@aryacode/dsh-local-llm 0.8.4 → 0.8.6

package/lib/index.d.ts

@@ -1610,10 +1610,13 @@ declare const LIBRARY_PROVIDER = "local-library";
 /**
  * The per-request output default this adapter publishes for one model.
  *
- * An engine's *maximum* response allowance is a ceiling it enforces when a
- * request asks for it, not something every request should ask for: this value
- * is materialized into requests that omit an output budget, so it stays small
- * (8K, and never more than a quarter of the model's own context window).
+ * It scales with the window the engine actually serves: a quarter of it, floored
+ * at 1,024 tokens and bounded by the engine's own allowance when one is known.
+ * A fixed 8K used to cap every model here, which silently truncated long answers
+ * and made a 100K- or 256K-token engine behave like an 8K one.
+ *
+ * @param context - the model's served context window in tokens.
+ * @param ceiling - the engine's maximum response allowance, when it declares one.
  */
 declare function outputDefault(context: number, ceiling?: number): number;
 /** Where the Strata engine's own OpenAI-compatible surface answers, and how to authorize it. */
@@ -1731,11 +1734,20 @@ declare class NInferController {
       distribution: string;
       binary: string;
       libraryPath: string;
+      terminateOnStop?: boolean;
     };
   });
   /** Resolve saved tuning or validated deployment defaults. */
   settings(): NInferSettings;
   private callWsl;
+  /**
+   * Shut the distro down after the engine stops, so the WSL VM gives its memory back.
+   *
+   * `wsl --terminate` is the only lever that returns the distro's RAM: stopping
+   * the binary leaves systemd and the utility VM resident. A distro that is
+   * already stopped makes the call a no-op, so a failure here is not an error.
+   */
+  private terminateDistro;
   /** A container this Host may reclaim; a missing Docker or an unrelated one is no takeover. */
   private external;
   private wslStatus;
@@ -1792,6 +1804,12 @@ declare class NInferController {
     ok: boolean;
     detail: string;
   }>;
+  /**
+   * Stop the engine, and by default hand the WSL distro's memory back.
+   *
+   * @param terminate - terminate the WSL distro after the engine stops; a start
+   *   or restart passes `false` because it is about to use that distro again.
+   */
   private stop;
   /** Stop the external container that is provably this installation's engine. */
   private takeOver;

package/lib/index.js

@@ -2605,13 +2605,16 @@ const LIBRARY_PROVIDER = "local-library";
 /**
 * The per-request output default this adapter publishes for one model.
 *
-* An engine's *maximum* response allowance is a ceiling it enforces when a
-* request asks for it, not something every request should ask for: this value
-* is materialized into requests that omit an output budget, so it stays small
-* (8K, and never more than a quarter of the model's own context window).
+* It scales with the window the engine actually serves: a quarter of it, floored
+* at 1,024 tokens and bounded by the engine's own allowance when one is known.
+* A fixed 8K used to cap every model here, which silently truncated long answers
+* and made a 100K- or 256K-token engine behave like an 8K one.
+*
+* @param context - the model's served context window in tokens.
+* @param ceiling - the engine's maximum response allowance, when it declares one.
 */
-function outputDefault(context, ceiling = 8192) {
-	return Math.max(1024, Math.min(ceiling, 8192, Math.floor(context / 4)));
+function outputDefault(context, ceiling = Number.POSITIVE_INFINITY) {
+	return Math.max(1024, Math.min(ceiling, Math.floor(context / 4)));
 }
 /** Thinking levels a chat-template engine can be asked for through its template kwargs. */
 const strataReasoning = {
@@ -3100,6 +3103,23 @@ var NInferController = class {
 			}));
 		});
 	}
+	/**
+	* Shut the distro down after the engine stops, so the WSL VM gives its memory back.
+	*
+	* `wsl --terminate` is the only lever that returns the distro's RAM: stopping
+	* the binary leaves systemd and the utility VM resident. A distro that is
+	* already stopped makes the call a no-op, so a failure here is not an error.
+	*/
+	async terminateDistro() {
+		const distribution = this.config.wsl?.distribution;
+		if (distribution === void 0) return;
+		try {
+			await execute$1("wsl.exe", ["--terminate", distribution], {
+				windowsHide: true,
+				timeout: 3e4
+			});
+		} catch (error) {}
+	}
 	/** A container this Host may reclaim; a missing Docker or an unrelated one is no takeover. */
 	async external() {
 		try {
@@ -3265,9 +3285,16 @@ var NInferController = class {
 			this.changing = false;
 		}
 	}
-	async stop() {
+	/**
+	* Stop the engine, and by default hand the WSL distro's memory back.
+	*
+	* @param terminate - terminate the WSL distro after the engine stops; a start
+	*   or restart passes `false` because it is about to use that distro again.
+	*/
+	async stop(terminate = true) {
 		if (this.config.wsl) {
 			await this.callWsl("stop");
+			if (terminate && this.config.wsl.terminateOnStop !== false) await this.terminateDistro();
 			return;
 		}
 		const container = await this.container();
@@ -3313,17 +3340,19 @@ var NInferController = class {
 				await this.takeOver();
 				takenOver = "Stopped the external NInfer container. ";
 			}
-			if (action === "stop" || action === "restart") await this.stop();
-			if (action === "stop") return {
-				ok: true,
-				detail: takenOver + "NInfer stopped."
-			};
+			if (action === "stop") {
+				await this.stop();
+				return {
+					ok: true,
+					detail: takenOver + "NInfer stopped."
+				};
+			}
 			if (action === "start" && ["running", "starting"].includes(status.phase)) return {
 				ok: true,
 				detail: "NInfer is already started."
 			};
 			if (!this.complete()) throw Error("Wait for the model checksum verification.");
-			await this.stop();
+			await this.stop(false);
 			if (this.config.wsl) {
 				const started = await this.callWsl("start");
 				return {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aryacode/dsh-local-llm",
-  "version": "0.8.4",
+  "version": "0.8.6",
   "description": "AryaStrata engine controls with the local NInfer Qwen3.8 dense model for AryaAI",
   "type": "module",
   "main": "lib/index.js",