feat(litellm): migrate litellm wrapper to integrations API

Move the LiteLLM tracing implementation from wrappers/litellm.py into
the integrations API under integrations/litellm/. This replaces the
manual LiteLLMWrapper class approach with FunctionWrapperPatcher
subclasses that use wrapt's wrap_function_wrapper for idempotent
monkey-patching.

Key changes:
- Add integrations/litellm/ with tracing.py, patchers.py, and
  integration.py following the standard integrations API pattern
- Add wrap_litellm() helper that instruments a specific litellm module
  object using the patchers, complementing the existing patch_litellm()
  which patches the globally-imported module
- Consolidate patcher list into _ALL_PATCHERS, shared by both
  LiteLLMIntegration.patchers and wrap_litellm()
- Reduce wrappers/litellm.py to compatibility re-exports
- Move tests and cassettes to integrations/litellm/
- Add cassettes_dir parameter to autoinstrument_test_context

Author: AbhiPrasad

py/noxfile.py

@@ -243,7 +243,7 @@ def test_litellm(session, version):
     # Install fastapi and orjson as they're required by litellm for proxy/responses operations
     session.install("openai<=1.99.9", "--force-reinstall", "fastapi", "orjson")
     _install(session, "litellm", version)
-    _run_tests(session, f"{WRAPPER_DIR}/test_litellm.py")
+    _run_tests(session, f"{INTEGRATION_DIR}/litellm/test_litellm.py")
     _run_core_tests(session)
 
 

py/src/braintrust/__init__.py

@@ -73,6 +73,9 @@ def is_equal(expected, output):
 from .integrations.anthropic import (
     wrap_anthropic,  # noqa: F401 # type: ignore[reportUnusedImport]
 )
+from .integrations.litellm import (
+    wrap_litellm,  # noqa: F401 # type: ignore[reportUnusedImport]
+)
 from .logger import *
 from .logger import (
     _internal_get_global_state,  # noqa: F401 # type: ignore[reportUnusedImport]
@@ -92,9 +95,6 @@ def is_equal(expected, output):
     BT_IS_ASYNC_ATTRIBUTE,  # noqa: F401 # type: ignore[reportUnusedImport]
     MarkAsyncWrapper,  # noqa: F401 # type: ignore[reportUnusedImport]
 )
-from .wrappers.litellm import (
-    wrap_litellm,  # noqa: F401 # type: ignore[reportUnusedImport]
-)
 from .wrappers.pydantic_ai import (
     setup_pydantic_ai,  # noqa: F401 # type: ignore[reportUnusedImport]
 )

py/src/braintrust/auto.py

@@ -14,6 +14,7 @@
     ClaudeAgentSDKIntegration,
     DSPyIntegration,
     GoogleGenAIIntegration,
+    LiteLLMIntegration,
 )
 
 
@@ -116,7 +117,7 @@ def auto_instrument(
     if anthropic:
         results["anthropic"] = _instrument_integration(AnthropicIntegration)
     if litellm:
-        results["litellm"] = _instrument_litellm()
+        results["litellm"] = _instrument_integration(LiteLLMIntegration)
     if pydantic_ai:
         results["pydantic_ai"] = _instrument_pydantic_ai()
     if google_genai:
@@ -147,14 +148,6 @@ def _instrument_integration(integration) -> bool:
     return False
 
 
-def _instrument_litellm() -> bool:
-    with _try_patch():
-        from braintrust.wrappers.litellm import patch_litellm
-
-        return patch_litellm()
-    return False
-
-
 def _instrument_pydantic_ai() -> bool:
     with _try_patch():
         from braintrust.wrappers.pydantic_ai import setup_pydantic_ai

py/src/braintrust/integrations/__init__.py

@@ -4,6 +4,7 @@
 from .claude_agent_sdk import ClaudeAgentSDKIntegration
 from .dspy import DSPyIntegration
 from .google_genai import GoogleGenAIIntegration
+from .litellm import LiteLLMIntegration
 
 
 __all__ = [
@@ -13,4 +14,5 @@
     "ClaudeAgentSDKIntegration",
     "DSPyIntegration",
     "GoogleGenAIIntegration",
+    "LiteLLMIntegration",
 ]

py/src/braintrust/integrations/auto_test_scripts/test_auto_litellm.py

@@ -1,24 +1,29 @@
 """Test auto_instrument for LiteLLM."""
 
+from pathlib import Path
+
 import litellm
 from braintrust.auto import auto_instrument
+from braintrust.integrations.litellm import LiteLLMIntegration
 from braintrust.wrappers.test_utils import autoinstrument_test_context
 
 
+_CASSETTES_DIR = Path(__file__).resolve().parent.parent / "litellm" / "cassettes"
+
 # 1. Verify not patched initially
-assert not hasattr(litellm, "_braintrust_wrapped")
+assert not LiteLLMIntegration.patchers[0].is_patched(litellm, None)
 
 # 2. Instrument
 results = auto_instrument()
 assert results.get("litellm") == True
-assert hasattr(litellm, "_braintrust_wrapped")
+assert LiteLLMIntegration.patchers[0].is_patched(litellm, None)
 
 # 3. Idempotent
 results2 = auto_instrument()
 assert results2.get("litellm") == True
 
 # 4. Make API call and verify span
-with autoinstrument_test_context("test_auto_litellm") as memory_logger:
+with autoinstrument_test_context("test_auto_litellm", cassettes_dir=_CASSETTES_DIR) as memory_logger:
     response = litellm.completion(
         model="gpt-4o-mini",
         messages=[{"role": "user", "content": "Say hi"}],

py/src/braintrust/integrations/auto_test_scripts/test_patch_litellm_aresponses.py

@@ -1,17 +1,20 @@
 """Test that patch_litellm() patches aresponses."""
 
 import asyncio
+from pathlib import Path
 
 import litellm
-from braintrust.wrappers.litellm import patch_litellm
+from braintrust.integrations.litellm import patch_litellm
 from braintrust.wrappers.test_utils import autoinstrument_test_context
 
 
+_CASSETTES_DIR = Path(__file__).resolve().parent.parent / "litellm" / "cassettes"
+
 patch_litellm()
 
 
 async def main():
-    with autoinstrument_test_context("test_patch_litellm_aresponses") as memory_logger:
+    with autoinstrument_test_context("test_patch_litellm_aresponses", cassettes_dir=_CASSETTES_DIR) as memory_logger:
         response = await litellm.aresponses(
             model="gpt-4o-mini",
             input="What's 12 + 12?",

py/src/braintrust/integrations/auto_test_scripts/test_patch_litellm_responses.py

@@ -1,13 +1,17 @@
 """Test that patch_litellm() patches responses."""
 
+from pathlib import Path
+
 import litellm
-from braintrust.wrappers.litellm import patch_litellm
+from braintrust.integrations.litellm import patch_litellm
 from braintrust.wrappers.test_utils import autoinstrument_test_context
 
 
+_CASSETTES_DIR = Path(__file__).resolve().parent.parent / "litellm" / "cassettes"
+
 patch_litellm()
 
-with autoinstrument_test_context("test_patch_litellm_responses") as memory_logger:
+with autoinstrument_test_context("test_patch_litellm_responses", cassettes_dir=_CASSETTES_DIR) as memory_logger:
     response = litellm.responses(
         model="gpt-4o-mini",
         input="What's 12 + 12?",

py/src/braintrust/integrations/dspy/tracing.py

@@ -50,7 +50,7 @@ class BraintrustDSpyCallback(BaseCallback):
         and disable DSPy's disk cache:
 
         ```python
-        from braintrust.wrappers.litellm import patch_litellm
+        from braintrust.integrations.litellm import patch_litellm
         patch_litellm()
 
         import dspy

py/src/braintrust/integrations/litellm/__init__.py

@@ -0,0 +1,40 @@
+"""Braintrust LiteLLM integration."""
+
+from .integration import LiteLLMIntegration
+from .patchers import wrap_litellm
+
+
+def patch_litellm() -> bool:
+    """Patch LiteLLM to add Braintrust tracing.
+
+    This wraps litellm.completion, litellm.acompletion, litellm.responses,
+    litellm.aresponses, litellm.embedding, and litellm.moderation to
+    automatically create Braintrust spans with detailed token metrics,
+    timing, and costs.
+
+    Returns:
+        True if LiteLLM was patched (or already patched), False if LiteLLM is not installed.
+
+    Example:
+        ```python
+        import braintrust
+        braintrust.patch_litellm()
+
+        import litellm
+        from braintrust import init_logger
+
+        logger = init_logger(project="my-project")
+        response = litellm.completion(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "Hello"}]
+        )
+        ```
+    """
+    return LiteLLMIntegration.setup()
+
+
+__all__ = [
+    "LiteLLMIntegration",
+    "patch_litellm",
+    "wrap_litellm",
+]