Description

After Express setup with local Ollama + Qwen3.6 35B on DGX Spark, OpenClaw TUI chats correctly using inference/qwen3.6:35b, but the status bar shows tokens ?/131k instead of a numeric token count. The DGX Spark test's "numeric token usage for local Ollama" assertion fails.

Component area: Inference / Agent&Skills / CLI&UX (TUI status bar, usage accounting).

Environment

Platform:        DGX Spark (Grace+Blackwell GB10), aarch64
OS:              FastOS (DGX Spark OS), latest OTA
Docker:          Docker CE, nvidia user in docker group
nemoclaw:        v0.0.46
openshell:       0.0.39
OpenClaw:        2026.4.24 (cbcfdf6)
Provider:        ollama-local
Model:           qwen3.6:35b
Policy tier:     Balanced (local-inference preset included)
Sandbox:         my-assistant (Ready)
Install method:  Express setup on DGX Spark

Steps to Reproduce

Preconditions: NemoClaw installed via Express setup; sandbox my-assistant Ready with provider ollama-local, model qwen3.6:35b, Balanced policy.

1. Connect to the sandbox from the host:

   nemoclaw my-assistant connect

2. Inside the sandbox, start the OpenClaw TUI:

   openclaw tui

3. In the TUI input, type hello and press Enter.
4. Observe:
   • The model/provider line in the status bar.
   • The token-usage portion of the status bar after the response completes.

Expected Result

After Express setup with local Ollama, "hello" prompt:

• Agent responds using local Ollama model, displayed as inference/qwen3.6:35b in the status bar.
• TUI status bar shows numeric token usage, e.g.:

  agent main | session main (openclaw-tui) | inference/qwen3.6:35b | tokens 37/131k
  

This confirms streamed token usage accounting is wired from the Ollama-backed provider into the OpenClaw gateway and exposed to the TUI on DGX Spark.

Actual Result

Agent responds correctly:

hello
Hello! 👋 How can I help you today?

Status line at the bottom of TUI shows:

connected | idle  agent main | session main (openclaw-tui) | inference/qwen3.6:35b | tokens ?/131k

• Provider/model: inference/qwen3.6:35b (correct for Express Ollama configuration).
• Token portion: tokens ?/131k — current usage is ?, only the max (131k) is shown.

The TUI succeeds functionally (chat works, provider is correct), but the usage accounting UI is incomplete for this local Ollama configuration.

Logs

nvidia@spark-dadc:~$ nemoclaw my-assistant connect
✓ Connecting to sandbox 'my-assistant'
sandbox@9c35e3cacb8d:~$ openclaw tui
# TUI banner appears (OpenClaw 2026.4.24, heartbeat info, etc.)
# In TUI:
hello
Hello! 👋 How can I help you today?
# Status bar:
connected | idle  agent main | session main (openclaw-tui) | inference/qwen3.6:35b | tokens ?/131k

Suggested Fix

• Ensure the Ollama-backed provider on DGX Spark reports token usage in a way the OpenClaw gateway can surface to the TUI, similar to cloud providers.
• Update the TUI so when using a local Ollama model like inference/qwen3.6:35b, the status bar shows numeric token counts (e.g. tokens 42/131k) instead of ?.
• If numeric token accounting is intentionally unavailable for local providers (e.g. due to Ollama API limitations):
  • Document this explicitly for DGX Spark + Ollama.
  • Update the DGX Spark Express QA test case to not require numeric token accounting for Ollama, or provide an alternative validation mechanism.

NVB#6196783