@bendyline/gezel-inference 0.1.0 → 1.0.1

package/dist/remotes/types.d.ts

@@ -1,3 +1,11 @@
+export interface RemoteKnownModel {
+    id: string;
+    name: string;
+    contextWindow?: number;
+    supportsTools?: boolean;
+    supportsReasoning?: boolean;
+    parameterSize?: string;
+}
 export interface PairedRemote {
     /** Stable id A assigns to this server = its full identity fingerprint. */
     remoteId: string;
@@ -29,6 +37,12 @@ export interface PairedRemote {
     pairedAt: number;
     /** Last time A successfully reached this server (in-memory unless flushed). */
     lastSeenAt?: number;
+    /**
+     * The server's chat models the last time A listed them, already namespaced
+     * for A. Kept so a picker can show them, marked unavailable, while the
+     * server is asleep or A is off its network.
+     */
+    knownModels?: RemoteKnownModel[];
     /**
      * Runtime-managed connections are never persisted and cannot be removed by
      * the ordinary pairing UI. Today the only managed peer is this computer's

package/dist/runtime-lock.js

@@ -9,7 +9,7 @@ export class SingleInstanceError extends Error {
     lockPath;
     constructor(holderPid, lockPath) {
         super(holderPid > 0
-            ? `Another gezel daemon is already running for this home (pid ${holderPid}). Stop it first with \`gezel stop --daemon\` (plain \`gezel stop\` only halts AI work), or use a different GEZEL_HOME.`
+            ? `Another gezel daemon is already running for this home (pid ${holderPid}). Stop it first with \`gezel stop\`, or use a different GEZEL_HOME.`
             : `Could not acquire the single-instance lock at ${lockPath}.`);
         this.holderPid = holderPid;
         this.lockPath = lockPath;

package/dist/system/memory.js

@@ -487,12 +487,22 @@ function isKnownUnifiedNvidiaDevice(name) {
     const normalized = name.toLowerCase();
     return /\bgb10\b/.test(normalized) || normalized.includes('dgx spark');
 }
+/**
+ * A hybrid-graphics laptop powers its discrete GPU down when idle, and the
+ * first `nvidia-smi` after that waits for it to wake: 2.25–2.36 s on an RTX
+ * 5070 Ti laptop (2026-10-15), against ~0.1 s warm. The old 2 s limit read
+ * that as "no NVIDIA GPU", so the machine was sized as CPU-only for the next
+ * minute and offered a 2.4 GB model with a 12 GB card idle beside it. Startup
+ * never awaits this probe, so the margin only costs a caller that asks for the
+ * profile while the driver is truly hung.
+ */
+const NVIDIA_MEMORY_PROBE_TIMEOUT_MS = 10_000;
 async function probeNvidiaMemory() {
     try {
         const { stdout } = await exec('nvidia-smi', ['--query-gpu=name,memory.total', '--format=csv,noheader,nounits'], 
         // Keep this short-lived console probe invisible without detaching it
         // from the daemon that owns and awaits it.
-        { timeout: 2000, ...windowsHeadlessSpawnOptions() });
+        { timeout: NVIDIA_MEMORY_PROBE_TIMEOUT_MS, ...windowsHeadlessSpawnOptions() });
         return parseNvidiaMemoryProbe(stdout);
     }
     catch {

package/dist/utils/service-worker-entry.d.ts

@@ -16,6 +16,10 @@ export declare const SERVICE_WORKER_ENTRIES: {
     readonly 'static-index': {
         readonly built: "index-store/static-index-worker.js";
     };
+    readonly 'dashboard-composite': {
+        readonly built: "ambient/dashboard-composite-worker.js";
+        readonly source: "ambient/dashboard-composite-worker.ts";
+    };
     readonly 'document-convert': {
         readonly built: "index-store/convert-worker.js";
         readonly source: "index-store/convert-worker.ts";

package/dist/utils/service-worker-entry.js

@@ -13,6 +13,10 @@ export const SERVICE_WORKER_ENTRIES = {
     embeddings: { built: 'memory/embed-worker.js' },
     'image-embeddings': { built: 'memory/image-embed-worker.js' },
     'static-index': { built: 'index-store/static-index-worker.js' },
+    'dashboard-composite': {
+        built: 'ambient/dashboard-composite-worker.js',
+        source: 'ambient/dashboard-composite-worker.ts',
+    },
     'document-convert': {
         built: 'index-store/convert-worker.js',
         source: 'index-store/convert-worker.ts',

package/package.json

@@ -1,16 +1,20 @@
 {
   "name": "@bendyline/gezel-inference",
-  "version": "0.1.0",
-  "description": "Shared on-device inference, model management and knowledge runtime for Gezel hosts.",
+  "version": "1.0.1",
+  "description": "Gezel's on-device inference runtime: native engines, local model management, embeddings, knowledge search, and OpenAI-compatible routes.",
   "license": "MIT",
   "keywords": [
     "gezel",
     "ai",
-    "agents",
     "local-first",
-    "daemon",
-    "server",
-    "llm"
+    "llm",
+    "inference",
+    "on-device",
+    "local-ai",
+    "llama.cpp",
+    "mlx",
+    "embeddings",
+    "openai-compatible"
   ],
   "homepage": "https://github.com/bendyline/gezel",
   "bugs": {
@@ -69,16 +73,16 @@
     "!dist/**/__pycache__"
   ],
   "dependencies": {
+    "@bendyline/gezel": "1.2.6",
+    "@bendyline/gezel-catalog": "1.2.6",
+    "@bendyline/gezel-client": "1.2.6",
+    "@bendyline/gezel-knowledge": "1.2.6",
     "hono": "4.13.9",
     "tar": "7.5.21",
     "undici": "8.11.2",
     "yauzl": "^3.4.0",
     "yazl": "^3.3.1",
-    "zod": "^4.4.3",
-    "@bendyline/gezel-client": "1.2.4",
-    "@bendyline/gezel-catalog": "1.2.4",
-    "@bendyline/gezel": "1.2.4",
-    "@bendyline/gezel-knowledge": "1.2.4"
+    "zod": "^4.4.3"
   },
   "peerDependencies": {
     "@huggingface/transformers": "^4.3.1"