@qvac/sdk 0.9.1 → 0.10.0

package/dist/schemas/get-model-info.js

@@ -1,40 +1,95 @@
 import { z } from "zod";
 export const getModelInfoParamsSchema = z.object({
-    name: z.string(),
+    name: z
+        .string()
+        .describe("The model's registry name (as found in the SDK's built-in catalog)."),
 });
 export const getModelInfoRequestSchema = getModelInfoParamsSchema.extend({
     type: z.literal("getModelInfo"),
 });
 export const loadedInstanceSchema = z.object({
-    registryId: z.string(),
-    loadedAt: z.coerce.date(),
-    config: z.unknown().optional(),
+    registryId: z
+        .string()
+        .describe("Identifier of the registered loaded instance."),
+    loadedAt: z
+        .coerce.date()
+        .describe("Timestamp when the instance was loaded."),
+    config: z
+        .unknown()
+        .optional()
+        .describe("Opaque model-specific configuration the instance was loaded with."),
 });
 export const cacheFileInfoSchema = z.object({
-    filename: z.string(),
-    path: z.string(),
-    expectedSize: z.number(),
-    sha256Checksum: z.string(),
-    isCached: z.boolean(),
-    actualSize: z.number().optional(),
-    cachedAt: z.coerce.date().optional(),
+    filename: z.string().describe("Name of the cached file on disk."),
+    path: z.string().describe("Absolute path to the cached file."),
+    expectedSize: z
+        .number()
+        .describe("Expected file size in bytes from the model metadata."),
+    sha256Checksum: z
+        .string()
+        .describe("Expected SHA-256 checksum of the file."),
+    isCached: z
+        .boolean()
+        .describe("Whether the file is currently present in the cache."),
+    actualSize: z
+        .number()
+        .optional()
+        .describe("Actual size of the cached file on disk, in bytes."),
+    cachedAt: z
+        .coerce.date()
+        .optional()
+        .describe("Timestamp when the file was last cached."),
 });
 export const modelInfoSchema = z.object({
-    name: z.string(),
-    modelId: z.string(),
+    name: z.string().describe("Catalog name of the model."),
+    modelId: z
+        .string()
+        .describe("Unique identifier used to reference the model in SDK calls."),
     // Registry-based fields
-    registryPath: z.string().optional(),
-    registrySource: z.string().optional(),
-    blobCoreKey: z.string().optional(),
-    blobBlockOffset: z.number().optional(),
-    blobBlockLength: z.number().optional(),
-    blobByteOffset: z.number().optional(),
-    engine: z.string().optional(),
-    quantization: z.string().optional(),
-    params: z.string().optional(),
-    expectedSize: z.number(),
-    sha256Checksum: z.string(),
-    addon: z.enum([
+    registryPath: z
+        .string()
+        .optional()
+        .describe("Registry-relative path (set for registry-backed models)."),
+    registrySource: z
+        .string()
+        .optional()
+        .describe("Registry source identifier, e.g. `huggingface`."),
+    blobCoreKey: z
+        .string()
+        .optional()
+        .describe("Hyperdrive blob core key for the model file."),
+    blobBlockOffset: z
+        .number()
+        .optional()
+        .describe("Starting block offset of the model file in the blob."),
+    blobBlockLength: z
+        .number()
+        .optional()
+        .describe("Number of blocks occupied by the model file in the blob."),
+    blobByteOffset: z
+        .number()
+        .optional()
+        .describe("Starting byte offset of the model file within its block."),
+    engine: z
+        .string()
+        .optional()
+        .describe("Inference engine identifier, e.g. `llamacpp-completion`."),
+    quantization: z
+        .string()
+        .optional()
+        .describe("Quantization identifier, e.g. `Q4_K_M`."),
+    params: z
+        .string()
+        .optional()
+        .describe("Parameter-count label for the model, e.g. `7B`."),
+    expectedSize: z
+        .number()
+        .describe("Expected total size of the model file in bytes."),
+    sha256Checksum: z
+        .string()
+        .describe("Expected SHA-256 checksum of the model file."),
+    addon: z
+        .enum([
         "llm",
         "whisper",
         "parakeet",
@@ -45,13 +100,29 @@ export const modelInfoSchema = z.object({
         "ocr",
         "diffusion",
         "other",
-    ]),
-    isCached: z.boolean(),
-    isLoaded: z.boolean(),
-    cacheFiles: z.array(cacheFileInfoSchema),
-    actualSize: z.number().optional(),
-    cachedAt: z.coerce.date().optional(),
-    loadedInstances: z.array(loadedInstanceSchema).optional(),
+    ])
+        .describe("Inference addon / capability category this model belongs to."),
+    isCached: z
+        .boolean()
+        .describe("Whether the model file is present in the local cache."),
+    isLoaded: z
+        .boolean()
+        .describe("Whether the model is currently loaded into memory."),
+    cacheFiles: z
+        .array(cacheFileInfoSchema)
+        .describe("Individual cache file entries that make up this model."),
+    actualSize: z
+        .number()
+        .optional()
+        .describe("Actual total size of the cached model on disk, in bytes."),
+    cachedAt: z
+        .coerce.date()
+        .optional()
+        .describe("Timestamp when the model was last cached."),
+    loadedInstances: z
+        .array(loadedInstanceSchema)
+        .optional()
+        .describe("Loaded instances associated with this model (one per live load)."),
 });
 export const getModelInfoResponseSchema = z.object({
     type: z.literal("getModelInfo"),

package/dist/schemas/heartbeat.d.ts

@@ -2,7 +2,6 @@ import { z } from "zod";
 export declare const heartbeatRequestSchema: z.ZodObject<{
     type: z.ZodLiteral<"heartbeat">;
     delegate: z.ZodOptional<z.ZodObject<{
-        topic: z.ZodString;
         providerPublicKey: z.ZodString;
         timeout: z.ZodOptional<z.ZodNumber>;
         healthCheckTimeout: z.ZodOptional<z.ZodNumber>;

package/dist/schemas/index.d.ts

@@ -1,7 +1,8 @@
 export * from "./archive";
 export * from "./cancel";
 export * from "./completion-stream";
-export * from "./tools";
+export * from "./completion-event";
+export { toolSchema, toolCallSchema, toolCallErrorSchema, TOOLS_MODE, type Tool, type ToolCall, type ToolCallError, type ToolCallWithCall, type ToolsMode, } from "./tools";
 export * from "./delegate";
 export * from "./delete-cache";
 export * from "./download-asset";
@@ -28,6 +29,7 @@ export * from "./sdcpp-config";
 export * from "./shard";
 export * from "./suspend";
 export * from "./resume";
+export * from "./state";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry";
@@ -35,6 +37,7 @@ export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, device
 export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, type ProfilerMode, type ServerBreakdown, type DelegationBreakdown, type OperationEvent, type ProfilingRequestMeta, type ProfilingResponseMeta, type PerCallProfiling, } from "./profiling";
 export { runtimeContextSchema, type RuntimeContext } from "./runtime-context";
 export * from "./get-model-info";
+export * from "./get-loaded-model-info";
 export * from "./model-src-utils";
 export * from "./json-schema";
 export { type McpClient, type McpClientInput } from "./mcp-adapter";

package/dist/schemas/index.js

@@ -2,7 +2,8 @@
 export * from "./archive.js";
 export * from "./cancel.js";
 export * from "./completion-stream.js";
-export * from "./tools.js";
+export * from "./completion-event.js";
+export { toolSchema, toolCallSchema, toolCallErrorSchema, TOOLS_MODE, } from "./tools.js";
 export * from "./delegate.js";
 export * from "./delete-cache.js";
 export * from "./download-asset.js";
@@ -29,6 +30,7 @@ export * from "./sdcpp-config.js";
 export * from "./shard.js";
 export * from "./suspend.js";
 export * from "./resume.js";
+export * from "./state.js";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client.js";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server.js";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry.js";
@@ -36,6 +38,7 @@ export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, device
 export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, } from "./profiling.js";
 export { runtimeContextSchema } from "./runtime-context.js";
 export * from "./get-model-info.js";
+export * from "./get-loaded-model-info.js";
 export * from "./model-src-utils.js";
 export * from "./json-schema.js";
 export {} from "./mcp-adapter.js";

package/dist/schemas/llamacpp-config.d.ts

@@ -24,6 +24,10 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    toolsMode: z.ZodOptional<z.ZodEnum<{
+        static: "static";
+        dynamic: "dynamic";
+    }>>;
     "cache-type-k": z.ZodOptional<z.ZodString>;
     "cache-type-v": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
@@ -39,6 +43,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
         expectedSize: z.ZodOptional<z.ZodNumber>;
         sha256Checksum: z.ZodOptional<z.ZodString>;
         addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            embeddings: "embeddings";
             "llamacpp-completion": "llamacpp-completion";
             "whispercpp-transcription": "whispercpp-transcription";
             "llamacpp-embedding": "llamacpp-embedding";
@@ -49,7 +54,6 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
-            embeddings: "embeddings";
             nmt: "nmt";
             parakeet: "parakeet";
             tts: "tts";
@@ -84,6 +88,10 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    toolsMode: z.ZodOptional<z.ZodEnum<{
+        static: "static";
+        dynamic: "dynamic";
+    }>>;
     "cache-type-k": z.ZodOptional<z.ZodString>;
     "cache-type-v": z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
@@ -99,6 +107,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         expectedSize: z.ZodOptional<z.ZodNumber>;
         sha256Checksum: z.ZodOptional<z.ZodString>;
         addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
+            embeddings: "embeddings";
             "llamacpp-completion": "llamacpp-completion";
             "whispercpp-transcription": "whispercpp-transcription";
             "llamacpp-embedding": "llamacpp-embedding";
@@ -109,7 +118,6 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
-            embeddings: "embeddings";
             nmt: "nmt";
             parakeet: "parakeet";
             tts: "tts";
@@ -136,6 +144,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    toolsMode?: "static" | "dynamic" | undefined;
     "cache-type-k"?: string | undefined;
     "cache-type-v"?: string | undefined;
     openclCacheDir?: string | undefined;
@@ -150,7 +159,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
+        addon?: "embeddings" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
 }, {
     ctx_size?: number | undefined;
@@ -171,6 +180,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    toolsMode?: "static" | "dynamic" | undefined;
     "cache-type-k"?: string | undefined;
     "cache-type-v"?: string | undefined;
     openclCacheDir?: string | undefined;
@@ -185,7 +195,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
+        addon?: "embeddings" | "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;

package/dist/schemas/llamacpp-config.js

@@ -1,5 +1,6 @@
 import { z } from "zod";
 import { modelSrcInputSchema } from "./model-src-utils.js";
+import { TOOLS_MODE } from "./tools.js";
 export const VERBOSITY = {
     ERROR: 0,
     WARN: 1,
@@ -38,6 +39,10 @@ export const llmConfigBaseSchema = z.object({
     stop_sequences: z.array(z.string()).optional(),
     n_discarded: z.number().optional(),
     tools: z.boolean().optional(),
+    toolsMode: z
+        .enum([TOOLS_MODE.static, TOOLS_MODE.dynamic])
+        .describe('Controls tool placement in the prompt. "static" (default) prepends the tool set once and reuses it across the session. "dynamic" anchors tools after the last user message and trims them from the kv-cache after the chain resolves so each user prompt can carry its own tools.')
+        .optional(),
     "cache-type-k": z.string().optional(),
     "cache-type-v": z.string().optional(),
     /**