fix(examples): robust fenced-JSON parsing and runner cleanup in llm review

Author: DNKYr

examples/skills_code_review_agent/review/llm_review.py

@@ -17,10 +17,10 @@
 from .findings import Finding
 from .redaction import redact_text
 
-_FENCE_RE = re.compile(r"```(?:json)?\s*(\{.*?\})\s*```", re.DOTALL)
+_FENCE_RE = re.compile(r"```(?:json)?\s*(\{.*\})\s*```", re.DOTALL)
 
 
-def parse_llm_output(text: str):
+def parse_llm_output(text: str) -> tuple[list[dict], str]:
     """Extract (findings_dicts, summary) from model text; ([], "") on failure."""
     candidates = [text.strip()]
     fence = _FENCE_RE.search(text)
@@ -36,24 +36,27 @@ def parse_llm_output(text: str):
     return [], ""
 
 
-async def run_llm_review(agent, diff_text: str, static_findings):
+async def run_llm_review(agent, diff_text: str, static_findings) -> tuple[list[Finding], str, list[str]]:
     """Run one review turn. Returns (llm_findings, summary, warnings)."""
-    warnings = []
+    warnings: list[str] = []
     runner = Runner(app_name="code_review_agent", agent=agent,
                     session_service=InMemorySessionService())
     prompt = REVIEW_REQUEST_TEMPLATE.format(
         findings_json=json.dumps([f.model_dump() for f in static_findings]),
         diff=redact_text(diff_text))
     message = Content(role="user", parts=[Part.from_text(text=prompt)])
     final_text = ""
-    async for event in runner.run_async(user_id="cr_user",
-                                        session_id=uuid.uuid4().hex,
-                                        new_message=message):
-        if event.partial or not event.content or not event.content.parts:
-            continue
-        for part in event.content.parts:
-            if getattr(part, "text", None) and not getattr(part, "thought", None):
-                final_text += part.text
+    try:
+        async for event in runner.run_async(user_id="cr_user",
+                                            session_id=uuid.uuid4().hex,
+                                            new_message=message):
+            if event.partial or not event.content or not event.content.parts:
+                continue
+            for part in event.content.parts:
+                if getattr(part, "text", None) and not getattr(part, "thought", None):
+                    final_text += part.text
+    finally:
+        await runner.close()
     raw_findings, summary = parse_llm_output(final_text)
     if not summary and final_text:
         warnings.append("llm output was not valid JSON; ignored")

examples/skills_code_review_agent/tests/test_llm_review.py

@@ -31,6 +31,14 @@ def test_parse_llm_output_fenced_json():
     assert summary == "s" and findings == []
 
 
+def test_parse_llm_output_fenced_nested_json():
+    text = ('```json\n{"summary": "s", "findings": [{"severity": "high", "category": "security",'
+            ' "file": "a.py", "line": 3, "title": "x", "confidence": 0.9}]}\n```')
+    findings, summary = parse_llm_output(text)
+    assert summary == "s"
+    assert findings[0]["severity"] == "high"
+
+
 def test_parse_llm_output_garbage_returns_empty():
     findings, summary = parse_llm_output("I could not review this.")
     assert findings == [] and summary == ""