From d751148dd0c0f10cf04a955929010309db2321bc Mon Sep 17 00:00:00 2001
From: Qian-Cheng-nju <Qian-Cheng-nju@users.noreply.github.com>
Date: Fri, 9 Oct 2026 12:16:13 +0000
Subject: [PATCH] Honor Claude Code terminal error flags

---
 src/evaluator/termination.py        |  7 ++---
 tests/evaluator/test_termination.py | 42 +++++++++++++++++++++++++++++
 2 files changed, 46 insertions(+), 3 deletions(-)

diff --git a/src/evaluator/termination.py b/src/evaluator/termination.py
index 7bea49c2..99dca5c8 100644
--- a/src/evaluator/termination.py
+++ b/src/evaluator/termination.py
@@ -240,8 +240,9 @@ def claude_code_result_error(ctx: TerminationContext) -> str | None:
     Background work can produce several ``result`` events. The final result
     must follow the last conversation activity; an earlier success cannot
     finalize a truncated follow-up turn. Malformed streams and execution-error
-    results are INFRA_ERROR. ``error_max_turns`` remains a turn-budget LIMIT,
-    not infrastructure, when it closes the stream.
+    results are INFRA_ERROR, including ``is_error`` on a ``success`` subtype.
+    ``error_max_turns`` remains a turn-budget LIMIT, not infrastructure, when
+    it closes the stream.
     """
     if ctx.backend != "claude_code":
         return None
@@ -261,7 +262,7 @@ def claude_code_result_error(ctx: TerminationContext) -> str | None:
         # No result closes the latest activity, including background turns.
         return TerminationReason.INFRA_ERROR
     subtype = last_result.get("subtype", "")
-    if subtype.startswith("error") and subtype != "error_max_turns":
+    if subtype != "error_max_turns" and (last_result.get("is_error") is True or subtype.startswith("error")):
         return TerminationReason.INFRA_ERROR
     return None
 
diff --git a/tests/evaluator/test_termination.py b/tests/evaluator/test_termination.py
index 392ebb4f..0cd3272c 100644
--- a/tests/evaluator/test_termination.py
+++ b/tests/evaluator/test_termination.py
@@ -17,6 +17,7 @@
 
 from evaluator.backends import get_backend
 from evaluator.backends.agentic import AgenticBackend
+from evaluator.score import is_non_genuine
 from evaluator.termination import (
     INFRA_RULES,
     TerminationContext,
@@ -1019,6 +1020,47 @@ def test_cc_exec_error_is_infra(tmp_path):
     assert classify(_ctx(p, backend="claude_code")) == TerminationReason.INFRA_ERROR
 
 
+@pytest.mark.parametrize("agent_exit", [0, 1])
+def test_cc_success_subtype_with_error_flag_is_infra(tmp_path, agent_exit):
+    events = [
+        *CC_SUCCESS,
+        {"type": "system", "subtype": "task_notification"},
+        {"type": "rate_limit_event", "rate_limit_info": {"status": "rejected", "rateLimitType": "five_hour"}},
+        {
+            "type": "assistant",
+            "error": "rate_limit",
+            "message": {"content": [{"type": "text", "text": "You've hit your session limit"}]},
+        },
+        {"type": "result", "subtype": "success", "is_error": True, "stop_reason": "stop_sequence"},
+    ]
+    p = _write_jsonl(tmp_path / "quota-after-background.jsonl", events)
+    reason = classify(_ctx(p, backend="claude_code", agent_exit=agent_exit))
+    assert reason == TerminationReason.INFRA_ERROR
+    assert is_non_genuine({"termination_reason": reason, "check_verdict": "TIMEOUT"})
+
+
+def test_cc_error_flag_does_not_require_quota_text(tmp_path):
+    p = _write_jsonl(tmp_path / "flagged-error.jsonl", [{"type": "result", "subtype": "success", "is_error": True}])
+    assert classify(_ctx(p, backend="claude_code")) == TerminationReason.INFRA_ERROR
+
+
+def test_cc_error_flag_recovered_by_later_success_is_ok(tmp_path):
+    events = [
+        {"type": "result", "subtype": "success", "is_error": True},
+        *CC_SUCCESS,
+    ]
+    p = _write_jsonl(tmp_path / "recovered-error-flag.jsonl", events)
+    assert classify(_ctx(p, backend="claude_code")) == TerminationReason.OK
+
+
+def test_cc_runner_timeout_precedes_result_error_flag(tmp_path):
+    p = _write_jsonl(
+        tmp_path / "timeout-error-flag.jsonl", [{"type": "result", "subtype": "success", "is_error": True}]
+    )
+    ctx = _ctx(p, backend="claude_code", agent_exit=-1, error="claude_code timeout after 7200s")
+    assert classify(ctx) == TerminationReason.TIMEOUT
+
+
 def test_cc_max_turns_is_not_infra(tmp_path):
     # error_max_turns is a turn-budget LIMIT, not infrastructure — must NOT flag.
     p = _write_jsonl(tmp_path / "maxturns.jsonl", CC_MAX_TURNS)