Skip to content

[Ubuntu 24.04][Agent&Skills] Ultra 550B content omits intermediate steps when no tools configured — only final command returned #4851

@hulynn

Description

@hulynn

Description

When nvidia/nemotron-3-ultra-550b-a55b is asked to perform a multi-step task (e.g. create a file then run it) with no file-write or bash-execute tools configured, the model's reasoning_content correctly plans all steps, but the content field only outputs the final step (the run command), silently skipping the file-creation step. The user receives an incomplete and non-functional response with no indication that intermediate steps were omitted. This differs from Nano 30B which hallucinates policy restrictions; Ultra 550B plans correctly but silently truncates its own action output.

Environment

Device:        Ubuntu 24.04 server (galaxy-sku2-018, aarch64)
OS:            Ubuntu 24.04.4 LTS
Architecture:  aarch64
Node.js:       v22.22.3
npm:           10.9.8
Docker:        29.5.2
OpenShell CLI: 0.0.44
NemoClaw:      v0.0.59
OpenClaw:      2026.5.27 (27ae826)
Model:         nvidia/nemotron-3-ultra-550b-a55b (sandbox: ultra-test)

Steps to Reproduce

  1. nemoclaw onboard with nvidia/nemotron-3-ultra-550b-a55b (NVIDIA Endpoints)
  2. nemoclaw ultra-test connect && openclaw tui
  3. Send: Create a file called hello.py in /tmp with a hello world script, then run it.
  4. Observe model response and API-level reasoning_content vs content fields

Expected Result

Either:

  • Model explains it lacks a file-write tool and shows the full code the user would need to run manually, or
  • All planned steps appear in content — both the file creation code and the run command.

Actual Result

API response fields (verified via direct curl to inference.local):

reasoning_content:

"The user wants me to:
 1. Create a file called hello.py in /tmp
 2. Put a hello world script in it
 3. Run the script
 Let me do this step by step."

content:

python3 /tmp/hello.py

finish_reason: stop

The reasoning plans 3 steps but content only outputs the final "run" command. Steps 1 and 2 (creating the file with content) are silently absent. The user sees only python3 /tmp/hello.py with no file to run.

Logs

Direct API verification (curl to inference.local, max_tokens=400):
  HTTP 200 | TTFB: 8.6s | finish_reason: stop
  usage: { "prompt_tokens": 35, "completion_tokens": 101, "total_tokens": 136 }

  reasoning_content:
    "The user wants me to: 1. Create a file... 2. Put a hello world script...
     3. Run the script. Let me do this step by step."

  content:
    "\n```bash\npython3 /tmp/hello.py\n```"

openclaw.json tools section:
  {"tools": {"toolSearch": true, "web": {"fetch": {"enabled": true}}}}
  No bash_execute or write_file tools configured.

NVB#6272828

Metadata

Metadata

Assignees

Labels

NV QABugs found by the NVIDIA QA Teamarea: inferenceInference routing, serving, model selection, or outputsplatform: ubuntuAffects Ubuntu Linux environmentsv0.0.64Release target

Type

No fields configured for Bug.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions