koishi-plugin-chatluna-google-gemini-adapter 1.2.6 → 1.2.8

package/lib/index.cjs

@@ -33,14 +33,14 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 // src/locales/zh-CN.schema.yml
 var require_zh_CN_schema = __commonJS({
   "src/locales/zh-CN.schema.yml"(exports2, module2) {
-    module2.exports = { $inner: [{}, { $desc: "请求选项", platform: "适配器的平台名。（不懂请不要修改）", apiKeys: { $inner: ["Gemini 的 API Key", "Gemini API 的请求地址"], $desc: "Gemini 的 API Key 和请求地址列表。" } }, { $desc: "模型配置", maxTokens: "输入的最大上下文 Token（16~2097000，必须是 16 的倍数）。注意：仅当您使用的模型最大 Token 为 8000 及以上时，才建议设置超过 2000 token。", temperature: "回复的随机性程度，数值越高，回复越随机（范围：0~2）。", googleSearch: "为模型启用谷歌搜索。", thinkingBudget: "思考预算，范围：(-1~24576)，设置的数值越大，思考时花费的 Token 越多,-1 为动态思考，0 为不思考。目前仅支持 gemini 2.5 系列模型。", groundingContentDisplay: "是否显示谷歌搜索结果。", imageGeneration: "为模型启用图像生成。目前仅支持 `gemini-2.0-flash-exp` 模型。", searchThreshold: "搜索的[置信度阈值](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval)，范围：0~1，设置的数值越低，则越倾向于使用谷歌搜索。（仅支持 `gemini-1.5` 系列模型。gemini 2.0 模型起使用动态的工具调用）", includeThoughts: "是否获取模型的思考内容。", codeExecution: "为模型启用代码执行工具。", urlContext: "为模型启用 URL 内容获取工具。" }] };
+    module2.exports = { $inner: [{}, { $desc: "请求选项", platform: "适配器的平台名。（不懂请不要修改）", apiKeys: { $inner: ["Gemini 的 API Key", "Gemini API 的请求地址"], $desc: "Gemini 的 API Key 和请求地址列表。" } }, { $desc: "模型配置", maxTokens: "输入的最大上下文 Token（16~2097000，必须是 16 的倍数）。注意：仅当您使用的模型最大 Token 为 8000 及以上时，才建议设置超过 2000 token。", temperature: "回复的随机性程度，数值越高，回复越随机（范围：0~2）。", googleSearch: "为模型启用谷歌搜索。", thinkingBudget: "思考预算，范围：(-1~24576)，设置的数值越大，思考时花费的 Token 越多,-1 为动态思考。目前仅支持 gemini 2.5 系列模型。", groundingContentDisplay: "是否显示谷歌搜索结果。", imageGeneration: "为模型启用图像生成。目前仅支持 `gemini-2.0-flash-exp` 模型。", searchThreshold: "搜索的[置信度阈值](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval)，范围：0~1，设置的数值越低，则越倾向于使用谷歌搜索。（仅支持 `gemini-1.5` 系列模型。gemini 2.0 模型起使用动态的工具调用）", includeThoughts: "是否获取模型的思考内容。", codeExecution: "为模型启用代码执行工具。", urlContext: "为模型启用 URL 内容获取工具。" }] };
   }
 });
 
 // src/locales/en-US.schema.yml
 var require_en_US_schema = __commonJS({
   "src/locales/en-US.schema.yml"(exports2, module2) {
-    module2.exports = { $inner: [{}, { $desc: "API Configuration", platform: "Adapter platform name. (Do not modify if you do not understand)", apiKeys: { $inner: ["Gemini API Key", "Gemini API Endpoint (optional)"], $desc: "Gemini API access credentials" } }, { $desc: "Model Parameters", maxTokens: "Max output tokens (16-2097000, multiple of 16). >2000 for 8k+ models", temperature: "Sampling temperature (0-2). Higher: more random, Lower: more deterministic", googleSearch: "Enable Google search", thinkingBudget: "Thinking budget (-1-24576). (-1: dynamic thinking, 0: no thinking) Higher: more tokens spent on thinking. Currently only supports `gemini-2.5` series models.", groundingContentDisplay: "Enable display of search results", imageGeneration: "Enable image generation (only for `gemini-2.0-flash-exp` model)", searchThreshold: "Search confidence [threshold](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval) (0-1). Lower: more likely to use Google search", includeThoughts: "Enable retrieval of model thoughts", codeExecution: "Enable code execution tool", urlContext: "Enable URL context retrieval tool" }] };
+    module2.exports = { $inner: [{}, { $desc: "API Configuration", platform: "Adapter platform name. (Do not modify if you do not understand)", apiKeys: { $inner: ["Gemini API Key", "Gemini API Endpoint (optional)"], $desc: "Gemini API access credentials" } }, { $desc: "Model Parameters", maxTokens: "Max output tokens (16-2097000, multiple of 16). >2000 for 8k+ models", temperature: "Sampling temperature (0-2). Higher: more random, Lower: more deterministic", googleSearch: "Enable Google search", thinkingBudget: "Thinking budget (-1-24576). (0: dynamic thinking) Higher: more tokens spent on thinking. Currently only supports `gemini-2.5` series models.", groundingContentDisplay: "Enable display of search results", imageGeneration: "Enable image generation (only for `gemini-2.0-flash-exp` model)", searchThreshold: "Search confidence [threshold](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval) (0-1). Lower: more likely to use Google search", includeThoughts: "Enable retrieval of model thoughts", codeExecution: "Enable code execution tool", urlContext: "Enable URL context retrieval tool" }] };
   }
 });
 
@@ -76,7 +76,7 @@ var import_stream = require("koishi-plugin-chatluna/utils/stream");
 var import_messages = require("@langchain/core/messages");
 var import_zod_to_json_schema = require("zod-to-json-schema");
 async function langchainMessageToGeminiMessage(messages, model) {
-  const mappedMessage = await Promise.all(
+  const mappedMessages = await Promise.all(
     messages.map(async (rawMessage) => {
       const role = messageTypeToGeminiRole(rawMessage.getType());
       if (role === "function" || rawMessage.additional_kwargs?.function_call != null) {
@@ -90,15 +90,15 @@ async function langchainMessageToGeminiMessage(messages, model) {
                   name: rawMessage.name,
                   content: (() => {
                     try {
-                      const result3 = JSON.parse(
+                      const result2 = JSON.parse(
                         rawMessage.content
                       );
-                      if (typeof result3 === "string") {
+                      if (typeof result2 === "string") {
                         return {
-                          response: result3
+                          response: result2
                         };
                       } else {
-                        return result3;
+                        return result2;
                       }
                     } catch (e) {
                       return {
@@ -112,15 +112,15 @@ async function langchainMessageToGeminiMessage(messages, model) {
                 name: rawMessage.additional_kwargs.function_call.name,
                 args: (() => {
                   try {
-                    const result3 = JSON.parse(
+                    const result2 = JSON.parse(
                       rawMessage.additional_kwargs.function_call.arguments
                     );
-                    if (typeof result3 === "string") {
+                    if (typeof result2 === "string") {
                       return {
-                        input: result3
+                        input: result2
                       };
                     } else {
-                      return result3;
+                      return result2;
                     }
                   } catch (e) {
                     return {
@@ -134,7 +134,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
         };
       }
       const images = rawMessage.additional_kwargs.images;
-      const result2 = {
+      const result = {
         role,
         parts: [
           {
@@ -146,7 +146,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
         for (const image of images) {
           const mineType = image.split(";")?.[0]?.split(":")?.[1];
           const data = image.replace(/^data:image\/\w+;base64,/, "");
-          result2.parts.push({
+          result.parts.push({
             inline_data: {
               // base64 image match type
               data,
@@ -154,7 +154,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
             }
           });
         }
-        result2.parts = result2.parts.filter((uncheckedPart) => {
+        result.parts = result.parts.filter((uncheckedPart) => {
           const part = partAsTypeCheck(
             uncheckedPart,
             (part2) => part2["text"] != null
@@ -162,54 +162,39 @@ async function langchainMessageToGeminiMessage(messages, model) {
           return part == null || part.text.length > 0;
         });
       }
-      return result2;
+      return result;
     })
   );
-  const result = [];
-  for (let i = 0; i < mappedMessage.length; i++) {
-    const message = mappedMessage[i];
-    if (message.role !== "system") {
-      result.push(message);
-      continue;
-    }
-    result.push({
-      role: "user",
-      parts: message.parts
-    });
-    const nextMessage = mappedMessage?.[i + 1];
-    if (nextMessage?.role === "model") {
-      continue;
-    }
-    if (nextMessage?.role === "user" || nextMessage?.role === "system") {
-      result.push({
-        role: "model",
-        parts: [{ text: "Okay, what do I need to do?" }]
-      });
-    }
-    if (nextMessage?.role === "system") {
-      result.push({
-        role: "user",
-        parts: [
-          {
-            text: "Continue what I said to you last message. Follow these instructions."
-          }
-        ]
-      });
+  return mappedMessages;
+}
+__name(langchainMessageToGeminiMessage, "langchainMessageToGeminiMessage");
+function extractSystemMessages(messages) {
+  let lastSystemMessage;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "system") {
+      lastSystemMessage = messages[i];
+      break;
     }
   }
-  if (result[result.length - 1].role === "model") {
-    result.push({
-      role: "user",
-      parts: [
-        {
-          text: "Continue what I said to you last message. Follow these instructions."
-        }
-      ]
-    });
+  if (lastSystemMessage == null) {
+    return [void 0, messages];
   }
-  return result;
+  const systemMessages = messages.slice(
+    0,
+    messages.indexOf(lastSystemMessage)
+  );
+  return [
+    {
+      role: "user",
+      parts: systemMessages.reduce((acc, cur) => {
+        acc.push(...cur.parts);
+        return acc;
+      }, [])
+    },
+    messages.slice(messages.indexOf(lastSystemMessage))
+  ];
 }
-__name(langchainMessageToGeminiMessage, "langchainMessageToGeminiMessage");
+__name(extractSystemMessages, "extractSystemMessages");
 function partAsType(part) {
   return part;
 }
@@ -355,16 +340,24 @@ var GeminiRequester = class extends import_api.ModelRequester {
       let model = params.model;
       let enabledThinking = null;
       if (model.includes("-thinking") && model.includes("gemini-2.5")) {
-        enabledThinking = !model.includes("-no-thinking");
-        model = model.replace("-no-thinking", "").replace("-thinking", "");
+        enabledThinking = !model.includes("-non-thinking");
+        model = model.replace("-nom-thinking", "").replace("-thinking", "");
+      }
+      const geminiMessages = await langchainMessageToGeminiMessage(
+        params.input,
+        model
+      );
+      const [systemInstruction, modelMessages] = extractSystemMessages(geminiMessages);
+      let thinkingBudget = this._pluginConfig.thinkingBudget ?? -1;
+      if (!enabledThinking && !model.includes("2.5-pro")) {
+        thinkingBudget = 0;
+      } else if (thinkingBudget >= 0 && thinkingBudget < 128) {
+        thinkingBudget = 128;
       }
       const response = await this._post(
         `models/${model}:streamGenerateContent?alt=sse`,
         {
-          contents: await langchainMessageToGeminiMessage(
-            params.input,
-            model
-          ),
+          contents: modelMessages,
           safetySettings: [
             {
               category: "HARM_CATEGORY_HARASSMENT",
@@ -397,10 +390,11 @@ var GeminiRequester = class extends import_api.ModelRequester {
               "gemini-2.0-flash-exp"
             ) && this._pluginConfig.imageGeneration ? ["TEXT", "IMAGE"] : void 0,
             thinkingConfig: enabledThinking != null || this._pluginConfig.includeThoughts ? {
-              thinkingBudget: enabledThinking ? this._pluginConfig.thinkingBudget ?? -1 : -1,
+              thinkingBudget,
               includeThoughts: this._pluginConfig.includeThoughts
             } : void 0
           },
+          systemInstruction: systemInstruction != null ? systemInstruction : void 0,
           tools: params.tools != null || this._pluginConfig.googleSearch || this._pluginConfig.codeExecution || this._pluginConfig.urlContext ? formatToolsToGeminiAITools(
             params.tools ?? [],
             this._pluginConfig,
@@ -413,7 +407,7 @@ var GeminiRequester = class extends import_api.ModelRequester {
       );
       let errorCount = 0;
       let groundingContent = "";
-      let currentGroudingIndex = 0;
+      let currentGroundingIndex = 0;
       await (0, import_sse.checkResponse)(response);
       const readableStream = new ReadableStream({
         async start(controller) {
@@ -444,7 +438,7 @@ var GeminiRequester = class extends import_api.ModelRequester {
               controller.enqueue(part);
             }
             for (const source of candidate.groundingMetadata?.groundingChunks ?? []) {
-              groundingContent += `[^${currentGroudingIndex++}]: [${source.web.title}](${source.web.uri})
+              groundingContent += `[^${currentGroundingIndex++}]: [${source.web.title}](${source.web.uri})
 `;
             }
           }
@@ -714,7 +708,7 @@ var GeminiClient = class extends import_client.PlatformModelAndEmbeddingsClient
         if (model.name.includes("gemini-2.5") && !model.name.includes("pro")) {
           if (!model.name.includes("-thinking")) {
             models.push(
-              { ...info, name: model.name + "-no-thinking" },
+              { ...info, name: model.name + "-nonthinking" },
               { ...info, name: model.name + "-thinking" },
               info
             );
@@ -814,7 +808,7 @@ var Config3 = import_koishi.Schema.intersect([
     googleSearch: import_koishi.Schema.boolean().default(false),
     codeExecution: import_koishi.Schema.boolean().default(false),
     urlContext: import_koishi.Schema.boolean().default(false),
-    thinkingBudget: import_koishi.Schema.number().min(128).max(24576).step(16).default(128),
+    thinkingBudget: import_koishi.Schema.number().min(-1).max(24576).default(-1),
     includeThoughts: import_koishi.Schema.boolean().default(false),
     imageGeneration: import_koishi.Schema.boolean().default(false),
     groundingContentDisplay: import_koishi.Schema.boolean().default(false),

package/lib/index.mjs

@@ -8,14 +8,14 @@ var __commonJS = (cb, mod) => function __require() {
 // src/locales/zh-CN.schema.yml
 var require_zh_CN_schema = __commonJS({
   "src/locales/zh-CN.schema.yml"(exports, module) {
-    module.exports = { $inner: [{}, { $desc: "请求选项", platform: "适配器的平台名。（不懂请不要修改）", apiKeys: { $inner: ["Gemini 的 API Key", "Gemini API 的请求地址"], $desc: "Gemini 的 API Key 和请求地址列表。" } }, { $desc: "模型配置", maxTokens: "输入的最大上下文 Token（16~2097000，必须是 16 的倍数）。注意：仅当您使用的模型最大 Token 为 8000 及以上时，才建议设置超过 2000 token。", temperature: "回复的随机性程度，数值越高，回复越随机（范围：0~2）。", googleSearch: "为模型启用谷歌搜索。", thinkingBudget: "思考预算，范围：(-1~24576)，设置的数值越大，思考时花费的 Token 越多,-1 为动态思考，0 为不思考。目前仅支持 gemini 2.5 系列模型。", groundingContentDisplay: "是否显示谷歌搜索结果。", imageGeneration: "为模型启用图像生成。目前仅支持 `gemini-2.0-flash-exp` 模型。", searchThreshold: "搜索的[置信度阈值](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval)，范围：0~1，设置的数值越低，则越倾向于使用谷歌搜索。（仅支持 `gemini-1.5` 系列模型。gemini 2.0 模型起使用动态的工具调用）", includeThoughts: "是否获取模型的思考内容。", codeExecution: "为模型启用代码执行工具。", urlContext: "为模型启用 URL 内容获取工具。" }] };
+    module.exports = { $inner: [{}, { $desc: "请求选项", platform: "适配器的平台名。（不懂请不要修改）", apiKeys: { $inner: ["Gemini 的 API Key", "Gemini API 的请求地址"], $desc: "Gemini 的 API Key 和请求地址列表。" } }, { $desc: "模型配置", maxTokens: "输入的最大上下文 Token（16~2097000，必须是 16 的倍数）。注意：仅当您使用的模型最大 Token 为 8000 及以上时，才建议设置超过 2000 token。", temperature: "回复的随机性程度，数值越高，回复越随机（范围：0~2）。", googleSearch: "为模型启用谷歌搜索。", thinkingBudget: "思考预算，范围：(-1~24576)，设置的数值越大，思考时花费的 Token 越多,-1 为动态思考。目前仅支持 gemini 2.5 系列模型。", groundingContentDisplay: "是否显示谷歌搜索结果。", imageGeneration: "为模型启用图像生成。目前仅支持 `gemini-2.0-flash-exp` 模型。", searchThreshold: "搜索的[置信度阈值](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval)，范围：0~1，设置的数值越低，则越倾向于使用谷歌搜索。（仅支持 `gemini-1.5` 系列模型。gemini 2.0 模型起使用动态的工具调用）", includeThoughts: "是否获取模型的思考内容。", codeExecution: "为模型启用代码执行工具。", urlContext: "为模型启用 URL 内容获取工具。" }] };
   }
 });
 
 // src/locales/en-US.schema.yml
 var require_en_US_schema = __commonJS({
   "src/locales/en-US.schema.yml"(exports, module) {
-    module.exports = { $inner: [{}, { $desc: "API Configuration", platform: "Adapter platform name. (Do not modify if you do not understand)", apiKeys: { $inner: ["Gemini API Key", "Gemini API Endpoint (optional)"], $desc: "Gemini API access credentials" } }, { $desc: "Model Parameters", maxTokens: "Max output tokens (16-2097000, multiple of 16). >2000 for 8k+ models", temperature: "Sampling temperature (0-2). Higher: more random, Lower: more deterministic", googleSearch: "Enable Google search", thinkingBudget: "Thinking budget (-1-24576). (-1: dynamic thinking, 0: no thinking) Higher: more tokens spent on thinking. Currently only supports `gemini-2.5` series models.", groundingContentDisplay: "Enable display of search results", imageGeneration: "Enable image generation (only for `gemini-2.0-flash-exp` model)", searchThreshold: "Search confidence [threshold](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval) (0-1). Lower: more likely to use Google search", includeThoughts: "Enable retrieval of model thoughts", codeExecution: "Enable code execution tool", urlContext: "Enable URL context retrieval tool" }] };
+    module.exports = { $inner: [{}, { $desc: "API Configuration", platform: "Adapter platform name. (Do not modify if you do not understand)", apiKeys: { $inner: ["Gemini API Key", "Gemini API Endpoint (optional)"], $desc: "Gemini API access credentials" } }, { $desc: "Model Parameters", maxTokens: "Max output tokens (16-2097000, multiple of 16). >2000 for 8k+ models", temperature: "Sampling temperature (0-2). Higher: more random, Lower: more deterministic", googleSearch: "Enable Google search", thinkingBudget: "Thinking budget (-1-24576). (0: dynamic thinking) Higher: more tokens spent on thinking. Currently only supports `gemini-2.5` series models.", groundingContentDisplay: "Enable display of search results", imageGeneration: "Enable image generation (only for `gemini-2.0-flash-exp` model)", searchThreshold: "Search confidence [threshold](https://ai.google.dev/gemini-api/docs/grounding?lang=rest#dynamic-retrieval) (0-1). Lower: more likely to use Google search", includeThoughts: "Enable retrieval of model thoughts", codeExecution: "Enable code execution tool", urlContext: "Enable URL context retrieval tool" }] };
   }
 });
 
@@ -59,7 +59,7 @@ import {
 } from "@langchain/core/messages";
 import { zodToJsonSchema } from "zod-to-json-schema";
 async function langchainMessageToGeminiMessage(messages, model) {
-  const mappedMessage = await Promise.all(
+  const mappedMessages = await Promise.all(
     messages.map(async (rawMessage) => {
       const role = messageTypeToGeminiRole(rawMessage.getType());
       if (role === "function" || rawMessage.additional_kwargs?.function_call != null) {
@@ -73,15 +73,15 @@ async function langchainMessageToGeminiMessage(messages, model) {
                   name: rawMessage.name,
                   content: (() => {
                     try {
-                      const result3 = JSON.parse(
+                      const result2 = JSON.parse(
                         rawMessage.content
                       );
-                      if (typeof result3 === "string") {
+                      if (typeof result2 === "string") {
                         return {
-                          response: result3
+                          response: result2
                         };
                       } else {
-                        return result3;
+                        return result2;
                       }
                     } catch (e) {
                       return {
@@ -95,15 +95,15 @@ async function langchainMessageToGeminiMessage(messages, model) {
                 name: rawMessage.additional_kwargs.function_call.name,
                 args: (() => {
                   try {
-                    const result3 = JSON.parse(
+                    const result2 = JSON.parse(
                       rawMessage.additional_kwargs.function_call.arguments
                     );
-                    if (typeof result3 === "string") {
+                    if (typeof result2 === "string") {
                       return {
-                        input: result3
+                        input: result2
                       };
                     } else {
-                      return result3;
+                      return result2;
                     }
                   } catch (e) {
                     return {
@@ -117,7 +117,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
         };
       }
       const images = rawMessage.additional_kwargs.images;
-      const result2 = {
+      const result = {
         role,
         parts: [
           {
@@ -129,7 +129,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
         for (const image of images) {
           const mineType = image.split(";")?.[0]?.split(":")?.[1];
           const data = image.replace(/^data:image\/\w+;base64,/, "");
-          result2.parts.push({
+          result.parts.push({
             inline_data: {
               // base64 image match type
               data,
@@ -137,7 +137,7 @@ async function langchainMessageToGeminiMessage(messages, model) {
             }
           });
         }
-        result2.parts = result2.parts.filter((uncheckedPart) => {
+        result.parts = result.parts.filter((uncheckedPart) => {
           const part = partAsTypeCheck(
             uncheckedPart,
             (part2) => part2["text"] != null
@@ -145,54 +145,39 @@ async function langchainMessageToGeminiMessage(messages, model) {
           return part == null || part.text.length > 0;
         });
       }
-      return result2;
+      return result;
     })
   );
-  const result = [];
-  for (let i = 0; i < mappedMessage.length; i++) {
-    const message = mappedMessage[i];
-    if (message.role !== "system") {
-      result.push(message);
-      continue;
-    }
-    result.push({
-      role: "user",
-      parts: message.parts
-    });
-    const nextMessage = mappedMessage?.[i + 1];
-    if (nextMessage?.role === "model") {
-      continue;
-    }
-    if (nextMessage?.role === "user" || nextMessage?.role === "system") {
-      result.push({
-        role: "model",
-        parts: [{ text: "Okay, what do I need to do?" }]
-      });
-    }
-    if (nextMessage?.role === "system") {
-      result.push({
-        role: "user",
-        parts: [
-          {
-            text: "Continue what I said to you last message. Follow these instructions."
-          }
-        ]
-      });
+  return mappedMessages;
+}
+__name(langchainMessageToGeminiMessage, "langchainMessageToGeminiMessage");
+function extractSystemMessages(messages) {
+  let lastSystemMessage;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "system") {
+      lastSystemMessage = messages[i];
+      break;
     }
   }
-  if (result[result.length - 1].role === "model") {
-    result.push({
-      role: "user",
-      parts: [
-        {
-          text: "Continue what I said to you last message. Follow these instructions."
-        }
-      ]
-    });
+  if (lastSystemMessage == null) {
+    return [void 0, messages];
   }
-  return result;
+  const systemMessages = messages.slice(
+    0,
+    messages.indexOf(lastSystemMessage)
+  );
+  return [
+    {
+      role: "user",
+      parts: systemMessages.reduce((acc, cur) => {
+        acc.push(...cur.parts);
+        return acc;
+      }, [])
+    },
+    messages.slice(messages.indexOf(lastSystemMessage))
+  ];
 }
-__name(langchainMessageToGeminiMessage, "langchainMessageToGeminiMessage");
+__name(extractSystemMessages, "extractSystemMessages");
 function partAsType(part) {
   return part;
 }
@@ -338,16 +323,24 @@ var GeminiRequester = class extends ModelRequester {
       let model = params.model;
       let enabledThinking = null;
       if (model.includes("-thinking") && model.includes("gemini-2.5")) {
-        enabledThinking = !model.includes("-no-thinking");
-        model = model.replace("-no-thinking", "").replace("-thinking", "");
+        enabledThinking = !model.includes("-non-thinking");
+        model = model.replace("-nom-thinking", "").replace("-thinking", "");
+      }
+      const geminiMessages = await langchainMessageToGeminiMessage(
+        params.input,
+        model
+      );
+      const [systemInstruction, modelMessages] = extractSystemMessages(geminiMessages);
+      let thinkingBudget = this._pluginConfig.thinkingBudget ?? -1;
+      if (!enabledThinking && !model.includes("2.5-pro")) {
+        thinkingBudget = 0;
+      } else if (thinkingBudget >= 0 && thinkingBudget < 128) {
+        thinkingBudget = 128;
       }
       const response = await this._post(
         `models/${model}:streamGenerateContent?alt=sse`,
         {
-          contents: await langchainMessageToGeminiMessage(
-            params.input,
-            model
-          ),
+          contents: modelMessages,
           safetySettings: [
             {
               category: "HARM_CATEGORY_HARASSMENT",
@@ -380,10 +373,11 @@ var GeminiRequester = class extends ModelRequester {
               "gemini-2.0-flash-exp"
             ) && this._pluginConfig.imageGeneration ? ["TEXT", "IMAGE"] : void 0,
             thinkingConfig: enabledThinking != null || this._pluginConfig.includeThoughts ? {
-              thinkingBudget: enabledThinking ? this._pluginConfig.thinkingBudget ?? -1 : -1,
+              thinkingBudget,
               includeThoughts: this._pluginConfig.includeThoughts
             } : void 0
           },
+          systemInstruction: systemInstruction != null ? systemInstruction : void 0,
           tools: params.tools != null || this._pluginConfig.googleSearch || this._pluginConfig.codeExecution || this._pluginConfig.urlContext ? formatToolsToGeminiAITools(
             params.tools ?? [],
             this._pluginConfig,
@@ -396,7 +390,7 @@ var GeminiRequester = class extends ModelRequester {
       );
       let errorCount = 0;
       let groundingContent = "";
-      let currentGroudingIndex = 0;
+      let currentGroundingIndex = 0;
       await checkResponse(response);
       const readableStream = new ReadableStream({
         async start(controller) {
@@ -427,7 +421,7 @@ var GeminiRequester = class extends ModelRequester {
               controller.enqueue(part);
             }
             for (const source of candidate.groundingMetadata?.groundingChunks ?? []) {
-              groundingContent += `[^${currentGroudingIndex++}]: [${source.web.title}](${source.web.uri})
+              groundingContent += `[^${currentGroundingIndex++}]: [${source.web.title}](${source.web.uri})
 `;
             }
           }
@@ -697,7 +691,7 @@ var GeminiClient = class extends PlatformModelAndEmbeddingsClient {
         if (model.name.includes("gemini-2.5") && !model.name.includes("pro")) {
           if (!model.name.includes("-thinking")) {
             models.push(
-              { ...info, name: model.name + "-no-thinking" },
+              { ...info, name: model.name + "-nonthinking" },
               { ...info, name: model.name + "-thinking" },
               info
             );
@@ -797,7 +791,7 @@ var Config3 = Schema.intersect([
     googleSearch: Schema.boolean().default(false),
     codeExecution: Schema.boolean().default(false),
     urlContext: Schema.boolean().default(false),
-    thinkingBudget: Schema.number().min(128).max(24576).step(16).default(128),
+    thinkingBudget: Schema.number().min(-1).max(24576).default(-1),
     includeThoughts: Schema.boolean().default(false),
     imageGeneration: Schema.boolean().default(false),
     groundingContentDisplay: Schema.boolean().default(false),

package/lib/utils.d.ts

@@ -3,6 +3,7 @@ import { StructuredTool } from '@langchain/core/tools';
 import { ChatCompletionFunction, ChatCompletionResponseMessage, ChatCompletionResponseMessageRoleEnum, ChatPart } from './types';
 import { Config } from '.';
 export declare function langchainMessageToGeminiMessage(messages: BaseMessage[], model?: string): Promise<ChatCompletionResponseMessage[]>;
+export declare function extractSystemMessages(messages: ChatCompletionResponseMessage[]): [ChatCompletionResponseMessage, ChatCompletionResponseMessage[]];
 export declare function partAsType<T extends ChatPart>(part: ChatPart): T;
 export declare function partAsTypeCheck<T extends ChatPart>(part: ChatPart, check: (part: ChatPart & unknown) => boolean): T | undefined;
 export declare function formatToolsToGeminiAITools(tools: StructuredTool[], config: Config, model: string): Record<string, any>;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "koishi-plugin-chatluna-google-gemini-adapter",
   "description": "google-gemini adapter for chatluna",
-  "version": "1.2.6",
+  "version": "1.2.8",
   "main": "lib/index.cjs",
   "module": "lib/index.mjs",
   "typings": "lib/index.d.ts",
@@ -72,7 +72,7 @@
   },
   "peerDependencies": {
     "koishi": "^4.18.7",
-    "koishi-plugin-chatluna": "^1.3.0-alpha.4"
+    "koishi-plugin-chatluna": "^1.3.0-alpha.7"
   },
   "koishi": {
     "description": {