@aryacode/dsh-local-llm 0.8.5 → 0.8.7

package/lib/index.d.ts

@@ -1610,12 +1610,18 @@ 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 share 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.
+ * @param divisor - the share of the window the answer may use: 4 for the Strata
+ *   engines; 3 for NInfer, whose 100K window needs 34,133 for a whole file (a
+ *   25,600 quarter cut a measured 25,690-token page write by 90 tokens).
  */
-declare function outputDefault(context: number, ceiling?: number): number;
+declare function outputDefault(context: number, ceiling?: number, divisor?: number): number;
 /** Where the Strata engine's own OpenAI-compatible surface answers, and how to authorize it. */
 interface StrataModelSource {
   /** Engine base URL, without a trailing slash, e.g. `http://127.0.0.1:8080`. */

package/lib/index.js

@@ -2605,13 +2605,19 @@ 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 share 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.
+* @param divisor - the share of the window the answer may use: 4 for the Strata
+*   engines; 3 for NInfer, whose 100K window needs 34,133 for a whole file (a
+*   25,600 quarter cut a measured 25,690-token page write by 90 tokens).
 */
-function outputDefault(context, ceiling = 8192) {
-	return Math.max(1024, Math.min(ceiling, 8192, Math.floor(context / 4)));
+function outputDefault(context, ceiling = Number.POSITIVE_INFINITY, divisor = 4) {
+	return Math.max(1024, Math.min(ceiling, Math.floor(context / divisor)));
 }
 /** Thinking levels a chat-template engine can be asked for through its template kwargs. */
 const strataReasoning = {
@@ -2660,9 +2666,9 @@ function libraryAdapter(controller, ensure, gate, checkIdle, attachments, strata
 	* The label the other Arya engines show: `qwen3.8-27b` becomes `Arya - Qwen3.8 27B`.
 	*
 	* The model list is what a person reads in the picker, so a bare artifact id is
-	* a poor label next to the existing `Arya - …` entries. Each segment is
-	* capitalised, and a lowercase letter that follows a trailing digit — a size or
-	* quantisation suffix such as `27b` — is uppercased the way these artifacts are
+	* a poor label next to the existing `Arya - …` entries. Each segment is
+	* capitalised, and a lowercase letter that follows a trailing digit — a size or
+	* quantisation suffix such as `27b` — is uppercased the way these artifacts are
 	* written on the tin.
 	*/
 	const aryaLabel = (name) => "Arya - " + name.split(/[-_]/u).filter((part) => part !== "").map((part) => {
@@ -2726,7 +2732,7 @@ function libraryAdapter(controller, ensure, gate, checkIdle, attachments, strata
 			if (typeof state.modelName === "string" && state.modelName !== "") {
 				const model = {
 					...info("ninfer:" + state.modelName, aryaLabel(state.modelName), Number(state.context) || 8192, state.vision === true, "ninfer", state.modelName),
-					defaultMaxTokens: outputDefault(Number(state.context) || 8192, ninfer.outputTokens()),
+					defaultMaxTokens: outputDefault(Number(state.context) || 8192, ninfer.outputTokens(), 3),
 					reasoning: ninferReasoning
 				};
 				models.push(model);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aryacode/dsh-local-llm",
-  "version": "0.8.5",
+  "version": "0.8.7",
   "description": "AryaStrata engine controls with the local NInfer Qwen3.8 dense model for AryaAI",
   "type": "module",
   "main": "lib/index.js",