourceKind: "file",
+      manifestPath: sourcePath
+    };
+  }
+  throw new Error("benchmark pack source must be a file or directory");
+}
+async function validateEvalBenchmarkPack(sourcePath, options) {
+  const trimmedSourcePath = sourcePath.trim();
+  if (trimmedSourcePath.length === 0) {
+    throw new Error("benchmark pack path must be a non-empty string");
+  }
+  const { manifestPath } = await resolveBenchmarkManifestPath(trimmedSourcePath);
+  const manifest = validateEvalBenchmarkManifest(await readJsonFile(manifestPath), {
+    memoryRedTeamBenchEnabled: options?.memoryRedTeamBenchEnabled
+  });
+  return {
+    sourcePath: trimmedSourcePath,
+    manifestPath,
+    benchmarkId: assertSafeBenchmarkId(manifest.benchmarkId),
+    benchmarkType: manifest.benchmarkType ?? "standard",
+    title: manifest.title,
+    attackClass: manifest.attackClass,
+    targetSurface: manifest.targetSurface,
+    totalCases: manifest.cases.length,
+    tags: [...manifest.tags ?? []],
+    sourceLinks: [...manifest.sourceLinks ?? []]
+  };
+}
+async function importEvalBenchmarkPack(options) {
+  const summary = await validateEvalBenchmarkPack(options.sourcePath, {
+    memoryRedTeamBenchEnabled: options.memoryRedTeamBenchEnabled
+  });
+  const rootDir = resolveEvalStoreDir(options.memoryDir, options.evalStoreDir);
+  const benchmarkDir = path.join(rootDir, "benchmarks");
+  const targetDir = path.join(benchmarkDir, summary.benchmarkId);
+  const { sourceKind, manifestPath } = await resolveBenchmarkManifestPath(summary.sourcePath);
+  let overwritten = false;
+  try {
+    await stat(targetDir);
+    if (options.force !== true) {
+      throw new Error(`benchmark pack already exists at ${targetDir}; rerun with force to replace it`);
+    }
+    overwritten = true;
+    await rm(targetDir, { recursive: true, force: true });
+  } catch (error) {
+    if (!(error instanceof Error) || !("code" in error) || error.code !== "ENOENT") {
+      throw error;
+    }
+  }
+  await mkdir(benchmarkDir, { recursive: true });
+  if (sourceKind === "directory") {
+    await cp(summary.sourcePath, targetDir, { recursive: true });
+  } else {
+    await mkdir(targetDir, { recursive: true });
+    await cp(manifestPath, path.join(targetDir, "manifest.json"));
+  }
+  return {
+    ...summary,
+    targetDir,
+    overwritten
+  };
+}
+async function recordEvalShadowRecall(options) {
+  const rootDir = resolveEvalStoreDir(options.memoryDir, options.evalStoreDir);
+  const validated = validateEvalShadowRecallRecord(options.record);
+  const day = validated.recordedAt.slice(0, 10);
+  const shadowRoot = path.join(rootDir, "shadow");
+  const shadowDir = path.join(shadowRoot, day);
+  const targetPath = path.join(shadowDir, `${validated.traceId}.json`);
+  assertPathWithin(shadowRoot, targetPath, "shadow recall record path");
+  await mkdir(shadowDir, { recursive: true });
+  await writeFile(targetPath, JSON.stringify(validated, null, 2), "utf-8");
+  return targetPath;
+}
+async function getEvalHarnessStatus(options) {
+  return (await collectEvalStoreSnapshot({
+    rootDir: resolveEvalStoreDir(options.memoryDir, options.evalStoreDir),
+    enabled: options.enabled,
+    shadowModeEnabled: options.shadowModeEnabled,
+    baselineSnapshotsEnabled: options.baselineSnapshotsEnabled,
+    memoryRedTeamBenchEnabled: options.memoryRedTeamBenchEnabled
+  })).status;
+}
+async function createEvalBaselineSnapshot(options) {
+  if (options.baselineSnapshotsEnabled !== true) {
+    throw new Error("benchmark baseline snapshots are disabled");
+  }
+  const snapshotId = assertSafePathSegment(assertString(options.snapshotId, "snapshotId"), "snapshotId");
+  const rootDir = resolveEvalStoreDir(options.memoryDir, options.evalStoreDir);
+  const store = await collectEvalStoreSnapshot({
+    rootDir,
+    enabled: true,
+    shadowModeEnabled: true,
+    baselineSnapshotsEnabled: true,
+    memoryRedTeamBenchEnabled: true
+  });
+  const latestRuns = latestCompletedRunsByBenchmark(store.runs);
+  const benchmarks = [...latestRuns.values()].sort((a, b) => a.benchmarkId.localeCompare(b.benchmarkId)).map((run) => ({
+    benchmarkId: run.benchmarkId,
+    runId: run.runId,
+    completedAt: run.completedAt,
+    gitRef: run.gitRef,
+    passRate: computePassRate(run),
+    metrics: run.metrics
+  }));
+  const snapshot = validateEvalBaselineSnapshot({
+    schemaVersion: 1,
+    snapshotId,
+    createdAt: options.createdAt ?? (/* @__PURE__ */ new Date()).toISOString(),
+    sourceRootDir: rootDir,
+    benchmarkCount: benchmarks.length,
+    benchmarks,
+    notes: options.notes,
+    gitRef: options.gitRef
+  });
+  const targetPath = path.join(rootDir, "baselines", `${snapshot.snapshotId}.json`);
+  await mkdir(path.dirname(targetPath), { recursive: true });
+  await writeFile(targetPath, JSON.stringify(snapshot, null, 2), "utf-8");
+  return { targetPath, snapshot };
+}
+async function runEvalBaselineDeltaReport(options) {
+  if (options.benchmarkDeltaReporterEnabled !== true) {
+    throw new Error("benchmark delta reporter is disabled");
+  }
+  const snapshotId = assertSafePathSegment(assertString(options.snapshotId, "snapshotId"), "snapshotId");
+  const candidateRootDir = resolveEvalStoreDir(options.memoryDir, options.evalStoreDir);
+  const candidateSnapshot = await collectEvalStoreSnapshot({
+    rootDir: candidateRootDir,
+    enabled: true,
+    shadowModeEnabled: true,
+    baselineSnapshotsEnabled: true,
+    memoryRedTeamBenchEnabled: true
+  });
+  const baselineSnapshot = candidateSnapshot.baselines.find((snapshot) => snapshot.snapshotId === snapshotId);
+  if (!baselineSnapshot) {
+    throw new Error(`benchmark baseline snapshot not found: ${snapshotId}`);
+  }
+  return buildEvalBaselineDeltaReport({
+    baselineSnapshot,
+    candidateSnapshot
+  });
+}
+async function runEvalStoredBaselineCiGate(options) {
+  const snapshotId = assertSafePathSegment(assertString(options.snapshotId, "snapshotId"), "snapshotId");
+  const baseRootDir = resolveEvalStoreDir(options.baseMemoryDir ?? options.baseEvalStoreDir, options.baseEvalStoreDir);
+  const candidateRootDir = resolveEvalStoreDir(
+    options.candidateMemoryDir ?? options.candidateEvalStoreDir,
+    options.candidateEvalStoreDir
+  );
+  const [baseSnapshot, candidateSnapshot] = await Promise.all([
+    collectEvalStoreSnapshot({
+      rootDir: baseRootDir,
+      enabled: true,
+      shadowModeEnabled: true,
+      baselineSnapshotsEnabled: true,
+      memoryRedTeamBenchEnabled: true
+    }),
+    collectEvalStoreSnapshot({
+      rootDir: candidateRootDir,
+      enabled: true,
+      shadowModeEnabled: true,
+      baselineSnapshotsEnabled: true,
+      memoryRedTeamBenchEnabled: true
+    })
+  ]);
+  const baselineSnapshot = baseSnapshot.baselines.find((snapshot) => snapshot.snapshotId === snapshotId) ?? candidateSnapshot.baselines.find((snapshot) => snapshot.snapshotId === snapshotId);
+  if (!baselineSnapshot) {
+    throw new Error(`benchmark baseline snapshot not found: ${snapshotId}`);
+  }
+  return {
+    baseRootDir,
+    baselineResolvedFrom: baseSnapshot.baselines.some((snapshot) => snapshot.snapshotId === snapshotId) ? "base" : "candidate",
+    ...buildEvalBaselineDeltaReport({
+      baselineSnapshot,
+      candidateSnapshot
+    })
+  };
+}
+function buildEvalBaselineDeltaReport(options) {
+  const { baselineSnapshot, candidateSnapshot } = options;
+  const regressions = [];
+  const improvements = [];
+  if (candidateSnapshot.status.invalidBenchmarks.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidBenchmarks.length} invalid benchmark manifest(s)`);
+  }
+  if (candidateSnapshot.status.invalidRuns.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidRuns.length} invalid run summary file(s)`);
+  }
+  if (candidateSnapshot.status.invalidShadows.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidShadows.length} invalid shadow record(s)`);
+  }
+  if (candidateSnapshot.status.invalidBaselines.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidBaselines.length} invalid baseline snapshot file(s)`);
+  }
+  const candidateRuns = latestCompletedRunsByBenchmark(candidateSnapshot.runs);
+  const baselineBenchmarks = new Map(
+    baselineSnapshot.benchmarks.map((benchmark) => [benchmark.benchmarkId, benchmark])
+  );
+  const missingCandidateBenchmarks = [...baselineBenchmarks.keys()].filter((benchmarkId) => !candidateRuns.has(benchmarkId)).sort();
+  for (const benchmarkId of missingCandidateBenchmarks) {
+    regressions.push(`candidate is missing latest completed benchmark run for ${benchmarkId}`);
+  }
+  const deltas = [];
+  for (const benchmarkId of [...baselineBenchmarks.keys()].sort()) {
+    const baseBenchmark = baselineBenchmarks.get(benchmarkId);
+    const candidateRun = candidateRuns.get(benchmarkId);
+    if (!baseBenchmark || !candidateRun) continue;
+    const passRateDelta = computePassRate(candidateRun) - baseBenchmark.passRate;
+    const delta = {
+      benchmarkId,
+      baseRunId: baseBenchmark.runId,
+      candidateRunId: candidateRun.runId,
+      basePassRate: baseBenchmark.passRate,
+      candidatePassRate: computePassRate(candidateRun),
+      passRateDelta,
+      metricDeltas: {},
+      regressions: [],
+      improvements: []
+    };
+    if (passRateDelta < 0) {
+      delta.regressions.push(`passRate ${baseBenchmark.passRate} -> ${delta.candidatePassRate}`);
+      regressions.push(`${benchmarkId} pass rate regressed (${baseBenchmark.passRate} -> ${delta.candidatePassRate})`);
+    } else if (passRateDelta > 0) {
+      delta.improvements.push(`passRate ${baseBenchmark.passRate} -> ${delta.candidatePassRate}`);
+      improvements.push(`${benchmarkId} pass rate improved (${baseBenchmark.passRate} -> ${delta.candidatePassRate})`);
+    }
+    const metricDelta = compareMetricDeltas(baseBenchmark.metrics, candidateRun.metrics);
+    delta.metricDeltas = metricDelta.deltas;
+    for (const regression of metricDelta.regressions) {
+      delta.regressions.push(regression);
+      regressions.push(`${benchmarkId} ${regression}`);
+    }
+    for (const improvement of metricDelta.improvements) {
+      delta.improvements.push(improvement);
+      improvements.push(`${benchmarkId} ${improvement}`);
+    }
+    deltas.push(delta);
+  }
+  const report = {
+    passed: regressions.length === 0,
+    baselineSnapshotId: baselineSnapshot.snapshotId,
+    baselineCreatedAt: baselineSnapshot.createdAt,
+    baselineSourceRootDir: baselineSnapshot.sourceRootDir,
+    candidateRootDir: candidateSnapshot.status.rootDir,
+    comparedBenchmarks: deltas.length,
+    missingCandidateBenchmarks,
+    invalidArtifacts: {
+      candidate: {
+        benchmarks: candidateSnapshot.status.invalidBenchmarks.length,
+        runs: candidateSnapshot.status.invalidRuns.length,
+        shadows: candidateSnapshot.status.invalidShadows.length,
+        baselines: candidateSnapshot.status.invalidBaselines.length
+      }
+    },
+    regressions,
+    improvements,
+    deltas,
+    markdownReport: ""
+  };
+  report.markdownReport = formatEvalBaselineDeltaMarkdown(report);
+  return report;
+}
+function resolveRequiredEvalStoreRoot(options, label) {
+  if (typeof options.evalStoreDir === "string" && options.evalStoreDir.trim().length > 0) {
+    return options.evalStoreDir.trim();
+  }
+  if (typeof options.memoryDir === "string" && options.memoryDir.trim().length > 0) {
+    return resolveEvalStoreDir(options.memoryDir.trim());
+  }
+  throw new Error(`${label} requires memoryDir or evalStoreDir`);
+}
+async function runEvalBenchmarkCiGate(options) {
+  const baseRootDir = resolveRequiredEvalStoreRoot(
+    { memoryDir: options.baseMemoryDir, evalStoreDir: options.baseEvalStoreDir },
+    "base"
+  );
+  const candidateRootDir = resolveRequiredEvalStoreRoot(
+    { memoryDir: options.candidateMemoryDir, evalStoreDir: options.candidateEvalStoreDir },
+    "candidate"
+  );
+  const baseSnapshot = await collectEvalStoreSnapshot({
+    rootDir: baseRootDir,
+    enabled: true,
+    shadowModeEnabled: true,
+    memoryRedTeamBenchEnabled: true
+  });
+  const candidateSnapshot = await collectEvalStoreSnapshot({
+    rootDir: candidateRootDir,
+    enabled: true,
+    shadowModeEnabled: true,
+    memoryRedTeamBenchEnabled: true
+  });
+  const regressions = [];
+  const improvements = [];
+  if (baseSnapshot.status.invalidBenchmarks.length > 0) {
+    regressions.push(`base store has ${baseSnapshot.status.invalidBenchmarks.length} invalid benchmark manifest(s)`);
+  }
+  if (baseSnapshot.status.invalidRuns.length > 0) {
+    regressions.push(`base store has ${baseSnapshot.status.invalidRuns.length} invalid run summary file(s)`);
+  }
+  if (baseSnapshot.status.invalidShadows.length > 0) {
+    regressions.push(`base store has ${baseSnapshot.status.invalidShadows.length} invalid shadow record(s)`);
+  }
+  if (candidateSnapshot.status.invalidBenchmarks.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidBenchmarks.length} invalid benchmark manifest(s)`);
+  }
+  if (candidateSnapshot.status.invalidRuns.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidRuns.length} invalid run summary file(s)`);
+  }
+  if (candidateSnapshot.status.invalidShadows.length > 0) {
+    regressions.push(`candidate store has ${candidateSnapshot.status.invalidShadows.length} invalid shadow record(s)`);
+  }
+  const baseRuns = latestCompletedRunsByBenchmark(baseSnapshot.runs);
+  const candidateRuns = latestCompletedRunsByBenchmark(candidateSnapshot.runs);
+  const missingCandidateBenchmarks = [...baseRuns.keys()].filter((benchmarkId) => !candidateRuns.has(benchmarkId)).sort();
+  for (const benchmarkId of missingCandidateBenchmarks) {
+    regressions.push(`candidate is missing latest completed benchmark run for ${benchmarkId}`);
+  }
+  const deltas = [];
+  for (const benchmarkId of [...baseRuns.keys()].sort()) {
+    const baseRun = baseRuns.get(benchmarkId);
+    const candidateRun = candidateRuns.get(benchmarkId);
+    if (!baseRun || !candidateRun) continue;
+    const basePassRate = computePassRate(baseRun);
+    const candidatePassRate = computePassRate(candidateRun);
+    const passRateDelta = candidatePassRate - basePassRate;
+    const delta = {
+      benchmarkId,
+      baseRunId: baseRun.runId,
+      candidateRunId: candidateRun.runId,
+      basePassRate,
+      candidatePassRate,
+      passRateDelta,
+      metricDeltas: {},
+      regressions: [],
+      improvements: []
+    };
+    if (passRateDelta < 0) {
+      delta.regressions.push(`passRate ${basePassRate} -> ${candidatePassRate}`);
+      regressions.push(`${benchmarkId} pass rate regressed (${basePassRate} -> ${candidatePassRate})`);
+    } else if (passRateDelta > 0) {
+      delta.improvements.push(`passRate ${basePassRate} -> ${candidatePassRate}`);
+      improvements.push(`${benchmarkId} pass rate improved (${basePassRate} -> ${candidatePassRate})`);
+    }
+    const metricDelta = compareMetricDeltas(baseRun.metrics, candidateRun.metrics);
+    delta.metricDeltas = metricDelta.deltas;
+    for (const regression of metricDelta.regressions) {
+      delta.regressions.push(regression);
+      regressions.push(`${benchmarkId} ${regression}`);
+    }
+    for (const improvement of metricDelta.improvements) {
+      delta.improvements.push(improvement);
+      improvements.push(`${benchmarkId} ${improvement}`);
+    }
+    deltas.push(delta);
+  }
+  return {
+    passed: regressions.length === 0,
+    baseRootDir: baseSnapshot.status.rootDir,
+    candidateRootDir: candidateSnapshot.status.rootDir,
+    comparedBenchmarks: deltas.length,
+    missingCandidateBenchmarks,
+    invalidArtifacts: {
+      base: {
+        benchmarks: baseSnapshot.status.invalidBenchmarks.length,
+        runs: baseSnapshot.status.invalidRuns.length,
+        shadows: baseSnapshot.status.invalidShadows.length
+      },
+      candidate: {
+        benchmarks: candidateSnapshot.status.invalidBenchmarks.length,
+        runs: candidateSnapshot.status.invalidRuns.length,
+        shadows: candidateSnapshot.status.invalidShadows.length
+      }
+    },
+    regressions,
+    improvements,
+    deltas
+  };
+}
+
+export {
+  resolveEvalStoreDir,
+  validateEvalBenchmarkManifest,
+  validateEvalRunSummary,
+  validateEvalBaselineSnapshot,
+  validateEvalShadowRecallRecord,
+  validateEvalBenchmarkPack,
+  importEvalBenchmarkPack,
+  recordEvalShadowRecall,
+  getEvalHarnessStatus,
+  createEvalBaselineSnapshot,
+  runEvalBaselineDeltaReport,
+  runEvalStoredBaselineCiGate,
+  runEvalBenchmarkCiGate
+};
+//# sourceMappingURL=chunk-D46YSIYX.js.map