fix(hosted-vllm): preserve reasoning_content in assistant history

[RishikeshVepura]

TLDR

Problem this solves:

• hosted_vllm drops replayed assistant reasoning_content - [Bug]: hosted_vllm drops reasoning_content from replayed assistant messages since 1.100.0 #41392
• Multi-turn reasoning agents lose prior reasoning since v1.100.0

How it solves it:

• Preserve reasoning_content in assistant message history
• Add regression coverage for the transformation

User Flow

Before: a developer’s tool-calling agent loses prior reasoning on its next turn

1. Their application calls a hosted vLLM reasoning model through LiteLLM
2. The assistant responds with reasoning content and a tool call
3. The application saves that assistant response and appends the tool result
4. The application makes its next call with the complete conversation history
5. The hosted vLLM model does not receive the assistant’s earlier reasoning content

After: the same agent retains prior reasoning on its next turn

1. Their application calls a hosted vLLM reasoning model through LiteLLM
2. The assistant responds with reasoning content and a tool call
3. The application saves that assistant response and appends the tool result
4. The application makes its next call with the complete conversation history
5. The hosted vLLM model receives the assistant’s earlier reasoning content

Relevant issues

Related to the Fireworks reasoning-content regression, but this affects the separate hosted_vllm provider transformation.

Affected release

Regression in v1.100.0

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The focused test file passes locally: 11 passed
• make lint passes locally
• TODO: Confirm GitHub Actions CI passes after pushing the PR
• 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 (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Screenshots / Proof of Fix

TODO: After committing this branch, run the local vLLM reproduction again and add the live before/after output here

Before

TODO: Add the output from LiteLLM v1.102.0 showing reasoning_content reached vLLM: False

After

TODO: Add the output from this PR branch showing reasoning_content reached vLLM: True

Type

🐛 Bug Fix

Caveats (if any)

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

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing RishikeshVepura:fix/hosted-vllm-preserve-reasoning-content (5e08695) with litellm_internal_staging (252c71c)}

[greptile-apps]

Greptile Summary

Preserves reasoning_content in replayed hosted vLLM assistant messages while continuing to remove thinking_blocks and normalize assistant content. The updated regression assertion verifies the field survives the production request transformation

Confidence Score: 5/5

The PR appears safe to merge, with the targeted behavior covered through the production transformation path

No actionable failures remain. The change directly removes the field deletion responsible for lost reasoning history, and the updated assertion would fail on the prior implementation

Important Files Changed

Filename	Overview
litellm/llms/hosted_vllm/chat/transformation.py	Stops removing assistant reasoning_content during hosted vLLM request transformation
tests/test_litellm/llms/hosted_vllm/chat/test_hosted_vllm_chat_transformation.py	Updates the mapped transformation test to verify that assistant reasoning history is preserved

_{Reviews (1): Last reviewed commit: "fix(hosted-vllm): preserve reasoning_con..." | Re-trigger Greptile}

[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.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

Closing as a duplicate of #43599, which carries the same hosted_vllm reasoning_content fix and merges cleanly against main. Thanks for the fix