@nevescloud/pip 3.9.0 → 3.9.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nevescloud/pip",
-  "version": "3.9.0",
+  "version": "3.9.1",
   "description": "Floating assistant bubble + panel + chat runtime. ESM, no build.",
   "type": "module",
   "main": "pip-core.esm.js",

package/providers/local.esm.js

@@ -257,9 +257,21 @@ export function createTransformersRenderer() {
     return splitThinking(buffer).answer || buffer.trim();
   }
 
+  // Background warm — starts the transformers.js import + tokenizer +
+  // model download without running inference. Hosts call this from
+  // requestIdleCallback after their page boots so the first user prompt
+  // doesn't pay the cold-start cost. Safe to call repeatedly; ensureLoaded
+  // is internally idempotent (returns the in-flight loadingPromise).
+  async function warm(turnEl) {
+    if (!config?.id) return;
+    try { await ensureLoaded(turnEl); }
+    catch { /* swallow — first real generate() will re-throw with the same error */ }
+  }
+
   return {
     setModel,
     generate,
+    warm,
     get currentModelId() { return config?.id || null; },
   };
 }
@@ -295,7 +307,7 @@ export function local({
   const renderer = createTransformersRenderer();
   if (model) renderer.setModel({ id: model, dtype, maxTokens, genParams, chatTemplate });
 
-  return ({ messages, signal, system, tools, turnEl, setReplyText }) => (async function* () {
+  const provider = ({ messages, signal, system, tools, turnEl, setReplyText }) => (async function* () {
     const effectiveSystem = system || systemPrompt || '';
     const augmentedSystem = buildToolSystemPrompt(effectiveSystem, tools);
 
@@ -347,6 +359,13 @@ export function local({
     if (error) throw error;
     yield { type: 'turn_end', stopReason: sawToolUse ? 'tool_use' : 'end_turn' };
   })();
+
+  // Surface the renderer's background warm() on the provider so hosts
+  // can trigger transformers.js import + weight download in idle time
+  // ahead of the first user message. Idempotent; no-op until setModel
+  // has run (covered above when `model` is passed).
+  provider.warm = (turnEl) => renderer.warm(turnEl);
+  return provider;
 }
 
 export { splitThinking };