@joshuaswarren/openclaw-engram 9.0.52 → 9.0.54

package/README.md

@@ -33,6 +33,7 @@ AI agents forget everything between conversations. Engram fixes that.
 - **Benchmark-first roadmap** — Engram now has an evaluation harness with live shadow recall recording and a CI benchmark delta gate, so memory improvements can be measured and regression-checked instead of argued from anecdotes.
 - **Baseline snapshot discipline** — Engram can now, when `benchmarkBaselineSnapshotsEnabled` is enabled, capture typed baseline snapshots of the latest completed benchmark runs so later PR delta reporting can compare candidates against a stable stored reference instead of an ad hoc branch state.
 - **Named baseline delta reporting** — Engram can now, when `benchmarkDeltaReporterEnabled` is enabled, compare the current eval store against a stored baseline snapshot, emit a machine-readable delta report plus markdown summary, and fail fast when a candidate regresses a benchmark that previously passed.
+- **Required CI baseline gate** — Engram's `eval-benchmark-gate` workflow now reads a named stored baseline snapshot from the base branch fixture store and blocks merges when the candidate branch regresses relative to that required baseline.
 - **Objective-state recall** — Engram can now store normalized file, process, and tool outcomes and, when `objectiveStateRecallEnabled` is enabled, inject the most relevant objective-state snapshots back into recall context as a separate `Objective State` section.
 - **Causal trajectory graph foundation** — Engram can now persist typed `goal -> action -> observation -> outcome -> follow-up` chains when `causalTrajectoryMemoryEnabled` is enabled and, with `actionGraphRecallEnabled`, emit deterministic action-conditioned edges into the causal graph for later trajectory-aware retrieval.
 - **Causal trajectory recall** — Engram can now, when `causalTrajectoryRecallEnabled` is enabled, inject prompt-relevant causal chains back into recall context as a separate `Causal Trajectories` section with lightweight match explainability.
@@ -215,6 +216,12 @@ Key settings:
 | `evalShadowModeEnabled` | `false` | Record live recall decisions to the eval store without changing injected output |
 | `benchmarkBaselineSnapshotsEnabled` | `false` | Enable versioned baseline snapshot artifacts for the latest completed benchmark runs |
 | `benchmarkDeltaReporterEnabled` | `false` | Enable named-baseline delta reports against the current eval store |
+
+The repo's required benchmark check uses the committed fixture snapshot at
+`tests/fixtures/eval-ci/store/baselines/required-main.json` as the stable
+release baseline for PR gating. During the rollout PR that first introduces
+that file, the gate bootstraps from the candidate branch snapshot once; after
+that, PRs resolve the required baseline from the base branch checkout.
 | `evalStoreDir` | `{memoryDir}/state/evals` | Root directory for benchmark packs, run summaries, and shadow recall records |
 | `objectiveStateMemoryEnabled` | `false` | Enable the objective-state memory foundation for normalized world/tool state snapshots |
 | `objectiveStateSnapshotWritesEnabled` | `false` | Permit objective-state snapshot writers to persist typed state records |

package/dist/index.js

@@ -6780,7 +6780,7 @@ function serializeFrontmatter(fm) {
       lines.push(`  - targetId: ${link.targetId}`);
       lines.push(`    linkType: ${link.linkType}`);
       lines.push(`    strength: ${link.strength}`);
-      if (link.reason) lines.push(`    reason: "${link.reason.replace(/"/g, '\\"')}"`);
+      if (link.reason) lines.push(`    reason: ${JSON.stringify(link.reason)}`);
     }
   }
   if (fm.intentGoal) lines.push(`intentGoal: ${fm.intentGoal}`);
@@ -6795,6 +6795,18 @@ function serializeFrontmatter(fm) {
   lines.push("---");
   return lines.join("\n");
 }
+function parseLinkReasonValue(rawValue) {
+  const legacyValue = rawValue.replace(/\\"/g, '"');
+  const looksLikeLegacyPath = !rawValue.includes("\\\\") && (/[A-Za-z]:\\[A-Za-z0-9._ -]+(?:\\[A-Za-z0-9._ -]+)+/.test(rawValue) || /\\[A-Za-z0-9._ -]+\\[A-Za-z0-9._ -]+/.test(rawValue));
+  if (looksLikeLegacyPath) {
+    return legacyValue;
+  }
+  try {
+    return JSON.parse(`"${rawValue}"`);
+  } catch {
+    return legacyValue;
+  }
+}
 function parseFrontmatter3(raw) {
   const match = raw.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
   if (!match) return null;
@@ -6897,14 +6909,14 @@ function parseFrontmatter3(raw) {
   if (fmBlock.includes("links:")) {
     const links = [];
     const linkMatches = fmBlock.matchAll(
-      /- targetId: (\S+)\s+linkType: (\S+)\s+strength: ([\d.]+)(?:\s+reason: "([^"]*)")?/g
+      /- targetId: (\S+)\s+linkType: (\S+)\s+strength: ([\d.]+)(?:\s+reason: "((?:\\.|[^"\\])*)")?/g
     );
     for (const match2 of linkMatches) {
       links.push({
         targetId: match2[1],
         linkType: match2[2],
         strength: parseFloat(match2[3]),
-        reason: match2[4] || void 0
+        reason: match2[4] ? parseLinkReasonValue(match2[4]) : void 0
       });
     }
     if (links.length > 0) {
@@ -12239,6 +12251,13 @@ async function runEvalBaselineDeltaReport(options) {
   if (!baselineSnapshot) {
     throw new Error(`benchmark baseline snapshot not found: ${snapshotId}`);
   }
+  return buildEvalBaselineDeltaReport({
+    baselineSnapshot,
+    candidateSnapshot
+  });
+}
+function buildEvalBaselineDeltaReport(options) {
+  const { baselineSnapshot, candidateSnapshot } = options;
   const regressions = [];
   const improvements = [];
   if (candidateSnapshot.status.invalidBenchmarks.length > 0) {