grepmax 0.17.0 → 0.17.1

package/dist/commands/mcp.js

@@ -521,7 +521,7 @@ exports.mcp = new commander_1.Command("mcp")
                         }
                     }
                 }
-                const result = yield searcher.search(query, limit, { rerank: true }, Object.keys(filters).length > 0 ? filters : undefined, pathPrefix);
+                const result = yield searcher.search(query, limit, { rerank: process.env.GMAX_RERANK === "1" }, Object.keys(filters).length > 0 ? filters : undefined, pathPrefix);
                 if (!result.data || result.data.length === 0) {
                     return ok("No matches found. Try broadening your query, using fewer keywords, or check `gmax status` to verify the project is indexed.");
                 }
@@ -1578,7 +1578,7 @@ exports.mcp = new commander_1.Command("mcp")
                 const rel = (p) => p.startsWith(`${projectRoot}/`) ? p.slice(projectRoot.length + 1) : p;
                 if (query) {
                     const searcher = getSearcher();
-                    const response = yield searcher.search(query, limit, { rerank: true }, {}, projectRoot);
+                    const response = yield searcher.search(query, limit, { rerank: process.env.GMAX_RERANK === "1" }, {}, projectRoot);
                     const changedSet = new Set(changedFiles);
                     let filtered = response.data.filter((r) => changedSet.has(String(r.path || "")));
                     if (role)
@@ -1753,7 +1753,7 @@ exports.mcp = new commander_1.Command("mcp")
             const limit = Math.min(Math.max(Number(args.limit) || 10, 1), 25);
             try {
                 const searcher = getSearcher();
-                const response = yield searcher.search(topic, limit, { rerank: true }, {}, projectRoot);
+                const response = yield searcher.search(topic, limit, { rerank: process.env.GMAX_RERANK === "1" }, {}, projectRoot);
                 if (response.data.length === 0)
                     return ok(`No results found for "${topic}".`);
                 const rel = (p) => p.startsWith(`${projectRoot}/`) ? p.slice(projectRoot.length + 1) : p;

package/dist/commands/search.js

@@ -571,7 +571,7 @@ Examples:
                             ? searchFilters
                             : undefined,
                         pathPrefix: pathFilter,
-                        rerank: true,
+                        rerank: process.env.GMAX_RERANK === "1",
                         explain: options.explain,
                         includeSkeletons: options.skeleton,
                         includeGraph: options.symbol,
@@ -668,7 +668,7 @@ Examples:
                 }
             }
             const searcher = new searcher_1.Searcher(vectorDb);
-            searchResult = yield searcher.search(pattern, parseInt(options.m, 10), { rerank: true, explain: options.explain }, Object.keys(searchFilters).length > 0
+            searchResult = yield searcher.search(pattern, parseInt(options.m, 10), { rerank: process.env.GMAX_RERANK === "1", explain: options.explain }, Object.keys(searchFilters).length > 0
                 ? searchFilters
                 : undefined, pathFilter);
         } // end if (!searchResult) — in-process fallback

package/dist/lib/daemon/daemon.js

@@ -109,6 +109,8 @@ class Daemon {
         this.startTime = Date.now();
         this.heartbeatInterval = null;
         this.idleInterval = null;
+        this.heartbeatTick = 0;
+        this.mlxRecoveryInFlight = false;
         this.shuttingDown = false;
         this.pendingOps = new Set();
         this.watcherFailCount = new Map();
@@ -289,6 +291,14 @@ class Daemon {
                 }
                 catch (_a) { }
                 (0, log_rotate_1.rotateLogFds)(path.join(config_1.PATHS.logsDir, "daemon.log"));
+                // Every 5 ticks (5 min), probe the MLX embed server and respawn if
+                // it's gone zombie (port held but /health unresponsive). Closes the
+                // 42h-degradation window where workers silently fell back to ONNX CPU
+                // after a frozen MLX process kept the port bound (v0.17.0 bug #1).
+                this.heartbeatTick++;
+                if (this.heartbeatTick % 5 === 0) {
+                    void this.checkMlxHealth();
+                }
             }, HEARTBEAT_INTERVAL_MS);
             // 10. Idle timeout (skip when disabled via env)
             if (IDLE_TIMEOUT_MS > 0) {
@@ -796,7 +806,7 @@ class Daemon {
             this.lastActivity = Date.now();
             let result;
             try {
-                result = yield searcher.search(payload.query, payload.limit, { rerank: payload.rerank !== false, explain: payload.explain === true }, payload.filters, payload.pathPrefix, undefined, signal);
+                result = yield searcher.search(payload.query, payload.limit, { rerank: payload.rerank === true, explain: payload.explain === true }, payload.filters, payload.pathPrefix, undefined, signal);
             }
             catch (err) {
                 if ((err === null || err === void 0 ? void 0 : err.name) === "AbortError") {
@@ -1185,6 +1195,31 @@ class Daemon {
             return null;
         }
     }
+    checkMlxHealth() {
+        return __awaiter(this, void 0, void 0, function* () {
+            if (this.shuttingDown || this.mlxRecoveryInFlight)
+                return;
+            if (yield this.isMlxServerUp())
+                return;
+            const port = parseInt(process.env.MLX_EMBED_PORT || "8100", 10);
+            const stalePid = this.getPortPid(port);
+            if (!stalePid)
+                return; // No process — let the next user-facing path spawn it.
+            this.mlxRecoveryInFlight = true;
+            try {
+                console.log(`[daemon] MLX zombie detected on port ${port} (PID ${stalePid}) — killing and respawning`);
+                yield (0, process_1.killProcess)(stalePid);
+                yield new Promise((r) => setTimeout(r, 500));
+                yield this.ensureMlxServer();
+            }
+            catch (err) {
+                console.error(`[daemon] MLX recovery failed: ${err instanceof Error ? err.message : String(err)}`);
+            }
+            finally {
+                this.mlxRecoveryInFlight = false;
+            }
+        });
+    }
     ensureMlxServer(mlxModel) {
         return __awaiter(this, void 0, void 0, function* () {
             if (yield this.isMlxServerUp()) {

package/dist/lib/daemon/ipc-handler.js

@@ -169,7 +169,7 @@ function handleCommand(daemon, cmd, conn) {
                             limit: limitRaw,
                             filters,
                             pathPrefix: typeof cmd.pathPrefix === "string" ? cmd.pathPrefix : undefined,
-                            rerank: cmd.rerank !== false,
+                            rerank: cmd.rerank === true,
                             explain: cmd.explain === true,
                             includeSkeletons: cmd.includeSkeletons === true,
                             skeletonLimit: skeletonLimitRaw,

package/dist/lib/search/searcher.js

@@ -340,7 +340,11 @@ class Searcher {
         return __awaiter(this, void 0, void 0, function* () {
             var _a, _b, _c, _d, _e, _f, _g, _h;
             const finalLimit = top_k !== null && top_k !== void 0 ? top_k : 10;
-            const doRerank = (_a = _search_options === null || _search_options === void 0 ? void 0 : _search_options.rerank) !== null && _a !== void 0 ? _a : true;
+            // ColBERT rerank is opt-in as of v0.17.1. On the 97-case eval it
+            // regresses MRR@10 by ~3% and doubles query latency; sweep across
+            // FUSED_WEIGHT ∈ {0,0.1,0.5,1,2} showed rerank scores dominate
+            // fused scores ~30:1 so blend tuning can't recover the loss.
+            const doRerank = (_a = _search_options === null || _search_options === void 0 ? void 0 : _search_options.rerank) !== null && _a !== void 0 ? _a : false;
             const explain = (_b = _search_options === null || _search_options === void 0 ? void 0 : _search_options.explain) !== null && _b !== void 0 ? _b : false;
             const searchIntent = intent || (0, intent_1.detectIntent)(query);
             const pool = (0, pool_1.getWorkerPool)();

package/dist/lib/workers/pool.js

@@ -83,6 +83,13 @@ const TASK_TIMEOUT_MS = (() => {
     return 120000;
 })();
 const FORCE_KILL_GRACE_MS = 200;
+// Longer grace for idle reaps: the worker isn't urgently in the way, and a
+// graceful SIGTERM lets ONNX free ~1GB of model memory. But if SIGTERM is
+// ignored (a worker burning 100% CPU inside a native ONNX matmul tight loop
+// won't service signals — the 42h zombie we saw in v0.17.0 validation),
+// escalate to SIGKILL. ~5s is well above ONNX teardown time but short
+// enough that the reap loop self-heals within a minute.
+const REAP_FORCE_KILL_GRACE_MS = 5000;
 class ProcessWorker {
     constructor(modulePath, execArgv, maxMemoryMb) {
         this.modulePath = modulePath;
@@ -444,11 +451,30 @@ class WorkerPool {
             w.child.removeAllListeners("message");
             w.child.removeAllListeners("exit");
             w.child.removeAllListeners("error");
+            const pid = w.child.pid;
             try {
                 w.child.kill("SIGTERM");
             }
             catch (_a) { }
             this.workers = this.workers.filter((x) => x !== w);
+            // SIGTERM is ignored by a worker stuck inside a native ONNX matmul
+            // tight loop. Escalate to SIGKILL if the process is still alive after
+            // the grace period. Rare; warn-level so it's visible if it fires.
+            if (pid !== undefined) {
+                setTimeout(() => {
+                    try {
+                        process.kill(pid, 0);
+                        (0, logger_1.log)("pool", `reap escalation: SIGTERM ignored by PID:${pid}, sending SIGKILL`);
+                        try {
+                            process.kill(pid, "SIGKILL");
+                        }
+                        catch (_a) { }
+                    }
+                    catch (_b) {
+                        // ESRCH — process already gone, nothing to do.
+                    }
+                }, REAP_FORCE_KILL_GRACE_MS);
+            }
         });
     }
     destroy() {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "author": "Robert Owens <78518764+reowens@users.noreply.github.com>",
   "homepage": "https://github.com/reowens/grepmax",
   "bugs": {

package/plugins/grepmax/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
   "author": {
     "name": "Robert Owens",