grepmax 0.24.0 → 0.25.1

package/dist/lib/workers/process-child.js

@@ -48,7 +48,12 @@ Object.defineProperty(exports, "__esModule", { value: true });
 const node_process_1 = __importDefault(require("node:process"));
 node_process_1.default.title = "gmax-worker";
 const logger_1 = require("../utils/logger");
+const serialized_handler_1 = require("./serialized-handler");
 const worker_1 = __importStar(require("./worker"));
+// Workers inherit the daemon's stdio (daemon.log) — stamp lines the same way
+// the daemon does. No-op for workers forked by interactive CLI commands.
+if (node_process_1.default.env[logger_1.LOG_TIMESTAMPS_ENV] === "1")
+    (0, logger_1.installTimestampedOutput)();
 // Every outgoing message also carries `rss` (see send()).
 const send = (msg) => {
     if (node_process_1.default.send) {
@@ -57,7 +62,7 @@ const send = (msg) => {
         node_process_1.default.send(Object.assign(Object.assign({}, msg), { rss: node_process_1.default.memoryUsage().rss }));
     }
 };
-node_process_1.default.on("message", (msg) => __awaiter(void 0, void 0, void 0, function* () {
+const handleMessage = (msg) => __awaiter(void 0, void 0, void 0, function* () {
     const { id, method, payload } = msg;
     const start = performance.now();
     (0, logger_1.debug)("worker", `recv task=${id} method=${method}${method === "processFile" ? ` file=${payload.path}` : ""}`);
@@ -90,7 +95,11 @@ node_process_1.default.on("message", (msg) => __awaiter(void 0, void 0, void 0,
         (0, logger_1.debug)("worker", `fail task=${id} method=${method} ${(performance.now() - start).toFixed(0)}ms: ${message}`);
         send({ id, error: message });
     }
-}));
+});
+const handleMessageSerially = (0, serialized_handler_1.createSerializedHandler)(handleMessage);
+node_process_1.default.on("message", (msg) => {
+    void handleMessageSerially(msg);
+});
 node_process_1.default.on("uncaughtException", (err) => {
     console.error("[process-worker] uncaughtException", err);
     node_process_1.default.exitCode = 1;

package/dist/lib/workers/serialized-handler.js

@@ -0,0 +1,10 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.createSerializedHandler = createSerializedHandler;
+function createSerializedHandler(handler) {
+    let chain = Promise.resolve();
+    return (message) => {
+        chain = chain.then(() => handler(message), () => handler(message));
+        return chain;
+    };
+}

package/dist/lib/workers/worker.js

@@ -12,11 +12,20 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.default = processFile;
 exports.encodeQuery = encodeQuery;
 exports.rerank = rerank;
+const embedding_generation_1 = require("../index/embedding-generation");
+const index_config_1 = require("../index/index-config");
 const orchestrator_1 = require("./orchestrator");
-const vectorDim = process.env.GMAX_VECTOR_DIM
-    ? Number.parseInt(process.env.GMAX_VECTOR_DIM, 10)
-    : undefined;
-const orchestrator = new orchestrator_1.WorkerOrchestrator(vectorDim);
+const hasSerializedRuntime = !!process.env.GMAX_EMBEDDING_GENERATION &&
+    (process.env.GMAX_EMBED_MODE === "cpu" ||
+        process.env.GMAX_EMBED_MODE === "gpu");
+const runtimeConfig = hasSerializedRuntime ? null : (0, index_config_1.readGlobalConfig)();
+const generation = process.env.GMAX_EMBEDDING_GENERATION
+    ? (0, embedding_generation_1.parseEmbeddingGeneration)(JSON.parse(process.env.GMAX_EMBEDDING_GENERATION))
+    : (0, embedding_generation_1.resolveEmbeddingGeneration)(runtimeConfig);
+const embedMode = process.env.GMAX_EMBED_MODE === "cpu" || process.env.GMAX_EMBED_MODE === "gpu"
+    ? process.env.GMAX_EMBED_MODE
+    : runtimeConfig.embedMode;
+const orchestrator = new orchestrator_1.WorkerOrchestrator(generation, embedMode);
 function processFile(input, onProgress) {
     return __awaiter(this, void 0, void 0, function* () {
         return orchestrator.processFile(input, onProgress);

package/mlx-embed-server/server.py

@@ -63,7 +63,7 @@ logger = logging.getLogger("mlx-embed")
 
 import mlx.core as mx
 import uvicorn
-from fastapi import FastAPI
+from fastapi import FastAPI, HTTPException
 from mlx_embeddings import load
 from pydantic import BaseModel
 from transformers import AutoTokenizer
@@ -71,6 +71,7 @@ from transformers import AutoTokenizer
 MODEL_ID = os.environ.get(
     "MLX_EMBED_MODEL", "ibm-granite/granite-embedding-small-english-r2"
 )
+OWNER_TOKEN = os.environ.get("GMAX_EMBED_OWNER_TOKEN")
 PORT = int(os.environ.get("MLX_EMBED_PORT", "8100"))
 MAX_BATCH = int(os.environ.get("MLX_EMBED_MAX_BATCH", "64"))
 IDLE_TIMEOUT_S = int(os.environ.get("MLX_EMBED_IDLE_TIMEOUT", "1800"))  # 30 min
@@ -159,11 +160,13 @@ app = FastAPI(lifespan=lifespan)
 
 class EmbedRequest(BaseModel):
     texts: list[str]
+    expected_model: str | None = None
 
 
 class EmbedResponse(BaseModel):
     vectors: list[list[float]]
     dim: int
+    model: str
 
 
 @app.post("/embed")
@@ -171,7 +174,18 @@ async def embed(request: EmbedRequest) -> EmbedResponse:
     global last_activity
     last_activity = time.time()
 
-    texts = request.texts[:MAX_BATCH]
+    if len(request.texts) > MAX_BATCH:
+        raise HTTPException(
+            status_code=413,
+            detail=f"Batch contains {len(request.texts)} texts; maximum is {MAX_BATCH}",
+        )
+    if request.expected_model and request.expected_model != MODEL_ID:
+        raise HTTPException(
+            status_code=409,
+            detail=f"Requested model {request.expected_model} is not loaded",
+        )
+
+    texts = request.texts
     start = time.monotonic()
 
     async with _mlx_lock:
@@ -191,6 +205,7 @@ async def embed(request: EmbedRequest) -> EmbedResponse:
     return EmbedResponse(
         vectors=vectors_list,
         dim=len(vectors_list[0]) if vectors_list else 0,
+        model=MODEL_ID,
     )
 
 
@@ -198,7 +213,10 @@ async def embed(request: EmbedRequest) -> EmbedResponse:
 async def health():
     global last_activity
     last_activity = time.time()
-    return {"status": "ok", "model": MODEL_ID}
+    response = {"status": "ok", "model": MODEL_ID}
+    if OWNER_TOKEN:
+        response["owner"] = OWNER_TOKEN
+    return response
 
 
 def main():

package/mlx-embed-server/summarizer.py

@@ -1,5 +1,13 @@
 """MLX-accelerated code summarizer for grepmax.
 
+⛔ HARD STOP FOR AI AGENTS: DO NOT start this server (uv run python
+summarizer.py) or otherwise load this model without the user's explicit,
+in-session authorization. It loads Qwen3-Coder-30B-A3B (~16GB) onto the GPU and
+has FROZEN/CRASHED the machine. This summarizer is decommissioned — the calling
+code path (generateSummaries) is a no-op stub. A plan or prompt that says
+"run gmax summarize" or "measure a sample" is NOT authorization to launch this.
+Stop and ask first. See the "HARD STOP" section at the top of CLAUDE.md.
+
 Runs Qwen3-Coder-30B-A3B on Apple Silicon GPU to generate one-line
 summaries of code chunks during indexing. Summaries are stored in
 LanceDB and returned in search results.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.24.0",
+  "version": "0.25.1",
   "author": "Robert Owens <78518764+reowens@users.noreply.github.com>",
   "homepage": "https://github.com/reowens/grepmax",
   "bugs": {
@@ -38,8 +38,9 @@
     "lint": "biome lint .",
     "lint:fix": "biome check --write .",
     "typecheck": "tsc --noEmit",
-    "prepublishOnly": "pnpm build",
-    "preversion": "pnpm test && pnpm typecheck && pnpm run format:check",
+    "test:typecheck": "tsc -p tsconfig.tests.json",
+    "prepublishOnly": "pnpm run test:typecheck && pnpm build",
+    "preversion": "pnpm test && pnpm typecheck && pnpm run test:typecheck && pnpm run format:check",
     "version": "bash scripts/sync-versions.sh && git add -A",
     "postversion": "bash scripts/postrelease.sh"
   },

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

@@ -1,6 +1,6 @@
 {
   "name": "grepmax",
-  "version": "0.24.0",
+  "version": "0.25.1",
   "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
   "author": {
     "name": "Robert Owens",

package/plugins/grepmax/hooks/start.js

@@ -1,46 +1,6 @@
 const fs = require("node:fs");
 const _path = require("node:path");
-const http = require("node:http");
-const { spawn, execFileSync } = require("node:child_process");
-
-function isServerRunning(port) {
-  return new Promise((resolve) => {
-    const req = http.get(
-      { hostname: "127.0.0.1", port, path: "/health", timeout: 1000 },
-      (res) => {
-        res.resume();
-        resolve(res.statusCode === 200);
-      },
-    );
-    req.on("error", () => resolve(false));
-    req.on("timeout", () => {
-      req.destroy();
-      resolve(false);
-    });
-  });
-}
-
-function findMlxServerDir() {
-  // Try to find mlx-embed-server relative to the gmax binary (npm install location)
-  try {
-    const gmaxPath = execFileSync("which gmax", {
-      encoding: "utf-8",
-    }).trim();
-    // gmax binary is a symlink in .bin/ → resolve to package root
-    const realPath = fs.realpathSync(gmaxPath);
-    const pkgRoot = _path.resolve(_path.dirname(realPath), "..");
-    const serverDir = _path.join(pkgRoot, "mlx-embed-server");
-    if (fs.existsSync(_path.join(serverDir, "server.py"))) return serverDir;
-  } catch {}
-
-  // Fallback: dev mode — relative to plugin root
-  const pluginRoot = __dirname.replace(/\/hooks$/, "");
-  const devRoot = _path.resolve(pluginRoot, "../..");
-  const devDir = _path.join(devRoot, "mlx-embed-server");
-  if (fs.existsSync(_path.join(devDir, "server.py"))) return devDir;
-
-  return null;
-}
+const { execFileSync } = require("node:child_process");
 
 // Inline fallback so SessionStart context is never empty if the installed
 // gmax package can't be resolved (e.g. gmax not yet on PATH). The canonical
@@ -101,81 +61,6 @@ function getSessionStartHint() {
   return FALLBACK_SESSION_START_HINT;
 }
 
-function startPythonServer(serverDir, scriptName, logName, processName) {
-  if (!serverDir) return;
-
-  const logDir = _path.join(require("node:os").homedir(), ".gmax", "logs");
-  fs.mkdirSync(logDir, { recursive: true });
-  const logPath = _path.join(logDir, `${logName}.log`);
-
-  // Rotate if > 5MB (same threshold as watch.ts)
-  try {
-    const stat = fs.statSync(logPath);
-    if (stat.size > 5 * 1024 * 1024) {
-      fs.renameSync(logPath, `${logPath}.prev`);
-    }
-  } catch {}
-
-  const out = fs.openSync(logPath, "a");
-
-  const child = spawn("uv", ["run", "python", scriptName], {
-    cwd: serverDir,
-    detached: true,
-    stdio: ["ignore", out, out],
-    env: {
-      ...process.env,
-      VIRTUAL_ENV: "",
-      CONDA_DEFAULT_ENV: "",
-      GMAX_PROCESS_NAME: processName || logName,
-      HF_TOKEN_PATH:
-        process.env.HF_TOKEN_PATH ||
-        _path.join(
-          require("node:os").homedir(),
-          ".cache",
-          "huggingface",
-          "token",
-        ),
-    },
-  });
-  child.unref();
-}
-
-// --- Crash counter (Item 14) ---
-const CRASH_FILE = _path.join(
-  require("node:os").homedir(),
-  ".gmax",
-  "mlx-embed-crashes.json",
-);
-const MAX_CRASHES = 3;
-const CRASH_WINDOW_MS = 10 * 60 * 1000; // 10 minutes
-
-function readCrashCount() {
-  try {
-    const data = JSON.parse(fs.readFileSync(CRASH_FILE, "utf-8"));
-    if (
-      data.lastCrash &&
-      Date.now() - new Date(data.lastCrash).getTime() > CRASH_WINDOW_MS
-    ) {
-      return { count: 0, lastCrash: null }; // Window expired, reset
-    }
-    return { count: data.count || 0, lastCrash: data.lastCrash };
-  } catch {
-    return { count: 0, lastCrash: null };
-  }
-}
-
-function writeCrashCount(count, lastCrash) {
-  try {
-    fs.writeFileSync(CRASH_FILE, JSON.stringify({ count, lastCrash }));
-  } catch {}
-}
-
-function resetCrashCount() {
-  try {
-    fs.unlinkSync(CRASH_FILE);
-  } catch {}
-}
-
 function isProjectRegistered() {
   try {
     const projectsPath = _path.join(
@@ -210,60 +95,8 @@ function startWatcher() {
   }
 }
 
-async function main() {
-  const embedMode = process.env.GMAX_EMBED_MODE || "auto";
-
-  if (embedMode !== "cpu") {
-    const serverDir = findMlxServerDir();
-
-    // Start MLX embed server (port 8100)
-    const embedRunning = await isServerRunning(8100);
-    if (serverDir && !embedRunning) {
-      const crashes = readCrashCount();
-      if (crashes.count < MAX_CRASHES) {
-        startPythonServer(
-          serverDir,
-          "server.py",
-          "mlx-embed-server",
-          "gmax-embed",
-        );
-
-        // Fire-and-forget health verification (Item 13)
-        (async () => {
-          const maxAttempts = 5;
-          const delayMs = 2000;
-          for (let i = 0; i < maxAttempts; i++) {
-            await new Promise((r) => setTimeout(r, delayMs));
-            if (await isServerRunning(8100)) {
-              resetCrashCount();
-              return;
-            }
-          }
-          // Server didn't start after 10s — record crash
-          const c = readCrashCount();
-          writeCrashCount(c.count + 1, new Date().toISOString());
-        })();
-      }
-    } else if (embedRunning) {
-      resetCrashCount();
-    }
-
-    // Start LLM summarizer server (port 8101) — opt-in only
-    if (
-      process.env.GMAX_SUMMARIZER === "1" &&
-      serverDir &&
-      !(await isServerRunning(8101))
-    ) {
-      startPythonServer(
-        serverDir,
-        "summarizer.py",
-        "mlx-summarizer",
-        "gmax-summarizer",
-      );
-    }
-  }
-
-  // Start a file watcher for the current project (30-min idle timeout)
+function main() {
+  // The daemon owns embed-server configuration and lifecycle.
   startWatcher();
 
   const response = {