@remnic/cli 9.52.1 → 9.52.2

package/dist/index.js

@@ -2731,6 +2731,9 @@ var RUN_VALUE_FLAGS = Object.freeze([
   "--calibration-dir",
   "--calibration-local-config-sha256",
   "--calibration-frontier-config-sha256",
+  "--source-result-id",
+  "--expected-answer-set-sha256",
+  "--expected-question-id-list-sha256",
   "--task-ids-file",
   "--expected-task-id-list-sha256",
   "--drain-timeout",
@@ -3852,6 +3855,25 @@ function parseBenchArgs(argv) {
   };
 }
 
+// src/bench-calibration-binding.ts
+function validateCalibrationProvenancePinSet(binding, hasBothConfigPins) {
+  const pins = [binding.sourceResultId, binding.expectedAnswerSetSha256, binding.expectedQuestionIdListSha256];
+  const hasAny = pins.some((pin) => pin !== void 0);
+  if (hasAny && (!pins.every((pin) => pin !== void 0) || !hasBothConfigPins)) {
+    throw new Error(
+      "--source-result-id, --expected-answer-set-sha256, and --expected-question-id-list-sha256 must be supplied together with both calibration configuration pins."
+    );
+  }
+  return hasAny;
+}
+function assertCalibrationProvenanceMatches(binding, state, benchmarkId) {
+  if (binding.sourceResultId !== void 0 && (binding.sourceResultId !== state.sourceResultId || binding.expectedAnswerSetSha256 !== state.answerSetHash || binding.expectedQuestionIdListSha256 !== state.orderedQuestionIdsHash)) {
+    throw new Error(
+      `Calibration provenance mismatch for ${benchmarkId}; refusing to attach calibration state from an unpinned source.`
+    );
+  }
+}
+
 // src/bench-status.ts
 import { mkdir, readFile, readdir as readdir2, rename, writeFile } from "fs/promises";
 import path8 from "path";
@@ -6586,11 +6608,12 @@ Options:
                            Calibration-only timeout for each frontier judge call
   --max-429-wait <ms>      Maximum cumulative 429 retry wait for provider calls
   --disable-thinking       Disable thinking for supported provider-backed models
-  --source-result-id <id>  Exact full stored result used by judge-calibrate
+  --source-result-id <id>  Exact stored result used by or expected for calibration
   --expected-answer-set-sha256 <sha256>
-                           Expected deterministic calibration-slice payload hash
+                           Expected deterministic calibration answer-set hash
   --expected-question-id-list-sha256 <sha256>
-                           Expected ordered source task-ID list hash
+                           Expected ordered source task-ID list hash; later runs may
+                           pass all three source pins with both config pins
   --calibration-dir <path> Private final-state and resumable-checkpoint directory
   --calibration-local-config-sha256 <sha256>
                            Required on later runs that attach calibration state
@@ -7931,10 +7954,10 @@ async function calibrateBenchJudges(parsed, rawArgs) {
     result,
     calibrationDir,
     calibrationIdentities,
-    { sourceResultId: loaded.meta.id, localJudgeConfigHash, frontierJudgeConfigHash }
+    { sourceResultId: loaded.meta.id, orderedQuestionIdsHash, localJudgeConfigHash, frontierJudgeConfigHash }
   );
   const persisted = await bench.loadJudgeCalibrationState(benchmarkId, calibrationDir);
-  if (!persisted || persisted.kappa !== result.kappa || persisted.warning !== result.warning || persisted.localJudgeConfigHash !== localJudgeConfigHash || persisted.frontierJudgeConfigHash !== frontierJudgeConfigHash) {
+  if (!persisted || persisted.kappa !== result.kappa || persisted.warning !== result.warning || persisted.localJudgeConfigHash !== localJudgeConfigHash || persisted.orderedQuestionIdsHash !== orderedQuestionIdsHash || persisted.frontierJudgeConfigHash !== frontierJudgeConfigHash) {
     console.error(
       `ERROR: calibration state round-trip failed for ${benchmarkId} (wrote kappa ${result.kappa}, read back ${persisted ? persisted.kappa : "nothing"}). Re-run judge-calibrate.`
     );
@@ -8658,8 +8681,10 @@ async function preparePersistedJudgeCalibrationAttachment(benchModule, benchmark
   const hasFrontierPin = Boolean(calibrationBinding.calibrationFrontierConfigSha256);
   const hasAnyPin = hasLocalPin || hasFrontierPin;
   const hasBothPins = hasLocalPin && hasFrontierPin;
+  const hasAnyProvenancePin = validateCalibrationProvenancePinSet(calibrationBinding, hasBothPins);
+  const hasAnyBindingPin = hasAnyPin || hasAnyProvenancePin;
   if (benchmarkId === "ama-bench" && calibrationBinding.amaBenchJudgeProtocol === "recommended") {
-    if (hasAnyPin) {
+    if (hasAnyBindingPin) {
       throw new Error(
         "AMA-Bench recommended-protocol runs cannot attach default-protocol judge calibration; remove both calibration pins or calibrate the recommended prompt contract separately."
       );
@@ -8676,7 +8701,7 @@ async function preparePersistedJudgeCalibrationAttachment(benchModule, benchmark
   );
   const state = await benchModule.loadJudgeCalibrationState?.(benchmarkId, calibrationDir);
   if (!state) {
-    if (hasAnyPin) {
+    if (hasAnyBindingPin) {
       throw new Error(
         `Calibration binding pins were supplied for ${benchmarkId}, but no valid calibration state could be loaded from ${calibrationDir}; rerun judge-calibrate or remove both pins.`
       );
@@ -8686,7 +8711,7 @@ async function preparePersistedJudgeCalibrationAttachment(benchModule, benchmark
   if (state.localJudgeProvider !== void 0 && state.localJudgeModel !== void 0) {
     const matchesLocal = runJudgeProvider?.provider === state.localJudgeProvider && runJudgeProvider.model === state.localJudgeModel;
     if (!matchesLocal) {
-      if (hasBothPins) {
+      if (hasAnyBindingPin) {
         throw new Error(
           `Calibration binding pins for ${benchmarkId} target the calibrated local judge ${state.localJudgeProvider}/${state.localJudgeModel}, but the run judge is ${runJudgeProvider?.provider ?? "unset"}/${runJudgeProvider?.model ?? "unset"}; refusing to ignore explicit pins.`
         );
@@ -8702,7 +8727,7 @@ async function preparePersistedJudgeCalibrationAttachment(benchModule, benchmark
   const resolvedRunJudgeConfigHash = hashCalibrationProviderConfig(runJudgeProvider);
   const hasCompleteLocalIdentity = state.localJudgeProvider !== void 0 && state.localJudgeModel !== void 0;
   if (!hasCompleteLocalIdentity && resolvedRunJudgeConfigHash !== state.localJudgeConfigHash) {
-    if (hasBothPins) {
+    if (hasAnyBindingPin) {
       throw new Error(
         `Calibration binding pins for ${benchmarkId} do not match the resolved run judge configuration; refusing to ignore explicit pins.`
       );
@@ -8717,6 +8742,7 @@ async function preparePersistedJudgeCalibrationAttachment(benchModule, benchmark
   if (calibrationBinding.calibrationLocalConfigSha256 !== state.localJudgeConfigHash || calibrationBinding.calibrationFrontierConfigSha256 !== state.frontierJudgeConfigHash) {
     throw new Error(`Calibration configuration hash mismatch for ${benchmarkId}; refusing to attach stale kappa.`);
   }
+  assertCalibrationProvenanceMatches(calibrationBinding, state, benchmarkId);
   if (resolvedRunJudgeConfigHash !== state.localJudgeConfigHash) {
     throw new Error(
       `Resolved run judge configuration hash mismatch for ${benchmarkId}; expected sha256:${state.localJudgeConfigHash}, got sha256:${resolvedRunJudgeConfigHash}. Refusing to attach stale kappa before benchmark dispatch.`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@remnic/cli",
-  "version": "9.52.1",
+  "version": "9.52.2",
   "description": "CLI for Remnic memory — init, query, doctor, daemon management",
   "type": "module",
   "main": "dist/index.js",
@@ -26,25 +26,25 @@
   },
   "dependencies": {
     "yaml": "^2.4.2",
-    "@remnic/plugin-pi": "^9.52.1",
-    "@remnic/server": "^9.52.1",
-    "@remnic/core": "^9.52.1"
+    "@remnic/server": "^9.52.2",
+    "@remnic/core": "^9.52.2",
+    "@remnic/plugin-pi": "^9.52.2"
   },
   "peerDependencies": {
-    "@remnic/bench": "^9.52.1",
-    "@remnic/plugin-openclaw": "^9.52.1",
-    "@remnic/export-weclone": "^9.52.1",
-    "@remnic/import-weclone": "^9.52.1",
-    "@remnic/import-chatgpt": "^9.52.1",
-    "@remnic/import-claude": "^9.52.1",
-    "@remnic/import-gemini": "^9.52.1",
-    "@remnic/import-lossless-claw": "^9.52.1",
-    "@remnic/import-mem0": "^9.52.1",
-    "@remnic/import-supermemory": "^9.52.1",
-    "@remnic/connector-limitless": "^9.52.1",
-    "@remnic/connector-bee": "^9.52.1",
-    "@remnic/connector-omi": "^9.52.1",
-    "@remnic/capture-audio": "^9.52.1"
+    "@remnic/bench": "^9.52.2",
+    "@remnic/plugin-openclaw": "^9.52.2",
+    "@remnic/export-weclone": "^9.52.2",
+    "@remnic/import-weclone": "^9.52.2",
+    "@remnic/import-chatgpt": "^9.52.2",
+    "@remnic/import-claude": "^9.52.2",
+    "@remnic/import-gemini": "^9.52.2",
+    "@remnic/import-lossless-claw": "^9.52.2",
+    "@remnic/import-mem0": "^9.52.2",
+    "@remnic/import-supermemory": "^9.52.2",
+    "@remnic/connector-limitless": "^9.52.2",
+    "@remnic/connector-bee": "^9.52.2",
+    "@remnic/connector-omi": "^9.52.2",
+    "@remnic/capture-audio": "^9.52.2"
   },
   "peerDependenciesMeta": {
     "@remnic/bench": {
@@ -93,19 +93,19 @@
   "devDependencies": {
     "tsup": "^8.5.1",
     "typescript": "^5.9.3",
-    "@remnic/bench": "9.52.1",
-    "@remnic/plugin-openclaw": "9.52.1",
-    "@remnic/import-weclone": "9.52.1",
-    "@remnic/import-chatgpt": "9.52.1",
-    "@remnic/export-weclone": "9.52.1",
-    "@remnic/import-gemini": "9.52.1",
-    "@remnic/import-lossless-claw": "9.52.1",
-    "@remnic/import-mem0": "9.52.1",
-    "@remnic/import-supermemory": "9.52.1",
-    "@remnic/import-claude": "9.52.1",
-    "@remnic/connector-limitless": "9.52.1",
-    "@remnic/connector-bee": "9.52.1",
-    "@remnic/connector-omi": "9.52.1"
+    "@remnic/plugin-openclaw": "9.52.2",
+    "@remnic/bench": "9.52.2",
+    "@remnic/export-weclone": "9.52.2",
+    "@remnic/import-chatgpt": "9.52.2",
+    "@remnic/import-lossless-claw": "9.52.2",
+    "@remnic/import-claude": "9.52.2",
+    "@remnic/import-weclone": "9.52.2",
+    "@remnic/import-mem0": "9.52.2",
+    "@remnic/import-supermemory": "9.52.2",
+    "@remnic/import-gemini": "9.52.2",
+    "@remnic/connector-limitless": "9.52.2",
+    "@remnic/connector-omi": "9.52.2",
+    "@remnic/connector-bee": "9.52.2"
   },
   "license": "MIT",
   "repository": {