@webmcp-auto-ui/agent 2.5.27 → 2.5.29

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@webmcp-auto-ui/agent",
-  "version": "2.5.27",
+  "version": "2.5.29",
   "description": "LLM agent loop + remote/WASM/local providers + MCP wrapper",
   "license": "AGPL-3.0-or-later",
   "type": "module",
@@ -11,7 +11,7 @@
       "import": "./src/index.ts"
     },
     "./server": {
-      "import": "./src/server/llmProxy.ts"
+      "import": "./src/server/index.ts"
     }
   },
   "scripts": {
@@ -27,5 +27,13 @@
     "@webmcp-auto-ui/core": "file:../core",
     "onnxruntime-web": "^1.24.3",
     "typescript": "^5.0.0"
+  },
+  "peerDependencies": {
+    "vega-embed": "^6.24.0"
+  },
+  "peerDependenciesMeta": {
+    "vega-embed": {
+      "optional": true
+    }
   }
 }

package/src/autoui-server.ts

@@ -5,21 +5,55 @@
 import { createWebMcpServer, parseFrontmatter } from '@webmcp-auto-ui/core';
 import { RAW_RECIPES } from './recipes/_generated.js';
 
-// Notebook widget recipes (vanilla renderers)
+// Notebook widget recipe (vanilla renderer) — moved to @webmcp-auto-ui/ui
 // @ts-ignore — Vite raw imports, not resolved by tsc
-import compactRecipe from './notebook-widgets/recipes/compact.md?raw';
-// @ts-ignore
-import workspaceRecipe from './notebook-widgets/recipes/workspace.md?raw';
-// @ts-ignore
-import documentRecipe from './notebook-widgets/recipes/document.md?raw';
-// @ts-ignore
-import editorialRecipe from './notebook-widgets/recipes/editorial.md?raw';
-
-// Notebook widget renderers (vanilla JS)
-import { render as renderCompact } from './notebook-widgets/compact.js';
-import { render as renderWorkspace } from './notebook-widgets/workspace.js';
-import { render as renderDocument } from './notebook-widgets/document.js';
-import { render as renderEditorial } from './notebook-widgets/editorial.js';
+import notebookRecipe from '@webmcp-auto-ui/ui/widgets/notebook/recipes/notebook.md?raw';
+
+// Notebook widget renderer (vanilla JS) — import via subpath to avoid pulling
+// the .svelte exports of the ui package root through tsc.
+import { render as renderNotebook } from '@webmcp-auto-ui/ui/widgets/notebook/notebook.js';
+import { render as renderRecipeBrowser } from '@webmcp-auto-ui/ui/widgets/notebook/recipe-browser.js';
+
+// Inline recipe for recipe-browser (real vanilla widget)
+const recipeBrowserRecipe = `---
+widget: recipe-browser
+description: Interactive recipe browser with search, kind/tag filters, preview and pick. Use when the user wants to browse, search, or select recipes from connected servers.
+group: rich
+schema:
+  type: object
+  required:
+    - recipes
+  properties:
+    recipes:
+      type: array
+      description: List of Recipe objects (id, name, description, body, servers, ...).
+      items:
+        type: object
+    filters:
+      type: object
+      description: Initial filters
+      properties:
+        q:
+          type: string
+        kind:
+          type: string
+          enum: [all, webmcp, mcp]
+        tags:
+          type: array
+          items:
+            type: string
+    layout:
+      type: string
+      enum: [list, grid]
+      description: Default layout (default list)
+---
+
+## When to use
+When the user wants to browse, search, or pick a recipe — for example "show me the available recipes" or "let me choose a recipe".
+
+## How to use
+Call widget_display({name: "recipe-browser", params: {recipes: [...], layout: "list"}}). The widget emits a bubbling 'widget:interact' CustomEvent with detail={action:"pick", payload: recipe} when the user clicks Pick.
+`;
 
 // ---------------------------------------------------------------------------
 // Inline recipes (frontmatter + body)
@@ -936,77 +970,18 @@ Pour des visualisations custom, animations, ou prototypes interactifs en JS pur.
 Call widget_display({name: "js-sandbox", params: {code: "document.getElementById('root').innerHTML = '<h1>Hello</h1>'"}}).
 `,
 
-  // ── recipe-browser ──────────────────────────────────────────────────────
-  `---
-widget: recipe-browser
-description: Displays available recipes as interactive cards and allows browsing each recipe's details.
-group: rich
-schema:
-  type: object
-  required:
-    - cards
-  properties:
-    title:
-      type: string
-    cards:
-      type: array
-      items:
-        type: object
-        required:
-          - title
-        properties:
-          title:
-            type: string
-          description:
-            type: string
-          tags:
-            type: array
-            items:
-              type: string
-          meta:
-            type: object
-            properties:
-              recipe_name:
-                type: string
-              server:
-                type: string
-    interactive:
-      type: boolean
----
-
-## When to use
-Quand l'utilisateur veut voir les recettes disponibles, explorer les possibilites du serveur, ou comprendre comment utiliser un widget specifique.
-
-## Comment
-
-### Etape 1 — Lister les recettes
-Appelle search_recipes() sur chaque serveur connecte (MCP et WebMCP) pour obtenir la liste des recettes.
-
-### Etape 2 — Afficher en cartes interactives
-Utilise widget_display({name: "cards", params: {...}}) avec le parametre interactive: true pour rendre les cartes cliquables :
-widget_display({name: "cards", params: {title: "Recettes disponibles", cards: [{title: "Nom", description: "Description", tags: ["serveur"], meta: {recipe_name: "nom_technique", server: "nom_serveur"}}], interactive: true}})
-
-Le champ meta est important : il sera renvoye dans l'evenement d'interaction quand l'utilisateur clique sur la carte.
-
-### Etape 3 — Reagir au clic
-Quand l'utilisateur clique sur une carte, tu recevras un message d'interaction contenant les donnees de meta. Utilise meta.recipe_name et meta.server pour :
-1. Appeler get_recipe(meta.recipe_name) sur le bon serveur
-2. Afficher le contenu dans un widget code avec lang: 'markdown'
-3. Lier les deux widgets : reutiliser le widget detail existant via canvas('update', ...) au lieu d'en creer un nouveau a chaque clic.
-
-## Common mistakes
-- Ne pas oublier interactive: true dans les cartes — sans ca, les clics ne remontent pas
-- Ne pas creer un nouveau widget detail a chaque clic — reutiliser l'existant via canvas('update', ...)
-- Les recettes MCP et WebMCP ont des noms de serveur differents — utiliser le bon prefixe pour get_recipe()
-`,
 ];
 
 // ---------------------------------------------------------------------------
 // Native widget names — derived from RECIPES frontmatter
 // ---------------------------------------------------------------------------
 
-/** Derived from RECIPES frontmatter — always in sync with registered widgets */
-export const NATIVE_WIDGET_NAMES = RECIPES.map(r => {
+/** Derived from RECIPES + notebook widget recipes — always in sync with registered widgets */
+const _NOTEBOOK_RECIPE_SOURCES: string[] = [
+  notebookRecipe as string,
+  recipeBrowserRecipe,
+];
+export const NATIVE_WIDGET_NAMES = [...RECIPES, ..._NOTEBOOK_RECIPE_SOURCES].map(r => {
   const match = r.match(/widget:\s*(\S+)/);
   return match ? match[1] : '';
 }).filter(Boolean) as string[];
@@ -1024,12 +999,10 @@ for (const recipe of RECIPES) {
   autoui.registerWidget(recipe, undefined);
 }
 
-// Notebook widgets — vanilla renderers (resolved via WidgetRenderer vanilla path)
+// Notebook widget — vanilla renderer (resolved via WidgetRenderer vanilla path)
 const NOTEBOOK_WIDGETS: Array<[string, (container: HTMLElement, data: any) => any]> = [
-  [compactRecipe as string, renderCompact],
-  [workspaceRecipe as string, renderWorkspace],
-  [documentRecipe as string, renderDocument],
-  [editorialRecipe as string, renderEditorial],
+  [notebookRecipe as string, renderNotebook],
+  [recipeBrowserRecipe, renderRecipeBrowser],
 ];
 for (const [recipe, renderer] of NOTEBOOK_WIDGETS) {
   autoui.registerWidget(recipe, renderer as any);

package/src/index.ts

@@ -11,12 +11,17 @@ export { TRANSFORMERS_MODELS, getTransformersModel, listTransformersModels } fro
 export type { TransformersModelEntry, TransformersFamily, ToolCallFormat } from './providers/transformers-models.js';
 export { parseToolCalls } from './prompts/tool-call-parsers.js';
 export type { ParseResult } from './prompts/tool-call-parsers.js';
-export { loadOrDownloadModel, clearModelCache } from './util/opfs-cache.js';
-export type { ModelFileSpec, CacheProgress } from './util/opfs-cache.js';
+export { loadOrDownloadModel, clearModelCache, listCachedModels, clearAllModelCaches, walkDirectoryStats } from './util/opfs-cache.js';
+export type { ModelFileSpec, CacheProgress, CachedModelInfo } from './util/opfs-cache.js';
+export { listAllStorage, deleteStorageEntry, clearAllStorage } from './util/storage-inventory.js';
+export type { StorageEntry, StorageSource } from './util/storage-inventory.js';
 export { buildGemmaPrompt } from './prompts/index.js';
 export type { BuildGemmaPromptInput } from './prompts/index.js';
 export { LocalLLMProvider } from './providers/local.js';
 export type { LocalLLMProviderOptions, LocalBackend } from './providers/local.js';
+export { HawkProvider } from './providers/hawk.js';
+export type { HawkLLMProviderOptions } from './providers/hawk.js';
+export { HAWK_MODELS, listHawkModels, type HawkModelEntry } from './providers/hawk-models.js';
 export { createProvider } from './providers/factory.js';
 export type { LLMConfig } from './providers/factory.js';
 

package/src/loop.ts

@@ -9,7 +9,7 @@ import type {
   LLMProvider, ProviderTool, McpToolDef, AgentCallbacks,
 } from './types.js';
 import type { ToolLayer, SchemaTransformOptions } from './tool-layers.js';
-import { buildToolsFromLayers, buildDiscoveryToolsWithAliases, activateServerTools, toProviderTools, sanitizeServerName, flattenPathMaps } from './tool-layers.js';
+import { buildToolsFromLayers, buildDiscoveryToolsWithAliases, activateServerTools, toProviderTools, sanitizeServerName } from './tool-layers.js';
 import { buildSystemPromptWithAliases, buildSystemPrompt } from './prompts/index.js';
 import type { DiscoveryCache } from './discovery-cache.js';
 import { unflattenParams, validateJsonSchema } from '@webmcp-auto-ui/core';
@@ -168,6 +168,9 @@ export async function runAgentLoop(
   // Use local alias maps (parallel-safe — no global singleton)
   const activatedServers = new Set<string>();
   const localAliasMap = new Map<string, string>();
+  // Local pathMaps for flattened schemas — populated as servers are lazily activated
+  // (see activateServerTools call below). Parallel-safe: scoped to this loop run.
+  const localPathMaps = new Map<string, Record<string, string[]>>();
   const trace = new PipelineTrace();
 
   const disc = buildDiscoveryToolsWithAliases(options.layers ?? [], schemaOptions, trace);
@@ -228,17 +231,24 @@ export async function runAgentLoop(
     // After 5+ iterations without render, inject a nudge message (once)
     // Merge into existing user message if the last message is already role=user (to avoid consecutive user messages)
     if (iterationsWithoutRender >= 5 && !hasRendered && !nudgedOnce) {
-      nudgedOnce = true;
       const nudgeText = 'STOP exploration. Use the data you already collected. Call widget_display() NOW to display results.';
       const lastMsg = messages[messages.length - 1];
-      if (lastMsg && lastMsg.role === 'user') {
-        if (typeof lastMsg.content === 'string') {
-          lastMsg.content = [{ type: 'text', text: lastMsg.content }, { type: 'text', text: nudgeText }];
-        } else if (Array.isArray(lastMsg.content)) {
-          (lastMsg.content as ContentBlock[]).push({ type: 'text', text: nudgeText });
+      // Skip if last turn carries tool_result blocks — mixing raw text with tool_response
+      // in one turn violates Gemma spec §7 (the serializer would emit text + <|tool_response|>
+      // together). Defer the nudge to a later iteration where the turn is pure-user.
+      const lastHasToolResult = lastMsg && Array.isArray(lastMsg.content)
+        && (lastMsg.content as ContentBlock[]).some(b => b.type === 'tool_result');
+      if (!lastHasToolResult) {
+        nudgedOnce = true;
+        if (lastMsg && lastMsg.role === 'user') {
+          if (typeof lastMsg.content === 'string') {
+            lastMsg.content = [{ type: 'text', text: lastMsg.content }, { type: 'text', text: nudgeText }];
+          } else if (Array.isArray(lastMsg.content)) {
+            (lastMsg.content as ContentBlock[]).push({ type: 'text', text: nudgeText });
+          }
+        } else {
+          messages.push({ role: 'user', content: nudgeText });
         }
-      } else {
-        messages.push({ role: 'user', content: nudgeText });
       }
     }
 
@@ -388,7 +398,10 @@ export async function runAgentLoop(
               activatedServers.add(serverKey);
               const layer = (options.layers ?? []).find(l => sanitizeServerName(l.serverName) === serverName && l.protocol === protocol);
               if (layer) {
-                activeTools = activateServerTools(activeTools, layer, schemaOptions, trace);
+                const act = activateServerTools(activeTools, layer, schemaOptions, trace);
+                activeTools = act.tools;
+                // Merge new pathMaps so unflattenParams works for lazily-activated tools.
+                for (const [k, v] of act.pathMaps) localPathMaps.set(k, v);
               }
             }
           }
@@ -401,8 +414,11 @@ export async function runAgentLoop(
           const protocol = tokenToProtocol(token);
 
           // Auto-repair + validate params before dispatch
+          // Resolve from the full activeTools (not iterationTools, which may be filtered
+          // to strip discovery tools after 4 iterations — would make toolDef undefined
+          // and silently skip auto-repair + schema validation).
           let toolInput = block.input as Record<string, unknown>;
-          const toolDef = iterationTools.find(t => t.name === block.name);
+          const toolDef = activeTools.find(t => t.name === block.name);
           if (toolDef?.input_schema) {
             const repair = autoRepairParams(toolInput, toolDef.input_schema, realToolName);
             if (repair.fixes.length > 0) {
@@ -451,8 +467,9 @@ export async function runAgentLoop(
                 result = `Error: no WebMCP server "${serverName}" found.`;
               } else {
                 // Unflatten params if schema was flattened
+                // Use the local snapshot (parallel-safe) rather than the global singleton.
                 if (schemaOptions?.flatten) {
-                  const pathMap = flattenPathMaps.get(block.name);
+                  const pathMap = localPathMaps.get(block.name);
                   if (pathMap) {
                     toolInput = unflattenParams(toolInput, pathMap);
                   }
@@ -638,18 +655,31 @@ export function trimConversationHistory(history: ChatMessage[], maxTokens: numbe
     total -= removed.reduce((s, m) => s + JSON.stringify(m).length, 0);
   }
 
-  // Remove orphaned tool_result messages at the start — these reference
-  // a tool_use in a message that was trimmed away, causing API errors.
-  while (trimmed.length > 0) {
-    const first = trimmed[0];
-    if (first.role === 'system') break; // preserve system messages at the front
-    const blocks = Array.isArray(first.content) ? first.content : [];
-    const hasToolResult = blocks.some((b: any) => b.type === 'tool_result');
-    if (hasToolResult) {
-      trimmed.shift();
-    } else {
-      break;
+  // Remove orphaned tool_result blocks anywhere in history — strict providers
+  // (Anthropic, etc.) reject tool_result blocks whose tool_use_id does not
+  // correspond to an earlier assistant tool_use. Head-only pruning misses
+  // internal orphans caused by mid-history trims.
+  const validToolUseIds = new Set<string>();
+  for (let i = 0; i < trimmed.length; i++) {
+    const msg = trimmed[i];
+    // Collect tool_use ids from assistant messages seen so far
+    if (msg.role === 'assistant' && Array.isArray(msg.content)) {
+      for (const b of msg.content as any[]) {
+        if (b?.type === 'tool_use' && typeof b.id === 'string') validToolUseIds.add(b.id);
+      }
     }
+    // Filter out orphan tool_result blocks in user messages
+    if (msg.role === 'user' && Array.isArray(msg.content)) {
+      msg.content = (msg.content as any[]).filter(b => {
+        if (b?.type !== 'tool_result') return true;
+        return typeof b.tool_use_id === 'string' && validToolUseIds.has(b.tool_use_id);
+      }) as any;
+    }
+  }
+  // Drop user messages that became empty after orphan-pruning
+  for (let i = trimmed.length - 1; i >= 0; i--) {
+    const c = trimmed[i].content;
+    if (Array.isArray(c) && c.length === 0) trimmed.splice(i, 1);
   }
 
   // Ensure the first non-system message is role=user (API requirement)

package/src/providers/factory.ts

@@ -3,18 +3,27 @@ import { RemoteLLMProvider } from './remote.js';
 import { WasmProvider } from './wasm.js';
 import { LocalLLMProvider, type LocalBackend } from './local.js';
 import { TransformersProvider } from './transformers.js';
+import { HawkProvider } from './hawk.js';
 
 export type LLMConfig =
   | { type: 'remote';       model?: RemoteModelId; proxyUrl?: string; apiKey?: string }
   | { type: 'wasm';         model?: WasmModelId;   onProgress?: (loaded: number, total: number) => void }
   | { type: 'transformers'; model: string;         onProgress?: (loaded: number, total: number) => void }
-  | { type: 'local';        model: string;         baseUrl: string; backend?: LocalBackend };
+  | { type: 'local';        model: string;         baseUrl: string; backend?: LocalBackend }
+  | { type: 'hawk';         model: string;         proxyUrl?: string };
 
 export function createProvider(config: LLMConfig): LLMProvider {
   const base = typeof window !== 'undefined' ? (document.querySelector('base') as HTMLBaseElement | null)?.href ?? '' : '';
 
   // Prefix-based dispatch: a `transformers-*` model routes to TransformersProvider
   // regardless of the declared type (defensive).
+  if ('model' in config && typeof config.model === 'string' && config.model.startsWith('hawk-')) {
+    return new HawkProvider({
+      proxyUrl: (config as { proxyUrl?: string }).proxyUrl ?? `${base}api/hawk`,
+      model: config.model.slice(5),
+    });
+  }
+
   if ('model' in config && typeof config.model === 'string' && config.model.startsWith('transformers-')) {
     const onProgress = (config as { onProgress?: (loaded: number, total: number) => void }).onProgress;
     return new TransformersProvider({
@@ -46,5 +55,10 @@ export function createProvider(config: LLMConfig): LLMProvider {
         model: config.model,
         backend: config.backend,
       });
+    case 'hawk':
+      return new HawkProvider({
+        proxyUrl: config.proxyUrl ?? `${base}api/hawk`,
+        model: config.model,
+      });
   }
 }

package/src/providers/hawk-models.ts

@@ -0,0 +1,23 @@
+export interface HawkModelEntry {
+  id: string;      // ID Hawk (sans préfixe)
+  label: string;   // Label humain pour le selector
+  tokps?: number;  // Tokens/sec estimés (warm, indicatif)
+}
+
+export const HAWK_MODELS: HawkModelEntry[] = [
+  { id: 'qwen35-2b',          label: 'Qwen 3.5 2B — 49 tok/s',         tokps: 49 },
+  { id: 'bielik-1.5b-v3',     label: 'Bielik 1.5B — 47 tok/s',         tokps: 47 },
+  { id: 'gemma4-e2b',         label: 'Gemma 4 E2B — 43 tok/s',         tokps: 43 },
+  { id: 'deepseek-coder-16b', label: 'DeepSeek Coder 16B — 41 tok/s',  tokps: 41 },
+  { id: 'ministral3-3b',      label: 'Ministral 3B — 35 tok/s',        tokps: 35 },
+  { id: 'qwen3-4b',           label: 'Qwen 3 4B — 28 tok/s',           tokps: 28 },
+  { id: 'gemma4-e4b',         label: 'Gemma 4 E4B — 26 tok/s',         tokps: 26 },
+  { id: 'qwen35-4b',          label: 'Qwen 3.5 4B — 23 tok/s',         tokps: 23 },
+  { id: 'qwen36-35b-a3b',     label: 'Qwen 3.6 35B MoE — 22 tok/s',    tokps: 22 },
+  { id: 'gemma4-26b-a4b',     label: 'Gemma 4 26B MoE — 20 tok/s',     tokps: 20 },
+  { id: 'ministral-8b',       label: 'Ministral 8B — 16 tok/s',        tokps: 16 },
+];
+
+export function listHawkModels(): HawkModelEntry[] {
+  return HAWK_MODELS;
+}

package/src/providers/hawk.ts

@@ -0,0 +1,181 @@
+import type { LLMProvider, LLMResponse, ChatMessage, ProviderTool, ContentBlock } from '../types.js';
+
+export interface HawkLLMProviderOptions {
+  proxyUrl: string;   // SvelteKit proxy endpoint, e.g. '/api/hawk'
+  model: string;      // e.g. 'qwen35-2b' (ID Hawk sans préfixe)
+}
+
+// ── OpenAI-compatible types ─────────────────────────────────────────
+
+interface OaiTool {
+  type: 'function';
+  function: { name: string; description: string; parameters: Record<string, unknown> };
+}
+
+interface OaiMessage {
+  role: 'system' | 'user' | 'assistant' | 'tool';
+  content?: string | null;
+  tool_calls?: { id: string; type: 'function'; function: { name: string; arguments: string } }[];
+  tool_call_id?: string;
+}
+
+interface OaiChoice {
+  message: {
+    content?: string | null;
+    tool_calls?: { id: string; type: 'function'; function: { name: string; arguments: string } }[];
+  };
+  finish_reason: string;
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────
+
+let _counter = 0;
+function hawkId(): string {
+  return 'hawk_' + (++_counter).toString(36) + '_' + Date.now().toString(36);
+}
+
+function toOaiTools(tools: ProviderTool[]): OaiTool[] {
+  return tools.map(t => ({
+    type: 'function' as const,
+    function: {
+      name: t.name,
+      description: t.description,
+      parameters: t.input_schema,
+    },
+  }));
+}
+
+function toOaiMessages(messages: ChatMessage[], system?: string): OaiMessage[] {
+  const out: OaiMessage[] = [];
+
+  if (system) out.push({ role: 'system', content: system });
+
+  for (const msg of messages) {
+    if (typeof msg.content === 'string') {
+      out.push({ role: msg.role === 'assistant' ? 'assistant' : 'user', content: msg.content });
+      continue;
+    }
+
+    const blocks = msg.content as ContentBlock[];
+    const textParts = blocks.filter(b => b.type === 'text').map(b => (b as { type: 'text'; text: string }).text);
+    const toolUses = blocks.filter(b => b.type === 'tool_use') as { type: 'tool_use'; id: string; name: string; input: Record<string, unknown> }[];
+    const toolResults = blocks.filter(b => b.type === 'tool_result') as { type: 'tool_result'; tool_use_id: string; content: string }[];
+
+    if (msg.role === 'assistant') {
+      const oai: OaiMessage = { role: 'assistant', content: textParts.join('\n') || null };
+      if (toolUses.length > 0) {
+        oai.tool_calls = toolUses.map(tu => ({
+          id: tu.id,
+          type: 'function' as const,
+          function: { name: tu.name, arguments: JSON.stringify(tu.input) },
+        }));
+      }
+      out.push(oai);
+    } else {
+      // User turn — may contain tool_result blocks (sent back after assistant tool_use)
+      for (const tr of toolResults) {
+        out.push({ role: 'tool', tool_call_id: tr.tool_use_id, content: tr.content });
+      }
+      if (textParts.length > 0) {
+        out.push({ role: 'user', content: textParts.join('\n') });
+      }
+      // If only tool_results and no text, we've already pushed them
+      if (toolResults.length === 0 && textParts.length === 0) {
+        out.push({ role: 'user', content: '' });
+      }
+    }
+  }
+  return out;
+}
+
+function parseArguments(raw: string): Record<string, unknown> {
+  try { return JSON.parse(raw); } catch { return { _raw: raw }; }
+}
+
+// ── Provider ────────────────────────────────────────────────────────
+
+export class HawkProvider implements LLMProvider {
+  readonly name = 'hawk';
+  readonly model: string;
+  private proxyUrl: string;
+
+  constructor(options: HawkLLMProviderOptions) {
+    this.model = options.model;
+    this.proxyUrl = options.proxyUrl;
+  }
+
+  async chat(
+    messages: ChatMessage[],
+    tools: ProviderTool[],
+    options?: { signal?: AbortSignal; system?: string; maxTokens?: number; temperature?: number },
+  ): Promise<LLMResponse> {
+    const oaiMessages = toOaiMessages(messages, options?.system);
+    const oaiTools = tools.length > 0 ? toOaiTools(tools) : undefined;
+
+    // NOTE: `model` is NOT sent in the body — the server proxy injects it
+    // from the X-Model header into the upstream Hawk request.
+    const body: Record<string, unknown> = {
+      messages: oaiMessages,
+      stream: false,
+    };
+    if (oaiTools) body.tools = oaiTools;
+    if (options?.maxTokens) body.max_tokens = options.maxTokens;
+    if (options?.temperature != null) body.temperature = options.temperature;
+
+    const response = await fetch(this.proxyUrl, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+        'X-Model': this.model,
+      },
+      body: JSON.stringify(body),
+      signal: options?.signal,
+    });
+
+    if (!response.ok) {
+      const txt = await response.text().catch(() => '');
+      throw new Error(`Hawk LLM ${response.status}${txt ? ': ' + txt.slice(0, 200) : ''}`);
+    }
+
+    const data = await response.json() as { choices?: OaiChoice[]; usage?: { prompt_tokens?: number; completion_tokens?: number } };
+    const choice = data.choices?.[0];
+    if (!choice) throw new Error('Hawk LLM returned no choices');
+
+    const content: ContentBlock[] = [];
+    const toolCalls = choice.message.tool_calls;
+
+    if (choice.message.content) {
+      content.push({ type: 'text', text: choice.message.content });
+    }
+
+    if (toolCalls && toolCalls.length > 0) {
+      for (const tc of toolCalls) {
+        content.push({
+          type: 'tool_use',
+          id: tc.id || hawkId(),
+          name: tc.function.name,
+          input: parseArguments(tc.function.arguments),
+        });
+      }
+    }
+
+    // Ensure at least one block
+    if (content.length === 0) {
+      content.push({ type: 'text', text: '' });
+    }
+
+    const hasToolUse = content.some(b => b.type === 'tool_use');
+    const stopReason = hasToolUse ? 'tool_use'
+      : choice.finish_reason === 'tool_calls' ? 'tool_use'
+      : 'end_turn';
+
+    return {
+      content,
+      stopReason,
+      usage: data.usage ? {
+        input_tokens: data.usage.prompt_tokens ?? 0,
+        output_tokens: data.usage.completion_tokens ?? 0,
+      } : undefined,
+    };
+  }
+}

package/src/providers/transformers.worker.ts

@@ -114,28 +114,13 @@ async function parseToolCalls(
 }
 
 // --------------------------------------------------------------------------
-// OPFS cache — loaded lazily with a best-effort fallback that defers entirely
-// to transformers.js's built-in HF cache (no OPFS intervention on our side).
+// Cache note: transformers.js manages its own cache via Cache Storage API
+// (enabled by `env.useBrowserCache = true` below). No OPFS pre-download from
+// this worker — the generic OPFS helper requires an explicit file list that
+// transformers.js doesn't expose. Progress is surfaced via `progress_callback`
+// in `fromPretrainedOpts`.
 // --------------------------------------------------------------------------
 
-async function loadOrDownloadModel(
-  _repo: string,
-  _onProgress: (fileProgress: number, totalProgress: number, status: string, loaded?: number, total?: number) => void,
-): Promise<void> {
-  try {
-    // Optional fallback import — module is shipped (../util/opfs-cache.ts);
-    // the try/catch is defensive only, guarding against bundler quirks or
-    // OPFS being unavailable in the worker (older browsers).
-    const mod: any = await import('../util/opfs-cache.js');
-    const fn = mod.loadOrDownloadModel ?? mod.default;
-    if (typeof fn === 'function') return await fn(_repo, _onProgress);
-  } catch {
-    // Import/OPFS unavailable — transformers.js falls back to its internal
-    // HTTP fetch + `caches` API. Progress arrives via from_pretrained's
-    // progress_callback below.
-  }
-}
-
 // --------------------------------------------------------------------------
 // Helpers
 // --------------------------------------------------------------------------
@@ -229,18 +214,6 @@ async function loadModel(modelEntry: TransformersModelEntry): Promise<void> {
 
   stoppingCriteria = new InterruptableStoppingCriteria();
 
-  // Pre-download (OPFS-aware when the cache module is available).
-  await loadOrDownloadModel(modelEntry.repo, (fp, tp, status, loaded, total) => {
-    post({
-      type: 'progress',
-      fileProgress: fp,
-      totalProgress: tp,
-      status,
-      loaded: loaded ?? 0,
-      total: total ?? modelEntry.size,
-    });
-  });
-
   // Aggregated progress callback — sums loaded/total across every file we see,
   // emitting a monotonic aggregate ratio. Two guards eliminate flicker:
   //   1. Files with total < 1_000_000 bytes are ignored (configs, tokenizers,