@kenkaiiii/gg-ai 4.3.2 → 4.3.4

package/dist/index.js

@@ -426,22 +426,37 @@ function normalizeOpenAIStopReason(reason) {
 }
 
 // src/providers/anthropic.ts
+var clientCache = /* @__PURE__ */ new Map();
+function getOrCreateClient(options) {
+  const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
+  const key = `${options.apiKey ?? ""}|${options.baseUrl ?? ""}|${isOAuth}`;
+  let client = clientCache.get(key);
+  if (!client) {
+    client = new Anthropic({
+      ...isOAuth ? { apiKey: null, authToken: options.apiKey } : { apiKey: options.apiKey },
+      ...options.baseUrl ? { baseURL: options.baseUrl } : {},
+      ...options.fetch ? { fetch: options.fetch } : {},
+      // Disable SDK-level retries — the agent loop handles retries itself with
+      // stall detection and context compaction.  SDK retries on abort just cycle
+      // through the already-aborted signal, wasting time.
+      maxRetries: 0,
+      ...isOAuth ? {
+        defaultHeaders: {
+          "user-agent": "claude-cli/2.1.75",
+          "x-app": "cli"
+        }
+      } : {}
+    });
+    clientCache.set(key, client);
+  }
+  return client;
+}
 function streamAnthropic(options) {
   return new StreamResult(runStream(options));
 }
 async function* runStream(options) {
+  const client = getOrCreateClient(options);
   const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
-  const client = new Anthropic({
-    ...isOAuth ? { apiKey: null, authToken: options.apiKey } : { apiKey: options.apiKey },
-    ...options.baseUrl ? { baseURL: options.baseUrl } : {},
-    ...options.fetch ? { fetch: options.fetch } : {},
-    ...isOAuth ? {
-      defaultHeaders: {
-        "user-agent": "claude-cli/2.1.75",
-        "x-app": "cli"
-      }
-    } : {}
-  });
   const cacheControl = toAnthropicCacheControl(options.cacheRetention, options.baseUrl);
   const { system: rawSystem, messages } = toAnthropicMessages(options.messages, cacheControl);
   const system = isOAuth ? [
@@ -705,16 +720,26 @@ function toError(err) {
 
 // src/providers/openai.ts
 import OpenAI from "openai";
+var clientCache2 = /* @__PURE__ */ new Map();
+function getOrCreateClient2(options) {
+  const key = `${options.apiKey ?? ""}|${options.baseUrl ?? ""}`;
+  let client = clientCache2.get(key);
+  if (!client) {
+    client = new OpenAI({
+      apiKey: options.apiKey,
+      ...options.baseUrl ? { baseURL: options.baseUrl } : {},
+      ...options.fetch ? { fetch: options.fetch } : {}
+    });
+    clientCache2.set(key, client);
+  }
+  return client;
+}
 function streamOpenAI(options) {
   return new StreamResult(runStream2(options));
 }
 async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
-  const client = new OpenAI({
-    apiKey: options.apiKey,
-    ...options.baseUrl ? { baseURL: options.baseUrl } : {},
-    ...options.fetch ? { fetch: options.fetch } : {}
-  });
+  const client = getOrCreateClient2(options);
   const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" || options.provider === "xiaomi";
   const messages = toOpenAIMessages(options.messages, { provider: options.provider });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
@@ -723,7 +748,7 @@ async function* runStream2(options) {
     model: options.model,
     messages,
     stream: true,
-    ...options.maxTokens ? options.provider === "xiaomi" ? { max_completion_tokens: options.maxTokens } : { max_tokens: options.maxTokens } : {},
+    ...options.maxTokens ? { max_tokens: options.maxTokens } : {},
     ...effectiveTemp != null && !options.thinking ? { temperature: effectiveTemp } : {},
     ...options.topP != null ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
@@ -743,7 +768,7 @@ async function* runStream2(options) {
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };
-    } else if (options.provider !== "xiaomi") {
+    } else {
       params.thinking = { type: "disabled" };
     }
   }
@@ -792,7 +817,9 @@ async function* runStream2(options) {
     const reasoningContent = delta.reasoning_content;
     if (typeof reasoningContent === "string" && reasoningContent) {
       thinkingAccum += reasoningContent;
-      yield { type: "thinking_delta", text: reasoningContent };
+      if (options.thinking) {
+        yield { type: "thinking_delta", text: reasoningContent };
+      }
     }
     if (delta.content) {
       textAccum += delta.content;