Description

Description

After switching the inference model via `openshell inference set`,
`nemoclaw status` correctly shows the new model, but the OpenClaw
agent inside the sandbox still reports the old model identity.

This is similar to the previously fixed bug 5997656/6008460
(Kimi→nemotron direction) but occurs in the reverse direction
(nemotron→kimi or kimi→nemotron after runtime switch).

The sandbox's openclaw.json or agent config caches the model set
at build time and does not pick up the gateway-side inference set
change without a rebuild.

Device:        MacBook Pro (Apple M4 Pro, 48 GB)
OS:            macOS 26.0.1
NemoClaw:      v0.0.36
OpenClaw:      2026.4.24
Original model: moonshotai/kimi-k2.6
Switched to:   nvidia/nemotron-3-super-120b-a12b

1. Onboard with Kimi K2.6
2. Run: openshell inference set -g nemoclaw --model nvidia/nemotron-3-super-120b-a12b --provider nvidia-prod
3. Verify: nemoclaw status → shows nemotron (correct)
4. Run: openclaw agent -m "What model are you?" --session-id post-switch
5. Agent responds: "I'm using model: inference/moonshotai/kimi-k2.6"

Agent should report the new model (nemotron) after inference set,
or at minimum the actual inference responses should come from the
new model.

nemoclaw status: nvidia/nemotron-3-super-120b-a12b (correct)
Agent self-report: inference/moonshotai/kimi-k2.6 (stale)

The openclaw.json inside the sandbox still has the build-time model.
openshell inference set only changes the gateway routing, not the
sandbox-internal agent config.

Related to previously fixed 5997656 / 6008460 (same symptom, different
direction). Those were closed as verified — this may be a regression
or a different code path (runtime switch vs onboard-time config).

Bug Details

[NVB#6154913]