[Brev][Onboard] NVIDIA Cloud default model differs across OpenClaw, Hermes, and Deep Code harnesses

[tedy-y]

Description

On the NemoClaw browser onboard page exported by a Brev launchable, selecting "NVIDIA Cloud" as the inference provider produces a different default model for each agent harness. Switching only the harness selection — with the provider left on NVIDIA Cloud — changes the NVIDIA model select to a different value each time, and for Hermes it falls back to the free-text "Other Model ID" entry instead of a named model.

Platform scope: Reproduced on Brev only; other platforms not tested
Regression: Unknown — earlier versions not tested

Environment

Device:        Brev launchable instance (n2d-standard-4, CPU only, no GPU)
OS:            Ubuntu 22.04.5 LTS
Architecture:  x86_64 (kernel 6.8.0-1064-gcp)
Node.js:       v22.23.2
npm:           10.9.8
Docker:        Docker version 29.1.3, build 29.1.3-0ubuntu3~22.04.2
OpenShell CLI: openshell 0.0.85
NemoClaw:      nemoclaw v0.0.97
OpenClaw:      N/A (defect is observed on the onboard page, before any
               OpenClaw sandbox is created)

Steps to Reproduce

1. Deploy the NemoClaw Brev launchable and open the /onboard page it exports.
2. On the first onboard screen, choose the middle option "Set up in this browser".
3. Set the inference provider to "NVIDIA Cloud".
4. Select agent harness "OpenClaw"; note the value in the NVIDIA model select.
5. Switch the agent harness to "Hermes"; note the value in the NVIDIA model select.
6. Switch the agent harness to "Deep Code"; note the value in the NVIDIA model select.

Expected Result

Under the NVIDIA Cloud provider, all supported agent harnesses default to the same model, unless a per-harness default is explicitly specified by the PRD.

Actual Result

Each harness defaults to a different model:

OpenClaw    -> "Nemotron 3 Super 120B"
Hermes      -> "Other Model ID" free-text field, pre-filled with
               nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
Deep Code   -> "Nemotron 3 Ultra 550B"

The Hermes default additionally does not resolve to any named entry in the NVIDIA model select; it lands on the "Other Model ID" fallback input.

The divergent default is not confined to the onboard UI — it is committed to the resulting sandbox. Completing the flow with the Deep Code harness on this instance produced a sandbox running model nvidia/nemotron-3-ultra-550b-a55b with provider nvidia-prod, matching the Ultra 550B default shown in the select.

Logs

Sandbox state after completing onboard with the Deep Code harness, showing the committed default model:

Sandboxes:
  deepagents-code *
    agent: langchain-deepagents-code  model: nvidia/nemotron-3-ultra-550b-a55b
    provider: nvidia-prod  CPU sandbox  policies: npm, pypi, huggingface, brew

Onboard-page model-select values were captured by direct UI observation; those steps produce no log output.

[hunglp6d]

@tedy-y
Different defaults per agent runtime are intentional. The Expected Result above does not match the accepted design. brevdev/nemoclaw-image#50 sets the Deep Agents Code default to Nemotron Ultra and states "OpenClaw and Hermes model defaults are unchanged." onboard-ui/lib/onboard-agents.test.ts and scripts/validate-image-contract.sh assert those per-runtime values. No PRD or documentation page records this design, which is why the divergence reads as a defect.

The Hermes default is the real defect. nvidia/nemotron-3-nano-omni-30b-a3b-reasoning was a member of CLOUD_MODEL_OPTIONS when it was chosen (#2628). #6074 removed it on 2026-07-01 when it refreshed that list from NVIDIA's featured-model catalog. The launchable pin was never updated, so the model select falls through to "Other Model ID". No check verifies that an agent's cloud model default is a catalog member, so a later model retirement can repeat this for any agent runtime.

The fix belongs in brevdev/nemoclaw-image, which builds the Brev Launchable and its /onboard page. I created brevdev/nemoclaw-image#114 there for the Hermes default and the missing check. I hold pull permission on that repository and no push permission, and brevdev/nemoclaw-image#41 records that the GCP image build and the Launchable smoke test require a branch in that repository, so a fork pull request cannot complete validation. I can supply the diff and the test on request.

Remaining work for this issue: record the per-runtime design in documentation.

[tedy-y]

Thanks for the update. @hunglp6d
So, it will include a docs fix?

[hunglp6d]

I checked the documentation again. docs/inference/choose-model.mdx has recorded this design since 2026-07-10, so this issue does not need a documentation change. My earlier comment was wrong to state that no page records it.

Nemotron 3 Super remains the shared default for OpenClaw and Hermes when it is present.
LangChain Deep Agents Code uses Nemotron 3 Ultra as its NVIDIA Endpoints default.

NemoClaw resolves the per-agent value from inference.default_model in the agent manifest (#6414). Only agents/langchain-deepagents-code/manifest.yaml sets that field, so OpenClaw and Hermes resolve to DEFAULT_CLOUD_MODEL, which is Nemotron 3 Super.

That sharpens the defect. The documented Hermes default is Nemotron 3 Super. The Brev Launchable pins nvidia/nemotron-3-nano-omni-30b-a3b-reasoning instead, so the Launchable contradicts documented NemoClaw behavior for Hermes. OpenClaw and Deep Agents Code match the documentation.

Remaining work is brevdev/nemoclaw-image#114. This issue can close when that change lands.