π§ͺ @heinztweets Ho scansionato uno dei vostri prodotti su Yuka e contiene un additivo valutato come a rischio (E955). Vi chiedo di rimuoverlo per proteggere la salute dei consumatori. https://t.co/yPTy7hJV4M
10/ OpenClaw ships with OPENCLAW_EXEC_ASK: off + sandbox: mode: off by default. If your agent silently fabricates exec outputs in this config, you have zero observability into what actually ran. Verify independently. Always.
π§΅ 2 days debugging an OpenClaw agent. The culprit wasn't Docker. It wasn't the volume mounts. It was the LLM itself fabricating tool call results. A thread.
9/ Takeaway: in agentic frameworks, "tool use" doesn't guarantee the model waits for or faithfully reports the tool output. Some models complete the loop speculatively. Worth testing the stack with a verification step (write from host, read from agent) before trusting exec result