[N1X Linux][Inference] Fresh OpenClaw TUI HTML task fails with incomplete or malformed tool call

[JoyceChenNV]

Description

In a newly created NemoClaw sandbox using its managed local vLLM route, a fresh OpenClaw TUI session fails on a frontend HTML task. The user receives an inference request error instead of an answer or generated page.

Platform scope: Reproduced on N1X Linux running Ubuntu 24.04 only; other platforms not tested.

Regression: Unknown — earlier versions not tested.

OpenShell issue: No; the failing operation is an OpenClaw agent inference request inside a NemoClaw managed sandbox.

The same prompt failed in the user's new session and in an independent fresh TUI session on the same machine.

Environment

Device:        N1X Linux target
OS:            Ubuntu 24.04
Architecture:  aarch64
Node.js:       v22.23.2
npm:           10.9.8
Docker:        29.2.1
OpenShell CLI: 0.0.116
NemoClaw:      v0.0.128
OpenClaw:      2026.9.1 (inside sandbox)
Inference:     NemoClaw managed local vLLM, nvidia/Qwen3.6-35B-A3B-NVFP4
Context:       32768 tokens; max output 4096 tokens

Steps to Reproduce

1. Install NemoClaw v0.0.128 and create a sandbox with the managed local vLLM route on N1X Linux.

2. Start a fresh OpenClaw TUI session in that sandbox:

   openshell sandbox exec --tty -n {sandbox} -- openclaw tui --session {new-session}

3. Send this exact prompt once:

   you are a frontend engineer, create a HTML page to demonstrate advantages and disadvantages of different agents
   

4. Wait for the agent response.

Expected Result

The agent completes the request and returns a usable HTML page, or reports a recoverable error with a way to retry.

Actual Result

After about 98 seconds, the TUI prints:

run error: ⚠️ inference/nvidia/Qwen3.6-35B-A3B-NVFP4 request failed.

The gateway reports:

rawError=Provider returned an incomplete or malformed tool call.

The fresh session started at 0/33k tokens and showed about 11k/33k at failure. The managed vLLM service returned HTTP 200 to a chat completion request during this run.

Logs

2026-09-23T12:15:50Z [agent/embedded] embedded run agent end: isError=true; error=⚠️ inference/nvidia/Qwen3.6-35B-A3B-NVFP4 request failed.; rawError=Provider returned an incomplete or malformed tool call