@atomic-chat/core 0.13.6 → 0.18.3

package/dist/contracts/events.d.ts

@@ -6,10 +6,10 @@
 import type { DiffusionErrorEvent, DiffusionJobEvent, DiffusionVideoJobEvent, DiffusionVideoProgressEvent, DiffusionProgressEvent, DiffusionStateEvent } from './diffusion.js';
 import type { EngineBuildChangedEvent } from './engine-builds.js';
 import type { EngineChangedEvent } from './engines.js';
+import type { FoundryLocalAcceleratorsEvent } from './foundry-local.js';
 import type { EnvironmentOperation, EnvironmentSnapshot } from './environment.js';
 import type { DecisionErrorEvent, DecisionStateEvent } from './decision.js';
 import type { EmbeddingErrorEvent, EmbeddingStateEvent } from './embedding.js';
-import type { ModelSetup } from './model-setup.js';
 import type { RemoteAccessStatus } from './remote-access.js';
 import type { LocalProviderId, RuntimeDeviceInfo, SessionInfo, SessionLoadStage } from './session.js';
 export type DownloadKind = 'model' | 'backend' | 'draft' | 'cudart';
@@ -24,6 +24,17 @@ export interface DownloadStage {
     attempt: number;
     maxAttempts: number;
 }
+/**
+ * One running download's combined progress, as its last `download:progress` event said: what
+ * `GET /atomic/v1/snapshot` lists under `downloads`. The core downloads engine builds only; model
+ * files are the clients' (ADR 2026-10-15-core-does-not-download-models).
+ */
+export interface DownloadProgress {
+    taskId: string;
+    transferred: number;
+    total: number;
+    percent: number;
+}
 export interface CoreEvents {
     'download:started': {
         taskId: string;
@@ -221,6 +232,12 @@ export interface CoreEvents {
      * the desktop releases that listen to it.
      */
     'engine:changed': EngineChangedEvent;
+    /**
+     * One accelerator's step of `POST /runtimes/foundry-local/accelerators/prepare` (change
+     * `add-foundry-local`, design D3): progress while the library downloads it, then `done` or
+     * `failed`. Re-read `GET /runtimes/foundry-local/accelerators` after the last one.
+     */
+    'foundry-local:accelerators': FoundryLocalAcceleratorsEvent;
     /**
      * The decision model (ADR 2026-09-30-the-decision-model-is-its-own-core-module). `state` on every
      * status change (the payload is the whole `DecisionStatus`, like `GET /decision/status`); `error`
@@ -235,11 +252,6 @@ export interface CoreEvents {
      */
     'embedding:state': EmbeddingStateEvent;
     'embedding:error': EmbeddingErrorEvent;
-    /**
-     * Every write of a model setup (`src/contracts/model-setup.ts`), the whole record: keep the one
-     * with the highest `revision` per `setup_id`.
-     */
-    'model-setup:changed': ModelSetup;
     /**
      * One request to the Local API Server, for the app's analytics window and its API screen
      * (PLAN.md §2 decision 15). `started` and `progress` are sent only while the app's inspector is

package/dist/contracts/foundry-local.d.ts

@@ -0,0 +1,77 @@
+/**
+ * Foundry Local's accelerators (openspec change `add-foundry-local`, task 6.4, design D3): what the
+ * library of the active build finds on this machine, and the explicit, client-ordered step that
+ * downloads them (NPU plugins through Windows ML, CUDA and WebGPU from Microsoft's CDN). Nothing
+ * downloads an accelerator on its own: not the start, not an install, not a model load.
+ *
+ * Browser-safe: types and constants only. snake_case on the wire, like `/engine-builds`.
+ */
+import type { FoundryLocalExecutionProvider } from './model-yml.js';
+export interface FoundryLocalAccelerator {
+    /** The library's name (`OpenVINOExecutionProvider`, `CUDAExecutionProvider`, …). */
+    name: string;
+    /** The conf/`model.yml` spelling; `null` for one this core runs no variant on (TensorRT-RTX). */
+    execution_provider: FoundryLocalExecutionProvider | null;
+    /** `gpu`, `npu`, `cpu`, or `gpu/npu` for OpenVINO, which Windows ML offers without an NPU too. */
+    device: string;
+    /** Downloaded and registered by a `prepare` on this machine. */
+    prepared: boolean;
+    /** Why the last `prepare` did not prepare it, in the library's words when it had any. */
+    reason?: string;
+}
+export interface FoundryLocalAccelerators {
+    /** `false`: no Foundry Local build is installed; the list is empty and nothing was probed. */
+    installed: boolean;
+    /** The build probed: `<tag>/<platform>` directory name pair, e.g. `v2.1.0/win-x64`. */
+    build: string | null;
+    /** The library's version string and C API version, from the probe. */
+    version: string | null;
+    c_api_version: number | null;
+    /** Milliseconds since the epoch of the probe this answer is; `null` when none ran. */
+    probed_at: number | null;
+    accelerators: FoundryLocalAccelerator[];
+    /** Why the probe failed (the library does not load, …); the list is then empty. */
+    error: string | null;
+}
+export interface FoundryLocalPrepareRequest {
+    /** The library names to prepare; absent or empty: every accelerator the library offers here. */
+    accelerators?: string[];
+}
+export interface FoundryLocalPrepareResult {
+    accelerators: Array<{
+        name: string;
+        prepared: boolean;
+        reason?: string;
+    }>;
+}
+export declare const FOUNDRY_LOCAL_PREPARE_STAGES: readonly ["started", "downloading", "done", "failed"];
+export type FoundryLocalPrepareStage = (typeof FOUNDRY_LOCAL_PREPARE_STAGES)[number];
+/**
+ * `foundry-local:accelerators`: one accelerator's step of a `prepare`. `fraction` is 0…1 while a CDN
+ * download runs; Windows ML accelerators report none, only `started` and `done`/`failed`, and a
+ * `failed` step has `fraction: null` (rulings R6 of the change).
+ */
+export interface FoundryLocalAcceleratorsEvent {
+    accelerator: string;
+    stage: FoundryLocalPrepareStage;
+    fraction: number | null;
+    reason?: string;
+}
+/** `POST /models/foundry-local/check`: the variant's accelerator and device, as conf and `model.yml` spell them. */
+export interface FoundryLocalCheckRequest {
+    execution_provider: string;
+    device: string;
+}
+export declare const FOUNDRY_LOCAL_CHECK_VERDICTS: readonly ["ok", "accelerator-not-prepared", "accelerator-unavailable", "platform-unsupported", "engine-not-installed"];
+export type FoundryLocalCheckVerdict = (typeof FOUNDRY_LOCAL_CHECK_VERDICTS)[number];
+/**
+ * Whether a Foundry Local variant would run here, from the platform, the installed build and the
+ * last accelerators probe — never the network (spec "Совместимость варианта проверяется в core").
+ */
+export interface FoundryLocalCheckResult {
+    verdict: FoundryLocalCheckVerdict;
+    /** Why, in words a card can show; `null` for `ok`. */
+    reason: string | null;
+    /** The library name of the accelerator the variant needs; `null` for the CPU. */
+    accelerator: string | null;
+}

package/dist/contracts/foundry-local.js

@@ -0,0 +1,17 @@
+/**
+ * Foundry Local's accelerators (openspec change `add-foundry-local`, task 6.4, design D3): what the
+ * library of the active build finds on this machine, and the explicit, client-ordered step that
+ * downloads them (NPU plugins through Windows ML, CUDA and WebGPU from Microsoft's CDN). Nothing
+ * downloads an accelerator on its own: not the start, not an install, not a model load.
+ *
+ * Browser-safe: types and constants only. snake_case on the wire, like `/engine-builds`.
+ */
+export const FOUNDRY_LOCAL_PREPARE_STAGES = ['started', 'downloading', 'done', 'failed'];
+export const FOUNDRY_LOCAL_CHECK_VERDICTS = [
+    'ok',
+    'accelerator-not-prepared',
+    'accelerator-unavailable',
+    'platform-unsupported',
+    'engine-not-installed',
+];
+//# sourceMappingURL=foundry-local.js.map

package/dist/contracts/foundry-local.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"foundry-local.js","sourceRoot":"","sources":["../../src/contracts/foundry-local.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAyCH,MAAM,CAAC,MAAM,4BAA4B,GAAG,CAAC,SAAS,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,CAAU,CAAA;AAqBjG,MAAM,CAAC,MAAM,4BAA4B,GAAG;IAC1C,IAAI;IACJ,0BAA0B;IAC1B,yBAAyB;IACzB,sBAAsB;IACtB,sBAAsB;CACd,CAAA"}

package/dist/contracts/index.d.ts

@@ -9,12 +9,13 @@ export * from './events.js';
 export * from './control-api.js';
 export * from './remote-access.js';
 export * from './diffusion.js';
+export * from './media.js';
 export * from './decision.js';
 export * from './embedding.js';
 export * from './hardware.js';
 export * from './backend-advisor.js';
 export * from './environment.js';
 export * from './model-compatibility.js';
-export * from './model-setup.js';
 export * from './engine-builds.js';
 export * from './engines.js';
+export * from './foundry-local.js';

package/dist/contracts/index.js

@@ -9,13 +9,14 @@ export * from './events.js';
 export * from './control-api.js';
 export * from './remote-access.js';
 export * from './diffusion.js';
+export * from './media.js';
 export * from './decision.js';
 export * from './embedding.js';
 export * from './hardware.js';
 export * from './backend-advisor.js';
 export * from './environment.js';
 export * from './model-compatibility.js';
-export * from './model-setup.js';
 export * from './engine-builds.js';
 export * from './engines.js';
+export * from './foundry-local.js';
 //# sourceMappingURL=index.js.map

package/dist/contracts/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/contracts/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AACH,cAAc,aAAa,CAAA;AAC3B,cAAc,cAAc,CAAA;AAC5B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,eAAe,CAAA;AAC7B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,oBAAoB,CAAA;AAClC,cAAc,gBAAgB,CAAA;AAC9B,cAAc,eAAe,CAAA;AAC7B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,eAAe,CAAA;AAC7B,cAAc,sBAAsB,CAAA;AACpC,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,cAAc,kBAAkB,CAAA;AAChC,cAAc,oBAAoB,CAAA;AAClC,cAAc,cAAc,CAAA"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/contracts/index.ts"],"names":[],"mappings":"AAAA;;GAEG;AACH,cAAc,aAAa,CAAA;AAC3B,cAAc,cAAc,CAAA;AAC5B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,eAAe,CAAA;AAC7B,cAAc,aAAa,CAAA;AAC3B,cAAc,kBAAkB,CAAA;AAChC,cAAc,oBAAoB,CAAA;AAClC,cAAc,gBAAgB,CAAA;AAC9B,cAAc,YAAY,CAAA;AAC1B,cAAc,eAAe,CAAA;AAC7B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,eAAe,CAAA;AAC7B,cAAc,sBAAsB,CAAA;AACpC,cAAc,kBAAkB,CAAA;AAChC,cAAc,0BAA0B,CAAA;AACxC,cAAc,oBAAoB,CAAA;AAClC,cAAc,cAAc,CAAA;AAC5B,cAAc,oBAAoB,CAAA"}

package/dist/contracts/media.d.ts

@@ -0,0 +1,186 @@
+/**
+ * Remote media: the image and video models cloud providers and self-hosted servers offer through the
+ * core (ADR 2026-10-13-remote-media-shares-the-diffusion-job-table). The catalog shapes are conf's
+ * `providers/media.json` as the core serves it (snake_case, like the cloud provider routes); what a
+ * generation request carries for a remote model sits in the camelCase diffusion requests
+ * (`target`, `remote`).
+ */
+/** The adapters the core has: one per media API it speaks. */
+export declare const MEDIA_DIALECTS: readonly ["openai-images", "xai-images", "openrouter-images", "gemini-image", "minimax-image", "dashscope-image", "hf-images", "nvidia-genai", "mistral-agent-image", "a1111", "openai-videos", "openrouter-videos", "xai-videos", "gemini-veo", "minimax-video", "minimax-video-v2", "dashscope-video", "aimlapi-videos", "hf-videos"];
+export type MediaDialect = (typeof MEDIA_DIALECTS)[number];
+export type MediaKind = 'image' | 'video';
+/** Which kind of media each dialect makes. */
+export declare const MEDIA_DIALECT_KIND: Readonly<Record<MediaDialect, MediaKind>>;
+/** What a remote image model can be asked to do; a subset of `ImageWorkflowId`. */
+export declare const REMOTE_IMAGE_WORKFLOWS: readonly ["create", "transform", "inpaint", "reference", "edit"];
+export type RemoteImageWorkflow = (typeof REMOTE_IMAGE_WORKFLOWS)[number];
+export declare const REMOTE_VIDEO_WORKFLOWS: readonly ["create", "image-to-video"];
+export type RemoteVideoWorkflow = (typeof REMOTE_VIDEO_WORKFLOWS)[number];
+/** A closed list of values a control offers, with the one it starts on. */
+export interface MediaChoice {
+    values: string[];
+    default?: string;
+}
+export interface MediaNumberRange {
+    min: number;
+    max: number;
+    default?: number;
+    step?: number;
+}
+/** Pixel size: a list of `WxH` (and `auto`), or a free range on a grid. */
+export type MediaSizeParam = {
+    kind: 'enum';
+    values: string[];
+    default?: string;
+} | {
+    kind: 'range';
+    min: number;
+    max: number;
+    multiple: number;
+    default_width?: number;
+    default_height?: number;
+};
+/** What the generation form offers for an image model; a missing key is a hidden control. */
+export interface MediaImageParams {
+    size?: MediaSizeParam;
+    aspect_ratios?: MediaChoice;
+    resolutions?: MediaChoice;
+    n?: {
+        max: number;
+    };
+    seed?: boolean;
+    negative_prompt?: boolean;
+    quality?: MediaChoice;
+    background?: MediaChoice;
+    output_format?: MediaChoice;
+    steps?: MediaNumberRange;
+    cfg?: MediaNumberRange;
+    guidance?: MediaNumberRange;
+    strength?: MediaNumberRange;
+    /** The server lists samplers live (`MediaModelDescriptor.samplers`). */
+    sampler?: boolean;
+    max_input_images?: number;
+}
+/** Clip length in seconds: a list of values, or a range. */
+export interface MediaDurations {
+    values?: number[];
+    min?: number;
+    max?: number;
+    default?: number;
+}
+export interface MediaVideoParams {
+    durations?: MediaDurations;
+    resolutions?: MediaChoice;
+    aspect_ratios?: MediaChoice;
+    sizes?: MediaChoice;
+    /** The model can generate sound; the form offers a switch. */
+    audio?: boolean;
+    seed?: boolean;
+    negative_prompt?: boolean;
+    /** Image to video also takes a last frame. */
+    last_frame?: boolean;
+}
+export type MediaModelOptions = Record<string, string | number | boolean>;
+interface MediaModelCommon {
+    id: string;
+    name: string;
+    description?: string;
+    /** The adapter that serves this model (the model's own, else its provider section's). */
+    api: MediaDialect;
+    /** Offered as soon as the provider is connected. */
+    default?: boolean;
+    options?: MediaModelOptions;
+    /** `catalog`: conf names it. `live`: only the provider's listing does; the dialect's defaults apply. */
+    source: 'catalog' | 'live';
+}
+export interface ImageModelDescriptor extends MediaModelCommon {
+    kind: 'image';
+    workflows: RemoteImageWorkflow[];
+    params: MediaImageParams;
+    /** Samplers the server listed, when `params.sampler` is set (A1111/Forge). */
+    samplers?: string[];
+}
+export interface VideoModelDescriptor extends MediaModelCommon {
+    kind: 'video';
+    workflows: RemoteVideoWorkflow[];
+    params: MediaVideoParams;
+}
+export type MediaModelDescriptor = ImageModelDescriptor | VideoModelDescriptor;
+export interface MediaDialectDefaults {
+    kind: MediaKind;
+    workflows: string[];
+    params: MediaImageParams | MediaVideoParams;
+}
+/** One provider's media, as `GET /atomic/v1/media/catalog` answers it. */
+export interface MediaProviderSection {
+    image?: {
+        api: MediaDialect;
+        models: ImageModelDescriptor[];
+    };
+    video?: {
+        api: MediaDialect;
+        models: VideoModelDescriptor[];
+    };
+}
+/** `GET /atomic/v1/media/catalog`. */
+export interface MediaCatalogView {
+    catalog_version: number;
+    /** `live` when the conf document was fetched, `cache` from disk, `baseline` the bundled copy. */
+    origin: 'live' | 'cache' | 'baseline';
+    dialects: Partial<Record<MediaDialect, MediaDialectDefaults>>;
+    providers: Record<string, MediaProviderSection>;
+}
+/** `GET /atomic/v1/media/providers/:provider/models?kind=`: the catalog's models plus the live listing. */
+export interface MediaProviderModels {
+    provider: string;
+    kind: MediaKind;
+    /** The adapter the provider's media of this kind go through; null when it has none. */
+    api: MediaDialect | null;
+    models: MediaModelDescriptor[];
+    /** Whether the provider was asked live; false for adapters without a listing. */
+    live: boolean;
+    /** Why the live listing failed; the catalog's models are still answered. */
+    error?: {
+        code: string;
+        message: string;
+    };
+}
+/** One enabled media model in a provider registration. */
+export interface CloudProviderMediaModel {
+    id: string;
+    kind: MediaKind;
+}
+/**
+ * The media part of `PUT /atomic/v1/cloud/providers/:provider`. Absent keeps what is stored, `null`
+ * clears it. `image_api` / `video_api` name the adapter for a provider the catalog does not know
+ * (a custom server); absent, the catalog's applies.
+ */
+export interface CloudProviderMedia {
+    image_api?: MediaDialect | null;
+    video_api?: MediaDialect | null;
+    models: CloudProviderMediaModel[];
+}
+/** Which provider and model a generation request is for; absent on a request for the local model. */
+export interface MediaTarget {
+    provider: string;
+    model: string;
+}
+/** Image settings only a remote model takes; every one optional, the model's default otherwise. */
+export interface RemoteImageOptions {
+    /** `auto` or `WxH`, for a model whose size is a list; a range model uses `width`/`height`. */
+    size?: string;
+    aspectRatio?: string;
+    resolution?: string;
+    quality?: string;
+    background?: string;
+    outputFormat?: string;
+}
+export interface RemoteVideoOptions {
+    durationSecs?: number;
+    resolution?: string;
+    aspectRatio?: string;
+    /** `WxH`, for a model whose size is a list. */
+    size?: string;
+    generateAudio?: boolean;
+}
+export {};

package/dist/contracts/media.js

@@ -0,0 +1,55 @@
+/**
+ * Remote media: the image and video models cloud providers and self-hosted servers offer through the
+ * core (ADR 2026-10-13-remote-media-shares-the-diffusion-job-table). The catalog shapes are conf's
+ * `providers/media.json` as the core serves it (snake_case, like the cloud provider routes); what a
+ * generation request carries for a remote model sits in the camelCase diffusion requests
+ * (`target`, `remote`).
+ */
+/** The adapters the core has: one per media API it speaks. */
+export const MEDIA_DIALECTS = [
+    'openai-images',
+    'xai-images',
+    'openrouter-images',
+    'gemini-image',
+    'minimax-image',
+    'dashscope-image',
+    'hf-images',
+    'nvidia-genai',
+    'mistral-agent-image',
+    'a1111',
+    'openai-videos',
+    'openrouter-videos',
+    'xai-videos',
+    'gemini-veo',
+    'minimax-video',
+    'minimax-video-v2',
+    'dashscope-video',
+    'aimlapi-videos',
+    'hf-videos',
+];
+/** Which kind of media each dialect makes. */
+export const MEDIA_DIALECT_KIND = {
+    'openai-images': 'image',
+    'xai-images': 'image',
+    'openrouter-images': 'image',
+    'gemini-image': 'image',
+    'minimax-image': 'image',
+    'dashscope-image': 'image',
+    'hf-images': 'image',
+    'nvidia-genai': 'image',
+    'mistral-agent-image': 'image',
+    'a1111': 'image',
+    'openai-videos': 'video',
+    'openrouter-videos': 'video',
+    'xai-videos': 'video',
+    'gemini-veo': 'video',
+    'minimax-video': 'video',
+    'minimax-video-v2': 'video',
+    'dashscope-video': 'video',
+    'aimlapi-videos': 'video',
+    'hf-videos': 'video',
+};
+/** What a remote image model can be asked to do; a subset of `ImageWorkflowId`. */
+export const REMOTE_IMAGE_WORKFLOWS = ['create', 'transform', 'inpaint', 'reference', 'edit'];
+export const REMOTE_VIDEO_WORKFLOWS = ['create', 'image-to-video'];
+//# sourceMappingURL=media.js.map

package/dist/contracts/media.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"media.js","sourceRoot":"","sources":["../../src/contracts/media.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,8DAA8D;AAC9D,MAAM,CAAC,MAAM,cAAc,GAAG;IAC5B,eAAe;IACf,YAAY;IACZ,mBAAmB;IACnB,cAAc;IACd,eAAe;IACf,iBAAiB;IACjB,WAAW;IACX,cAAc;IACd,qBAAqB;IACrB,OAAO;IACP,eAAe;IACf,mBAAmB;IACnB,YAAY;IACZ,YAAY;IACZ,eAAe;IACf,kBAAkB;IAClB,iBAAiB;IACjB,gBAAgB;IAChB,WAAW;CACH,CAAA;AAKV,8CAA8C;AAC9C,MAAM,CAAC,MAAM,kBAAkB,GAA8C;IAC3E,eAAe,EAAE,OAAO;IACxB,YAAY,EAAE,OAAO;IACrB,mBAAmB,EAAE,OAAO;IAC5B,cAAc,EAAE,OAAO;IACvB,eAAe,EAAE,OAAO;IACxB,iBAAiB,EAAE,OAAO;IAC1B,WAAW,EAAE,OAAO;IACpB,cAAc,EAAE,OAAO;IACvB,qBAAqB,EAAE,OAAO;IAC9B,OAAO,EAAE,OAAO;IAChB,eAAe,EAAE,OAAO;IACxB,mBAAmB,EAAE,OAAO;IAC5B,YAAY,EAAE,OAAO;IACrB,YAAY,EAAE,OAAO;IACrB,eAAe,EAAE,OAAO;IACxB,kBAAkB,EAAE,OAAO;IAC3B,iBAAiB,EAAE,OAAO;IAC1B,gBAAgB,EAAE,OAAO;IACzB,WAAW,EAAE,OAAO;CACrB,CAAA;AAED,mFAAmF;AACnF,MAAM,CAAC,MAAM,sBAAsB,GAAG,CAAC,QAAQ,EAAE,WAAW,EAAE,SAAS,EAAE,WAAW,EAAE,MAAM,CAAU,CAAA;AAGtG,MAAM,CAAC,MAAM,sBAAsB,GAAG,CAAC,QAAQ,EAAE,gBAAgB,CAAU,CAAA"}

package/dist/contracts/model-yml.d.ts

@@ -28,6 +28,8 @@ export interface ModelYml {
      * llama.cpp provider; the load gate still reads the header either way.
      */
     atomic_runtime?: ModelYmlAtomicRuntime;
+    /** A Foundry Local variant (`<data>/foundry-local/models/<variant_id>/`), written by the app's installer. */
+    foundry_local?: ModelYmlFoundryLocal;
 }
 export interface ModelYmlAtomicRuntime {
     provider: 'atomic-prism';
@@ -38,5 +40,31 @@ export interface ModelYmlAtomicRuntime {
     /** Conf model-rules family id. */
     family?: string;
 }
+/** The accelerators a Foundry Local variant may name, as conf's model list spells them. */
+export declare const FOUNDRY_LOCAL_EXECUTION_PROVIDERS: readonly ["cpu", "cuda", "webgpu", "openvino", "qnn", "vitisai"];
+export type FoundryLocalExecutionProvider = (typeof FOUNDRY_LOCAL_EXECUTION_PROVIDERS)[number];
+export declare const FOUNDRY_LOCAL_DEVICES: readonly ["cpu", "gpu", "npu"];
+export type FoundryLocalDevice = (typeof FOUNDRY_LOCAL_DEVICES)[number];
+/**
+ * The `foundry_local` block (openspec change `add-foundry-local`, design D4): which variant the folder
+ * holds and what the host registers it with. The accelerator is passed to the library explicitly.
+ */
+export interface ModelYmlFoundryLocal {
+    /** The model the variant belongs to (conf list id, or the catalog's parent model). */
+    model_id: string;
+    variant_id: string;
+    execution_provider: FoundryLocalExecutionProvider;
+    device: FoundryLocalDevice;
+    /** `azureml://registries/<r>/models/<name>/versions/<v>`. */
+    asset_id?: string;
+    asset_version?: string;
+    /** The variant's context, when it is smaller than the model's (NPU builds). */
+    context_length?: number;
+    tools: boolean;
+    reasoning: boolean;
+    /** The markers the library splits reasoning on; without them `<think>` lands in the text (operator decision Q1). */
+    reasoning_start?: string;
+    reasoning_end?: string;
+}
 /** Key order the app's serde_yaml writer produces; `writeModelYml` must emit keys in this order. */
 export declare const MODEL_YML_KEY_ORDER: readonly (keyof ModelYml)[];

package/dist/contracts/model-yml.js

@@ -2,6 +2,16 @@
  * `<data>/llamacpp/models/<id>/model.yml` — the file that defines a model. Shared with the app;
  * see PLAN.md §8.1. Canonical shape: guest-js/types.ts:142-171 (+ `capabilities` from cli/mod.rs).
  */
+/** The accelerators a Foundry Local variant may name, as conf's model list spells them. */
+export const FOUNDRY_LOCAL_EXECUTION_PROVIDERS = [
+    'cpu',
+    'cuda',
+    'webgpu',
+    'openvino',
+    'qnn',
+    'vitisai',
+];
+export const FOUNDRY_LOCAL_DEVICES = ['cpu', 'gpu', 'npu'];
 /** Key order the app's serde_yaml writer produces; `writeModelYml` must emit keys in this order. */
 export const MODEL_YML_KEY_ORDER = [
     'model_path',
@@ -20,5 +30,6 @@ export const MODEL_YML_KEY_ORDER = [
     'source',
     'capabilities',
     'atomic_runtime',
+    'foundry_local',
 ];
 //# sourceMappingURL=model-yml.js.map

package/dist/contracts/model-yml.js.map

@@ -1 +1 @@
-{"version":3,"file":"model-yml.js","sourceRoot":"","sources":["../../src/contracts/model-yml.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAkDH,oGAAoG;AACpG,MAAM,CAAC,MAAM,mBAAmB,GAAgC;IAC9D,YAAY;IACZ,aAAa;IACb,MAAM;IACN,YAAY;IACZ,cAAc;IACd,kBAAkB;IAClB,eAAe;IACf,mBAAmB;IACnB,WAAW;IACX,kBAAkB;IAClB,iBAAiB;IACjB,gBAAgB;IAChB,mBAAmB;IACnB,QAAQ;IACR,cAAc;IACd,gBAAgB;CACjB,CAAA"}
+{"version":3,"file":"model-yml.js","sourceRoot":"","sources":["../../src/contracts/model-yml.ts"],"names":[],"mappings":"AAAA;;;GAGG;AAoDH,2FAA2F;AAC3F,MAAM,CAAC,MAAM,iCAAiC,GAAG;IAC/C,KAAK;IACL,MAAM;IACN,QAAQ;IACR,UAAU;IACV,KAAK;IACL,SAAS;CACD,CAAA;AAGV,MAAM,CAAC,MAAM,qBAAqB,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAU,CAAA;AAyBnE,oGAAoG;AACpG,MAAM,CAAC,MAAM,mBAAmB,GAAgC;IAC9D,YAAY;IACZ,aAAa;IACb,MAAM;IACN,YAAY;IACZ,cAAc;IACd,kBAAkB;IAClB,eAAe;IACf,mBAAmB;IACnB,WAAW;IACX,kBAAkB;IAClB,iBAAiB;IACjB,gBAAgB;IAChB,mBAAmB;IACnB,QAAQ;IACR,cAAc;IACd,gBAAgB;IAChB,eAAe;CAChB,CAAA"}

package/dist/contracts/session.d.ts

@@ -10,8 +10,21 @@
  * `atomic-prism` is PrismML's llama.cpp fork, the only engine that runs Bonsai's Prism packings
  * (`PQ2_0`, `PTQ1_0`, the Hadamard transform). It shares the GGUF model tree with the two other
  * llama.cpp providers and never becomes the default for an ordinary model.
+ *
+ * `foundry-local` is Microsoft Foundry Local (ONNX Runtime GenAI: NPU, CUDA, WebGPU, CPU), run by a
+ * host process of this binary per model (openspec change `add-foundry-local`); registered only on
+ * the platforms its archives support.
+ */
+export type LocalProviderId = 'llamacpp-upstream' | 'llamacpp' | 'atomic-prism' | 'mlx' | 'foundation-models' | 'foundry-local' | ManagedProviderId;
+/**
+ * The managed engines — each one a container core runs from a model of the managed model store — in
+ * the order their registry lists them (change `add-sglang-runtime`, D1). The one list every place
+ * that must tell a managed provider from the others reads, rather than naming them one by one; the
+ * managed engine registry refuses, at startup, an engine missing from it.
  */
-export type LocalProviderId = 'llamacpp-upstream' | 'llamacpp' | 'atomic-prism' | 'mlx' | 'foundation-models' | 'tensorrt-llm' | 'vllm';
+export declare const MANAGED_PROVIDER_IDS: readonly ["tensorrt-llm", "vllm", "sglang"];
+export type ManagedProviderId = (typeof MANAGED_PROVIDER_IDS)[number];
+export declare function isManagedProviderId(provider: string): provider is ManagedProviderId;
 export interface RuntimeDeviceInfo {
     /** Backends in load order, deduped, e.g. ["CUDA", "CPU"]. */
     loaded_backends: string[];

package/dist/contracts/session.js

@@ -3,6 +3,16 @@
  * and by the Rust agent (`core_sessions` mirror). Byte-identical with
  * src-tauri/plugins/tauri-plugin-llamacpp-upstream/guest-js/types.ts:122-132 and state.rs:9-23.
  */
+/**
+ * The managed engines — each one a container core runs from a model of the managed model store — in
+ * the order their registry lists them (change `add-sglang-runtime`, D1). The one list every place
+ * that must tell a managed provider from the others reads, rather than naming them one by one; the
+ * managed engine registry refuses, at startup, an engine missing from it.
+ */
+export const MANAGED_PROVIDER_IDS = ['tensorrt-llm', 'vllm', 'sglang'];
+export function isManagedProviderId(provider) {
+    return MANAGED_PROVIDER_IDS.includes(provider);
+}
 /**
  * Stage of a managed-runtime model load, reported on `session:load-progress` (spec
  * `tensorrt-llm-runtime`, design D8): `stopping-previous` evicts whatever currently holds the GPU,

package/dist/contracts/session.js.map

@@ -1 +1 @@
-{"version":3,"file":"session.js","sourceRoot":"","sources":["../../src/contracts/session.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAyDH;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAG;IACjC,mBAAmB;IACnB,oBAAoB;IACpB,qBAAqB;IACrB,OAAO;CACC,CAAA"}
+{"version":3,"file":"session.js","sourceRoot":"","sources":["../../src/contracts/session.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAuBH;;;;;GAKG;AACH,MAAM,CAAC,MAAM,oBAAoB,GAAG,CAAC,cAAc,EAAE,MAAM,EAAE,QAAQ,CAAU,CAAA;AAG/E,MAAM,UAAU,mBAAmB,CAAC,QAAgB;IAClD,OAAQ,oBAA0C,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAA;AACvE,CAAC;AA8CD;;;;;;;GAOG;AACH,MAAM,CAAC,MAAM,mBAAmB,GAAG;IACjC,mBAAmB;IACnB,oBAAoB;IACpB,qBAAqB;IACrB,OAAO;CACC,CAAA"}

package/dist/contracts/settings.d.ts

@@ -80,6 +80,10 @@ export interface MlxConfig {
 export interface FoundationModelsConfig {
     timeout: number;
 }
+/** `foundry-local` (change `add-foundry-local`): seconds a model may take to load in its host. */
+export interface FoundryLocalConfig {
+    timeout: number;
+}
 /**
  * `tensorrt-llm` as the settings store holds it (`src/settings/schema/tensorrt-llm.json`). `gpu_id: ''`
  * lets a load pick the card with the most free memory and `load_timeout_seconds: 0` keeps the engine's own
@@ -123,7 +127,42 @@ export interface VllmProviderConfig {
     kv_cache_dtype: 'auto' | 'fp8';
     load_timeout_seconds: number;
 }
-export type ProviderSettings<P extends LocalProviderId> = P extends 'mlx' ? MlxConfig : P extends 'foundation-models' ? FoundationModelsConfig : P extends 'tensorrt-llm' ? TensorrtLlmProviderConfig : P extends 'vllm' ? VllmProviderConfig : LlamacppConfig;
+/**
+ * `sglang` (change `add-sglang-runtime`, spec `sglang-runtime`, design D6): stored values. `0` and `''`
+ * are "not set" and reach the engine as nothing: SGLang's own value, or core's own memory share.
+ */
+export interface SglangProviderConfig {
+    /** The card's UUID; `''` picks the card with the most free memory at load time. */
+    gpu_id: string;
+    context_length: number;
+    /** Enforced per request by the session gateway and the engine's default output limit. */
+    max_output_tokens: number;
+    /** `--max-running-requests`: requests served at once. */
+    max_running_requests: number;
+    /** `--mem-fraction-static`; `0` (auto) lets core compute it from the card's free memory. */
+    mem_fraction_static: number;
+    /** `--max-total-tokens`, the KV cache in tokens; `0` lets SGLang size it from its share. */
+    max_total_tokens: number;
+    /** `auto` (graphs captured up to `max_running_requests`), `on` (SGLang's own) or `off`. */
+    cuda_graphs: 'auto' | 'on' | 'off';
+    /** `auto` (the model's own precision) or `fp8` (Ada and newer; ignored on older cards). */
+    kv_cache_dtype: 'auto' | 'fp8';
+    load_timeout_seconds: number;
+    /** `false` passes `--disable-radix-cache`. */
+    enable_prefix_caching: boolean;
+    /** `--chunked-prefill-size`; `0` leaves SGLang's own. */
+    chunked_prefill_size: number;
+    dtype: 'auto' | 'float16' | 'bfloat16' | 'float32';
+    /** `--random-seed`; `0` leaves SGLang's own. */
+    seed: number;
+    /** Sampling defaults for a request that sets none; `''` is the model's own. */
+    default_temperature: string;
+    default_top_p: string;
+    default_top_k: string;
+    default_min_p: string;
+    default_repetition_penalty: string;
+}
+export type ProviderSettings<P extends LocalProviderId> = P extends 'mlx' ? MlxConfig : P extends 'foundation-models' ? FoundationModelsConfig : P extends 'tensorrt-llm' ? TensorrtLlmProviderConfig : P extends 'vllm' ? VllmProviderConfig : P extends 'sglang' ? SglangProviderConfig : P extends 'foundry-local' ? FoundryLocalConfig : LlamacppConfig;
 /** UI descriptor of one setting (mirrors the app's SettingComponentProps). */
 export interface SettingDescriptor {
     key: string;

package/dist/contracts/settings.js.map

@@ -1 +1 @@
-{"version":3,"file":"settings.js","sourceRoot":"","sources":["../../src/contracts/settings.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AA8KH,MAAM,CAAC,MAAM,uBAAuB,GAAmB;IACrD,IAAI,EAAE,WAAW;IACjB,IAAI,EAAE,IAAI;IACV,MAAM,EAAE,KAAK;IACb,OAAO,EAAE,EAAE;IACX,aAAa,EAAE,EAAE;IACjB,YAAY,EAAE,IAAI;IAClB,gBAAgB,EAAE,OAAO;IACzB,iBAAiB,EAAE,KAAK;IACxB,YAAY,EAAE,KAAK;CACpB,CAAA"}
+{"version":3,"file":"settings.js","sourceRoot":"","sources":["../../src/contracts/settings.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AA2NH,MAAM,CAAC,MAAM,uBAAuB,GAAmB;IACrD,IAAI,EAAE,WAAW;IACjB,IAAI,EAAE,IAAI;IACV,MAAM,EAAE,KAAK;IACb,OAAO,EAAE,EAAE;IACX,aAAa,EAAE,EAAE;IACjB,YAAY,EAAE,IAAI;IAClB,gBAAgB,EAAE,OAAO;IACzB,iBAAiB,EAAE,KAAK;IACxB,YAAY,EAAE,KAAK;CACpB,CAAA"}