@aliou/pi-neuralwatt 0.16.1 → 0.16.2

package/extensions/provider/api/openai-completions.ts

@@ -1,3 +1,4 @@
+import type { StreamOptions } from "@earendil-works/pi-ai";
 import { stream, streamSimple } from "@earendil-works/pi-ai/compat";
 import {
   NEURALWATT_BASE_URL,
@@ -8,9 +9,56 @@ import type { NeuralwattModel } from "../models/catalog";
 import type { AnyStreamSimple } from "../stream-simple";
 import type { NeuralwattApiHandler } from "./types";
 
+type OpenAiCompletionsBody = {
+  messages?: Array<Record<string, unknown>>;
+  [key: string]: unknown;
+};
+
+/**
+ * Neuralwatt streams chain-of-thought in the `reasoning` field, and pi-ai
+ * replays prior thinking under the field name it recorded from the stream —
+ * also `reasoning`. The served chat templates only render `reasoning_content`
+ * (verified against the Kimi K3 template), so replayed thinking never reaches
+ * the model. Rename the replayed field on the wire. This is a move, not a
+ * copy: sending an empty `reasoning_content` next to a populated `reasoning`
+ * makes the gateway prefer the empty field and silently drops the replay.
+ */
+function makeReasoningReplayInjector(
+  upstream?: StreamOptions["onPayload"],
+): NonNullable<StreamOptions["onPayload"]> {
+  return async (payload, model) => {
+    const next = await upstream?.(payload, model);
+    const body = (next !== undefined ? next : payload) as OpenAiCompletionsBody;
+    const messages = body.messages;
+    if (!Array.isArray(messages)) return body;
+
+    return {
+      ...body,
+      messages: messages.map((message) => {
+        if (message?.role !== "assistant" || !("reasoning" in message)) {
+          return message;
+        }
+        const { reasoning, ...rest } = message;
+        // A pre-set non-empty reasoning_content wins; drop the duplicate.
+        return typeof rest.reasoning_content === "string" &&
+          rest.reasoning_content.length > 0
+          ? rest
+          : { ...rest, reasoning_content: reasoning };
+      }),
+    };
+  };
+}
+
 export function createOpenAiCompletionsApi(options?: {
   streamSimple?: AnyStreamSimple;
 }): NeuralwattApiHandler {
+  const withReasoningReplay = (options?: {
+    onPayload?: StreamOptions["onPayload"];
+  }) => ({
+    ...options,
+    onPayload: makeReasoningReplayInjector(options?.onPayload),
+  });
+
   return {
     stampModels: (models: NeuralwattModel[]) =>
       models.map((model) => {
@@ -24,7 +72,12 @@ export function createOpenAiCompletionsApi(options?: {
         };
       }),
     stream: (model, context, streamOptions) =>
-      stream(model, context, streamOptions as never),
-    streamSimple: options?.streamSimple ?? streamSimple,
+      stream(model, context, withReasoningReplay(streamOptions) as never),
+    streamSimple: (model, context, simpleOptions) =>
+      (options?.streamSimple ?? streamSimple)(
+        model,
+        context,
+        withReasoningReplay(simpleOptions) as never,
+      ),
   };
 }

package/extensions/provider/models/public-models.ts

@@ -308,21 +308,21 @@ const FAMILIES: [NeuralwattModelFamily, NeuralwattVariantSpec[]][] = [
       {
         id: "qwen3.6-35b",
         name: "Qwen3.6 35B",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: true,
       },
       {
         id: "qwen3.6-35b-fast",
         name: "Qwen3.6 35B Fast",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: false,
       },
       {
         id: "qwen3.6-35b-flex",
         name: "Qwen3.6 35B (flex)",
-        contextWindow: 131056,
+        contextWindow: 262128,
         maxOutputTokens: null,
         reasoning: true,
         costMultiplier: 0.65,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.16.1",
+  "version": "0.16.2",
   "license": "MIT",
   "type": "module",
   "private": false,