@aparte/provider-transformers 0.19.3 → 0.20.0

package/README.md

@@ -77,11 +77,12 @@ registerModel({ id: 'my-org/my-model', name: 'Mine', capabilities: ['streaming']
 // runners/mine.js — served by your app; the worker imports it by URL
 export async function createRunner(ctx) {
   // ctx.transformers is the Transformers.js the page installed or mapped — never import your own
-  const { AutoTokenizer, AutoModelForCausalLM, TextStreamer } = ctx.transformers;
+  const { AutoTokenizer, AutoModelForCausalLM, TextStreamer, InterruptableStoppingCriteria } = ctx.transformers;
   const tokenizer = await AutoTokenizer.from_pretrained(ctx.modelId);
   const model = await AutoModelForCausalLM.from_pretrained(ctx.modelId, { dtype: ctx.dtype, device: ctx.device });
   return {
-    async generate({ messages, options, emit }) {
+    async generate({ messages, options, emit, signal }) {
+      if (signal.aborted) return;
       // messages arrive WITH their content parts — render them the way this model wants;
       // this one reads text, so the parts are flattened and tool turns left out
       const turns = messages
@@ -92,7 +93,10 @@ export async function createRunner(ctx) {
         skip_prompt: true, skip_special_tokens: true,
         callback_function: (text) => emit({ type: 'text', delta: text }),
       });
-      await model.generate({ ...inputs, max_new_tokens: options.maxTokens ?? 512, streamer });
+      // Stop: without this the model generates on to its ceiling, and the next reply waits behind it
+      const stopping = new InterruptableStoppingCriteria();
+      signal.addEventListener('abort', () => stopping.interrupt(), { once: true });
+      await model.generate({ ...inputs, max_new_tokens: options.maxTokens ?? 8192, streamer, stopping_criteria: stopping });
       // `done` is optional — the provider closes the stream when generate() resolves
     },
     // anything that is not a generation: the provider passes it through, untouched
@@ -121,14 +125,16 @@ one WebGPU pipeline, so a worker per chat would mean N copies resident in one ta
 the **same** model is the case this is for: they share the load, for free.
 
 Two chats on **different** models is the case that costs. They serialize on the one pipeline,
-and at the default budget of one cached model each turn can evict and reload gigabytes. The
-provider warns once when it sees it:
+so each turn unloads one model and loads the other. The cache budget adds to that: it is how
+many models stay **downloaded** in the browser's Cache API (disk, not memory: one model is
+loaded at a time whatever the budget), and at the default of one, a switch can delete the other
+model's weights and download them again. The provider warns once when it sees it:
 
 ```ts
 import { setMaxCachedModels, getMaxCachedModels } from '@aparte/provider-transformers';
 
-setMaxCachedModels(2);            // keep both resident, if the machine has the memory
-setMaxCachedModels(0);            // no limit — you are managing memory yourself
+setMaxCachedModels(2);            // keep both downloaded, if the browser's storage has room
+setMaxCachedModels(0);            // no limit — the browser's storage quota is the only bound
 getMaxCachedModels();             // the current budget
 ```
 

package/dist/index.d.ts

@@ -66,6 +66,14 @@ export interface TransformersModelConfig {
      * the same Transformers.js instance the built-ins use. Wins over `task`.
      */
     runner?: string;
+    /**
+     * The most tokens a reply from this model may run to, unless the request sets its own
+     * `maxTokens`. Left unset, a built-in runner stops at 8192. With a built-in runner, whatever
+     * the ceiling, a reply also ends where the model's context window does, prompt included, and
+     * Stop interrupts it. A runner of your own receives `options.maxTokens` only when the request
+     * or this setting gives one, so it sets its own default (the README's example uses 8192).
+     */
+    maxTokens?: number;
     /** ONNX dtype or per-part dtype map (e.g. `'q4'` or `{ decoder_model_merged: 'q4' }`). */
     dtype?: Dtype;
     /** Not read: where models run is one setting for the page, {@link setComputeDevice}. */
@@ -77,8 +85,13 @@ export interface TransformersModelConfig {
  */
 export declare function registerModel(config: TransformersModelConfig): void;
 /**
- * Set the maximum number of models to keep in cache. When exceeded after a new
- * model is ready, the oldest models are evicted. 0 = unlimited.
+ * Set the maximum number of models to keep downloaded in the browser's Cache API. When
+ * exceeded after a new model is ready, the oldest models' weights are deleted. 0 = unlimited.
+ * It bounds disk, not memory: one model is loaded at a time whatever the budget.
+ *
+ * The budget counts and evicts only the models given to {@link registerModel} (and the
+ * one just loaded). Other Transformers.js weights in the origin's cache, from your own
+ * pipelines or another library, are never counted and never deleted.
  */
 export declare function setMaxCachedModels(max: number): void;
 /** Returns the current max-cached-models setting. */

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAEH,OAAO,KAAK,EACR,gBAAgB,EAChB,aAAa,EAKhB,MAAM,cAAc,CAAC;AAEtB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAOvE,MAAM,WAAW,eAAe;IAC5B,MAAM,EAAE,OAAO,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,KAAK,GAAG,KAAK,GAAG,MAAM,CAAC;IAC7B,kBAAkB,EAAE,MAAM,CAAC;CAC9B;AAKD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,KAAK,EAAE;IAAE,GAAG,EAAE,MAAM,CAAC;IAAC,GAAG,CAAC,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAAG,IAAI,CAE9F;AAED,wBAAsB,cAAc,IAAI,OAAO,CAAC,eAAe,CAAC,CA8B/D;AAMD,8DAA8D;AAC9D,MAAM,WAAW,uBAAuB;IACpC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,aAAa,CAAC,cAAc,CAAC,CAAC;IAC5C;;;OAGG;IACH,IAAI,CAAC,EAAE,aAAa,CAAC;IACrB;;;;OAIG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,0FAA0F;IAC1F,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,wFAAwF;IACxF,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACtC;AAQD;;GAEG;AACH,wBAAgB,aAAa,CAAC,MAAM,EAAE,uBAAuB,GAAG,IAAI,CAUnE;AAaD;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAEpD;AAED,qDAAqD;AACrD,wBAAgB,kBAAkB,IAAI,MAAM,CAE3C;AAED;;;;;GAKG;AACH,MAAM,MAAM,aAAa,GAAG,MAAM,GAAG,QAAQ,GAAG,MAAM,CAAC;AAGvD,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,aAAa,GAAG,IAAI,CAEvD;AAED,wBAAgB,gBAAgB,IAAI,aAAa,CAEhD;AAqXD;;;;;;;;;;;;;GAaG;AACH,eAAO,MAAM,oBAAoB,EAAE,gBAAgB,GAC7C,QAAQ,CAAC,IAAI,CAAC,gBAAgB,EAAE,cAAc,GAAG,gBAAgB,GAAG,MAAM,CAAC,CA0KhF,CAAC;AAEF,eAAe,oBAAoB,CAAC;AACpC,YAAY,EAAE,gBAAgB,EAAE,aAAa,EAAE,WAAW,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AACpG,YAAY,EACR,kBAAkB,EAClB,aAAa,EACb,mBAAmB,EACnB,cAAc,EACd,YAAY,EACZ,YAAY,EACZ,aAAa,EACb,kBAAkB,GACrB,MAAM,oBAAoB,CAAC;AAM5B,8EAA8E;AAC9E,wBAAgB,gBAAgB,IAAI,MAAM,GAAG,IAAI,CAEhD;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAa/F;AAED,oFAAoF;AACpF,wBAAgB,eAAe,IAAI,IAAI,CA+BtC;AAED,MAAM,WAAW,gBAAgB;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;IACb,6EAA6E;IAC7E,SAAS,EAAE,MAAM,CAAC;IAClB,2DAA2D;IAC3D,MAAM,EAAE,OAAO,CAAC;CACnB;AAED;;;GAGG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAsDpE;AAED;;;GAGG;AACH,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAoBtE"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAgCG;AAEH,OAAO,KAAK,EACR,gBAAgB,EAChB,aAAa,EAKhB,MAAM,cAAc,CAAC;AAEtB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAOvE,MAAM,WAAW,eAAe;IAC5B,MAAM,EAAE,OAAO,CAAC;IAChB,KAAK,EAAE,MAAM,CAAC;IACd,IAAI,EAAE,KAAK,GAAG,KAAK,GAAG,MAAM,CAAC;IAC7B,kBAAkB,EAAE,MAAM,CAAC;CAC9B;AAKD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,KAAK,EAAE;IAAE,GAAG,EAAE,MAAM,CAAC;IAAC,GAAG,CAAC,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAAG,IAAI,CAE9F;AAED,wBAAsB,cAAc,IAAI,OAAO,CAAC,eAAe,CAAC,CA8B/D;AAMD,8DAA8D;AAC9D,MAAM,WAAW,uBAAuB;IACpC,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,aAAa,CAAC,cAAc,CAAC,CAAC;IAC5C;;;OAGG;IACH,IAAI,CAAC,EAAE,aAAa,CAAC;IACrB;;;;OAIG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB;;;;;;OAMG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,0FAA0F;IAC1F,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,wFAAwF;IACxF,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACtC;AAQD;;GAEG;AACH,wBAAgB,aAAa,CAAC,MAAM,EAAE,uBAAuB,GAAG,IAAI,CAUnE;AAKD;;;;;;;;GAQG;AACH,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,MAAM,GAAG,IAAI,CAEpD;AAED,qDAAqD;AACrD,wBAAgB,kBAAkB,IAAI,MAAM,CAE3C;AAED;;;;;GAKG;AACH,MAAM,MAAM,aAAa,GAAG,MAAM,GAAG,QAAQ,GAAG,MAAM,CAAC;AAGvD,wBAAgB,gBAAgB,CAAC,CAAC,EAAE,aAAa,GAAG,IAAI,CAEvD;AAED,wBAAgB,gBAAgB,IAAI,aAAa,CAEhD;AAkXD;;;;;;;;;;;;;GAaG;AACH,eAAO,MAAM,oBAAoB,EAAE,gBAAgB,GAC7C,QAAQ,CAAC,IAAI,CAAC,gBAAgB,EAAE,cAAc,GAAG,gBAAgB,GAAG,MAAM,CAAC,CA6KhF,CAAC;AAEF,eAAe,oBAAoB,CAAC;AACpC,YAAY,EAAE,gBAAgB,EAAE,aAAa,EAAE,WAAW,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AACpG,YAAY,EACR,kBAAkB,EAClB,aAAa,EACb,mBAAmB,EACnB,cAAc,EACd,YAAY,EACZ,YAAY,EACZ,aAAa,EACb,kBAAkB,GACrB,MAAM,oBAAoB,CAAC;AAM5B,8EAA8E;AAC9E,wBAAgB,gBAAgB,IAAI,MAAM,GAAG,IAAI,CAEhD;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAa/F;AAED,oFAAoF;AACpF,wBAAgB,eAAe,IAAI,IAAI,CA+BtC;AAgBD,MAAM,WAAW,gBAAgB;IAC7B,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,CAAC;IACb,6EAA6E;IAC7E,SAAS,EAAE,MAAM,CAAC;IAClB,2DAA2D;IAC3D,MAAM,EAAE,OAAO,CAAC;CACnB;AAED;;;GAGG;AACH,wBAAsB,gBAAgB,IAAI,OAAO,CAAC,gBAAgB,EAAE,CAAC,CA+CpE;AAED;;;GAGG;AACH,wBAAsB,iBAAiB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAmBtE"}

package/dist/index.js

@@ -38,12 +38,6 @@ function registerModel(config) {
     }];
   }
 }
-function _modelFromCacheEntry(modelId) {
-  const config = _registeredModels.get(modelId);
-  if (config) return { id: config.id, name: config.name, description: config.description, capabilities: config.capabilities };
-  const name = (modelId.split("/").pop() ?? modelId).replace(/-/g, " ");
-  return { id: modelId, name, capabilities: ["streaming"] };
-}
 let _maxCachedModels = 1;
 function setMaxCachedModels(max) {
   _maxCachedModels = max;
@@ -61,7 +55,7 @@ function getComputeDevice() {
 async function _enforceMaxCachedModels(keepModelId) {
   if (_maxCachedModels === 0) return;
   try {
-    const cached = await listCachedModels();
+    const cached = (await listCachedModels()).filter((e) => _registeredModels.has(e.modelId) || e.modelId === keepModelId);
     const inUse = new Set(_queuedModelIds.values());
     const others = cached.filter((e) => e.modelId !== keepModelId && !inUse.has(e.modelId));
     const excess = cached.length - _maxCachedModels;
@@ -72,17 +66,6 @@ async function _enforceMaxCachedModels(keepModelId) {
   } catch {
   }
 }
-async function _refreshKnownModels() {
-  try {
-    const cached = await listCachedModels();
-    for (const entry of cached) {
-      if (!_knownModels.find((m) => m.id === entry.modelId)) {
-        _knownModels = [..._knownModels, _modelFromCacheEntry(entry.modelId)];
-      }
-    }
-  } catch {
-  }
-}
 function _selection(modelId) {
   const config = _registeredModels.get(modelId);
   const runner = config?.runner;
@@ -110,8 +93,9 @@ function _warnIfContended(requested) {
   const other = _contendingModelId(requested);
   if (!other) return;
   _warnedModelContention = true;
+  const deletes = _maxCachedModels === 1 ? ` With a cache budget of 1, a switch also deletes the other model's downloaded weights and downloads them again; setMaxCachedModels(2) keeps both, if the browser's storage has room.` : "";
   console.warn(
-    `[Aparte] Two chats are driving different local models at once ("${requested}" behind "${other}"). Transformers.js runs one pipeline per tab, so these generates are serialized, and with a cache budget of ${_maxCachedModels} each switch can evict and reload gigabytes of weights. Point both chats at one model, or raise the budget with setMaxCachedModels(2) if the machine has the memory. This warns once.`
+    `[Aparte] Two chats are driving different local models at once ("${requested}" behind "${other}"). Transformers.js runs one pipeline per tab, so these generates are serialized: each switch unloads one model and loads the other.${deletes} Point both chats at one model to avoid it. This warns once.`
   );
 }
 function _releaseGenerateSlot(id) {
@@ -245,7 +229,7 @@ function _handleWorkerMessage(event) {
     case "pipeline-ready": {
       _loadedModelId = msg.modelId;
       _preparingModelId = null;
-      void _enforceMaxCachedModels(msg.modelId).then(() => _refreshKnownModels());
+      void _enforceMaxCachedModels(msg.modelId);
       break;
     }
     case "gen-event": {
@@ -304,14 +288,18 @@ const TransformersProvider = {
   getModels() {
     return _knownModels;
   },
+  /**
+   * The registered models. Only those: the Cache API is the origin's, and listing what it
+   * held offered an app's own embedder, or another library's model, as a chat model to
+   * pick. `listCachedModels()` still reports everything downloaded, for a settings panel.
+   */
   async fetchModels() {
-    await _refreshKnownModels();
     return _knownModels;
   },
   async chat(request, _config, ctx) {
     const requestId = uuid();
     const options = {
-      maxTokens: request.maxTokens,
+      maxTokens: request.maxTokens ?? _registeredModels.get(request.modelId)?.maxTokens,
       temperature: request.temperature,
       seed: request.seed
     };
@@ -401,12 +389,11 @@ const TransformersProvider = {
     if (_preparingModelId === modelId) return "cached";
     if ("caches" in globalThis) {
       try {
-        const encodedId = encodeURIComponent(modelId);
         const names = await caches.keys();
         for (const name of names) {
           const cache = await caches.open(name);
           const keys = await cache.keys();
-          if (keys.some((r) => r.url.includes(encodedId) || r.url.includes(modelId + "/"))) {
+          if (keys.some((r) => _modelIdOfUrl(r.url) === modelId)) {
             return "cached";
           }
         }
@@ -479,13 +466,13 @@ function terminateWorker() {
   _queuedModelIds.clear();
   _warnedModelContention = false;
 }
+function _modelIdOfUrl(url) {
+  const m = url.match(/huggingface\.co\/([^/]+\/[^/]+)\/resolve\//);
+  return m ? decodeURIComponent(m[1]) : null;
+}
 async function listCachedModels() {
   if (!("caches" in globalThis)) return [];
   const found = /* @__PURE__ */ new Map();
-  function extractModelId(url) {
-    const m = url.match(/huggingface\.co\/([^/]+\/[^/]+)\/resolve\//);
-    return m ? decodeURIComponent(m[1]) : null;
-  }
   function modelName(modelId) {
     const config = _registeredModels.get(modelId);
     if (config) return config.name;
@@ -498,7 +485,7 @@ async function listCachedModels() {
         const cache = await caches.open(cacheName);
         const requests = await cache.keys();
         for (const req of requests) {
-          const modelId = extractModelId(req.url);
+          const modelId = _modelIdOfUrl(req.url);
           if (!modelId) continue;
           if (!found.has(modelId)) {
             found.set(modelId, { name: modelName(modelId), sizeBytes: 0 });
@@ -540,9 +527,8 @@ async function deleteCachedModel(modelId) {
       try {
         const cache = await caches.open(cacheName);
         const requests = await cache.keys();
-        const encoded = encodeURIComponent(modelId);
         await Promise.all(
-          requests.filter((r) => r.url.includes(modelId) || r.url.includes(encoded)).map((r) => cache.delete(r))
+          requests.filter((r) => _modelIdOfUrl(r.url) === modelId).map((r) => cache.delete(r))
         );
       } catch {
       }

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: 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\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    /** Not read: where models run is one setting for the page, {@link setComputeDevice}. */\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/**\n * Evict models from cache until count <= _maxCachedModels. `keepModelId` is never\n * evicted, and neither is a model a generate is queued or running on: the budget\n * fires from `pipeline-ready`, i.e. when ANOTHER model finishes loading, so it\n * used to delete the weights of the model that was answering — and\n * `deleteCachedModel` terminates the worker when that model is the loaded one.\n */\nasync function _enforceMaxCachedModels(keepModelId: string): Promise<void> {\n    if (_maxCachedModels === 0) return; // unlimited\n    try {\n        const cached = await listCachedModels();\n        const inUse = new Set(_queuedModelIds.values());\n        const others = cached.filter(e => e.modelId !== keepModelId && !inUse.has(e.modelId));\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/**\n * The worker this package publishes, beside `dist/index.js`.\n *\n * It used to be `import workerUrl from './worker.ts?worker&url'`, which handed the\n * emit to Vite's worker plugin — and that plugin rewrote the one shape a consumer's\n * bundler can detect into `new URL(\"assets/worker-<hash>.js\", import.meta.url).href`\n * behind a `@vite-ignore`. Nothing static was left to detect, so a bundled app copied\n * the chunk as an opaque asset without ever processing it as a module, and the\n * `import('@huggingface/transformers')` inside it stayed a bare specifier no browser\n * can resolve: every model load failed. The worker is a second lib entry now, at a\n * stable `dist/worker.js`, and this package constructs it itself.\n */\nconst WORKER_FILE = './worker.js';\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    // Behind a constant, and that is not style either. Vite's asset transform rewrites\n    // `new URL(<literal>, import.meta.url)` to a base of its own; behind a variable it\n    // does not look, so this line keeps the REAL module URL — which is what the origin\n    // comparison and the blob body below have to be built from. The build drops that\n    // transform anyway (see `vite.config.ts`), so in the shipped bytes the two forms\n    // resolve identically; under a dev server and under vitest they do not.\n    const url = new URL(WORKER_FILE, 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, and it is written out rather than reusing\n    // `WORKER_FILE` above. `new Worker(new URL('./worker.js', import.meta.url))` is the\n    // 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 —\n    // which is how `@huggingface/transformers` gets resolved inside it. Behind a variable\n    // the file is copied as an opaque asset and its imports are never touched, so\n    // hoisting this line to reuse `url` would fix a CDN page by breaking every bundled\n    // app. That the SHIPPED bytes still carry it is asserted by\n    // `src/__tests__/published-shape.test.ts`: the claim was true of this source and\n    // false of the artifact for as long as the build owned the emit.\n    if (sameOrigin || !canMintBlob) return new Worker(new URL('./worker.js', 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            const ctrl = _pendingGenerates.get(requestId);\n            _pendingGenerates.delete(requestId);\n            if (posted) {\n                _getWorker().postMessage({ type: 'cancel', id: requestId });\n                // The reply ends HERE, not when the worker gets round to answering:\n                // a token already in flight was otherwise enqueued into a stream the\n                // user had stopped. `done` and not an error — a stop is not a failure,\n                // and it is what settles the non-streaming path too. The worker's own\n                // `gen-done` still releases the queue slot: that happens before it\n                // looks the controller up.\n                try { ctrl?.enqueue({ type: 'done' as const }); ctrl?.close(); } catch { /* already closed */ }\n                return;\n            }\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":";AA0DA,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;AASA,eAAe,wBAAwB,aAAoC;AACvE,MAAI,qBAAqB,EAAG;AAC5B,MAAI;AACA,UAAM,SAAS,MAAM,iBAAA;AACrB,UAAM,QAAQ,IAAI,IAAI,gBAAgB,QAAQ;AAC9C,UAAM,SAAS,OAAO,OAAO,CAAA,MAAK,EAAE,YAAY,eAAe,CAAC,MAAM,IAAI,EAAE,OAAO,CAAC;AACpF,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;AAcvC,MAAM,cAAc;AAGpB,IAAI,iBAAgC;AAmBpC,SAAS,eAAuB;AAO5B,QAAM,MAAM,IAAI,IAAI,aAAa,YAAY,GAAG;AAChD,QAAM,aAAa,OAAO,aAAa,eAAe,IAAI,WAAW,SAAS;AAM9E,QAAM,cAAc,OAAO,SAAS,cAAc,OAAO,IAAI,oBAAoB;AAWjF,MAAI,cAAc,CAAC,YAAa,QAAO,IAAI,OAAO,IAAI,IAAI,eAAe,YAAY,GAAG,GAAG,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,YAAM,OAAO,kBAAkB,IAAI,SAAS;AAC5C,wBAAkB,OAAO,SAAS;AAClC,UAAI,QAAQ;AACR,mBAAA,EAAa,YAAY,EAAE,MAAM,UAAU,IAAI,WAAW;AAO1D,YAAI;AAAE,gBAAM,QAAQ,EAAE,MAAM,OAAA,CAAiB;AAAG,gBAAM,MAAA;AAAA,QAAS,QAAQ;AAAA,QAAuB;AAC9F;AAAA,MACJ;AACA,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;"}
+{"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\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    /**\n     * The most tokens a reply from this model may run to, unless the request sets its own\n     * `maxTokens`. Left unset, a built-in runner stops at 8192. With a built-in runner, whatever\n     * the ceiling, a reply also ends where the model's context window does, prompt included, and\n     * Stop interrupts it. A runner of your own receives `options.maxTokens` only when the request\n     * or this setting gives one, so it sets its own default (the README's example uses 8192).\n     */\n    maxTokens?: number;\n    /** ONNX dtype or per-part dtype map (e.g. `'q4'` or `{ decoder_model_merged: 'q4' }`). */\n    dtype?: Dtype;\n    /** Not read: where models run is one setting for the page, {@link setComputeDevice}. */\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(). */\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/** 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 downloaded in the browser's Cache API. When\n * exceeded after a new model is ready, the oldest models' weights are deleted. 0 = unlimited.\n * It bounds disk, not memory: one model is loaded at a time whatever the budget.\n *\n * The budget counts and evicts only the models given to {@link registerModel} (and the\n * one just loaded). Other Transformers.js weights in the origin's cache, from your own\n * pipelines or another library, are never counted and never deleted.\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/**\n * Evict models from cache until count <= _maxCachedModels. `keepModelId` is never\n * evicted, and neither is a model a generate is queued or running on: the budget\n * fires from `pipeline-ready`, i.e. when ANOTHER model finishes loading, so it\n * used to delete the weights of the model that was answering — and\n * `deleteCachedModel` terminates the worker when that model is the loaded one.\n */\nasync function _enforceMaxCachedModels(keepModelId: string): Promise<void> {\n    if (_maxCachedModels === 0) return; // unlimited\n    try {\n        // Only the models this provider was given count: the Cache API is the origin's,\n        // and an app's own Transformers.js pipelines (an embedder, a second library) keep\n        // their weights there too. Counting them evicted a model we never loaded.\n        const cached = (await listCachedModels())\n            .filter(e => _registeredModels.has(e.modelId) || e.modelId === keepModelId);\n        const inUse = new Set(_queuedModelIds.values());\n        const others = cached.filter(e => e.modelId !== keepModelId && !inUse.has(e.modelId));\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/**\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    // Only a budget of one deletes a model the other chat needs: 0 is no limit, and two or\n    // more keep both chats' models downloaded.\n    const deletes = _maxCachedModels === 1\n        ? ` With a cache budget of 1, a switch also deletes the other model's downloaded weights `\n          + `and downloads them again; setMaxCachedModels(2) keeps both, if the browser's storage has room.`\n        : '';\n    console.warn(\n        `[Aparte] Two chats are driving different local models at once (\"${requested}\" behind `\n        + `\"${other}\"). Transformers.js runs one pipeline per tab, so these generates are `\n        + `serialized: each switch unloads one model and loads the other.${deletes} Point both `\n        + `chats at one model to avoid it. 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/**\n * The worker this package publishes, beside `dist/index.js`.\n *\n * It used to be `import workerUrl from './worker.ts?worker&url'`, which handed the\n * emit to Vite's worker plugin — and that plugin rewrote the one shape a consumer's\n * bundler can detect into `new URL(\"assets/worker-<hash>.js\", import.meta.url).href`\n * behind a `@vite-ignore`. Nothing static was left to detect, so a bundled app copied\n * the chunk as an opaque asset without ever processing it as a module, and the\n * `import('@huggingface/transformers')` inside it stayed a bare specifier no browser\n * can resolve: every model load failed. The worker is a second lib entry now, at a\n * stable `dist/worker.js`, and this package constructs it itself.\n */\nconst WORKER_FILE = './worker.js';\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    // Behind a constant, and that is not style either. Vite's asset transform rewrites\n    // `new URL(<literal>, import.meta.url)` to a base of its own; behind a variable it\n    // does not look, so this line keeps the REAL module URL — which is what the origin\n    // comparison and the blob body below have to be built from. The build drops that\n    // transform anyway (see `vite.config.ts`), so in the shipped bytes the two forms\n    // resolve identically; under a dev server and under vitest they do not.\n    const url = new URL(WORKER_FILE, 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, and it is written out rather than reusing\n    // `WORKER_FILE` above. `new Worker(new URL('./worker.js', import.meta.url))` is the\n    // 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 —\n    // which is how `@huggingface/transformers` gets resolved inside it. Behind a variable\n    // the file is copied as an opaque asset and its imports are never touched, so\n    // hoisting this line to reuse `url` would fix a CDN page by breaking every bundled\n    // app. That the SHIPPED bytes still carry it is asserted by\n    // `src/__tests__/published-shape.test.ts`: the claim was true of this source and\n    // false of the artifact for as long as the build owned the emit.\n    if (sameOrigin || !canMintBlob) return new Worker(new URL('./worker.js', 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);\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    /**\n     * The registered models. Only those: the Cache API is the origin's, and listing what it\n     * held offered an app's own embedder, or another library's model, as a chat model to\n     * pick. `listCachedModels()` still reports everything downloaded, for a settings panel.\n     */\n    async fetchModels(): Promise<AparteAIModel[]> {\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 ?? _registeredModels.get(request.modelId)?.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            const ctrl = _pendingGenerates.get(requestId);\n            _pendingGenerates.delete(requestId);\n            if (posted) {\n                _getWorker().postMessage({ type: 'cancel', id: requestId });\n                // The reply ends HERE, not when the worker gets round to answering:\n                // a token already in flight was otherwise enqueued into a stream the\n                // user had stopped. `done` and not an error — a stop is not a failure,\n                // and it is what settles the non-streaming path too. The worker's own\n                // `gen-done` still releases the queue slot: that happens before it\n                // looks the controller up.\n                try { ctrl?.enqueue({ type: 'done' as const }); ctrl?.close(); } catch { /* already closed */ }\n                return;\n            }\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 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 => _modelIdOfUrl(r.url) === 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\n/**\n * The model a cached file belongs to, read from its Hugging Face resolve path — or null.\n * Every cache lookup matches on this, by equality: a substring match took a longer sibling\n * with it (`Qwen2.5-0.5B` also matched `Qwen2.5-0.5B-Instruct`), so deleting one model\n * deleted the other's weights.\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 */\nfunction _modelIdOfUrl(url: string): string | null {\n    const m = url.match(/huggingface\\.co\\/([^/]+\\/[^/]+)\\/resolve\\//);\n    return m ? decodeURIComponent(m[1]!) : null;\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    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 = _modelIdOfUrl(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                await Promise.all(\n                    requests\n                        .filter(r => _modelIdOfUrl(r.url) === modelId)\n                        .map(r => cache.delete(r)),\n                );\n            } catch { /* skip */ }\n        }));\n    } catch { /* Cache API unavailable */ }\n}\n"],"names":[],"mappings":";AA0DA,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;AAuCA,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,IAAI,mBAAmB;AAWhB,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;AASA,eAAe,wBAAwB,aAAoC;AACvE,MAAI,qBAAqB,EAAG;AAC5B,MAAI;AAIA,UAAM,UAAU,MAAM,iBAAA,GACjB,OAAO,CAAA,MAAK,kBAAkB,IAAI,EAAE,OAAO,KAAK,EAAE,YAAY,WAAW;AAC9E,UAAM,QAAQ,IAAI,IAAI,gBAAgB,QAAQ;AAC9C,UAAM,SAAS,OAAO,OAAO,CAAA,MAAK,EAAE,YAAY,eAAe,CAAC,MAAM,IAAI,EAAE,OAAO,CAAC;AACpF,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;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;AAGzB,QAAM,UAAU,qBAAqB,IAC/B,yLAEA;AACN,UAAQ;AAAA,IACJ,mEAAmE,SAAS,aACtE,KAAK,uIACwD,OAAO;AAAA,EAAA;AAGlF;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;AAcvC,MAAM,cAAc;AAGpB,IAAI,iBAAgC;AAmBpC,SAAS,eAAuB;AAO5B,QAAM,MAAM,IAAI,IAAI,aAAa,YAAY,GAAG;AAChD,QAAM,aAAa,OAAO,aAAa,eAAe,IAAI,WAAW,SAAS;AAM9E,QAAM,cAAc,OAAO,SAAS,cAAc,OAAO,IAAI,oBAAoB;AAWjF,MAAI,cAAc,CAAC,YAAa,QAAO,IAAI,OAAO,IAAI,IAAI,eAAe,YAAY,GAAG,GAAG,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;AACxC;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;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,cAAwC;AAC1C,WAAO;AAAA,EACX;AAAA,EAEA,MAAM,KACF,SACA,SACA,KAC2B;AAC3B,UAAM,YAAY,KAAA;AAClB,UAAM,UAAU;AAAA,MACZ,WAAW,QAAQ,aAAa,kBAAkB,IAAI,QAAQ,OAAO,GAAG;AAAA,MACxE,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,YAAM,OAAO,kBAAkB,IAAI,SAAS;AAC5C,wBAAkB,OAAO,SAAS;AAClC,UAAI,QAAQ;AACR,mBAAA,EAAa,YAAY,EAAE,MAAM,UAAU,IAAI,WAAW;AAO1D,YAAI;AAAE,gBAAM,QAAQ,EAAE,MAAM,OAAA,CAAiB;AAAG,gBAAM,MAAA;AAAA,QAAS,QAAQ;AAAA,QAAuB;AAC9F;AAAA,MACJ;AACA,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,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,cAAc,EAAE,GAAG,MAAM,OAAO,GAAG;AAClD,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;AAWA,SAAS,cAAc,KAA4B;AAC/C,QAAM,IAAI,IAAI,MAAM,4CAA4C;AAChE,SAAO,IAAI,mBAAmB,EAAE,CAAC,CAAE,IAAI;AAC3C;AAeA,eAAsB,mBAAgD;AAClE,MAAI,EAAE,YAAY,YAAa,QAAO,CAAA;AAEtC,QAAM,4BAAY,IAAA;AAElB,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,cAAc,IAAI,GAAG;AACrC,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,QAAQ;AAAA,UACV,SACK,OAAO,CAAA,MAAK,cAAc,EAAE,GAAG,MAAM,OAAO,EAC5C,IAAI,CAAA,MAAK,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.map

@@ -1 +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;;;;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,CA6BtI;AAcD,eAAO,MAAM,YAAY,EAAE,YAsC1B,CAAC"}
+{"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;;;;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,CA6BtI;AAeD,eAAO,MAAM,YAAY,EAAE,YAuC1B,CAAC"}

package/dist/runners/image-text-to-text.js

@@ -1,4 +1,4 @@
-import { l as loadOptions, i as interruptOn, t as textStreamer, g as generationOptions, T as TOOL_TURNS_DROPPED, U as UNSUPPORTED_PARTS_DROPPED } from "./shared.js";
+import { l as loadOptions, c as contextLength, i as interruptOn, s as stoppingCriteria, t as textStreamer, g as generationOptions, T as TOOL_TURNS_DROPPED, U as UNSUPPORTED_PARTS_DROPPED } from "./shared.js";
 function toChatTemplate(messages, warn) {
   const chat = [];
   const images = [];
@@ -35,6 +35,7 @@ const createRunner = async (ctx) => {
     tf.AutoProcessor.from_pretrained(ctx.modelId),
     tf.AutoModelForImageTextToText.from_pretrained(ctx.modelId, loadOptions(ctx))
   ]);
+  const contextSize = contextLength(model.config);
   return {
     async generate({ messages, options, emit, signal }) {
       const { stopping, release } = interruptOn(signal, ctx.transformers);
@@ -46,7 +47,7 @@ const createRunner = async (ctx) => {
           ...inputs,
           ...generationOptions(options),
           streamer: textStreamer(ctx.transformers, processor.tokenizer, emit),
-          stopping_criteria: stopping
+          stopping_criteria: stoppingCriteria(stopping, ctx.transformers, contextSize)
         });
       } finally {
         release();

package/dist/runners/image-text-to-text.js.map

@@ -1 +1 @@
-{"version":3,"file":"image-text-to-text.js","sources":["../../src/runners/image-text-to-text.ts"],"sourcesContent":["/**\n * The built-in vision runner — a model that reads images and text and writes text\n * (SmolVLM, Qwen2-VL, LFM2-VL, Gemma 3, LLaVA…: everything `AutoModelForImageTextToText`\n * resolves).\n *\n * Transformers.js 4.x has no `image-text-to-text` PIPELINE, so this runner goes through\n * the model classes themselves, the way the SmolVLM examples do: `AutoProcessor` renders\n * the chat template with `{ type: 'image' }` placeholders, the images are decoded beside\n * the prompt in the same order, the processor turns both into tensors, and `generate()`\n * streams through a `TextStreamer`. Tool turns are dropped with the shared warning.\n */\n\nimport type { AparteChatMessage } from '@aparte/core';\nimport type { CreateRunner, RunnerGenerateInput } from './types.js';\nimport { TOOL_TURNS_DROPPED, UNSUPPORTED_PARTS_DROPPED, generationOptions, interruptOn, loadOptions, textStreamer } from './shared.js';\n\ntype HFPart = { type: 'image' } | { type: 'text'; text: string };\ntype HFMessage = { role: 'user' | 'assistant' | 'system'; content: HFPart[] };\n\n/**\n * The conversation in the HF chat shape the processor's template expects — every turn's\n * content as parts, an `{ type: 'image' }` placeholder where a picture goes — plus the\n * pictures themselves, in order of appearance, for the processor to pair with them.\n */\nexport function toChatTemplate(messages: AparteChatMessage[], warn: (message: string) => void): { chat: HFMessage[]; images: string[] } {\n    const chat: HFMessage[] = [];\n    const images: string[] = [];\n    let droppedToolTurns = 0;\n    let droppedParts = 0;\n    for (const m of messages) {\n        if (m.role !== 'user' && m.role !== 'assistant' && m.role !== 'system') { droppedToolTurns++; continue; }\n        // An assistant's calls ride on an `assistant` message, which passes the role test\n        // above and, when the model said nothing before them, carries no part either — so\n        // without this the whole turn leaves the prompt with nothing said.\n        if (m.toolCalls?.length) droppedToolTurns++;\n        const parts: HFPart[] = [];\n        if (typeof m.content === 'string') {\n            if (m.content) parts.push({ type: 'text', text: m.content });\n        } else {\n            for (const p of m.content) {\n                if (p.type === 'text') { if (p.text) parts.push({ type: 'text', text: p.text }); }\n                else if (p.type === 'image') { images.push(p.image); parts.push({ type: 'image' }); }\n                // Not an arm the union can produce, and the mirror of the role-axis guard the\n                // wire mappers carry: an app built against an older aparte can still hand us a\n                // `file` part, and pushing its absent `image` gave `load_image(undefined)`.\n                else droppedParts++;\n            }\n        }\n        if (parts.length > 0) chat.push({ role: m.role, content: parts });\n    }\n    if (droppedToolTurns > 0) warn(TOOL_TURNS_DROPPED);\n    if (droppedParts > 0) warn(UNSUPPORTED_PARTS_DROPPED);\n    return { chat, images };\n}\n\n/** The processor and model, as this runner calls them (Transformers.js types them through `Callable`). */\ninterface VisionProcessor {\n    (text: string, images: unknown[]): Promise<Record<string, unknown>>;\n    /** The text half alone — callable, for a turn that carries no picture. */\n    tokenizer: (text: string) => Record<string, unknown> | Promise<Record<string, unknown>>;\n    apply_chat_template(messages: HFMessage[], options: Record<string, unknown>): unknown;\n}\ninterface VisionModel {\n    generate(args: Record<string, unknown>): Promise<unknown>;\n    dispose(): Promise<unknown>;\n}\n\nexport const createRunner: CreateRunner = async (ctx) => {\n    const tf = ctx.transformers as unknown as {\n        AutoProcessor: { from_pretrained(model: string, options?: unknown): Promise<VisionProcessor> };\n        AutoModelForImageTextToText: { from_pretrained(model: string, options?: unknown): Promise<VisionModel> };\n        load_image(source: string): Promise<unknown>;\n    };\n    // The processor carries no weights: dtype and device are the model's alone.\n    const [processor, model] = await Promise.all([\n        tf.AutoProcessor.from_pretrained(ctx.modelId),\n        tf.AutoModelForImageTextToText.from_pretrained(ctx.modelId, loadOptions(ctx)),\n    ]);\n\n    return {\n        async generate({ messages, options, emit, signal }: RunnerGenerateInput): Promise<void> {\n            const { stopping, release } = interruptOn(signal, ctx.transformers);\n            try {\n                const { chat, images } = toChatTemplate(messages, ctx.warn);\n                const prompt = processor.apply_chat_template(chat, { add_generation_prompt: true, tokenize: false }) as string;\n                // The processor pairs the prompt with pictures and wants at least one (Idefics3's\n                // reads `images.rows` — \"hello\" as a first message crashed it on a real SmolVLM).\n                // A turn with no image is text: the tokenizer alone takes it.\n                const inputs = images.length > 0\n                    ? await processor(prompt, await Promise.all(images.map((src) => tf.load_image(src))))\n                    : await processor.tokenizer(prompt);\n                await model.generate({\n                    ...inputs,\n                    ...generationOptions(options),\n                    streamer: textStreamer(ctx.transformers, processor.tokenizer, emit),\n                    stopping_criteria: stopping,\n                });\n            } finally {\n                release();\n            }\n        },\n        async dispose() {\n            await model.dispose();\n        },\n    };\n};\n"],"names":[],"mappings":";AAwBO,SAAS,eAAe,UAA+B,MAA0E;AACpI,QAAM,OAAoB,CAAA;AAC1B,QAAM,SAAmB,CAAA;AACzB,MAAI,mBAAmB;AACvB,MAAI,eAAe;AACnB,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,UAAU,EAAE,SAAS,eAAe,EAAE,SAAS,UAAU;AAAE;AAAoB;AAAA,IAAU;AAIxG,QAAI,EAAE,WAAW,OAAQ;AACzB,UAAM,QAAkB,CAAA;AACxB,QAAI,OAAO,EAAE,YAAY,UAAU;AAC/B,UAAI,EAAE,QAAS,OAAM,KAAK,EAAE,MAAM,QAAQ,MAAM,EAAE,SAAS;AAAA,IAC/D,OAAO;AACH,iBAAW,KAAK,EAAE,SAAS;AACvB,YAAI,EAAE,SAAS,QAAQ;AAAE,cAAI,EAAE,KAAM,OAAM,KAAK,EAAE,MAAM,QAAQ,MAAM,EAAE,MAAM;AAAA,QAAG,WACxE,EAAE,SAAS,SAAS;AAAE,iBAAO,KAAK,EAAE,KAAK;AAAG,gBAAM,KAAK,EAAE,MAAM,QAAA,CAAS;AAAA,QAAG,MAI/E;AAAA,MACT;AAAA,IACJ;AACA,QAAI,MAAM,SAAS,EAAG,MAAK,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAA,CAAO;AAAA,EACpE;AACA,MAAI,mBAAmB,EAAG,MAAK,kBAAkB;AACjD,MAAI,eAAe,EAAG,MAAK,yBAAyB;AACpD,SAAO,EAAE,MAAM,OAAA;AACnB;AAcO,MAAM,eAA6B,OAAO,QAAQ;AACrD,QAAM,KAAK,IAAI;AAMf,QAAM,CAAC,WAAW,KAAK,IAAI,MAAM,QAAQ,IAAI;AAAA,IACzC,GAAG,cAAc,gBAAgB,IAAI,OAAO;AAAA,IAC5C,GAAG,4BAA4B,gBAAgB,IAAI,SAAS,YAAY,GAAG,CAAC;AAAA,EAAA,CAC/E;AAED,SAAO;AAAA,IACH,MAAM,SAAS,EAAE,UAAU,SAAS,MAAM,UAA8C;AACpF,YAAM,EAAE,UAAU,QAAA,IAAY,YAAY,QAAQ,IAAI,YAAY;AAClE,UAAI;AACA,cAAM,EAAE,MAAM,OAAA,IAAW,eAAe,UAAU,IAAI,IAAI;AAC1D,cAAM,SAAS,UAAU,oBAAoB,MAAM,EAAE,uBAAuB,MAAM,UAAU,OAAO;AAInG,cAAM,SAAS,OAAO,SAAS,IACzB,MAAM,UAAU,QAAQ,MAAM,QAAQ,IAAI,OAAO,IAAI,CAAC,QAAQ,GAAG,WAAW,GAAG,CAAC,CAAC,CAAC,IAClF,MAAM,UAAU,UAAU,MAAM;AACtC,cAAM,MAAM,SAAS;AAAA,UACjB,GAAG;AAAA,UACH,GAAG,kBAAkB,OAAO;AAAA,UAC5B,UAAU,aAAa,IAAI,cAAc,UAAU,WAAW,IAAI;AAAA,UAClE,mBAAmB;AAAA,QAAA,CACtB;AAAA,MACL,UAAA;AACI,gBAAA;AAAA,MACJ;AAAA,IACJ;AAAA,IACA,MAAM,UAAU;AACZ,YAAM,MAAM,QAAA;AAAA,IAChB;AAAA,EAAA;AAER;"}
+{"version":3,"file":"image-text-to-text.js","sources":["../../src/runners/image-text-to-text.ts"],"sourcesContent":["/**\n * The built-in vision runner — a model that reads images and text and writes text\n * (SmolVLM, Qwen2-VL, LFM2-VL, Gemma 3, LLaVA…: everything `AutoModelForImageTextToText`\n * resolves).\n *\n * Transformers.js 4.x has no `image-text-to-text` PIPELINE, so this runner goes through\n * the model classes themselves, the way the SmolVLM examples do: `AutoProcessor` renders\n * the chat template with `{ type: 'image' }` placeholders, the images are decoded beside\n * the prompt in the same order, the processor turns both into tensors, and `generate()`\n * streams through a `TextStreamer`. Tool turns are dropped with the shared warning.\n */\n\nimport type { AparteChatMessage } from '@aparte/core';\nimport type { CreateRunner, RunnerGenerateInput } from './types.js';\nimport { TOOL_TURNS_DROPPED, UNSUPPORTED_PARTS_DROPPED, contextLength, generationOptions, interruptOn, loadOptions, stoppingCriteria, textStreamer } from './shared.js';\n\ntype HFPart = { type: 'image' } | { type: 'text'; text: string };\ntype HFMessage = { role: 'user' | 'assistant' | 'system'; content: HFPart[] };\n\n/**\n * The conversation in the HF chat shape the processor's template expects — every turn's\n * content as parts, an `{ type: 'image' }` placeholder where a picture goes — plus the\n * pictures themselves, in order of appearance, for the processor to pair with them.\n */\nexport function toChatTemplate(messages: AparteChatMessage[], warn: (message: string) => void): { chat: HFMessage[]; images: string[] } {\n    const chat: HFMessage[] = [];\n    const images: string[] = [];\n    let droppedToolTurns = 0;\n    let droppedParts = 0;\n    for (const m of messages) {\n        if (m.role !== 'user' && m.role !== 'assistant' && m.role !== 'system') { droppedToolTurns++; continue; }\n        // An assistant's calls ride on an `assistant` message, which passes the role test\n        // above and, when the model said nothing before them, carries no part either — so\n        // without this the whole turn leaves the prompt with nothing said.\n        if (m.toolCalls?.length) droppedToolTurns++;\n        const parts: HFPart[] = [];\n        if (typeof m.content === 'string') {\n            if (m.content) parts.push({ type: 'text', text: m.content });\n        } else {\n            for (const p of m.content) {\n                if (p.type === 'text') { if (p.text) parts.push({ type: 'text', text: p.text }); }\n                else if (p.type === 'image') { images.push(p.image); parts.push({ type: 'image' }); }\n                // Not an arm the union can produce, and the mirror of the role-axis guard the\n                // wire mappers carry: an app built against an older aparte can still hand us a\n                // `file` part, and pushing its absent `image` gave `load_image(undefined)`.\n                else droppedParts++;\n            }\n        }\n        if (parts.length > 0) chat.push({ role: m.role, content: parts });\n    }\n    if (droppedToolTurns > 0) warn(TOOL_TURNS_DROPPED);\n    if (droppedParts > 0) warn(UNSUPPORTED_PARTS_DROPPED);\n    return { chat, images };\n}\n\n/** The processor and model, as this runner calls them (Transformers.js types them through `Callable`). */\ninterface VisionProcessor {\n    (text: string, images: unknown[]): Promise<Record<string, unknown>>;\n    /** The text half alone — callable, for a turn that carries no picture. */\n    tokenizer: (text: string) => Record<string, unknown> | Promise<Record<string, unknown>>;\n    apply_chat_template(messages: HFMessage[], options: Record<string, unknown>): unknown;\n}\ninterface VisionModel {\n    generate(args: Record<string, unknown>): Promise<unknown>;\n    dispose(): Promise<unknown>;\n    config?: unknown;\n}\n\nexport const createRunner: CreateRunner = async (ctx) => {\n    const tf = ctx.transformers as unknown as {\n        AutoProcessor: { from_pretrained(model: string, options?: unknown): Promise<VisionProcessor> };\n        AutoModelForImageTextToText: { from_pretrained(model: string, options?: unknown): Promise<VisionModel> };\n        load_image(source: string): Promise<unknown>;\n    };\n    // The processor carries no weights: dtype and device are the model's alone.\n    const [processor, model] = await Promise.all([\n        tf.AutoProcessor.from_pretrained(ctx.modelId),\n        tf.AutoModelForImageTextToText.from_pretrained(ctx.modelId, loadOptions(ctx)),\n    ]);\n    const contextSize = contextLength(model.config);\n\n    return {\n        async generate({ messages, options, emit, signal }: RunnerGenerateInput): Promise<void> {\n            const { stopping, release } = interruptOn(signal, ctx.transformers);\n            try {\n                const { chat, images } = toChatTemplate(messages, ctx.warn);\n                const prompt = processor.apply_chat_template(chat, { add_generation_prompt: true, tokenize: false }) as string;\n                // The processor pairs the prompt with pictures and wants at least one (Idefics3's\n                // reads `images.rows` — \"hello\" as a first message crashed it on a real SmolVLM).\n                // A turn with no image is text: the tokenizer alone takes it.\n                const inputs = images.length > 0\n                    ? await processor(prompt, await Promise.all(images.map((src) => tf.load_image(src))))\n                    : await processor.tokenizer(prompt);\n                await model.generate({\n                    ...inputs,\n                    ...generationOptions(options),\n                    streamer: textStreamer(ctx.transformers, processor.tokenizer, emit),\n                    stopping_criteria: stoppingCriteria(stopping, ctx.transformers, contextSize),\n                });\n            } finally {\n                release();\n            }\n        },\n        async dispose() {\n            await model.dispose();\n        },\n    };\n};\n"],"names":[],"mappings":";AAwBO,SAAS,eAAe,UAA+B,MAA0E;AACpI,QAAM,OAAoB,CAAA;AAC1B,QAAM,SAAmB,CAAA;AACzB,MAAI,mBAAmB;AACvB,MAAI,eAAe;AACnB,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,UAAU,EAAE,SAAS,eAAe,EAAE,SAAS,UAAU;AAAE;AAAoB;AAAA,IAAU;AAIxG,QAAI,EAAE,WAAW,OAAQ;AACzB,UAAM,QAAkB,CAAA;AACxB,QAAI,OAAO,EAAE,YAAY,UAAU;AAC/B,UAAI,EAAE,QAAS,OAAM,KAAK,EAAE,MAAM,QAAQ,MAAM,EAAE,SAAS;AAAA,IAC/D,OAAO;AACH,iBAAW,KAAK,EAAE,SAAS;AACvB,YAAI,EAAE,SAAS,QAAQ;AAAE,cAAI,EAAE,KAAM,OAAM,KAAK,EAAE,MAAM,QAAQ,MAAM,EAAE,MAAM;AAAA,QAAG,WACxE,EAAE,SAAS,SAAS;AAAE,iBAAO,KAAK,EAAE,KAAK;AAAG,gBAAM,KAAK,EAAE,MAAM,QAAA,CAAS;AAAA,QAAG,MAI/E;AAAA,MACT;AAAA,IACJ;AACA,QAAI,MAAM,SAAS,EAAG,MAAK,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAA,CAAO;AAAA,EACpE;AACA,MAAI,mBAAmB,EAAG,MAAK,kBAAkB;AACjD,MAAI,eAAe,EAAG,MAAK,yBAAyB;AACpD,SAAO,EAAE,MAAM,OAAA;AACnB;AAeO,MAAM,eAA6B,OAAO,QAAQ;AACrD,QAAM,KAAK,IAAI;AAMf,QAAM,CAAC,WAAW,KAAK,IAAI,MAAM,QAAQ,IAAI;AAAA,IACzC,GAAG,cAAc,gBAAgB,IAAI,OAAO;AAAA,IAC5C,GAAG,4BAA4B,gBAAgB,IAAI,SAAS,YAAY,GAAG,CAAC;AAAA,EAAA,CAC/E;AACD,QAAM,cAAc,cAAc,MAAM,MAAM;AAE9C,SAAO;AAAA,IACH,MAAM,SAAS,EAAE,UAAU,SAAS,MAAM,UAA8C;AACpF,YAAM,EAAE,UAAU,QAAA,IAAY,YAAY,QAAQ,IAAI,YAAY;AAClE,UAAI;AACA,cAAM,EAAE,MAAM,OAAA,IAAW,eAAe,UAAU,IAAI,IAAI;AAC1D,cAAM,SAAS,UAAU,oBAAoB,MAAM,EAAE,uBAAuB,MAAM,UAAU,OAAO;AAInG,cAAM,SAAS,OAAO,SAAS,IACzB,MAAM,UAAU,QAAQ,MAAM,QAAQ,IAAI,OAAO,IAAI,CAAC,QAAQ,GAAG,WAAW,GAAG,CAAC,CAAC,CAAC,IAClF,MAAM,UAAU,UAAU,MAAM;AACtC,cAAM,MAAM,SAAS;AAAA,UACjB,GAAG;AAAA,UACH,GAAG,kBAAkB,OAAO;AAAA,UAC5B,UAAU,aAAa,IAAI,cAAc,UAAU,WAAW,IAAI;AAAA,UAClE,mBAAmB,iBAAiB,UAAU,IAAI,cAAc,WAAW;AAAA,QAAA,CAC9E;AAAA,MACL,UAAA;AACI,gBAAA;AAAA,MACJ;AAAA,IACJ;AAAA,IACA,MAAM,UAAU;AACZ,YAAM,MAAM,QAAA;AAAA,IAChB;AAAA,EAAA;AAER;"}

package/dist/runners/shared.d.ts

@@ -31,9 +31,33 @@ export declare function interruptOn(signal: AbortSignal, transformers: Transform
 };
 /** 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. */
+/**
+ * The reply's ceiling when no `maxTokens` is set: long enough that an answer is not cut
+ * (some 6,000 words), short enough that a small model stuck repeating itself stops. Decoding
+ * is greedy by default, and in the browser the attention cache grows with every token — a
+ * loop left to run to a 131,072-token window ends in a GPU out-of-memory, not a clean stop.
+ */
+export declare const DEFAULT_MAX_NEW_TOKENS = 8192;
+/**
+ * How many tokens the model's window holds, prompt and reply together, read from its
+ * config — a vision model keeps it on the text half. `undefined` when the config says nothing.
+ */
+export declare function contextLength(config: unknown): number | undefined;
+/**
+ * Sampling options in Transformers.js' vocabulary, from the request's. The reply's ceiling is
+ * the caller's `maxTokens`, else {@link DEFAULT_MAX_NEW_TOKENS} — always set: left out, the
+ * text-generation pipeline would apply its own default of 256.
+ */
 export declare function generationOptions(options: {
     maxTokens?: number;
     temperature?: number;
 }): Record<string, unknown>;
+/**
+ * The stopping criteria a generate runs with: the Stop signal's, plus — when the model's
+ * window is known — one that ends the reply when prompt and reply fill that window. A ceiling
+ * larger than what the window has left would otherwise run the model past its context, where
+ * it degrades; this bounds every ceiling, the default and the caller's, without having to
+ * count the prompt's tokens first.
+ */
+export declare function stoppingCriteria(stopping: unknown, transformers: TransformersModule, contextSize: number | undefined): unknown;
 //# sourceMappingURL=shared.d.ts.map

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

@@ -1 +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,QAI6C,CAAC;AAE7E;;;;;;;GAOG;AACH,eAAO,MAAM,yBAAyB,QAIJ,CAAC;AAEnC;;;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"}
+{"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,QAI6C,CAAC;AAE7E;;;;;;;GAOG;AACH,eAAO,MAAM,yBAAyB,QAIJ,CAAC;AAEnC;;;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;;;;;GAKG;AACH,eAAO,MAAM,sBAAsB,OAAO,CAAC;AAE3C;;;GAGG;AACH,wBAAgB,aAAa,CAAC,MAAM,EAAE,OAAO,GAAG,MAAM,GAAG,SAAS,CAOjE;AAED;;;;GAIG;AACH,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;AAED;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,QAAQ,EAAE,OAAO,EAAE,YAAY,EAAE,kBAAkB,EAAE,WAAW,EAAE,MAAM,GAAG,SAAS,GAAG,OAAO,CAI9H"}

package/dist/runners/shared.js

@@ -32,20 +32,36 @@ function textStreamer(transformers, tokenizer, emit) {
     }
   });
 }
+const DEFAULT_MAX_NEW_TOKENS = 8192;
+function contextLength(config) {
+  const c = config;
+  const text = c?.["text_config"];
+  for (const value of [c?.["max_position_embeddings"], text?.["max_position_embeddings"], c?.["n_positions"]]) {
+    if (typeof value === "number" && value > 0) return value;
+  }
+  return void 0;
+}
 function generationOptions(options) {
   const temperature = options.temperature ?? 0;
   return {
-    max_new_tokens: options.maxTokens ?? 512,
+    max_new_tokens: options.maxTokens ?? DEFAULT_MAX_NEW_TOKENS,
     do_sample: temperature > 0,
     temperature: temperature > 0 ? temperature : void 0
   };
 }
+function stoppingCriteria(stopping, transformers, contextSize) {
+  if (contextSize === void 0) return stopping;
+  const MaxLengthCriteria = transformers.MaxLengthCriteria;
+  return [stopping, new MaxLengthCriteria(contextSize)];
+}
 export {
   TOOL_TURNS_DROPPED as T,
   UNSUPPORTED_PARTS_DROPPED as U,
+  contextLength as c,
   generationOptions as g,
   interruptOn as i,
   loadOptions as l,
+  stoppingCriteria as s,
   textStreamer as t
 };
 //# sourceMappingURL=shared.js.map

package/dist/runners/shared.js.map

@@ -1 +1 @@
-{"version":3,"file":"shared.js","sources":["../../src/runners/shared.ts"],"sourcesContent":["/**\n * What the two built-in runners share — kept in one place so the two cannot drift on\n * how progress is reported, how a stop reaches the model, or what a dropped tool turn\n * says. Types only from core (see `text-generation.ts` for why).\n */\n\nimport type { AparteStreamEvent } from '@aparte/core';\nimport type { RunnerContext, TransformersModule } from './types.js';\n\nexport const TOOL_TURNS_DROPPED =\n    'Dropped the tool call(s) and their results from the prompt: this runner does not support '\n    + 'tool calling, so the model sees neither the call the assistant made nor the result that '\n    + 'answered it. What the assistant SAID before calling stays in the prompt. Use an '\n    + 'OpenAI-compatible endpoint for tools, or a runner that renders them.';\n\n/**\n * A content part neither built-in runner can carry. The union is text and image, so this\n * is unreachable from typed code — and reachable all the same from an app built against an\n * older aparte, which declared a third `file` part nothing ever filled. The wire mappers\n * guard the same case on the ROLE axis (`openai-compat`'s removed `tool_call`/`tool_result`);\n * the part axis had no guard, so the part was pushed into the image list as `undefined` and\n * `load_image` threw at generate time instead of the part being counted and named.\n */\nexport const UNSUPPORTED_PARTS_DROPPED =\n    'Dropped content part(s) this runner cannot carry: it takes text and image parts only, so '\n    + 'anything else leaves the prompt with nothing in its place. A `file` part is the case to '\n    + 'expect — it was removed from `AparteContentPart`; inline what you want the model to read '\n    + 'as text, or send an image.';\n\n/**\n * The options a `from_pretrained` / `pipeline()` call takes from the context: download\n * progress forwarded to the page (percentages, rounded), dtype and device when set.\n */\nexport function loadOptions(ctx: RunnerContext): Record<string, unknown> {\n    const opts: Record<string, unknown> = {\n        progress_callback: (p: { status?: string; file?: string; progress?: number }) => {\n            if (p.status === 'progress') ctx.progress({ status: 'downloading', file: p.file, progress: Math.round(p.progress ?? 0) });\n            else if (p.status === 'done') ctx.progress({ status: 'loading', file: p.file });\n        },\n    };\n    if (ctx.dtype) opts['dtype'] = ctx.dtype;\n    if (ctx.device && ctx.device !== 'auto') opts['device'] = ctx.device;\n    return opts;\n}\n\n/**\n * A stopping criteria the signal interrupts — so a Stop actually STOPS the model, not just\n * the read; otherwise generation runs to `max_new_tokens` off-thread, spending exactly the\n * CPU/GPU/battery this provider exists to save. Call `release()` in a `finally`.\n */\nexport function interruptOn(signal: AbortSignal, transformers: TransformersModule): { stopping: unknown; release(): void } {\n    const stopping = new transformers.InterruptableStoppingCriteria();\n    const onAbort = (): void => { stopping.interrupt(); };\n    if (signal.aborted) onAbort();\n    else signal.addEventListener('abort', onAbort, { once: true });\n    return { stopping, release: () => { signal.removeEventListener('abort', onAbort); } };\n}\n\n/** A `TextStreamer` that emits each decoded token as a `text` event, prompt skipped. */\nexport function textStreamer(transformers: TransformersModule, tokenizer: unknown, emit: (event: AparteStreamEvent) => void): unknown {\n    const TextStreamer = transformers.TextStreamer as unknown as new (tokenizer: unknown, options: Record<string, unknown>) => unknown;\n    return new TextStreamer(tokenizer, {\n        skip_prompt: true,\n        skip_special_tokens: true,\n        callback_function: (text: string) => { if (text) emit({ type: 'text', delta: text }); },\n    });\n}\n\n/** Sampling options in Transformers.js' vocabulary, from the request's. */\nexport function generationOptions(options: { maxTokens?: number; temperature?: number }): Record<string, unknown> {\n    const temperature = options.temperature ?? 0;\n    return {\n        max_new_tokens: options.maxTokens ?? 512,\n        do_sample: temperature > 0,\n        temperature: temperature > 0 ? temperature : undefined,\n    };\n}\n"],"names":[],"mappings":"AASO,MAAM,qBACT;AAaG,MAAM,4BACT;AASG,SAAS,YAAY,KAA6C;AACrE,QAAM,OAAgC;AAAA,IAClC,mBAAmB,CAAC,MAA6D;AAC7E,UAAI,EAAE,WAAW,gBAAgB,SAAS,EAAE,QAAQ,eAAe,MAAM,EAAE,MAAM,UAAU,KAAK,MAAM,EAAE,YAAY,CAAC,GAAG;AAAA,eAC/G,EAAE,WAAW,OAAQ,KAAI,SAAS,EAAE,QAAQ,WAAW,MAAM,EAAE,KAAA,CAAM;AAAA,IAClF;AAAA,EAAA;AAEJ,MAAI,IAAI,MAAO,MAAK,OAAO,IAAI,IAAI;AACnC,MAAI,IAAI,UAAU,IAAI,WAAW,OAAQ,MAAK,QAAQ,IAAI,IAAI;AAC9D,SAAO;AACX;AAOO,SAAS,YAAY,QAAqB,cAA0E;AACvH,QAAM,WAAW,IAAI,aAAa,8BAAA;AAClC,QAAM,UAAU,MAAY;AAAE,aAAS,UAAA;AAAA,EAAa;AACpD,MAAI,OAAO,QAAS,SAAA;AAAA,cACR,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM;AAC7D,SAAO,EAAE,UAAU,SAAS,MAAM;AAAE,WAAO,oBAAoB,SAAS,OAAO;AAAA,EAAG,EAAA;AACtF;AAGO,SAAS,aAAa,cAAkC,WAAoB,MAAmD;AAClI,QAAM,eAAe,aAAa;AAClC,SAAO,IAAI,aAAa,WAAW;AAAA,IAC/B,aAAa;AAAA,IACb,qBAAqB;AAAA,IACrB,mBAAmB,CAAC,SAAiB;AAAE,UAAI,KAAM,MAAK,EAAE,MAAM,QAAQ,OAAO,MAAM;AAAA,IAAG;AAAA,EAAA,CACzF;AACL;AAGO,SAAS,kBAAkB,SAAgF;AAC9G,QAAM,cAAc,QAAQ,eAAe;AAC3C,SAAO;AAAA,IACH,gBAAgB,QAAQ,aAAa;AAAA,IACrC,WAAW,cAAc;AAAA,IACzB,aAAa,cAAc,IAAI,cAAc;AAAA,EAAA;AAErD;"}
+{"version":3,"file":"shared.js","sources":["../../src/runners/shared.ts"],"sourcesContent":["/**\n * What the two built-in runners share — kept in one place so the two cannot drift on\n * how progress is reported, how a stop reaches the model, or what a dropped tool turn\n * says. Types only from core (see `text-generation.ts` for why).\n */\n\nimport type { AparteStreamEvent } from '@aparte/core';\nimport type { RunnerContext, TransformersModule } from './types.js';\n\nexport const TOOL_TURNS_DROPPED =\n    'Dropped the tool call(s) and their results from the prompt: this runner does not support '\n    + 'tool calling, so the model sees neither the call the assistant made nor the result that '\n    + 'answered it. What the assistant SAID before calling stays in the prompt. Use an '\n    + 'OpenAI-compatible endpoint for tools, or a runner that renders them.';\n\n/**\n * A content part neither built-in runner can carry. The union is text and image, so this\n * is unreachable from typed code — and reachable all the same from an app built against an\n * older aparte, which declared a third `file` part nothing ever filled. The wire mappers\n * guard the same case on the ROLE axis (`openai-compat`'s removed `tool_call`/`tool_result`);\n * the part axis had no guard, so the part was pushed into the image list as `undefined` and\n * `load_image` threw at generate time instead of the part being counted and named.\n */\nexport const UNSUPPORTED_PARTS_DROPPED =\n    'Dropped content part(s) this runner cannot carry: it takes text and image parts only, so '\n    + 'anything else leaves the prompt with nothing in its place. A `file` part is the case to '\n    + 'expect — it was removed from `AparteContentPart`; inline what you want the model to read '\n    + 'as text, or send an image.';\n\n/**\n * The options a `from_pretrained` / `pipeline()` call takes from the context: download\n * progress forwarded to the page (percentages, rounded), dtype and device when set.\n */\nexport function loadOptions(ctx: RunnerContext): Record<string, unknown> {\n    const opts: Record<string, unknown> = {\n        progress_callback: (p: { status?: string; file?: string; progress?: number }) => {\n            if (p.status === 'progress') ctx.progress({ status: 'downloading', file: p.file, progress: Math.round(p.progress ?? 0) });\n            else if (p.status === 'done') ctx.progress({ status: 'loading', file: p.file });\n        },\n    };\n    if (ctx.dtype) opts['dtype'] = ctx.dtype;\n    if (ctx.device && ctx.device !== 'auto') opts['device'] = ctx.device;\n    return opts;\n}\n\n/**\n * A stopping criteria the signal interrupts — so a Stop actually STOPS the model, not just\n * the read; otherwise generation runs to `max_new_tokens` off-thread, spending exactly the\n * CPU/GPU/battery this provider exists to save. Call `release()` in a `finally`.\n */\nexport function interruptOn(signal: AbortSignal, transformers: TransformersModule): { stopping: unknown; release(): void } {\n    const stopping = new transformers.InterruptableStoppingCriteria();\n    const onAbort = (): void => { stopping.interrupt(); };\n    if (signal.aborted) onAbort();\n    else signal.addEventListener('abort', onAbort, { once: true });\n    return { stopping, release: () => { signal.removeEventListener('abort', onAbort); } };\n}\n\n/** A `TextStreamer` that emits each decoded token as a `text` event, prompt skipped. */\nexport function textStreamer(transformers: TransformersModule, tokenizer: unknown, emit: (event: AparteStreamEvent) => void): unknown {\n    const TextStreamer = transformers.TextStreamer as unknown as new (tokenizer: unknown, options: Record<string, unknown>) => unknown;\n    return new TextStreamer(tokenizer, {\n        skip_prompt: true,\n        skip_special_tokens: true,\n        callback_function: (text: string) => { if (text) emit({ type: 'text', delta: text }); },\n    });\n}\n\n/**\n * The reply's ceiling when no `maxTokens` is set: long enough that an answer is not cut\n * (some 6,000 words), short enough that a small model stuck repeating itself stops. Decoding\n * is greedy by default, and in the browser the attention cache grows with every token — a\n * loop left to run to a 131,072-token window ends in a GPU out-of-memory, not a clean stop.\n */\nexport const DEFAULT_MAX_NEW_TOKENS = 8192;\n\n/**\n * How many tokens the model's window holds, prompt and reply together, read from its\n * config — a vision model keeps it on the text half. `undefined` when the config says nothing.\n */\nexport function contextLength(config: unknown): number | undefined {\n    const c = config as Record<string, unknown> | undefined;\n    const text = c?.['text_config'] as Record<string, unknown> | undefined;\n    for (const value of [c?.['max_position_embeddings'], text?.['max_position_embeddings'], c?.['n_positions']]) {\n        if (typeof value === 'number' && value > 0) return value;\n    }\n    return undefined;\n}\n\n/**\n * Sampling options in Transformers.js' vocabulary, from the request's. The reply's ceiling is\n * the caller's `maxTokens`, else {@link DEFAULT_MAX_NEW_TOKENS} — always set: left out, the\n * text-generation pipeline would apply its own default of 256.\n */\nexport function generationOptions(options: { maxTokens?: number; temperature?: number }): Record<string, unknown> {\n    const temperature = options.temperature ?? 0;\n    return {\n        max_new_tokens: options.maxTokens ?? DEFAULT_MAX_NEW_TOKENS,\n        do_sample: temperature > 0,\n        temperature: temperature > 0 ? temperature : undefined,\n    };\n}\n\n/**\n * The stopping criteria a generate runs with: the Stop signal's, plus — when the model's\n * window is known — one that ends the reply when prompt and reply fill that window. A ceiling\n * larger than what the window has left would otherwise run the model past its context, where\n * it degrades; this bounds every ceiling, the default and the caller's, without having to\n * count the prompt's tokens first.\n */\nexport function stoppingCriteria(stopping: unknown, transformers: TransformersModule, contextSize: number | undefined): unknown {\n    if (contextSize === undefined) return stopping;\n    const MaxLengthCriteria = transformers.MaxLengthCriteria as unknown as new (maxLength: number) => unknown;\n    return [stopping, new MaxLengthCriteria(contextSize)];\n}\n"],"names":[],"mappings":"AASO,MAAM,qBACT;AAaG,MAAM,4BACT;AASG,SAAS,YAAY,KAA6C;AACrE,QAAM,OAAgC;AAAA,IAClC,mBAAmB,CAAC,MAA6D;AAC7E,UAAI,EAAE,WAAW,gBAAgB,SAAS,EAAE,QAAQ,eAAe,MAAM,EAAE,MAAM,UAAU,KAAK,MAAM,EAAE,YAAY,CAAC,GAAG;AAAA,eAC/G,EAAE,WAAW,OAAQ,KAAI,SAAS,EAAE,QAAQ,WAAW,MAAM,EAAE,KAAA,CAAM;AAAA,IAClF;AAAA,EAAA;AAEJ,MAAI,IAAI,MAAO,MAAK,OAAO,IAAI,IAAI;AACnC,MAAI,IAAI,UAAU,IAAI,WAAW,OAAQ,MAAK,QAAQ,IAAI,IAAI;AAC9D,SAAO;AACX;AAOO,SAAS,YAAY,QAAqB,cAA0E;AACvH,QAAM,WAAW,IAAI,aAAa,8BAAA;AAClC,QAAM,UAAU,MAAY;AAAE,aAAS,UAAA;AAAA,EAAa;AACpD,MAAI,OAAO,QAAS,SAAA;AAAA,cACR,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM;AAC7D,SAAO,EAAE,UAAU,SAAS,MAAM;AAAE,WAAO,oBAAoB,SAAS,OAAO;AAAA,EAAG,EAAA;AACtF;AAGO,SAAS,aAAa,cAAkC,WAAoB,MAAmD;AAClI,QAAM,eAAe,aAAa;AAClC,SAAO,IAAI,aAAa,WAAW;AAAA,IAC/B,aAAa;AAAA,IACb,qBAAqB;AAAA,IACrB,mBAAmB,CAAC,SAAiB;AAAE,UAAI,KAAM,MAAK,EAAE,MAAM,QAAQ,OAAO,MAAM;AAAA,IAAG;AAAA,EAAA,CACzF;AACL;AAQO,MAAM,yBAAyB;AAM/B,SAAS,cAAc,QAAqC;AAC/D,QAAM,IAAI;AACV,QAAM,OAAO,IAAI,aAAa;AAC9B,aAAW,SAAS,CAAC,IAAI,yBAAyB,GAAG,OAAO,yBAAyB,GAAG,IAAI,aAAa,CAAC,GAAG;AACzG,QAAI,OAAO,UAAU,YAAY,QAAQ,EAAG,QAAO;AAAA,EACvD;AACA,SAAO;AACX;AAOO,SAAS,kBAAkB,SAAgF;AAC9G,QAAM,cAAc,QAAQ,eAAe;AAC3C,SAAO;AAAA,IACH,gBAAgB,QAAQ,aAAa;AAAA,IACrC,WAAW,cAAc;AAAA,IACzB,aAAa,cAAc,IAAI,cAAc;AAAA,EAAA;AAErD;AASO,SAAS,iBAAiB,UAAmB,cAAkC,aAA0C;AAC5H,MAAI,gBAAgB,OAAW,QAAO;AACtC,QAAM,oBAAoB,aAAa;AACvC,SAAO,CAAC,UAAU,IAAI,kBAAkB,WAAW,CAAC;AACxD;"}

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

@@ -1 +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,CAoBlH;AASD,eAAO,MAAM,YAAY,EAAE,YAqB1B,CAAC"}
+{"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,CAoBlH;AAUD,eAAO,MAAM,YAAY,EAAE,YAsB1B,CAAC"}

package/dist/runners/text-generation.js

@@ -1,4 +1,4 @@
-import { l as loadOptions, i as interruptOn, t as textStreamer, g as generationOptions, T as TOOL_TURNS_DROPPED } from "./shared.js";
+import { l as loadOptions, c as contextLength, i as interruptOn, s as stoppingCriteria, t as textStreamer, g as generationOptions, T as TOOL_TURNS_DROPPED } from "./shared.js";
 const IMAGES_DROPPED = 'This model has no vision runner: image parts were dropped from the prompt, so the model answers as if there were none. Register the model with task: "image-text-to-text", or point `runner` at a module of your own.';
 function textOf(content) {
   if (typeof content === "string") return content;
@@ -25,6 +25,7 @@ function flattenForChatTemplate(messages, warn) {
 const createRunner = async (ctx) => {
   const pipeline = ctx.transformers.pipeline;
   const pipe = await pipeline("text-generation", ctx.modelId, loadOptions(ctx));
+  const contextSize = contextLength(pipe.model?.config);
   return {
     async generate({ messages, options, emit, signal }) {
       const { stopping, release } = interruptOn(signal, ctx.transformers);
@@ -32,7 +33,7 @@ const createRunner = async (ctx) => {
         await pipe(flattenForChatTemplate(messages, ctx.warn), {
           ...generationOptions(options),
           streamer: textStreamer(ctx.transformers, pipe.tokenizer, emit),
-          stopping_criteria: stopping
+          stopping_criteria: stoppingCriteria(stopping, ctx.transformers, contextSize)
         });
       } finally {
         release();

package/dist/runners/text-generation.js.map

@@ -1 +1 @@
-{"version":3,"file":"text-generation.js","sources":["../../src/runners/text-generation.ts"],"sourcesContent":["/**\n * The built-in text runner — the generic `pipeline('text-generation')` path this\n * provider has always run, extracted from the worker so it is one runner among others.\n *\n * It flattens the conversation to `{ role, content: string }` turns (the tokenizer applies\n * the chat template). Two things it cannot carry, it SAYS: tool turns (their wire syntax is\n * model-specific) and image parts (a text model has no eyes). The second used to vanish\n * silently — a photo attached to a text model produced an answer that pretended — and\n * that silence, not the limitation, was the defect.\n */\n\n// Types only. A runner runs INSIDE the worker, and the worker bundle has no way to\n// resolve `@aparte/core` at runtime (an import map does not reach a worker) — so a value\n// import here is not externalised, it is inlined: the first build that imported\n// `contentToText` shipped all of core, components included, in a 426 kB runner chunk.\nimport type { AparteChatMessage, AparteContentPart } from '@aparte/core';\nimport type { CreateRunner, RunnerContext, RunnerGenerateInput } from './types.js';\nimport { TOOL_TURNS_DROPPED, generationOptions, interruptOn, loadOptions, textStreamer } from './shared.js';\n\ntype SimpleMessage = { role: 'user' | 'assistant' | 'system'; content: string };\n\nexport const IMAGES_DROPPED =\n    'This model has no vision runner: image parts were dropped from the prompt, so the model '\n    + 'answers as if there were none. Register the model with task: \"image-text-to-text\", or '\n    + 'point `runner` at a module of your own.';\n\n/** The text parts of a message, joined — what a text-only chat template can take. */\nfunction textOf(content: string | AparteContentPart[]): string {\n    if (typeof content === 'string') return content;\n    return content\n        .filter((p): p is Extract<AparteContentPart, { type: 'text' }> => p.type === 'text')\n        .map((p) => p.text)\n        .join('');\n}\n\n/** Flatten to what the chat template takes; say what was left out. */\nexport function flattenForChatTemplate(messages: AparteChatMessage[], warn: RunnerContext['warn']): SimpleMessage[] {\n    const result: SimpleMessage[] = [];\n    let droppedImages = 0;\n    let droppedToolTurns = 0;\n    for (const m of messages) {\n        if (m.role === 'user' || m.role === 'assistant' || m.role === 'system') {\n            // An assistant's calls ride on an `assistant` message, which passes the role test\n            // above and, when the model said nothing before them, carries no text either — so\n            // without this the whole turn leaves the prompt with nothing said.\n            if (m.toolCalls?.length) droppedToolTurns++;\n            if (Array.isArray(m.content)) droppedImages += m.content.filter((p) => p.type === 'image').length;\n            const text = textOf(m.content);\n            if (text) result.push({ role: m.role, content: text });\n        } else {\n            droppedToolTurns++;\n        }\n    }\n    if (droppedImages > 0) warn(IMAGES_DROPPED);\n    if (droppedToolTurns > 0) warn(TOOL_TURNS_DROPPED);\n    return result;\n}\n\n/** The pipeline, as this runner calls it. Transformers.js types it per task; this is the one task. */\ninterface TextPipeline {\n    (messages: SimpleMessage[], options: Record<string, unknown>): Promise<unknown>;\n    tokenizer: unknown;\n    dispose?: () => Promise<void>;\n}\n\nexport const createRunner: CreateRunner = async (ctx) => {\n    const pipeline = ctx.transformers.pipeline as unknown as (task: string, model: string, options: unknown) => Promise<TextPipeline>;\n    const pipe = await pipeline('text-generation', ctx.modelId, loadOptions(ctx));\n\n    return {\n        async generate({ messages, options, emit, signal }: RunnerGenerateInput): Promise<void> {\n            const { stopping, release } = interruptOn(signal, ctx.transformers);\n            try {\n                await pipe(flattenForChatTemplate(messages, ctx.warn), {\n                    ...generationOptions(options),\n                    streamer: textStreamer(ctx.transformers, pipe.tokenizer, emit),\n                    stopping_criteria: stopping,\n                });\n            } finally {\n                release();\n            }\n        },\n        dispose() {\n            return pipe.dispose?.();\n        },\n    };\n};\n"],"names":[],"mappings":";AAqBO,MAAM,iBACT;AAKJ,SAAS,OAAO,SAA+C;AAC3D,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,SAAO,QACF,OAAO,CAAC,MAAyD,EAAE,SAAS,MAAM,EAClF,IAAI,CAAC,MAAM,EAAE,IAAI,EACjB,KAAK,EAAE;AAChB;AAGO,SAAS,uBAAuB,UAA+B,MAA8C;AAChH,QAAM,SAA0B,CAAA;AAChC,MAAI,gBAAgB;AACpB,MAAI,mBAAmB;AACvB,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,UAAU,EAAE,SAAS,eAAe,EAAE,SAAS,UAAU;AAIpE,UAAI,EAAE,WAAW,OAAQ;AACzB,UAAI,MAAM,QAAQ,EAAE,OAAO,EAAG,kBAAiB,EAAE,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO,EAAE;AAC3F,YAAM,OAAO,OAAO,EAAE,OAAO;AAC7B,UAAI,aAAa,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAM;AAAA,IACzD,OAAO;AACH;AAAA,IACJ;AAAA,EACJ;AACA,MAAI,gBAAgB,EAAG,MAAK,cAAc;AAC1C,MAAI,mBAAmB,EAAG,MAAK,kBAAkB;AACjD,SAAO;AACX;AASO,MAAM,eAA6B,OAAO,QAAQ;AACrD,QAAM,WAAW,IAAI,aAAa;AAClC,QAAM,OAAO,MAAM,SAAS,mBAAmB,IAAI,SAAS,YAAY,GAAG,CAAC;AAE5E,SAAO;AAAA,IACH,MAAM,SAAS,EAAE,UAAU,SAAS,MAAM,UAA8C;AACpF,YAAM,EAAE,UAAU,QAAA,IAAY,YAAY,QAAQ,IAAI,YAAY;AAClE,UAAI;AACA,cAAM,KAAK,uBAAuB,UAAU,IAAI,IAAI,GAAG;AAAA,UACnD,GAAG,kBAAkB,OAAO;AAAA,UAC5B,UAAU,aAAa,IAAI,cAAc,KAAK,WAAW,IAAI;AAAA,UAC7D,mBAAmB;AAAA,QAAA,CACtB;AAAA,MACL,UAAA;AACI,gBAAA;AAAA,MACJ;AAAA,IACJ;AAAA,IACA,UAAU;AACN,aAAO,KAAK,UAAA;AAAA,IAChB;AAAA,EAAA;AAER;"}
+{"version":3,"file":"text-generation.js","sources":["../../src/runners/text-generation.ts"],"sourcesContent":["/**\n * The built-in text runner — the generic `pipeline('text-generation')` path this\n * provider has always run, extracted from the worker so it is one runner among others.\n *\n * It flattens the conversation to `{ role, content: string }` turns (the tokenizer applies\n * the chat template). Two things it cannot carry, it SAYS: tool turns (their wire syntax is\n * model-specific) and image parts (a text model has no eyes). The second used to vanish\n * silently — a photo attached to a text model produced an answer that pretended — and\n * that silence, not the limitation, was the defect.\n */\n\n// Types only. A runner runs INSIDE the worker, and the worker bundle has no way to\n// resolve `@aparte/core` at runtime (an import map does not reach a worker) — so a value\n// import here is not externalised, it is inlined: the first build that imported\n// `contentToText` shipped all of core, components included, in a 426 kB runner chunk.\nimport type { AparteChatMessage, AparteContentPart } from '@aparte/core';\nimport type { CreateRunner, RunnerContext, RunnerGenerateInput } from './types.js';\nimport { TOOL_TURNS_DROPPED, contextLength, generationOptions, interruptOn, loadOptions, stoppingCriteria, textStreamer } from './shared.js';\n\ntype SimpleMessage = { role: 'user' | 'assistant' | 'system'; content: string };\n\nexport const IMAGES_DROPPED =\n    'This model has no vision runner: image parts were dropped from the prompt, so the model '\n    + 'answers as if there were none. Register the model with task: \"image-text-to-text\", or '\n    + 'point `runner` at a module of your own.';\n\n/** The text parts of a message, joined — what a text-only chat template can take. */\nfunction textOf(content: string | AparteContentPart[]): string {\n    if (typeof content === 'string') return content;\n    return content\n        .filter((p): p is Extract<AparteContentPart, { type: 'text' }> => p.type === 'text')\n        .map((p) => p.text)\n        .join('');\n}\n\n/** Flatten to what the chat template takes; say what was left out. */\nexport function flattenForChatTemplate(messages: AparteChatMessage[], warn: RunnerContext['warn']): SimpleMessage[] {\n    const result: SimpleMessage[] = [];\n    let droppedImages = 0;\n    let droppedToolTurns = 0;\n    for (const m of messages) {\n        if (m.role === 'user' || m.role === 'assistant' || m.role === 'system') {\n            // An assistant's calls ride on an `assistant` message, which passes the role test\n            // above and, when the model said nothing before them, carries no text either — so\n            // without this the whole turn leaves the prompt with nothing said.\n            if (m.toolCalls?.length) droppedToolTurns++;\n            if (Array.isArray(m.content)) droppedImages += m.content.filter((p) => p.type === 'image').length;\n            const text = textOf(m.content);\n            if (text) result.push({ role: m.role, content: text });\n        } else {\n            droppedToolTurns++;\n        }\n    }\n    if (droppedImages > 0) warn(IMAGES_DROPPED);\n    if (droppedToolTurns > 0) warn(TOOL_TURNS_DROPPED);\n    return result;\n}\n\n/** The pipeline, as this runner calls it. Transformers.js types it per task; this is the one task. */\ninterface TextPipeline {\n    (messages: SimpleMessage[], options: Record<string, unknown>): Promise<unknown>;\n    tokenizer: unknown;\n    model?: { config?: unknown };\n    dispose?: () => Promise<void>;\n}\n\nexport const createRunner: CreateRunner = async (ctx) => {\n    const pipeline = ctx.transformers.pipeline as unknown as (task: string, model: string, options: unknown) => Promise<TextPipeline>;\n    const pipe = await pipeline('text-generation', ctx.modelId, loadOptions(ctx));\n    const contextSize = contextLength(pipe.model?.config);\n\n    return {\n        async generate({ messages, options, emit, signal }: RunnerGenerateInput): Promise<void> {\n            const { stopping, release } = interruptOn(signal, ctx.transformers);\n            try {\n                await pipe(flattenForChatTemplate(messages, ctx.warn), {\n                    ...generationOptions(options),\n                    streamer: textStreamer(ctx.transformers, pipe.tokenizer, emit),\n                    stopping_criteria: stoppingCriteria(stopping, ctx.transformers, contextSize),\n                });\n            } finally {\n                release();\n            }\n        },\n        dispose() {\n            return pipe.dispose?.();\n        },\n    };\n};\n"],"names":[],"mappings":";AAqBO,MAAM,iBACT;AAKJ,SAAS,OAAO,SAA+C;AAC3D,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,SAAO,QACF,OAAO,CAAC,MAAyD,EAAE,SAAS,MAAM,EAClF,IAAI,CAAC,MAAM,EAAE,IAAI,EACjB,KAAK,EAAE;AAChB;AAGO,SAAS,uBAAuB,UAA+B,MAA8C;AAChH,QAAM,SAA0B,CAAA;AAChC,MAAI,gBAAgB;AACpB,MAAI,mBAAmB;AACvB,aAAW,KAAK,UAAU;AACtB,QAAI,EAAE,SAAS,UAAU,EAAE,SAAS,eAAe,EAAE,SAAS,UAAU;AAIpE,UAAI,EAAE,WAAW,OAAQ;AACzB,UAAI,MAAM,QAAQ,EAAE,OAAO,EAAG,kBAAiB,EAAE,QAAQ,OAAO,CAAC,MAAM,EAAE,SAAS,OAAO,EAAE;AAC3F,YAAM,OAAO,OAAO,EAAE,OAAO;AAC7B,UAAI,aAAa,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,MAAM;AAAA,IACzD,OAAO;AACH;AAAA,IACJ;AAAA,EACJ;AACA,MAAI,gBAAgB,EAAG,MAAK,cAAc;AAC1C,MAAI,mBAAmB,EAAG,MAAK,kBAAkB;AACjD,SAAO;AACX;AAUO,MAAM,eAA6B,OAAO,QAAQ;AACrD,QAAM,WAAW,IAAI,aAAa;AAClC,QAAM,OAAO,MAAM,SAAS,mBAAmB,IAAI,SAAS,YAAY,GAAG,CAAC;AAC5E,QAAM,cAAc,cAAc,KAAK,OAAO,MAAM;AAEpD,SAAO;AAAA,IACH,MAAM,SAAS,EAAE,UAAU,SAAS,MAAM,UAA8C;AACpF,YAAM,EAAE,UAAU,QAAA,IAAY,YAAY,QAAQ,IAAI,YAAY;AAClE,UAAI;AACA,cAAM,KAAK,uBAAuB,UAAU,IAAI,IAAI,GAAG;AAAA,UACnD,GAAG,kBAAkB,OAAO;AAAA,UAC5B,UAAU,aAAa,IAAI,cAAc,KAAK,WAAW,IAAI;AAAA,UAC7D,mBAAmB,iBAAiB,UAAU,IAAI,cAAc,WAAW;AAAA,QAAA,CAC9E;AAAA,MACL,UAAA;AACI,gBAAA;AAAA,MACJ;AAAA,IACJ;AAAA,IACA,UAAU;AACN,aAAO,KAAK,UAAA;AAAA,IAChB;AAAA,EAAA;AAER;"}

package/dist/runners/types.d.ts

@@ -12,6 +12,10 @@
  * what is not a generation (swapping an adapter, warming something up), reports progress
  * with a free-form `detail`, and can dispose. A research-grade provider that today
  * carries its own worker fits here without giving up anything of its own.
+ *
+ * `tool_use` is in what a runner can emit, not in what it is given: `RunnerGenerateInput`
+ * carries no tool definitions, because the provider sends a model none. A runner that calls
+ * tools has to know them itself.
  */
 import type { AparteChatMessage, AparteStreamEvent } from '@aparte/core';
 /** The Transformers.js module, as the worker resolved it — a runner never imports its own copy. */

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

@@ -1 +1 @@
-{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/runners/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AAEzE,mGAAmG;AACnG,MAAM,MAAM,kBAAkB,GAAG,cAAc,2BAA2B,CAAC,CAAC;AAE5E,wEAAwE;AACxE,MAAM,MAAM,KAAK,GAAG,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;AACpD,MAAM,MAAM,MAAM,GAAG,QAAQ,GAAG,MAAM,GAAG,MAAM,CAAC;AAEhD,oFAAoF;AACpF,MAAM,MAAM,aAAa,GAAG,iBAAiB,GAAG,oBAAoB,CAAC;AAErE,MAAM,WAAW,cAAc;IAC3B,MAAM,EAAE,aAAa,GAAG,SAAS,GAAG,QAAQ,GAAG,OAAO,CAAC;IACvD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,aAAa;IACb,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,+FAA+F;IAC/F,MAAM,CAAC,EAAE,OAAO,CAAC;CACpB;AAED,MAAM,WAAW,iBAAiB;IAC9B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,IAAI,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,qCAAqC;AACrC,MAAM,WAAW,aAAa;IAC1B,YAAY,EAAE,kBAAkB,CAAC;IACjC,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,4GAA4G;IAC5G,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,uDAAuD;IACvD,QAAQ,CAAC,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;IAClC,kGAAkG;IAClG,IAAI,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CAC/B;AAED,MAAM,WAAW,mBAAmB;IAChC,kFAAkF;IAClF,QAAQ,EAAE,iBAAiB,EAAE,CAAC;IAC9B,OAAO,EAAE,iBAAiB,CAAC;IAC3B,yFAAyF;IACzF,IAAI,CAAC,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACrC,gEAAgE;IAChE,MAAM,EAAE,WAAW,CAAC;CACvB;AAED,MAAM,WAAW,kBAAkB;IAC/B,wGAAwG;IACxG,QAAQ,CAAC,KAAK,EAAE,mBAAmB,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACpD,uFAAuF;IACvF,OAAO,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;IAC3D,OAAO,CAAC,IAAI,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACpC;AAED,MAAM,MAAM,YAAY,GAAG,CAAC,GAAG,EAAE,aAAa,KAAK,OAAO,CAAC,kBAAkB,CAAC,CAAC;AAE/E,oCAAoC;AACpC,MAAM,WAAW,YAAY;IACzB,YAAY,EAAE,YAAY,CAAC;CAC9B"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/runners/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC;AAEzE,mGAAmG;AACnG,MAAM,MAAM,kBAAkB,GAAG,cAAc,2BAA2B,CAAC,CAAC;AAE5E,wEAAwE;AACxE,MAAM,MAAM,KAAK,GAAG,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;AACpD,MAAM,MAAM,MAAM,GAAG,QAAQ,GAAG,MAAM,GAAG,MAAM,CAAC;AAEhD,oFAAoF;AACpF,MAAM,MAAM,aAAa,GAAG,iBAAiB,GAAG,oBAAoB,CAAC;AAErE,MAAM,WAAW,cAAc;IAC3B,MAAM,EAAE,aAAa,GAAG,SAAS,GAAG,QAAQ,GAAG,OAAO,CAAC;IACvD,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,aAAa;IACb,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,+FAA+F;IAC/F,MAAM,CAAC,EAAE,OAAO,CAAC;CACpB;AAED,MAAM,WAAW,iBAAiB;IAC9B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,IAAI,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,qCAAqC;AACrC,MAAM,WAAW,aAAa;IAC1B,YAAY,EAAE,kBAAkB,CAAC;IACjC,OAAO,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,4GAA4G;IAC5G,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,uDAAuD;IACvD,QAAQ,CAAC,CAAC,EAAE,cAAc,GAAG,IAAI,CAAC;IAClC,kGAAkG;IAClG,IAAI,CAAC,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CAC/B;AAED,MAAM,WAAW,mBAAmB;IAChC,kFAAkF;IAClF,QAAQ,EAAE,iBAAiB,EAAE,CAAC;IAC9B,OAAO,EAAE,iBAAiB,CAAC;IAC3B,yFAAyF;IACzF,IAAI,CAAC,KAAK,EAAE,iBAAiB,GAAG,IAAI,CAAC;IACrC,gEAAgE;IAChE,MAAM,EAAE,WAAW,CAAC;CACvB;AAED,MAAM,WAAW,kBAAkB;IAC/B,wGAAwG;IACxG,QAAQ,CAAC,KAAK,EAAE,mBAAmB,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IACpD,uFAAuF;IACvF,OAAO,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;IAC3D,OAAO,CAAC,IAAI,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CACpC;AAED,MAAM,MAAM,YAAY,GAAG,CAAC,GAAG,EAAE,aAAa,KAAK,OAAO,CAAC,kBAAkB,CAAC,CAAC;AAE/E,oCAAoC;AACpC,MAAM,WAAW,YAAY;IACzB,YAAY,EAAE,YAAY,CAAC;CAC9B"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aparte/provider-transformers",
-  "version": "0.19.3",
+  "version": "0.20.0",
   "homepage": "https://apartejs.dev/providers/ai/transformers/",
   "description": "Run LLMs 100% in the browser via Transformers.js (WebGPU/WASM) — a local, keyless AI provider for aparté. Streams tokens off the main thread in a Web Worker.",
   "type": "module",
@@ -25,7 +25,7 @@
     "node": ">=18"
   },
   "peerDependencies": {
-    "@aparte/core": ">=0.19.3 <1.0.0",
+    "@aparte/core": ">=0.20.0 <1.0.0",
     "@huggingface/transformers": "^4.2.0"
   },
   "devDependencies": {
@@ -33,7 +33,7 @@
     "@types/node": "^22.0.0",
     "typescript": "^5.4.0",
     "vite": "^6.0.0",
-    "@aparte/core": "0.19.3"
+    "@aparte/core": "0.20.0"
   },
   "keywords": [
     "ai",
@@ -56,7 +56,7 @@
   },
   "scripts": {
     "dev": "vite",
-    "build": "vite build && tsc -b --emitDeclarationOnly --force",
+    "build": "vite build && tsc -b --emitDeclarationOnly --force && node ../../../../scripts/stamp-dist.mjs",
     "preview": "vite preview",
     "test": "vitest",
     "test:run": "vitest run",