@kenkaiiii/gg-ai 5.17.0 → 5.18.0

package/dist/index.js

@@ -1832,7 +1832,11 @@ async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
   const useStreaming = options.streaming !== false;
   const client = createClient2(options);
-  const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" || options.provider === "xiaomi";
+  const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
+  const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").endsWith("/coding/v1") === true;
+  const isKimiK27 = options.provider === "moonshot" && options.model.startsWith("kimi-k2.7-code");
+  const hasFixedKimiSampling = isKimiK3 || isKimiK27;
+  const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" && !isKimiK3 && !isKimiK27 || options.provider === "xiaomi";
   const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
   const downgradedMessages = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
   if (options.provider === "moonshot") {
@@ -1844,7 +1848,9 @@ async function* runStream2(options) {
   }
   const messages = toOpenAIMessages(downgradedMessages, {
     provider: options.provider,
-    thinking: !!options.thinking,
+    // K3 and K2.7 preserve reasoning even when the user hides thinking in the
+    // UI; keep assistant tool-call history wire-valid in that display mode.
+    thinking: isKimiK3 || isKimiK27 || !!options.thinking,
     supportsImages: options.supportsImages
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
@@ -1854,10 +1860,10 @@ async function* runStream2(options) {
     messages,
     stream: useStreaming,
     ...options.maxTokens ? { max_completion_tokens: options.maxTokens } : {},
-    ...effectiveTemp != null && !options.thinking ? { temperature: effectiveTemp } : {},
-    ...options.topP != null ? { top_p: options.topP } : {},
+    ...effectiveTemp != null && !options.thinking && !hasFixedKimiSampling ? { temperature: effectiveTemp } : {},
+    ...options.topP != null && !hasFixedKimiSampling ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
-    ...options.thinking && !usesThinkingParam ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
+    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
     ...options.tools?.length ? { tools: toOpenAITools(options.tools) } : {},
     ...options.toolChoice && options.tools?.length ? { tool_choice: toOpenAIToolChoice(options.toolChoice) } : {},
     ...useStreaming ? { stream_options: { include_usage: true } } : {}
@@ -1867,13 +1873,21 @@ async function* runStream2(options) {
     paramsAny.prompt_cache_key = normalizePromptCacheKey(options.promptCacheKey ?? "ggcoder");
     if (options.provider === "openai" && options.model.startsWith("gpt-5.6")) {
       paramsAny.prompt_cache_options = { mode: "implicit", ttl: "30m" };
-    } else if ((options.cacheRetention ?? "short") === "long") {
+    } else if (!isKimiK3 && (options.cacheRetention ?? "short") === "long") {
       paramsAny.prompt_cache_retention = "24h";
     }
   }
   if (options.provider === "openai" && options.serviceTier) {
     params.service_tier = options.serviceTier;
   }
+  if (isKimiK3) {
+    const paramsAny = params;
+    if (isManagedKimiK3) {
+      paramsAny.thinking = { type: "enabled", effort: "max", keep: "all" };
+    } else {
+      paramsAny.reasoning_effort = "max";
+    }
+  }
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };