fix(guardrails/peyeeye): wrap json decode errors, simplify session cleanup

- Move response.json() inside the try/except in _post so JSON decode
  failures surface as PEyeEyeGuardrailAPIError rather than a raw
  ValueError.
- Drop the redundant ses_-prefix check in async_post_call_success_hook;
  peyeeye_session_mode == 'stateful' is already the gate, and the prefix
  check skipped legitimate cleanup if the API ever returned a different
  id shape.
- Add test_redact_json_decode_error_raises_typed covering the new
  wrapping branch.

litellm/proxy/guardrails/guardrail_hooks/peyeeye/peyeeye.py

@@ -211,8 +211,9 @@ class PEyeEyeGuardrail(CustomGuardrail):
                             new_parts.append(part)
                     message.content = new_parts
 
-        # Clean up: drop the session server-side and from cache.
-        if self.peyeeye_session_mode == "stateful" and session_id.startswith("ses_"):
+        # Clean up: drop the stateful session server-side. Stateless
+        # ``skey_…`` blobs hold no server-side state, so skip the DELETE.
+        if self.peyeeye_session_mode == "stateful":
             try:
                 await self._delete_session(session_id)
             except Exception as e:
@@ -288,9 +289,9 @@ class PEyeEyeGuardrail(CustomGuardrail):
                 url=url, headers=self._headers(), json=body, timeout=15.0
             )
             response.raise_for_status()
+            return response.json()
         except Exception as e:
             self._reraise_api_error(e, path)
-        return response.json()
 
     def _headers(self) -> Dict[str, str]:
         return {

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_peyeeye.py

@@ -236,6 +236,30 @@ async def test_redact_api_error_raises_typed():
         )
 
 
+@pytest.mark.asyncio
+async def test_redact_json_decode_error_raises_typed():
+    """If response.json() blows up, surface a typed PEyeEyeGuardrailAPIError."""
+    g = PEyeEyeGuardrail(peyeeye_api_key="pk", guardrail_name="t")
+    g.async_handler = MagicMock()
+
+    bad = MagicMock()
+    bad.raise_for_status = MagicMock()
+    bad.json.side_effect = ValueError("not json")
+    g.async_handler.post = AsyncMock(return_value=bad)
+
+    cache = DualCache()
+    with pytest.raises(PEyeEyeGuardrailAPIError):
+        await g.async_pre_call_hook(
+            UserAPIKeyAuth(api_key="x"),
+            cache,
+            {
+                "messages": [{"role": "user", "content": "hi"}],
+                "litellm_call_id": "json-err",
+            },
+            "completion",
+        )
+
+
 @pytest.mark.asyncio
 async def test_pre_call_raises_on_length_mismatch():
     """If /v1/redact returns fewer texts than sent, refuse to forward."""