fix: parse qwen thinking json output

Co-authored-by: Codex <noreply@openai.com>

model_runtime.py

@@ -252,14 +252,29 @@ def _loads_lenient(content: str) -> dict[str, Any]:
     try:
         parsed: Any = json.loads(text)
     except json.JSONDecodeError:
-        start, end = text.find("{"), text.rfind("}")
-        if start == -1 or end == -1 or end <= start:
+        candidates = list(_json_object_candidates(text))
+        if not candidates:
             raise ValueError("Model response was not valid JSON.")
-        try:
-            parsed = json.loads(text[start : end + 1])
-        except json.JSONDecodeError as exc:
-            raise ValueError("Model response was not valid JSON.") from exc
+        parsed = candidates[-1]
 
     if not isinstance(parsed, dict):
         raise ValueError("Model response was not a JSON object.")
     return parsed
+
+
+def _json_object_candidates(text: str) -> list[dict[str, Any]]:
+    decoder = json.JSONDecoder()
+    candidates: list[dict[str, Any]] = []
+    cursor = 0
+    while True:
+        start = text.find("{", cursor)
+        if start == -1:
+            return candidates
+        try:
+            parsed, consumed = decoder.raw_decode(text[start:])
+        except json.JSONDecodeError:
+            cursor = start + 1
+            continue
+        if isinstance(parsed, dict):
+            candidates.append(parsed)
+        cursor = start + max(consumed, 1)

tests/test_model_runtime.py

@@ -74,6 +74,16 @@ class ModelRuntimeTests(unittest.TestCase):
 
         self.assertEqual(memo["outcome_audit_memo"], MEMO_JSON["outcome_audit_memo"])
 
+    def test_parse_model_json_uses_final_object_after_thinking_braces(self) -> None:
+        raw = (
+            "<think>Draft {not json} and a scratch object "
+            '{"draft": "ignore this"} before the final answer.</think>\n'
+            + json.dumps(MEMO_JSON)
+        )
+        memo = parse_model_json(raw)
+
+        self.assertEqual(memo["executive_memo"], MEMO_JSON["executive_memo"])
+
     def test_run_model_assist_uses_selected_model(self) -> None:
         result, narrative = analyze_trace_file(Path("examples/sample_trace_redacted.jsonl"))
         generate = RecordingGenerator()