fix(proxy): preserve reasoning fields on streaming cache hits

[jesco-absolut]

Relevant issues

Fixes #32068

Linear ticket

N/A

Pre-Submission checklist

• I have added meaningful tests
• 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
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

Streaming cache hits were replayed through convert_to_streaming_response_async, which rebuilt Delta with only content, role, function_call, and tool_calls. Cached reasoning_content and thinking_blocks were present in the stored response but dropped during replay

The regression test failed before the fix with:

AttributeError: 'Delta' object has no attribute 'reasoning_content'

After this change, async streaming replay preserves cached reasoning fields on the first replay chunk while later content-slice chunks still avoid duplicating metadata

Live proxy proof runbook for a reviewer with Redis and a reasoning-capable model configured:

curl -N -i http://localhost:4000/v1/chat/completions \
  -H 'Authorization: Bearer sk-1234' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "REASONING_MODEL",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Think briefly, then answer with the word done."}
    ]
  }'

Run the same request twice. Expected behavior: the first request is a cache miss and includes streamed delta.reasoning_content; the second request is a cache hit and still includes streamed delta.reasoning_content

I could not run this live proxy proof locally because this checkout does not have a Redis-backed proxy and real reasoning provider key configured

Type

Bug Fix
Test

Changes

Follow-up to #31022 in the same reasoning-streaming area

This changes async cached streaming replay to construct Delta from the full cached message dict, matching the sync cached replay path. That preserves reasoning_content, thinking_blocks, and any future Delta fields instead of hard-coding a small subset

Regression coverage verifies that cached async streaming replay preserves reasoning metadata on the first replay chunk and does not duplicate it on later content-slice chunks

Validation run:

uv run --no-sync pytest tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py::TestConvertToStreamingResponseAsync::test_preserves_reasoning_fields -q
uv run --no-sync pytest tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py -q
uv run --no-sync black --check tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
uv run --no-sync ruff check litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py
uv run --no-sync ruff check --ignore T201,F401 tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
uv run --no-sync python -m py_compile litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py
git --no-pager diff --check

Results:

74 passed
All checks passed for the changed production file and the new test path

[greptile-apps]

Greptile Summary

This PR fixes async streaming cache-hit replay to preserve reasoning_content and thinking_blocks by constructing Delta from the full cached message dict rather than an explicit four-field subset, bringing it in line with the already-correct sync path.

• Production change (1 line): Delta(**choice["message"]) replaces the hard-coded Delta(content=..., role=..., function_call=..., tool_calls=...) in convert_to_streaming_response_async, matching convert_to_streaming_response at line 299.
• Test addition: test_preserves_reasoning_fields verifies that cached reasoning_content and thinking_blocks appear on the first replay chunk and are correctly stripped from subsequent content-slice chunks by the existing _clear_later_replay_slice_metadata helper.

Confidence Score: 5/5

Safe to merge — a one-line change that aligns the async replay path with the already-shipped sync path, accompanied by a targeted regression test.

The async path was constructing Delta with four hard-coded fields while the sync counterpart already used Delta(**choice["message"]). The fix is a single-line alignment; the sync path has been production-stable, and the new test confirms first-chunk field preservation and correct later-chunk stripping via the existing _clear_later_replay_slice_metadata helper. No existing tests were modified or weakened.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/llm_response_utils/convert_dict_to_response.py	Replaces explicit four-field Delta construction in the async streaming replay path with Delta(**choice["message"]), matching the existing sync path (line 299). Preserves reasoning_content, thinking_blocks, and any future Delta fields from cached responses.
tests/llm_translation/test_llm_response_utils/test_convert_dict_to_chat_completion.py	Adds test_preserves_reasoning_fields covering that the first replay chunk carries reasoning_content and thinking_blocks from the cache while later content-slice chunks do not duplicate them. All new tests; no existing test coverage weakened.

_{Reviews (1): Last reviewed commit: "fix(proxy): preserve reasoning fields on..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing jesco-absolut:litellm_fix_streaming_cache_reasoning (33f60b6) with litellm_internal_staging (ecba48d)}