-
Notifications
You must be signed in to change notification settings - Fork 2.2k
[Bug] Agent fabricates successful execution output after tool failure (exec: python not found) #1452
Description
Description
When a tool call fails inside the NemoClaw sandbox, the OpenClaw agent
silently discards the error and fabricates a plausible-looking success
response — including fake file paths, fake execution output, and a
confirmation that work was saved. The user has no way to distinguish
real results from hallucinated ones.
Root cause likely in OpenClaw's agent loop, but surfacing here because
operator-side visibility of tool failures is within NemoClaw/OpenShell's
scope. A python → python3 symlink in the sandbox base image would also
prevent this specific trigger.
Reproduction Steps
-
Run
nemoclaw onboardand connect to the sandbox -
Inside the sandbox, run:```bash
openclaw agent --agent main --local
-m "write a python script that reads a CSV file and outputs basic statistics"
3. Observe agent output
Tool call fails:
[tools] exec failed: /bin/bash: line 1: python: command not found
Agent then outputs a full fabricated success response — fake column
statistics, fake row counts — and confirms *"the script is ready to
use and saved in your workspace."
4. Verify the filesystem:```bash
ls -la /sandbox/csv_stats.py 2>/dev/null || echo "FILE NOT FOUND"
ls -la /sandbox/sample.csv 2>/dev/null || echo "FILE NOT FOUND"
Environment
- NemoClaw: early preview (March 2026)
- OpenClaw: 2026.3.11 (29dc654)
- OpenShell: 0.0.10
- Platform: Ubuntu (Brev cloud instance), no GPU
- Inference: NVIDIA Endpoints (cloud)
- Sandbox: Landlock + seccomp + netns
Debug Output
Logs
Checklist
- I confirmed this bug is reproducible
- I searched existing issues and this is not a duplicate