Severity: High — blocks the natural NVIDIA-on-Nebius integration path

Discovered: 2026-05-06 during tutorial development (Nebius Token Factory + NemoClaw integration)

Description

NemoClaw's Option 3 (OpenAI-compatible endpoint) hardcodes NEMOCLAW_REASONING=false and reads responses from choices[0].message.content. When a user selects a reasoning-only model (e.g. any NVIDIA Nemotron model available via a third-party provider like Nebius Token Factory), the response is returned in reasoning_content with an empty content field.

This causes two failures:

1. Smoke check failure during onboarding — the 32-token budget is exhausted by the reasoning trace; content is null/empty, causing the check to fail
2. 400 errors on tool calls — reasoning models do not support tool calls via the OpenAI-compatible wrapper

Root cause

NemoClaw sets NEMOCLAW_REASONING=true only when using NVIDIA's own endpoint (Option 1). Option 3 providers receive no reasoning-mode signal, so there is no way for the user to configure this at runtime.

Workaround

Use a non-reasoning model for Option 3 — e.g. deepseek-ai/DeepSeek-V3.2, meta-llama/Llama-3.3-70B-Instruct, or NousResearch/Hermes-4-70B.

Expected fix

Allow NEMOCLAW_REASONING to be set as a runtime flag during Option 3 onboarding, or detect reasoning-model responses (non-empty reasoning_content, empty content) and handle them gracefully.

Note: a companion issue has been filed with Nebius requesting a non-reasoning inference mode for Nemotron models via their OpenAI-compatible endpoint.