fix(anthropic): preserve text delta when opening stream block

[its-amann]

Relevant issues

Fixes #30014

Summary

Anthropic stream wrapping was dropping the first non empty delta that opened a new content block unless that delta was a tool argument. In the /v1/messages to /v1/chat/completions path, that can make the first text chunk after a block transition disappear from the client stream.

This keeps the existing tool argument behavior and also queues non empty text, thinking, signature, and tool argument deltas when they trigger a content block transition. I added sync and async regression coverage for switching from thinking output back to text output.

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 and received a Confidence Score of at least 4/5 before requesting a maintainer review

Tests

• uv run pytest tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py::TestAnthropicStreamWrapperTextDeltas -q
• uv run pytest tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py::TestAnthropicStreamWrapperToolArgs tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py::TestAnthropicStreamWrapperTextDeltas -q
• uv run pytest tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py -q
• uv run ruff check litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
• git diff --cached --check

[CLAassistant]

All committers have signed the CLA.

[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 fixes a bug where the first non-empty delta on a newly opened content block was silently dropped during Anthropic stream wrapping, unless that delta was input_json_delta. The fix generalizes the existing "queue delta at block transition" logic to cover text_delta, thinking_delta, and signature_delta in addition to input_json_delta, and extracts the guard into a shared helper used by both the sync and async paths.

• _should_queue_block_transition_delta replaces two identical inline conditions in __next__ and __anext__, making the extension one-line rather than requiring matched edits in two places.
• New sync/async regression tests cover thinking→text and text→thinking transitions, and the existing interleaved-tool-calls test is strengthened with explicit assertions on the text delta values that were previously dropped.

Confidence Score: 5/5

Safe to merge — the change is isolated to the block-transition queueing logic and is fully covered by new regression tests in both sync and async paths.

The fix is a narrow, well-scoped generalization of an existing guard: the old code only preserved input_json_delta at block transitions; the new helper extends that to text_delta, thinking_delta, and signature_delta as well. Both sync and async paths are updated symmetrically. New regression tests cover both directions (thinking→text and text→thinking) in sync and async flavors, and the existing interleaved-tool-calls test is now stricter rather than looser.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py	Adds _should_queue_block_transition_delta helper and wires it into both the sync and async block-transition paths; replaces the prior input_json_delta-only guard cleanly.
tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py	Adds TestAnthropicStreamWrapperTextDeltas class with sync/async tests for both the thinking→text and text→thinking block transitions; also unit-tests the helper's edge cases directly.
tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_parallel_tool_calls.py	Strengthens the interleaved-tool-calls-and-text test by adding the two text content_block_delta events that were previously dropped to expected_types and asserting their text values.

_{Reviews (2): Last reviewed commit: "test(anthropic): cover block transition ..." | Re-trigger Greptile}

[Sameerlite]

Thanks for the fix to preserve the opening delta! Two small things:

• New commits have landed since Greptile last reviewed, so I'm re-triggering a fresh review (below).
• Could you paste the captured test output (or a short before/after stream transcript)? Listing the commands is helpful, but seeing the result confirms it.

@greptileai

Once that's in we'll take another look!

[its-amann]

Thanks for taking a look, and sorry I missed this before the PR was closed.

I see #30024 has since landed and closed #30014, so this PR is duplicate now. For completeness, here is the captured output from the final branch, since you asked for it:

$ uv run pytest tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_parallel_tool_calls.py --cov=litellm --cov-report=xml -q
83 passed in 48.48s

$ uv run diff-cover coverage.xml --compare-branch origin/litellm_internal_staging --include 'litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py'
streaming_iterator.py (100%)
Total: 14 lines
Missing: 0 lines
Coverage: 100%

$ uv run --no-sync mypy litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
Success: no issues found in 1 source file

$ uv run ruff check litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_parallel_tool_calls.py
All checks passed

$ uv run black --check litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_parallel_tool_calls.py
3 files would be left unchanged

The behavior covered on this branch was that the transition chunk is emitted as a content_block_delta after content_block_start for both thinking to text and text to thinking transitions, instead of being swallowed. No action needed here if #30024 is the accepted fix.