feat: v2.2.63 — 한국어 오타 최소화 (채팅 temperature 설정 + anti-glitch 샘플링)

- streamer.ts: LM Studio SDK 호출에 topP/topK/minP/repeatPenalty 추가
  — 저확률 오답 토큰을 잘라 한글 음절 깨짐(붕괴→붕점) 억제
- 채팅 기본 temperature 0.7 → 0.3 (분석/업무형 답변 안정화)
- 신규 설정 g1nation.chatTemperature — Settings 패널 '고급' 섹션에서
  조절 가능 (config.ts / settingsPanelProvider / settings-panel.html+js)

chronicle 기록(ADR-0022, ADR-0023) 포함.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/agent.ts

@@ -437,7 +437,7 @@ export class AgentExecutor {
             brainEnabled = false,
             loopDepth = 0,
             visionContent,
-            temperature = 0.7,
+            temperature = getConfig().chatTemperature,
             systemPrompt = getSystemPrompt()
         } = options;
         const { ollamaUrl, defaultModel: configDefaultModel, timeout, multiAgentEnabled } = getConfig();

src/config.ts

@@ -24,6 +24,8 @@ export interface IAgentConfig {
     activeBrainId: string;
     maxContextSize: number;
     maxAutoSteps: number;
+    /** 채팅 응답 생성 temperature. 낮을수록 한국어 오타·깨진 토큰이 줄어든다. */
+    chatTemperature: number;
     dryRun: boolean;
     multiAgentEnabled: boolean;
     memoryEnabled: boolean;
@@ -204,6 +206,7 @@ export function getConfig(): IAgentConfig {
         activeBrainId: activeBrain.id,
         maxContextSize: cfg.get<number>('maxContextSize', 12000),
         maxAutoSteps: cfg.get<number>('maxAutoSteps', 50),
+        chatTemperature: Math.min(2, Math.max(0, cfg.get<number>('chatTemperature', 0.3))),
         dryRun: cfg.get<boolean>('dryRun', false),
         multiAgentEnabled: cfg.get<boolean>('multiAgentEnabled', false),
         memoryEnabled: cfg.get<boolean>('memoryEnabled', true),

src/features/settings/settingsPanelProvider.ts

@@ -82,6 +82,7 @@ interface SettingsState {
         multiAgentEnabled: boolean;
         maxAutoSteps: number;
         maxContextSize: number;
+        chatTemperature: number;
     };
     datacollect: {
         bridgeUrl: string;
@@ -581,6 +582,9 @@ export class SettingsPanelProvider implements vscode.WebviewViewProvider {
         if (typeof msg.maxContextSize === 'number' && Number.isFinite(msg.maxContextSize)) {
             await this._safeConfigUpdate('maxContextSize', Math.max(1000, Math.floor(msg.maxContextSize)));
         }
+        if (typeof msg.chatTemperature === 'number' && Number.isFinite(msg.chatTemperature)) {
+            await this._safeConfigUpdate('chatTemperature', Math.max(0, Math.min(2, msg.chatTemperature)));
+        }
     }
 
     // ────────────── Datacollect (slash 명령) ──────────────
@@ -652,6 +656,7 @@ export class SettingsPanelProvider implements vscode.WebviewViewProvider {
                 multiAgentEnabled: cfg.get<boolean>('multiAgentEnabled', false),
                 maxAutoSteps: cfg.get<number>('maxAutoSteps', 50) ?? 50,
                 maxContextSize: cfg.get<number>('maxContextSize', 32000) ?? 32000,
+                chatTemperature: cfg.get<number>('chatTemperature', 0.3) ?? 0.3,
             },
             datacollect: {
                 bridgeUrl: cfg.get<string>('datacollectBridgeUrl', '') || '',

src/lmstudio/streamer.ts

@@ -75,6 +75,15 @@ export class LMStudioStreamer implements IChatStreamer {
             const prediction = (model as any).respond(req.messages, {
                 temperature: req.temperature,
                 maxTokens: req.maxTokens ?? 4096,
+                // Glitch suppression: a small / quantized model samples wrong
+                // neighbour tokens (Korean syllable corruption like 붕괴→붕점,
+                // 핵심→핵점) when the distribution is left wide. A tight nucleus
+                // + top-k and a min-p floor cut the low-probability tail;
+                // repeatPenalty curbs stutter (것입니다서입니다).
+                topPSampling: 0.9,
+                topKSampling: 20,
+                minPSampling: 0.05,
+                repeatPenalty: 1.1,
                 // Safety net: if our own token budgeting still underestimated and the prompt
                 // exceeds the model's context window, decide whether the SDK should fail
                 // loudly (stopAtLimit — default) or silently drop content.