Helix 0.8.1 runs tool calls with cut-off arguments; the upstream fix is merged in NAT develop but not in the 1.9 line Helix pins #3693

Summary
NeMo Helix 0.8.1 resolves nvidia-nat-langchain 1.9.0. With a tool_calling_agent workflow, a tool call whose streamed arguments were cut off is still executed when the stream ends without finish_reason or the Responses API returns status "incomplete". In my test, a 40-line write_file call wrote 16 lines and the agent answered "Done."

Measured (local mock model server, nhx agents invoke, 3 runs per case)

• Stream ends without finish_reason: tool ran, 16 of 40 lines (3/3)
• Responses status "incomplete": tool ran, 16 of 40 lines (3/3)
• Normal completion, chat and Responses (controls): 40 of 40 lines (3/3)

Upstream status
The fix is merged in NAT develop: NVIDIA/NeMo-Agent-Toolkit#2551 (tool_calling_agent). The react_agent fix is merged too: NVIDIA/NeMo-Agent-Toolkit#2553. Helix pins nvidia-nat-langchain>=1.9.0a20260811,<1.10, while develop builds 1.10, so Helix does not get the fix.

Backport tested on Helix
I ported the fix to release/1.9 and ran the same cases through Helix 0.8.1 with it installed: both cut-off cases no longer run the tool (3/3), and the complete cases still write 40 of 40 lines (3/3), including a reply without finish_reason.

Question
Would the Helix team prefer a 1.9.x release of nvidia-nat-langchain with the backport, or raising the pin once 1.10 ships? I can open the backport PR against release/1.9.

Found with https://github.com/simin-yuan/agent-cutoff-bench