fix(logging): preserve ModelResponse choices format in redacted standard_logging_object + add Charity Engine provider endpoint

[Sameerlite]

Summary

Fixes failing redaction tests and adds Charity Engine provider endpoint support.

Changes

1. Redaction fix (litellm/litellm_core_utils/redact_messages.py)

• Fix: Handle dict representation of ModelResponse (from model_dump()) in perform_redaction
• When response_obj is a dict with choices, preserve the full structure and redact content/audio in place
• Add _redact_standard_logging_object helper for redacting the standard_logging_object field in different formats (ModelResponse, ResponsesAPIResponse, etc.)

2. Test updates (tests/logging_callback_tests/test_logging_redaction_e2e_test.py)

• Update assertions to expect choices format instead of {"text": "redacted-by-litellm"}
• Aligns with expected behavior since commit d84e5e3

3. Provider endpoint (provider_endpoints_support.json)

• Add Charity Engine provider endpoint with documentation URL and supported endpoints

Tests fixed

• test_standard_logging_payload[True-ft:gpt-3.5-turbo:my-org:custom_suffix:id]
• test_standard_logging_payload_audio[True-True]
• test_standard_logging_payload_audio[True-False]
• All test_logging_redaction_e2e_test tests

Motivation

The tests in test_custom_callback_input.py were updated (commit a50a84c) to expect the new choices format, but the code to support this was reverted. This PR restores the correct redaction behavior.

Made with Cursor

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 10, 2026 4:55am

[greptile-apps]

Greptile Summary

This PR fixes redaction of ModelResponse in dict form (from model_dump()) within perform_redaction, updates test assertions to match the choices-based format already produced by CustomLogger, and registers the Charity Engine provider endpoint. The core logic fix is solid, but a new helper function _redact_standard_logging_object is added and never called, making it dead code.

Key changes:

• perform_redaction now correctly handles a dict with a "choices" key (e.g. a model_dump()-serialised ModelResponse) — this is the real fix for the failing tests
• Test assertions updated from response == {"text": "redacted-by-litellm"} to response["choices"][0]["message"]["content"] == "redacted-by-litellm", aligned with existing custom_logger.py behaviour
• New _redact_standard_logging_object helper is defined but never called anywhere in the codebase — the actual standard_logging_object redaction continues to be handled by CustomLogger.redact_standard_logging_payload_from_model_call_details() in custom_logger.py
• The dead helper also has an incomplete redaction: it omits reasoning_content and thinking_blocks handling present in the live perform_redaction dict branch
• charity_engine added to provider_endpoints_support.json with chat completions, messages, and responses marked as supported

Confidence Score: 4/5

• Safe to merge with minor cleanup: the functional fix in perform_redaction is correct, the dead-code helper should be removed or wired up.
• The perform_redaction dict-choices branch is the actual functional fix and is correctly implemented. Test assertions are updated to match existing custom_logger.py behaviour. The only concern is _redact_standard_logging_object being dead code — it doesn't break anything at runtime, but it adds maintenance noise and has an incomplete redaction (missing reasoning_content/thinking_blocks) that could become a real bug if someone later wires it in.
• litellm/litellm_core_utils/redact_messages.py — the _redact_standard_logging_object function should either be connected to the call chain or removed.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/redact_messages.py	Two changes: (1) new _redact_standard_logging_object helper that is never called (dead code); (2) new dict-choices branch in perform_redaction that correctly handles model_dump()-style dicts — this is the real fix. The dead helper also misses reasoning_content/thinking_blocks redaction.
tests/logging_callback_tests/test_logging_redaction_e2e_test.py	Test assertions updated from response == {"text": "redacted-by-litellm"} to response["choices"][0]["message"]["content"] == "redacted-by-litellm", aligning with the ModelResponse-style dict now produced by custom_logger.py. All tests use mock_response, so no real network calls are made.
provider_endpoints_support.json	Adds a charity_engine entry with chat completions, messages, and responses endpoints enabled. Structurally consistent with surrounding entries.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[perform_redaction called] --> B{result type?}
    B -->|litellm.ModelResponse| C[_redact_choice_content on each choice]
    B -->|dict with choices key - NEW| D[Inline dict redaction: content\nreasoning_content, thinking_blocks, audio]
    B -->|litellm.ResponsesAPIResponse| E[_redact_responses_api_output\nnullify reasoning field]
    B -->|litellm.EmbeddingResponse| F[Clear data list]
    B -->|other| G[Return plain redacted text dict]

    H[CustomLogger.redact_standard_logging_payload\n_from_model_call_details] --> I{response format?}
    I -->|dict with output key| J[Deep-copy and redact text in output items]
    I -->|other dict| K[Replace with new ModelResponse dict\nwith single redacted choice]

    L[_redact_standard_logging_object - NEW DEAD CODE] -.->|never called| M[Would redact standard_logging_object\nresponse field in-place]

Loading

_{Last reviewed commit: b084458}

[greptile-apps]

_redact_standard_logging_object is dead code — never called

This function is defined but never invoked anywhere in the codebase. The redaction of standard_logging_object already happens inside CustomLogger.redact_standard_logging_payload_from_model_call_details() in litellm/integrations/custom_logger.py (lines 773–869). That method is the one actually called during the logging pipeline.

Because _redact_standard_logging_object is never called from perform_redaction or redact_message_input_output_from_logging, it has no effect on runtime behavior. If it is intended to replace or supplement the logic in custom_logger.py, it needs to be wired in; otherwise it should be removed to avoid confusion for future maintainers.

[greptile-apps]

Missing reasoning_content and thinking_blocks redaction

In _redact_standard_logging_object, the "choices" dict branch only redacts content and audio. The parallel dict branch added to perform_redaction (lines 164–186) also redacts reasoning_content and thinking_blocks:

if "reasoning_content" in choice["message"]:
    choice["message"]["reasoning_content"] = "redacted-by-litellm"
if "thinking_blocks" in choice["message"]:
    choice["message"]["thinking_blocks"] = None

If this function is ever wired up, reasoning content and thinking blocks in the standard_logging_object response would not be redacted, creating an inconsistency with the behaviour of perform_redaction.