[All Platforms][Onboard] re-onboard with different provider/model does not trigger drift detection or sandbox recreation prompt

[lyral-commits]

Description

During nemoclaw onboard for an existing sandbox, when the user selects a different inference provider or model than what the sandbox was originally created with, no drift detection prompt appears and the sandbox config is updated silently without offering to recreate the sandbox. The expected "Sandbox was created with X but you selected Y. Recreate? [y/N]" interactive prompt is never shown.

Platform scope: Reproduced on macOS (Apple M4 Pro, docker-desktop), DGX Spark (Linux arm64), and WSL2 (NVIDIA RTX PRO 6000 Blackwell Server Edition). All three platforms exhibit the same silent behavior.
Regression: Unknown — earlier versions not tested.

Environment

Device:        Apple Mac Studio (M4 Pro) / DGX Spark (Linux arm64) / WSL2 on Windows (NVIDIA RTX PRO 6000 Blackwell Server Edition)
OS:            macOS (Darwin arm64) / Ubuntu Linux arm64 / Windows WSL2
Architecture:  arm64 (macOS, Spark) / x86_64 (WSL2)
Node.js:       v22.22.3 (macOS) / v22.23.1 (Spark, WSL2)
npm:           10.9.8
Docker:        29.5.2 build 79eb04c7d8 (macOS, docker-desktop) / 29.5.3 build d1c06ef (WSL2)
OpenShell CLI: openshell 0.0.72
NemoClaw:      v0.0.81
OpenClaw:      2026.6.10

Steps to Reproduce

1. Run: nemoclaw onboard --fresh --name drift-test
2. Select OpenClaw agent, select NVIDIA Endpoints, select model nvidia/nemotron-3-super-120b-a12b
3. Complete onboard successfully
4. Run: nemoclaw drift-test status — confirm Provider: nvidia-prod, Model: nvidia/nemotron-3-super-120b-a12b
5. Run: nemoclaw onboard --fresh --name drift-test
6. Select OpenClaw agent, select NVIDIA Endpoints, select a different model (e.g. deepseek-ai/deepseek-v4-pro)
7. Accept configuration when prompted
8. Observe output between steps 5 and 7 for any drift/recreation prompt

Expected Result

At step 6, after selecting the different model, onboard detects drift from the stored sandbox config and presents a recreation prompt:

Sandbox 'drift-test' was created with nvidia-prod/nvidia/nemotron-3-super-120b-a12b
but you selected nvidia-prod/deepseek-ai/deepseek-v4-pro. Recreate? [y/N]

In non-interactive mode, the sandbox should be automatically recreated with:

[non-interactive] Recreating sandbox due to provider/model drift.

Actual Result

No drift detection prompt appears at any point. The onboard proceeds silently through all steps and updates the inference route to the new model without offering to recreate the sandbox. nemoclaw drift-test status after re-onboard shows the model updated but the sandbox was not recreated.

Logs

Not captured.

[rluo8]

This issue is not reproduced on current main using Ubuntu 24.04 and WSL2. The warning message could show with the reproduce steps.
Sandbox 'drift-test' exists but requested inference selection changed.
Current: provider=nvidia-prod model=nvidia/nemotron-3-ultra-550b-a55b
Requested: provider=nvidia-prod model=deepseek-ai/deepseek-v4-pro
Recreating the sandbox is required to apply this change to the running OpenClaw UI.
Recreate sandbox 'drift-test' now? [y/N]: y

[nvshaxie]

Confirmed fixed on current main. Re-onboarding with a changed provider/model does now trigger drift detection and the recreation prompt — the check lives in src/lib/onboard/sandbox-lifecycle.ts (drift warning + Recreate sandbox '<name>' now? [y/N]), and the recreate path was hardened by #6626.

@rluo8 already reproduced the working behavior on current main (Ubuntu 24.04 / WSL2) on 2026-07-14:

  Sandbox 'drift-test' exists but requested inference selection changed.
  Current:   provider=nvidia-prod  model=nvidia/nemotron-3-ultra-550b-a55b
  Requested: provider=nvidia-prod  model=deepseek-ai/deepseek-v4-pro
  Recreating the sandbox is required to apply this change to the running OpenClaw UI.
  Recreate sandbox 'drift-test' now? [y/N]: y

Closing as completed. Please reopen if you can still reproduce a missing drift prompt on v0.0.83.