@kenkaiiii/gg-ai 4.3.6 → 4.3.8

package/dist/index.cjs

@@ -472,36 +472,29 @@ function normalizeOpenAIStopReason(reason) {
 }
 
 // src/providers/anthropic.ts
-var clientCache = /* @__PURE__ */ new Map();
-function getOrCreateClient(options) {
+function createClient(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 import_sdk.default({
-      ...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;
+  return new import_sdk.default({
+    ...isOAuth ? { apiKey: null, authToken: options.apiKey } : { apiKey: options.apiKey },
+    ...options.baseUrl ? { baseURL: options.baseUrl } : {},
+    ...options.fetch ? { fetch: options.fetch } : {},
+    // Allow SDK retries for connection-level failures (socket hang up, 500s,
+    // connection refused).  Our stall detection handles abort-initiated retries
+    // separately — SDK retries only fire on genuine transport errors.
+    maxRetries: 2,
+    ...isOAuth ? {
+      defaultHeaders: {
+        "user-agent": "claude-cli/2.1.75",
+        "x-app": "cli"
+      }
+    } : {}
+  });
 }
 function streamAnthropic(options) {
   return new StreamResult(runStream(options));
 }
 async function* runStream(options) {
-  const client = getOrCreateClient(options);
+  const client = createClient(options);
   const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
   const cacheControl = toAnthropicCacheControl(options.cacheRetention, options.baseUrl);
   const { system: rawSystem, messages } = toAnthropicMessages(options.messages, cacheControl);
@@ -550,10 +543,13 @@ async function* runStream(options) {
     })(),
     stream: true
   };
+  const hasAdaptiveThinking = options.model.includes("opus-4-6") || options.model.includes("opus-4.6") || options.model.includes("sonnet-4-6") || options.model.includes("sonnet-4.6");
   const betaHeaders = [
     ...isOAuth ? ["claude-code-20250219", "oauth-2025-04-20"] : [],
     ...options.compaction ? ["compact-2026-01-12"] : [],
-    ...options.clearToolUses ? ["context-management-2025-06-27"] : []
+    ...options.clearToolUses ? ["context-management-2025-06-27"] : [],
+    "fine-grained-tool-streaming-2025-05-14",
+    ...!hasAdaptiveThinking ? ["interleaved-thinking-2025-05-14"] : []
   ];
   const stream2 = client.messages.stream(params, {
     signal: options.signal ?? void 0,
@@ -766,26 +762,19 @@ function toError(err) {
 
 // src/providers/openai.ts
 var import_openai = __toESM(require("openai"), 1);
-var clientCache2 = /* @__PURE__ */ new Map();
-function getOrCreateClient2(options) {
-  const key = `${options.apiKey ?? ""}|${options.baseUrl ?? ""}`;
-  let client = clientCache2.get(key);
-  if (!client) {
-    client = new import_openai.default({
-      apiKey: options.apiKey,
-      ...options.baseUrl ? { baseURL: options.baseUrl } : {},
-      ...options.fetch ? { fetch: options.fetch } : {}
-    });
-    clientCache2.set(key, client);
-  }
-  return client;
+function createClient2(options) {
+  return new import_openai.default({
+    apiKey: options.apiKey,
+    ...options.baseUrl ? { baseURL: options.baseUrl } : {},
+    ...options.fetch ? { fetch: options.fetch } : {}
+  });
 }
 function streamOpenAI(options) {
   return new StreamResult(runStream2(options));
 }
 async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
-  const client = getOrCreateClient2(options);
+  const client = createClient2(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;
@@ -794,7 +783,7 @@ async function* runStream2(options) {
     model: options.model,
     messages,
     stream: true,
-    ...options.maxTokens ? { max_tokens: options.maxTokens } : {},
+    ...options.maxTokens ? { max_completion_tokens: options.maxTokens } : {},
     ...effectiveTemp != null && !options.thinking ? { temperature: effectiveTemp } : {},
     ...options.topP != null ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
@@ -939,6 +928,12 @@ function toError2(err, provider = "openai") {
     let msg = err.message;
     const body = err.error;
     if (body) {
+      const modelName = body.model || "";
+      const _code = body.code || "";
+      const message = body.message || "";
+      if (modelName === "codex-mini-latest" || message.includes("codex-mini-latest")) {
+        msg = `codex-mini-latest requires an OpenAI Pro or Max subscription. You currently have access to GPT-5.4 and GPT-5.4 Mini with your account.`;
+      }
       msg += ` | body: ${JSON.stringify(body)}`;
     }
     return new ProviderError(provider, msg, {
@@ -1008,6 +1003,11 @@ async function* runStream3(options) {
       message += `
 
 Hint: Codex models require a ChatGPT Plus ($20/mo) or Pro ($200/mo) subscription. The "codex-spark" variants require ChatGPT Pro. Ensure your account has an active subscription at https://chatgpt.com/settings`;
+    }
+    if (response.status === 404 && text.includes("does not exist")) {
+      message += `
+
+Hint: codex-mini-latest requires an OpenAI Pro ($200/mo) or Max subscription. GPT-5.4 and GPT-5.4 Mini work with any active ChatGPT plan.`;
     }
     throw new ProviderError("openai", message, {
       statusCode: response.status