@qvac/sdk 0.12.1 → 0.13.0

package/dist/server/rpc/handlers/rag.js

@@ -111,13 +111,13 @@ function omitOnProgress(obj) {
  * "still mine?" guard so an older op's deferred cleanup cannot stomp
  * a newer op's mapping.
  */
-function beginRagContext(workspace, requestId) {
+async function beginRagContext(workspace, requestId) {
     const registry = getRequestRegistry();
     const prev = getActiveRagRequest(workspace);
     if (prev !== undefined && prev !== requestId) {
         registry.cancel({ requestId: prev, reason: "rag-workspace-preempt" });
     }
-    const ctx = registry.begin({
+    const ctx = await registry.begin({
         requestId,
         kind: "rag",
     });
@@ -146,7 +146,7 @@ async function handleRagInternal(request, onProgress) {
             try {
                 const workspace = request.workspace ?? DEFAULT_WORKSPACE;
                 const requestId = request.requestId ?? generateServerRequestId();
-                const ctx = __addDisposableResource(env_1, beginRagContext(workspace, requestId), true);
+                const ctx = __addDisposableResource(env_1, await beginRagContext(workspace, requestId), true);
                 const log = withRequestContext(getServerLogger(), ctx);
                 log.debug("ingest start");
                 const handlerOptions = createHandlerOptions("ingest", workspace, ctx.signal, onProgress);
@@ -175,7 +175,7 @@ async function handleRagInternal(request, onProgress) {
             try {
                 const workspace = request.workspace ?? DEFAULT_WORKSPACE;
                 const requestId = request.requestId ?? generateServerRequestId();
-                const ctx = __addDisposableResource(env_2, beginRagContext(workspace, requestId), true);
+                const ctx = __addDisposableResource(env_2, await beginRagContext(workspace, requestId), true);
                 const log = withRequestContext(getServerLogger(), ctx);
                 log.debug("saveEmbeddings start");
                 const handlerOptions = createHandlerOptions("saveEmbeddings", workspace, ctx.signal, onProgress);
@@ -220,7 +220,7 @@ async function handleRagInternal(request, onProgress) {
             try {
                 const workspace = request.workspace ?? DEFAULT_WORKSPACE;
                 const requestId = request.requestId ?? generateServerRequestId();
-                const ctx = __addDisposableResource(env_3, beginRagContext(workspace, requestId), true);
+                const ctx = __addDisposableResource(env_3, await beginRagContext(workspace, requestId), true);
                 const log = withRequestContext(getServerLogger(), ctx);
                 log.debug("reindex start");
                 const handlerOptions = createHandlerOptions("reindex", workspace, ctx.signal, onProgress);

package/dist/server/utils/checksum.js

@@ -1,10 +1,11 @@
 import fs from "bare-fs";
 import crypto from "bare-crypto";
 import { nowMs } from "../../profiling/index.js";
+import {} from "bare-buffer";
 export async function calculateFileChecksum(filePath) {
     return new Promise((resolve, reject) => {
         try {
-            const hash = crypto.createHash("sha256");
+            const hash = crypto.createHash("sha-256");
             const stream = fs.createReadStream(filePath);
             stream.on("data", (chunk) => {
                 hash.update(chunk);

package/dist/server/utils/formatting.js

@@ -1,10 +1,11 @@
 import crypto from "bare-crypto";
+import { Buffer } from "bare-buffer";
 /**
  * Generate a short hash (16 characters) from any input string
  */
 export function generateShortHash(input) {
     const hash = crypto
-        .createHash("sha256")
+        .createHash("sha-256")
         .update(Buffer.from(input, "utf8"))
         .digest("hex");
     return hash.substring(0, 16);

package/dist/server/utils/multi-gpu-mobile.d.ts

@@ -0,0 +1,5 @@
+export declare const MULTI_GPU_KEYS: readonly ["main-gpu", "split-mode", "tensor-split"];
+type MultiGpuKey = (typeof MULTI_GPU_KEYS)[number];
+export declare function stripMultiGpuKeys(config: Record<string, unknown>): readonly MultiGpuKey[];
+export {};
+//# sourceMappingURL=multi-gpu-mobile.d.ts.map

package/dist/server/utils/multi-gpu-mobile.js

@@ -0,0 +1,10 @@
+export const MULTI_GPU_KEYS = ["main-gpu", "split-mode", "tensor-split"];
+// Deletes any multi-GPU keys from `config` IN PLACE and returns the removed
+// keys. Mutation is intentional: callers strip before the config reaches the
+// model constructor. gpu_layers (single-GPU layer offload) is intentionally
+// absent from MULTI_GPU_KEYS.
+export function stripMultiGpuKeys(config) {
+    const stripped = MULTI_GPU_KEYS.filter((k) => k in config);
+    stripped.forEach((k) => delete config[k]);
+    return stripped;
+}

package/dist/server/worker.js

@@ -4,7 +4,7 @@
 import { initializeWorkerCore, ensureRPCSetup } from "./worker-core.js";
 import { registerPlugins } from "./plugins/index.js";
 import { getServerLogger } from "../logging/index.js";
-import { llmPlugin, embeddingsPlugin, whisperPlugin, parakeetPlugin, nmtPlugin, ttsPlugin, ocrPlugin, diffusionPlugin, vlaPlugin, classificationPlugin, } from "./bare/plugins/index.js";
+import { llmPlugin, embeddingsPlugin, whisperPlugin, bciPlugin, parakeetPlugin, nmtPlugin, ttsPlugin, ocrPlugin, diffusionPlugin, vlaPlugin, classificationPlugin, } from "./bare/plugins/index.js";
 const { hasRPCConfig } = initializeWorkerCore();
 const logger = getServerLogger();
 logger.info("🐻 Hello from Bare");
@@ -12,6 +12,7 @@ registerPlugins([
     llmPlugin,
     embeddingsPlugin,
     whisperPlugin,
+    bciPlugin,
     parakeetPlugin,
     nmtPlugin,
     ttsPlugin,

package/dist/utils/aggregate-events.d.ts

@@ -1,4 +1,4 @@
-import type { CompletionError, CompletionEvent, CompletionFinal, CompletionStats, ToolCall } from "../schemas/index";
+import type { CompletionError, CompletionEvent, CompletionFinal, CompletionStats, StopReason, ToolCall } from "../schemas/index";
 import { type ToolHandlerMap } from "./tool-helpers";
 export type AggregatedEvents = {
     contentText: string;
@@ -7,6 +7,7 @@ export type AggregatedEvents = {
     toolCalls: ToolCall[];
     rawFullText: string | undefined;
     error: CompletionError | undefined;
+    stopReason: StopReason | undefined;
     /**
      * True when the terminal `completionDone` carried
      * `stopReason: "cancelled"`. The client wrapper rejects the

package/dist/utils/aggregate-events.js

@@ -6,6 +6,7 @@ export function aggregateEvents(events) {
     let stats;
     let rawFullText;
     let error;
+    let stopReason;
     let cancelled = false;
     const toolCalls = [];
     for (const event of events) {
@@ -31,8 +32,13 @@ export function aggregateEvents(events) {
             if (event.stopReason === "error" && "error" in event) {
                 error = event.error;
             }
-            else if (event.stopReason === "cancelled") {
-                cancelled = true;
+            else {
+                if (event.stopReason !== undefined) {
+                    stopReason = event.stopReason;
+                }
+                if (event.stopReason === "cancelled") {
+                    cancelled = true;
+                }
             }
         }
     }
@@ -43,11 +49,12 @@ export function aggregateEvents(events) {
         toolCalls,
         rawFullText,
         error,
+        stopReason,
         cancelled,
     };
 }
 export function buildFinalFromEvents(events, handlers) {
-    const { contentText, thinkingText, stats, toolCalls, rawFullText, error, cancelled, } = aggregateEvents(events);
+    const { contentText, thinkingText, stats, toolCalls, rawFullText, error, stopReason, cancelled, } = aggregateEvents(events);
     const attachedToolCalls = attachHandlersToToolCalls(toolCalls, handlers);
     const fullText = rawFullText ?? contentText;
     const cacheableAssistantContent = attachedToolCalls.length === 0
@@ -62,6 +69,7 @@ export function buildFinalFromEvents(events, handlers) {
         ...(cacheableAssistantContent !== undefined && {
             cacheableAssistantContent,
         }),
+        ...(stopReason !== undefined && { stopReason }),
     };
     return { final, error, cancelled };
 }

package/dist/utils/errors-server.d.ts

@@ -128,10 +128,12 @@ export declare class RequestNotFoundError extends QvacErrorBase {
     toErrorResponseFields(): Record<string, unknown>;
 }
 /**
- * Thrown by `RequestRegistry.begin(...)` when a registered concurrency
- * policy rejects the request (e.g. `oneAtATimePerModel` for the
- * `completion` kind). Distinct from `RequestIdConflictError`, which
- * only fires on UUID collisions.
+ * Thrown by `await RequestRegistry.begin(...)` when a registered concurrency
+ * policy refuses the request. Under the default queue policy a same-model
+ * request waits FIFO rather than rejecting, so this fires on the bounded-queue
+ * cases: an explicit `onOverflow: "reject"`, the per-model queue-depth cap, or
+ * a `queueTimeoutMs` elapsing while waiting for a slot. Distinct from
+ * `RequestIdConflictError`, which only fires on UUID collisions.
  */
 export declare class RequestRejectedByPolicyError extends QvacErrorBase {
     readonly requestId: string;

package/dist/utils/errors-server.js

@@ -191,10 +191,12 @@ export class RequestNotFoundError extends QvacErrorBase {
     }
 }
 /**
- * Thrown by `RequestRegistry.begin(...)` when a registered concurrency
- * policy rejects the request (e.g. `oneAtATimePerModel` for the
- * `completion` kind). Distinct from `RequestIdConflictError`, which
- * only fires on UUID collisions.
+ * Thrown by `await RequestRegistry.begin(...)` when a registered concurrency
+ * policy refuses the request. Under the default queue policy a same-model
+ * request waits FIFO rather than rejecting, so this fires on the bounded-queue
+ * cases: an explicit `onOverflow: "reject"`, the per-model queue-depth cap, or
+ * a `queueTimeoutMs` elapsing while waiting for a slot. Distinct from
+ * `RequestIdConflictError`, which only fires on UUID collisions.
  */
 export class RequestRejectedByPolicyError extends QvacErrorBase {
     requestId;