diff --git a/tarot-lab/app/pipeline.py b/tarot-lab/app/pipeline.py
index d5dc21c..0704a67 100644
--- a/tarot-lab/app/pipeline.py
+++ b/tarot-lab/app/pipeline.py
@@ -28,6 +28,15 @@ class TarotError(Exception):
     pass
 
 
+class TarotTruncated(Exception):
+    """Claude 응답이 max_tokens에서 잘림 — reroll feedback에 길이 단축 요청 필요."""
+
+    def __init__(self, raw_text: str, limit: int):
+        self.raw_text = raw_text
+        self.limit = limit
+        super().__init__(f"Claude 응답이 max_tokens={limit}에서 잘림")
+
+
 def calc_cost(tokens_in: int, tokens_out: int) -> float:
     return (
         tokens_in / 1_000_000 * TAROT_COST_INPUT_PER_M
@@ -64,9 +73,10 @@ async def _call_claude(user_text: str, feedback: str = "") -> tuple[dict, dict,
         raise TarotError("ANTHROPIC_API_KEY missing")
     if feedback:
         user_text = f"이전 응답이 다음 이유로 거절됨: {feedback}\n올바른 스키마(시스템 지침)로 다시 응답.\n\n{user_text}"
+    max_tokens = 2800  # 3-card spread JSON(summary + cards×3 evidence + interactions + advice) 안전 마진
     payload = {
         "model": TAROT_MODEL,
-        "max_tokens": 1400,  # 응답 시간 단축 — 3-card spread evidence·interactions 포함 충분
+        "max_tokens": max_tokens,
         "system": [{"type": "text", "text": SYSTEM_PROMPT,
                     "cache_control": {"type": "ephemeral"}}],
         "messages": [{"role": "user", "content": [{"type": "text", "text": user_text}]}],
@@ -89,7 +99,15 @@ async def _call_claude(user_text: str, feedback: str = "") -> tuple[dict, dict,
     usage = resp.get("usage", {}) or {}
     tokens_in = int(usage.get("input_tokens", 0) or 0)
     tokens_out = int(usage.get("output_tokens", 0) or 0)
-    logger.info("tarot claude call: latency=%dms, in=%d, out=%d", latency_ms, tokens_in, tokens_out)
+    stop_reason = resp.get("stop_reason", "")
+    logger.info(
+        "tarot claude call: latency=%dms, in=%d, out=%d, stop=%s",
+        latency_ms, tokens_in, tokens_out, stop_reason,
+    )
+    if stop_reason == "max_tokens":
+        # truncation은 JSON 파싱 시도 전에 단락 — _extract_json은 거의 항상 실패하고
+        # 일반 JSONDecodeError로 reroll되면 모델이 길이를 줄일 힌트를 못 받음.
+        raise TarotTruncated(raw_text, max_tokens)
     parsed = _extract_json(raw_text)
     meta = {
         "tokens_in": tokens_in,
@@ -116,8 +134,15 @@ async def interpret(req: TarotInterpretRequest) -> Dict[str, Any]:
             parsed, meta, _raw = await _call_claude(user_text, feedback=last_error)
         except httpx.HTTPError as e:
             raise TarotError(f"Claude HTTP error: {e}") from e
+        except TarotTruncated as e:
+            last_error = (
+                f"이전 응답이 max_tokens={e.limit}에서 잘렸습니다. "
+                "각 카드의 interpretation·evidence·advice를 1~2문장으로 축약해 "
+                "전체 JSON 길이를 줄여 다시 응답하세요."
+            )
+            continue
         except json.JSONDecodeError as e:
-            last_error = f"JSON 파싱 실패: {e}"
+            last_error = f"JSON 파싱 실패: {e}. 모든 문자열을 닫고 유효한 JSON으로 재작성."
             continue
         total_in += meta["tokens_in"]
         total_out += meta["tokens_out"]
diff --git a/tarot-lab/tests/test_pipeline.py b/tarot-lab/tests/test_pipeline.py
index 49f1983..878e30e 100644
--- a/tarot-lab/tests/test_pipeline.py
+++ b/tarot-lab/tests/test_pipeline.py
@@ -47,10 +47,11 @@ def _valid_response_json():
     }
 
 
-def _claude_envelope(text: str, in_tok=100, out_tok=200):
+def _claude_envelope(text: str, in_tok=100, out_tok=200, stop_reason="end_turn"):
     return {
         "content": [{"type": "text", "text": text}],
         "usage": {"input_tokens": in_tok, "output_tokens": out_tok},
+        "stop_reason": stop_reason,
     }
 
 
@@ -112,3 +113,33 @@ async def test_interpret_http_error():
 def test_calc_cost():
     cost = pipeline.calc_cost(1_000_000, 1_000_000)
     assert cost == pipeline.TAROT_COST_INPUT_PER_M + pipeline.TAROT_COST_OUTPUT_PER_M
+
+
+@respx.mock
+async def test_interpret_truncated_then_success():
+    """1차 응답이 max_tokens에서 잘림 → 2차에서 정상 JSON 반환."""
+    truncated_text = '{"summary": "흐름이 있음", "cards": [{"position": "과거", "card": "the-fool", "reversed": false, "interpretation": "끝나지 않은 문장'
+    valid = json.dumps(_valid_response_json())
+    respx.post("https://api.anthropic.com/v1/messages").mock(
+        side_effect=[
+            httpx.Response(200, json=_claude_envelope(truncated_text, stop_reason="max_tokens")),
+            httpx.Response(200, json=_claude_envelope(valid)),
+        ]
+    )
+    result = await pipeline.interpret(_req())
+    assert result["reroll_count"] == 1
+    assert "interpretation_json" in result
+
+
+@respx.mock
+async def test_interpret_truncated_twice_raises():
+    """두 번 모두 max_tokens 잘림 → TarotError, 메시지에 'max_tokens' 포함."""
+    truncated_text = '{"summary": "...", "cards": [{"position":'
+    respx.post("https://api.anthropic.com/v1/messages").mock(
+        return_value=httpx.Response(
+            200, json=_claude_envelope(truncated_text, stop_reason="max_tokens")
+        )
+    )
+    with pytest.raises(pipeline.TarotError) as exc_info:
+        await pipeline.interpret(_req())
+    assert "max_tokens" in str(exc_info.value)