@aryacode/dsh-local-llm 0.8.5 → 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. */

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 = {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aryacode/dsh-local-llm",
-  "version": "0.8.5",
+  "version": "0.8.6",
   "description": "AryaStrata engine controls with the local NInfer Qwen3.8 dense model for AryaAI",
   "type": "module",
   "main": "lib/index.js",