@aliou/pi-neuralwatt 0.7.2 → 0.7.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aliou/pi-neuralwatt",
-  "version": "0.7.2",
+  "version": "0.7.4",
   "license": "MIT",
   "type": "module",
   "private": false,

package/src/extensions/provider/index.ts

@@ -46,6 +46,11 @@ function registerNeuralwattProvider(
   const { includeLegacyModelIds, includeHiddenModels } =
     configLoader.getConfig();
 
+  const publicModels = getNeuralwattModels({ includeLegacyModelIds });
+  const resolvedHiddenModels = includeHiddenModels
+    ? dedupeHiddenModels(hiddenModels, publicModels)
+    : [];
+
   const config: Parameters<ExtensionAPI["registerProvider"]>[1] = {
     baseUrl: "https://api.neuralwatt.com/v1",
     apiKey: "$NEURALWATT_API_KEY",
@@ -55,10 +60,7 @@ function registerNeuralwattProvider(
       Referer: "https://pi.dev",
       "X-Title": "npm:@aliou/pi-neuralwatt",
     },
-    models: [
-      ...getNeuralwattModels({ includeLegacyModelIds }),
-      ...(includeHiddenModels ? hiddenModels : []),
-    ],
+    models: [...publicModels, ...resolvedHiddenModels],
   };
 
   const provider = getApiProvider("openai-completions");
@@ -73,6 +75,23 @@ function registerNeuralwattProvider(
   pi.registerProvider("neuralwatt", config);
 }
 
+/**
+ * Drop any hidden model whose ID collides with a public or legacy model.
+ *
+ * Models can graduate from hidden (authenticated /v1/models only) to public
+ * (unauthenticated list). When that happens, a stale on-disk cache may still
+ * list the now-public ID, which would register it twice and make Pi treat the
+ * scoped model as ambiguous ("No models match pattern"). Dedupe against the
+ * public list so a stale cache can never shadow a public model.
+ */
+function dedupeHiddenModels(
+  hiddenModels: ProviderModelConfig[],
+  publicModels: ProviderModelConfig[],
+): ProviderModelConfig[] {
+  const publicIds = new Set(publicModels.map((m) => m.id));
+  return hiddenModels.filter((m) => !publicIds.has(m.id));
+}
+
 export default async function (pi: ExtensionAPI) {
   await configLoader.load();
 
@@ -171,12 +190,21 @@ export default async function (pi: ExtensionAPI) {
       return { message };
     }
 
+    // Fallback for 429s where no layer-specific headers were captured. The
+    // streamSimple wrap (wrapNeuralwattStreamSimple) already formats a
+    // detailed message via formatRateLimitError when it captures headers;
+    // detect that case by the `"429 rate limit:"` prefix it emits and leave
+    // it untouched. This branch only fires for genuinely headerless 429s
+    // (e.g. anonymous playground limits, or a 429 from infra in front of
+    // Neuralwatt), since after_provider_response cannot observe 429s — the
+    // OpenAI SDK throws before Pi's onResponse hook runs.
     if (
       event.message.role === "assistant" &&
       event.message.stopReason === "error" &&
       (event.message.provider === "neuralwatt" ||
         ctx.model?.provider === "neuralwatt") &&
-      event.message.errorMessage?.includes("429")
+      event.message.errorMessage?.includes("429") &&
+      !event.message.errorMessage.startsWith("429 rate limit:")
     ) {
       return {
         message: normalizeNeuralwattRateLimitError(event.message, {
@@ -245,14 +273,14 @@ export default async function (pi: ExtensionAPI) {
         ctx.modelRegistry.authStorage,
         hiddenModelsAbort.signal,
       );
-      // Persist for the next startup so scoped models resolve without warnings
-      // on Pi's subsequent launches.
-      if (fetched.length > 0) {
-        hiddenModels = fetched;
-        await writeHiddenModelsCache(hiddenModels);
-        if (!hiddenModelsAbort.signal.aborted) {
-          registerNeuralwattProvider(pi, handleSseQuota, hiddenModels);
-        }
+      // Persist for the next startup so scoped models resolve without
+      // warnings on Pi's subsequent launches. Always write the cache (even
+      // when empty) and re-register, so graduated or removed hidden models
+      // are purged from both the cache and the provider's model list.
+      hiddenModels = fetched;
+      await writeHiddenModelsCache(hiddenModels);
+      if (!hiddenModelsAbort.signal.aborted) {
+        registerNeuralwattProvider(pi, handleSseQuota, hiddenModels);
       }
     }
 

package/src/extensions/provider/models/hidden.ts

@@ -10,19 +10,10 @@ import { NEURALWATT_MODELS } from "./public-models";
 // 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.
+// Previously hidden models that have since gone public now live in public-models.ts.
 const HIDDEN_MODEL_OVERRIDES: Partial<
   Record<string, Partial<ProviderModelConfig>>
-> = {
-  "glm-5.2-short": {
-    thinkingLevelMap: {
-      minimal: null,
-      low: null,
-      medium: null,
-      high: "high",
-      xhigh: "max",
-    },
-  },
-};
+> = {};
 
 function buildHiddenModel(apiModel: NeuralwattApiModel): ProviderModelConfig {
   const meta = apiModel.metadata;

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

@@ -49,6 +49,52 @@ export const NEURALWATT_MODELS: ProviderModelConfig[] = [
       maxTokensField: "max_tokens",
     },
   },
+  // GLM-5.2 Short - ZhipuAI (200K context, bounded reasoning budget)
+  {
+    id: "glm-5.2-short",
+    name: "GLM-5.2 Short",
+    reasoning: true,
+    input: ["text"],
+    cost: {
+      input: 1.45,
+      output: 4.5,
+      cacheRead: 0.3625,
+      cacheWrite: 0,
+    },
+    contextWindow: 199984,
+    maxTokens: 65536,
+    thinkingLevelMap: {
+      minimal: null,
+      low: null,
+      medium: null,
+      high: "high",
+      xhigh: "max",
+    },
+    compat: {
+      supportsDeveloperRole: false,
+      maxTokensField: "max_tokens",
+      requiresReasoningContentOnAssistantMessages: true,
+    },
+  },
+  // GLM-5.2 Short Fast - ZhipuAI (200K context, reasoning disabled)
+  {
+    id: "glm-5.2-short-fast",
+    name: "GLM-5.2 Short Fast",
+    reasoning: false,
+    input: ["text"],
+    cost: {
+      input: 1.45,
+      output: 4.5,
+      cacheRead: 0.3625,
+      cacheWrite: 0,
+    },
+    contextWindow: 199984,
+    maxTokens: 65536,
+    compat: {
+      supportsDeveloperRole: false,
+      maxTokensField: "max_tokens",
+    },
+  },
   // Kimi K2.5 - MoonshotAI
   {
     id: "moonshotai/Kimi-K2.5",