@aparte/provider-transformers 0.16.3 → 0.16.5

package/dist/index.js

@@ -1,5 +1,5 @@
-import { uuid, contentToText } from "@aparte/core";
-const workerUrl = "" + new URL("assets/worker-Dz5bU2S7.js", import.meta.url).href;
+import { uuid } from "@aparte/core";
+const workerUrl = "" + new URL("assets/worker-Bm0eBd4l.js", import.meta.url).href;
 let _hardwareTiers = null;
 function setHardwareTierModels(tiers) {
   _hardwareTiers = tiers;
@@ -83,29 +83,20 @@ async function _refreshKnownModels() {
   } catch {
   }
 }
-let _warnedToolTurnsDropped = false;
-function toMessages(messages) {
-  const result = [];
-  let droppedToolTurns = 0;
-  for (const m of messages) {
-    if (m.role === "user" || m.role === "assistant" || m.role === "system") {
-      const text = contentToText(m.content);
-      if (text) result.push({ role: m.role, content: text });
-    } else {
-      droppedToolTurns++;
-    }
-  }
-  if (droppedToolTurns > 0 && !_warnedToolTurnsDropped) {
-    _warnedToolTurnsDropped = true;
-    console.warn(
-      `[transformers] Dropped ${droppedToolTurns} tool turn(s) from the prompt: this provider does not support tool calling (v1), so the model will not see the call or its result. Use an OpenAI-compatible endpoint for tools, or render the turns yourself before sending.`
-    );
-  }
-  return result;
+function _selection(modelId) {
+  const config = _registeredModels.get(modelId);
+  const runner = config?.runner;
+  return {
+    task: config?.task ?? "text-generation",
+    ...runner ? { runner: typeof location === "undefined" ? runner : new URL(runner, location.href).href } : {},
+    dtype: config?.dtype,
+    device: _computeDevice
+  };
 }
 let _worker = null;
 const _pendingPrepares = /* @__PURE__ */ new Map();
 const _pendingGenerates = /* @__PURE__ */ new Map();
+const _pendingCommands = /* @__PURE__ */ new Map();
 let _generateChain = Promise.resolve();
 const _generateDoneResolvers = /* @__PURE__ */ new Map();
 const _queuedModelIds = /* @__PURE__ */ new Map();
@@ -140,7 +131,7 @@ function _spawnWorker() {
   const canMintBlob = typeof Blob === "function" && typeof URL.createObjectURL === "function";
   if (sameOrigin || !canMintBlob) return new Worker(new URL(
     /* @vite-ignore */
-    "" + new URL("assets/worker-Dz5bU2S7.js", import.meta.url).href,
+    "" + new URL("assets/worker-Bm0eBd4l.js", import.meta.url).href,
     import.meta.url
   ), { type: "module" });
   _workerBlobUrl = URL.createObjectURL(
@@ -207,6 +198,8 @@ function _handleWorkerError(e) {
     }
   }
   _pendingGenerates.clear();
+  for (const c of _pendingCommands.values()) c.reject(new Error(message));
+  _pendingCommands.clear();
   for (const resolve of _generateDoneResolvers.values()) {
     try {
       resolve();
@@ -258,10 +251,23 @@ function _handleWorkerMessage(event) {
       void _enforceMaxCachedModels(msg.modelId).then(() => _refreshKnownModels());
       break;
     }
-    case "gen-chunk": {
+    case "gen-event": {
       const ctrl = _pendingGenerates.get(msg.id);
       if (!ctrl) break;
-      ctrl.enqueue({ type: msg.chunkType, delta: msg.delta });
+      ctrl.enqueue(msg.event);
+      break;
+    }
+    case "warning": {
+      console.warn(`[transformers] ${msg.message}`);
+      break;
+    }
+    case "command-result": {
+      _releaseGenerateSlot(msg.id);
+      const pending = _pendingCommands.get(msg.id);
+      if (!pending) break;
+      _pendingCommands.delete(msg.id);
+      if (msg.error !== void 0) pending.reject(new Error(msg.error));
+      else pending.resolve(msg.result);
       break;
     }
     case "gen-done": {
@@ -305,35 +311,59 @@ const TransformersProvider = {
     await _refreshKnownModels();
     return _knownModels;
   },
-  async chat(request) {
-    const messages = toMessages(request.messages);
+  async chat(request, _config, ctx) {
     const requestId = uuid();
     const options = {
       maxTokens: request.maxTokens,
       temperature: request.temperature,
       seed: request.seed
     };
-    const task = _registeredModels.get(request.modelId)?.task ?? "text-generation";
+    const signal = ctx?.signal;
     _warnIfContended(request.modelId);
     _queuedModelIds.set(requestId, request.modelId);
     const prevGenerate = _generateChain;
     _generateChain = new Promise((resolveSlot) => {
       _generateDoneResolvers.set(requestId, resolveSlot);
     });
+    let posted = false;
+    let stopped = false;
+    const stop = () => {
+      if (stopped) return;
+      stopped = true;
+      signal?.removeEventListener("abort", stop);
+      if (posted) {
+        _getWorker().postMessage({ type: "cancel", id: requestId });
+        return;
+      }
+      const ctrl = _pendingGenerates.get(requestId);
+      _pendingGenerates.delete(requestId);
+      if (!ctrl) return;
+      try {
+        ctrl.enqueue({ type: "error", message: "Generation cancelled before it started" });
+        ctrl.close();
+      } catch {
+      }
+    };
     const postGenerate = () => {
+      if (stopped) {
+        _releaseGenerateSlot(requestId);
+        return;
+      }
+      posted = true;
       _getWorker().postMessage({
         type: "generate",
         id: requestId,
         modelId: request.modelId,
-        messages,
+        // The conversation as it is, parts included: which parts a model can take
+        // is the runner's knowledge, not this thread's.
+        messages: request.messages,
         options,
-        task,
-        dtype: _registeredModels.get(request.modelId)?.dtype,
-        device: _computeDevice
+        ..._selection(request.modelId)
       });
     };
+    let response;
     if (request.stream === false) {
-      return new Promise((resolve, reject) => {
+      response = new Promise((resolve, reject) => {
         let result = "";
         const fakeCtrl = {
           enqueue: (chunk) => {
@@ -347,18 +377,22 @@ const TransformersProvider = {
         _pendingGenerates.set(requestId, fakeCtrl);
         void prevGenerate.then(postGenerate);
       });
+    } else {
+      response = new ReadableStream({
+        async start(controller) {
+          _pendingGenerates.set(requestId, controller);
+          await prevGenerate;
+          postGenerate();
+        },
+        cancel() {
+          _pendingGenerates.delete(requestId);
+          stop();
+        }
+      });
     }
-    return new ReadableStream({
-      async start(controller) {
-        _pendingGenerates.set(requestId, controller);
-        await prevGenerate;
-        postGenerate();
-      },
-      cancel() {
-        _pendingGenerates.delete(requestId);
-        _getWorker().postMessage({ type: "cancel", id: requestId });
-      }
-    });
+    if (signal?.aborted) stop();
+    else signal?.addEventListener("abort", stop, { once: true });
+    return response;
   },
   async getModelStatus(modelId) {
     if (_loadedModelId === modelId) return "ready";
@@ -386,11 +420,9 @@ const TransformersProvider = {
     }
     const requestId = uuid();
     _preparingModelId = modelId;
-    const task = _registeredModels.get(modelId)?.task ?? "text-generation";
-    const dtype = _registeredModels.get(modelId)?.dtype;
     return new Promise((resolve, reject) => {
       _pendingPrepares.set(requestId, { modelId, onProgress, resolve, reject });
-      _getWorker().postMessage({ type: "prepare", id: requestId, modelId, task, dtype, device: _computeDevice });
+      _getWorker().postMessage({ type: "prepare", id: requestId, modelId, ..._selection(modelId) });
     });
   },
   async deleteModel(modelId) {
@@ -400,6 +432,20 @@ const TransformersProvider = {
 function getLoadedModelId() {
   return _loadedModelId;
 }
+function runnerCommand(modelId, name, payload) {
+  const requestId = uuid();
+  _queuedModelIds.set(requestId, modelId);
+  const previous = _generateChain;
+  _generateChain = new Promise((resolveSlot) => {
+    _generateDoneResolvers.set(requestId, resolveSlot);
+  });
+  return new Promise((resolve, reject) => {
+    _pendingCommands.set(requestId, { resolve, reject });
+    void previous.then(() => {
+      _getWorker().postMessage({ type: "command", id: requestId, modelId, name, payload, ..._selection(modelId) });
+    });
+  });
+}
 function terminateWorker() {
   _worker?.terminate();
   _worker = null;
@@ -418,6 +464,8 @@ function terminateWorker() {
     }
   }
   _pendingGenerates.clear();
+  for (const c of _pendingCommands.values()) c.reject(new Error("Worker terminated"));
+  _pendingCommands.clear();
   for (const resolve of _generateDoneResolvers.values()) {
     try {
       resolve();
@@ -510,6 +558,7 @@ export {
   getMaxCachedModels,
   listCachedModels,
   registerModel,
+  runnerCommand,
   setComputeDevice,
   setHardwareTierModels,
   setMaxCachedModels,

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sources":["../src/index.ts"],"sourcesContent":["/**\n * @aparte/provider-transformers — run LLMs 100% in the browser via Transformers.js.\n *\n * A local, keyless `AparteAIProvider`: it owns its I/O (inference runs off the main\n * thread in a Web Worker) so `AparteDirectTransport` delegates to its `chat()`. Model\n * weights download once and persist in the Cache API.\n *\n * Scope (v1): generic **text-generation** streaming. Tool-calling for local models is\n * model-specific (every family has its own wire format) and is out of scope here — the\n * app registers models and streams plain replies. Vision / embeddings can follow on demand.\n *\n * ## This provider's state is TAB-scoped, on purpose\n *\n * Everything below the \"Worker bridge\" heading — the worker, the loaded model, the\n * generate chain — plus `setComputeDevice`, `setMaxCachedModels` and\n * `setHardwareTierModels`, is module-level and therefore shared by every chat on the\n * page. That is deliberate, and it is the opposite of what the rest of the suite does:\n * a plugin's providers scope to one chat, this one cannot.\n *\n * The reason is the resource, not the design. A local model is 1–2 GB of weights and one\n * WebGPU pipeline. Handing each chat its own worker would mean N copies resident in one\n * tab — which is the failure this package exists to avoid, not a capability. The\n * settings above describe the *machine* (which backend, how many models to keep\n * cached), so per-chat values would not mean anything either.\n *\n * What the constraint costs: two chats on the page driving DIFFERENT local models take\n * turns on one pipeline, so each turn may evict and reload gigabytes. That used to\n * happen silently — a multi-second stall with nothing to read. It now warns once, from\n * `chat()`, when a generate is queued for a model other than the one already in flight.\n * Same model in both chats is free and correct: they share the load.\n */\n\nimport type {\n    AparteAIProvider,\n    AparteAIModel,\n    AparteChatRequest,\n    AparteChatResponse,\n    AparteChatMessage,\n    ModelStatus,\n    ModelLoadProgress,\n} from '@aparte/core';\nimport { contentToText, uuid } from '@aparte/core';\n\n// The worker's URL, not the worker itself: this package constructs it by hand because a\n// cross-origin copy has to go through a blob (see `_spawnWorker`). `?worker&url` is what\n// keeps Vite emitting the worker as its own chunk — the `new Worker(new URL(...))` form\n// it detects by pattern was the only other way, and moving the URL out of that call made\n// the build inline the worker's raw TypeScript as a data: URL instead. Caught by a\n// two-origin browser probe, not by any test.\nimport workerUrl from './worker.ts?worker&url';\n\n/** The minimal chat shape passed to the worker (the tokenizer applies the chat template). */\ntype SimpleMessage = { role: 'user' | 'assistant' | 'system'; content: string };\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Hardware detection\n// ─────────────────────────────────────────────────────────────────────────────\n\nexport interface HardwareProfile {\n    hasGpu: boolean;\n    ramGb: number;\n    tier: 'low' | 'mid' | 'high';\n    recommendedModelId: string;\n}\n\n/** Hardware-tier model overrides — set by the app via setHardwareTierModels(). */\nlet _hardwareTiers: { low: string; mid?: string; high: string } | null = null;\n\n/**\n * Set the model IDs to use per hardware tier. Call before detectHardware() is used\n * to pick a default model — the provider ships no model knowledge of its own.\n */\nexport function setHardwareTierModels(tiers: { low: string; mid?: string; high: string }): void {\n    _hardwareTiers = tiers;\n}\n\nexport async function detectHardware(): Promise<HardwareProfile> {\n    // navigator.deviceMemory: W3C API, Chromium only, capped at 8 GB for privacy\n    // (1 | 2 | 4 | 8). Falls back to 4 on Firefox/Safari.\n    const ramGb: number = (navigator as unknown as { deviceMemory?: number }).deviceMemory ?? 4;\n\n    // Real WebGPU check: requestAdapter() returns null if no capable GPU is present.\n    let hasGpu = false;\n    if ('gpu' in navigator) {\n        try {\n            const adapter = await (navigator as unknown as { gpu: { requestAdapter(): Promise<unknown> } }).gpu.requestAdapter();\n            hasGpu = adapter !== null;\n        } catch {\n            hasGpu = false;\n        }\n    }\n\n    let tier: 'low' | 'mid' | 'high';\n    if (!hasGpu || ramGb < 4) {\n        tier = 'low';\n    } else if (ramGb < 8) {\n        tier = 'mid';\n    } else {\n        tier = 'high';\n    }\n\n    const recommendedModelId = _hardwareTiers\n        ? (_hardwareTiers[tier] ?? _hardwareTiers.high ?? '')\n        : '';\n\n    return { hasGpu, ramGb, tier, recommendedModelId };\n}\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Model catalog — all model knowledge lives in the app, not the provider.\n// ─────────────────────────────────────────────────────────────────────────────\n\n/** Configuration for a model registered with the provider. */\nexport interface TransformersModelConfig {\n    id: string;\n    name: string;\n    description?: string;\n    capabilities: AparteAIModel['capabilities'];\n    /** Transformers.js pipeline task — determines the model architecture / load path. */\n    task: 'text-generation';\n    /** ONNX dtype or per-part dtype map (e.g. `'q4'` or `{ decoder_model_merged: 'q4' }`). */\n    dtype?: string | Record<string, string>;\n    /** Preferred device. Defaults to WebGPU when available, else WASM. */\n    device?: 'webgpu' | 'wasm' | 'auto';\n    metadata?: Record<string, unknown>;\n}\n\n/** Models registered by the app via registerModel(). */\nconst _registeredModels = new Map<string, TransformersModelConfig>();\n\n/** Mutable model list — populated by registerModel() and cache discovery. */\nlet _knownModels: AparteAIModel[] = [];\n\n/**\n * Register a model with the provider. Call before the model is used for inference.\n */\nexport function registerModel(config: TransformersModelConfig): void {\n    _registeredModels.set(config.id, config);\n    if (!_knownModels.find(m => m.id === config.id)) {\n        _knownModels = [..._knownModels, {\n            id: config.id,\n            name: config.name,\n            description: config.description,\n            capabilities: config.capabilities,\n        }];\n    }\n}\n\n/** Build an AparteAIModel entry from a cache-discovered modelId not in the registry. */\nfunction _modelFromCacheEntry(modelId: string): AparteAIModel {\n    const config = _registeredModels.get(modelId);\n    if (config) return { id: config.id, name: config.name, description: config.description, capabilities: config.capabilities };\n    const name = (modelId.split('/').pop() ?? modelId).replace(/-/g, ' ');\n    return { id: modelId, name, capabilities: ['streaming'] };\n}\n\n/** Max number of models to keep in cache. 0 = unlimited. Default: 1. */\nlet _maxCachedModels = 1;\n\n/**\n * Set the maximum number of models to keep in cache. When exceeded after a new\n * model is ready, the oldest models are evicted. 0 = unlimited.\n */\nexport function setMaxCachedModels(max: number): void {\n    _maxCachedModels = max;\n}\n\n/** Returns the current max-cached-models setting. */\nexport function getMaxCachedModels(): number {\n    return _maxCachedModels;\n}\n\n/**\n * User's preferred compute backend for local inference.\n *   'auto'   → WebGPU when available, else WASM (default)\n *   'webgpu' → force WebGPU\n *   'wasm'   → force WASM CPU\n */\nexport type ComputeDevice = 'auto' | 'webgpu' | 'wasm';\nlet _computeDevice: ComputeDevice = 'auto';\n\nexport function setComputeDevice(d: ComputeDevice): void {\n    _computeDevice = d;\n}\n\nexport function getComputeDevice(): ComputeDevice {\n    return _computeDevice;\n}\n\n/** Evict models from cache until count <= _maxCachedModels; `keepModelId` is never evicted. */\nasync function _enforceMaxCachedModels(keepModelId: string): Promise<void> {\n    if (_maxCachedModels === 0) return; // unlimited\n    try {\n        const cached = await listCachedModels();\n        const others = cached.filter(e => e.modelId !== keepModelId);\n        const excess = cached.length - _maxCachedModels;\n        if (excess <= 0) return;\n        // Delete the excess models (oldest first — they appear first in cache scan order).\n        for (let i = 0; i < excess && i < others.length; i++) {\n            await deleteCachedModel(others[i]!.modelId);\n        }\n    } catch { /* cache unavailable */ }\n}\n\n/** Merge cached models into _knownModels (idempotent). Called by fetchModels(). */\nasync function _refreshKnownModels(): Promise<void> {\n    try {\n        const cached = await listCachedModels();\n        for (const entry of cached) {\n            if (!_knownModels.find(m => m.id === entry.modelId)) {\n                _knownModels = [..._knownModels, _modelFromCacheEntry(entry.modelId)];\n            }\n        }\n    } catch { /* cache unavailable */ }\n}\n\n/** Warn at most once per session that tool turns were left out of the prompt. */\nlet _warnedToolTurnsDropped = false;\n\n/** AparteChatMessage[] → plain chat turns (the tokenizer's chat template does the rest). */\nfunction toMessages(messages: AparteChatMessage[]): SimpleMessage[] {\n    const result: SimpleMessage[] = [];\n    let droppedToolTurns = 0;\n    for (const m of messages) {\n        if (m.role === 'user' || m.role === 'assistant' || m.role === 'system') {\n            const text = contentToText(m.content);\n            if (text) result.push({ role: m.role, content: text });\n        } else {\n            // tool_call / tool_result are not supported by this generic provider (v1):\n            // rendering them needs a model-specific tool syntax. Dropping them\n            // silently meant an app with registered tools got a model that never saw\n            // the call or its result, with nothing to explain the behaviour.\n            droppedToolTurns++;\n        }\n    }\n    if (droppedToolTurns > 0 && !_warnedToolTurnsDropped) {\n        _warnedToolTurnsDropped = true;\n        console.warn(\n            `[transformers] Dropped ${droppedToolTurns} tool turn(s) from the prompt: this provider ` +\n            'does not support tool calling (v1), so the model will not see the call or its result. ' +\n            'Use an OpenAI-compatible endpoint for tools, or render the turns yourself before sending.',\n        );\n    }\n    return result;\n}\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Worker bridge\n// ─────────────────────────────────────────────────────────────────────────────\n\nlet _worker: Worker | null = null;\n\ninterface PendingPrepare {\n    modelId: string;\n    onProgress: (p: ModelLoadProgress) => void;\n    resolve: () => void;\n    reject: (err: Error) => void;\n}\nconst _pendingPrepares = new Map<string, PendingPrepare>();\nconst _pendingGenerates = new Map<string, ReadableStreamDefaultController>();\n\n// ── Generate serialization ──────────────────────────────────────────────────\n// The worker holds ONE pipeline: two concurrent generates would corrupt each\n// other. Each chat() chains its `generate` behind the previous generate's\n// completion (gen-done / gen-error).\nlet _generateChain: Promise<void> = Promise.resolve();\nconst _generateDoneResolvers = new Map<string, () => void>();\n\n// ── Contention on the one pipeline ──────────────────────────────────────────\n// Serialization is correct but invisible: two chats driving DIFFERENT local\n// models take turns, and with `maxCachedModels` at its default of 1 each turn\n// can evict and reload gigabytes. The user sees a stall; the developer sees\n// nothing. These two track just enough to say so, once.\nconst _queuedModelIds = new Map<string, string>();\nlet _warnedModelContention = false;\n\n/** Model ids of generates currently queued or running on the single pipeline. */\nfunction _contendingModelId(requested: string): string | undefined {\n    for (const id of _queuedModelIds.values()) if (id !== requested) return id;\n    return undefined;\n}\n\n/**\n * Warn once when a generate has to queue behind another chat's DIFFERENT model.\n * Not a warning about switching models in one chat — that is a deliberate act\n * with visible feedback. This fires only when two are in flight at once.\n */\nfunction _warnIfContended(requested: string): void {\n    if (_warnedModelContention) return;\n    const other = _contendingModelId(requested);\n    if (!other) return;\n    _warnedModelContention = true;\n    console.warn(\n        `[Aparte] Two chats are driving different local models at once (\"${requested}\" behind `\n        + `\"${other}\"). Transformers.js runs one pipeline per tab, so these generates are `\n        + `serialized, and with a cache budget of ${_maxCachedModels} each switch can evict and `\n        + `reload gigabytes of weights. Point both chats at one model, or raise the budget with `\n        + `setMaxCachedModels(2) if the machine has the memory. This warns once.`,\n    );\n}\n\n/** Settle the serialization slot for a finished generate. */\nfunction _releaseGenerateSlot(id: string): void {\n    _queuedModelIds.delete(id);\n    const resolve = _generateDoneResolvers.get(id);\n    if (resolve) {\n        _generateDoneResolvers.delete(id);\n        resolve();\n    }\n}\n\n/** Model known to be loaded (main-thread view). */\nlet _loadedModelId: string | null = null;\n/** Model currently being prepared (for the getModelStatus 'cached' path). */\nlet _preparingModelId: string | null = null;\n\n/** The blob URL the worker was built from, if it needed one. Revoked with the worker. */\nlet _workerBlobUrl: string | null = null;\n\n/**\n * Build the worker — including when this package is served from another origin.\n *\n * `new Worker()` refuses a cross-origin script outright, and that is not an exotic\n * case: it is every deploy whose JavaScript lives on a CDN or an asset host while the\n * page lives somewhere else, with or without a bundler. Reproduced with the package on\n * one port and the page on another: `SecurityError: Script at '…/assets/worker-*.js'\n * cannot be accessed from origin '…'`.\n *\n * A blob inherits the ORIGIN OF THE DOCUMENT THAT CREATES IT, so a one-line blob whose\n * body imports the real worker by absolute URL is same-origin by construction, and the\n * import inside it is a normal cross-origin module fetch, which is allowed. It is the\n * shim ffmpeg.wasm and tesseract.js use for the same reason.\n *\n * Same-origin keeps the direct path: no blob, nothing to revoke, and a stack trace that\n * names the real file.\n */\nfunction _spawnWorker(): Worker {\n    const url = new URL(workerUrl, import.meta.url);\n    const sameOrigin = typeof location === 'undefined' || url.origin === location.origin;\n    // A blob is the only way across an origin, so an environment that cannot mint one has\n    // nothing to gain from trying: construct directly and let the platform say what it\n    // thinks. jsdom is that environment — it has `Blob` and no `URL.createObjectURL` — and\n    // every test in this package went through the blob path and threw before this line\n    // existed.\n    const canMintBlob = typeof Blob === 'function' && typeof URL.createObjectURL === 'function';\n    // The literal below is not style. `new Worker(new URL('./worker.ts', import.meta.url))`\n    // is the exact shape Vite's worker detection and webpack's WorkerPlugin match on, and\n    // matching it is what makes a CONSUMER's bundler process the worker as a module — which\n    // is how `@huggingface/transformers` gets resolved inside it today. Behind a variable\n    // the chunk is copied as an opaque asset and its imports are never touched, so hoisting\n    // this line to reuse it for the blob would fix a CDN page by breaking every bundled app.\n    if (sameOrigin || !canMintBlob) return new Worker(new URL('./worker.ts', import.meta.url), { type: 'module' });\n\n    _workerBlobUrl = URL.createObjectURL(\n        new Blob([`import ${JSON.stringify(url.href)};`], { type: 'text/javascript' }),\n    );\n    try {\n        return new Worker(_workerBlobUrl, { type: 'module' });\n    } catch (error) {\n        // A page with `worker-src 'self'` (or `script-src` without `blob:`) blocks the\n        // shim, and the direct URL was already refused for its origin — so there is\n        // nothing left to try. Say which of the two walls was hit, because the browser's\n        // own message does not distinguish them.\n        URL.revokeObjectURL(_workerBlobUrl);\n        _workerBlobUrl = null;\n        throw new Error(\n            `@aparte/provider-transformers is served from ${url.origin}, which is not this page's origin, `\n            + 'so its worker has to be started through a blob: URL — and this page\\'s Content-Security-Policy '\n            + 'refuses that. Allow `blob:` in `worker-src` (or `script-src`), or serve the package from your '\n            + `own origin. Original error: ${String(error)}`,\n        );\n    }\n}\n\nfunction _releaseWorkerBlob(): void {\n    if (_workerBlobUrl) {\n        URL.revokeObjectURL(_workerBlobUrl);\n        _workerBlobUrl = null;\n    }\n}\n\n/**\n * Where the page says Transformers.js lives, if it says so at all.\n *\n * The worker cannot ask: an import map is the DOCUMENT's, and by spec it does not reach\n * a worker. The main thread can, and does it the platform's way — `import.meta.resolve`\n * consults that same map — so a page that already maps `@huggingface/transformers` (it\n * has to, to import this package by name at all) is telling us where its copy is. That\n * map is the CDN consumer's manifest: the version pin stays with the consumer, which is\n * the whole point of a peer dependency, and this package invents no second place to say\n * it.\n *\n * `undefined` under a bundler, where the specifier is resolved at build time and the\n * worker's own `import('@huggingface/transformers')` is the path that runs.\n */\nfunction _peerModuleUrl(): string | undefined {\n    const resolve = (import.meta as unknown as { resolve?: (specifier: string) => string }).resolve;\n    if (typeof resolve === 'function') {\n        try {\n            const href = resolve('@huggingface/transformers');\n            if (href && /^https?:/i.test(href)) return href;\n        } catch { /* not in the map — fall through */ }\n    }\n    // Older engines have no `import.meta.resolve`; read the map they do have.\n    try {\n        const el = document.querySelector('script[type=\"importmap\"]');\n        const map = el?.textContent ? JSON.parse(el.textContent) as { imports?: Record<string, string> } : null;\n        const href = map?.imports?.['@huggingface/transformers'];\n        if (href) return new URL(href, location.href).href;\n    } catch { /* no document, or a map that is not JSON */ }\n    return undefined;\n}\n\nfunction _getWorker(): Worker {\n    if (!_worker) {\n        _worker = _spawnWorker();\n        _worker.addEventListener('message', _handleWorkerMessage);\n        _worker.addEventListener('error', _handleWorkerError);\n        _worker.addEventListener('messageerror', _handleWorkerError);\n        // First message, before any work: postMessage keeps order, so the worker has it\n        // by the time a prepare or a generate needs the module.\n        _worker.postMessage({ type: 'init', transformersUrl: _peerModuleUrl() });\n    }\n    return _worker;\n}\n\n/**\n * Worker crashed (uncaught error / WASM init failure / OOM). Reject every in-flight\n * prepare and close every open generate stream so the UI doesn't hang. Subsequent\n * calls rebuild the worker.\n */\nfunction _handleWorkerError(e: Event): void {\n    const message = (e as ErrorEvent)?.message || 'Worker crashed unexpectedly';\n\n    for (const p of _pendingPrepares.values()) {\n        try { p.reject(new Error(message)); } catch { /* ignore */ }\n    }\n    _pendingPrepares.clear();\n\n    for (const ctrl of _pendingGenerates.values()) {\n        try { ctrl.enqueue({ type: 'error' as const, message }); ctrl.close(); }\n        catch { /* ignore */ }\n    }\n    _pendingGenerates.clear();\n\n    // Release every serialization slot so the generate chain doesn't deadlock.\n    for (const resolve of _generateDoneResolvers.values()) {\n        try { resolve(); } catch { /* ignore */ }\n    }\n    _generateDoneResolvers.clear();\n    _generateChain = Promise.resolve();\n    _queuedModelIds.clear();\n\n    _loadedModelId = null;\n    _preparingModelId = null;\n    try { _worker?.terminate(); } catch { /* ignore */ }\n    _worker = null;\n    _releaseWorkerBlob();\n}\n\nfunction _handleWorkerMessage(event: MessageEvent): void {\n    const msg = event.data;\n\n    switch (msg.type) {\n        case 'progress': {\n            const pending = _pendingPrepares.get(msg.id);\n            if (!pending) break;\n            if (msg.status === 'ready') {\n                pending.onProgress({ status: 'ready' });\n                pending.resolve();\n                _pendingPrepares.delete(msg.id);\n            } else if (msg.status === 'loading') {\n                pending.onProgress({ status: 'loading' });\n            } else if (msg.status === 'cached') {\n                pending.onProgress({ status: 'cached', file: msg.file, progress: msg.progress });\n            } else {\n                pending.onProgress({ status: 'downloading', file: msg.file, progress: msg.progress });\n            }\n            break;\n        }\n        case 'prepare-error': {\n            const pending = _pendingPrepares.get(msg.id);\n            if (!pending) break;\n            pending.reject(new Error(msg.message));\n            _pendingPrepares.delete(msg.id);\n            if (_preparingModelId === pending.modelId) _preparingModelId = null;\n            break;\n        }\n        case 'pipeline-ready': {\n            _loadedModelId = msg.modelId;\n            _preparingModelId = null;\n            // Evict models over the cache limit, then refresh the known list.\n            void _enforceMaxCachedModels(msg.modelId).then(() => _refreshKnownModels());\n            break;\n        }\n        case 'gen-chunk': {\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue({ type: msg.chunkType as 'text' | 'thinking', delta: msg.delta });\n            break;\n        }\n        case 'gen-done': {\n            _releaseGenerateSlot(msg.id);\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue({ type: 'done' as const, ...(msg.usage ? { usage: msg.usage } : {}) });\n            ctrl.close();\n            _pendingGenerates.delete(msg.id);\n            break;\n        }\n        case 'gen-error': {\n            _releaseGenerateSlot(msg.id);\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue({ type: 'error' as const, message: msg.message });\n            ctrl.close();\n            _pendingGenerates.delete(msg.id);\n            break;\n        }\n    }\n}\n\n/**\n * Narrowed so the two members the docs tell you to CALL are not optional.\n *\n * `AparteAIProvider` declares `prepareModel` and `getModelStatus` optional (most\n * providers have nothing to download), and widening to it made both\n * possibly-undefined — so the documented `TransformersProvider.prepareModel(...)`\n * needed a `!` or a guard in every strict consumer. Same technique openai-compat\n * already used for its own always-present members.\n *\n * `chat` joined the list once `AparteAIProvider` became a union: it is optional on\n * the format-adapter arm, and this provider IS its `chat()` — running inference\n * locally is the whole package. Narrowing it here says so once, instead of every\n * caller writing `provider.chat!(...)`.\n */\nexport const TransformersProvider: AparteAIProvider\n    & Required<Pick<AparteAIProvider, 'prepareModel' | 'getModelStatus' | 'chat'>> = {\n    id: 'transformers',\n\n    getMetadata() {\n        return {\n            id: 'transformers',\n            name: 'Transformers.js',\n            icon: `<svg viewBox=\"0 0 24 24\" fill=\"none\" xmlns=\"http://www.w3.org/2000/svg\"><path d=\"M12 2L2 7l10 5 10-5-10-5z\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/><path d=\"M2 17l10 5 10-5\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/><path d=\"M2 12l10 5 10-5\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/></svg>`,\n            color: '#f59e0b',\n            description: 'Run LLMs directly in your browser via WebGPU or WASM — no API, no key',\n            hasFreeModels: true,\n            isLocal: true,\n            helpUrl: 'https://huggingface.co/docs/transformers.js',\n        };\n    },\n\n    getModels(): AparteAIModel[] {\n        return _knownModels;\n    },\n\n    async fetchModels(): Promise<AparteAIModel[]> {\n        await _refreshKnownModels();\n        return _knownModels;\n    },\n\n    async chat(request: AparteChatRequest): Promise<AparteChatResponse> {\n        const messages = toMessages(request.messages);\n        const requestId = uuid();\n        const options = {\n            maxTokens: request.maxTokens,\n            temperature: request.temperature,\n            seed: request.seed,\n        };\n        const task = _registeredModels.get(request.modelId)?.task ?? 'text-generation';\n\n        // ── Reserve a serialization slot ─────────────────────────────────────\n        // Chain this generate behind the previous one; the worker has a single\n        // pipeline, so generates MUST NOT overlap.\n        _warnIfContended(request.modelId);\n        _queuedModelIds.set(requestId, request.modelId);\n        const prevGenerate = _generateChain;\n        _generateChain = new Promise<void>((resolveSlot) => {\n            _generateDoneResolvers.set(requestId, resolveSlot);\n        });\n        const postGenerate = (): void => {\n            _getWorker().postMessage({\n                type: 'generate',\n                id: requestId,\n                modelId: request.modelId,\n                messages,\n                options,\n                task,\n                dtype: _registeredModels.get(request.modelId)?.dtype,\n                device: _computeDevice,\n            });\n        };\n\n        if (request.stream === false) {\n            return new Promise<string>((resolve, reject) => {\n                let result = '';\n                const fakeCtrl = {\n                    enqueue: (chunk: { type: string; delta?: string; message?: string }) => {\n                        if (chunk.type === 'text') result += chunk.delta ?? '';\n                        else if (chunk.type === 'done') resolve(result);\n                        else if (chunk.type === 'error') reject(new Error(chunk.message));\n                    },\n                    close: () => { /* no-op */ },\n                } as unknown as ReadableStreamDefaultController;\n                _pendingGenerates.set(requestId, fakeCtrl);\n                void prevGenerate.then(postGenerate);\n            });\n        }\n\n        return new ReadableStream({\n            async start(controller) {\n                _pendingGenerates.set(requestId, controller);\n                await prevGenerate;\n                postGenerate();\n            },\n            cancel() {\n                _pendingGenerates.delete(requestId);\n                // Actually STOP the model (not just detach the reader): tell the worker\n                // to interrupt this generate. The serialization slot is still released\n                // by the resulting gen-done/gen-error, so a queued generate can't start\n                // before the worker has stopped this one.\n                _getWorker().postMessage({ type: 'cancel', id: requestId });\n            },\n        });\n    },\n\n    async getModelStatus(modelId: string): Promise<ModelStatus> {\n        if (_loadedModelId === modelId) return 'ready';\n        if (_preparingModelId === modelId) return 'cached';\n        if ('caches' in globalThis) {\n            try {\n                const encodedId = encodeURIComponent(modelId);\n                const names = await caches.keys();\n                for (const name of names) {\n                    const cache = await caches.open(name);\n                    const keys = await cache.keys();\n                    if (keys.some(r => r.url.includes(encodedId) || r.url.includes(modelId + '/'))) {\n                        return 'cached';\n                    }\n                }\n            } catch {\n                // Cache API unavailable\n            }\n        }\n        return 'not-downloaded';\n    },\n\n    async prepareModel(modelId: string, onProgress: (p: ModelLoadProgress) => void): Promise<void> {\n        if (_loadedModelId === modelId) {\n            onProgress({ status: 'ready' });\n            return;\n        }\n\n        const requestId = uuid();\n        _preparingModelId = modelId;\n\n        const task = _registeredModels.get(modelId)?.task ?? 'text-generation';\n        const dtype = _registeredModels.get(modelId)?.dtype;\n        return new Promise<void>((resolve, reject) => {\n            _pendingPrepares.set(requestId, { modelId, onProgress, resolve, reject });\n            _getWorker().postMessage({ type: 'prepare', id: requestId, modelId, task, dtype, device: _computeDevice });\n        });\n    },\n\n    async deleteModel(modelId: string): Promise<void> {\n        await deleteCachedModel(modelId);\n    },\n};\n\nexport default TransformersProvider;\nexport type { AparteAIProvider, AparteAIModel, ModelStatus, ModelLoadProgress } from '@aparte/core';\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Cache utilities (settings panels, etc.)\n// ─────────────────────────────────────────────────────────────────────────────\n\n/** Returns the modelId currently loaded in the worker's pipeline, or null. */\nexport function getLoadedModelId(): string | null {\n    return _loadedModelId;\n}\n\n/** Terminate the shared worker and reset in-memory state. Safe to call any time. */\nexport function terminateWorker(): void {\n    _worker?.terminate();\n    _worker = null;\n    _releaseWorkerBlob();\n    _loadedModelId = null;\n    _preparingModelId = null;\n    for (const [, p] of _pendingPrepares) {\n        p.reject(new Error('Worker terminated'));\n    }\n    _pendingPrepares.clear();\n    for (const [, ctrl] of _pendingGenerates) {\n        try { ctrl.enqueue({ type: 'error' as const, message: 'Worker terminated' }); ctrl.close(); } catch { /* already closed */ }\n    }\n    _pendingGenerates.clear();\n\n    // Release every serialization slot and reset the chain — the same three lines\n    // the worker-error handler above already carried, with the same reason. Without\n    // them, terminating mid-generate left `_generateChain` pending on a resolver\n    // that had just been dropped, so the NEXT chat() awaited a promise that could\n    // never settle: no error, no rejection, the stream simply never started again\n    // for the life of the page.\n    for (const resolve of _generateDoneResolvers.values()) {\n        try { resolve(); } catch { /* ignore */ }\n    }\n    _generateDoneResolvers.clear();\n    _generateChain = Promise.resolve();\n    _queuedModelIds.clear();\n    // A terminated worker is a fresh situation; let the contention warning speak again.\n    _warnedModelContention = false;\n}\n\nexport interface CachedModelEntry {\n    modelId: string;\n    name: string;\n    /** Total size in bytes of all cached files for this model. -1 if unknown. */\n    sizeBytes: number;\n    /** True if the model is currently loaded in the worker. */\n    loaded: boolean;\n}\n\n/**\n * Scan the Cache API to find which Transformers.js models have been downloaded,\n * by matching cache entry URLs against the Hugging Face resolve path.\n */\nexport async function listCachedModels(): Promise<CachedModelEntry[]> {\n    if (!('caches' in globalThis)) return [];\n\n    const found = new Map<string, { name: string; sizeBytes: number }>();\n\n    // e.g. https://huggingface.co/onnx-community/Qwen2.5-0.5B/resolve/main/config.json\n    //   → onnx-community/Qwen2.5-0.5B\n    function extractModelId(url: string): string | null {\n        const m = url.match(/huggingface\\.co\\/([^/]+\\/[^/]+)\\/resolve\\//);\n        return m ? decodeURIComponent(m[1]!) : null;\n    }\n\n    function modelName(modelId: string): string {\n        const config = _registeredModels.get(modelId);\n        if (config) return config.name;\n        return (modelId.split('/').pop() ?? modelId).replace(/-/g, ' ');\n    }\n\n    try {\n        const cacheNames = await caches.keys();\n        await Promise.all(cacheNames.map(async (cacheName) => {\n            try {\n                const cache = await caches.open(cacheName);\n                const requests = await cache.keys();\n                for (const req of requests) {\n                    const modelId = extractModelId(req.url);\n                    if (!modelId) continue;\n                    if (!found.has(modelId)) {\n                        found.set(modelId, { name: modelName(modelId), sizeBytes: 0 });\n                    }\n                    const response = await cache.match(req);\n                    if (!response) continue;\n                    const contentLength = response.headers.get('content-length');\n                    if (contentLength) {\n                        found.get(modelId)!.sizeBytes += parseInt(contentLength, 10);\n                    } else {\n                        try {\n                            const blob = await response.clone().blob();\n                            found.get(modelId)!.sizeBytes += blob.size;\n                        } catch { /* skip */ }\n                    }\n                }\n            } catch { /* skip inaccessible cache */ }\n        }));\n    } catch {\n        return [];\n    }\n\n    return Array.from(found.entries()).map(([modelId, { name, sizeBytes }]) => ({\n        modelId,\n        name,\n        sizeBytes,\n        loaded: _loadedModelId === modelId,\n    }));\n}\n\n/**\n * Delete all cached files for a modelId from the Cache API, terminating the worker\n * first if that model is currently loaded.\n */\nexport async function deleteCachedModel(modelId: string): Promise<void> {\n    if (_loadedModelId === modelId || _preparingModelId === modelId) {\n        terminateWorker();\n    }\n    if (!('caches' in globalThis)) return;\n    try {\n        const cacheNames = await caches.keys();\n        await Promise.all(cacheNames.map(async (cacheName) => {\n            try {\n                const cache = await caches.open(cacheName);\n                const requests = await cache.keys();\n                const encoded = encodeURIComponent(modelId);\n                await Promise.all(\n                    requests\n                        .filter(r => r.url.includes(modelId) || r.url.includes(encoded))\n                        .map(r => cache.delete(r)),\n                );\n            } catch { /* skip */ }\n        }));\n    } catch { /* Cache API unavailable */ }\n}\n"],"names":[],"mappings":";;AAkEA,IAAI,iBAAqE;AAMlE,SAAS,sBAAsB,OAA0D;AAC5F,mBAAiB;AACrB;AAEA,eAAsB,iBAA2C;AAG7D,QAAM,QAAiB,UAAmD,gBAAgB;AAG1F,MAAI,SAAS;AACb,MAAI,SAAS,WAAW;AACpB,QAAI;AACA,YAAM,UAAU,MAAO,UAAyE,IAAI,eAAA;AACpG,eAAS,YAAY;AAAA,IACzB,QAAQ;AACJ,eAAS;AAAA,IACb;AAAA,EACJ;AAEA,MAAI;AACJ,MAAI,CAAC,UAAU,QAAQ,GAAG;AACtB,WAAO;AAAA,EACX,WAAW,QAAQ,GAAG;AAClB,WAAO;AAAA,EACX,OAAO;AACH,WAAO;AAAA,EACX;AAEA,QAAM,qBAAqB,iBACpB,eAAe,IAAI,KAAK,eAAe,QAAQ,KAChD;AAEN,SAAO,EAAE,QAAQ,OAAO,MAAM,mBAAA;AAClC;AAsBA,MAAM,wCAAwB,IAAA;AAG9B,IAAI,eAAgC,CAAA;AAK7B,SAAS,cAAc,QAAuC;AACjE,oBAAkB,IAAI,OAAO,IAAI,MAAM;AACvC,MAAI,CAAC,aAAa,KAAK,CAAA,MAAK,EAAE,OAAO,OAAO,EAAE,GAAG;AAC7C,mBAAe,CAAC,GAAG,cAAc;AAAA,MAC7B,IAAI,OAAO;AAAA,MACX,MAAM,OAAO;AAAA,MACb,aAAa,OAAO;AAAA,MACpB,cAAc,OAAO;AAAA,IAAA,CACxB;AAAA,EACL;AACJ;AAGA,SAAS,qBAAqB,SAAgC;AAC1D,QAAM,SAAS,kBAAkB,IAAI,OAAO;AAC5C,MAAI,OAAQ,QAAO,EAAE,IAAI,OAAO,IAAI,MAAM,OAAO,MAAM,aAAa,OAAO,aAAa,cAAc,OAAO,aAAA;AAC7G,QAAM,QAAQ,QAAQ,MAAM,GAAG,EAAE,SAAS,SAAS,QAAQ,MAAM,GAAG;AACpE,SAAO,EAAE,IAAI,SAAS,MAAM,cAAc,CAAC,WAAW,EAAA;AAC1D;AAGA,IAAI,mBAAmB;AAMhB,SAAS,mBAAmB,KAAmB;AAClD,qBAAmB;AACvB;AAGO,SAAS,qBAA6B;AACzC,SAAO;AACX;AASA,IAAI,iBAAgC;AAE7B,SAAS,iBAAiB,GAAwB;AACrD,mBAAiB;AACrB;AAEO,SAAS,mBAAkC;AAC9C,SAAO;AACX;AAGA,eAAe,wBAAwB,aAAoC;AACvE,MAAI,qBAAqB,EAAG;AAC5B,MAAI;AACA,UAAM,SAAS,MAAM,iBAAA;AACrB,UAAM,SAAS,OAAO,OAAO,CAAA,MAAK,EAAE,YAAY,WAAW;AAC3D,UAAM,SAAS,OAAO,SAAS;AAC/B,QAAI,UAAU,EAAG;AAEjB,aAAS,IAAI,GAAG,IAAI,UAAU,IAAI,OAAO,QAAQ,KAAK;AAClD,YAAM,kBAAkB,OAAO,CAAC,EAAG,OAAO;AAAA,IAC9C;AAAA,EACJ,QAAQ;AAAA,EAA0B;AACtC;AAGA,eAAe,sBAAqC;AAChD,MAAI;AACA,UAAM,SAAS,MAAM,iBAAA;AACrB,eAAW,SAAS,QAAQ;AACxB,UAAI,CAAC,aAAa,KAAK,CAAA,MAAK,EAAE,OAAO,MAAM,OAAO,GAAG;AACjD,uBAAe,CAAC,GAAG,cAAc,qBAAqB,MAAM,OAAO,CAAC;AAAA,MACxE;AAAA,IACJ;AAAA,EACJ,QAAQ;AAAA,EAA0B;AACtC;AAGA,IAAI,0BAA0B;AAG9B,SAAS,WAAW,UAAgD;AAChE,QAAM,SAA0B,CAAA;AAChC,MAAI,mBAAmB;AACvB,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,UAAU,EAAE,SAAS,eAAe,EAAE,SAAS,UAAU;AACpE,YAAM,OAAO,cAAc,EAAE,OAAO;AACpC,UAAI,aAAa,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAM;AAAA,IACzD,OAAO;AAKH;AAAA,IACJ;AAAA,EACJ;AACA,MAAI,mBAAmB,KAAK,CAAC,yBAAyB;AAClD,8BAA0B;AAC1B,YAAQ;AAAA,MACJ,0BAA0B,gBAAgB;AAAA,IAAA;AAAA,EAIlD;AACA,SAAO;AACX;AAMA,IAAI,UAAyB;AAQ7B,MAAM,uCAAuB,IAAA;AAC7B,MAAM,wCAAwB,IAAA;AAM9B,IAAI,iBAAgC,QAAQ,QAAA;AAC5C,MAAM,6CAA6B,IAAA;AAOnC,MAAM,sCAAsB,IAAA;AAC5B,IAAI,yBAAyB;AAG7B,SAAS,mBAAmB,WAAuC;AAC/D,aAAW,MAAM,gBAAgB,OAAA,EAAU,KAAI,OAAO,UAAW,QAAO;AACxE,SAAO;AACX;AAOA,SAAS,iBAAiB,WAAyB;AAC/C,MAAI,uBAAwB;AAC5B,QAAM,QAAQ,mBAAmB,SAAS;AAC1C,MAAI,CAAC,MAAO;AACZ,2BAAyB;AACzB,UAAQ;AAAA,IACJ,mEAAmE,SAAS,aACtE,KAAK,gHACiC,gBAAgB;AAAA,EAAA;AAIpE;AAGA,SAAS,qBAAqB,IAAkB;AAC5C,kBAAgB,OAAO,EAAE;AACzB,QAAM,UAAU,uBAAuB,IAAI,EAAE;AAC7C,MAAI,SAAS;AACT,2BAAuB,OAAO,EAAE;AAChC,YAAA;AAAA,EACJ;AACJ;AAGA,IAAI,iBAAgC;AAEpC,IAAI,oBAAmC;AAGvC,IAAI,iBAAgC;AAmBpC,SAAS,eAAuB;AAC5B,QAAM,MAAM,IAAI,IAAI,WAAW,YAAY,GAAG;AAC9C,QAAM,aAAa,OAAO,aAAa,eAAe,IAAI,WAAW,SAAS;AAM9E,QAAM,cAAc,OAAO,SAAS,cAAc,OAAO,IAAI,oBAAoB;AAOjF,MAAI,cAAc,CAAC,YAAa,QAAO,IAAI,OAAO,IAAA;AAAA;AAAA,IAAA,KAAA,IAAA,IAAA,6BAAA,YAAA,GAAA,EAAA;AAAA,IAAA,YAAA;AAAA,EAAA,GAAyC,EAAE,MAAM,UAAU;AAE7G,mBAAiB,IAAI;AAAA,IACjB,IAAI,KAAK,CAAC,UAAU,KAAK,UAAU,IAAI,IAAI,CAAC,GAAG,GAAG,EAAE,MAAM,mBAAmB;AAAA,EAAA;AAEjF,MAAI;AACA,WAAO,IAAI,OAAO,gBAAgB,EAAE,MAAM,UAAU;AAAA,EACxD,SAAS,OAAO;AAKZ,QAAI,gBAAgB,cAAc;AAClC,qBAAiB;AACjB,UAAM,IAAI;AAAA,MACN,gDAAgD,IAAI,MAAM,oQAGzB,OAAO,KAAK,CAAC;AAAA,IAAA;AAAA,EAEtD;AACJ;AAEA,SAAS,qBAA2B;AAChC,MAAI,gBAAgB;AAChB,QAAI,gBAAgB,cAAc;AAClC,qBAAiB;AAAA,EACrB;AACJ;AAgBA,SAAS,iBAAqC;AAC1C,QAAM,UAAW,YAAuE;AACxF,MAAI,OAAO,YAAY,YAAY;AAC/B,QAAI;AACA,YAAM,OAAO,QAAQ,2BAA2B;AAChD,UAAI,QAAQ,YAAY,KAAK,IAAI,EAAG,QAAO;AAAA,IAC/C,QAAQ;AAAA,IAAsC;AAAA,EAClD;AAEA,MAAI;AACA,UAAM,KAAK,SAAS,cAAc,0BAA0B;AAC5D,UAAM,MAAM,IAAI,cAAc,KAAK,MAAM,GAAG,WAAW,IAA4C;AACnG,UAAM,OAAO,KAAK,UAAU,2BAA2B;AACvD,QAAI,KAAM,QAAO,IAAI,IAAI,MAAM,SAAS,IAAI,EAAE;AAAA,EAClD,QAAQ;AAAA,EAA+C;AACvD,SAAO;AACX;AAEA,SAAS,aAAqB;AAC1B,MAAI,CAAC,SAAS;AACV,cAAU,aAAA;AACV,YAAQ,iBAAiB,WAAW,oBAAoB;AACxD,YAAQ,iBAAiB,SAAS,kBAAkB;AACpD,YAAQ,iBAAiB,gBAAgB,kBAAkB;AAG3D,YAAQ,YAAY,EAAE,MAAM,QAAQ,iBAAiB,eAAA,GAAkB;AAAA,EAC3E;AACA,SAAO;AACX;AAOA,SAAS,mBAAmB,GAAgB;AACxC,QAAM,UAAW,GAAkB,WAAW;AAE9C,aAAW,KAAK,iBAAiB,UAAU;AACvC,QAAI;AAAE,QAAE,OAAO,IAAI,MAAM,OAAO,CAAC;AAAA,IAAG,QAAQ;AAAA,IAAe;AAAA,EAC/D;AACA,mBAAiB,MAAA;AAEjB,aAAW,QAAQ,kBAAkB,UAAU;AAC3C,QAAI;AAAE,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS;AAAG,WAAK,MAAA;AAAA,IAAS,QACjE;AAAA,IAAe;AAAA,EACzB;AACA,oBAAkB,MAAA;AAGlB,aAAW,WAAW,uBAAuB,UAAU;AACnD,QAAI;AAAE,cAAA;AAAA,IAAW,QAAQ;AAAA,IAAe;AAAA,EAC5C;AACA,yBAAuB,MAAA;AACvB,mBAAiB,QAAQ,QAAA;AACzB,kBAAgB,MAAA;AAEhB,mBAAiB;AACjB,sBAAoB;AACpB,MAAI;AAAE,aAAS,UAAA;AAAA,EAAa,QAAQ;AAAA,EAAe;AACnD,YAAU;AACV,qBAAA;AACJ;AAEA,SAAS,qBAAqB,OAA2B;AACrD,QAAM,MAAM,MAAM;AAElB,UAAQ,IAAI,MAAA;AAAA,IACR,KAAK,YAAY;AACb,YAAM,UAAU,iBAAiB,IAAI,IAAI,EAAE;AAC3C,UAAI,CAAC,QAAS;AACd,UAAI,IAAI,WAAW,SAAS;AACxB,gBAAQ,WAAW,EAAE,QAAQ,QAAA,CAAS;AACtC,gBAAQ,QAAA;AACR,yBAAiB,OAAO,IAAI,EAAE;AAAA,MAClC,WAAW,IAAI,WAAW,WAAW;AACjC,gBAAQ,WAAW,EAAE,QAAQ,UAAA,CAAW;AAAA,MAC5C,WAAW,IAAI,WAAW,UAAU;AAChC,gBAAQ,WAAW,EAAE,QAAQ,UAAU,MAAM,IAAI,MAAM,UAAU,IAAI,SAAA,CAAU;AAAA,MACnF,OAAO;AACH,gBAAQ,WAAW,EAAE,QAAQ,eAAe,MAAM,IAAI,MAAM,UAAU,IAAI,SAAA,CAAU;AAAA,MACxF;AACA;AAAA,IACJ;AAAA,IACA,KAAK,iBAAiB;AAClB,YAAM,UAAU,iBAAiB,IAAI,IAAI,EAAE;AAC3C,UAAI,CAAC,QAAS;AACd,cAAQ,OAAO,IAAI,MAAM,IAAI,OAAO,CAAC;AACrC,uBAAiB,OAAO,IAAI,EAAE;AAC9B,UAAI,sBAAsB,QAAQ,QAAS,qBAAoB;AAC/D;AAAA,IACJ;AAAA,IACA,KAAK,kBAAkB;AACnB,uBAAiB,IAAI;AACrB,0BAAoB;AAEpB,WAAK,wBAAwB,IAAI,OAAO,EAAE,KAAK,MAAM,qBAAqB;AAC1E;AAAA,IACJ;AAAA,IACA,KAAK,aAAa;AACd,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,EAAE,MAAM,IAAI,WAAkC,OAAO,IAAI,OAAO;AAC7E;AAAA,IACJ;AAAA,IACA,KAAK,YAAY;AACb,2BAAqB,IAAI,EAAE;AAC3B,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,EAAE,MAAM,QAAiB,GAAI,IAAI,QAAQ,EAAE,OAAO,IAAI,MAAA,IAAU,CAAA,GAAK;AAClF,WAAK,MAAA;AACL,wBAAkB,OAAO,IAAI,EAAE;AAC/B;AAAA,IACJ;AAAA,IACA,KAAK,aAAa;AACd,2BAAqB,IAAI,EAAE;AAC3B,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS,IAAI,SAAS;AAC7D,WAAK,MAAA;AACL,wBAAkB,OAAO,IAAI,EAAE;AAC/B;AAAA,IACJ;AAAA,EAAA;AAER;AAgBO,MAAM,uBACwE;AAAA,EACjF,IAAI;AAAA,EAEJ,cAAc;AACV,WAAO;AAAA,MACH,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,MAAM;AAAA,MACN,OAAO;AAAA,MACP,aAAa;AAAA,MACb,eAAe;AAAA,MACf,SAAS;AAAA,MACT,SAAS;AAAA,IAAA;AAAA,EAEjB;AAAA,EAEA,YAA6B;AACzB,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,cAAwC;AAC1C,UAAM,oBAAA;AACN,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,KAAK,SAAyD;AAChE,UAAM,WAAW,WAAW,QAAQ,QAAQ;AAC5C,UAAM,YAAY,KAAA;AAClB,UAAM,UAAU;AAAA,MACZ,WAAW,QAAQ;AAAA,MACnB,aAAa,QAAQ;AAAA,MACrB,MAAM,QAAQ;AAAA,IAAA;AAElB,UAAM,OAAO,kBAAkB,IAAI,QAAQ,OAAO,GAAG,QAAQ;AAK7D,qBAAiB,QAAQ,OAAO;AAChC,oBAAgB,IAAI,WAAW,QAAQ,OAAO;AAC9C,UAAM,eAAe;AACrB,qBAAiB,IAAI,QAAc,CAAC,gBAAgB;AAChD,6BAAuB,IAAI,WAAW,WAAW;AAAA,IACrD,CAAC;AACD,UAAM,eAAe,MAAY;AAC7B,iBAAA,EAAa,YAAY;AAAA,QACrB,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,SAAS,QAAQ;AAAA,QACjB;AAAA,QACA;AAAA,QACA;AAAA,QACA,OAAO,kBAAkB,IAAI,QAAQ,OAAO,GAAG;AAAA,QAC/C,QAAQ;AAAA,MAAA,CACX;AAAA,IACL;AAEA,QAAI,QAAQ,WAAW,OAAO;AAC1B,aAAO,IAAI,QAAgB,CAAC,SAAS,WAAW;AAC5C,YAAI,SAAS;AACb,cAAM,WAAW;AAAA,UACb,SAAS,CAAC,UAA8D;AACpE,gBAAI,MAAM,SAAS,OAAQ,WAAU,MAAM,SAAS;AAAA,qBAC3C,MAAM,SAAS,OAAQ,SAAQ,MAAM;AAAA,qBACrC,MAAM,SAAS,QAAS,QAAO,IAAI,MAAM,MAAM,OAAO,CAAC;AAAA,UACpE;AAAA,UACA,OAAO,MAAM;AAAA,UAAc;AAAA,QAAA;AAE/B,0BAAkB,IAAI,WAAW,QAAQ;AACzC,aAAK,aAAa,KAAK,YAAY;AAAA,MACvC,CAAC;AAAA,IACL;AAEA,WAAO,IAAI,eAAe;AAAA,MACtB,MAAM,MAAM,YAAY;AACpB,0BAAkB,IAAI,WAAW,UAAU;AAC3C,cAAM;AACN,qBAAA;AAAA,MACJ;AAAA,MACA,SAAS;AACL,0BAAkB,OAAO,SAAS;AAKlC,mBAAA,EAAa,YAAY,EAAE,MAAM,UAAU,IAAI,WAAW;AAAA,MAC9D;AAAA,IAAA,CACH;AAAA,EACL;AAAA,EAEA,MAAM,eAAe,SAAuC;AACxD,QAAI,mBAAmB,QAAS,QAAO;AACvC,QAAI,sBAAsB,QAAS,QAAO;AAC1C,QAAI,YAAY,YAAY;AACxB,UAAI;AACA,cAAM,YAAY,mBAAmB,OAAO;AAC5C,cAAM,QAAQ,MAAM,OAAO,KAAA;AAC3B,mBAAW,QAAQ,OAAO;AACtB,gBAAM,QAAQ,MAAM,OAAO,KAAK,IAAI;AACpC,gBAAM,OAAO,MAAM,MAAM,KAAA;AACzB,cAAI,KAAK,KAAK,CAAA,MAAK,EAAE,IAAI,SAAS,SAAS,KAAK,EAAE,IAAI,SAAS,UAAU,GAAG,CAAC,GAAG;AAC5E,mBAAO;AAAA,UACX;AAAA,QACJ;AAAA,MACJ,QAAQ;AAAA,MAER;AAAA,IACJ;AACA,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,aAAa,SAAiB,YAA2D;AAC3F,QAAI,mBAAmB,SAAS;AAC5B,iBAAW,EAAE,QAAQ,SAAS;AAC9B;AAAA,IACJ;AAEA,UAAM,YAAY,KAAA;AAClB,wBAAoB;AAEpB,UAAM,OAAO,kBAAkB,IAAI,OAAO,GAAG,QAAQ;AACrD,UAAM,QAAQ,kBAAkB,IAAI,OAAO,GAAG;AAC9C,WAAO,IAAI,QAAc,CAAC,SAAS,WAAW;AAC1C,uBAAiB,IAAI,WAAW,EAAE,SAAS,YAAY,SAAS,QAAQ;AACxE,iBAAA,EAAa,YAAY,EAAE,MAAM,WAAW,IAAI,WAAW,SAAS,MAAM,OAAO,QAAQ,eAAA,CAAgB;AAAA,IAC7G,CAAC;AAAA,EACL;AAAA,EAEA,MAAM,YAAY,SAAgC;AAC9C,UAAM,kBAAkB,OAAO;AAAA,EACnC;AACJ;AAUO,SAAS,mBAAkC;AAC9C,SAAO;AACX;AAGO,SAAS,kBAAwB;AACpC,WAAS,UAAA;AACT,YAAU;AACV,qBAAA;AACA,mBAAiB;AACjB,sBAAoB;AACpB,aAAW,CAAA,EAAG,CAAC,KAAK,kBAAkB;AAClC,MAAE,OAAO,IAAI,MAAM,mBAAmB,CAAC;AAAA,EAC3C;AACA,mBAAiB,MAAA;AACjB,aAAW,CAAA,EAAG,IAAI,KAAK,mBAAmB;AACtC,QAAI;AAAE,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS,qBAAqB;AAAG,WAAK,MAAA;AAAA,IAAS,QAAQ;AAAA,IAAuB;AAAA,EAC/H;AACA,oBAAkB,MAAA;AAQlB,aAAW,WAAW,uBAAuB,UAAU;AACnD,QAAI;AAAE,cAAA;AAAA,IAAW,QAAQ;AAAA,IAAe;AAAA,EAC5C;AACA,yBAAuB,MAAA;AACvB,mBAAiB,QAAQ,QAAA;AACzB,kBAAgB,MAAA;AAEhB,2BAAyB;AAC7B;AAeA,eAAsB,mBAAgD;AAClE,MAAI,EAAE,YAAY,YAAa,QAAO,CAAA;AAEtC,QAAM,4BAAY,IAAA;AAIlB,WAAS,eAAe,KAA4B;AAChD,UAAM,IAAI,IAAI,MAAM,4CAA4C;AAChE,WAAO,IAAI,mBAAmB,EAAE,CAAC,CAAE,IAAI;AAAA,EAC3C;AAEA,WAAS,UAAU,SAAyB;AACxC,UAAM,SAAS,kBAAkB,IAAI,OAAO;AAC5C,QAAI,eAAe,OAAO;AAC1B,YAAQ,QAAQ,MAAM,GAAG,EAAE,SAAS,SAAS,QAAQ,MAAM,GAAG;AAAA,EAClE;AAEA,MAAI;AACA,UAAM,aAAa,MAAM,OAAO,KAAA;AAChC,UAAM,QAAQ,IAAI,WAAW,IAAI,OAAO,cAAc;AAClD,UAAI;AACA,cAAM,QAAQ,MAAM,OAAO,KAAK,SAAS;AACzC,cAAM,WAAW,MAAM,MAAM,KAAA;AAC7B,mBAAW,OAAO,UAAU;AACxB,gBAAM,UAAU,eAAe,IAAI,GAAG;AACtC,cAAI,CAAC,QAAS;AACd,cAAI,CAAC,MAAM,IAAI,OAAO,GAAG;AACrB,kBAAM,IAAI,SAAS,EAAE,MAAM,UAAU,OAAO,GAAG,WAAW,GAAG;AAAA,UACjE;AACA,gBAAM,WAAW,MAAM,MAAM,MAAM,GAAG;AACtC,cAAI,CAAC,SAAU;AACf,gBAAM,gBAAgB,SAAS,QAAQ,IAAI,gBAAgB;AAC3D,cAAI,eAAe;AACf,kBAAM,IAAI,OAAO,EAAG,aAAa,SAAS,eAAe,EAAE;AAAA,UAC/D,OAAO;AACH,gBAAI;AACA,oBAAM,OAAO,MAAM,SAAS,MAAA,EAAQ,KAAA;AACpC,oBAAM,IAAI,OAAO,EAAG,aAAa,KAAK;AAAA,YAC1C,QAAQ;AAAA,YAAa;AAAA,UACzB;AAAA,QACJ;AAAA,MACJ,QAAQ;AAAA,MAAgC;AAAA,IAC5C,CAAC,CAAC;AAAA,EACN,QAAQ;AACJ,WAAO,CAAA;AAAA,EACX;AAEA,SAAO,MAAM,KAAK,MAAM,QAAA,CAAS,EAAE,IAAI,CAAC,CAAC,SAAS,EAAE,MAAM,UAAA,CAAW,OAAO;AAAA,IACxE;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,mBAAmB;AAAA,EAAA,EAC7B;AACN;AAMA,eAAsB,kBAAkB,SAAgC;AACpE,MAAI,mBAAmB,WAAW,sBAAsB,SAAS;AAC7D,oBAAA;AAAA,EACJ;AACA,MAAI,EAAE,YAAY,YAAa;AAC/B,MAAI;AACA,UAAM,aAAa,MAAM,OAAO,KAAA;AAChC,UAAM,QAAQ,IAAI,WAAW,IAAI,OAAO,cAAc;AAClD,UAAI;AACA,cAAM,QAAQ,MAAM,OAAO,KAAK,SAAS;AACzC,cAAM,WAAW,MAAM,MAAM,KAAA;AAC7B,cAAM,UAAU,mBAAmB,OAAO;AAC1C,cAAM,QAAQ;AAAA,UACV,SACK,OAAO,CAAA,MAAK,EAAE,IAAI,SAAS,OAAO,KAAK,EAAE,IAAI,SAAS,OAAO,CAAC,EAC9D,IAAI,OAAK,MAAM,OAAO,CAAC,CAAC;AAAA,QAAA;AAAA,MAErC,QAAQ;AAAA,MAAa;AAAA,IACzB,CAAC,CAAC;AAAA,EACN,QAAQ;AAAA,EAA8B;AAC1C;"}
+{"version":3,"file":"index.js","sources":["../src/index.ts"],"sourcesContent":["/**\n * @aparte/provider-transformers — run LLMs 100% in the browser via Transformers.js.\n *\n * A local, keyless `AparteAIProvider`: it owns its I/O (inference runs off the main\n * thread in a Web Worker) so `AparteDirectTransport` delegates to its `chat()`. Model\n * weights download once and persist in the Cache API.\n *\n * Scope: the worker runs a **runner** — the built-in `text-generation` (any chat model\n * behind Transformers.js' `pipeline()`), or a module of the app's own named by\n * `TransformersModelConfig.runner` (see `runners/types.ts` for the contract). Tool-calling\n * for local models is model-specific (every family has its own wire format), so the\n * built-in drops tool turns and says so; a custom runner may render them.\n *\n * ## This provider's state is TAB-scoped, on purpose\n *\n * Everything below the \"Worker bridge\" heading — the worker, the loaded model, the\n * generate chain — plus `setComputeDevice`, `setMaxCachedModels` and\n * `setHardwareTierModels`, is module-level and therefore shared by every chat on the\n * page. That is deliberate, and it is the opposite of what the rest of the suite does:\n * a plugin's providers scope to one chat, this one cannot.\n *\n * The reason is the resource, not the design. A local model is 1–2 GB of weights and one\n * WebGPU pipeline. Handing each chat its own worker would mean N copies resident in one\n * tab — which is the failure this package exists to avoid, not a capability. The\n * settings above describe the *machine* (which backend, how many models to keep\n * cached), so per-chat values would not mean anything either.\n *\n * What the constraint costs: two chats on the page driving DIFFERENT local models take\n * turns on one pipeline, so each turn may evict and reload gigabytes. That used to\n * happen silently — a multi-second stall with nothing to read. It now warns once, from\n * `chat()`, when a generate is queued for a model other than the one already in flight.\n * Same model in both chats is free and correct: they share the load.\n */\n\nimport type {\n    AparteAIProvider,\n    AparteAIModel,\n    AparteChatRequest,\n    AparteChatResponse,\n    ModelStatus,\n    ModelLoadProgress,\n} from '@aparte/core';\nimport { uuid } from '@aparte/core';\nimport type { BuiltInRunner, Device, Dtype } from './runners/types.js';\n\n// The worker's URL, not the worker itself: this package constructs it by hand because a\n// cross-origin copy has to go through a blob (see `_spawnWorker`). `?worker&url` is what\n// keeps Vite emitting the worker as its own chunk — the `new Worker(new URL(...))` form\n// it detects by pattern was the only other way, and moving the URL out of that call made\n// the build inline the worker's raw TypeScript as a data: URL instead. Caught by a\n// two-origin browser probe, not by any test.\nimport workerUrl from './worker.ts?worker&url';\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Hardware detection\n// ─────────────────────────────────────────────────────────────────────────────\n\nexport interface HardwareProfile {\n    hasGpu: boolean;\n    ramGb: number;\n    tier: 'low' | 'mid' | 'high';\n    recommendedModelId: string;\n}\n\n/** Hardware-tier model overrides — set by the app via setHardwareTierModels(). */\nlet _hardwareTiers: { low: string; mid?: string; high: string } | null = null;\n\n/**\n * Set the model IDs to use per hardware tier. Call before detectHardware() is used\n * to pick a default model — the provider ships no model knowledge of its own.\n */\nexport function setHardwareTierModels(tiers: { low: string; mid?: string; high: string }): void {\n    _hardwareTiers = tiers;\n}\n\nexport async function detectHardware(): Promise<HardwareProfile> {\n    // navigator.deviceMemory: W3C API, Chromium only, capped at 8 GB for privacy\n    // (1 | 2 | 4 | 8). Falls back to 4 on Firefox/Safari.\n    const ramGb: number = (navigator as unknown as { deviceMemory?: number }).deviceMemory ?? 4;\n\n    // Real WebGPU check: requestAdapter() returns null if no capable GPU is present.\n    let hasGpu = false;\n    if ('gpu' in navigator) {\n        try {\n            const adapter = await (navigator as unknown as { gpu: { requestAdapter(): Promise<unknown> } }).gpu.requestAdapter();\n            hasGpu = adapter !== null;\n        } catch {\n            hasGpu = false;\n        }\n    }\n\n    let tier: 'low' | 'mid' | 'high';\n    if (!hasGpu || ramGb < 4) {\n        tier = 'low';\n    } else if (ramGb < 8) {\n        tier = 'mid';\n    } else {\n        tier = 'high';\n    }\n\n    const recommendedModelId = _hardwareTiers\n        ? (_hardwareTiers[tier] ?? _hardwareTiers.high ?? '')\n        : '';\n\n    return { hasGpu, ramGb, tier, recommendedModelId };\n}\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Model catalog — all model knowledge lives in the app, not the provider.\n// ─────────────────────────────────────────────────────────────────────────────\n\n/** Configuration for a model registered with the provider. */\nexport interface TransformersModelConfig {\n    id: string;\n    name: string;\n    description?: string;\n    capabilities: AparteAIModel['capabilities'];\n    /**\n     * Which built-in runner loads and drives the model. `'text-generation'` (the default)\n     * is any chat model behind Transformers.js' `pipeline()`. Ignored when `runner` is set.\n     */\n    task?: BuiltInRunner;\n    /**\n     * A runner of your own: the URL of an ES module exporting `createRunner` (see\n     * `TransformersRunner`). Resolved against the page, imported by the worker, and handed\n     * the same Transformers.js instance the built-ins use. Wins over `task`.\n     */\n    runner?: string;\n    /** ONNX dtype or per-part dtype map (e.g. `'q4'` or `{ decoder_model_merged: 'q4' }`). */\n    dtype?: Dtype;\n    /** Preferred device. Defaults to WebGPU when available, else WASM. */\n    device?: Device;\n    metadata?: Record<string, unknown>;\n}\n\n/** Models registered by the app via registerModel(). */\nconst _registeredModels = new Map<string, TransformersModelConfig>();\n\n/** Mutable model list — populated by registerModel() and cache discovery. */\nlet _knownModels: AparteAIModel[] = [];\n\n/**\n * Register a model with the provider. Call before the model is used for inference.\n */\nexport function registerModel(config: TransformersModelConfig): void {\n    _registeredModels.set(config.id, config);\n    if (!_knownModels.find(m => m.id === config.id)) {\n        _knownModels = [..._knownModels, {\n            id: config.id,\n            name: config.name,\n            description: config.description,\n            capabilities: config.capabilities,\n        }];\n    }\n}\n\n/** Build an AparteAIModel entry from a cache-discovered modelId not in the registry. */\nfunction _modelFromCacheEntry(modelId: string): AparteAIModel {\n    const config = _registeredModels.get(modelId);\n    if (config) return { id: config.id, name: config.name, description: config.description, capabilities: config.capabilities };\n    const name = (modelId.split('/').pop() ?? modelId).replace(/-/g, ' ');\n    return { id: modelId, name, capabilities: ['streaming'] };\n}\n\n/** Max number of models to keep in cache. 0 = unlimited. Default: 1. */\nlet _maxCachedModels = 1;\n\n/**\n * Set the maximum number of models to keep in cache. When exceeded after a new\n * model is ready, the oldest models are evicted. 0 = unlimited.\n */\nexport function setMaxCachedModels(max: number): void {\n    _maxCachedModels = max;\n}\n\n/** Returns the current max-cached-models setting. */\nexport function getMaxCachedModels(): number {\n    return _maxCachedModels;\n}\n\n/**\n * User's preferred compute backend for local inference.\n *   'auto'   → WebGPU when available, else WASM (default)\n *   'webgpu' → force WebGPU\n *   'wasm'   → force WASM CPU\n */\nexport type ComputeDevice = 'auto' | 'webgpu' | 'wasm';\nlet _computeDevice: ComputeDevice = 'auto';\n\nexport function setComputeDevice(d: ComputeDevice): void {\n    _computeDevice = d;\n}\n\nexport function getComputeDevice(): ComputeDevice {\n    return _computeDevice;\n}\n\n/** Evict models from cache until count <= _maxCachedModels; `keepModelId` is never evicted. */\nasync function _enforceMaxCachedModels(keepModelId: string): Promise<void> {\n    if (_maxCachedModels === 0) return; // unlimited\n    try {\n        const cached = await listCachedModels();\n        const others = cached.filter(e => e.modelId !== keepModelId);\n        const excess = cached.length - _maxCachedModels;\n        if (excess <= 0) return;\n        // Delete the excess models (oldest first — they appear first in cache scan order).\n        for (let i = 0; i < excess && i < others.length; i++) {\n            await deleteCachedModel(others[i]!.modelId);\n        }\n    } catch { /* cache unavailable */ }\n}\n\n/** Merge cached models into _knownModels (idempotent). Called by fetchModels(). */\nasync function _refreshKnownModels(): Promise<void> {\n    try {\n        const cached = await listCachedModels();\n        for (const entry of cached) {\n            if (!_knownModels.find(m => m.id === entry.modelId)) {\n                _knownModels = [..._knownModels, _modelFromCacheEntry(entry.modelId)];\n            }\n        }\n    } catch { /* cache unavailable */ }\n}\n\n/**\n * How the worker should load `modelId`: which runner, which weights, which device.\n *\n * A custom `runner` is made absolute HERE, not in the worker: a worker's base URL is its\n * own script's, not the page's, and the blob shim `_spawnWorker` may build has no\n * meaningful base at all — so a relative path would resolve against the wrong place or\n * fail outright. The page is the one place that knows what the app meant.\n */\nfunction _selection(modelId: string): { task: BuiltInRunner; runner?: string; dtype?: Dtype; device: ComputeDevice } {\n    const config = _registeredModels.get(modelId);\n    const runner = config?.runner;\n    return {\n        task: config?.task ?? 'text-generation',\n        ...(runner ? { runner: typeof location === 'undefined' ? runner : new URL(runner, location.href).href } : {}),\n        dtype: config?.dtype,\n        device: _computeDevice,\n    };\n}\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Worker bridge\n// ─────────────────────────────────────────────────────────────────────────────\n\nlet _worker: Worker | null = null;\n\ninterface PendingPrepare {\n    modelId: string;\n    onProgress: (p: ModelLoadProgress) => void;\n    resolve: () => void;\n    reject: (err: Error) => void;\n}\nconst _pendingPrepares = new Map<string, PendingPrepare>();\nconst _pendingGenerates = new Map<string, ReadableStreamDefaultController>();\nconst _pendingCommands = new Map<string, { resolve: (value: unknown) => void; reject: (err: Error) => void }>();\n\n// ── Generate serialization ──────────────────────────────────────────────────\n// The worker holds ONE pipeline: two concurrent generates would corrupt each\n// other. Each chat() chains its `generate` behind the previous generate's\n// completion (gen-done / gen-error).\nlet _generateChain: Promise<void> = Promise.resolve();\nconst _generateDoneResolvers = new Map<string, () => void>();\n\n// ── Contention on the one pipeline ──────────────────────────────────────────\n// Serialization is correct but invisible: two chats driving DIFFERENT local\n// models take turns, and with `maxCachedModels` at its default of 1 each turn\n// can evict and reload gigabytes. The user sees a stall; the developer sees\n// nothing. These two track just enough to say so, once.\nconst _queuedModelIds = new Map<string, string>();\nlet _warnedModelContention = false;\n\n/** Model ids of generates currently queued or running on the single pipeline. */\nfunction _contendingModelId(requested: string): string | undefined {\n    for (const id of _queuedModelIds.values()) if (id !== requested) return id;\n    return undefined;\n}\n\n/**\n * Warn once when a generate has to queue behind another chat's DIFFERENT model.\n * Not a warning about switching models in one chat — that is a deliberate act\n * with visible feedback. This fires only when two are in flight at once.\n */\nfunction _warnIfContended(requested: string): void {\n    if (_warnedModelContention) return;\n    const other = _contendingModelId(requested);\n    if (!other) return;\n    _warnedModelContention = true;\n    console.warn(\n        `[Aparte] Two chats are driving different local models at once (\"${requested}\" behind `\n        + `\"${other}\"). Transformers.js runs one pipeline per tab, so these generates are `\n        + `serialized, and with a cache budget of ${_maxCachedModels} each switch can evict and `\n        + `reload gigabytes of weights. Point both chats at one model, or raise the budget with `\n        + `setMaxCachedModels(2) if the machine has the memory. This warns once.`,\n    );\n}\n\n/** Settle the serialization slot for a finished generate. */\nfunction _releaseGenerateSlot(id: string): void {\n    _queuedModelIds.delete(id);\n    const resolve = _generateDoneResolvers.get(id);\n    if (resolve) {\n        _generateDoneResolvers.delete(id);\n        resolve();\n    }\n}\n\n/** Model known to be loaded (main-thread view). */\nlet _loadedModelId: string | null = null;\n/** Model currently being prepared (for the getModelStatus 'cached' path). */\nlet _preparingModelId: string | null = null;\n\n/** The blob URL the worker was built from, if it needed one. Revoked with the worker. */\nlet _workerBlobUrl: string | null = null;\n\n/**\n * Build the worker — including when this package is served from another origin.\n *\n * `new Worker()` refuses a cross-origin script outright, and that is not an exotic\n * case: it is every deploy whose JavaScript lives on a CDN or an asset host while the\n * page lives somewhere else, with or without a bundler. Reproduced with the package on\n * one port and the page on another: `SecurityError: Script at '…/assets/worker-*.js'\n * cannot be accessed from origin '…'`.\n *\n * A blob inherits the ORIGIN OF THE DOCUMENT THAT CREATES IT, so a one-line blob whose\n * body imports the real worker by absolute URL is same-origin by construction, and the\n * import inside it is a normal cross-origin module fetch, which is allowed. It is the\n * shim ffmpeg.wasm and tesseract.js use for the same reason.\n *\n * Same-origin keeps the direct path: no blob, nothing to revoke, and a stack trace that\n * names the real file.\n */\nfunction _spawnWorker(): Worker {\n    const url = new URL(workerUrl, import.meta.url);\n    const sameOrigin = typeof location === 'undefined' || url.origin === location.origin;\n    // A blob is the only way across an origin, so an environment that cannot mint one has\n    // nothing to gain from trying: construct directly and let the platform say what it\n    // thinks. jsdom is that environment — it has `Blob` and no `URL.createObjectURL` — and\n    // every test in this package went through the blob path and threw before this line\n    // existed.\n    const canMintBlob = typeof Blob === 'function' && typeof URL.createObjectURL === 'function';\n    // The literal below is not style. `new Worker(new URL('./worker.ts', import.meta.url))`\n    // is the exact shape Vite's worker detection and webpack's WorkerPlugin match on, and\n    // matching it is what makes a CONSUMER's bundler process the worker as a module — which\n    // is how `@huggingface/transformers` gets resolved inside it today. Behind a variable\n    // the chunk is copied as an opaque asset and its imports are never touched, so hoisting\n    // this line to reuse it for the blob would fix a CDN page by breaking every bundled app.\n    if (sameOrigin || !canMintBlob) return new Worker(new URL('./worker.ts', import.meta.url), { type: 'module' });\n\n    _workerBlobUrl = URL.createObjectURL(\n        new Blob([`import ${JSON.stringify(url.href)};`], { type: 'text/javascript' }),\n    );\n    try {\n        return new Worker(_workerBlobUrl, { type: 'module' });\n    } catch (error) {\n        // A page with `worker-src 'self'` (or `script-src` without `blob:`) blocks the\n        // shim, and the direct URL was already refused for its origin — so there is\n        // nothing left to try. Say which of the two walls was hit, because the browser's\n        // own message does not distinguish them.\n        URL.revokeObjectURL(_workerBlobUrl);\n        _workerBlobUrl = null;\n        throw new Error(\n            `@aparte/provider-transformers is served from ${url.origin}, which is not this page's origin, `\n            + 'so its worker has to be started through a blob: URL — and this page\\'s Content-Security-Policy '\n            + 'refuses that. Allow `blob:` in `worker-src` (or `script-src`), or serve the package from your '\n            + `own origin. Original error: ${String(error)}`,\n        );\n    }\n}\n\nfunction _releaseWorkerBlob(): void {\n    if (_workerBlobUrl) {\n        URL.revokeObjectURL(_workerBlobUrl);\n        _workerBlobUrl = null;\n    }\n}\n\n/**\n * Where the page says Transformers.js lives, if it says so at all.\n *\n * The worker cannot ask: an import map is the DOCUMENT's, and by spec it does not reach\n * a worker. The main thread can, and does it the platform's way — `import.meta.resolve`\n * consults that same map — so a page that already maps `@huggingface/transformers` (it\n * has to, to import this package by name at all) is telling us where its copy is. That\n * map is the CDN consumer's manifest: the version pin stays with the consumer, which is\n * the whole point of a peer dependency, and this package invents no second place to say\n * it.\n *\n * `undefined` under a bundler, where the specifier is resolved at build time and the\n * worker's own `import('@huggingface/transformers')` is the path that runs.\n */\nfunction _peerModuleUrl(): string | undefined {\n    const resolve = (import.meta as unknown as { resolve?: (specifier: string) => string }).resolve;\n    if (typeof resolve === 'function') {\n        try {\n            const href = resolve('@huggingface/transformers');\n            if (href && /^https?:/i.test(href)) return href;\n        } catch { /* not in the map — fall through */ }\n    }\n    // Older engines have no `import.meta.resolve`; read the map they do have.\n    try {\n        const el = document.querySelector('script[type=\"importmap\"]');\n        const map = el?.textContent ? JSON.parse(el.textContent) as { imports?: Record<string, string> } : null;\n        const href = map?.imports?.['@huggingface/transformers'];\n        if (href) return new URL(href, location.href).href;\n    } catch { /* no document, or a map that is not JSON */ }\n    return undefined;\n}\n\nfunction _getWorker(): Worker {\n    if (!_worker) {\n        _worker = _spawnWorker();\n        _worker.addEventListener('message', _handleWorkerMessage);\n        _worker.addEventListener('error', _handleWorkerError);\n        _worker.addEventListener('messageerror', _handleWorkerError);\n        // First message, before any work: postMessage keeps order, so the worker has it\n        // by the time a prepare or a generate needs the module.\n        _worker.postMessage({ type: 'init', transformersUrl: _peerModuleUrl() });\n    }\n    return _worker;\n}\n\n/**\n * Worker crashed (uncaught error / WASM init failure / OOM). Reject every in-flight\n * prepare and close every open generate stream so the UI doesn't hang. Subsequent\n * calls rebuild the worker.\n */\nfunction _handleWorkerError(e: Event): void {\n    const message = (e as ErrorEvent)?.message || 'Worker crashed unexpectedly';\n\n    for (const p of _pendingPrepares.values()) {\n        try { p.reject(new Error(message)); } catch { /* ignore */ }\n    }\n    _pendingPrepares.clear();\n\n    for (const ctrl of _pendingGenerates.values()) {\n        try { ctrl.enqueue({ type: 'error' as const, message }); ctrl.close(); }\n        catch { /* ignore */ }\n    }\n    _pendingGenerates.clear();\n    for (const c of _pendingCommands.values()) c.reject(new Error(message));\n    _pendingCommands.clear();\n\n    // Release every serialization slot so the generate chain doesn't deadlock.\n    for (const resolve of _generateDoneResolvers.values()) {\n        try { resolve(); } catch { /* ignore */ }\n    }\n    _generateDoneResolvers.clear();\n    _generateChain = Promise.resolve();\n    _queuedModelIds.clear();\n\n    _loadedModelId = null;\n    _preparingModelId = null;\n    try { _worker?.terminate(); } catch { /* ignore */ }\n    _worker = null;\n    _releaseWorkerBlob();\n}\n\nfunction _handleWorkerMessage(event: MessageEvent): void {\n    const msg = event.data;\n\n    switch (msg.type) {\n        case 'progress': {\n            const pending = _pendingPrepares.get(msg.id);\n            if (!pending) break;\n            if (msg.status === 'ready') {\n                pending.onProgress({ status: 'ready' });\n                pending.resolve();\n                _pendingPrepares.delete(msg.id);\n            } else if (msg.status === 'loading') {\n                pending.onProgress({ status: 'loading' });\n            } else if (msg.status === 'cached') {\n                pending.onProgress({ status: 'cached', file: msg.file, progress: msg.progress });\n            } else {\n                pending.onProgress({ status: 'downloading', file: msg.file, progress: msg.progress });\n            }\n            break;\n        }\n        case 'prepare-error': {\n            const pending = _pendingPrepares.get(msg.id);\n            if (!pending) break;\n            pending.reject(new Error(msg.message));\n            _pendingPrepares.delete(msg.id);\n            if (_preparingModelId === pending.modelId) _preparingModelId = null;\n            break;\n        }\n        case 'pipeline-ready': {\n            _loadedModelId = msg.modelId;\n            _preparingModelId = null;\n            // Evict models over the cache limit, then refresh the known list.\n            void _enforceMaxCachedModels(msg.modelId).then(() => _refreshKnownModels());\n            break;\n        }\n        case 'gen-event': {\n            // The runner speaks the stream vocabulary itself; nothing to translate.\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue(msg.event);\n            break;\n        }\n        case 'warning': {\n            // Already said once per text by the worker; the page just carries the voice.\n            console.warn(`[transformers] ${msg.message}`);\n            break;\n        }\n        case 'command-result': {\n            _releaseGenerateSlot(msg.id);\n            const pending = _pendingCommands.get(msg.id);\n            if (!pending) break;\n            _pendingCommands.delete(msg.id);\n            if (msg.error !== undefined) pending.reject(new Error(msg.error));\n            else pending.resolve(msg.result);\n            break;\n        }\n        case 'gen-done': {\n            _releaseGenerateSlot(msg.id);\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue({ type: 'done' as const, ...(msg.usage ? { usage: msg.usage } : {}) });\n            ctrl.close();\n            _pendingGenerates.delete(msg.id);\n            break;\n        }\n        case 'gen-error': {\n            _releaseGenerateSlot(msg.id);\n            const ctrl = _pendingGenerates.get(msg.id);\n            if (!ctrl) break;\n            ctrl.enqueue({ type: 'error' as const, message: msg.message });\n            ctrl.close();\n            _pendingGenerates.delete(msg.id);\n            break;\n        }\n    }\n}\n\n/**\n * Narrowed so the two members the docs tell you to CALL are not optional.\n *\n * `AparteAIProvider` declares `prepareModel` and `getModelStatus` optional (most\n * providers have nothing to download), and widening to it made both\n * possibly-undefined — so the documented `TransformersProvider.prepareModel(...)`\n * needed a `!` or a guard in every strict consumer. Same technique openai-compat\n * already used for its own always-present members.\n *\n * `chat` joined the list once `AparteAIProvider` became a union: it is optional on\n * the format-adapter arm, and this provider IS its `chat()` — running inference\n * locally is the whole package. Narrowing it here says so once, instead of every\n * caller writing `provider.chat!(...)`.\n */\nexport const TransformersProvider: AparteAIProvider\n    & Required<Pick<AparteAIProvider, 'prepareModel' | 'getModelStatus' | 'chat'>> = {\n    id: 'transformers',\n\n    getMetadata() {\n        return {\n            id: 'transformers',\n            name: 'Transformers.js',\n            icon: `<svg viewBox=\"0 0 24 24\" fill=\"none\" xmlns=\"http://www.w3.org/2000/svg\"><path d=\"M12 2L2 7l10 5 10-5-10-5z\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/><path d=\"M2 17l10 5 10-5\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/><path d=\"M2 12l10 5 10-5\" stroke=\"currentColor\" stroke-width=\"2\" stroke-linecap=\"round\" stroke-linejoin=\"round\"/></svg>`,\n            color: '#f59e0b',\n            description: 'Run LLMs directly in your browser via WebGPU or WASM — no API, no key',\n            hasFreeModels: true,\n            isLocal: true,\n            helpUrl: 'https://huggingface.co/docs/transformers.js',\n        };\n    },\n\n    getModels(): AparteAIModel[] {\n        return _knownModels;\n    },\n\n    async fetchModels(): Promise<AparteAIModel[]> {\n        await _refreshKnownModels();\n        return _knownModels;\n    },\n\n    async chat(\n        request: AparteChatRequest,\n        _config?: string | Record<string, string>,\n        ctx?: { providerId: string; signal?: AbortSignal },\n    ): Promise<AparteChatResponse> {\n        const requestId = uuid();\n        const options = {\n            maxTokens: request.maxTokens,\n            temperature: request.temperature,\n            seed: request.seed,\n        };\n        const signal = ctx?.signal;\n\n        // ── Reserve a serialization slot ─────────────────────────────────────\n        // Chain this generate behind the previous one; the worker has a single\n        // pipeline, so generates MUST NOT overlap.\n        _warnIfContended(request.modelId);\n        _queuedModelIds.set(requestId, request.modelId);\n        const prevGenerate = _generateChain;\n        _generateChain = new Promise<void>((resolveSlot) => {\n            _generateDoneResolvers.set(requestId, resolveSlot);\n        });\n        // ── Stop, from either side ───────────────────────────────────────────\n        // The transport's `ctx.signal` (the user's Stop, which the provider contract\n        // says a bridge MUST honour — this one read it nowhere) and the stream's own\n        // `cancel()` say the same thing, and the worker hears it once. Before the\n        // generate has been posted there is nothing to interrupt: the stream is\n        // settled here, and the slot is released when its turn in the chain comes —\n        // not earlier, or the next generate would start over the one still running.\n        let posted = false;\n        let stopped = false;\n        const stop = (): void => {\n            if (stopped) return;\n            stopped = true;\n            signal?.removeEventListener('abort', stop);\n            if (posted) {\n                _getWorker().postMessage({ type: 'cancel', id: requestId });\n                return;\n            }\n            const ctrl = _pendingGenerates.get(requestId);\n            _pendingGenerates.delete(requestId);\n            if (!ctrl) return;\n            try { ctrl.enqueue({ type: 'error' as const, message: 'Generation cancelled before it started' }); ctrl.close(); }\n            catch { /* already closed */ }\n        };\n        const postGenerate = (): void => {\n            if (stopped) { _releaseGenerateSlot(requestId); return; }\n            posted = true;\n            _getWorker().postMessage({\n                type: 'generate',\n                id: requestId,\n                modelId: request.modelId,\n                // The conversation as it is, parts included: which parts a model can take\n                // is the runner's knowledge, not this thread's.\n                messages: request.messages,\n                options,\n                ..._selection(request.modelId),\n            });\n        };\n\n        let response: AparteChatResponse | Promise<string>;\n        if (request.stream === false) {\n            response = new Promise<string>((resolve, reject) => {\n                let result = '';\n                const fakeCtrl = {\n                    enqueue: (chunk: { type: string; delta?: string; message?: string }) => {\n                        if (chunk.type === 'text') result += chunk.delta ?? '';\n                        else if (chunk.type === 'done') resolve(result);\n                        else if (chunk.type === 'error') reject(new Error(chunk.message));\n                    },\n                    close: () => { /* no-op */ },\n                } as unknown as ReadableStreamDefaultController;\n                _pendingGenerates.set(requestId, fakeCtrl);\n                void prevGenerate.then(postGenerate);\n            });\n        } else {\n            response = new ReadableStream({\n                async start(controller) {\n                    _pendingGenerates.set(requestId, controller);\n                    await prevGenerate;\n                    postGenerate();\n                },\n                cancel() {\n                    // The reader is gone, so nothing may be enqueued for it again — and the\n                    // model actually STOPS (not just the read): the worker interrupts this\n                    // generate, and the slot is still released by the resulting\n                    // gen-done/gen-error, so a queued generate cannot start before that.\n                    _pendingGenerates.delete(requestId);\n                    stop();\n                },\n            });\n        }\n\n        // `start` has run by now, so the controller is registered and a stop settles it.\n        if (signal?.aborted) stop();\n        else signal?.addEventListener('abort', stop, { once: true });\n        return response;\n    },\n\n    async getModelStatus(modelId: string): Promise<ModelStatus> {\n        if (_loadedModelId === modelId) return 'ready';\n        if (_preparingModelId === modelId) return 'cached';\n        if ('caches' in globalThis) {\n            try {\n                const encodedId = encodeURIComponent(modelId);\n                const names = await caches.keys();\n                for (const name of names) {\n                    const cache = await caches.open(name);\n                    const keys = await cache.keys();\n                    if (keys.some(r => r.url.includes(encodedId) || r.url.includes(modelId + '/'))) {\n                        return 'cached';\n                    }\n                }\n            } catch {\n                // Cache API unavailable\n            }\n        }\n        return 'not-downloaded';\n    },\n\n    async prepareModel(modelId: string, onProgress: (p: ModelLoadProgress) => void): Promise<void> {\n        if (_loadedModelId === modelId) {\n            onProgress({ status: 'ready' });\n            return;\n        }\n\n        const requestId = uuid();\n        _preparingModelId = modelId;\n\n        return new Promise<void>((resolve, reject) => {\n            _pendingPrepares.set(requestId, { modelId, onProgress, resolve, reject });\n            _getWorker().postMessage({ type: 'prepare', id: requestId, modelId, ..._selection(modelId) });\n        });\n    },\n\n    async deleteModel(modelId: string): Promise<void> {\n        await deleteCachedModel(modelId);\n    },\n};\n\nexport default TransformersProvider;\nexport type { AparteAIProvider, AparteAIModel, ModelStatus, ModelLoadProgress } from '@aparte/core';\nexport type {\n    TransformersRunner,\n    RunnerContext,\n    RunnerGenerateInput,\n    RunnerProgress,\n    RunnerModule,\n    CreateRunner,\n    BuiltInRunner,\n    TransformersModule,\n} from './runners/types.js';\n\n// ─────────────────────────────────────────────────────────────────────────────\n// Cache utilities (settings panels, etc.)\n// ─────────────────────────────────────────────────────────────────────────────\n\n/** Returns the modelId currently loaded in the worker's pipeline, or null. */\nexport function getLoadedModelId(): string | null {\n    return _loadedModelId;\n}\n\n/**\n * Send a runner something that is not a generation — swap an adapter, warm a cache, ask\n * a capability — and get its answer. The name and payload are the runner's vocabulary\n * (the built-in runners answer none). Queued behind the generates in flight: the worker\n * holds one runner, and a command on it mid-stream would race the stream.\n */\nexport function runnerCommand(modelId: string, name: string, payload: unknown): Promise<unknown> {\n    const requestId = uuid();\n    _queuedModelIds.set(requestId, modelId);\n    const previous = _generateChain;\n    _generateChain = new Promise<void>((resolveSlot) => {\n        _generateDoneResolvers.set(requestId, resolveSlot);\n    });\n    return new Promise<unknown>((resolve, reject) => {\n        _pendingCommands.set(requestId, { resolve, reject });\n        void previous.then(() => {\n            _getWorker().postMessage({ type: 'command', id: requestId, modelId, name, payload, ..._selection(modelId) });\n        });\n    });\n}\n\n/** Terminate the shared worker and reset in-memory state. Safe to call any time. */\nexport function terminateWorker(): void {\n    _worker?.terminate();\n    _worker = null;\n    _releaseWorkerBlob();\n    _loadedModelId = null;\n    _preparingModelId = null;\n    for (const [, p] of _pendingPrepares) {\n        p.reject(new Error('Worker terminated'));\n    }\n    _pendingPrepares.clear();\n    for (const [, ctrl] of _pendingGenerates) {\n        try { ctrl.enqueue({ type: 'error' as const, message: 'Worker terminated' }); ctrl.close(); } catch { /* already closed */ }\n    }\n    _pendingGenerates.clear();\n    for (const c of _pendingCommands.values()) c.reject(new Error('Worker terminated'));\n    _pendingCommands.clear();\n\n    // Release every serialization slot and reset the chain — the same three lines\n    // the worker-error handler above already carried, with the same reason. Without\n    // them, terminating mid-generate left `_generateChain` pending on a resolver\n    // that had just been dropped, so the NEXT chat() awaited a promise that could\n    // never settle: no error, no rejection, the stream simply never started again\n    // for the life of the page.\n    for (const resolve of _generateDoneResolvers.values()) {\n        try { resolve(); } catch { /* ignore */ }\n    }\n    _generateDoneResolvers.clear();\n    _generateChain = Promise.resolve();\n    _queuedModelIds.clear();\n    // A terminated worker is a fresh situation; let the contention warning speak again.\n    _warnedModelContention = false;\n}\n\nexport interface CachedModelEntry {\n    modelId: string;\n    name: string;\n    /** Total size in bytes of all cached files for this model. -1 if unknown. */\n    sizeBytes: number;\n    /** True if the model is currently loaded in the worker. */\n    loaded: boolean;\n}\n\n/**\n * Scan the Cache API to find which Transformers.js models have been downloaded,\n * by matching cache entry URLs against the Hugging Face resolve path.\n */\nexport async function listCachedModels(): Promise<CachedModelEntry[]> {\n    if (!('caches' in globalThis)) return [];\n\n    const found = new Map<string, { name: string; sizeBytes: number }>();\n\n    // e.g. https://huggingface.co/onnx-community/Qwen2.5-0.5B/resolve/main/config.json\n    //   → onnx-community/Qwen2.5-0.5B\n    function extractModelId(url: string): string | null {\n        const m = url.match(/huggingface\\.co\\/([^/]+\\/[^/]+)\\/resolve\\//);\n        return m ? decodeURIComponent(m[1]!) : null;\n    }\n\n    function modelName(modelId: string): string {\n        const config = _registeredModels.get(modelId);\n        if (config) return config.name;\n        return (modelId.split('/').pop() ?? modelId).replace(/-/g, ' ');\n    }\n\n    try {\n        const cacheNames = await caches.keys();\n        await Promise.all(cacheNames.map(async (cacheName) => {\n            try {\n                const cache = await caches.open(cacheName);\n                const requests = await cache.keys();\n                for (const req of requests) {\n                    const modelId = extractModelId(req.url);\n                    if (!modelId) continue;\n                    if (!found.has(modelId)) {\n                        found.set(modelId, { name: modelName(modelId), sizeBytes: 0 });\n                    }\n                    const response = await cache.match(req);\n                    if (!response) continue;\n                    const contentLength = response.headers.get('content-length');\n                    if (contentLength) {\n                        found.get(modelId)!.sizeBytes += parseInt(contentLength, 10);\n                    } else {\n                        try {\n                            const blob = await response.clone().blob();\n                            found.get(modelId)!.sizeBytes += blob.size;\n                        } catch { /* skip */ }\n                    }\n                }\n            } catch { /* skip inaccessible cache */ }\n        }));\n    } catch {\n        return [];\n    }\n\n    return Array.from(found.entries()).map(([modelId, { name, sizeBytes }]) => ({\n        modelId,\n        name,\n        sizeBytes,\n        loaded: _loadedModelId === modelId,\n    }));\n}\n\n/**\n * Delete all cached files for a modelId from the Cache API, terminating the worker\n * first if that model is currently loaded.\n */\nexport async function deleteCachedModel(modelId: string): Promise<void> {\n    if (_loadedModelId === modelId || _preparingModelId === modelId) {\n        terminateWorker();\n    }\n    if (!('caches' in globalThis)) return;\n    try {\n        const cacheNames = await caches.keys();\n        await Promise.all(cacheNames.map(async (cacheName) => {\n            try {\n                const cache = await caches.open(cacheName);\n                const requests = await cache.keys();\n                const encoded = encodeURIComponent(modelId);\n                await Promise.all(\n                    requests\n                        .filter(r => r.url.includes(modelId) || r.url.includes(encoded))\n                        .map(r => cache.delete(r)),\n                );\n            } catch { /* skip */ }\n        }));\n    } catch { /* Cache API unavailable */ }\n}\n"],"names":[],"mappings":";;AAiEA,IAAI,iBAAqE;AAMlE,SAAS,sBAAsB,OAA0D;AAC5F,mBAAiB;AACrB;AAEA,eAAsB,iBAA2C;AAG7D,QAAM,QAAiB,UAAmD,gBAAgB;AAG1F,MAAI,SAAS;AACb,MAAI,SAAS,WAAW;AACpB,QAAI;AACA,YAAM,UAAU,MAAO,UAAyE,IAAI,eAAA;AACpG,eAAS,YAAY;AAAA,IACzB,QAAQ;AACJ,eAAS;AAAA,IACb;AAAA,EACJ;AAEA,MAAI;AACJ,MAAI,CAAC,UAAU,QAAQ,GAAG;AACtB,WAAO;AAAA,EACX,WAAW,QAAQ,GAAG;AAClB,WAAO;AAAA,EACX,OAAO;AACH,WAAO;AAAA,EACX;AAEA,QAAM,qBAAqB,iBACpB,eAAe,IAAI,KAAK,eAAe,QAAQ,KAChD;AAEN,SAAO,EAAE,QAAQ,OAAO,MAAM,mBAAA;AAClC;AA+BA,MAAM,wCAAwB,IAAA;AAG9B,IAAI,eAAgC,CAAA;AAK7B,SAAS,cAAc,QAAuC;AACjE,oBAAkB,IAAI,OAAO,IAAI,MAAM;AACvC,MAAI,CAAC,aAAa,KAAK,CAAA,MAAK,EAAE,OAAO,OAAO,EAAE,GAAG;AAC7C,mBAAe,CAAC,GAAG,cAAc;AAAA,MAC7B,IAAI,OAAO;AAAA,MACX,MAAM,OAAO;AAAA,MACb,aAAa,OAAO;AAAA,MACpB,cAAc,OAAO;AAAA,IAAA,CACxB;AAAA,EACL;AACJ;AAGA,SAAS,qBAAqB,SAAgC;AAC1D,QAAM,SAAS,kBAAkB,IAAI,OAAO;AAC5C,MAAI,OAAQ,QAAO,EAAE,IAAI,OAAO,IAAI,MAAM,OAAO,MAAM,aAAa,OAAO,aAAa,cAAc,OAAO,aAAA;AAC7G,QAAM,QAAQ,QAAQ,MAAM,GAAG,EAAE,SAAS,SAAS,QAAQ,MAAM,GAAG;AACpE,SAAO,EAAE,IAAI,SAAS,MAAM,cAAc,CAAC,WAAW,EAAA;AAC1D;AAGA,IAAI,mBAAmB;AAMhB,SAAS,mBAAmB,KAAmB;AAClD,qBAAmB;AACvB;AAGO,SAAS,qBAA6B;AACzC,SAAO;AACX;AASA,IAAI,iBAAgC;AAE7B,SAAS,iBAAiB,GAAwB;AACrD,mBAAiB;AACrB;AAEO,SAAS,mBAAkC;AAC9C,SAAO;AACX;AAGA,eAAe,wBAAwB,aAAoC;AACvE,MAAI,qBAAqB,EAAG;AAC5B,MAAI;AACA,UAAM,SAAS,MAAM,iBAAA;AACrB,UAAM,SAAS,OAAO,OAAO,CAAA,MAAK,EAAE,YAAY,WAAW;AAC3D,UAAM,SAAS,OAAO,SAAS;AAC/B,QAAI,UAAU,EAAG;AAEjB,aAAS,IAAI,GAAG,IAAI,UAAU,IAAI,OAAO,QAAQ,KAAK;AAClD,YAAM,kBAAkB,OAAO,CAAC,EAAG,OAAO;AAAA,IAC9C;AAAA,EACJ,QAAQ;AAAA,EAA0B;AACtC;AAGA,eAAe,sBAAqC;AAChD,MAAI;AACA,UAAM,SAAS,MAAM,iBAAA;AACrB,eAAW,SAAS,QAAQ;AACxB,UAAI,CAAC,aAAa,KAAK,CAAA,MAAK,EAAE,OAAO,MAAM,OAAO,GAAG;AACjD,uBAAe,CAAC,GAAG,cAAc,qBAAqB,MAAM,OAAO,CAAC;AAAA,MACxE;AAAA,IACJ;AAAA,EACJ,QAAQ;AAAA,EAA0B;AACtC;AAUA,SAAS,WAAW,SAAiG;AACjH,QAAM,SAAS,kBAAkB,IAAI,OAAO;AAC5C,QAAM,SAAS,QAAQ;AACvB,SAAO;AAAA,IACH,MAAM,QAAQ,QAAQ;AAAA,IACtB,GAAI,SAAS,EAAE,QAAQ,OAAO,aAAa,cAAc,SAAS,IAAI,IAAI,QAAQ,SAAS,IAAI,EAAE,KAAA,IAAS,CAAA;AAAA,IAC1G,OAAO,QAAQ;AAAA,IACf,QAAQ;AAAA,EAAA;AAEhB;AAMA,IAAI,UAAyB;AAQ7B,MAAM,uCAAuB,IAAA;AAC7B,MAAM,wCAAwB,IAAA;AAC9B,MAAM,uCAAuB,IAAA;AAM7B,IAAI,iBAAgC,QAAQ,QAAA;AAC5C,MAAM,6CAA6B,IAAA;AAOnC,MAAM,sCAAsB,IAAA;AAC5B,IAAI,yBAAyB;AAG7B,SAAS,mBAAmB,WAAuC;AAC/D,aAAW,MAAM,gBAAgB,OAAA,EAAU,KAAI,OAAO,UAAW,QAAO;AACxE,SAAO;AACX;AAOA,SAAS,iBAAiB,WAAyB;AAC/C,MAAI,uBAAwB;AAC5B,QAAM,QAAQ,mBAAmB,SAAS;AAC1C,MAAI,CAAC,MAAO;AACZ,2BAAyB;AACzB,UAAQ;AAAA,IACJ,mEAAmE,SAAS,aACtE,KAAK,gHACiC,gBAAgB;AAAA,EAAA;AAIpE;AAGA,SAAS,qBAAqB,IAAkB;AAC5C,kBAAgB,OAAO,EAAE;AACzB,QAAM,UAAU,uBAAuB,IAAI,EAAE;AAC7C,MAAI,SAAS;AACT,2BAAuB,OAAO,EAAE;AAChC,YAAA;AAAA,EACJ;AACJ;AAGA,IAAI,iBAAgC;AAEpC,IAAI,oBAAmC;AAGvC,IAAI,iBAAgC;AAmBpC,SAAS,eAAuB;AAC5B,QAAM,MAAM,IAAI,IAAI,WAAW,YAAY,GAAG;AAC9C,QAAM,aAAa,OAAO,aAAa,eAAe,IAAI,WAAW,SAAS;AAM9E,QAAM,cAAc,OAAO,SAAS,cAAc,OAAO,IAAI,oBAAoB;AAOjF,MAAI,cAAc,CAAC,YAAa,QAAO,IAAI,OAAO,IAAA;AAAA;AAAA,IAAA,KAAA,IAAA,IAAA,6BAAA,YAAA,GAAA,EAAA;AAAA,IAAA,YAAA;AAAA,EAAA,GAAyC,EAAE,MAAM,UAAU;AAE7G,mBAAiB,IAAI;AAAA,IACjB,IAAI,KAAK,CAAC,UAAU,KAAK,UAAU,IAAI,IAAI,CAAC,GAAG,GAAG,EAAE,MAAM,mBAAmB;AAAA,EAAA;AAEjF,MAAI;AACA,WAAO,IAAI,OAAO,gBAAgB,EAAE,MAAM,UAAU;AAAA,EACxD,SAAS,OAAO;AAKZ,QAAI,gBAAgB,cAAc;AAClC,qBAAiB;AACjB,UAAM,IAAI;AAAA,MACN,gDAAgD,IAAI,MAAM,oQAGzB,OAAO,KAAK,CAAC;AAAA,IAAA;AAAA,EAEtD;AACJ;AAEA,SAAS,qBAA2B;AAChC,MAAI,gBAAgB;AAChB,QAAI,gBAAgB,cAAc;AAClC,qBAAiB;AAAA,EACrB;AACJ;AAgBA,SAAS,iBAAqC;AAC1C,QAAM,UAAW,YAAuE;AACxF,MAAI,OAAO,YAAY,YAAY;AAC/B,QAAI;AACA,YAAM,OAAO,QAAQ,2BAA2B;AAChD,UAAI,QAAQ,YAAY,KAAK,IAAI,EAAG,QAAO;AAAA,IAC/C,QAAQ;AAAA,IAAsC;AAAA,EAClD;AAEA,MAAI;AACA,UAAM,KAAK,SAAS,cAAc,0BAA0B;AAC5D,UAAM,MAAM,IAAI,cAAc,KAAK,MAAM,GAAG,WAAW,IAA4C;AACnG,UAAM,OAAO,KAAK,UAAU,2BAA2B;AACvD,QAAI,KAAM,QAAO,IAAI,IAAI,MAAM,SAAS,IAAI,EAAE;AAAA,EAClD,QAAQ;AAAA,EAA+C;AACvD,SAAO;AACX;AAEA,SAAS,aAAqB;AAC1B,MAAI,CAAC,SAAS;AACV,cAAU,aAAA;AACV,YAAQ,iBAAiB,WAAW,oBAAoB;AACxD,YAAQ,iBAAiB,SAAS,kBAAkB;AACpD,YAAQ,iBAAiB,gBAAgB,kBAAkB;AAG3D,YAAQ,YAAY,EAAE,MAAM,QAAQ,iBAAiB,eAAA,GAAkB;AAAA,EAC3E;AACA,SAAO;AACX;AAOA,SAAS,mBAAmB,GAAgB;AACxC,QAAM,UAAW,GAAkB,WAAW;AAE9C,aAAW,KAAK,iBAAiB,UAAU;AACvC,QAAI;AAAE,QAAE,OAAO,IAAI,MAAM,OAAO,CAAC;AAAA,IAAG,QAAQ;AAAA,IAAe;AAAA,EAC/D;AACA,mBAAiB,MAAA;AAEjB,aAAW,QAAQ,kBAAkB,UAAU;AAC3C,QAAI;AAAE,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS;AAAG,WAAK,MAAA;AAAA,IAAS,QACjE;AAAA,IAAe;AAAA,EACzB;AACA,oBAAkB,MAAA;AAClB,aAAW,KAAK,iBAAiB,OAAA,KAAY,OAAO,IAAI,MAAM,OAAO,CAAC;AACtE,mBAAiB,MAAA;AAGjB,aAAW,WAAW,uBAAuB,UAAU;AACnD,QAAI;AAAE,cAAA;AAAA,IAAW,QAAQ;AAAA,IAAe;AAAA,EAC5C;AACA,yBAAuB,MAAA;AACvB,mBAAiB,QAAQ,QAAA;AACzB,kBAAgB,MAAA;AAEhB,mBAAiB;AACjB,sBAAoB;AACpB,MAAI;AAAE,aAAS,UAAA;AAAA,EAAa,QAAQ;AAAA,EAAe;AACnD,YAAU;AACV,qBAAA;AACJ;AAEA,SAAS,qBAAqB,OAA2B;AACrD,QAAM,MAAM,MAAM;AAElB,UAAQ,IAAI,MAAA;AAAA,IACR,KAAK,YAAY;AACb,YAAM,UAAU,iBAAiB,IAAI,IAAI,EAAE;AAC3C,UAAI,CAAC,QAAS;AACd,UAAI,IAAI,WAAW,SAAS;AACxB,gBAAQ,WAAW,EAAE,QAAQ,QAAA,CAAS;AACtC,gBAAQ,QAAA;AACR,yBAAiB,OAAO,IAAI,EAAE;AAAA,MAClC,WAAW,IAAI,WAAW,WAAW;AACjC,gBAAQ,WAAW,EAAE,QAAQ,UAAA,CAAW;AAAA,MAC5C,WAAW,IAAI,WAAW,UAAU;AAChC,gBAAQ,WAAW,EAAE,QAAQ,UAAU,MAAM,IAAI,MAAM,UAAU,IAAI,SAAA,CAAU;AAAA,MACnF,OAAO;AACH,gBAAQ,WAAW,EAAE,QAAQ,eAAe,MAAM,IAAI,MAAM,UAAU,IAAI,SAAA,CAAU;AAAA,MACxF;AACA;AAAA,IACJ;AAAA,IACA,KAAK,iBAAiB;AAClB,YAAM,UAAU,iBAAiB,IAAI,IAAI,EAAE;AAC3C,UAAI,CAAC,QAAS;AACd,cAAQ,OAAO,IAAI,MAAM,IAAI,OAAO,CAAC;AACrC,uBAAiB,OAAO,IAAI,EAAE;AAC9B,UAAI,sBAAsB,QAAQ,QAAS,qBAAoB;AAC/D;AAAA,IACJ;AAAA,IACA,KAAK,kBAAkB;AACnB,uBAAiB,IAAI;AACrB,0BAAoB;AAEpB,WAAK,wBAAwB,IAAI,OAAO,EAAE,KAAK,MAAM,qBAAqB;AAC1E;AAAA,IACJ;AAAA,IACA,KAAK,aAAa;AAEd,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,IAAI,KAAK;AACtB;AAAA,IACJ;AAAA,IACA,KAAK,WAAW;AAEZ,cAAQ,KAAK,kBAAkB,IAAI,OAAO,EAAE;AAC5C;AAAA,IACJ;AAAA,IACA,KAAK,kBAAkB;AACnB,2BAAqB,IAAI,EAAE;AAC3B,YAAM,UAAU,iBAAiB,IAAI,IAAI,EAAE;AAC3C,UAAI,CAAC,QAAS;AACd,uBAAiB,OAAO,IAAI,EAAE;AAC9B,UAAI,IAAI,UAAU,OAAW,SAAQ,OAAO,IAAI,MAAM,IAAI,KAAK,CAAC;AAAA,UAC3D,SAAQ,QAAQ,IAAI,MAAM;AAC/B;AAAA,IACJ;AAAA,IACA,KAAK,YAAY;AACb,2BAAqB,IAAI,EAAE;AAC3B,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,EAAE,MAAM,QAAiB,GAAI,IAAI,QAAQ,EAAE,OAAO,IAAI,MAAA,IAAU,CAAA,GAAK;AAClF,WAAK,MAAA;AACL,wBAAkB,OAAO,IAAI,EAAE;AAC/B;AAAA,IACJ;AAAA,IACA,KAAK,aAAa;AACd,2BAAqB,IAAI,EAAE;AAC3B,YAAM,OAAO,kBAAkB,IAAI,IAAI,EAAE;AACzC,UAAI,CAAC,KAAM;AACX,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS,IAAI,SAAS;AAC7D,WAAK,MAAA;AACL,wBAAkB,OAAO,IAAI,EAAE;AAC/B;AAAA,IACJ;AAAA,EAAA;AAER;AAgBO,MAAM,uBACwE;AAAA,EACjF,IAAI;AAAA,EAEJ,cAAc;AACV,WAAO;AAAA,MACH,IAAI;AAAA,MACJ,MAAM;AAAA,MACN,MAAM;AAAA,MACN,OAAO;AAAA,MACP,aAAa;AAAA,MACb,eAAe;AAAA,MACf,SAAS;AAAA,MACT,SAAS;AAAA,IAAA;AAAA,EAEjB;AAAA,EAEA,YAA6B;AACzB,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,cAAwC;AAC1C,UAAM,oBAAA;AACN,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,KACF,SACA,SACA,KAC2B;AAC3B,UAAM,YAAY,KAAA;AAClB,UAAM,UAAU;AAAA,MACZ,WAAW,QAAQ;AAAA,MACnB,aAAa,QAAQ;AAAA,MACrB,MAAM,QAAQ;AAAA,IAAA;AAElB,UAAM,SAAS,KAAK;AAKpB,qBAAiB,QAAQ,OAAO;AAChC,oBAAgB,IAAI,WAAW,QAAQ,OAAO;AAC9C,UAAM,eAAe;AACrB,qBAAiB,IAAI,QAAc,CAAC,gBAAgB;AAChD,6BAAuB,IAAI,WAAW,WAAW;AAAA,IACrD,CAAC;AAQD,QAAI,SAAS;AACb,QAAI,UAAU;AACd,UAAM,OAAO,MAAY;AACrB,UAAI,QAAS;AACb,gBAAU;AACV,cAAQ,oBAAoB,SAAS,IAAI;AACzC,UAAI,QAAQ;AACR,mBAAA,EAAa,YAAY,EAAE,MAAM,UAAU,IAAI,WAAW;AAC1D;AAAA,MACJ;AACA,YAAM,OAAO,kBAAkB,IAAI,SAAS;AAC5C,wBAAkB,OAAO,SAAS;AAClC,UAAI,CAAC,KAAM;AACX,UAAI;AAAE,aAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS,0CAA0C;AAAG,aAAK,MAAA;AAAA,MAAS,QAC3G;AAAA,MAAuB;AAAA,IACjC;AACA,UAAM,eAAe,MAAY;AAC7B,UAAI,SAAS;AAAE,6BAAqB,SAAS;AAAG;AAAA,MAAQ;AACxD,eAAS;AACT,iBAAA,EAAa,YAAY;AAAA,QACrB,MAAM;AAAA,QACN,IAAI;AAAA,QACJ,SAAS,QAAQ;AAAA;AAAA;AAAA,QAGjB,UAAU,QAAQ;AAAA,QAClB;AAAA,QACA,GAAG,WAAW,QAAQ,OAAO;AAAA,MAAA,CAChC;AAAA,IACL;AAEA,QAAI;AACJ,QAAI,QAAQ,WAAW,OAAO;AAC1B,iBAAW,IAAI,QAAgB,CAAC,SAAS,WAAW;AAChD,YAAI,SAAS;AACb,cAAM,WAAW;AAAA,UACb,SAAS,CAAC,UAA8D;AACpE,gBAAI,MAAM,SAAS,OAAQ,WAAU,MAAM,SAAS;AAAA,qBAC3C,MAAM,SAAS,OAAQ,SAAQ,MAAM;AAAA,qBACrC,MAAM,SAAS,QAAS,QAAO,IAAI,MAAM,MAAM,OAAO,CAAC;AAAA,UACpE;AAAA,UACA,OAAO,MAAM;AAAA,UAAc;AAAA,QAAA;AAE/B,0BAAkB,IAAI,WAAW,QAAQ;AACzC,aAAK,aAAa,KAAK,YAAY;AAAA,MACvC,CAAC;AAAA,IACL,OAAO;AACH,iBAAW,IAAI,eAAe;AAAA,QAC1B,MAAM,MAAM,YAAY;AACpB,4BAAkB,IAAI,WAAW,UAAU;AAC3C,gBAAM;AACN,uBAAA;AAAA,QACJ;AAAA,QACA,SAAS;AAKL,4BAAkB,OAAO,SAAS;AAClC,eAAA;AAAA,QACJ;AAAA,MAAA,CACH;AAAA,IACL;AAGA,QAAI,QAAQ,QAAS,MAAA;AAAA,iBACR,iBAAiB,SAAS,MAAM,EAAE,MAAM,MAAM;AAC3D,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,eAAe,SAAuC;AACxD,QAAI,mBAAmB,QAAS,QAAO;AACvC,QAAI,sBAAsB,QAAS,QAAO;AAC1C,QAAI,YAAY,YAAY;AACxB,UAAI;AACA,cAAM,YAAY,mBAAmB,OAAO;AAC5C,cAAM,QAAQ,MAAM,OAAO,KAAA;AAC3B,mBAAW,QAAQ,OAAO;AACtB,gBAAM,QAAQ,MAAM,OAAO,KAAK,IAAI;AACpC,gBAAM,OAAO,MAAM,MAAM,KAAA;AACzB,cAAI,KAAK,KAAK,CAAA,MAAK,EAAE,IAAI,SAAS,SAAS,KAAK,EAAE,IAAI,SAAS,UAAU,GAAG,CAAC,GAAG;AAC5E,mBAAO;AAAA,UACX;AAAA,QACJ;AAAA,MACJ,QAAQ;AAAA,MAER;AAAA,IACJ;AACA,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,aAAa,SAAiB,YAA2D;AAC3F,QAAI,mBAAmB,SAAS;AAC5B,iBAAW,EAAE,QAAQ,SAAS;AAC9B;AAAA,IACJ;AAEA,UAAM,YAAY,KAAA;AAClB,wBAAoB;AAEpB,WAAO,IAAI,QAAc,CAAC,SAAS,WAAW;AAC1C,uBAAiB,IAAI,WAAW,EAAE,SAAS,YAAY,SAAS,QAAQ;AACxE,mBAAa,YAAY,EAAE,MAAM,WAAW,IAAI,WAAW,SAAS,GAAG,WAAW,OAAO,EAAA,CAAG;AAAA,IAChG,CAAC;AAAA,EACL;AAAA,EAEA,MAAM,YAAY,SAAgC;AAC9C,UAAM,kBAAkB,OAAO;AAAA,EACnC;AACJ;AAoBO,SAAS,mBAAkC;AAC9C,SAAO;AACX;AAQO,SAAS,cAAc,SAAiB,MAAc,SAAoC;AAC7F,QAAM,YAAY,KAAA;AAClB,kBAAgB,IAAI,WAAW,OAAO;AACtC,QAAM,WAAW;AACjB,mBAAiB,IAAI,QAAc,CAAC,gBAAgB;AAChD,2BAAuB,IAAI,WAAW,WAAW;AAAA,EACrD,CAAC;AACD,SAAO,IAAI,QAAiB,CAAC,SAAS,WAAW;AAC7C,qBAAiB,IAAI,WAAW,EAAE,SAAS,QAAQ;AACnD,SAAK,SAAS,KAAK,MAAM;AACrB,iBAAA,EAAa,YAAY,EAAE,MAAM,WAAW,IAAI,WAAW,SAAS,MAAM,SAAS,GAAG,WAAW,OAAO,GAAG;AAAA,IAC/G,CAAC;AAAA,EACL,CAAC;AACL;AAGO,SAAS,kBAAwB;AACpC,WAAS,UAAA;AACT,YAAU;AACV,qBAAA;AACA,mBAAiB;AACjB,sBAAoB;AACpB,aAAW,CAAA,EAAG,CAAC,KAAK,kBAAkB;AAClC,MAAE,OAAO,IAAI,MAAM,mBAAmB,CAAC;AAAA,EAC3C;AACA,mBAAiB,MAAA;AACjB,aAAW,CAAA,EAAG,IAAI,KAAK,mBAAmB;AACtC,QAAI;AAAE,WAAK,QAAQ,EAAE,MAAM,SAAkB,SAAS,qBAAqB;AAAG,WAAK,MAAA;AAAA,IAAS,QAAQ;AAAA,IAAuB;AAAA,EAC/H;AACA,oBAAkB,MAAA;AAClB,aAAW,KAAK,iBAAiB,OAAA,KAAY,OAAO,IAAI,MAAM,mBAAmB,CAAC;AAClF,mBAAiB,MAAA;AAQjB,aAAW,WAAW,uBAAuB,UAAU;AACnD,QAAI;AAAE,cAAA;AAAA,IAAW,QAAQ;AAAA,IAAe;AAAA,EAC5C;AACA,yBAAuB,MAAA;AACvB,mBAAiB,QAAQ,QAAA;AACzB,kBAAgB,MAAA;AAEhB,2BAAyB;AAC7B;AAeA,eAAsB,mBAAgD;AAClE,MAAI,EAAE,YAAY,YAAa,QAAO,CAAA;AAEtC,QAAM,4BAAY,IAAA;AAIlB,WAAS,eAAe,KAA4B;AAChD,UAAM,IAAI,IAAI,MAAM,4CAA4C;AAChE,WAAO,IAAI,mBAAmB,EAAE,CAAC,CAAE,IAAI;AAAA,EAC3C;AAEA,WAAS,UAAU,SAAyB;AACxC,UAAM,SAAS,kBAAkB,IAAI,OAAO;AAC5C,QAAI,eAAe,OAAO;AAC1B,YAAQ,QAAQ,MAAM,GAAG,EAAE,SAAS,SAAS,QAAQ,MAAM,GAAG;AAAA,EAClE;AAEA,MAAI;AACA,UAAM,aAAa,MAAM,OAAO,KAAA;AAChC,UAAM,QAAQ,IAAI,WAAW,IAAI,OAAO,cAAc;AAClD,UAAI;AACA,cAAM,QAAQ,MAAM,OAAO,KAAK,SAAS;AACzC,cAAM,WAAW,MAAM,MAAM,KAAA;AAC7B,mBAAW,OAAO,UAAU;AACxB,gBAAM,UAAU,eAAe,IAAI,GAAG;AACtC,cAAI,CAAC,QAAS;AACd,cAAI,CAAC,MAAM,IAAI,OAAO,GAAG;AACrB,kBAAM,IAAI,SAAS,EAAE,MAAM,UAAU,OAAO,GAAG,WAAW,GAAG;AAAA,UACjE;AACA,gBAAM,WAAW,MAAM,MAAM,MAAM,GAAG;AACtC,cAAI,CAAC,SAAU;AACf,gBAAM,gBAAgB,SAAS,QAAQ,IAAI,gBAAgB;AAC3D,cAAI,eAAe;AACf,kBAAM,IAAI,OAAO,EAAG,aAAa,SAAS,eAAe,EAAE;AAAA,UAC/D,OAAO;AACH,gBAAI;AACA,oBAAM,OAAO,MAAM,SAAS,MAAA,EAAQ,KAAA;AACpC,oBAAM,IAAI,OAAO,EAAG,aAAa,KAAK;AAAA,YAC1C,QAAQ;AAAA,YAAa;AAAA,UACzB;AAAA,QACJ;AAAA,MACJ,QAAQ;AAAA,MAAgC;AAAA,IAC5C,CAAC,CAAC;AAAA,EACN,QAAQ;AACJ,WAAO,CAAA;AAAA,EACX;AAEA,SAAO,MAAM,KAAK,MAAM,QAAA,CAAS,EAAE,IAAI,CAAC,CAAC,SAAS,EAAE,MAAM,UAAA,CAAW,OAAO;AAAA,IACxE;AAAA,IACA;AAAA,IACA;AAAA,IACA,QAAQ,mBAAmB;AAAA,EAAA,EAC7B;AACN;AAMA,eAAsB,kBAAkB,SAAgC;AACpE,MAAI,mBAAmB,WAAW,sBAAsB,SAAS;AAC7D,oBAAA;AAAA,EACJ;AACA,MAAI,EAAE,YAAY,YAAa;AAC/B,MAAI;AACA,UAAM,aAAa,MAAM,OAAO,KAAA;AAChC,UAAM,QAAQ,IAAI,WAAW,IAAI,OAAO,cAAc;AAClD,UAAI;AACA,cAAM,QAAQ,MAAM,OAAO,KAAK,SAAS;AACzC,cAAM,WAAW,MAAM,MAAM,KAAA;AAC7B,cAAM,UAAU,mBAAmB,OAAO;AAC1C,cAAM,QAAQ;AAAA,UACV,SACK,OAAO,CAAA,MAAK,EAAE,IAAI,SAAS,OAAO,KAAK,EAAE,IAAI,SAAS,OAAO,CAAC,EAC9D,IAAI,OAAK,MAAM,OAAO,CAAC,CAAC;AAAA,QAAA;AAAA,MAErC,QAAQ;AAAA,MAAa;AAAA,IACzB,CAAC,CAAC;AAAA,EACN,QAAQ;AAAA,EAA8B;AAC1C;"}

package/dist/runners/image-text-to-text.d.ts

@@ -0,0 +1,36 @@
+/**
+ * The built-in vision runner — a model that reads images and text and writes text
+ * (SmolVLM, Qwen2-VL, LFM2-VL, Gemma 3, LLaVA…: everything `AutoModelForImageTextToText`
+ * resolves).
+ *
+ * Transformers.js 4.x has no `image-text-to-text` PIPELINE, so this runner goes through
+ * the model classes themselves, the way the SmolVLM examples do: `AutoProcessor` renders
+ * the chat template with `{ type: 'image' }` placeholders, the images are decoded beside
+ * the prompt in the same order, the processor turns both into tensors, and `generate()`
+ * streams through a `TextStreamer`. Tool turns are dropped with the shared warning.
+ */
+import type { AparteChatMessage } from '@aparte/core';
+import type { CreateRunner } from './types.js';
+type HFPart = {
+    type: 'image';
+} | {
+    type: 'text';
+    text: string;
+};
+type HFMessage = {
+    role: 'user' | 'assistant' | 'system';
+    content: HFPart[];
+};
+export declare const UNSUPPORTED_PARTS_DROPPED = "Dropped content part(s) this vision runner cannot carry (only text and image parts reach the model).";
+/**
+ * The conversation in the HF chat shape the processor's template expects — every turn's
+ * content as parts, an `{ type: 'image' }` placeholder where a picture goes — plus the
+ * pictures themselves, in order of appearance, for the processor to pair with them.
+ */
+export declare function toChatTemplate(messages: AparteChatMessage[], warn: (message: string) => void): {
+    chat: HFMessage[];
+    images: string[];
+};
+export declare const createRunner: CreateRunner;
+export {};
+//# sourceMappingURL=image-text-to-text.d.ts.map

package/dist/runners/image-text-to-text.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"image-text-to-text.d.ts","sourceRoot":"","sources":["../../src/runners/image-text-to-text.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AACtD,OAAO,KAAK,EAAE,YAAY,EAAuB,MAAM,YAAY,CAAC;AAGpE,KAAK,MAAM,GAAG;IAAE,IAAI,EAAE,OAAO,CAAA;CAAE,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AACjE,KAAK,SAAS,GAAG;IAAE,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,QAAQ,CAAC;IAAC,OAAO,EAAE,MAAM,EAAE,CAAA;CAAE,CAAC;AAE9E,eAAO,MAAM,yBAAyB,yGACoE,CAAC;AAE3G;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,QAAQ,EAAE,iBAAiB,EAAE,EAAE,IAAI,EAAE,CAAC,OAAO,EAAE,MAAM,KAAK,IAAI,GAAG;IAAE,IAAI,EAAE,SAAS,EAAE,CAAC;IAAC,MAAM,EAAE,MAAM,EAAE,CAAA;CAAE,CAsBtI;AAcD,eAAO,MAAM,YAAY,EAAE,YAsC1B,CAAC"}

package/dist/runners/shared.d.ts

@@ -0,0 +1,30 @@
+/**
+ * What the two built-in runners share — kept in one place so the two cannot drift on
+ * how progress is reported, how a stop reaches the model, or what a dropped tool turn
+ * says. Types only from core (see `text-generation.ts` for why).
+ */
+import type { AparteStreamEvent } from '@aparte/core';
+import type { RunnerContext, TransformersModule } from './types.js';
+export declare const TOOL_TURNS_DROPPED: string;
+/**
+ * The options a `from_pretrained` / `pipeline()` call takes from the context: download
+ * progress forwarded to the page (percentages, rounded), dtype and device when set.
+ */
+export declare function loadOptions(ctx: RunnerContext): Record<string, unknown>;
+/**
+ * A stopping criteria the signal interrupts — so a Stop actually STOPS the model, not just
+ * the read; otherwise generation runs to `max_new_tokens` off-thread, spending exactly the
+ * CPU/GPU/battery this provider exists to save. Call `release()` in a `finally`.
+ */
+export declare function interruptOn(signal: AbortSignal, transformers: TransformersModule): {
+    stopping: unknown;
+    release(): void;
+};
+/** A `TextStreamer` that emits each decoded token as a `text` event, prompt skipped. */
+export declare function textStreamer(transformers: TransformersModule, tokenizer: unknown, emit: (event: AparteStreamEvent) => void): unknown;
+/** Sampling options in Transformers.js' vocabulary, from the request's. */
+export declare function generationOptions(options: {
+    maxTokens?: number;
+    temperature?: number;
+}): Record<string, unknown>;
+//# sourceMappingURL=shared.d.ts.map

package/dist/runners/shared.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"shared.d.ts","sourceRoot":"","sources":["../../src/runners/shared.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AACtD,OAAO,KAAK,EAAE,aAAa,EAAE,kBAAkB,EAAE,MAAM,YAAY,CAAC;AAEpE,eAAO,MAAM,kBAAkB,QAGO,CAAC;AAEvC;;;GAGG;AACH,wBAAgB,WAAW,CAAC,GAAG,EAAE,aAAa,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAUvE;AAED;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,MAAM,EAAE,WAAW,EAAE,YAAY,EAAE,kBAAkB,GAAG;IAAE,QAAQ,EAAE,OAAO,CAAC;IAAC,OAAO,IAAI,IAAI,CAAA;CAAE,CAMzH;AAED,wFAAwF;AACxF,wBAAgB,YAAY,CAAC,YAAY,EAAE,kBAAkB,EAAE,SAAS,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,KAAK,EAAE,iBAAiB,KAAK,IAAI,GAAG,OAAO,CAOpI;AAED,2EAA2E;AAC3E,wBAAgB,iBAAiB,CAAC,OAAO,EAAE;IAAE,SAAS,CAAC,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAOhH"}

package/dist/runners/text-generation.d.ts

@@ -0,0 +1,22 @@
+/**
+ * The built-in text runner — the generic `pipeline('text-generation')` path this
+ * provider has always run, extracted from the worker so it is one runner among others.
+ *
+ * It flattens the conversation to `{ role, content: string }` turns (the tokenizer applies
+ * the chat template). Two things it cannot carry, it SAYS: tool turns (their wire syntax is
+ * model-specific) and image parts (a text model has no eyes). The second used to vanish
+ * silently — a photo attached to a text model produced an answer that pretended — and
+ * that silence, not the limitation, was the defect.
+ */
+import type { AparteChatMessage } from '@aparte/core';
+import type { CreateRunner, RunnerContext } from './types.js';
+type SimpleMessage = {
+    role: 'user' | 'assistant' | 'system';
+    content: string;
+};
+export declare const IMAGES_DROPPED: string;
+/** Flatten to what the chat template takes; say what was left out. */
+export declare function flattenForChatTemplate(messages: AparteChatMessage[], warn: RunnerContext['warn']): SimpleMessage[];
+export declare const createRunner: CreateRunner;
+export {};
+//# sourceMappingURL=text-generation.d.ts.map

package/dist/runners/text-generation.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"text-generation.d.ts","sourceRoot":"","sources":["../../src/runners/text-generation.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAMH,OAAO,KAAK,EAAE,iBAAiB,EAAqB,MAAM,cAAc,CAAC;AACzE,OAAO,KAAK,EAAE,YAAY,EAAE,aAAa,EAAuB,MAAM,YAAY,CAAC;AAGnF,KAAK,aAAa,GAAG;IAAE,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,QAAQ,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAAC;AAEhF,eAAO,MAAM,cAAc,QAGoB,CAAC;AAWhD,sEAAsE;AACtE,wBAAgB,sBAAsB,CAAC,QAAQ,EAAE,iBAAiB,EAAE,EAAE,IAAI,EAAE,aAAa,CAAC,MAAM,CAAC,GAAG,aAAa,EAAE,CAgBlH;AASD,eAAO,MAAM,YAAY,EAAE,YAqB1B,CAAC"}