feat: models.profiles[<model>].think — 모델별 think 모드 오버라이드

glm-5.3-flash:cloud 실측(2026-09-22): 오케스트레이터 턴(multiAgentActive)에서
think:true로 돌며 라운드당 reasoning 3~6K 토큰 소모(뉴스 요약 턴 최종 호출
5,868토큰, 72초) — gemma4는 같은 턴을 ~100토큰에 처리. completion 중앙값 331 vs
93(3.5배). 플래시인데도 체감이 느린 진범은 속도가 아니라 출력량.

- server.ts에 getModelProfileThink 신설(forceToolChoice/fixedNumCtx와 같은
  프로파일 패턴) — boolean 또는 high/medium/low
- handle-chat.ts의 think 결정에 프로파일 오버라이드 적용(라운드 루프 안에서
  effectiveModel 기준으로 판정)
- config: glm-5.3-flash:cloud에 think:false 세팅

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

src/gateway/chat/handle-chat.ts

@@ -127,6 +127,7 @@ export interface HandleChatDeps {
   goalLikelyNeedsTextInput: (goal: string) => boolean;
   getModelProfileForceToolChoice: (modelName: string) => boolean;
   getModelProfileFixedNumCtx: (modelName: string) => number | undefined;
+  getModelProfileThink: (modelName: string) => boolean | 'high' | 'medium' | 'low' | undefined;
   stripExplicitThinkTags: (text: string) => { cleaned: string; thinking: string };
   separateThinkingFromContent: (text: string) => { reply: string; thinking: string };
   isOrchestrationSkillEnabled: () => boolean;
@@ -181,6 +182,7 @@ export function createHandleChat(deps: HandleChatDeps) {
     goalLikelyNeedsTextInput,
     getModelProfileForceToolChoice,
     getModelProfileFixedNumCtx,
+    getModelProfileThink,
     stripExplicitThinkTags,
     separateThinkingFromContent,
     isOrchestrationSkillEnabled,
@@ -1434,7 +1436,12 @@ async function handleChat(
           // mid-session reloads.
           num_ctx: getModelProfileFixedNumCtx(resolvedModelThisRound || getModelForRole('executor')),
           num_predict: (isCodeAiSession || isProjSession) ? 32768 : (needsLongOutput ? 32768 : 6000),
-          think: primaryThinkMode,
+          // models.profiles[<model>].think로 모델별 오버라이드 가능 (예: glm-5.3-flash는
+          // 오케스트레이터 턴에서도 reasoning을 3~6K 토큰씩 써서 응답이 1분+ — 2026-09-22 실측)
+          think: (() => {
+            const pt = getModelProfileThink(effectiveModel || getModelForRole('executor'));
+            return pt !== undefined ? pt : primaryThinkMode;
+          })(),
           model: effectiveModel,
         });
 

src/gateway/server.ts

@@ -1824,6 +1824,30 @@ function getModelProfileFixedNumCtx(modelName: string): number | undefined {
 }
 
 
+// Per-model thinking override for the chat pipeline's primaryThinkMode. Default behavior:
+// orchestrator turns (multiAgentActive) think, everything else doesn't. Some models think
+// too verbosely to be usable there — glm-5.3-flash:cloud measured 2026-09-22 writing 3–6K
+// reasoning tokens per orchestrator round (12K chars of content-inline reasoning on a news
+// digest turn, ~72s for the final call) where gemma4 spends ~100 on the same turns. Set
+// models.profiles[<model>].think = false (or "high"/"medium"/"low") to pin the think mode
+// for that model; omit the field to keep the default behavior.
+function getModelProfileThink(modelName: string): boolean | 'high' | 'medium' | 'low' | undefined {
+  const m = String(modelName || '').trim();
+  if (!m) return undefined;
+  try {
+    const raw = getConfig().getConfig() as any;
+    const v = raw?.models?.profiles?.[m]?.think;
+    if (typeof v === 'boolean') return v;
+    if (typeof v === 'string' && ['high', 'medium', 'low'].includes(v.toLowerCase())) {
+      return v.toLowerCase() as 'high' | 'medium' | 'low';
+    }
+    return undefined;
+  } catch {
+    return undefined;
+  }
+}
+
+
 function resolveWorkspaceFilePath(workspacePath: string, filename: string): string {
   if (!filename) return '';
   if (path.isAbsolute(filename)) return filename;
@@ -2084,6 +2108,7 @@ const { handleChat, handleCodeChat } = createHandleChat({
   goalLikelyNeedsTextInput,
   getModelProfileForceToolChoice,
   getModelProfileFixedNumCtx,
+  getModelProfileThink,
   stripExplicitThinkTags,
   separateThinkingFromContent,
   isOrchestrationSkillEnabled,