@ai-matrx/desktop-protocol 0.21.42 → 0.21.44

package/CHANGELOG.md

@@ -1,9 +1,30 @@
 # Changelog
 
+## 0.21.44
+### Unreleased
+
+- **Protocol 1.49 — local models come from the platform catalog.** `TextModelId` and
+  `TranscriptionModelId` become `ModelKey` (a string key, `<model>` or `<model>/<variant>`); every 1.48 id
+  is still valid. New service op `ml.catalog_set` (`MlCatalogSetParams`: per model the selected variant's
+  files with URL, size and SHA-256; the llama-server build; the default transcription model knob) — the
+  `ml` service no longer holds a model list. Consumer action: a Python service built from this contract
+  must serve `ml.catalog_set`; code that switched over the old enums must handle any key.
+
+- **Protocol 1.48 — model start carries the app's knobs, and the service echoes what it applied.**
+  `ModelsLlmStartParams` gains optional `gpu_layers` (-1..999), `threads` (1..512), `load_timeout_s` (10..3600)
+  beside `context_length`; `LlmRunning` gains optional `context_length`, `gpu_layers`, `threads` (what was
+  actually applied, for config attestation). All optional and listed in `FIELDS_SINCE` "1.48"; Matrx 2 makes
+  them required when the Python service and every caller send them. The false "the service picks a size"
+  comment on `context_length` is corrected (the service uses a fixed 4096 when it is omitted).
+
 ## 0.21.42
 
-UNVERIFIED: this release has no Unreleased entry and the source comparison could not be made (the caller supplied no source comparison). Whether source changed is unknown; consumer action is unknown.
-Check the package's gitHead on npm against this version's gitHead: `git diff <previous gitHead> <this gitHead> -- apps/shared/desktop-protocol/src`.
+SOURCE CHANGED WITH NO RELEASE NOTE. This release shipped source changes that nobody described in an Unreleased entry,
+so they may include breaking API changes. Consumer action is UNKNOWN: read the diff before upgrading.
+Diff: `git diff bdadfd428f a5da075c1d -- apps/shared/desktop-protocol/src`
+
+Changed shipped source files (1):
+- `apps/shared/desktop-protocol/src/schema/apps-instances.ts`
 
 ## 0.21.39
 

package/dist/client.cjs

@@ -144,7 +144,7 @@ var import_zod2 = require("zod");
 var import_zod = require("zod");
 var import_uuid = require("@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 = import_zod.z.string().regex(/^[1-9][0-9]{0,2}\.[0-9]{1,3}$/);
 var RequestId = import_zod.z.string().min(1).max(64).regex(/^[A-Za-z0-9_-]+$/);
@@ -796,7 +796,8 @@ var SessionControlChangedEvent = import_zod8.z.strictObject({ controller_client_
 
 // src/schema/models.ts
 var import_zod9 = require("zod");
-var TranscriptionModelId = import_zod9.z.enum(["tiny", "base", "small", "large-v3-turbo"]);
+var ModelKey = import_zod9.z.string().min(1).max(128).regex(/^[a-z0-9][a-z0-9._-]*(\/[a-z0-9][a-z0-9._-]*)?$/);
+var TranscriptionModelId = ModelKey;
 var MlRuntimeState = import_zod9.z.enum(["not_installed", "installing", "installed", "failed"]);
 var MlInstallPhase = import_zod9.z.enum(["packages", "model", "verifying"]);
 var MlInstallProgress = import_zod9.z.strictObject({
@@ -856,18 +857,7 @@ var MlProgress = import_zod9.z.strictObject({
   /** The segment just transcribed (transcribe only): text appears as it is recognised. */
   segment: TranscriptSegment.nullable()
 });
-var TextModelId = import_zod9.z.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 = import_zod9.z.strictObject({
   id: TextModelId,
   /** Download size of the weights. */
@@ -875,7 +865,14 @@ var TextModelInfo = import_zod9.z.strictObject({
   installed: import_zod9.z.boolean()
 });
 var LlmServerState = import_zod9.z.enum(["stopped", "starting", "running", "failed"]);
-var LlmRunning = import_zod9.z.strictObject({ id: TextModelId, port: import_zod9.z.int().min(1024).max(65535) });
+var LlmRunning = import_zod9.z.strictObject({
+  id: TextModelId,
+  port: import_zod9.z.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: import_zod9.z.int().min(512).max(131072).optional(),
+  gpu_layers: import_zod9.z.int().min(-1).max(999).optional(),
+  threads: import_zod9.z.int().min(1).max(512).optional()
+});
 var ModelsLlmStatusParams = import_zod9.z.strictObject({});
 var ModelsLlmStatusResult = import_zod9.z.strictObject({
   state: LlmServerState,
@@ -893,12 +890,57 @@ var ModelsLlmInstallParams = import_zod9.z.strictObject({ model: TextModelId });
 var ModelsLlmInstallResult = import_zod9.z.strictObject({ model: TextModelId, elapsed_ms: nonNegInt });
 var ModelsLlmStartParams = import_zod9.z.strictObject({
   model: TextModelId,
-  /** Tokens of context; the service picks a size that fits this computer when omitted. */
-  context_length: import_zod9.z.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: import_zod9.z.int().min(512).max(131072).optional(),
+  /** (1.48) Layers offloaded to the GPU; -1 = all. From the knob desktop.models/gpu_layers. */
+  gpu_layers: import_zod9.z.int().min(-1).max(999).optional(),
+  /** (1.48) CPU threads, computed by the app from the knobs. */
+  threads: import_zod9.z.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: import_zod9.z.int().min(10).max(3600).optional()
 });
 var ModelsLlmStartResult = import_zod9.z.strictObject({ running: LlmRunning, elapsed_ms: nonNegInt });
 var ModelsLlmStopParams = import_zod9.z.strictObject({});
 var ModelsLlmStopResult = import_zod9.z.strictObject({ stopped: import_zod9.z.boolean() });
+var Sha256Hex = import_zod9.z.string().regex(/^[0-9a-f]{64}$/);
+var CatalogFile = import_zod9.z.strictObject({
+  /** Path inside the model's folder (the artifact's path_in_repo). */
+  path: import_zod9.z.string().min(1).max(512),
+  /** The pinned https URL it is downloaded from. */
+  url: import_zod9.z.string().url().max(2048),
+  size_bytes: nonNegInt,
+  sha256: Sha256Hex,
+  /** weights, weights_shard, mmproj, config, tokenizer, … (ai.model_artifact.role). */
+  role: import_zod9.z.string().max(32)
+});
+var CatalogModelEntry = import_zod9.z.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: import_zod9.z.string().min(1).max(64),
+  /** The upstream revision (or the variant id when there is none): a change re-downloads. */
+  revision: import_zod9.z.string().min(1).max(64),
+  multilingual: import_zod9.z.boolean(),
+  /** Exactly the files this computer fetches for this model: never more, never fewer. */
+  files: import_zod9.z.array(CatalogFile).min(1).max(256)
+});
+var LlamaServerBuild = import_zod9.z.strictObject({
+  version: import_zod9.z.string().min(1).max(32),
+  url: import_zod9.z.string().url().max(2048),
+  size_bytes: nonNegInt,
+  sha256: Sha256Hex
+});
+var MlCatalogSetParams = import_zod9.z.strictObject({
+  text: import_zod9.z.array(CatalogModelEntry).max(512),
+  transcription: import_zod9.z.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: import_zod9.z.string().max(40)
+});
+var MlCatalogSetResult = import_zod9.z.strictObject({ text: nonNegInt, transcription: nonNegInt });
 var DownloadKind = import_zod9.z.enum(["speech_model", "text_model", "runtime"]);
 var DownloadRow = import_zod9.z.strictObject({
   /** The job's key in the service: "model:base", "llm:smollm2-135m", "llm-server", "runtime". */