fix: keep content: null on an assistant tool-call turn

[likalight]

Fixes #37711.

Root cause

I traced the reporter's repro to cleanup_none_field_in_message in
litellm/utils.py:

new_message: Final = message.copy()
return {k: v for k, v in new_message.items() if v is not None}

It drops every key whose value is None. For an assistant message carrying
tool_calls, content: null is exactly the shape the OpenAI spec prescribes
for a tool-call-only turn, so the key is removed and providers that require it
to be present reject the request.

The helper's own docstring says the intent is removing stray keys like
{"function": None} that trip provider validation — content was collateral.

It is reached via validate_and_fix_openai_messages, which is why this
reproduces with a plain litellm.completion() call: no proxy, no provider, no
API key.

The fix

Exempt content when the assistant message carries tool_calls or a legacy
function_call. Everything else is stripped exactly as before — including a
None content on an assistant message that has neither, so the original
behaviour is preserved wherever it was not the bug.

Verification

Six tests in tests/test_litellm/test_cleanup_none_field_in_message.py:

• content: null survives on a tool_calls turn, and on a legacy
  function_call turn
• other None fields on that same turn (function, name) are still
  stripped — the helper's original purpose is intact
• content: null is still stripped from an assistant message with no tool call
• a user message is unaffected
• an end-to-end case through validate_and_fix_openai_messages, the caller
  that was actually dropping the field

4 of the 6 fail without the source change, and the reporter's own repro script
now prints content key present: True.

Regression check. I ran the provider suites most likely to depend on the old
behaviour — tests/test_litellm/llms/{openai,anthropic,databricks,mistral} —
2169 passed. There are 54 failures in that set, but the identical 54 fail on a
clean checkout with nothing applied (verified by stashing and re-running:
54 failed, 268 passed on both), so none are caused by this change.

• ruff format --check clean on both touched files
• ruff check litellm/utils.py: same findings as base, none added
• Type-discipline checker: identical counts to base on every rule

---

On the red code-quality check: it is not from this PR. It fails on every
PR against litellm_internal_staging right now, including ones that touch no
workflow files, because three unit shards in .github/workflows/test-unit.yml
cap the job below the startup-safety invariant. Fixed independently in #38046;
this PR needs no change for it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR adjusts OpenAI-message normalization so assistant tool-call and legacy function-call turns retain an explicit content: null, while other null fields continue to be removed.

• Adds focused coverage for tool calls, legacy function calls, unrelated null fields, non-tool assistant messages, user messages, and the full validation path.
• The implementation is narrowly scoped to assistant turns carrying a truthy tool-call payload.

Confidence Score: 4/5

The PR appears safe to merge functionally, with only non-blocking repository-convention cleanup needed for the new tests.

The normalization change is narrowly guarded and the added cases cover its intended behavior; the remaining concern is test organization and disallowed commentary rather than runtime correctness.

Files Needing Attention: tests/test_litellm/test_cleanup_none_field_in_message.py

Important Files Changed

Filename	Overview
litellm/utils.py	Preserves null content for assistant tool/function-call turns without changing cleanup of other null-valued fields.
tests/test_litellm/test_cleanup_none_field_in_message.py	Adds comprehensive regression coverage, but does so in a new test file and includes an ordinary explanatory comment contrary to repository conventions.

_{Reviews (1): Last reviewed commit: "fix: keep content: null on an assistant ..." | Re-trigger Greptile}

[greptile-apps]

Regression tests are misplaced

This bug fix creates a standalone test module for an existing helper and adds an ordinary explanatory comment at line 34, fragmenting the helper’s regression coverage and violating the repository’s test-placement and source-comment conventions. Move these cases into the existing mapped test module and remove the redundant comment.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing likalight:fix/preserve-null-content-on-tool-call-turn (9e7c25d) with litellm_internal_staging (d447be1)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (f005afa) during the generation of this report, so d447be1 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩