@compr/opscontext-mcp 2.6.0 → 2.6.1

package/CHANGELOG.md

@@ -4,6 +4,21 @@ All notable changes to OpsContext for AI Agents (previously ContextEngine — MC
 
 > Entries for 2.2.0 through 2.4.0 were not backfilled here; see `docs/sessions/SESSION_19` through `SESSION_21` for those releases.
 
+## [2.6.1] 2026-09-06: readers never embed; a store record without tags no longer breaks search
+
+### Fixed
+
+- **`search_context` and `score_project` answered "Cannot read properties of undefined (reading
+  'join')"** on every server, 2.5.9 included, while `save_learning` worked. `searchLearnings`
+  dereferenced `tags.join` on 20 agent-saved store records that have no `tags` array, inside the
+  firewall's learning injection, which runs for every tool that passes a context hint. Optional
+  fields are now read with fallbacks (`[LEARNING-FIELDS-ARE-OPTIONAL-ON-READ]`); the records are
+  not rewritten.
+- **A reader that starts before its indexer has published** (every reader when all windows
+  restart at once) used to build locally and embed the whole corpus itself. It now builds
+  keyword-only, embeds nothing, and adopts the shared index the moment it appears. The trial
+  script starts all three servers together and asserts zero embed runs on the readers.
+
 ## [2.6.0] 2026-09-05: One indexer, many readers
 
 Every Claude Code chat spawns its own MCP server, plus launchd, plus VS Code. Measured on

package/dist/cli.js

@@ -831,7 +831,7 @@ async function cliSaveLearning(args) {
         console.log(`   Project:  ${learning.project}`);
     if (learning.context)
         console.log(`   Context:  ${learning.context}`);
-    console.log(`   Tags:     ${learning.tags.join(", ")}`);
+    console.log(`   Tags:     ${(learning.tags ?? []).join(", ")}`); // [LOCK] [LEARNING-FIELDS-ARE-OPTIONAL-ON-READ]
 }
 async function cliDeleteLearning(id) {
     if (!id) {

package/dist/index.js

@@ -1083,7 +1083,7 @@ server.tool("save_learning", "Save a permanent operational rule learned during a
             `- **ID:** \`${learning.id}\``,
             `- **Category:** ${category}`,
             project ? `- **Project:** ${project}` : "",
-            `- **Tags:** ${learning.tags.join(", ")}`,
+            `- **Tags:** ${(learning.tags ?? []).join(", ")}`, // [LOCK] [LEARNING-FIELDS-ARE-OPTIONAL-ON-READ]
             ``,
             `📊 Store: ${stats.total} learnings across ${Object.keys(stats.categories).length} categories`,
             ``,
@@ -1378,7 +1378,7 @@ async function main() {
     //    always loads for whoever embeds or answers queries. [LOCK] [EMBEDDINGS-ARE-CONTENT-ADDRESSED]
     //    A reader that adopted the index already has its vectors and loads the model on its first
     //    semantic query, not before: 300 MB per process is worth waiting for.
-    if (!adopted) {
+    if (!adopted && role === "indexer") {
         vectorStore = loadEmbeddingStore().vectors;
         if (vectorStore.size > 0)
             console.error(`[ContextEngine] 💾 Embedding store: ${vectorStore.size} vectors`);
@@ -1390,6 +1390,12 @@ async function main() {
             publishIndex();
         });
     }
+    else if (!adopted) {
+        // A reader that started before its indexer published (every window restarted at once):
+        // keyword search from its own build, no embedding of its own, the index adopted the moment
+        // it appears. Only indexers embed; that is the whole point. [LOCK] [ONE-INDEXER-MANY-READERS]
+        console.error(`[ContextEngine] ⏳ Reader without an index: keyword search only until pid ${indexerPid ?? "?"} publishes`);
+    }
     // 5. Watch (indexer) or poll (reader), and keep the election running
     if (role === "indexer")
         startWatching();

package/dist/learnings.js

@@ -448,7 +448,16 @@ export function searchLearnings(query) {
         return store.learnings;
     const scored = [];
     for (const learning of store.learnings) {
-        const text = `${learning.category} ${learning.rule} ${learning.context} ${learning.project || ""} ${learning.tags.join(" ")}`.toLowerCase();
+        // [LOCKED] [LEARNING-FIELDS-ARE-OPTIONAL-ON-READ] 2026-09-06
+        // [NEVER] dereference an optional field of a store record (tags, context, project) without a
+        //         fallback on a read path, and [NEVER] "fix" such a record by rewriting the store.
+        // WHY: 20 agent-saved records (s57_*, s78_*, kept on purpose in the SESSION_25 prune) have no
+        //      `tags` array. `learning.tags.join` threw here, inside the firewall's learning injection,
+        //      so every tool that passes a context hint (search_context, score_project, ...) answered
+        //      "Cannot read properties of undefined (reading 'join')" on every server, on 2.5.9 as on
+        //      2.6.0, while save_learning (exempt from the firewall) worked. Another chat noticed.
+        // FIX: read with fallbacks; the records stay as they are (the store is data, not ours to edit).
+        const text = `${learning.category} ${learning.rule} ${learning.context ?? ""} ${learning.project || ""} ${(learning.tags ?? []).join(" ")}`.toLowerCase();
         let score = 0;
         for (const token of tokens) {
             if (text.includes(token)) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@compr/opscontext-mcp",
-  "version": "2.6.0",
+  "version": "2.6.1",
   "description": "OpsContext for AI Agents — read-only fleet visibility (PM2/nginx/Docker/git/cron) + tamper-evident audit log + policy-as-code hooks. The ops + compliance layer Claude Code can't grow natively.",
   "type": "module",
   "main": "dist/index.js",