fix(logging): redact assistant tool call arguments in spend logs

[devin-ai-integration]

Relevant issues

Fixes #33107

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

See the "Changes" section for what leaked and what is now redacted. Reproduction is a streaming /chat/completions call that emits tool calls while message redaction is on (store_prompts_in_spend_logs: false / turn_off_message_logging); before the fix the stored response kept the assistant tool call function.arguments verbatim, after the fix they read redacted-by-litellm

Type

🐛 Bug Fix

Changes

When message redaction is on, perform_redaction scrubbed content, reasoning_content, thinking_blocks and audio on each choice but left assistant tool_calls (and legacy function_call) untouched, so the model-generated function.arguments still landed in LiteLLM_SpendLogs unredacted. The arguments are prompt-derived and are exactly the kind of payload the redaction is meant to strip

The redaction helpers now also blank the tool call arguments across every response shape they already handle, keeping the surrounding structure (id, name, type, token counts) intact so metrics stay visible:

# object form (Choices.message / StreamingChoices.delta)
for tc in message.tool_calls: tc.function.arguments = "redacted-by-litellm"
message.function_call.arguments = "redacted-by-litellm"

# dict form (model_dump'd choices, incl. standard_logging_object.response)
choice["message"]["tool_calls"][i]["function"]["arguments"] = redacted_str
choice["message"]["function_call"]["arguments"] = redacted_str

# Responses API output items
function_call_item.arguments = "redacted-by-litellm"

The streaming path (complete_streaming_response) is covered because it flows through the same _redact_choice_content helper

QA runbook

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

Link to Devin session: https://app.devin.ai/sessions/d4704cbfc1564c8ab7a3ece8dd6d98c3

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes a gap in the message redaction logic where assistant tool call arguments (function.arguments) were not being blanked when turn_off_message_logging/store_prompts_in_spend_logs: false was active, causing prompt-derived data to land in LiteLLM_SpendLogs unredacted.

• Adds _redact_tool_calls / _redact_function_call helpers for object-form choices and wires them into _redact_choice_content (covering both Choices and StreamingChoices).
• Adds _redact_tool_calls_dict for dict-form choices and applies it in _redact_model_response_dict_choices (message + delta paths) and in the Responses API dict helpers.
• Six new unit tests cover dict, object, streaming, standard_logging_object, and Responses API paths; no existing tests were altered.

Confidence Score: 4/5

Safe to merge; the fix correctly extends redaction to all tool-call argument paths and is well-tested.

The logic is sound and the new tests cover all major redaction paths. One minor inconsistency exists in _redact_tool_calls_dict where arguments is written even when absent from the dict, unlike every other guard in the same file. This is unlikely to cause problems in practice (spend-log entries carry complete responses), but it is a small behavioural divergence from the object-form helper.

litellm/litellm_core_utils/redact_messages.py — specifically the _redact_tool_calls_dict helper around line 159.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/redact_messages.py	Extends perform_redaction to blank tool call arguments (object and dict forms) across all redaction paths; minor inconsistency in dict helper where arguments is always set even when absent.
tests/test_litellm/litellm_core_utils/test_redact_messages.py	Adds 6 new focused tests covering all redaction paths for tool call arguments; no existing tests were modified.

_{Reviews (1): Last reviewed commit: "fix(logging): redact assistant tool call..." | Re-trigger Greptile}

[greptile-apps]

In _redact_tool_calls_dict, arguments is unconditionally written to every function dict that exists, even if the key wasn't present (e.g., a streaming partial chunk where only name has arrived so far). The companion object-form helper correctly guards with hasattr(function, "arguments"), and the function_call branch in the same function also guards with "arguments" in function_call. An absent arguments being turned into "redacted-by-litellm" could confuse consumers that distinguish between "no arguments received yet" and "arguments were redacted".

Suggested change

	if isinstance(tool_calls, list):
	for tool_call in tool_calls:
	if isinstance(tool_call, dict) and isinstance(tool_call.get("function"), dict):
	tool_call["function"]["arguments"] = redacted_str
	if isinstance(tool_calls, list):
	for tool_call in tool_calls:
	if isinstance(tool_call, dict) and isinstance(tool_call.get("function"), dict):
	if "arguments" in tool_call["function"]:
	tool_call["function"]["arguments"] = redacted_str

[codecov]

Codecov Report

❌ Patch coverage is 93.10345% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/litellm_core_utils/redact_messages.py	93.10%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_redact_assistant_tool_calls (681b5c3) with litellm_internal_staging (f1f33f5)}