fix(streaming): map reasoning to reasoning_content in Delta for gpt-oss providers

[Chesars]

Relevant issues

Fixes #13300

Pre-Submission checklist

• I have Added testing in the tests/litellm/ directory
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

Providers like Cerebras return delta.reasoning in streaming responses for gpt-oss models, but LiteLLM's Delta class expects reasoning_content. This causes reasoning content to be silently dropped during streaming.

The fix adds a 3-line mapping in Delta.__init__ that converts reasoning → reasoning_content when reasoning_content is not already set. This is generic and works for any OpenAI-compatible provider that uses the reasoning field name.

Tested with:

• Cerebras gpt-oss-120b (streaming + non-streaming)
• Groq gpt-oss-20b (streaming + non-streaming)
• NVIDIA NIM gpt-oss-20b (streaming + non-streaming)

Files changed:

• litellm/types/utils.py: Map reasoning kwarg to reasoning_content in Delta.__init__
• tests/test_litellm/types/test_types_utils.py: Add unit test for the mapping

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 4, 2026 8:45pm

[greptile-apps]

Greptile Summary

This PR fixes a bug where reasoning content was silently dropped during streaming for OpenAI-compatible providers (e.g., Cerebras, Groq gpt-oss models) that return delta.reasoning instead of delta.reasoning_content. The fix adds a 3-line field name mapping in Delta.__init__ that normalizes reasoning → reasoning_content before the parent class is initialized.

• litellm/types/utils.py: Added reasoning → reasoning_content mapping in Delta.__init__, placed before super().__init__() to prevent field leakage
• tests/test_litellm/types/test_types_utils.py: Added comprehensive unit test covering mapping, precedence, non-leakage, and no-reasoning scenarios
• poetry.lock: Routine regeneration (Poetry 2.2.0 → 2.3.2), unrelated to the bug fix
• Note: Some providers (Groq, OpenAI GPT streaming handler) already have their own reasoning → reasoning_content mapping at the transformation layer. This Delta-level mapping acts as a generic safety net for providers that lack such mapping (e.g., Cerebras).

Confidence Score: 4/5

• This PR is safe to merge — it adds a small, defensive field mapping with no backwards-incompatible behavior changes.
• The change is minimal (3 lines of logic), well-tested, and follows existing patterns in the codebase. It only activates when reasoning_content is absent and reasoning is present, so it cannot break existing behavior. The mapping is placed before super().__init__() to prevent field leakage. One minor edge case remains: when both reasoning_content and reasoning are provided simultaneously, reasoning is still passed to the parent — but this is a theoretical edge case unlikely in practice.
• No files require special attention — the core change in litellm/types/utils.py is clean and well-scoped.

Important Files Changed

Filename	Overview
litellm/types/utils.py	Added reasoning → reasoning_content field mapping in Delta.__init__ before super().__init__(). Clean, minimal change that normalizes field names for gpt-oss providers.
tests/test_litellm/types/test_types_utils.py	Added comprehensive unit test covering: basic mapping, direct reasoning_content, precedence when both present, non-leakage assertion, and absent-field case. No network calls — pure mock tests.
poetry.lock	Routine regeneration from Poetry 2.2.0 to 2.3.2 with markers format changes. Unrelated to the bug fix.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Provider Streaming Response<br/>(e.g., Cerebras, Groq)"] -->|"delta.reasoning = '...'"| B["Delta.__init__(**kwargs)"]
    B --> C{"reasoning_content<br/>is None?"}
    C -->|"Yes"| D["Pop 'reasoning' from params<br/>Set reasoning_content = reasoning"]
    C -->|"No"| E["Keep existing<br/>reasoning_content"]
    D --> F["super().__init__(**params)<br/>(reasoning removed from params)"]
    E --> F
    F --> G{"reasoning_content<br/>is not None?"}
    G -->|"Yes"| H["self.reasoning_content = value"]
    G -->|"No"| I["del self.reasoning_content<br/>(OpenAI spec compliance)"]

Loading

_{Last reviewed commit: e48b7ae}

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ Chesars
❌ github-actions[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Chesars]

@greptile