@webmcp-auto-ui/agent 2.5.23 → 2.5.25

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@webmcp-auto-ui/agent",
-  "version": "2.5.23",
+  "version": "2.5.25",
   "description": "LLM agent loop + remote/WASM/local providers + MCP wrapper",
   "license": "AGPL-3.0-or-later",
   "type": "module",

package/src/discovery-cache.ts

@@ -6,6 +6,9 @@
 
 import type { ProviderTool } from './types.js';
 
+/** Tool names that are resolved locally from cache — hidden from user-facing browsers. */
+export const DISCOVERY_TOOL_NAMES = new Set(['list_recipes', 'search_recipes', 'get_recipe', 'list_tools', 'search_tools']);
+
 export interface CachedRecipe {
   name: string;
   description?: string;
@@ -79,6 +82,12 @@ export class DiscoveryCache {
     return this.servers.get(serverPrefix)?.tools.length ?? 0;
   }
 
+  /** Tool count excluding discovery tools (hidden from user-facing browsers) */
+  browsableToolCount(serverPrefix: string): number {
+    const tools = this.servers.get(serverPrefix)?.tools ?? [];
+    return tools.filter(t => !DISCOVERY_TOOL_NAMES.has(t.name)).length;
+  }
+
   /**
    * Try to resolve a discovery tool call from cache.
    * Returns the result string if handled, or null if not a discovery tool.

package/src/index.ts

@@ -29,7 +29,7 @@ export { buildToolsFromLayers, buildDiscoveryTools, buildDiscoveryToolsWithAlias
 export type { ToolLayer, McpLayer, WebMcpLayer, SystemPromptResult, DiscoveryToolsResult, SchemaTransformOptions, BuildToolsResult } from './tool-layers.js';
 
 // Discovery cache
-export { DiscoveryCache } from './discovery-cache.js';
+export { DiscoveryCache, DISCOVERY_TOOL_NAMES } from './discovery-cache.js';
 export type { CachedRecipe, ServerCache } from './discovery-cache.js';
 
 // Re-export core WebMCP types

package/src/loop.ts

@@ -92,6 +92,8 @@ export interface AgentLoopOptions {
   maxIterations?: number;
   maxTokens?: number;
   maxTools?: number;
+  /** WASM-only: cap on conversation messages sent to the model. Default: derived from contextSize. */
+  maxMessages?: number;
   temperature?: number;
   topK?: number;
   cacheEnabled?: boolean;
@@ -125,6 +127,7 @@ export async function runAgentLoop(
     maxIterations = 5,
     maxTokens,
     maxTools,
+    maxMessages,
     temperature,
     topK,
     cacheEnabled = true,
@@ -264,7 +267,7 @@ export async function runAgentLoop(
     const t0 = performance.now();
     let streamingText = '';
     const response = await provider.chat(messages, iterationTools, {
-      signal, cacheEnabled, system: iterationSystemPrompt, maxTokens, maxTools, temperature, topK,
+      signal, cacheEnabled, system: iterationSystemPrompt, maxTokens, maxTools, maxMessages, temperature, topK,
       onToken: callbacks.onToken ? (token) => {
         callbacks.onToken!(token);
         streamingText += token;

package/src/providers/wasm.ts

@@ -176,7 +176,7 @@ export class WasmProvider implements LLMProvider {
   async chat(
     messages: ChatMessage[],
     tools: ProviderTool[],
-    options?: { signal?: AbortSignal; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; system?: string; maxTools?: number }
+    options?: { signal?: AbortSignal; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; system?: string; maxTools?: number; maxMessages?: number }
   ): Promise<LLMResponse> {
     if (this.status !== 'ready') await this.initialize();
     if (!this.inference) throw new Error('Model not initialized');
@@ -202,7 +202,7 @@ export class WasmProvider implements LLMProvider {
   private async _chat(
     messages: ChatMessage[],
     tools: ProviderTool[],
-    options?: { signal?: AbortSignal; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; system?: string; maxTools?: number }
+    options?: { signal?: AbortSignal; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; system?: string; maxTools?: number; maxMessages?: number }
   ): Promise<LLMResponse> {
     // Apply per-request options
     if (options?.maxTokens || options?.temperature || options?.topK) {
@@ -222,7 +222,7 @@ export class WasmProvider implements LLMProvider {
 
     // Aggressive clipping: Gemma struggles with long conversations — dynamic cap based on context size
     const contextTokens = this.opts.contextSize ?? 4096;
-    const MAX_MESSAGES = contextTokens <= 4096 ? 8 : contextTokens <= 8192 ? 16 : 32;
+    const MAX_MESSAGES = options?.maxMessages ?? Math.max(4, Math.floor(contextTokens / 512));
     while (messages.length > MAX_MESSAGES) {
       messages = messages.slice(1);
     }
@@ -660,9 +660,23 @@ export class WasmProvider implements LLMProvider {
   private buildPrompt(messages: ChatMessage[], tools: ProviderTool[], systemPrompt?: string, maxTools?: number): string {
     const systemParts: string[] = [];
 
-    // Inject system prompt from settings if provided
+    // Inject system prompt from settings if provided.
+    // Rewrite paren syntax "tool_name()" / "tool_name(args)" to Gemma 4 native call syntax.
+    // Without this, Gemma mimics the paren syntax as plain text (regression from commit 2724b9e).
     if (systemPrompt) {
-      systemParts.push(systemPrompt);
+      const gemmaPrompt = systemPrompt.replace(
+        /\b([a-zA-Z_][a-zA-Z0-9_]*)\(([^)]*)\)/g,
+        (_full, name, args) => {
+          const trimmed = args.trim();
+          if (!trimmed) return `<|tool_call>call:${name}{}<tool_call|>`;
+          const argBody = trimmed
+            .split(',')
+            .map((a: string) => `${a.trim()}:<|"|>...<|"|>`)
+            .join(',');
+          return `<|tool_call>call:${name}{${argBody}}<tool_call|>`;
+        }
+      );
+      systemParts.push(gemmaPrompt);
     }
 
     if (tools.length > 0) {

package/src/tool-layers.ts

@@ -465,6 +465,8 @@ Unless a recipe specifies otherwise, use these tools to display your responses o
 
 ${actionTools.join('\n')}
 
+widget_display may ONLY be called with data returned by a non-autoui DATA tool actually invoked in the current session. Fabricating IDs, URLs, names, dates, or any content not returned by a tool is a critical violation. If no DATA tool has been called yet, go back to STEP 1.
+
 STEP 5 — Fallback
 
 If previous steps failed, fall back to a classic chat without tool calling.`;

package/src/types.ts

@@ -51,7 +51,7 @@ export interface LLMProvider {
   chat(
     messages: ChatMessage[],
     tools: ProviderTool[],
-    options?: { signal?: AbortSignal; cacheEnabled?: boolean; system?: string; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; maxTools?: number }
+    options?: { signal?: AbortSignal; cacheEnabled?: boolean; system?: string; maxTokens?: number; temperature?: number; topK?: number; onToken?: (token: string) => void; maxTools?: number; maxMessages?: number }
   ): Promise<LLMResponse>;
 }