@aparte/provider-transformers 0.24.0 → 0.26.0

package/README.md

@@ -106,9 +106,11 @@ export async function createRunner(ctx) {
 }
 ```
 
+`input.tools` are the tools the request allows (a `RunnerTool`: `name`, `description`,
+`inputSchema`) and `input.toolChoice` its `'auto'` / `'none'` / `{ name }` — absent when it has none.
 `emit` speaks aparté's stream vocabulary (`text`, `thinking`, `tool_use`, `done`, `error`),
 `signal` fires on Stop, `ctx.progress` (a `RunnerProgress`) / `ctx.warn` reach the page. The contract is
-exported: `TransformersRunner`, `RunnerContext`, `RunnerGenerateInput`, `CreateRunner`, `RunnerModule` (what
+exported: `TransformersRunner`, `RunnerContext`, `RunnerGenerateInput`, `RunnerTool`, `CreateRunner`, `RunnerModule` (what
 the module exports), `BuiltInRunner` (the two `task` names) and `TransformersModule` (the type of
 `ctx.transformers`).
 `runnerCommand(modelId, name, payload)` reaches a runner's `command()` from the page, queued

package/dist/index.d.ts

@@ -122,7 +122,7 @@ export declare function getComputeDevice(): ComputeDevice;
 export declare const TransformersProvider: AparteAIProvider & Required<Pick<AparteAIProvider, 'prepareModel' | 'getModelStatus' | 'chat'>>;
 export default TransformersProvider;
 export type { AparteAIProvider, AparteAIModel, ModelStatus, ModelLoadProgress } from '@aparte/core';
-export type { TransformersRunner, RunnerContext, RunnerGenerateInput, RunnerProgress, RunnerModule, CreateRunner, BuiltInRunner, TransformersModule, } from './runners/types.js';
+export type { TransformersRunner, RunnerContext, RunnerGenerateInput, RunnerTool, RunnerProgress, RunnerModule, CreateRunner, BuiltInRunner, TransformersModule, } from './runners/types.js';
 /** Returns the modelId currently loaded in the worker's pipeline, or null. */
 export declare function getLoadedModelId(): string | null;
 /**

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;;;;;;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"}
+{"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,CAiLhF,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,UAAU,EACV,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

@@ -348,6 +348,10 @@ const TransformersProvider = {
         // is the runner's knowledge, not this thread's.
         messages: request.messages,
         options,
+        // The tools this turn allows, as plain definitions (what crosses to a worker
+        // is cloned), and the request's choice among them.
+        ...request.tools?.length ? { tools: request.tools.map(({ name, description, inputSchema }) => ({ name, description, inputSchema })) } : {},
+        ...request.toolChoice ? { toolChoice: request.toolChoice } : {},
         ..._selection(request.modelId)
       });
     };

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    /**\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;"}
+{"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                // The tools this turn allows, as plain definitions (what crosses to a worker\n                // is cloned), and the request's choice among them.\n                ...(request.tools?.length ? { tools: request.tools.map(({ name, description, inputSchema }) => ({ name, description, inputSchema })) } : {}),\n                ...(request.toolChoice ? { toolChoice: request.toolChoice } : {}),\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    RunnerTool,\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;AAAA;AAAA,QAGA,GAAI,QAAQ,OAAO,SAAS,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC,EAAE,MAAM,aAAa,YAAA,OAAmB,EAAE,MAAM,aAAa,YAAA,EAAc,EAAA,IAAM,CAAA;AAAA,QACzI,GAAI,QAAQ,aAAa,EAAE,YAAY,QAAQ,WAAA,IAAe,CAAA;AAAA,QAC9D,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;AAqBO,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/types.d.ts

@@ -13,11 +13,14 @@
  * 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.
+ * A runner is given the tools the request allows (`tools`, when the model declares
+ * `function_calling` and the app registered some) and the request's `toolChoice`, and can emit
+ * `tool_use` for them. The built-in runners take neither: they warn when a turn needs a tool
+ * call they cannot carry.
  */
-import type { AparteChatMessage, AparteStreamEvent } from '@aparte/core';
+import type { AparteChatMessage, AparteChatRequest, AparteStreamEvent, AparteTool } from '@aparte/core';
+/** A tool as a runner sees it: what a model needs to call it, no more. */
+export type RunnerTool = Pick<AparteTool, 'name' | 'description' | 'inputSchema'>;
 /** The Transformers.js module, as the worker resolved it — a runner never imports its own copy. */
 export type TransformersModule = typeof import('@huggingface/transformers');
 /** ONNX dtype, or a per-part map (`{ decoder_model_merged: 'q4' }`). */
@@ -54,6 +57,10 @@ export interface RunnerGenerateInput {
     /** The conversation, WITH its content parts — the runner decides what to keep. */
     messages: AparteChatMessage[];
     options: GenerationOptions;
+    /** The tools this turn may call — the app's registration, not a list the runner keeps. Absent when there are none. */
+    tools?: RunnerTool[];
+    /** `'auto'`, `'none'` (answer directly this turn) or a tool the model must call. Absent when the request sets none. */
+    toolChoice?: AparteChatRequest['toolChoice'];
     /** The stream, in aparté's vocabulary. `done` and `error` are the host's to close on. */
     emit(event: AparteStreamEvent): void;
     /** Fired when the page cancels; wire it to the model's stop. */

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

@@ -1 +1 @@
-{"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"}
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/runners/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,UAAU,EAAE,MAAM,cAAc,CAAC;AAExG,0EAA0E;AAC1E,MAAM,MAAM,UAAU,GAAG,IAAI,CAAC,UAAU,EAAE,MAAM,GAAG,aAAa,GAAG,aAAa,CAAC,CAAC;AAElF,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,sHAAsH;IACtH,KAAK,CAAC,EAAE,UAAU,EAAE,CAAC;IACrB,uHAAuH;IACvH,UAAU,CAAC,EAAE,iBAAiB,CAAC,YAAY,CAAC,CAAC;IAC7C,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/dist/worker-host.d.ts

@@ -9,7 +9,7 @@
  *
  *   → init            { transformersUrl? }                      once, first
  *   → prepare         { id, modelId, task?, runner?, dtype?, device? }
- *   → generate        { id, modelId, messages, options, task?, runner?, dtype?, device? }
+ *   → generate        { id, modelId, messages, options, tools?, toolChoice?, task?, runner?, dtype?, device? }
  *   → cancel          { id }
  *   → command         { id, modelId, name, payload, task?, runner?, dtype?, device? }
  *   ← progress        { id, status, file?, progress?, detail? }
@@ -26,7 +26,7 @@
  * kept — a local model is gigabytes, and two of them resident is the failure to avoid.
  */
 import type { AparteChatMessage, AparteStreamEvent, AparteUsage } from '@aparte/core';
-import type { BuiltInRunner, Device, Dtype, GenerationOptions, RunnerModule, RunnerProgress, TransformersModule } from './runners/types.js';
+import type { BuiltInRunner, Device, Dtype, GenerationOptions, RunnerGenerateInput, RunnerModule, RunnerProgress, RunnerTool, TransformersModule } from './runners/types.js';
 interface RunnerSelection {
     modelId: string;
     task?: BuiltInRunner;
@@ -45,6 +45,8 @@ export type InMessage = {
     id: string;
     messages: AparteChatMessage[];
     options: GenerationOptions;
+    tools?: RunnerTool[];
+    toolChoice?: RunnerGenerateInput['toolChoice'];
 } & RunnerSelection) | {
     type: 'cancel';
     id: string;

package/dist/worker-host.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"worker-host.d.ts","sourceRoot":"","sources":["../src/worker-host.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AACtF,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,EAAE,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,cAAc,EAAE,kBAAkB,EAAsB,MAAM,oBAAoB,CAAC;AAEhK,UAAU,eAAe;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,CAAC,EAAE,aAAa,CAAC;IACrB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,MAAM,SAAS,GACf;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,eAAe,CAAC,EAAE,MAAM,CAAA;CAAE,GAC1C,CAAC;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,eAAe,CAAC,GACnD,CAAC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,iBAAiB,EAAE,CAAC;IAAC,OAAO,EAAE,iBAAiB,CAAA;CAAE,GAAG,eAAe,CAAC,GAC/G;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAC9B,CAAC;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,OAAO,CAAA;CAAE,GAAG,eAAe,CAAC,CAAC;AAE1F,MAAM,MAAM,UAAU,GAChB,CAAC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,cAAc,CAAC,GACnD;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAC3C;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GACtD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,iBAAiB,CAAA;CAAE,GAC3D;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,WAAW,CAAA;CAAE,GACrD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAClD;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GACpC;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC;AAE/E,MAAM,WAAW,cAAc;IAC3B,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,IAAI,CAAC;IAChC,0FAA0F;IAC1F,gBAAgB,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAAC;IAClE,sGAAsG;IACtG,YAAY,CAAC,IAAI,EAAE;QAAE,IAAI,EAAE,aAAa,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,YAAY,CAAC,CAAC;IACpF;;;OAGG;IACH,SAAS,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,CAAC;CAClC;AAWD,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,cAAc,GAAG;IAAE,SAAS,CAAC,GAAG,EAAE,SAAS,GAAG,IAAI,CAAA;CAAE,CAiK1F"}
+{"version":3,"file":"worker-host.d.ts","sourceRoot":"","sources":["../src/worker-host.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,OAAO,KAAK,EAAE,iBAAiB,EAAE,iBAAiB,EAAE,WAAW,EAAE,MAAM,cAAc,CAAC;AACtF,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,EAAE,KAAK,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,YAAY,EAAE,cAAc,EAAE,UAAU,EAAE,kBAAkB,EAAsB,MAAM,oBAAoB,CAAC;AAEjM,UAAU,eAAe;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,CAAC,EAAE,aAAa,CAAC;IACrB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,MAAM,CAAC,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,MAAM,SAAS,GACf;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,eAAe,CAAC,EAAE,MAAM,CAAA;CAAE,GAC1C,CAAC;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,eAAe,CAAC,GACnD,CAAC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,iBAAiB,EAAE,CAAC;IAAC,OAAO,EAAE,iBAAiB,CAAC;IAAC,KAAK,CAAC,EAAE,UAAU,EAAE,CAAC;IAAC,UAAU,CAAC,EAAE,mBAAmB,CAAC,YAAY,CAAC,CAAA;CAAE,GAAG,eAAe,CAAC,GACrL;IAAE,IAAI,EAAE,QAAQ,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAC9B,CAAC;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,OAAO,CAAA;CAAE,GAAG,eAAe,CAAC,CAAC;AAE1F,MAAM,MAAM,UAAU,GAChB,CAAC;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GAAG,cAAc,CAAC,GACnD;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAC3C;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GACtD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,iBAAiB,CAAA;CAAE,GAC3D;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,WAAW,CAAA;CAAE,GACrD;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAClD;IAAE,IAAI,EAAE,SAAS,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GACpC;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,OAAO,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,CAAC;AAE/E,MAAM,WAAW,cAAc;IAC3B,IAAI,CAAC,OAAO,EAAE,UAAU,GAAG,IAAI,CAAC;IAChC,0FAA0F;IAC1F,gBAAgB,CAAC,SAAS,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,kBAAkB,CAAC,CAAC;IAClE,sGAAsG;IACtG,YAAY,CAAC,IAAI,EAAE;QAAE,IAAI,EAAE,aAAa,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,YAAY,CAAC,CAAC;IACpF;;;OAGG;IACH,SAAS,CAAC,IAAI,OAAO,CAAC,OAAO,CAAC,CAAC;CAClC;AAWD,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,cAAc,GAAG;IAAE,SAAS,CAAC,GAAG,EAAE,SAAS,GAAG,IAAI,CAAA;CAAE,CAmK1F"}

package/dist/worker.js

@@ -80,6 +80,8 @@ function createWorkerHost(deps) {
       await runner.generate({
         messages: msg.messages,
         options: msg.options,
+        ...msg.tools ? { tools: msg.tools } : {},
+        ...msg.toolChoice ? { toolChoice: msg.toolChoice } : {},
         signal: controller.signal,
         emit: (event) => {
           if (closed) return;

package/dist/worker.js.map

@@ -1 +1 @@
-{"version":3,"file":"worker.js","sources":["../src/worker-host.ts","../src/worker.ts"],"sourcesContent":["/**\n * The worker's logic, as a function of its three seams.\n *\n * `worker.ts` is the two-line shell that binds this to `self`; everything it decides is\n * here, with `post`, `loadTransformers` and `importRunner` injected — so the protocol is\n * tested with fakes where no Worker and no model can load (`worker-host.test.ts`).\n *\n * The protocol (main ⇄ worker):\n *\n *   → init            { transformersUrl? }                      once, first\n *   → prepare         { id, modelId, task?, runner?, dtype?, device? }\n *   → generate        { id, modelId, messages, options, task?, runner?, dtype?, device? }\n *   → cancel          { id }\n *   → command         { id, modelId, name, payload, task?, runner?, dtype?, device? }\n *   ← progress        { id, status, file?, progress?, detail? }\n *   ← pipeline-ready  { modelId }                                a runner is up\n *   ← prepare-error   { id, message }\n *   ← gen-event       { id, event }                              text | thinking | tool_use\n *   ← gen-done        { id, usage? }                             the host closes; `done` is never forwarded raw\n *   ← gen-error       { id, message }                            emitted or thrown\n *   ← warning         { message }                                once per distinct text\n *   ← command-result  { id, result } | { id, error }\n *\n * One runner is resident at a time, keyed by model AND runner: a switch disposes the\n * previous one first, which is the \"one pipeline per tab\" rule this package has always\n * kept — a local model is gigabytes, and two of them resident is the failure to avoid.\n */\n\nimport type { AparteChatMessage, AparteStreamEvent, AparteUsage } from '@aparte/core';\nimport type { BuiltInRunner, Device, Dtype, GenerationOptions, RunnerModule, RunnerProgress, TransformersModule, TransformersRunner } from './runners/types.js';\n\ninterface RunnerSelection {\n    modelId: string;\n    task?: BuiltInRunner;\n    runner?: string;\n    dtype?: Dtype;\n    device?: Device;\n}\n\nexport type InMessage =\n    | { type: 'init'; transformersUrl?: string }\n    | ({ type: 'prepare'; id: string } & RunnerSelection)\n    | ({ type: 'generate'; id: string; messages: AparteChatMessage[]; options: GenerationOptions } & RunnerSelection)\n    | { type: 'cancel'; id: string }\n    | ({ type: 'command'; id: string; name: string; payload: unknown } & RunnerSelection);\n\nexport type OutMessage =\n    | ({ type: 'progress'; id: string } & RunnerProgress)\n    | { type: 'pipeline-ready'; modelId: string }\n    | { type: 'prepare-error'; id: string; message: string }\n    | { type: 'gen-event'; id: string; event: AparteStreamEvent }\n    | { type: 'gen-done'; id: string; usage?: AparteUsage }\n    | { type: 'gen-error'; id: string; message: string }\n    | { type: 'warning'; message: string }\n    | { type: 'command-result'; id: string; result?: unknown; error?: string };\n\nexport interface WorkerHostDeps {\n    post(message: OutMessage): void;\n    /** Resolve Transformers.js — the bundled specifier first, else the URL `init` carried. */\n    loadTransformers(moduleUrl?: string): Promise<TransformersModule>;\n    /** Import the runner module a selection names: a built-in by `task`, a custom one by `runner` URL. */\n    importRunner(spec: { task: BuiltInRunner; runner?: string }): Promise<RunnerModule>;\n    /**\n     * Whether this worker can run on WebGPU: an adapter answers. Asked once, the first time a\n     * model loads on `'auto'`. Defaults to `navigator.gpu.requestAdapter()`.\n     */\n    hasWebGpu?(): Promise<boolean>;\n}\n\n/** WebGPU is usable when the API exists AND an adapter answers — the API alone can be there with no GPU behind it. */\nasync function defaultHasWebGpu(): Promise<boolean> {\n    const gpu = (globalThis.navigator as { gpu?: { requestAdapter(): Promise<unknown> } } | undefined)?.gpu;\n    if (!gpu) return false;\n    try { return (await gpu.requestAdapter()) != null; } catch { return false; }\n}\n\nconst errorText = (err: unknown, fallback: string): string => (err instanceof Error && err.message) || fallback;\n\nexport function createWorkerHost(deps: WorkerHostDeps): { onMessage(msg: InMessage): void } {\n    let moduleUrl: string | undefined;\n    let current: { key: string; runner: TransformersRunner } | null = null;\n    const generates = new Map<string, AbortController>();\n    const warned = new Set<string>();\n    let webGpu: Promise<boolean> | undefined;\n\n    // `'auto'` is resolved here rather than passed on: Transformers.js would try WebNN before\n    // WebGPU, and dropping it (as the runners used to) left its browser default, WASM.\n    const resolveDevice = async (device: Device | undefined): Promise<Device | undefined> => {\n        if (device !== 'auto') return device;\n        webGpu ??= (deps.hasWebGpu ?? defaultHasWebGpu)();\n        return (await webGpu) ? 'webgpu' : 'wasm';\n    };\n\n    const warn = (message: string): void => {\n        if (warned.has(message)) return;\n        warned.add(message);\n        deps.post({ type: 'warning', message });\n    };\n\n    async function ensureRunner(sel: RunnerSelection, progressId?: string): Promise<TransformersRunner> {\n        const task = sel.task ?? 'text-generation';\n        const key = `${sel.modelId}::${sel.runner ?? task}`;\n        if (current?.key === key) return current.runner;\n        if (current) {\n            const previous = current;\n            current = null;\n            await previous.runner.dispose?.();\n        }\n        const transformers = await deps.loadTransformers(moduleUrl);\n        const { createRunner } = await deps.importRunner({ task, runner: sel.runner });\n        const device = await resolveDevice(sel.device);\n        const create = (on: Device | undefined): Promise<TransformersRunner> => createRunner({\n            transformers,\n            modelId: sel.modelId,\n            dtype: sel.dtype,\n            device: on,\n            progress: (p) => { if (progressId) deps.post({ type: 'progress', id: progressId, ...p }); },\n            warn,\n        });\n        let runner: TransformersRunner;\n        try {\n            runner = await create(device);\n        } catch (err) {\n            // An adapter that answers is not a GPU that runs every model: a blocklisted driver,\n            // an operator the WebGPU backend lacks. When the page asked for `'auto'`, WASM is\n            // still an answer, as it always was before `'auto'` meant WebGPU. A forced\n            // `'webgpu'` fails as asked.\n            if (sel.device !== 'auto' || device !== 'webgpu') throw err;\n            warn(`[provider-transformers] ${sel.modelId} could not load on WebGPU (${errorText(err, 'unknown error')}); running it on WASM.`);\n            runner = await create('wasm');\n        }\n        current = { key, runner };\n        deps.post({ type: 'pipeline-ready', modelId: sel.modelId });\n        return runner;\n    }\n\n    async function handlePrepare(msg: Extract<InMessage, { type: 'prepare' }>): Promise<void> {\n        try {\n            await ensureRunner(msg, msg.id);\n            deps.post({ type: 'progress', id: msg.id, status: 'ready' });\n        } catch (err) {\n            deps.post({ type: 'prepare-error', id: msg.id, message: errorText(err, 'Failed to load model') });\n        }\n    }\n\n    async function handleGenerate(msg: Extract<InMessage, { type: 'generate' }>): Promise<void> {\n        // `onMessage` registered the controller when the message arrived; this is the\n        // same one, so a `cancel` that landed while the generate waited its turn is\n        // already recorded on it.\n        const controller = generates.get(msg.id) ?? new AbortController();\n        generates.set(msg.id, controller);\n        let usage: AparteUsage | undefined;\n        let closed = false;\n        try {\n            if (controller.signal.aborted) {\n                // Cancelled before it started: do not load a model in order to\n                // interrupt it a moment later. `gen-done` still goes out — it is what\n                // releases the main thread's queue slot.\n                deps.post({ type: 'gen-done', id: msg.id });\n                return;\n            }\n            const runner = await ensureRunner(msg, msg.id);\n            await runner.generate({\n                messages: msg.messages,\n                options: msg.options,\n                signal: controller.signal,\n                emit: (event) => {\n                    if (closed) return;\n                    // `done` and `error` close the stream, and closing is the host's: it\n                    // releases the queue slot on the main thread through gen-done/gen-error.\n                    if (event.type === 'done') { usage = event.usage; return; }\n                    if (event.type === 'error') { closed = true; deps.post({ type: 'gen-error', id: msg.id, message: event.message }); return; }\n                    deps.post({ type: 'gen-event', id: msg.id, event });\n                },\n            });\n            if (!closed) deps.post({ type: 'gen-done', id: msg.id, ...(usage ? { usage } : {}) });\n        } catch (err) {\n            if (!closed) deps.post({ type: 'gen-error', id: msg.id, message: errorText(err, 'Generation failed') });\n        } finally {\n            generates.delete(msg.id);\n        }\n    }\n\n    async function handleCommand(msg: Extract<InMessage, { type: 'command' }>): Promise<void> {\n        try {\n            const runner = await ensureRunner(msg);\n            if (!runner.command) {\n                deps.post({ type: 'command-result', id: msg.id, error: `This runner has no command handler (asked for \"${msg.name}\")` });\n                return;\n            }\n            const result = await runner.command(msg.name, msg.payload);\n            deps.post({ type: 'command-result', id: msg.id, result });\n        } catch (err) {\n            deps.post({ type: 'command-result', id: msg.id, error: errorText(err, 'Command failed') });\n        }\n    }\n\n    /*\n     * ONE chain for every message that touches the resident runner.\n     *\n     * `prepare`, `generate` and `command` all go through `ensureRunner`, which on a\n     * key change disposes the previous runner — and they used to run concurrently,\n     * so a `prepare` for another model disposed the pipeline an in-flight\n     * `generate` was executing (with the real runner, `pipe.dispose()` on the ONNX\n     * session mid-token), and two `prepare`s in one tick each saw `current === null`\n     * across three awaits and left two multi-GB models resident, the first\n     * unreachable for cleanup. The main thread already chains its generates for the\n     * same reason; `prepare` was on no chain at all.\n     *\n     * `cancel` is deliberately NOT chained: its whole job is to reach the generate\n     * that is running now.\n     *\n     * Every handler settles its own errors (each posts a `*-error` message), so the\n     * chain cannot be poisoned; the trailing catch is belt and braces.\n     */\n    let chain: Promise<void> = Promise.resolve();\n    const serialize = (work: () => Promise<void>): void => {\n        chain = chain.then(work).catch(() => { /* handlers report their own failures */ });\n    };\n\n    return {\n        onMessage(msg: InMessage): void {\n            switch (msg.type) {\n                case 'init': moduleUrl = msg.transformersUrl; break;\n                case 'prepare': serialize(() => handlePrepare(msg)); break;\n                /*\n                 * The abort intent is registered HERE, when the message arrives, not\n                 * where the generate runs: a generate is chained, so it can wait for\n                 * as long as a model load takes, and a `cancel` reaching the worker in\n                 * that window found no controller to abort. The generate then started\n                 * as if nothing had happened and streamed tokens into a reply the user\n                 * had already stopped.\n                 */\n                case 'generate': generates.set(msg.id, new AbortController()); serialize(() => handleGenerate(msg)); break;\n                case 'cancel': generates.get(msg.id)?.abort(); break;\n                case 'command': serialize(() => handleCommand(msg)); break;\n            }\n        },\n    };\n}\n","/**\n * The Transformers.js inference worker — the shell.\n *\n * Runs entirely off the main thread. What it decides lives in `worker-host.ts` (the\n * protocol, the one-runner-at-a-time rule, cancel, warnings); what a model IS lives in a\n * runner (`runners/`). This file binds the two to `self` and owns the two things only a\n * real worker can do: resolve Transformers.js, and import a runner module.\n */\n\nimport type { BuiltInRunner, RunnerModule, TransformersModule } from './runners/types.js';\nimport { createWorkerHost } from './worker-host.js';\n\n/**\n * Where Transformers.js comes from, resolved once, from whichever path has it.\n *\n * A static `import … from '@huggingface/transformers'` is unresolvable in a worker\n * served without a bundler: an import map lives on the DOCUMENT and, by spec, does not\n * reach a worker — so the page can map the specifier for itself and the worker still\n * cannot. The two paths, in order:\n *\n *   1. `import('@huggingface/transformers')` — a bare specifier, statically visible, so a\n *      consumer's bundler resolves and bundles the peer exactly as it did before.\n *   2. the absolute URL the main thread read from the page's own import map and sent in\n *      the first message — the CDN path, where that map is the consumer's manifest.\n *\n * The order matters: a bundled app must never reach for the network copy. Whichever\n * path won is the module every runner receives as `ctx.transformers` — one copy per\n * worker, the version the CONSUMER installed or pinned.\n */\nlet _tf: Promise<TransformersModule> | null = null;\n\nfunction loadTransformers(moduleUrl?: string): Promise<TransformersModule> {\n    _tf ??= (async () => {\n        let mod: TransformersModule;\n        try {\n            mod = await import('@huggingface/transformers');\n        } catch (bundlerPathFailed) {\n            if (!moduleUrl) throw bundlerPathFailed;\n            mod = await import(/* @vite-ignore */ moduleUrl) as TransformersModule;\n        }\n        // Fetch weights from the Hugging Face hub (not local paths) and cache them in the\n        // browser Cache API — this is what `listCachedModels()` scans on the main thread.\n        mod.env.allowLocalModels = false;\n        mod.env.useBrowserCache = true;\n        return mod;\n    })();\n    return _tf;\n}\n\n/**\n * The runners this package ships, each behind a dynamic import so the bundler splits\n * it into its own chunk and a page loads only the one its model asks for. The imports\n * are RELATIVE on purpose: a module script resolves them against its own URL, so they\n * follow the worker wherever it is served from — the same origin, a CDN, or through the\n * blob shim `_spawnWorker` builds for a cross-origin copy.\n */\nconst BUILT_IN: Record<BuiltInRunner, () => Promise<RunnerModule>> = {\n    'text-generation': () => import('./runners/text-generation.js'),\n    'image-text-to-text': () => import('./runners/image-text-to-text.js'),\n};\n\nfunction importRunner({ task, runner }: { task: BuiltInRunner; runner?: string }): Promise<RunnerModule> {\n    // A custom runner is an absolute URL by the time it gets here (the main thread\n    // resolved it against the page), and it wins over `task`.\n    if (runner) return import(/* @vite-ignore */ runner) as Promise<RunnerModule>;\n    return BUILT_IN[task]();\n}\n\n// DOM's `Worker` interface types `postMessage` + typed `addEventListener('message')`,\n// which is enough for the worker scope — avoids pulling the WebWorker lib (it clashes\n// with DOM's global `postMessage`).\nconst ctx = self as unknown as Worker;\n\nconst host = createWorkerHost({\n    post: (message) => { ctx.postMessage(message); },\n    loadTransformers,\n    importRunner,\n});\n\nctx.addEventListener('message', (event: MessageEvent) => { host.onMessage(event.data); });\n"],"names":[],"mappings":"AAsEA,eAAe,mBAAqC;AAChD,QAAM,MAAO,WAAW,WAA4E;AACpG,MAAI,CAAC,IAAK,QAAO;AACjB,MAAI;AAAE,WAAQ,MAAM,IAAI,eAAA,KAAqB;AAAA,EAAM,QAAQ;AAAE,WAAO;AAAA,EAAO;AAC/E;AAEA,MAAM,YAAY,CAAC,KAAc,aAA8B,eAAe,SAAS,IAAI,WAAY;AAEhG,SAAS,iBAAiB,MAA2D;AACxF,MAAI;AACJ,MAAI,UAA8D;AAClE,QAAM,gCAAgB,IAAA;AACtB,QAAM,6BAAa,IAAA;AACnB,MAAI;AAIJ,QAAM,gBAAgB,OAAO,WAA4D;AACrF,QAAI,WAAW,OAAQ,QAAO;AAC9B,gBAAY,KAAK,aAAa,kBAAA;AAC9B,WAAQ,MAAM,SAAU,WAAW;AAAA,EACvC;AAEA,QAAM,OAAO,CAAC,YAA0B;AACpC,QAAI,OAAO,IAAI,OAAO,EAAG;AACzB,WAAO,IAAI,OAAO;AAClB,SAAK,KAAK,EAAE,MAAM,WAAW,SAAS;AAAA,EAC1C;AAEA,iBAAe,aAAa,KAAsB,YAAkD;AAChG,UAAM,OAAO,IAAI,QAAQ;AACzB,UAAM,MAAM,GAAG,IAAI,OAAO,KAAK,IAAI,UAAU,IAAI;AACjD,QAAI,SAAS,QAAQ,IAAK,QAAO,QAAQ;AACzC,QAAI,SAAS;AACT,YAAM,WAAW;AACjB,gBAAU;AACV,YAAM,SAAS,OAAO,UAAA;AAAA,IAC1B;AACA,UAAM,eAAe,MAAM,KAAK,iBAAiB,SAAS;AAC1D,UAAM,EAAE,iBAAiB,MAAM,KAAK,aAAa,EAAE,MAAM,QAAQ,IAAI,QAAQ;AAC7E,UAAM,SAAS,MAAM,cAAc,IAAI,MAAM;AAC7C,UAAM,SAAS,CAAC,OAAwD,aAAa;AAAA,MACjF;AAAA,MACA,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,QAAQ;AAAA,MACR,UAAU,CAAC,MAAM;AAAE,YAAI,WAAY,MAAK,KAAK,EAAE,MAAM,YAAY,IAAI,YAAY,GAAG,GAAG;AAAA,MAAG;AAAA,MAC1F;AAAA,IAAA,CACH;AACD,QAAI;AACJ,QAAI;AACA,eAAS,MAAM,OAAO,MAAM;AAAA,IAChC,SAAS,KAAK;AAKV,UAAI,IAAI,WAAW,UAAU,WAAW,SAAU,OAAM;AACxD,WAAK,2BAA2B,IAAI,OAAO,8BAA8B,UAAU,KAAK,eAAe,CAAC,wBAAwB;AAChI,eAAS,MAAM,OAAO,MAAM;AAAA,IAChC;AACA,cAAU,EAAE,KAAK,OAAA;AACjB,SAAK,KAAK,EAAE,MAAM,kBAAkB,SAAS,IAAI,SAAS;AAC1D,WAAO;AAAA,EACX;AAEA,iBAAe,cAAc,KAA6D;AACtF,QAAI;AACA,YAAM,aAAa,KAAK,IAAI,EAAE;AAC9B,WAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI,QAAQ,SAAS;AAAA,IAC/D,SAAS,KAAK;AACV,WAAK,KAAK,EAAE,MAAM,iBAAiB,IAAI,IAAI,IAAI,SAAS,UAAU,KAAK,sBAAsB,EAAA,CAAG;AAAA,IACpG;AAAA,EACJ;AAEA,iBAAe,eAAe,KAA8D;AAIxF,UAAM,aAAa,UAAU,IAAI,IAAI,EAAE,KAAK,IAAI,gBAAA;AAChD,cAAU,IAAI,IAAI,IAAI,UAAU;AAChC,QAAI;AACJ,QAAI,SAAS;AACb,QAAI;AACA,UAAI,WAAW,OAAO,SAAS;AAI3B,aAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI;AAC1C;AAAA,MACJ;AACA,YAAM,SAAS,MAAM,aAAa,KAAK,IAAI,EAAE;AAC7C,YAAM,OAAO,SAAS;AAAA,QAClB,UAAU,IAAI;AAAA,QACd,SAAS,IAAI;AAAA,QACb,QAAQ,WAAW;AAAA,QACnB,MAAM,CAAC,UAAU;AACb,cAAI,OAAQ;AAGZ,cAAI,MAAM,SAAS,QAAQ;AAAE,oBAAQ,MAAM;AAAO;AAAA,UAAQ;AAC1D,cAAI,MAAM,SAAS,SAAS;AAAE,qBAAS;AAAM,iBAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,SAAS,MAAM,QAAA,CAAS;AAAG;AAAA,UAAQ;AAC3H,eAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,OAAO;AAAA,QACtD;AAAA,MAAA,CACH;AACD,UAAI,CAAC,OAAQ,MAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI,GAAI,QAAQ,EAAE,UAAU,CAAA,GAAK;AAAA,IACxF,SAAS,KAAK;AACV,UAAI,CAAC,OAAQ,MAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,SAAS,UAAU,KAAK,mBAAmB,GAAG;AAAA,IAC1G,UAAA;AACI,gBAAU,OAAO,IAAI,EAAE;AAAA,IAC3B;AAAA,EACJ;AAEA,iBAAe,cAAc,KAA6D;AACtF,QAAI;AACA,YAAM,SAAS,MAAM,aAAa,GAAG;AACrC,UAAI,CAAC,OAAO,SAAS;AACjB,aAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,OAAO,kDAAkD,IAAI,IAAI,KAAA,CAAM;AACvH;AAAA,MACJ;AACA,YAAM,SAAS,MAAM,OAAO,QAAQ,IAAI,MAAM,IAAI,OAAO;AACzD,WAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,QAAQ;AAAA,IAC5D,SAAS,KAAK;AACV,WAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,OAAO,UAAU,KAAK,gBAAgB,EAAA,CAAG;AAAA,IAC7F;AAAA,EACJ;AAoBA,MAAI,QAAuB,QAAQ,QAAA;AACnC,QAAM,YAAY,CAAC,SAAoC;AACnD,YAAQ,MAAM,KAAK,IAAI,EAAE,MAAM,MAAM;AAAA,IAA2C,CAAC;AAAA,EACrF;AAEA,SAAO;AAAA,IACH,UAAU,KAAsB;AAC5B,cAAQ,IAAI,MAAA;AAAA,QACR,KAAK;AAAQ,sBAAY,IAAI;AAAiB;AAAA,QAC9C,KAAK;AAAW,oBAAU,MAAM,cAAc,GAAG,CAAC;AAAG;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,QASrD,KAAK;AAAY,oBAAU,IAAI,IAAI,IAAI,IAAI,iBAAiB;AAAG,oBAAU,MAAM,eAAe,GAAG,CAAC;AAAG;AAAA,QACrG,KAAK;AAAU,oBAAU,IAAI,IAAI,EAAE,GAAG,MAAA;AAAS;AAAA,QAC/C,KAAK;AAAW,oBAAU,MAAM,cAAc,GAAG,CAAC;AAAG;AAAA,MAAA;AAAA,IAE7D;AAAA,EAAA;AAER;AClNA,IAAI,MAA0C;AAE9C,SAAS,iBAAiB,WAAiD;AACvE,WAAS,YAAY;AACjB,QAAI;AACJ,QAAI;AACA,YAAM,MAAM,OAAO,2BAA2B;AAAA,IAClD,SAAS,mBAAmB;AACxB,UAAI,CAAC,UAAW,OAAM;AACtB,YAAM,MAAM;AAAA;AAAA,QAA0B;AAAA;AAAA,IAC1C;AAGA,QAAI,IAAI,mBAAmB;AAC3B,QAAI,IAAI,kBAAkB;AAC1B,WAAO;AAAA,EACX,GAAA;AACA,SAAO;AACX;AASA,MAAM,WAA+D;AAAA,EACjE,mBAAmB,MAAM,OAAO,8BAA8B;AAAA,EAC9D,sBAAsB,MAAM,OAAO,iCAAiC;AACxE;AAEA,SAAS,aAAa,EAAE,MAAM,UAA2E;AAGrG,MAAI,OAAQ,QAAO;AAAA;AAAA,IAA0B;AAAA;AAC7C,SAAO,SAAS,IAAI,EAAA;AACxB;AAKA,MAAM,MAAM;AAEZ,MAAM,OAAO,iBAAiB;AAAA,EAC1B,MAAM,CAAC,YAAY;AAAE,QAAI,YAAY,OAAO;AAAA,EAAG;AAAA,EAC/C;AAAA,EACA;AACJ,CAAC;AAED,IAAI,iBAAiB,WAAW,CAAC,UAAwB;AAAE,OAAK,UAAU,MAAM,IAAI;AAAG,CAAC;"}
+{"version":3,"file":"worker.js","sources":["../src/worker-host.ts","../src/worker.ts"],"sourcesContent":["/**\n * The worker's logic, as a function of its three seams.\n *\n * `worker.ts` is the two-line shell that binds this to `self`; everything it decides is\n * here, with `post`, `loadTransformers` and `importRunner` injected — so the protocol is\n * tested with fakes where no Worker and no model can load (`worker-host.test.ts`).\n *\n * The protocol (main ⇄ worker):\n *\n *   → init            { transformersUrl? }                      once, first\n *   → prepare         { id, modelId, task?, runner?, dtype?, device? }\n *   → generate        { id, modelId, messages, options, tools?, toolChoice?, task?, runner?, dtype?, device? }\n *   → cancel          { id }\n *   → command         { id, modelId, name, payload, task?, runner?, dtype?, device? }\n *   ← progress        { id, status, file?, progress?, detail? }\n *   ← pipeline-ready  { modelId }                                a runner is up\n *   ← prepare-error   { id, message }\n *   ← gen-event       { id, event }                              text | thinking | tool_use\n *   ← gen-done        { id, usage? }                             the host closes; `done` is never forwarded raw\n *   ← gen-error       { id, message }                            emitted or thrown\n *   ← warning         { message }                                once per distinct text\n *   ← command-result  { id, result } | { id, error }\n *\n * One runner is resident at a time, keyed by model AND runner: a switch disposes the\n * previous one first, which is the \"one pipeline per tab\" rule this package has always\n * kept — a local model is gigabytes, and two of them resident is the failure to avoid.\n */\n\nimport type { AparteChatMessage, AparteStreamEvent, AparteUsage } from '@aparte/core';\nimport type { BuiltInRunner, Device, Dtype, GenerationOptions, RunnerGenerateInput, RunnerModule, RunnerProgress, RunnerTool, TransformersModule, TransformersRunner } from './runners/types.js';\n\ninterface RunnerSelection {\n    modelId: string;\n    task?: BuiltInRunner;\n    runner?: string;\n    dtype?: Dtype;\n    device?: Device;\n}\n\nexport type InMessage =\n    | { type: 'init'; transformersUrl?: string }\n    | ({ type: 'prepare'; id: string } & RunnerSelection)\n    | ({ type: 'generate'; id: string; messages: AparteChatMessage[]; options: GenerationOptions; tools?: RunnerTool[]; toolChoice?: RunnerGenerateInput['toolChoice'] } & RunnerSelection)\n    | { type: 'cancel'; id: string }\n    | ({ type: 'command'; id: string; name: string; payload: unknown } & RunnerSelection);\n\nexport type OutMessage =\n    | ({ type: 'progress'; id: string } & RunnerProgress)\n    | { type: 'pipeline-ready'; modelId: string }\n    | { type: 'prepare-error'; id: string; message: string }\n    | { type: 'gen-event'; id: string; event: AparteStreamEvent }\n    | { type: 'gen-done'; id: string; usage?: AparteUsage }\n    | { type: 'gen-error'; id: string; message: string }\n    | { type: 'warning'; message: string }\n    | { type: 'command-result'; id: string; result?: unknown; error?: string };\n\nexport interface WorkerHostDeps {\n    post(message: OutMessage): void;\n    /** Resolve Transformers.js — the bundled specifier first, else the URL `init` carried. */\n    loadTransformers(moduleUrl?: string): Promise<TransformersModule>;\n    /** Import the runner module a selection names: a built-in by `task`, a custom one by `runner` URL. */\n    importRunner(spec: { task: BuiltInRunner; runner?: string }): Promise<RunnerModule>;\n    /**\n     * Whether this worker can run on WebGPU: an adapter answers. Asked once, the first time a\n     * model loads on `'auto'`. Defaults to `navigator.gpu.requestAdapter()`.\n     */\n    hasWebGpu?(): Promise<boolean>;\n}\n\n/** WebGPU is usable when the API exists AND an adapter answers — the API alone can be there with no GPU behind it. */\nasync function defaultHasWebGpu(): Promise<boolean> {\n    const gpu = (globalThis.navigator as { gpu?: { requestAdapter(): Promise<unknown> } } | undefined)?.gpu;\n    if (!gpu) return false;\n    try { return (await gpu.requestAdapter()) != null; } catch { return false; }\n}\n\nconst errorText = (err: unknown, fallback: string): string => (err instanceof Error && err.message) || fallback;\n\nexport function createWorkerHost(deps: WorkerHostDeps): { onMessage(msg: InMessage): void } {\n    let moduleUrl: string | undefined;\n    let current: { key: string; runner: TransformersRunner } | null = null;\n    const generates = new Map<string, AbortController>();\n    const warned = new Set<string>();\n    let webGpu: Promise<boolean> | undefined;\n\n    // `'auto'` is resolved here rather than passed on: Transformers.js would try WebNN before\n    // WebGPU, and dropping it (as the runners used to) left its browser default, WASM.\n    const resolveDevice = async (device: Device | undefined): Promise<Device | undefined> => {\n        if (device !== 'auto') return device;\n        webGpu ??= (deps.hasWebGpu ?? defaultHasWebGpu)();\n        return (await webGpu) ? 'webgpu' : 'wasm';\n    };\n\n    const warn = (message: string): void => {\n        if (warned.has(message)) return;\n        warned.add(message);\n        deps.post({ type: 'warning', message });\n    };\n\n    async function ensureRunner(sel: RunnerSelection, progressId?: string): Promise<TransformersRunner> {\n        const task = sel.task ?? 'text-generation';\n        const key = `${sel.modelId}::${sel.runner ?? task}`;\n        if (current?.key === key) return current.runner;\n        if (current) {\n            const previous = current;\n            current = null;\n            await previous.runner.dispose?.();\n        }\n        const transformers = await deps.loadTransformers(moduleUrl);\n        const { createRunner } = await deps.importRunner({ task, runner: sel.runner });\n        const device = await resolveDevice(sel.device);\n        const create = (on: Device | undefined): Promise<TransformersRunner> => createRunner({\n            transformers,\n            modelId: sel.modelId,\n            dtype: sel.dtype,\n            device: on,\n            progress: (p) => { if (progressId) deps.post({ type: 'progress', id: progressId, ...p }); },\n            warn,\n        });\n        let runner: TransformersRunner;\n        try {\n            runner = await create(device);\n        } catch (err) {\n            // An adapter that answers is not a GPU that runs every model: a blocklisted driver,\n            // an operator the WebGPU backend lacks. When the page asked for `'auto'`, WASM is\n            // still an answer, as it always was before `'auto'` meant WebGPU. A forced\n            // `'webgpu'` fails as asked.\n            if (sel.device !== 'auto' || device !== 'webgpu') throw err;\n            warn(`[provider-transformers] ${sel.modelId} could not load on WebGPU (${errorText(err, 'unknown error')}); running it on WASM.`);\n            runner = await create('wasm');\n        }\n        current = { key, runner };\n        deps.post({ type: 'pipeline-ready', modelId: sel.modelId });\n        return runner;\n    }\n\n    async function handlePrepare(msg: Extract<InMessage, { type: 'prepare' }>): Promise<void> {\n        try {\n            await ensureRunner(msg, msg.id);\n            deps.post({ type: 'progress', id: msg.id, status: 'ready' });\n        } catch (err) {\n            deps.post({ type: 'prepare-error', id: msg.id, message: errorText(err, 'Failed to load model') });\n        }\n    }\n\n    async function handleGenerate(msg: Extract<InMessage, { type: 'generate' }>): Promise<void> {\n        // `onMessage` registered the controller when the message arrived; this is the\n        // same one, so a `cancel` that landed while the generate waited its turn is\n        // already recorded on it.\n        const controller = generates.get(msg.id) ?? new AbortController();\n        generates.set(msg.id, controller);\n        let usage: AparteUsage | undefined;\n        let closed = false;\n        try {\n            if (controller.signal.aborted) {\n                // Cancelled before it started: do not load a model in order to\n                // interrupt it a moment later. `gen-done` still goes out — it is what\n                // releases the main thread's queue slot.\n                deps.post({ type: 'gen-done', id: msg.id });\n                return;\n            }\n            const runner = await ensureRunner(msg, msg.id);\n            await runner.generate({\n                messages: msg.messages,\n                options: msg.options,\n                ...(msg.tools ? { tools: msg.tools } : {}),\n                ...(msg.toolChoice ? { toolChoice: msg.toolChoice } : {}),\n                signal: controller.signal,\n                emit: (event) => {\n                    if (closed) return;\n                    // `done` and `error` close the stream, and closing is the host's: it\n                    // releases the queue slot on the main thread through gen-done/gen-error.\n                    if (event.type === 'done') { usage = event.usage; return; }\n                    if (event.type === 'error') { closed = true; deps.post({ type: 'gen-error', id: msg.id, message: event.message }); return; }\n                    deps.post({ type: 'gen-event', id: msg.id, event });\n                },\n            });\n            if (!closed) deps.post({ type: 'gen-done', id: msg.id, ...(usage ? { usage } : {}) });\n        } catch (err) {\n            if (!closed) deps.post({ type: 'gen-error', id: msg.id, message: errorText(err, 'Generation failed') });\n        } finally {\n            generates.delete(msg.id);\n        }\n    }\n\n    async function handleCommand(msg: Extract<InMessage, { type: 'command' }>): Promise<void> {\n        try {\n            const runner = await ensureRunner(msg);\n            if (!runner.command) {\n                deps.post({ type: 'command-result', id: msg.id, error: `This runner has no command handler (asked for \"${msg.name}\")` });\n                return;\n            }\n            const result = await runner.command(msg.name, msg.payload);\n            deps.post({ type: 'command-result', id: msg.id, result });\n        } catch (err) {\n            deps.post({ type: 'command-result', id: msg.id, error: errorText(err, 'Command failed') });\n        }\n    }\n\n    /*\n     * ONE chain for every message that touches the resident runner.\n     *\n     * `prepare`, `generate` and `command` all go through `ensureRunner`, which on a\n     * key change disposes the previous runner — and they used to run concurrently,\n     * so a `prepare` for another model disposed the pipeline an in-flight\n     * `generate` was executing (with the real runner, `pipe.dispose()` on the ONNX\n     * session mid-token), and two `prepare`s in one tick each saw `current === null`\n     * across three awaits and left two multi-GB models resident, the first\n     * unreachable for cleanup. The main thread already chains its generates for the\n     * same reason; `prepare` was on no chain at all.\n     *\n     * `cancel` is deliberately NOT chained: its whole job is to reach the generate\n     * that is running now.\n     *\n     * Every handler settles its own errors (each posts a `*-error` message), so the\n     * chain cannot be poisoned; the trailing catch is belt and braces.\n     */\n    let chain: Promise<void> = Promise.resolve();\n    const serialize = (work: () => Promise<void>): void => {\n        chain = chain.then(work).catch(() => { /* handlers report their own failures */ });\n    };\n\n    return {\n        onMessage(msg: InMessage): void {\n            switch (msg.type) {\n                case 'init': moduleUrl = msg.transformersUrl; break;\n                case 'prepare': serialize(() => handlePrepare(msg)); break;\n                /*\n                 * The abort intent is registered HERE, when the message arrives, not\n                 * where the generate runs: a generate is chained, so it can wait for\n                 * as long as a model load takes, and a `cancel` reaching the worker in\n                 * that window found no controller to abort. The generate then started\n                 * as if nothing had happened and streamed tokens into a reply the user\n                 * had already stopped.\n                 */\n                case 'generate': generates.set(msg.id, new AbortController()); serialize(() => handleGenerate(msg)); break;\n                case 'cancel': generates.get(msg.id)?.abort(); break;\n                case 'command': serialize(() => handleCommand(msg)); break;\n            }\n        },\n    };\n}\n","/**\n * The Transformers.js inference worker — the shell.\n *\n * Runs entirely off the main thread. What it decides lives in `worker-host.ts` (the\n * protocol, the one-runner-at-a-time rule, cancel, warnings); what a model IS lives in a\n * runner (`runners/`). This file binds the two to `self` and owns the two things only a\n * real worker can do: resolve Transformers.js, and import a runner module.\n */\n\nimport type { BuiltInRunner, RunnerModule, TransformersModule } from './runners/types.js';\nimport { createWorkerHost } from './worker-host.js';\n\n/**\n * Where Transformers.js comes from, resolved once, from whichever path has it.\n *\n * A static `import … from '@huggingface/transformers'` is unresolvable in a worker\n * served without a bundler: an import map lives on the DOCUMENT and, by spec, does not\n * reach a worker — so the page can map the specifier for itself and the worker still\n * cannot. The two paths, in order:\n *\n *   1. `import('@huggingface/transformers')` — a bare specifier, statically visible, so a\n *      consumer's bundler resolves and bundles the peer exactly as it did before.\n *   2. the absolute URL the main thread read from the page's own import map and sent in\n *      the first message — the CDN path, where that map is the consumer's manifest.\n *\n * The order matters: a bundled app must never reach for the network copy. Whichever\n * path won is the module every runner receives as `ctx.transformers` — one copy per\n * worker, the version the CONSUMER installed or pinned.\n */\nlet _tf: Promise<TransformersModule> | null = null;\n\nfunction loadTransformers(moduleUrl?: string): Promise<TransformersModule> {\n    _tf ??= (async () => {\n        let mod: TransformersModule;\n        try {\n            mod = await import('@huggingface/transformers');\n        } catch (bundlerPathFailed) {\n            if (!moduleUrl) throw bundlerPathFailed;\n            mod = await import(/* @vite-ignore */ moduleUrl) as TransformersModule;\n        }\n        // Fetch weights from the Hugging Face hub (not local paths) and cache them in the\n        // browser Cache API — this is what `listCachedModels()` scans on the main thread.\n        mod.env.allowLocalModels = false;\n        mod.env.useBrowserCache = true;\n        return mod;\n    })();\n    return _tf;\n}\n\n/**\n * The runners this package ships, each behind a dynamic import so the bundler splits\n * it into its own chunk and a page loads only the one its model asks for. The imports\n * are RELATIVE on purpose: a module script resolves them against its own URL, so they\n * follow the worker wherever it is served from — the same origin, a CDN, or through the\n * blob shim `_spawnWorker` builds for a cross-origin copy.\n */\nconst BUILT_IN: Record<BuiltInRunner, () => Promise<RunnerModule>> = {\n    'text-generation': () => import('./runners/text-generation.js'),\n    'image-text-to-text': () => import('./runners/image-text-to-text.js'),\n};\n\nfunction importRunner({ task, runner }: { task: BuiltInRunner; runner?: string }): Promise<RunnerModule> {\n    // A custom runner is an absolute URL by the time it gets here (the main thread\n    // resolved it against the page), and it wins over `task`.\n    if (runner) return import(/* @vite-ignore */ runner) as Promise<RunnerModule>;\n    return BUILT_IN[task]();\n}\n\n// DOM's `Worker` interface types `postMessage` + typed `addEventListener('message')`,\n// which is enough for the worker scope — avoids pulling the WebWorker lib (it clashes\n// with DOM's global `postMessage`).\nconst ctx = self as unknown as Worker;\n\nconst host = createWorkerHost({\n    post: (message) => { ctx.postMessage(message); },\n    loadTransformers,\n    importRunner,\n});\n\nctx.addEventListener('message', (event: MessageEvent) => { host.onMessage(event.data); });\n"],"names":[],"mappings":"AAsEA,eAAe,mBAAqC;AAChD,QAAM,MAAO,WAAW,WAA4E;AACpG,MAAI,CAAC,IAAK,QAAO;AACjB,MAAI;AAAE,WAAQ,MAAM,IAAI,eAAA,KAAqB;AAAA,EAAM,QAAQ;AAAE,WAAO;AAAA,EAAO;AAC/E;AAEA,MAAM,YAAY,CAAC,KAAc,aAA8B,eAAe,SAAS,IAAI,WAAY;AAEhG,SAAS,iBAAiB,MAA2D;AACxF,MAAI;AACJ,MAAI,UAA8D;AAClE,QAAM,gCAAgB,IAAA;AACtB,QAAM,6BAAa,IAAA;AACnB,MAAI;AAIJ,QAAM,gBAAgB,OAAO,WAA4D;AACrF,QAAI,WAAW,OAAQ,QAAO;AAC9B,gBAAY,KAAK,aAAa,kBAAA;AAC9B,WAAQ,MAAM,SAAU,WAAW;AAAA,EACvC;AAEA,QAAM,OAAO,CAAC,YAA0B;AACpC,QAAI,OAAO,IAAI,OAAO,EAAG;AACzB,WAAO,IAAI,OAAO;AAClB,SAAK,KAAK,EAAE,MAAM,WAAW,SAAS;AAAA,EAC1C;AAEA,iBAAe,aAAa,KAAsB,YAAkD;AAChG,UAAM,OAAO,IAAI,QAAQ;AACzB,UAAM,MAAM,GAAG,IAAI,OAAO,KAAK,IAAI,UAAU,IAAI;AACjD,QAAI,SAAS,QAAQ,IAAK,QAAO,QAAQ;AACzC,QAAI,SAAS;AACT,YAAM,WAAW;AACjB,gBAAU;AACV,YAAM,SAAS,OAAO,UAAA;AAAA,IAC1B;AACA,UAAM,eAAe,MAAM,KAAK,iBAAiB,SAAS;AAC1D,UAAM,EAAE,iBAAiB,MAAM,KAAK,aAAa,EAAE,MAAM,QAAQ,IAAI,QAAQ;AAC7E,UAAM,SAAS,MAAM,cAAc,IAAI,MAAM;AAC7C,UAAM,SAAS,CAAC,OAAwD,aAAa;AAAA,MACjF;AAAA,MACA,SAAS,IAAI;AAAA,MACb,OAAO,IAAI;AAAA,MACX,QAAQ;AAAA,MACR,UAAU,CAAC,MAAM;AAAE,YAAI,WAAY,MAAK,KAAK,EAAE,MAAM,YAAY,IAAI,YAAY,GAAG,GAAG;AAAA,MAAG;AAAA,MAC1F;AAAA,IAAA,CACH;AACD,QAAI;AACJ,QAAI;AACA,eAAS,MAAM,OAAO,MAAM;AAAA,IAChC,SAAS,KAAK;AAKV,UAAI,IAAI,WAAW,UAAU,WAAW,SAAU,OAAM;AACxD,WAAK,2BAA2B,IAAI,OAAO,8BAA8B,UAAU,KAAK,eAAe,CAAC,wBAAwB;AAChI,eAAS,MAAM,OAAO,MAAM;AAAA,IAChC;AACA,cAAU,EAAE,KAAK,OAAA;AACjB,SAAK,KAAK,EAAE,MAAM,kBAAkB,SAAS,IAAI,SAAS;AAC1D,WAAO;AAAA,EACX;AAEA,iBAAe,cAAc,KAA6D;AACtF,QAAI;AACA,YAAM,aAAa,KAAK,IAAI,EAAE;AAC9B,WAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI,QAAQ,SAAS;AAAA,IAC/D,SAAS,KAAK;AACV,WAAK,KAAK,EAAE,MAAM,iBAAiB,IAAI,IAAI,IAAI,SAAS,UAAU,KAAK,sBAAsB,EAAA,CAAG;AAAA,IACpG;AAAA,EACJ;AAEA,iBAAe,eAAe,KAA8D;AAIxF,UAAM,aAAa,UAAU,IAAI,IAAI,EAAE,KAAK,IAAI,gBAAA;AAChD,cAAU,IAAI,IAAI,IAAI,UAAU;AAChC,QAAI;AACJ,QAAI,SAAS;AACb,QAAI;AACA,UAAI,WAAW,OAAO,SAAS;AAI3B,aAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI;AAC1C;AAAA,MACJ;AACA,YAAM,SAAS,MAAM,aAAa,KAAK,IAAI,EAAE;AAC7C,YAAM,OAAO,SAAS;AAAA,QAClB,UAAU,IAAI;AAAA,QACd,SAAS,IAAI;AAAA,QACb,GAAI,IAAI,QAAQ,EAAE,OAAO,IAAI,MAAA,IAAU,CAAA;AAAA,QACvC,GAAI,IAAI,aAAa,EAAE,YAAY,IAAI,WAAA,IAAe,CAAA;AAAA,QACtD,QAAQ,WAAW;AAAA,QACnB,MAAM,CAAC,UAAU;AACb,cAAI,OAAQ;AAGZ,cAAI,MAAM,SAAS,QAAQ;AAAE,oBAAQ,MAAM;AAAO;AAAA,UAAQ;AAC1D,cAAI,MAAM,SAAS,SAAS;AAAE,qBAAS;AAAM,iBAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,SAAS,MAAM,QAAA,CAAS;AAAG;AAAA,UAAQ;AAC3H,eAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,OAAO;AAAA,QACtD;AAAA,MAAA,CACH;AACD,UAAI,CAAC,OAAQ,MAAK,KAAK,EAAE,MAAM,YAAY,IAAI,IAAI,IAAI,GAAI,QAAQ,EAAE,UAAU,CAAA,GAAK;AAAA,IACxF,SAAS,KAAK;AACV,UAAI,CAAC,OAAQ,MAAK,KAAK,EAAE,MAAM,aAAa,IAAI,IAAI,IAAI,SAAS,UAAU,KAAK,mBAAmB,GAAG;AAAA,IAC1G,UAAA;AACI,gBAAU,OAAO,IAAI,EAAE;AAAA,IAC3B;AAAA,EACJ;AAEA,iBAAe,cAAc,KAA6D;AACtF,QAAI;AACA,YAAM,SAAS,MAAM,aAAa,GAAG;AACrC,UAAI,CAAC,OAAO,SAAS;AACjB,aAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,OAAO,kDAAkD,IAAI,IAAI,KAAA,CAAM;AACvH;AAAA,MACJ;AACA,YAAM,SAAS,MAAM,OAAO,QAAQ,IAAI,MAAM,IAAI,OAAO;AACzD,WAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,QAAQ;AAAA,IAC5D,SAAS,KAAK;AACV,WAAK,KAAK,EAAE,MAAM,kBAAkB,IAAI,IAAI,IAAI,OAAO,UAAU,KAAK,gBAAgB,EAAA,CAAG;AAAA,IAC7F;AAAA,EACJ;AAoBA,MAAI,QAAuB,QAAQ,QAAA;AACnC,QAAM,YAAY,CAAC,SAAoC;AACnD,YAAQ,MAAM,KAAK,IAAI,EAAE,MAAM,MAAM;AAAA,IAA2C,CAAC;AAAA,EACrF;AAEA,SAAO;AAAA,IACH,UAAU,KAAsB;AAC5B,cAAQ,IAAI,MAAA;AAAA,QACR,KAAK;AAAQ,sBAAY,IAAI;AAAiB;AAAA,QAC9C,KAAK;AAAW,oBAAU,MAAM,cAAc,GAAG,CAAC;AAAG;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,QASrD,KAAK;AAAY,oBAAU,IAAI,IAAI,IAAI,IAAI,iBAAiB;AAAG,oBAAU,MAAM,eAAe,GAAG,CAAC;AAAG;AAAA,QACrG,KAAK;AAAU,oBAAU,IAAI,IAAI,EAAE,GAAG,MAAA;AAAS;AAAA,QAC/C,KAAK;AAAW,oBAAU,MAAM,cAAc,GAAG,CAAC;AAAG;AAAA,MAAA;AAAA,IAE7D;AAAA,EAAA;AAER;ACpNA,IAAI,MAA0C;AAE9C,SAAS,iBAAiB,WAAiD;AACvE,WAAS,YAAY;AACjB,QAAI;AACJ,QAAI;AACA,YAAM,MAAM,OAAO,2BAA2B;AAAA,IAClD,SAAS,mBAAmB;AACxB,UAAI,CAAC,UAAW,OAAM;AACtB,YAAM,MAAM;AAAA;AAAA,QAA0B;AAAA;AAAA,IAC1C;AAGA,QAAI,IAAI,mBAAmB;AAC3B,QAAI,IAAI,kBAAkB;AAC1B,WAAO;AAAA,EACX,GAAA;AACA,SAAO;AACX;AASA,MAAM,WAA+D;AAAA,EACjE,mBAAmB,MAAM,OAAO,8BAA8B;AAAA,EAC9D,sBAAsB,MAAM,OAAO,iCAAiC;AACxE;AAEA,SAAS,aAAa,EAAE,MAAM,UAA2E;AAGrG,MAAI,OAAQ,QAAO;AAAA;AAAA,IAA0B;AAAA;AAC7C,SAAO,SAAS,IAAI,EAAA;AACxB;AAKA,MAAM,MAAM;AAEZ,MAAM,OAAO,iBAAiB;AAAA,EAC1B,MAAM,CAAC,YAAY;AAAE,QAAI,YAAY,OAAO;AAAA,EAAG;AAAA,EAC/C;AAAA,EACA;AACJ,CAAC;AAED,IAAI,iBAAiB,WAAW,CAAC,UAAwB;AAAE,OAAK,UAAU,MAAM,IAAI;AAAG,CAAC;"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aparte/provider-transformers",
-  "version": "0.24.0",
+  "version": "0.26.0",
   "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": ">=0.24.0 <1.0.0",
+    "@aparte/core": ">=0.26.0 <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": "0.24.0"
+    "@aparte/core": "0.26.0"
   },
   "keywords": [
     "ai",