20260126

.vscode/launch.json

@@ -1,6 +1,6 @@
 {
   "version": "0.2.0",
-  "configurations": [
+  "configurations": [    
     {
       "type": "node",
       "request": "launch",

README.md

@@ -8,6 +8,9 @@ English / [Japanese](./README_jp.md)
 > The "Generator" setting will be moved to the Avatar Settings section, and the existing "Generator" will be renamed to "Model."  
 > We plan to clarify the relationship between Avatar AI, which is centered around context and is composed of a collection of multiple generators.  
 
+> As a temporary measure, we have added cutoffChatLimit to the LLM settings.  
+> This will reduce token consumption by forcibly truncating the context.  
+
 > Supported LM Studio API (OpenAI compatibility mode).
 
 > Note: If you are using a previous version and are experiencing unstable operation, please initialize the settings by going to SystemSetting > initialize.  

README_jp.md

@@ -10,6 +10,9 @@ Japanese / [English](./README.md)
 > 「ジェネレーター」の設定はアバター設定側に移動し、既存の「ジェネレーター」は「モデル」に変更します。  
 > アバターAIがコンテキストを中心として複数のジェネレーターの集合体で構成される関係性をより明確化する予定です。
 
+> 暫定的な対策としてLLM設定に cutoffChatLimit を追加しました。  
+> 強制的にコンテキストを切り詰めることでトークンの消費量を抑制します。    
+
 > LM Studio API (OpenAI互換モード)をサポートしました。
 
 > Note: 以前の版を使っている方で、動作が不安定になった方はSystemSetting > initialize で一旦設定の初期化をお願いします。  

package.json

@@ -1,7 +1,7 @@
 {
   "name": "avatar-shell",
   "description": "Avatar/Visually-based MCP client",
-  "version": "0.2.5",
+  "version": "0.2.6",
   "license": "Apache-2.0",
   "private": true,
   "type": "module",
@@ -84,5 +84,8 @@
     "sound-play": "^1.1.0",
     "ws": "^8.18.2",
     "zod": "^3.25.34"
+  },
+  "volta": {
+    "node": "24.13.0"
   }
 }

packages/common/DefGenerators.ts

@@ -105,6 +105,7 @@ export const ContextGeneratorSettingSchema = Schema.partial(Schema.Struct({
   toContext:Schema.optional(AsContextLinesSchema),
   noTool: Schema.optional(Schema.Boolean),
   useModel: Schema.optional(Schema.String),
+  cutoffChatLimit: Schema.optional(Schema.Number),
   debug: Schema.Any,  //  デバッグ用汎用
 }))
 
@@ -174,7 +175,8 @@ Generator 全体設定(api keyなど全体で設定するもの
 
 const generatorCommonConfigSchema = Schema.Struct({
   maxTokenThreshold:Schema.Number,
-  summarizePrompt:Schema.String
+  summarizePrompt:Schema.String,
+  cutoffChatLimit:Schema.optional(Schema.Number)
 })
 
 //    openAi  openAiText, openAiImage, openAiVoice

packages/main/package.json

@@ -27,5 +27,8 @@
     "electron-devtools-installer": "4.0.0",
     "typescript": "5.8.3",
     "vite": "7.0.5"
+  },
+  "volta": {
+    "node": "24.13.0"
   }
 }

packages/main/src/AvatarState.ts

@@ -855,7 +855,7 @@ export class AvatarState {
     });
   }
 
-  MaxGen = 2;  //  TODO 世代の最適値は
+  MaxGen = 20;  //  TODO 世代の最適値は
 
   enterInner(inner: GenInner, addToBuffer = true) {
     const it = this;
@@ -893,6 +893,7 @@ export class AvatarState {
           // }
           yield* it.resetPrevBuffer();
           it.clearStreamingText();
+          //  gen.setContextDelimiter() ここは次に回すgenが未確定だからデリミダ付けられないか。。
           return;  //  func の無限ループを防ぐ
         }
         //  Generator処理
@@ -922,6 +923,7 @@ export class AvatarState {
           yield* Queue.offerAll(it.outerQueue, io);
         } else {
           it.clearStreamingText();
+          gen.setContextDelimiter()
         }
       }),
       step: b => loop,
@@ -1057,11 +1059,12 @@ export class AvatarState {
         toGenerator: gen,
         input: message,
         genNum: 1,  //  この入力はまだcontextに追加されていないので1から開始して追加させる
-        setting: {
-          toClass: setting.toClass,
-          toRole: setting.toRole,
-          toContext: setting.toContext,
-        },
+        setting: setting,
+        // setting: {
+        //   toClass: setting.toClass,
+        //   toRole: setting.toRole,
+        //   toContext: setting.toContext,
+        // },
       }, addToBuffer);
       // console.log('start loop');
 

packages/main/src/generators/ClaudeGenerator.ts

@@ -37,9 +37,9 @@ export abstract class ClaudeBaseGenerator extends ContextGenerator {
   protected abstract genName: GeneratorProvider;
   protected maxModelContextSize = 200000;  //  TODO Claudeの最大コンテキスト長も20Kぐらいらしい
 
-  protected get previousContexts() {
-    return this.previousNativeContexts as Anthropic.Messages.MessageParam[];
-  }
+  // protected get previousContexts() {
+  //   return this.previousNativeContexts as Anthropic.Messages.MessageParam[];
+  // }
 
   static generatorInfo: ContextGeneratorInfo = {
     usePreviousContext: true,
@@ -58,6 +58,10 @@ export abstract class ClaudeBaseGenerator extends ContextGenerator {
     });
   }
 
+  getPreviousNativeContexts():(Anthropic.Messages.MessageParam | null)[] {
+    return this.previousNativeContexts as (Anthropic.Messages.MessageParam | null)[];
+  }
+
   // filterToolRes(value: ContentBlock ) {
   //   //  sysConfigでexperimental.mcpUiFilterDisabledがtrueの場合フィルタしない
   //   if(this.sysSetting.experimental.mcpUiFilterDisabled) return [value];
@@ -279,7 +283,32 @@ export class ClaudeTextGenerator extends ClaudeBaseGenerator {
     return Effect.gen(function* () {
       //  TODO prev contextを抽出(AsMessage履歴から合成またはコンテキストキャッシュから再生)
       const prevMake = yield* it.makePreviousContext(avatarState,current);
-      const prev = Array.from(it.previousContexts)  // ユーザからのmcpExternalで
+      let prev: Anthropic.Messages.MessageParam[] = []
+      const lessCutoff = it.sysSetting.generators.anthropic.common?.cutoffChatLimit || Number.MAX_SAFE_INTEGER;
+      if (lessCutoff !== Number.MAX_SAFE_INTEGER) {
+        const sum = Array.from(it.getPreviousNativeContexts()).reduce((previousValue, currentValue) => {
+          if(currentValue === null){
+            previousValue.list.push(previousValue.buf)
+            return {list:previousValue.list,buf:[]}
+          }
+          return {list:previousValue.list,buf:previousValue.buf.concat(currentValue)}
+        },{list:[] as Anthropic.Messages.MessageParam[][],buf:[] as Anthropic.Messages.MessageParam[]})
+        if(sum.buf.length > 0) {
+          sum.list.push(sum.buf)
+        }
+
+        const cutoff = sum.list.reverse().reduce((previousValue, currentValue) => {
+          if (previousValue.count <= 0) {
+            return previousValue;
+          }
+          const next = previousValue.out.concat(currentValue.reverse());
+          previousValue.count-= currentValue.length
+          return {out:next,count:previousValue.count}
+        },{out:[] as Anthropic.Messages.MessageParam[][],count:lessCutoff})
+        prev = cutoff.out.reverse().flat()
+      } else {
+        prev = Array.from(it.getPreviousNativeContexts()).filter((value):value is  Anthropic.Messages.MessageParam => value !== null);  // ユーザからのmcpExternalで
+      }
       //  入力current GenInnerからcurrent contextを調整(input textまはたMCP responses)
       const mes = yield* it.makeCurrentContext(current);
 
@@ -317,11 +346,11 @@ export class ClaudeTextGenerator extends ClaudeBaseGenerator {
         },
       });
 
-      it.previousContexts.push({
+      it.getPreviousNativeContexts().push({
         role: mes.role,
         content:mes.content,
       })
-      it.previousContexts.push({
+      it.getPreviousNativeContexts().push({
         role: message.role,
         content:message.content,
       } as Anthropic.Messages.MessageParam)

packages/main/src/generators/ContextGenerator.ts

@@ -74,6 +74,14 @@ export abstract class ContextGenerator {
     avatarState.clearStreamingText()
   }
 
+  /**
+   * prevContextBufferに分離可能なデリミダを追加する
+   * 今のところnullを使う
+   */
+  setContextDelimiter() {
+    this.previousNativeContexts.push(null);
+  }
+
   filterForLlmPrevContext(asMes:AsMessage[],current?:AsMessage) {
     //  TODO ここは妥当な条件を再検討が必要。。
     //  currentを含まないこと、currentよりもtickが古いこと

packages/main/src/generators/GeminiGenerator.ts

@@ -55,7 +55,7 @@ export abstract class GeminiBaseGenerator extends ContextGenerator {
   }
 
   protected get previousContexts() {
-    return this.previousNativeContexts as Content[];
+    return this.previousNativeContexts as (Content | null)[];
   }
 
   constructor(sysConfig: SysConfig, settings?: GeminiSettings) {
@@ -155,7 +155,7 @@ export abstract class GeminiBaseGenerator extends ContextGenerator {
         //  geminiの場合、画像ファイルはinput_imageとしてbase64で送る
         const media = yield* DocService.readDocMedia(current.input.content.mediaUrl);
         const b1 = yield* it.shrinkImage(Buffer.from(media, 'base64').buffer, it.geminiSettings?.inWidth);
-        const blob = new Blob([b1], {type: 'image/png'});
+        const blob = new Blob([new Uint8Array(b1).buffer], {type: 'image/png'});
         const myfile = yield* Effect.tryPromise(() => it.ai.files.upload({
           file: blob,
           config: {mimeType: 'image/png'},
@@ -211,7 +211,33 @@ export class GeminiTextGenerator extends GeminiBaseGenerator {
     return Effect.gen(function* () {
       //  prev contextを抽出(AsMessage履歴から合成またはコンテキストキャッシュから再生)
       const prevMake = yield* it.makePreviousContext(avatarState,current);
-      const prev = Array.from(it.previousContexts)
+      let prev:Content[] = []
+      const lessCutoff = it.sysSetting.generators.gemini.common?.cutoffChatLimit || Number.MAX_SAFE_INTEGER;
+      if (lessCutoff !== Number.MAX_SAFE_INTEGER) {
+        const sum = Array.from(it.previousContexts).reduce((previousValue, currentValue) => {
+          if(currentValue === null){
+            previousValue.list.push(previousValue.buf)
+            return {list:previousValue.list,buf:[]}
+          }
+          return {list:previousValue.list,buf:previousValue.buf.concat(currentValue)}
+        },{list:[] as Content[][],buf:[] as Content[]})
+        if(sum.buf.length > 0) {
+          sum.list.push(sum.buf)
+        }
+
+        const cutoff = sum.list.reverse().reduce((previousValue, currentValue) => {
+          if (previousValue.count <= 0) {
+            return previousValue;
+          }
+          const next = previousValue.out.concat(currentValue.reverse());
+          previousValue.count-= currentValue.length
+          return {out:next,count:previousValue.count}
+        },{out:[] as Content[][],count:lessCutoff})
+        prev = cutoff.out.reverse().flat()
+      } else {
+        prev = Array.from(it.previousContexts).filter((value):value is  Content => value !== null);  // ユーザからのmcpExternalで
+      }
+      // const prev = Array.from(it.previousContexts)
       //  入力current GenInnerからcurrent contextを調整(input textまはたMCP responses)
       const mes = yield* it.makeCurrentContext(current);
 

packages/main/src/generators/LmStudioGenerator.ts

@@ -47,6 +47,10 @@ export abstract class LmStudioBaseGenerator extends ContextGenerator {
   protected override maxModelContextSize = -1;
   protected baseUrl=''
 
+  getPreviousNativeContexts():(ResponseInputItem | null)[] {
+    return this.previousNativeContexts as (ResponseInputItem | null)[];
+  }
+
   static generatorInfo: ContextGeneratorInfo = {
     usePreviousContext: true,
     defaultPrevContextSize: 100,
@@ -263,7 +267,35 @@ export class LmStudioTextGenerator extends LmStudioBaseGenerator {
 
       //  prev contextを抽出(AsMessage履歴から合成またはコンテキストキャッシュから再生)
       const prevMake = yield* it.makePreviousContext(avatarState, current);
-      const prev = Array.from(it.previousNativeContexts).filter(value => !(value.type === 'message' && value.role === 'developer'))
+      let prev:ResponseInputItem[] = []
+      const lessCutoff = it.sysSetting.generators.lmStudio.common?.cutoffChatLimit || Number.MAX_SAFE_INTEGER;
+      // const lessCutoff = Math.min(current.setting?.cutoffChatLimit || Number.MAX_SAFE_INTEGER, it.lmStudioSettings?.cutoffChatLimit || Number.MAX_SAFE_INTEGER);
+      if (lessCutoff !== Number.MAX_SAFE_INTEGER) {
+        const sum = Array.from(it.getPreviousNativeContexts()).reduce((previousValue, currentValue) => {
+          if(currentValue === null){
+            previousValue.list.push(previousValue.buf)
+            return {list:previousValue.list,buf:[]}
+          }
+          return {list:previousValue.list,buf:previousValue.buf.concat(currentValue)}
+        },{list:[] as ResponseInputItem[][],buf:[] as ResponseInputItem[]})
+        if(sum.buf.length > 0) {
+          sum.list.push(sum.buf)
+        }
+
+        const cutoff = sum.list.reverse().reduce((previousValue, currentValue) => {
+          if (previousValue.count <= 0) {
+            return previousValue;
+          }
+          const next = previousValue.out.concat(currentValue.reverse());
+          previousValue.count-= currentValue.length
+          return {out:next,count:previousValue.count}
+        },{out:[] as ResponseInputItem[][],count:lessCutoff})
+        prev = cutoff.out.reverse().flat().filter(value => !(value.type === 'message' && value.role === 'developer'))
+      } else {
+        prev = Array.from(it.getPreviousNativeContexts()).filter((value):value is ResponseInputItem  => value !== null && !(value.type === 'message' && value.role === 'developer'))
+        // prev = Array.from(it.getPreviousNativeContexts()).filter((value):value is  Anthropic.Messages.MessageParam => value !== null);  // ユーザからのmcpExternalで
+      }
+      // const prev = Array.from(it.getPreviousNativeContexts()).filter((value):value is ResponseInputItem  => value !== null && !(value.type === 'message' && value.role === 'developer'))
       //  TODO prevMakeとprevの差分チェックは後々必要
       //  入力current GenInnerからcurrent contextを調整(input textまはたMCP responses)
       const mes = yield* it.makeCurrentContext(current);
@@ -290,7 +322,7 @@ export class LmStudioTextGenerator extends LmStudioBaseGenerator {
         // store:true,
       };
       mes.forEach(a => {
-        it.previousNativeContexts.push(a)
+        it.getPreviousNativeContexts().push(a)
       })
 
       const res = yield* Effect.tryPromise({
@@ -328,7 +360,7 @@ export class LmStudioTextGenerator extends LmStudioBaseGenerator {
       ).map(a => a.response.output).flat();
       console.log('responseOut:', JSON.stringify(responseOut));
       responseOut.forEach(a => {
-        it.previousNativeContexts.push(a) //  TODO ResponseOutputItemをResponseInputItemに変換する必要があるケースがあったはず。。
+        it.getPreviousNativeContexts().push(a) //  TODO ResponseOutputItemをResponseInputItemに変換する必要があるケースがあったはず。。
       })
       const textOut = responseOut.filter(b => b.type === 'message').map((b: ResponseOutputMessage) => {
         //  TODO mesIdは1件のはず?

packages/main/src/generators/OllamaGenerator.ts

@@ -25,10 +25,11 @@ export class OllamaTextGenerator extends ContextGenerator {
   protected model = 'llama3.2';
   private ollama: Ollama;
   protected systemPrompt?:Message[];
+  protected ollamaSettings: ContextGeneratorSetting | undefined;
   //  TODO ollamaの場合最大コンテキストサイズの取得は?
 
   protected get previousContexts() {
-    return this.previousNativeContexts as Message[];
+    return this.previousNativeContexts as (Message | null)[];
   }
 
   static make(sysConfig:SysConfig, settings?: ContextGeneratorSetting) {
@@ -37,6 +38,7 @@ export class OllamaTextGenerator extends ContextGenerator {
 
   constructor(setting: SysConfig,settings?: ContextGeneratorSetting) {
     super(setting);
+    this.ollamaSettings = settings;
     this.model = settings?.useModel || setting.generators.ollama?.model || 'llama3.1';
     this.ollama = new Ollama({
       host: setting.generators.ollama?.host || 'http://localhost:11434',
@@ -64,7 +66,33 @@ export class OllamaTextGenerator extends ContextGenerator {
           images:undefined, //  TODO 画像は送るべきか?
         } as Message]
       })
-      const prev = Array.from(it.previousContexts)
+      let prev:Message[] = []
+      const lessCutoff = it.sysSetting.generators.ollama.common?.cutoffChatLimit || Number.MAX_SAFE_INTEGER;
+      if (lessCutoff !== Number.MAX_SAFE_INTEGER) {
+        const sum = Array.from(it.previousContexts).reduce((previousValue, currentValue) => {
+          if(currentValue === null){
+            previousValue.list.push(previousValue.buf)
+            return {list:previousValue.list,buf:[]}
+          }
+          return {list:previousValue.list,buf:previousValue.buf.concat(currentValue)}
+        },{list:[] as Message[][],buf:[] as Message[]})
+        if(sum.buf.length > 0) {
+          sum.list.push(sum.buf)
+        }
+        const cutoff = sum.list.reverse().reduce((previousValue, currentValue) => {
+          if (previousValue.count <= 0) {
+            return previousValue;
+          }
+          const next = previousValue.out.concat(currentValue.reverse());
+          previousValue.count-= currentValue.length
+          return {out:next,count:previousValue.count}
+        },{out:[] as Message[][],count:lessCutoff})
+        prev = cutoff.out.reverse().flat()
+      } else {
+        prev = Array.from(it.previousContexts).filter((value):value is  Message => value !== null);  // ユーザからのmcpExternalで
+      }
+
+      // const prev = Array.from(it.previousContexts)
       //  入力current GenInnerからcurrent contextを調整(input textまはたMCP responses)
       const mes:Message = { role: 'user', content: '' };
       if (current.input?.content.text) {