@aliou/pi-neuralwatt 0.16.0 → 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/build.ts

@@ -161,10 +161,6 @@ export function buildNeuralwattModel(
     supportsDeveloperRole: false,
     maxTokensField: "max_tokens",
   };
-  if (variant.reasoning) {
-    compat.requiresReasoningContentOnAssistantMessages = true;
-  }
-
   const multiplier = variant.costMultiplier ?? 1;
   const scale = (value: number): number =>
     multiplier === 1 ? value : Number((value * multiplier).toFixed(6));

package/extensions/provider/models/catalog.ts

@@ -35,9 +35,8 @@ function apiModelToProviderModel(model: NeuralwattApiModel): NeuralwattModel {
   const compat: NonNullable<ProviderModelConfig["compat"]> = {
     supportsDeveloperRole: meta.capabilities.developer_role,
     maxTokensField: "max_tokens",
+    ...COMPAT_OVERRIDES[model.id],
   };
-  if (reasoning) compat.requiresReasoningContentOnAssistantMessages = true;
-  Object.assign(compat, COMPAT_OVERRIDES[model.id]);
 
   const contextWindow = model.max_model_len;
 

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

@@ -9,10 +9,12 @@ import {
 // Pricing, capabilities, and limits are sourced from the API metadata fields;
 // `maxTokens` is `metadata.limits.max_output_tokens ?? max_model_len`.
 //
-// Each reasoning family snapshots its `reasoning.supported_efforts` +
-// `reasoning.mandatory` from the API; `buildThinkingLevelMap` turns that into
-// the Pi thinking level map by identity (no aliasing). See `models.test.ts`
-// for the drift check against the live catalog.
+// Each reasoning family snapshots its `reasoning.supported_efforts`,
+// `reasoning.mandatory`, and `reasoning.effort_aliases` from the API;
+// `buildThinkingLevelMap` turns that into the Pi thinking level map by
+// identity (no aliasing), while the anthropic-messages surface map resolves
+// unsupported levels through the aliases. See `models.test.ts` for the drift
+// check against the live catalog.
 
 // DeepSeek V4 Flash: efforts max/high/none, not mandatory.
 // https://api-docs.deepseek.com/guides/thinking_mode/
@@ -22,6 +24,24 @@ const DEEPSEEK_V4_FLASH: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["max", "high", "none"],
     mandatory: false,
+    effort_aliases: {
+      xhigh: "max",
+      medium: "high",
+      low: "high",
+      minimal: "high",
+    },
+  },
+};
+
+// DeepSeek. V4.1 Flash is vision-capable, unlike V4 Flash, and admits
+// reasoning levels up to `xhigh`. Not mandatory (off defaults).
+const DEEPSEEK_V4_1_FLASH: NeuralwattModelFamily = {
+  cost: { input: 0.15, output: 0.6, cacheRead: 0.015 },
+  vision: true,
+  reasoningMetadata: {
+    supported_efforts: ["max", "xhigh", "high", "low", "none"],
+    mandatory: false,
+    effort_aliases: { medium: "high", minimal: "low" },
   },
 };
 
@@ -37,6 +57,13 @@ const GEMMA_4: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["max", "none"],
     mandatory: false,
+    effort_aliases: {
+      xhigh: "max",
+      high: "max",
+      medium: "max",
+      low: "max",
+      minimal: "max",
+    },
   },
 };
 
@@ -48,6 +75,7 @@ const GLM_5_3: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["max", "high", "low"],
     mandatory: true,
+    effort_aliases: { xhigh: "max", medium: "high", minimal: "low" },
   },
 };
 
@@ -60,6 +88,7 @@ const GLM_5_3_FLASH: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["max", "high", "low"],
     mandatory: true,
+    effort_aliases: { xhigh: "max", medium: "high", minimal: "low" },
   },
 };
 
@@ -72,6 +101,7 @@ const KIMI_K3: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["max", "high", "low", "none"],
     mandatory: false,
+    effort_aliases: { xhigh: "max", medium: "high", minimal: "low" },
   },
 };
 
@@ -93,6 +123,13 @@ const QWEN_3_6_35B: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["high", "none"],
     mandatory: false,
+    effort_aliases: {
+      max: "high",
+      xhigh: "high",
+      medium: "high",
+      low: "high",
+      minimal: "high",
+    },
   },
 };
 
@@ -105,6 +142,7 @@ const QWEN_3_8_27B: NeuralwattModelFamily = {
   reasoningMetadata: {
     supported_efforts: ["xhigh", "medium", "low", "none"],
     mandatory: false,
+    effort_aliases: { max: "xhigh", high: "xhigh", minimal: "low" },
   },
 };
 
@@ -116,14 +154,14 @@ const FAMILIES: [NeuralwattModelFamily, NeuralwattVariantSpec[]][] = [
         id: "deepseek-v4-flash",
         name: "DeepSeek V4 Flash",
         contextWindow: 1048560,
-        maxOutputTokens: 65536,
+        maxOutputTokens: 393216,
         reasoning: true,
       },
       {
         id: "deepseek-v4-flash-flex",
         name: "DeepSeek V4 Flash (flex)",
         contextWindow: 1048560,
-        maxOutputTokens: 65536,
+        maxOutputTokens: 393216,
         reasoning: true,
         costMultiplier: 0.65,
       },
@@ -131,9 +169,29 @@ const FAMILIES: [NeuralwattModelFamily, NeuralwattVariantSpec[]][] = [
         id: "deepseek-v4-flash-speed",
         name: "DeepSeek V4 Flash (Speed)",
         contextWindow: 1048560,
-        maxOutputTokens: 65536,
+        maxOutputTokens: 393216,
+        reasoning: true,
+      },
+    ],
+  ],
+  [
+    DEEPSEEK_V4_1_FLASH,
+    [
+      {
+        id: "deepseek-v4.1-flash",
+        name: "DeepSeek V4.1 Flash",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
         reasoning: true,
       },
+      {
+        id: "deepseek-v4.1-flash-flex",
+        name: "DeepSeek V4.1 Flash (flex)",
+        contextWindow: 1048560,
+        maxOutputTokens: 393216,
+        reasoning: true,
+        costMultiplier: 0.65,
+      },
     ],
   ],
   [
@@ -250,21 +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/extensions/provider/models/refresh.ts

@@ -19,7 +19,7 @@ export const MODEL_STORE_TTL_MS = 4 * 60 * 60 * 1000;
  * must not be replayed for an anonymous user. Matches the anonymous-key
  * convention in src/lib/neuralwatt-api.ts (authHeaders).
  */
-const CATALOG_SCOPE_VERSION = "v1";
+const CATALOG_SCOPE_VERSION = "v2";
 type CatalogScope = "public" | "key";
 
 type ScopedModelsStoreEntry = ModelsStoreEntry & { catalogKey?: string };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.16.0",
+  "version": "0.16.2",
   "license": "MIT",
   "type": "module",
   "private": false,