@ai-matrx/desktop-protocol 0.21.42 → 0.21.44

package/dist/client.d.cts

@@ -1,8 +1,8 @@
 import { BackoffOptions } from '@ai-matrx/realtime';
 export { RECONNECT_ALARM_ATTEMPTS } from '@ai-matrx/realtime';
 import { z } from 'zod';
-import { W as Welcome, D as DesktopProtocolError, C as CloseOutcome, O as OpName, a as OpParams, b as OpResult, c as OPS, S as StreamMeta, E as EndReason, d as EVENTS, e as ClientType, M as ModelsTranscribeResult, T as TranscriptionModelId, f as MlProgress } from './models-Ut2nuRzV.cjs';
-export { i as isDesktopProtocolError } from './models-Ut2nuRzV.cjs';
+import { W as Welcome, D as DesktopProtocolError, C as CloseOutcome, O as OpName, a as OpParams, b as OpResult, c as OPS, S as StreamMeta, E as EndReason, d as EVENTS, e as ClientType, M as ModelsTranscribeResult, T as TranscriptionModelId, f as MlProgress } from './models-CCRnwbz8.cjs';
+export { i as isDesktopProtocolError } from './models-CCRnwbz8.cjs';
 import { FrameHeader } from './frame.cjs';
 
 /** The subset of the WHATWG WebSocket the client uses — browser, React Native, `ws` and Workers all fit. */

package/dist/client.d.ts

@@ -1,8 +1,8 @@
 import { BackoffOptions } from '@ai-matrx/realtime';
 export { RECONNECT_ALARM_ATTEMPTS } from '@ai-matrx/realtime';
 import { z } from 'zod';
-import { W as Welcome, D as DesktopProtocolError, C as CloseOutcome, O as OpName, a as OpParams, b as OpResult, c as OPS, S as StreamMeta, E as EndReason, d as EVENTS, e as ClientType, M as ModelsTranscribeResult, T as TranscriptionModelId, f as MlProgress } from './models-Ut2nuRzV.js';
-export { i as isDesktopProtocolError } from './models-Ut2nuRzV.js';
+import { W as Welcome, D as DesktopProtocolError, C as CloseOutcome, O as OpName, a as OpParams, b as OpResult, c as OPS, S as StreamMeta, E as EndReason, d as EVENTS, e as ClientType, M as ModelsTranscribeResult, T as TranscriptionModelId, f as MlProgress } from './models-CCRnwbz8.js';
+export { i as isDesktopProtocolError } from './models-CCRnwbz8.js';
 import { FrameHeader } from './frame.js';
 
 /** The subset of the WHATWG WebSocket the client uses — browser, React Native, `ws` and Workers all fit. */

package/dist/client.js

@@ -112,7 +112,7 @@ import { z as z2 } from "zod";
 import { z } from "zod";
 import { UUID_SHAPE_PATTERN } from "@ai-matrx/kit/uuid";
 var PROTOCOL_MAJOR = 1;
-var PROTOCOL_MINOR = 47;
+var PROTOCOL_MINOR = 49;
 var PROTOCOL_VERSION = `${PROTOCOL_MAJOR}.${PROTOCOL_MINOR}`;
 var ProtocolVersion = z.string().regex(/^[1-9][0-9]{0,2}\.[0-9]{1,3}$/);
 var RequestId = z.string().min(1).max(64).regex(/^[A-Za-z0-9_-]+$/);
@@ -764,7 +764,8 @@ var SessionControlChangedEvent = z8.strictObject({ controller_client_id: z8.stri
 
 // src/schema/models.ts
 import { z as z9 } from "zod";
-var TranscriptionModelId = z9.enum(["tiny", "base", "small", "large-v3-turbo"]);
+var ModelKey = z9.string().min(1).max(128).regex(/^[a-z0-9][a-z0-9._-]*(\/[a-z0-9][a-z0-9._-]*)?$/);
+var TranscriptionModelId = ModelKey;
 var MlRuntimeState = z9.enum(["not_installed", "installing", "installed", "failed"]);
 var MlInstallPhase = z9.enum(["packages", "model", "verifying"]);
 var MlInstallProgress = z9.strictObject({
@@ -824,18 +825,7 @@ var MlProgress = z9.strictObject({
   /** The segment just transcribed (transcribe only): text appears as it is recognised. */
   segment: TranscriptSegment.nullable()
 });
-var TextModelId = z9.enum([
-  "smollm2-135m",
-  "llama-3.2-3b",
-  "phi-4-mini",
-  "gemma-3-4b",
-  "llama-3.1-8b",
-  "qwen-2.5-7b",
-  "mistral-small-3.1",
-  "gpt-oss-20b",
-  "qwen-2.5-coder-7b",
-  "qwen-2.5-coder-14b"
-]);
+var TextModelId = ModelKey;
 var TextModelInfo = z9.strictObject({
   id: TextModelId,
   /** Download size of the weights. */
@@ -843,7 +833,14 @@ var TextModelInfo = z9.strictObject({
   installed: z9.boolean()
 });
 var LlmServerState = z9.enum(["stopped", "starting", "running", "failed"]);
-var LlmRunning = z9.strictObject({ id: TextModelId, port: z9.int().min(1024).max(65535) });
+var LlmRunning = z9.strictObject({
+  id: TextModelId,
+  port: z9.int().min(1024).max(65535),
+  /** (1.48) What the service actually applied, echoed so the app can attest "read and applied". Optional until the service sends them. */
+  context_length: z9.int().min(512).max(131072).optional(),
+  gpu_layers: z9.int().min(-1).max(999).optional(),
+  threads: z9.int().min(1).max(512).optional()
+});
 var ModelsLlmStatusParams = z9.strictObject({});
 var ModelsLlmStatusResult = z9.strictObject({
   state: LlmServerState,
@@ -861,12 +858,57 @@ var ModelsLlmInstallParams = z9.strictObject({ model: TextModelId });
 var ModelsLlmInstallResult = z9.strictObject({ model: TextModelId, elapsed_ms: nonNegInt });
 var ModelsLlmStartParams = z9.strictObject({
   model: TextModelId,
-  /** Tokens of context; the service picks a size that fits this computer when omitted. */
-  context_length: z9.int().min(512).max(131072).optional()
+  /** Tokens of context, from the knob desktop.models/context_length. Omitted today means the service's fixed 4096 (optional until callers send it). */
+  context_length: z9.int().min(512).max(131072).optional(),
+  /** (1.48) Layers offloaded to the GPU; -1 = all. From the knob desktop.models/gpu_layers. */
+  gpu_layers: z9.int().min(-1).max(999).optional(),
+  /** (1.48) CPU threads, computed by the app from the knobs. */
+  threads: z9.int().min(1).max(512).optional(),
+  /** (1.48) Seconds to wait for the model to load, computed by the app from the knobs. */
+  load_timeout_s: z9.int().min(10).max(3600).optional()
 });
 var ModelsLlmStartResult = z9.strictObject({ running: LlmRunning, elapsed_ms: nonNegInt });
 var ModelsLlmStopParams = z9.strictObject({});
 var ModelsLlmStopResult = z9.strictObject({ stopped: z9.boolean() });
+var Sha256Hex = z9.string().regex(/^[0-9a-f]{64}$/);
+var CatalogFile = z9.strictObject({
+  /** Path inside the model's folder (the artifact's path_in_repo). */
+  path: z9.string().min(1).max(512),
+  /** The pinned https URL it is downloaded from. */
+  url: z9.string().url().max(2048),
+  size_bytes: nonNegInt,
+  sha256: Sha256Hex,
+  /** weights, weights_shard, mmproj, config, tokenizer, … (ai.model_artifact.role). */
+  role: z9.string().max(32)
+});
+var CatalogModelEntry = z9.strictObject({
+  /** The key callers name this model by (the 1.48 id for a seeded model). */
+  key: ModelKey,
+  /** The selected variant's key (`q4_k_m`, `ct2`). */
+  variant: z9.string().min(1).max(64),
+  /** The upstream revision (or the variant id when there is none): a change re-downloads. */
+  revision: z9.string().min(1).max(64),
+  multilingual: z9.boolean(),
+  /** Exactly the files this computer fetches for this model: never more, never fewer. */
+  files: z9.array(CatalogFile).min(1).max(256)
+});
+var LlamaServerBuild = z9.strictObject({
+  version: z9.string().min(1).max(32),
+  url: z9.string().url().max(2048),
+  size_bytes: nonNegInt,
+  sha256: Sha256Hex
+});
+var MlCatalogSetParams = z9.strictObject({
+  text: z9.array(CatalogModelEntry).max(512),
+  transcription: z9.array(CatalogModelEntry).max(64),
+  /** Null where the catalog has no build for this os/arch (installs then say so). */
+  llama_server: LlamaServerBuild.nullable(),
+  /** The knob desktop.models/default_transcription_model. */
+  default_transcription: TranscriptionModelId,
+  /** When the catalog was last changed (catalog_updated_at), for the service's log. */
+  catalog_updated_at: z9.string().max(40)
+});
+var MlCatalogSetResult = z9.strictObject({ text: nonNegInt, transcription: nonNegInt });
 var DownloadKind = z9.enum(["speech_model", "text_model", "runtime"]);
 var DownloadRow = z9.strictObject({
   /** The job's key in the service: "model:base", "llm:smollm2-135m", "llm-server", "runtime". */