Troubleshooting Guide#

To share feedback or ask questions about this release, access our NVIDIA TAO Developer Forum.

NGC#

The TAO Execution SDK pulls container images from nvcr.io and resolves model assets from NGC. The TAO Execution SDK reads its credentials from the shell that launches the agent. Set them up once per shell, before you start the agent:

  • Configure NGC and log in to the registry:

    ngc config set
docker login nvcr.io

  • When running ngc config set, the NGC CLI may not prompt you to configure the team and org. In that case you’ll see:

    Missing org - If apikey is set, org is also required.

    Back up your existing NGC API key from ~/.ngc/config and clear the configuration:

    ngc config clear

    Then re-run ngc config set and supply the team and org when prompted.

Agent and Plugin Issues#

  • /plugin install finished but the skills aren’t visible to the agent. Run /reload-plugins in the same session. /plugin install only registers the plugin; the reload makes the skills, slash commands, and hooks active.

  • The agent reports a backend is “not configured”: the environment variable that backend needs is not visible to the agent process. Environment variables must be exported in the shell before you launched the agent. Exit the agent, re-export the variables in the same shell (refer to the Getting Started with NVIDIA TAO Toolkit), and restart the agent with claude --resume (or relaunch Codex) in that same shell.

  • The agent cites a kwarg, action name, or endpoint you can’t find in the docs. Ask it to cite the SKILL.md file it read. The skill bank is the source of truth; if there is no file to cite, the agent hallucinated and you should ignore the suggestion.

  • A job dispatches but fails immediately. Confirm the credentials for the backend you picked: NGC_KEY (always), HF_TOKEN (gated models), and ACCESS_KEY/SECRET_KEY (for s3:// URIs). For the local Docker backend, also confirm docker login nvcr.io succeeded on the host.