Merge 6de98912d3 into 8d6b8dab3d

litellm/proxy/pass_through_endpoints/jsonpath_extractor.py

@@ -9,6 +9,22 @@ from typing import Any, Final
 from litellm._logging import verbose_proxy_logger
 
 
+def _normalize_wildcard_segments(expr: str) -> tuple[str, ...]:
+    """Normalize ``key[*]`` into ``key`` + ``[*]`` segments.
+
+    Idempotent: the recursion in :meth:`JsonPathExtractor.evaluate` passes
+    already-normalized paths (e.g. ``content.[*].text``) back through here.
+    Re-running ``expr.replace("[*]", ".[*]")`` on those injects empty
+    segments (``content..[*].text``); dropping empty segments afterwards
+    keeps the normalization stable for raw (``a[*].b``), adjacent
+    (``a[*][*].b``) and already-normalized inputs alike, so no
+    nested-wildcard expression is silently dropped and skips the
+    guardrail (#46302).
+    """
+    segments: Final = tuple(segment for segment in expr.replace("[*]", ".[*]").split(".") if segment)
+    return segments
+
+
 class JsonPathExtractor:
     """Extracts field values from data using JSONPath-like expressions."""
 
@@ -55,7 +71,7 @@ class JsonPathExtractor:
         if not expr or not data:
             return None
 
-        parts: Final = expr.replace("[*]", ".[*]").split(".")
+        parts: Final = _normalize_wildcard_segments(expr)
         current: Any = data
 
         for i, part in enumerate(parts):

tests/unit/proxy/pass_through_endpoints/test_passthrough_guardrails.py

@@ -155,6 +155,26 @@ class TestJsonPathExtractorEvaluate:
         result = JsonPathExtractor.evaluate(data, "messages[*].content")
         assert result == ["You are helpful", "Hello"]
 
+    def test_nested_array_wildcards(self):
+        """Should extract values through two array wildcards.
+
+        Regression test for #46302: the recursive wildcard branch passes the
+        already-normalized remaining path (e.g. "content.[*].text") back
+        through evaluate, where the old replace-based normalization injected
+        an empty segment ("content..[*].text") and silently dropped every
+        nested-wildcard expression - so a pass-through guardrail whose only
+        request_fields entry is messages[*].content[*].text never saw any
+        text and skipped the check entirely.
+        """
+        data = {
+            "messages": [
+                {"role": "user", "content": [{"type": "text", "text": "item1"}]},
+                {"role": "assistant", "content": [{"type": "text", "text": "item2"}]},
+            ]
+        }
+        assert JsonPathExtractor.evaluate(data, "messages[*].content[*].text") == ["item1", "item2"]
+        assert JsonPathExtractor.extract_fields(data, ["messages[*].content[*].text"]) == "item1" + chr(10) + "item2"
+
     def test_missing_key_returns_none(self):
         """Should return None for missing key."""
         data = {"query": "test"}