fix(openai): handle Azure stream chunks without delta

[hassiebp]

Summary

Fixes Azure OpenAI streaming traces when Azure emits a trailing content-filter chunk whose chat choice has delta=None.

The OpenAI stream extractor previously dereferenced delta.__dict__ unconditionally for OpenAI v1 responses. In Azure deployments with custom content filters, the trailing metadata-only chunk can have no delta, causing final stream extraction to fail internally. Because finalization catches extraction errors, the generation span ended without assistant output or usage details.

This PR:

• treats delta=None as an empty delta while continuing to process prior streamed content
• preserves the last non-null usage and finish_reason across trailing metadata-only chunks
• adds a regression test with an Azure-style trailing content-filter chunk

Verification

• uv run --frozen pytest tests/unit/test_openai.py -q
• uv run --frozen ruff check langfuse/openai.py tests/unit/test_openai.py
• uv run --frozen ruff format --check langfuse/openai.py tests/unit/test_openai.py
• uv run --frozen mypy langfuse --no-error-summary

Disclaimer: Experimental PR review

Greptile Summary

Fixes a crash in Azure OpenAI streaming when Azure emits a trailing content-filter chunk with delta=None and finish_reason=None, which previously caused finalization to produce spans without output or usage data.

• Adds a delta is not None guard before calling delta.__dict__ for OpenAI v1 objects, then falls back to an empty dict so downstream logic runs safely.
• Changes usage and finish_reason accumulation to a "last non-null wins" pattern, preserving values from earlier normal chunks when a trailing metadata-only chunk clears them.
• Adds a regression test with an Azure-style three-chunk stream (content chunk → stop chunk with usage → trailing content-filter chunk with delta=None).

Confidence Score: 5/5

Safe to merge — the fix is narrowly scoped to the stream extractor, does not alter any public API or data model, and is covered by a direct regression test.

The change is a minimal three-part defensive fix in one function, each part independently safe: a null guard on delta before dict conversion, an empty-dict fallback so the rest of the loop body runs unchanged, and a non-null-only assignment pattern for usage and finish_reason. The regression test exercises exactly the failing scenario end-to-end and checks all three affected outputs (content, finish_reason, usage).

No files require special attention.

Sequence Diagram

sequenceDiagram
    participant Azure as Azure OpenAI API
    participant Extractor as _extract_streamed_openai_response
    participant Span as Langfuse Span

    Azure->>Extractor: "chunk 1 — delta with content, finish_reason=None, usage=None"
    Note over Extractor: model set, content accumulated
    Azure->>Extractor: "chunk 2 — delta empty, finish_reason="stop", usage={tokens}"
    Note over Extractor: finish_reason preserved (non-null), usage preserved (non-null)
    Azure->>Extractor: "chunk 3 — delta=None, finish_reason=None, usage=None (content-filter)"
    Note over Extractor: delta=None → skip __dict__ → treat as {}<br/>finish_reason=None → keep "stop"<br/>usage=None → keep prior usage
    Extractor->>Span: "finalize with content, finish_reason="stop", usage={tokens}"

Loading

_{Reviews (1): Last reviewed commit: "fix(openai): handle Azure stream chunks ..." | Re-trigger Greptile}

[github-actions]

@claude review