@prestyj/ai 5.12.0 → 5.16.0

package/README.md

@@ -42,7 +42,7 @@ Tool parameters are Zod schemas. Converted to JSON Schema at the provider bounda
 |---|---|---|
 | `anthropic` | Claude Opus 5, Sonnet 5, Haiku 4.5 | Extended thinking, prompt caching, server-side compaction |
 | `openai` | GPT-4.1, o3, o4-mini | Supports OAuth (codex endpoint) and API keys |
-| `glm` | GLM-5.1, GLM-4.7 | Z.AI platform, OpenAI-compatible |
+| `glm` | GLM-5.3 | Z.AI platform, OpenAI-compatible |
 | `moonshot` | Kimi K3, Kimi K2.7 Code | Moonshot platform, OpenAI-compatible |
 
 ---

package/dist/index.cjs

@@ -53,6 +53,7 @@ __export(index_exports, {
   redactText: () => redactText,
   redactValue: () => redactValue,
   registerPalsuProvider: () => registerPalsuProvider,
+  resolveToolSchema: () => resolveToolSchema,
   sanitizeMessagesForWire: () => sanitizeMessagesForWire,
   setProviderDiagnostic: () => setProviderDiagnostic,
   sliceHead: () => sliceHead,
@@ -125,6 +126,7 @@ var PROVIDER_DISPLAY = {
   openrouter: "OpenRouter",
   sakana: "Sakana",
   xai: "xAI (Grok)",
+  huggingface: "Hugging Face",
   xiaomi: "Xiaomi (MiMo)",
   minimax: "MiniMax"
 };
@@ -192,7 +194,7 @@ function formatError(err) {
         provider: err.provider,
         statusCode: err.statusCode,
         ...err.requestId ? { requestId: err.requestId } : {},
-        guidance: "Request access via your Anthropic account team (see platform.claude.com/docs/en/about-claude/models/overview), or switch to Claude Fable 5 via the model selector \u2014 same underlying model, generally available."
+        guidance: "Request access via your Anthropic account team (see platform.claude.com/docs/en/about-claude/models/overview), or switch to Claude Fable 5.1 via the model selector \u2014 same underlying model, generally available."
       };
     }
     if (isUsageLimitError(err)) {
@@ -1171,6 +1173,9 @@ function toLocalReasoningEffort(level) {
   if (level === "max" || level === "ultra" || level === "xhigh") return "max";
   return level;
 }
+function toGlmReasoningEffort(level) {
+  return level === "ultra" ? "max" : level;
+}
 function toOpenAIReasoningEffort(level, model) {
   const effort = level === "max" || level === "ultra" ? "xhigh" : level;
   if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
@@ -1617,7 +1622,15 @@ async function* runStream(options) {
           yield keepalive;
           break;
         }
-        // message_stop — loop exits naturally
+        // message_stop — loop exits naturally.
+        //
+        // Deliberately NOT breaking early here. Breaking makes the SDK iterator
+        // run `if (!done) controller.abort()` in its `finally`
+        // (core/streaming.js:97), which tears the connection down instead of
+        // returning it to the keep-alive pool — every turn would then pay a
+        // fresh TLS handshake. Draining to the end is what every other Anthropic
+        // client does, and the stall it guards against is handled by the agent
+        // loop's idle timeout.
         default:
           yield keepalive;
           break;
@@ -2040,6 +2053,11 @@ async function* runStream2(options) {
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };
+      if (options.provider === "glm") {
+        params.reasoning_effort = toGlmReasoningEffort(
+          options.thinking
+        );
+      }
     } else {
       params.thinking = { type: "disabled" };
     }
@@ -3022,6 +3040,7 @@ var CODE_ASSIST_SUPPORTED_MODELS = /* @__PURE__ */ new Set([
   "gemini-3.5-flash",
   "gemini-3-flash",
   "gemini-3.1-flash-lite",
+  "gemini-3.7-flash",
   "gemini-2.5-pro",
   "gemini-2.5-flash",
   "gemma-4-31b-it",
@@ -3044,13 +3063,14 @@ var ACCOUNT_GATED_MODELS = /* @__PURE__ */ new Set([
   "gemini-3-flash",
   "gemini-3.5-flash",
   "gemini-3.1-pro-preview",
-  "gemini-3.1-pro-preview-customtools"
+  "gemini-3.1-pro-preview-customtools",
+  "gemini-3.7-flash"
 ]);
 function accountGatedMessage(model) {
   return `Your Google account isn't entitled to "${model}" over Gemini Code Assist OAuth, so the API reports it as not found. This is an account-access limit, not a ezcoder bug.`;
 }
 function accountGatedHint() {
-  return `Newer Gemini models (3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin \u2014 free/personal accounts usually can't call them. Switch to Gemini 3.1 Flash Lite (it works on this account) with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
+  return `Newer Gemini models (3.7 Flash, 3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin \u2014 free/personal accounts usually can't call them. Switch to Gemini 3.1 Flash Lite (it works on this account) with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
 }
 function formatErrorMessage(status, body, model) {
   if (status === 404 && !CODE_ASSIST_SUPPORTED_MODELS.has(model)) {
@@ -3734,6 +3754,18 @@ providerRegistry.register("openrouter", {
     baseUrl: options.baseUrl ?? "https://openrouter.ai/api/v1"
   })
 });
+providerRegistry.register("huggingface", {
+  // Hugging Face Inference Providers router — one HF token (hf.co/settings/tokens,
+  // "Make calls to Inference Providers" permission) routes to whichever hosted
+  // backend serves each open model. Chat Completions-compatible; model ids are
+  // Hub repo paths ("Qwen/Qwen3-Coder-480B-A35B-Instruct"), optionally with an
+  // ":auto"/":fastest"/":cheapest" provider-selection suffix. Billing follows
+  // each backend's per-token rates on the HF account (small free tier).
+  stream: (options) => streamOpenAI({
+    ...options,
+    baseUrl: options.baseUrl ?? "https://router.huggingface.co/v1"
+  })
+});
 providerRegistry.register("sakana", {
   // Sakana Fugu is a multi-agent system exposed as a standard LLM through the
   // OpenAI-compatible Sakana API. We ride the Chat Completions transport (the
@@ -4229,6 +4261,7 @@ function registerPalsuProvider(config) {
   redactText,
   redactValue,
   registerPalsuProvider,
+  resolveToolSchema,
   sanitizeMessagesForWire,
   setProviderDiagnostic,
   sliceHead,