Summary
When using openclaw agent to process simple chat messages, the model returns empty content (payloads: []) with stopReason: "stop" and only ~11 output tokens. The agent reports "completed" but sends no actual text back to the caller.
Root Cause
openclaw agent sends 23 tool definitions to the NVIDIA API alongside every user message. When using nvidia/llama-3.3-nemotron-super-49b-v1 (and nemotron-3-super-120b-a12b), the model interprets these tools as available actions and generates a tool call (11 tokens) instead of a text response. The openclaw agent then captures an empty content array.
Reproduction
# This returns empty content via openclaw:
openclaw agent --agent main -m "hello" --session-id test --json
# Output: payloads: [], stopReason: "stop", outputTokens: 11
# But the same model responds perfectly when called directly without tools:
curl -X POST "https://integrate.api.nvidia.com/v1/chat/completions" \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-d '{"model":"nvidia/llama-3.3-nemotron-super-49b-v1","messages":[{"role":"user","content":"hello"}],"max_tokens":200}'
# Output: {"content": "Hi! How can I help you today?", ...}Environment
- NemoClaw: installed via
nemoclaw install --non-interactive (March 2026)
- OpenClaw: 2026.3.11
- Server: Ubuntu 22.04 on Hetzner VPS
- Models tested:
nvidia/llama-3.3-nemotron-super-49b-v1, nvidia/nemotron-3-super-120b-a12b
Expected Behavior
Simple chat messages should return text responses, not tool calls. Either:
- Tool definitions should not be sent for simple conversational queries, or
- The model should be prompted/instructed to prefer text responses over tool calls when the user message is conversational, or
- The agent should fall back to extracting text if the model returns a tool call for a non-tool-related query
Workaround
Calling the NVIDIA API directly without tool definitions produces correct text responses. We built a custom Telegram bridge that bypasses openclaw agent entirely.
Summary
When using
openclaw agentto process simple chat messages, the model returns empty content (payloads: []) withstopReason: "stop"and only ~11 output tokens. The agent reports "completed" but sends no actual text back to the caller.Root Cause
openclaw agentsends 23 tool definitions to the NVIDIA API alongside every user message. When usingnvidia/llama-3.3-nemotron-super-49b-v1(andnemotron-3-super-120b-a12b), the model interprets these tools as available actions and generates a tool call (11 tokens) instead of a text response. The openclaw agent then captures an empty content array.Reproduction
Environment
nemoclaw install --non-interactive(March 2026)nvidia/llama-3.3-nemotron-super-49b-v1,nvidia/nemotron-3-super-120b-a12bExpected Behavior
Simple chat messages should return text responses, not tool calls. Either:
Workaround
Calling the NVIDIA API directly without tool definitions produces correct text responses. We built a custom Telegram bridge that bypasses
openclaw agententirely.