Repository navigation
Chat Completions streaming computes assistant output_index with a bool expression #3109
Description
Please read this first
- Have you read the docs? Agents SDK docs
- Have you searched for related issues? Others may have faced similar issues.
Describe the bug
The Chat Completions streaming handler sometimes passes a boolean value as output_index for assistant message events after a reasoning item has been emitted.
Several event constructors use this expression:
output_index=state.reasoning_content_index_and_output is not NoneWhen reasoning content exists, this produces True rather than the integer 1. Python treats bool as a subclass of int, so this can pass local numeric checks, but downstream JSON consumers and event snapshots can observe the wrong type.
Relevant code:
src/agents/models/chatcmpl_stream_handler.py: text, refusal, content-part-done, and output-item-done events use the boolean expression above.
Debug information
- Agents SDK version: current
mainat3854c124 - Python version: Python 3.12.1
Repro steps
Stream Chat Completions chunks where reasoning content is emitted before regular text content. Inspect the output_index on the text/message events that follow the reasoning item.
Current behavior: some events use True as output_index.
Expected behavior
All streamed response events should use integer output_index values, for example 0 for the reasoning item and 1 for the assistant message that follows it. No event should expose a boolean output_index.
Activity
[-]Chat Completions streaming emits boolean output_index values[/-][+]Chat Completions streaming computes assistant output_index with a bool expression[/+]Aphroq commentedon May 4, 2026
Update:
I re-tested this on
main.The source observation is still valid: several event constructors pass:
However, I could not reproduce the originally described downstream-visible behavior. The OpenAI SDK/Pydantic response event models coerce this value to
int.Observed stream events:
Direct event construction also coerces
output_index=Trueto1, andmodel_dump_json()emits"output_index":1.So this issue should be treated as a code clarity / defensive consistency cleanup, not as a confirmed runtime bug where consumers observe
bool.fix: #3109 stabilize chat completions stream output indexes
seratch commentedon May 7, 2026
Fixed y #3176
fix: #3109 stabilize chat completions stream output indexes (#3176)
fix: openai#3109 stabilize chat completions stream output indexes (op…