@aliou/pi-neuralwatt 0.9.0 → 0.10.2

package/extensions/provider/models/hidden.ts

@@ -3,6 +3,11 @@ import { fetchNeuralwattModels } from "../../../src/lib/neuralwatt-api";
 import type { NeuralwattApiModel } from "../../../src/types/models-api";
 import { NEURALWATT_MODELS } from "./public-models";
 
+// Hidden aliases that work for authorized accounts but are omitted from the
+// authenticated /v1/models response. Keep these gated by includeHiddenModels.
+// Move an entry to public-models.ts once Neuralwatt advertises it publicly.
+export const HIDDEN_NEURALWATT_MODELS: ProviderModelConfig[] = [];
+
 // Per-ID overrides for known hidden models. The authenticated /v1/models endpoint
 // exposes pricing and capabilities, but some Pi-specific behavior (thinking levels,
 // compat flags) has to be supplied by hand.

package/extensions/provider/models/index.ts

@@ -2,7 +2,7 @@ import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
 import { buildLegacyNeuralwattModels } from "./legacy";
 import { NEURALWATT_MODELS } from "./public-models";
 
-export { loadHiddenModels } from "./hidden";
+export { HIDDEN_NEURALWATT_MODELS, loadHiddenModels } from "./hidden";
 export {
   buildLegacyNeuralwattModels,
   LEGACY_MODEL_ALIAS_MAP,

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

@@ -3,7 +3,29 @@ import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
 // Public models returned by https://api.neuralwatt.com/v1/models (unauthenticated view).
 // Pricing, capabilities, and limits are sourced from the API metadata fields.
 export const NEURALWATT_MODELS: ProviderModelConfig[] = [
+  // Gemma 4 31B - Google, served from NVIDIA's NVFP4 checkpoint
+  {
+    id: "gemma-4-31b",
+    name: "Gemma 4 31B",
+    reasoning: false,
+    input: ["text", "image"],
+    cost: {
+      input: 0.144,
+      output: 0.42,
+      cacheRead: 0.036,
+      cacheWrite: 0,
+    },
+    contextWindow: 262128,
+    maxTokens: 16384,
+    compat: {
+      supportsDeveloperRole: false,
+      maxTokensField: "max_tokens",
+    },
+  },
   // GLM-5.2 - ZhipuAI
+  // Native reasoning tiers: off (none), high, max. Exposes Pi's `max` level
+  // (introduced in Pi 0.80.6) for GLM's top tier; `xhigh` is an unsupported
+  // hole between `high` and `max`.
   {
     id: "glm-5.2",
     name: "GLM-5.2",
@@ -23,7 +45,8 @@ export const NEURALWATT_MODELS: ProviderModelConfig[] = [
       low: null,
       medium: null,
       high: "high",
-      xhigh: "max",
+      xhigh: null,
+      max: "max",
     },
     compat: {
       supportsDeveloperRole: false,
@@ -70,7 +93,8 @@ export const NEURALWATT_MODELS: ProviderModelConfig[] = [
       low: null,
       medium: null,
       high: "high",
-      xhigh: "max",
+      xhigh: null,
+      max: "max",
     },
     compat: {
       supportsDeveloperRole: false,
@@ -283,7 +307,8 @@ export const NEURALWATT_MODELS: ProviderModelConfig[] = [
       low: null,
       medium: null,
       high: "high",
-      xhigh: "max",
+      xhigh: null,
+      max: "max",
     },
     compat: {
       supportsDeveloperRole: false,
@@ -366,7 +391,8 @@ export const NEURALWATT_MODELS: ProviderModelConfig[] = [
       low: null,
       medium: null,
       high: "high",
-      xhigh: "max",
+      xhigh: null,
+      max: "max",
     },
     compat: {
       supportsDeveloperRole: false,

package/extensions/provider/models/refresh.ts

@@ -5,7 +5,7 @@ import type {
   RefreshModelsContext,
 } from "@earendil-works/pi-ai";
 import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
-import { loadHiddenModels } from "./hidden";
+import { HIDDEN_NEURALWATT_MODELS, loadHiddenModels } from "./hidden";
 import { buildLegacyNeuralwattModels } from "./legacy";
 import { NEURALWATT_MODELS } from "./public-models";
 
@@ -67,6 +67,22 @@ function dedupeHiddenModels(
   return hiddenModels.filter((model) => !baselineIds.has(model.id));
 }
 
+function configuredHiddenModels(
+  discoveredModels: ProviderModelConfig[],
+  baselineModels: ProviderModelConfig[],
+): ProviderModelConfig[] {
+  const hardcodedIds = new Set(
+    HIDDEN_NEURALWATT_MODELS.map((model) => model.id),
+  );
+  return dedupeHiddenModels(
+    [
+      ...HIDDEN_NEURALWATT_MODELS,
+      ...discoveredModels.filter((model) => !hardcodedIds.has(model.id)),
+    ],
+    baselineModels,
+  );
+}
+
 function cachedHiddenModels(
   stored: ModelsStoreEntry | undefined,
 ): ProviderModelConfig[] {
@@ -104,7 +120,10 @@ export async function refreshNeuralwattModels(
     return baseline;
   }
 
-  const cachedHidden = dedupeHiddenModels(cachedHiddenModels(stored), baseline);
+  const cachedHidden = configuredHiddenModels(
+    cachedHiddenModels(stored),
+    baseline,
+  );
   const cachedCatalog = [...baseline, ...cachedHidden];
 
   if (!context.allowNetwork || context.signal?.aborted) {
@@ -124,7 +143,7 @@ export async function refreshNeuralwattModels(
     throw new Error("Neuralwatt model catalog refresh failed");
   }
 
-  const catalog = [...baseline, ...dedupeHiddenModels(hidden, baseline)];
+  const catalog = [...baseline, ...configuredHiddenModels(hidden, baseline)];
   await persistModels(context, catalog);
   return catalog;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.9.0",
+  "version": "0.10.2",
   "license": "MIT",
   "type": "module",
   "private": false,