grepmax 0.26.4 → 0.26.6

package/dist/lib/workers/pool.js

@@ -58,26 +58,6 @@ const config_1 = require("../../config");
 const embedding_generation_1 = require("../index/embedding-generation");
 const index_config_1 = require("../index/index-config");
 const logger_1 = require("../utils/logger");
-function reviveBufferLike(input) {
-    if (input &&
-        typeof input === "object" &&
-        "type" in input &&
-        input.type === "Buffer" &&
-        Array.isArray(input.data)) {
-        return Buffer.from(input.data);
-    }
-    return input;
-}
-function reviveProcessFileResult(result) {
-    if (!result || !Array.isArray(result.vectors))
-        return result;
-    const vectors = result.vectors.map((v) => {
-        const revived = reviveBufferLike(v.colbert);
-        return revived && (Buffer.isBuffer(revived) || revived instanceof Int8Array)
-            ? Object.assign(Object.assign({}, v), { colbert: revived }) : v;
-    });
-    return Object.assign(Object.assign({}, result), { vectors });
-}
 const TASK_TIMEOUT_MS = (() => {
     var _a;
     const fromEnv = Number.parseInt((_a = process.env.GMAX_WORKER_TASK_TIMEOUT_MS) !== null && _a !== void 0 ? _a : "", 10);
@@ -155,6 +135,7 @@ class ProcessWorker {
         this.child = childProcess.fork(modulePath, {
             execArgv: [...memArgs, ...execArgv],
             env: Object.assign(Object.assign({}, process.env), embeddingEnvironment),
+            serialization: "advanced",
         });
     }
 }
@@ -182,13 +163,13 @@ const MIN_KEEP_WORKERS = 1;
 // Recycle a worker whose RSS has grown past this. ONNX native memory lives
 // outside V8, so --max-old-space-size can't bound it; replacing the worker is
 // the only reclaim. The threshold must sit ABOVE the healthy working set or
-// the pool executes warm workers for normal behavior: measured steady state
-// is ~700-800 MB with MLX serving dense embeds and ~900-1050 MB in ONNX CPU
-// fallback (granite session loaded in-process), flat across task count and
-// file size. The old 800 default was calibrated for the pre-
-// enableCpuMemArena:false era (arenas pinned ~2 GB) and sat inside the
-// healthy range — every busy worker was recycled after nearly every task,
-// paying a full model reload each time. 0 (or negative) disables the check.
+// the pool executes warm workers for normal behavior. A post-tree-cleanup
+// soak on 2026-08-04 measured MLX p50/p95/max at 1056/1110/1122 MB (900 tasks)
+// and ONNX at 1030/1182/1195 MB (600 tasks). Applying the 1.25x ONNX-p95 rule
+// and rounding to a 128 MB boundary retains 1536 MB. The old 800 default sat
+// inside the healthy range and recycled warm workers after normal tasks; the
+// previously proposed 1024 MLX-only override is also below measured p50. 0
+// (or negative) disables the check.
 const WORKER_RSS_RECYCLE_MB = (() => {
     var _a;
     const fromEnv = Number.parseInt((_a = process.env.GMAX_WORKER_RSS_RECYCLE_MB) !== null && _a !== void 0 ? _a : "", 10);
@@ -255,6 +236,12 @@ class WorkerPool {
             .map((w) => w.child.pid)
             .filter((pid) => pid !== undefined);
     }
+    /** Latest worker-reported RSS, sampled after each IPC message. */
+    getWorkerMemoryStats() {
+        return this.workers.flatMap((worker) => worker.child.pid === undefined
+            ? []
+            : [{ pid: worker.child.pid, rssBytes: worker.lastRssBytes }]);
+    }
     clearTaskTimeout(task) {
         if (task.timeout) {
             clearTimeout(task.timeout);
@@ -378,16 +365,12 @@ class WorkerPool {
                 task.reject(new Error(msg.error));
             }
             else {
-                let result = msg.result;
-                if (task.method === "processFile") {
-                    result = reviveProcessFileResult(result);
-                }
                 const elapsed = task.startTime
                     ? `${Date.now() - task.startTime}ms`
                     : "?ms";
                 const filePath = (_h = (_f = (_e = task.payload) === null || _e === void 0 ? void 0 : _e.path) !== null && _f !== void 0 ? _f : (_g = task.payload) === null || _g === void 0 ? void 0 : _g.absolutePath) !== null && _h !== void 0 ? _h : "";
                 (0, logger_1.debug)("pool", `complete task=${task.id} method=${task.method} ${elapsed}${filePath ? ` file=${filePath}` : ""}`);
-                task.resolve(result);
+                task.resolve(msg.result);
             }
             this.completeTask(task, worker);
             this.consecutiveRespawns = 0;

package/mlx-embed-server/server.py

@@ -13,6 +13,7 @@ GPU operations.
 import asyncio
 import logging
 import os
+import re
 import signal
 import socket
 import time
@@ -90,12 +91,22 @@ def is_port_in_use(port: int) -> bool:
         return s.connect_ex(("127.0.0.1", port)) == 0
 
 
+# JSON can carry lone UTF-16 surrogates (a client slicing mid-codepoint);
+# the HF fast tokenizer rejects them with a TypeError that fails the whole
+# batch. Replace with U+FFFD so one bad text can't take down a batch.
+_LONE_SURROGATE = re.compile("[\ud800-\udfff]")
+
+
 def embed_texts(texts: list[str]) -> mx.array:
     """Tokenize, forward pass, L2 normalize.
 
     mlx_embeddings model already does mean pooling internally —
     last_hidden_state is (batch, dim), not (batch, seq, dim).
     """
+    texts = [
+        _LONE_SURROGATE.sub("�", t) if _LONE_SURROGATE.search(t) else t
+        for t in texts
+    ]
     encoded = tokenizer(
         texts, padding=True, truncation=True, max_length=256, return_tensors="np"
     )

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.26.4",
+  "version": "0.26.6",
   "author": "Robert Owens <78518764+reowens@users.noreply.github.com>",
   "homepage": "https://github.com/reowens/grepmax",
   "bugs": {
@@ -79,8 +79,8 @@
   "dependencies": {
     "@clack/prompts": "^1.6.0",
     "@huggingface/transformers": "^4.2.0",
-    "@lancedb/lancedb": "^0.30.0",
-    "@modelcontextprotocol/sdk": "^1.29.0",
+    "@lancedb/lancedb": "0.31.0",
+    "@modelcontextprotocol/sdk": "^1.30.0",
     "@parcel/watcher": "^2.5.6",
     "apache-arrow": "^18.1.0",
     "chalk": "^5.6.2",
@@ -99,11 +99,6 @@
     "web-tree-sitter": "^0.26.9",
     "zod": "^4.4.3"
   },
-  "pnpm": {
-    "overrides": {
-      "mathjs": "15.2.0"
-    }
-  },
   "devDependencies": {
     "@biomejs/biome": "2.4.10",
     "@types/node": "^26.0.1",

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

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.26.4",
+  "version": "0.26.6",
   "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
   "author": {
     "name": "Robert Owens",