grepmax 0.26.58 → 0.26.60

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
       "name": "grepmax",
       "source": "./plugins/grepmax",
       "description": "Semantic code search for Claude Code. Automatically indexes your project and provides intelligent search capabilities.",
-      "version": "0.26.58",
+      "version": "0.26.60",
       "author": {
         "name": "Robert Owens",
         "email": "robowens@me.com"

package/README.md

@@ -453,6 +453,7 @@ fixtures/
 | --- | --- | --- |
 | `GMAX_EMBED_MODE` | Force `cpu` or `gpu` | Auto-detect |
 | `GMAX_WORKER_THREADS` | Worker processes for embedding; overrides `gmax config --worker-threads` | `min(cores, 4, max(2, floor(cores/2)))` |
+| `GMAX_RESOURCE_BUDGET_MB` | macOS aggregate admission budget across clients, stores, workers and models; accepts 256–6144 MiB | `6144` |
 | `GMAX_WORKER_RSS_RECYCLE_MB` | Recycle workers that remain above this RSS; `0` disables the check | `1536` |
 | `GMAX_DEBUG` | Debug logging | Off |
 | `GMAX_RERANK` | Force ColBERT rerank (`1`); concentrated candidates can also enable it automatically ([why](docs/known-limitations.md)) | Off |
@@ -474,7 +475,7 @@ gmax watch restart            # Restart daemon
 
 `gmax doctor` and `gmax status` report persistent safety stops and daemon startup quarantine. `doctor` exits with code 2 when startup is blocked; `status --json` includes `daemon.startupBlockedReason`. Installing an update preserves these stops. macOS kernel-zone probes use `zprint -L` to omit the wired-memory report and kernel symbolication while retaining the same zone counters and pressure thresholds.
 
-When heavy work is paused, the daemon keeps metadata, bounded symbol/file lookups, stored skeletons by exact file path, and keyword search available. Search uses the existing FTS index and excludes unindexed rows, loads no embedding model, returns at most 20 results, and reports keyword-only scores and retained-index freshness. Native queries have a two-second deadline and run one at a time with 48 MiB of combined Lance cache capacity. Indexing, watchers, vector/graph scans, and model launches remain paused. Missing FTS indexes return an error; paused search never rebuilds an index or falls back to a corpus scan.
+When heavy work is paused, the daemon keeps metadata, bounded symbol/file lookups, stored skeletons by exact file path, and keyword search available. Search uses the existing FTS index and excludes unindexed rows, loads no embedding model, returns at most 20 results, and reports keyword-only scores and retained-index freshness. CLI agent output preserves the keyword-only warning, and daemon reads skip client-side model setup. Native queries have a two-second deadline and run one at a time with 48 MiB of combined Lance cache capacity. Indexing, watchers, vector/graph scans, and model launches remain paused. Missing FTS indexes return an error; paused search never rebuilds an index or falls back to a corpus scan.
 
 To serve bounded reads under an existing startup quarantine, run `gmax watch --daemon --read-only -b`. This preserves the safety markers. `status --json` reports `daemon.service.mode` and its reason; `doctor` reports paused availability and skips native corpus diagnostics. Pausing is sticky for that daemon process. Full service requires a deliberate restart after host recovery and quarantine review. Known critical OS or kernel pressure still stops even the read-only service.
 
@@ -552,3 +553,7 @@ grepmax is built upon the foundation of [mgrep](https://github.com/mixedbread-ai
 ## License
 
 Licensed under the Apache License, Version 2.0. See [LICENSE](LICENSE).
+
+On macOS, heavy work uses a shared resource ledger and a bounded batch footprint probe covering every gmax process and its children. Native stores, workers and model launches reserve headroom before allocating it. Warning or unknown measurements pause heavy work; known critical pressure persists a stop. The 6 GiB admission budget is a conservative scheduling threshold, not an OS-enforced memory cap. It may be lowered with `GMAX_RESOURCE_BUDGET_MB`. Existing safety markers remain in place after upgrades.
+
+MCP sessions started before this resource policy must reconnect after upgrading before heavy work can resume. Keyword search through the bounded read service stays available while those sessions are running. Cached resource snapshots in `gmax status --json` distinguish aggregate physical footprint from daemon RSS.

package/dist/commands/doctor.js

@@ -358,16 +358,6 @@ exports.doctor = new commander_1.Command("doctor")
     }
     if (!opts.agent)
         console.log("gmax Doctor\n");
-    const startupBlockedReason = (0, autostart_1.daemonStartDeniedReason)();
-    if (startupBlockedReason !== null) {
-        const reason = startupBlockedReason.replace(/[\r\n\t]/g, " ");
-        console.log(opts.agent
-            ? `daemon_startup\tblocked=true\treason=${reason}`
-            : `FAIL  Daemon startup blocked: ${reason}`);
-        // Inspect the retained index, but do not report a healthy installation
-        // while the persistent host stop or autostart quarantine prevents use.
-        process.exitCode = 2;
-    }
     // Metadata-only diagnostics while paused: never open a second native
     // reader or scan the retained corpus on a pressured host.
     const { sendDaemonCommand } = yield Promise.resolve().then(() => __importStar(require("../lib/utils/daemon-client")));
@@ -384,6 +374,16 @@ exports.doctor = new commander_1.Command("doctor")
         process.exitCode = 2;
         return;
     }
+    const startupBlockedReason = (0, autostart_1.daemonStartDeniedReason)();
+    if (startupBlockedReason !== null) {
+        const reason = startupBlockedReason.replace(/[\r\n\t]/g, " ");
+        console.log(opts.agent
+            ? `daemon_startup\tblocked=true\treason=${reason}`
+            : `FAIL  Daemon startup blocked: ${reason}`);
+        // Inspect the retained index, but do not report a healthy installation
+        // while the persistent host stop or autostart quarantine prevents use.
+        process.exitCode = 2;
+    }
     const root = config_1.PATHS.globalRoot;
     const models = config_1.PATHS.models;
     const grammars = config_1.PATHS.grammars;

package/dist/commands/mcp.js

@@ -101,6 +101,7 @@ const mcp_execution_1 = require("../lib/utils/mcp-execution");
 const path_containment_1 = require("../lib/utils/path-containment");
 const project_registry_1 = require("../lib/utils/project-registry");
 const project_root_1 = require("../lib/utils/project-root");
+const resource_budget_1 = require("../lib/utils/resource-budget");
 const scope_filter_1 = require("../lib/utils/scope-filter");
 const store_access_1 = require("../lib/utils/store-access");
 const store_context_1 = require("../lib/utils/store-context");
@@ -368,6 +369,18 @@ exports.mcp = new commander_1.Command("mcp")
     .description("Start MCP server (stdio, auto-started by plugins)")
     .action((_optsArg, _cmd) => __awaiter(void 0, void 0, void 0, function* () {
     process.title = "gmax-mcp";
+    try {
+        const client = resource_budget_1.resourceBudget.registerClient();
+        process.once("exit", () => {
+            try {
+                client.release();
+            }
+            catch (_a) { }
+        });
+    }
+    catch (error) {
+        console.error(`gmax: resource coordination unavailable; heavy work paused (${error instanceof Error ? error.message : "unknown"})`);
+    }
     // --- Lifecycle ---
     // No per-session VectorDB. Every read tool goes through the daemon's read
     // verbs, and the only store this process may open is the short-lived one

package/dist/commands/search-output.js

@@ -210,9 +210,11 @@ function renderSearchOutput(params) {
         var _a, _b, _c;
         var _d, _e, _f, _g, _h, _j;
         const { searchResult, options, minScore, crossProject, pattern, effectiveRoot, projectRoot, vectorDb, precomputedSkeletons, precomputedGraph, indexState, } = params;
-        if (!options.agent && ((_a = searchResult.warnings) === null || _a === void 0 ? void 0 : _a.length)) {
+        if ((_a = searchResult.warnings) === null || _a === void 0 ? void 0 : _a.length) {
             for (const w of searchResult.warnings) {
-                console.warn(`Warning: ${w}`);
+                console.warn(options.agent
+                    ? `warn\t${w.replace(/[\r\n\t]/g, " ")}`
+                    : `Warning: ${w}`);
             }
         }
         // Partial-index signal (Phase 6): when the index is mid-catchup, results

package/dist/commands/search-run.js

@@ -53,6 +53,7 @@ const sync_helpers_1 = require("../lib/index/sync-helpers");
 const syncer_1 = require("../lib/index/syncer");
 const per_file_1 = require("../lib/search/per-file");
 const searcher_1 = require("../lib/search/searcher");
+const setup_helpers_1 = require("../lib/setup/setup-helpers");
 const vector_db_1 = require("../lib/store/vector-db");
 const lock_1 = require("../lib/utils/lock");
 const project_registry_1 = require("../lib/utils/project-registry");
@@ -160,6 +161,8 @@ function runSearch(params) {
             // In-process fallback: open VectorDB, ensure index, run Searcher.
             // Only entered when the daemon path didn't produce results.
             if (!searchResult) {
+                // A daemon read needs no client-side model download or loading.
+                yield (0, setup_helpers_1.ensureSetup)();
                 const vectorDb = new vector_db_1.VectorDB(paths.lancedbDir);
                 openedDb = vectorDb;
                 // Check for active indexing lock and warn if present

package/dist/commands/search.js

@@ -46,7 +46,6 @@ exports.search = void 0;
 const path = __importStar(require("node:path"));
 const commander_1 = require("commander");
 const per_file_1 = require("../lib/search/per-file");
-const setup_helpers_1 = require("../lib/setup/setup-helpers");
 const cross_project_1 = require("../lib/utils/cross-project");
 const exit_1 = require("../lib/utils/exit");
 const path_containment_1 = require("../lib/utils/path-containment");
@@ -238,7 +237,6 @@ Examples:
             return;
     }
     try {
-        yield (0, setup_helpers_1.ensureSetup)();
         let checkRoot = cwdProjectRoot;
         if (options.root) {
             const resolved = resolvedOptionRoot !== undefined

package/dist/lib/daemon/daemon.js

@@ -67,6 +67,7 @@ const autostart_1 = require("../utils/autostart");
 const blocked_roots_1 = require("../utils/blocked-roots");
 const daemon_client_1 = require("../utils/daemon-client");
 const daemon_launcher_1 = require("../utils/daemon-launcher");
+const host_resource_1 = require("../utils/host-resource");
 const kernel_zone_1 = require("../utils/kernel-zone");
 const keyed_mutex_1 = require("../utils/keyed-mutex");
 const log_rotate_1 = require("../utils/log-rotate");
@@ -76,6 +77,7 @@ const process_1 = require("../utils/process");
 const process_footprint_1 = require("../utils/process-footprint");
 const project_registry_1 = require("../utils/project-registry");
 const query_timeout_1 = require("../utils/query-timeout");
+const resource_budget_1 = require("../utils/resource-budget");
 const resource_snapshot_1 = require("../utils/resource-snapshot");
 const root_availability_1 = require("../utils/root-availability");
 const safety_latch_1 = require("../utils/safety-latch");
@@ -165,6 +167,7 @@ class Daemon {
         // resources must be gated on this to avoid hitting null stores mid-startup.
         this.ready = false;
         this.pausedReason = null;
+        this.lastHostResources = null;
         this.pausePromise = null;
         this.pausedReadGate = new async_semaphore_1.AsyncSemaphore(1);
         this.processManager = new process_manager_1.ProcessManager({
@@ -246,7 +249,7 @@ class Daemon {
     }
     start() {
         return __awaiter(this, arguments, void 0, function* (options = {}) {
-            var _a, _b;
+            var _a, _b, _c;
             const denied = (0, autostart_1.daemonStartDeniedReason)();
             if (denied !== null && !options.readOnly)
                 throw new Error(`gmax: ${denied}; preserving containment`);
@@ -260,6 +263,19 @@ class Daemon {
                 this.pausedReason = (_a = denied !== null && denied !== void 0 ? denied : pressure === null || pressure === void 0 ? void 0 : pressure.reason) !== null && _a !== void 0 ? _a : "read-only service requested";
             }
             process.title = "gmax-daemon";
+            if (this.pausedReason === null) {
+                try {
+                    this.lastHostResources = resource_budget_1.resourceBudget.check();
+                }
+                catch (error) {
+                    if (error instanceof resource_budget_1.ResourceAdmissionError && error.critical)
+                        throw error;
+                    this.pausedReason =
+                        error instanceof Error
+                            ? error.message
+                            : "aggregate resource admission unavailable";
+                }
+            }
             // Read verbs register here rather than at read-verbs.ts module scope: the
             // registry is asserted on exactly in tests/ipc-read-verbs.test.ts, and a
             // module-scope registration would make it non-empty on import. Register
@@ -311,7 +327,7 @@ class Daemon {
             try {
                 fs.unlinkSync(config_1.PATHS.daemonSocket);
             }
-            catch (_c) { }
+            catch (_d) { }
             this.server = net.createServer((conn) => {
                 (0, logger_1.debug)("daemon", "client connected");
                 this.connections.add(conn);
@@ -405,9 +421,13 @@ class Daemon {
                     }
                     catch (error) {
                         const freshPressure = this.hostPressureDeniedReason();
-                        if (freshPressure === null || freshPressure.critical)
+                        if (freshPressure === null || freshPressure === void 0 ? void 0 : freshPressure.critical)
+                            throw error;
+                        if (freshPressure === null &&
+                            !(error instanceof resource_budget_1.ResourceAdmissionError) &&
+                            (error === null || error === void 0 ? void 0 : error.code) !== "HOST_SAFETY")
                             throw error;
-                        this.pauseWork(freshPressure.reason);
+                        this.pauseWork((_b = freshPressure === null || freshPressure === void 0 ? void 0 : freshPressure.reason) !== null && _b !== void 0 ? _b : error.message);
                         yield this.pausePromise;
                     }
                     if (this.pausedReason === null) {
@@ -427,7 +447,7 @@ class Daemon {
             // 6. LLM server manager (constructed, not started — starts on first request)
             this.llmServer = new server_1.LlmServer();
             // 6b. MLX embed server — start if GPU mode is active
-            const globalConfig = (_b = this.activeConfig) !== null && _b !== void 0 ? _b : (0, index_config_1.readGlobalConfig)();
+            const globalConfig = (_c = this.activeConfig) !== null && _c !== void 0 ? _c : (0, index_config_1.readGlobalConfig)();
             const isAppleSilicon = process.arch === "arm64" && process.platform === "darwin";
             if (this.pausedReason === null &&
                 isAppleSilicon &&
@@ -510,6 +530,8 @@ class Daemon {
                         this.logResourceSnapshot("paused-heartbeat", null);
                     return;
                 }
+                if (!this.checkAggregateResources())
+                    return;
                 // 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
@@ -758,6 +780,7 @@ class Daemon {
         });
     }
     assertHeavyOperationAdmission(name) {
+        var _a;
         const denied = (0, autostart_1.daemonStartDeniedReason)();
         if (denied !== null)
             throw new Error(`${name} refused: ${denied}`);
@@ -765,6 +788,8 @@ class Daemon {
         if (!this.checkKernelZonePressure()) {
             throw new Error(`${name} refused: host pressure is unsafe or unknown${this.pausedReason ? `: ${this.pausedReason}` : ""}`);
         }
+        if (!this.checkAggregateResources())
+            throw new Error(`${name} refused: ${(_a = this.pausedReason) !== null && _a !== void 0 ? _a : "aggregate resources unavailable"}`);
         // Another process can quarantine the machine while the probes run.
         const finalDenied = (0, autostart_1.daemonStartDeniedReason)();
         if (finalDenied !== null)
@@ -2037,6 +2062,8 @@ class Daemon {
         // Both ordinary probes and deferred-recycle ticks check safety first.
         if (!this.checkKernelZonePressure())
             return;
+        if (sampleFootprint && !this.checkAggregateResources())
+            return;
         this.maybeRecycle(sampleFootprint);
         if (this.shuttingDown || this.recycling || !sampleFootprint)
             return;
@@ -2048,6 +2075,8 @@ class Daemon {
     logResourceSnapshot(reason, footprintMb) {
         var _a, _b;
         var _c, _d;
+        if (reason.startsWith("paused-"))
+            this.lastHostResources = (0, host_resource_1.sampleHostResources)();
         let lanceCacheBytes = null;
         try {
             lanceCacheBytes = (_c = (_a = this.vectorDb) === null || _a === void 0 ? void 0 : _a.cacheSizeBytes()) !== null && _c !== void 0 ? _c : null;
@@ -2061,6 +2090,7 @@ class Daemon {
             pendingFiles: [...this.processors.values()].reduce((sum, processor) => sum + processor.progress.pendingFiles, 0),
             operations: this.operations.activeCount,
             maintenance: (_d = (_b = this.vectorDb) === null || _b === void 0 ? void 0 : _b.isMaintenanceActive()) !== null && _d !== void 0 ? _d : false,
+            host: this.lastHostResources,
         });
         console.log(`[daemon] Resource snapshot: ${JSON.stringify(this.lastResourceSnapshot)}`);
     }
@@ -2165,6 +2195,24 @@ class Daemon {
         }
         return this.pausedReason === null;
     }
+    checkAggregateResources() {
+        if (this.shuttingDown || this.recycling || this.pausedReason !== null)
+            return false;
+        try {
+            this.lastHostResources = resource_budget_1.resourceBudget.check();
+            return true;
+        }
+        catch (error) {
+            const reason = error instanceof Error
+                ? error.message
+                : "aggregate resource admission unavailable";
+            if (error instanceof resource_budget_1.ResourceAdmissionError && error.critical)
+                this.stopForSafety(reason);
+            else
+                this.pauseWork(reason);
+            return false;
+        }
+    }
     /** Unknown/warning pauses resource expansion; only known critical shuts down. */
     hostPressureDeniedReason() {
         if (process.platform !== "darwin")

package/dist/lib/daemon/mlx-server-manager.js

@@ -52,6 +52,7 @@ const config_1 = require("../../config");
 const log_rotate_1 = require("../utils/log-rotate");
 const mlx_hf_cache_1 = require("../utils/mlx-hf-cache");
 const process_1 = require("../utils/process");
+const resource_budget_1 = require("../utils/resource-budget");
 const STARTUP_ATTEMPTS = 30;
 const STARTUP_POLL_MS = 1000;
 const MAX_HEALTH_BODY_BYTES = 16 * 1024;
@@ -151,11 +152,18 @@ class MlxServerManager {
     terminateOwned(server) {
         return __awaiter(this, void 0, void 0, function* () {
             var _a;
-            const terminate = (_a = this.deps.terminateGroup) !== null && _a !== void 0 ? _a : process_1.terminateProcessGroup;
+            var _b;
+            const terminate = (_b = this.deps.terminateGroup) !== null && _b !== void 0 ? _b : process_1.terminateProcessGroup;
             const stopped = yield terminate(server.pid);
             if (!stopped) {
                 console.error(`[daemon] Failed to fully stop owned MLX process group ${server.pid}`);
             }
+            else {
+                try {
+                    (_a = server.reservation) === null || _a === void 0 ? void 0 : _a.release();
+                }
+                catch (_c) { }
+            }
             server.child.removeAllListeners("error");
             server.child.removeAllListeners("exit");
             if (this.stable === server)
@@ -204,7 +212,7 @@ class MlxServerManager {
     ensureInternal(model) {
         return __awaiter(this, void 0, void 0, function* () {
             var _a, _b, _c, _d, _e;
-            var _f, _g, _h, _j, _k;
+            var _f, _g, _h, _j, _k, _l;
             this.phase = "probing";
             const health = yield this.probeHealth(this.port);
             if (this.shouldStop())
@@ -223,6 +231,7 @@ class MlxServerManager {
                     this.stable.lastHealthyAt = Date.now();
                     return;
                 }
+                resource_budget_1.resourceBudget.check(this.getPortPid(this.port));
                 if (this.stable.kind === "owned")
                     yield this.terminateOwned(this.stable);
                 this.stable = { kind: "adopted", model, lastHealthyAt: Date.now() };
@@ -278,17 +287,41 @@ class MlxServerManager {
             const logFd = openLog(path.join(config_1.PATHS.logsDir, "mlx-embed-server.log"));
             const env = Object.assign(Object.assign({}, process.env), { MLX_EMBED_MODEL: model, GMAX_EMBED_OWNER_TOKEN: owner, HF_HOME: (0, mlx_hf_cache_1.resolveMlxHfHome)(model) });
             this.phase = "starting";
-            let child;
+            let child = null;
+            let reservation;
             try {
+                reservation = resource_budget_1.resourceBudget.reserve(resource_budget_1.EMBEDDING_RESOURCE_RESERVE_MB, "embedding");
                 child = ((_j = this.deps.spawn) !== null && _j !== void 0 ? _j : node_child_process_1.spawn)("uv", ["run", "python", "server.py"], {
                     cwd: serverDir,
                     detached: true,
                     stdio: ["ignore", logFd, logFd],
                     env,
                 });
+                if (!child.pid) {
+                    reservation.release();
+                    throw new Error("spawned MLX process has no PID");
+                }
+                try {
+                    reservation.attach(child.pid, true);
+                }
+                catch (error) {
+                    const stopped = yield ((_k = this.deps.terminateGroup) !== null && _k !== void 0 ? _k : process_1.terminateProcessGroup)(child.pid);
+                    if (stopped)
+                        reservation.release();
+                    throw error;
+                }
             }
             catch (error) {
+                // If no child was created, there is no native group to keep charged.
+                if (!child) {
+                    try {
+                        reservation === null || reservation === void 0 ? void 0 : reservation.release();
+                    }
+                    catch (_m) { }
+                }
                 this.lastError = error instanceof Error ? error.message : String(error);
+                if (error instanceof resource_budget_1.ResourceAdmissionError)
+                    this.enabled = false;
                 this.phase = "failed";
                 console.error(`[daemon] MLX embed server failed to spawn: ${this.lastError} — falling back to CPU embeddings`);
                 return;
@@ -297,9 +330,9 @@ class MlxServerManager {
                 try {
                     closeFd(logFd);
                 }
-                catch (_l) { }
+                catch (_o) { }
             }
-            if (!child.pid) {
+            if (!(child === null || child === void 0 ? void 0 : child.pid)) {
                 this.lastError = "spawned MLX process has no PID";
                 this.phase = "failed";
                 return;
@@ -312,6 +345,7 @@ class MlxServerManager {
                 model,
                 owner,
                 lastHealthyAt: 0,
+                reservation,
             };
             const startup = { error: null, exited: false };
             const onError = (error) => {
@@ -338,6 +372,7 @@ class MlxServerManager {
                 child.off("exit", onExit);
                 if (readiness.owner !== owner) {
                     yield this.terminateOwned(candidate);
+                    resource_budget_1.resourceBudget.check(this.getPortPid(this.port));
                     this.stable = {
                         kind: "adopted",
                         model,
@@ -356,7 +391,7 @@ class MlxServerManager {
             child.off("error", onError);
             child.off("exit", onExit);
             yield this.terminateOwned(candidate);
-            this.lastError = (_k = (_e = startup.error) === null || _e === void 0 ? void 0 : _e.message) !== null && _k !== void 0 ? _k : (this.shouldStop() ? "MLX startup cancelled" : "MLX startup timed out");
+            this.lastError = (_l = (_e = startup.error) === null || _e === void 0 ? void 0 : _e.message) !== null && _l !== void 0 ? _l : (this.shouldStop() ? "MLX startup cancelled" : "MLX startup timed out");
             if (!this.shouldStop()) {
                 this.phase = "failed";
                 console.error(`[daemon] ${this.lastError} — falling back to CPU embeddings`);

package/dist/lib/llm/server.js

@@ -50,6 +50,8 @@ const path = __importStar(require("node:path"));
 const config_1 = require("../../config");
 const index_config_1 = require("../index/index-config");
 const log_rotate_1 = require("../utils/log-rotate");
+const process_1 = require("../utils/process");
+const resource_budget_1 = require("../utils/resource-budget");
 const config_2 = require("./config");
 // ⛔ HARD STOP FOR AI AGENTS: This starts llama-server with a multi-GB local
 // LLM (Qwen3.5-35B GGUF ≈21GB by default). DO NOT start it — via start(),
@@ -124,6 +126,7 @@ class LlmServer {
                 throw new Error("LLM is disabled. Run `gmax llm on` to enable.");
             }
             if (yield this.healthy()) {
+                resource_budget_1.resourceBudget.check(this.readPid());
                 // Adopt an existing server (e.g. after daemon crash + restart)
                 this.lastRequestTime = Date.now();
                 this.startTime = Date.now();
@@ -148,7 +151,17 @@ class LlmServer {
             if (!fs.existsSync(this.config.model)) {
                 throw new Error(`Model file not found: "${this.config.model}". Set GMAX_LLM_MODEL to a valid .gguf path`);
             }
-            const logFd = (0, log_rotate_1.openRotatedLog)(config_1.PATHS.llmLogFile);
+            // Reserve the model's on-disk size plus runtime headroom before starting
+            // it. The default multi-GB model cannot fit the containment budget.
+            const reservation = resource_budget_1.resourceBudget.reserve(Math.ceil(fs.statSync(this.config.model).size / 1048576) + 512, "llm");
+            let logFd;
+            try {
+                logFd = (0, log_rotate_1.openRotatedLog)(config_1.PATHS.llmLogFile);
+            }
+            catch (error) {
+                reservation.release();
+                throw error;
+            }
             const args = [
                 "-m",
                 this.config.model,
@@ -164,16 +177,36 @@ class LlmServer {
             if (this.config.reasoningFormat) {
                 args.push("--reasoning-format", this.config.reasoningFormat);
             }
-            const child = (0, node_child_process_1.spawn)(binary, args, {
-                detached: true,
-                stdio: ["ignore", logFd, logFd],
-            });
+            let child;
+            try {
+                child = (0, node_child_process_1.spawn)(binary, args, {
+                    detached: true,
+                    stdio: ["ignore", logFd, logFd],
+                });
+            }
+            catch (error) {
+                reservation.release();
+                throw error;
+            }
+            finally {
+                fs.closeSync(logFd);
+            }
             child.unref();
-            fs.closeSync(logFd);
             const pid = child.pid;
             if (!pid) {
+                reservation.release();
                 throw new Error("Failed to spawn llama-server — no PID returned");
             }
+            // Group reservations are reaped only after the whole detached group is
+            // confirmed dead. A launcher exit cannot make surviving children free.
+            try {
+                reservation.attach(pid, true);
+            }
+            catch (error) {
+                if (yield (0, process_1.terminateProcessGroup)(pid))
+                    reservation.release();
+                throw error;
+            }
             fs.writeFileSync(config_1.PATHS.llmPidFile, String(pid));
             console.log(`[llm] Starting llama-server (PID: ${pid}, port: ${this.config.port})`);
             // Poll until ready
@@ -257,6 +290,7 @@ class LlmServer {
                 throw new Error("LLM is disabled. Run `gmax llm on` to enable.");
             }
             if (yield this.healthy()) {
+                resource_budget_1.resourceBudget.check(this.readPid());
                 this.touchIdle();
                 return;
             }

package/dist/lib/store/vector-db.js

@@ -63,6 +63,7 @@ const cleanup_1 = require("../utils/cleanup");
 const filter_builder_1 = require("../utils/filter-builder");
 const logger_1 = require("../utils/logger");
 const query_timeout_1 = require("../utils/query-timeout");
+const resource_budget_1 = require("../utils/resource-budget");
 const ann_config_1 = require("./ann-config");
 const compaction_result_1 = require("./compaction-result");
 const lancedb = __importStar(require("./lance-sdk"));
@@ -252,6 +253,8 @@ class VectorDB {
         this.lastDiskCheckMs = 0;
         this.lastLoggedPressure = "ok";
         this.leasePromise = null;
+        this.resourceReservation = null;
+        this.connecting = null;
         this.leaseAbort = new AbortController();
         this.closePromise = null;
         this.releaseLeaseOnClose = true;
@@ -374,6 +377,8 @@ class VectorDB {
     }
     getDb() {
         return __awaiter(this, void 0, void 0, function* () {
+            var _a;
+            var _b, _c;
             if (this.closed) {
                 throw new Error("VectorDB connection is closed");
             }
@@ -387,17 +392,38 @@ class VectorDB {
                 throw new Error("VectorDB connection is closed");
             }
             if (!this.db) {
+                if (this.connecting)
+                    return this.connecting;
                 if (creatingStore || !fs.existsSync(this.lancedbDir)) {
                     this.assertMutationAllowed();
                     (0, maintenance_policy_1.assertFreshDiskMutationAllowed)(this.lancedbDir);
                 }
                 fs.mkdirSync(this.lancedbDir, { recursive: true });
+                // The explicit bounded reader remains available under warning/unknown
+                // pressure. All ordinary clients reserve their native cache capacity in
+                // one shared ledger before opening a connection, including offline MCP.
+                if (!this.options.readOnly) {
+                    this.resourceReservation = resource_budget_1.resourceBudget.reserve(((_b = this.options.indexCacheMb) !== null && _b !== void 0 ? _b : LANCE_INDEX_CACHE_MB) +
+                        ((_c = this.options.metadataCacheMb) !== null && _c !== void 0 ? _c : LANCE_METADATA_CACHE_MB), "native-store");
+                }
                 this.session = createLanceSession(this.options);
                 // 0.38 accepts the legacy third argument in its types but drops it in
                 // the JS wrapper. Put the session in native ConnectionOptions instead.
-                this.db = yield lancedb.connect(this.lancedbDir, {
+                this.connecting = lancedb.connect(this.lancedbDir, {
                     session: this.session,
                 });
+                try {
+                    this.db = yield this.connecting;
+                }
+                catch (error) {
+                    (_a = this.resourceReservation) === null || _a === void 0 ? void 0 : _a.release();
+                    this.resourceReservation = null;
+                    this.session = null;
+                    throw error;
+                }
+                finally {
+                    this.connecting = null;
+                }
             }
             return this.db;
         });
@@ -1874,7 +1900,8 @@ class VectorDB {
     }
     finishClose() {
         return __awaiter(this, void 0, void 0, function* () {
-            var _a, _b;
+            var _a, _b, _c;
+            yield (_a = this.connecting) === null || _a === void 0 ? void 0 : _a.catch(() => { });
             if (this.maintenanceTimer) {
                 clearInterval(this.maintenanceTimer);
                 this.maintenanceTimer = null;
@@ -1893,12 +1920,22 @@ class VectorDB {
                 ]);
                 this.maintenancePromise = null;
             }
-            (_a = this.unregisterCleanup) === null || _a === void 0 ? void 0 : _a.call(this);
+            (_b = this.unregisterCleanup) === null || _b === void 0 ? void 0 : _b.call(this);
             this.unregisterCleanup = undefined;
             let closeError;
-            if ((_b = this.db) === null || _b === void 0 ? void 0 : _b.close) {
+            let nativeClosed = !this.db;
+            if ((_c = this.db) === null || _c === void 0 ? void 0 : _c.close) {
                 try {
                     const closing = this.db.close();
+                    void Promise.resolve(closing).then(() => {
+                        var _a;
+                        nativeClosed = true;
+                        try {
+                            (_a = this.resourceReservation) === null || _a === void 0 ? void 0 : _a.release();
+                            this.resourceReservation = null;
+                        }
+                        catch (_b) { }
+                    }, () => { });
                     if (this.requireClosedOnClose)
                         yield closing;
                     else {
@@ -1914,12 +1951,21 @@ class VectorDB {
             }
             this.db = null;
             this.session = null;
+            if (nativeClosed && this.resourceReservation) {
+                try {
+                    this.resourceReservation.release();
+                    this.resourceReservation = null;
+                }
+                catch (error) {
+                    closeError !== null && closeError !== void 0 ? closeError : (closeError = error);
+                }
+            }
             if (this.leasePromise) {
                 let lease = null;
                 try {
                     lease = yield this.leasePromise;
                 }
-                catch (_c) {
+                catch (_d) {
                     this.leasePromise = null;
                 }
                 if (lease) {