diff --git a/src/providers/ollama-adapter.ts b/src/providers/ollama-adapter.ts
index 3e875a3..2e7b2ce 100644
--- a/src/providers/ollama-adapter.ts
+++ b/src/providers/ollama-adapter.ts
@@ -343,6 +343,39 @@ function repairJson(input: string): string {
   return s;
 }
 
+/**
+ * The ordered `think` values to retry with when a call comes back empty.
+ *
+ * The retry ladder exists because some models (e.g. glm-5.1:cloud) default to thinking-only output
+ * when `think` is omitted, emitting reasoning and no content — nudging them to an explicit mode
+ * gets a real answer out.
+ *
+ * `think: false` is the one request that must not climb that ladder. It is an operator decision
+ * ("this model wastes its budget thinking"), so silently re-enabling thinking on an empty response
+ * would undo exactly what was asked for — and for a model whose thinking ratio is 85%, that turns a
+ * deliberate 7x speedup back into the slow path without a word. Explicit `false` therefore falls
+ * back only to the model's own default, and only as a last resort so an empty reply is not a dead
+ * end. Measured 2026-08-12 on muse-glimmer: think:false produced 0 empty answers in 20 questions,
+ * so this fallback should stay cold in practice.
+ */
+export function buildThinkCandidates(requested?: boolean | 'high' | 'medium' | 'low') {
+  const candidates: Array<boolean | 'high' | 'medium' | 'low' | undefined> = [];
+  const push = (v: boolean | 'high' | 'medium' | 'low' | undefined) => {
+    if (!candidates.some(x => x === v)) candidates.push(v);
+  };
+  if (requested === false) {
+    push(false);
+    push(undefined);
+    return candidates;
+  }
+  push(requested);
+  if (requested !== 'low') push('low');
+  push(undefined);
+  if (requested !== true) push(true);
+  push('medium');
+  return candidates;
+}
+
 export class OllamaAdapter implements LLMProvider {
   readonly id: 'ollama' | 'ollama_local';
   private client: Ollama;
@@ -864,15 +897,6 @@ export class OllamaAdapter implements LLMProvider {
   }
 
   private buildThinkCandidates(requested?: boolean | 'high' | 'medium' | 'low') {
-    const candidates: Array<boolean | 'high' | 'medium' | 'low' | undefined> = [];
-    const push = (v: boolean | 'high' | 'medium' | 'low' | undefined) => {
-      if (!candidates.some(x => x === v)) candidates.push(v);
-    };
-    push(requested);
-    if (requested !== 'low') push('low');
-    push(undefined);
-    if (requested !== true) push(true);
-    push('medium');
-    return candidates;
+    return buildThinkCandidates(requested);
   }
 }
diff --git a/tests/think-candidates.test.ts b/tests/think-candidates.test.ts
new file mode 100644
index 0000000..034ccdc
--- /dev/null
+++ b/tests/think-candidates.test.ts
@@ -0,0 +1,53 @@
+/**
+ * buildThinkCandidates — think 재시도 사다리
+ *
+ * 빈 응답이 오면 다른 think 값으로 재시도한다. glm-5.1:cloud처럼 think를 생략하면 생각만 하고
+ * content를 안 내놓는 모델을 건지려고 만든 장치다.
+ *
+ * 문제는 `think:false`였다. 2026-08-12 muse-glimmer(생각비중 85%, 체감 3.1 tok/s) 측정 후
+ * 생각을 끄기로 했는데, 사다리가 빈 응답 한 번에 thinking을 도로 켜버리면 그 결정이 조용히
+ * 뒤집힌다. 명시적 false는 모델 기본값까지만 물러선다.
+ */
+
+import { test, describe } from 'node:test';
+import assert from 'node:assert/strict';
+import { buildThinkCandidates } from '../src/providers/ollama-adapter';
+
+describe('think:false는 사다리를 오르지 않는다', () => {
+  test('false → [false, undefined]까지만', () => {
+    assert.deepEqual(buildThinkCandidates(false), [false, undefined]);
+  });
+
+  test('생각을 켜는 값이 하나도 없다', () => {
+    const c = buildThinkCandidates(false);
+    for (const v of [true, 'low', 'medium', 'high']) {
+      assert.ok(!c.includes(v as any), `${v}가 포함되면 안 됨`);
+    }
+  });
+
+  test('빈 응답이 막다른 길이 되지 않게 기본값 폴백은 남긴다', () => {
+    assert.ok(buildThinkCandidates(false).includes(undefined));
+  });
+});
+
+describe('나머지 경로는 그대로', () => {
+  test('요청값이 맨 앞에 온다', () => {
+    assert.equal(buildThinkCandidates(true)[0], true);
+    assert.equal(buildThinkCandidates('low')[0], 'low');
+    assert.equal(buildThinkCandidates(undefined)[0], undefined);
+  });
+
+  test('true 요청은 여전히 여러 모드로 재시도한다', () => {
+    const c = buildThinkCandidates(true);
+    assert.ok(c.includes('low'));
+    assert.ok(c.includes(undefined));
+    assert.ok(c.includes('medium'));
+  });
+
+  test('중복이 없다', () => {
+    for (const req of [undefined, true, false, 'low', 'medium', 'high'] as const) {
+      const c = buildThinkCandidates(req);
+      assert.equal(new Set(c).size, c.length, `중복 발생: ${JSON.stringify(c)}`);
+    }
+  });
+});