fix(streaming): preserve unsigned thinking blocks in stream_chunk_builder

[abhay-codes07]

Relevant issues

Fixes #33034

Pre-Submission checklist

• I have added meaningful tests
• My PR passes lint / format / unit tests locally
• My PR's scope is as isolated as possible; it only solves 1 specific problem

Screenshots / Proof of Fix

Reasoning-content data loss during streamed-response assembly, reproducible offline (no keys/network).

Before (base 3d63eda) — unsigned thinking text is dropped:

thinking_blocks (unsigned stream): None

After (this PR) — text preserved; signed/redacted paths unchanged:

UNSIGNED: [{'type': 'thinking', 'thinking': 'Let me reason about this.'}]
SIGNED  : [{'type': 'thinking', 'thinking': 'reasoning', 'signature': 'sig123'}]
REDACTED: [{'type': 'redacted_thinking', 'data': 'abc'}]

New regression test (fails on base, passes here) + existing thinking tests still green:

tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py  17 passed
  (incl. test_get_combined_thinking_content_preserves_unsigned_block)

Type

🐛 Bug Fix

Changes

ChunkProcessor.get_combined_thinking_content required both accumulated text and a signature to emit a thinking block, so a block whose signature never arrived (truncated stream, or a provider without per-block signatures) lost all its reasoning text. Emit the block whenever text exists and attach the signature only when present. signature is Optional on ChatCompletionThinkingBlock, and the send-side prompt transforms already drop unsignable blocks before replaying them, so this does not risk invalid replay. Signed and redacted paths are unchanged.

---

cc @ishaan-jaff @krrish-berri-2 — isolated data-loss fix in stream assembly with a regression test. Thanks!

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[greptile-apps]

Greptile Summary

This PR fixes a silent data-loss bug in ChunkProcessor.get_combined_thinking_content where accumulated thinking text was dropped whenever no signature arrived — as happens with a truncated stream or a provider that emits thinking without per-block signatures. The condition guarding _flush_thinking_block previously required both text and a signature; the fix relaxes that to text-only and attaches the signature only when present, matching the Optional[str] typing on ChatCompletionThinkingBlock.

• Core fix (streaming_chunk_builder_utils.py): _flush_thinking_block now emits a block as long as there is accumulated text; signature is added conditionally. Signed and redacted paths are unchanged.
• Regression test (test_streaming_chunk_builder_utils.py): A new test verifies that unsigned thinking text is preserved after assembly, failing on the base commit and passing here. Existing signed/interleaved/redacted tests continue to pass.

Confidence Score: 5/5

Safe to merge — the change is minimal, well-scoped, and strictly additive: callers that previously received unsigned thinking blocks as None will now receive the accumulated text instead, with no impact on signed or redacted paths.

The one-line condition change in _flush_thinking_block is straightforward, the TypedDict already declares signature as Optional[str], and the production caller in main.py uses the same chunk-filtering pattern that the new test exercises. No existing test expectations were weakened.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/litellm_core_utils/streaming_chunk_builder_utils.py	Fixes silent data loss in _flush_thinking_block by emitting unsigned thinking blocks; signature is attached only when present, matching the Optional[str] typing on ChatCompletionThinkingBlock.
tests/test_litellm/litellm_core_utils/test_streaming_chunk_builder_utils.py	Adds a focused regression test for the unsigned-thinking-block bug; mirrors the same chunk-filtering pattern used by the production caller in main.py and by the existing interleaved-block test.

_{Reviews (1): Last reviewed commit: "fix(streaming): preserve unsigned thinki..." | Re-trigger Greptile}

[abhay-codes07]

Heads up: the red auth-and-jwt check here is a pre-existing failure on the litellm_oss_daily_2026_07_10 base (the test_can_key_call_model_wildcard_access/test_can_team_access_model bedrockz/* cases), unrelated to this change. Root-caused and fixed separately in #33031; this PR's own changes are fully covered and green. Once #33031 lands in the base, this check goes green on re-run.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!