[l2ysho]

Note

TL;DR
autofix tests

Mechanism

ReActAgent(tools=..., llm=llm) leaves streaming at its default of True, so every LLM call goes through llama_index.llms.openai.base.OpenAI._astream_chat. That method carries @llm_retry_decorator, but the network call lives inside an inner async def gen() that the caller iterates after the decorated function has returned. Tenacity therefore wraps a function that performs no I/O and can never fire.

The openai SDK's own max_retries=3 is out of reach for the same shape of reason: with stream=True the retry-guarded _send_request returns as soon as headers arrive, and the body is consumed later in AsyncStream.__stream__, which has no transport-error handling.

So a mid-stream connection drop propagates as a raw httpx.ReadError out of await handler in agent.py, into run_query's except in my_actor/main.py, which calls Actor.fail() and exits 1.

With streaming=False the call routes to _achat, where the request sits directly in the decorated function body and a stream=False response has its body read inside the SDK's retry-guarded _send_request. Both retry layers engage.

Verified against the installed library

• ReActAgent(tools=[], llm=llm).streaming is True; passing streaming=False sets it to False.
• The template does not consume AgentStream events — its stream_events() loop handles only ToolCall and ToolCallResult, which MultiAgentWorkflow.call_tool emits independently of self.streaming. Verbose logging is unchanged.
• The two branches of ReActAgent.take_step differ only in whether AgentStream deltas are written; both return the same AgentOutput with the same tool_calls.

Evidence

Run 38986102096, job 117257819030 (2026-10-17, windows-latest / Python 3.12). The traceback runs react_agent.py:158 → _get_streaming_response → llama_index/llms/openai/base.py:831 in gen → openai/_streaming.py → httpx.ReadError, with no tenacity frame in it.

Honesty about confidence

This is n=1: one failure in 30 days, and the same SHA passed on the nightly schedule on 10-16, 10-18 and 10-19. The fix is unproven against a live failure — nobody can reproduce a dropped connection on demand, and I did not run the template against the OpenAI API. What is proven is the code path: the retry decorator demonstrably wraps no I/O on the streaming path, and does wrap it on the non-streaming path. The change removes a retry layer that was silently doing nothing; whether that particular ReadError would have been retried away is an inference, not a measurement.

Archives

dist/templates/python-llamaindex-agent.zip is rebuilt automatically — the "Build template archives" workflow runs on every push to master and commits chore: Update template archives [skip ci]. Not hand-edited here, matching #1073.

Checks run locally

ruff check, ruff format --check, ty check on the template; pnpm run lint, pnpm run build, pnpm run test-without-templates at the repo root. All pass. (pnpm run format:check flags an untracked local .claude/settings.local.json, unrelated to this change.)

🤖 Generated with Claude Code

[apify-factory]

Looks good, nothing blocking.

🏭 Apify Factory auto approved ✅ — we looked twice, just to be sure

Every gate passed, and neither the reviewer nor a juror found anything a human needs to look at.

_{Approval locked to 4671eb454 · workflow run}