The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 23, 2026 10:31pm

All committers have signed the CLA.

Merging this PR will not alter performance

✅ 16 untouched benchmarks

_{Comparing quora:fix/missing-content-part-added (1716956) with main (63425b4)}

Greptile Summary

This PR fixes a spec compliance gap in LiteLLMCompletionStreamingIterator where the required response.content_part.added SSE event was never emitted for non-OpenAI providers (Claude, Gemini, etc.). Downstream parsers need this event to initialise the text-part structure before any response.output_text.delta events arrive.

Changes:

• streaming_iterator.py: Calls create_content_part_added_event() inside _ensure_output_item_for_chunk() immediately after enqueuing output_item.added for message items. A flag guard (sent_content_part_added_event) prevents double-emission. The sequence number increments correctly via the existing create_content_part_added_event() helper.
• test_anthropic_completion.py: The existing integration test is updated to include CONTENT_PART_ADDED in its expected event sequence, making it correctly verify the new behaviour.
• test_litellm_completion_responses.py: Three new unit tests (using __new__ + MagicMock, no network calls) cover: message items emit the event, reasoning items do not, and the event is not duplicated on repeated calls.

The fix is narrowly scoped to message items: reasoning-first output items still return early without emitting content_part.added, which is consistent with the spec and existing design.

Confidence Score: 4/5

• This PR is safe to merge — it adds a missing required SSE event with no breaking changes to existing consumers.
• The fix is minimal and well-targeted: a single if block added in one code path, guarded by an existing flag. Three new unit tests validate the added, not-added, and idempotency cases. The integration test is correctly updated. No backwards-incompatible changes; consumers that previously ignored the missing event will now receive the required event in spec-correct order. One point deducted because the inner if not self.sent_content_part_added_event guard is logically redundant (already covered by the outer sent_output_item_added_event gate), leaving a small ambiguity in the code's intent that could confuse future readers.
• No files require special attention; all changes are straightforward and well-tested.

Sequence Diagram

sequenceDiagram
    participant Provider as Non-OpenAI Provider<br/>(Claude / Gemini)
    participant Iterator as LiteLLMCompletionStreamingIterator
    participant Consumer as Downstream Parser

    Note over Iterator,Consumer: Before fix (broken)
    Provider->>Iterator: streaming chunk (text delta)
    Iterator->>Consumer: response.created
    Iterator->>Consumer: response.in_progress
    Iterator->>Consumer: response.output_item.added (message)
    Iterator->>Consumer: response.output_text.delta
    Note over Consumer: ❌ Fails — text part not initialised

    Note over Iterator,Consumer: After fix (correct)
    Provider->>Iterator: streaming chunk (text delta)
    Iterator->>Consumer: response.created
    Iterator->>Consumer: response.in_progress
    Iterator->>Consumer: response.output_item.added (message)
    Iterator->>Consumer: response.content_part.added ← NEW
    Iterator->>Consumer: response.output_text.delta
    Note over Consumer: ✅ Works — text part already initialised
    Iterator->>Consumer: response.output_text.done
    Iterator->>Consumer: response.content_part.done
    Iterator->>Consumer: response.output_item.done
    Iterator->>Consumer: response.completed

_{Reviews (3): Last reviewed commit: "test(responses): update expected events ..." | Re-trigger Greptile}