Block publish when LLM rewrite quality degrades

tests/test_legacy_script_delegation.py

@@ -52,6 +52,19 @@ class LegacyScriptDelegationTests(unittest.TestCase):
         self.assertEqual(calls[0]["source_mode"], "mock")
         self.assertEqual(calls[0]["llm_mode"], "mock")
 
+    def test_main_exits_nonzero_when_new_pipeline_blocks_publish(self):
+        module = load_pipeline_module()
+
+        def fake_run_daily_report(**kwargs):
+            return {"reports": {"stage8": {"status": "blocked", "error": "rewrite_fallback_ratio_exceeded"}}}
+
+        with patch.object(module, "load_env", return_value={}):
+            with patch("ai_daily_report.runner.run_daily_report", side_effect=fake_run_daily_report):
+                with self.assertRaises(SystemExit) as raised:
+                    module.main()
+
+        self.assertEqual(raised.exception.code, 2)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_stage0_to_8_pipeline.py

@@ -1,5 +1,6 @@
 import json
 import unittest
+from urllib.error import HTTPError
 
 from ai_daily_report.pipeline import run_stage0_to_stage8
 
@@ -74,6 +75,65 @@ class Stage0To8PipelineTests(unittest.TestCase):
         self.assertIn("stage8", result["reports"])
         self.assertEqual(result["reports"]["stage8"]["status"], "ok")
 
+    def test_run_stage0_to_stage8_blocks_publish_when_rewrite_quality_gate_fails(self):
+        configs = [{"name": "AI HOT", "type": "fake", "role": "primary", "priority": 10}]
+
+        def fetcher(config, run_date):
+            return [
+                {
+                    "title_raw": f"News {index}",
+                    "summary_raw": f"Summary {index}",
+                    "url": f"https://example.com/{index}",
+                    "source_label": "Example",
+                    "section_hint": "模型发布/更新",
+                }
+                for index in range(6)
+            ]
+
+        def semantic_llm_call(prompt):
+            return json.dumps({"duplicate_groups": [], "not_duplicates": [], "uncertain": []})
+
+        def rewrite_llm_call(prompt):
+            raise HTTPError(
+                url="https://llm.example/v1/chat/completions",
+                code=503,
+                msg="Service Unavailable",
+                hdrs=None,
+                fp=None,
+            )
+
+        def guide_llm_call(prompt):
+            payload = json.loads(prompt)
+            return json.dumps(
+                {
+                    "theme": "模型能力继续更新。",
+                    "threads": [
+                        {
+                            "title": "模型更新",
+                            "text": "多条模型新闻更新。",
+                            "item_ids": [payload["items"][0]["id"]],
+                            "kind": "thread",
+                        }
+                    ],
+                }
+            )
+
+        result = run_stage0_to_stage8(
+            configs,
+            "2026-06-04",
+            fetcher=fetcher,
+            semantic_llm_call=semantic_llm_call,
+            rewrite_llm_call=rewrite_llm_call,
+            guide_llm_call=guide_llm_call,
+            mode="publish",
+            base_url="https://blog.example",
+            client=None,
+        )
+
+        self.assertEqual(result["publish"].status, "blocked")
+        self.assertIn("rewrite_fallback_ratio_exceeded", result["reports"]["stage7"]["blocking_errors"])
+        self.assertIn("rewrite_fallback_ratio_exceeded", result["reports"]["stage8"]["error"])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_stage4_rewrite.py

@@ -1,5 +1,6 @@
 import json
 import unittest
+from urllib.error import HTTPError
 
 from ai_daily_report.models import NewsItem
 from ai_daily_report.rewrite import rewrite_items
@@ -91,6 +92,29 @@ class Stage4RewriteTests(unittest.TestCase):
         self.assertEqual(report["fallback_count"], 0)
         self.assertEqual(calls, [["a", "b"], ["a"], ["b"]])
 
+    def test_rewrite_items_does_not_retry_single_items_after_transient_http_error(self):
+        items = [news_item("a"), news_item("b")]
+        calls = 0
+
+        def llm_call(prompt):
+            nonlocal calls
+            calls += 1
+            raise HTTPError(
+                url="https://llm.example/v1/chat/completions",
+                code=503,
+                msg="Service Unavailable",
+                hdrs=None,
+                fp=None,
+            )
+
+        rewritten, report = rewrite_items(items, llm_call=llm_call, batch_size=2)
+
+        self.assertEqual(calls, 1)
+        self.assertEqual([item.title for item in rewritten], ["OpenAI launches GPT-5 API", "OpenAI launches GPT-5 API"])
+        self.assertEqual(report["fallback_count"], 2)
+        self.assertTrue(report["quality_gate_failed"])
+        self.assertIn("rewrite_fallback_ratio_exceeded", report["blocking_errors"])
+
 
 if __name__ == "__main__":
     unittest.main()