elder roost
I’ve got OpenClaw running locally with Ollama and the WhatsApp channel linked (self‑chat to my own number works, messages go through both ways).
Problem: instead of answering conversationally, my WhatsApp agent mostly responds by analyzing OpenClaw logs/metadata and describing tools like read, write, web_fetch, sessions_list, etc. It will even “answer” with JSON tool calls like a fake web_fetch to an API that doesn’t exist, instead of talking to me.
In openclaw.json I’ve already set "tools": { "profile": "minimal" }, dmPolicy: "allowlist" with my number in allowFrom, and a binding that routes my WhatsApp DM to a whatsapp-chat agent whose AGENTS.md prompt is explicitly conversational. The WhatsApp channel status shows linked, running, connected, and test messages from the CLI reach my phone fine.
Question: what else do I need to change so the WhatsApp agent stops treating internal runtime/tool metadata as the thing to analyze, and just behaves like a normal chat assistant? Is there a known setting or pattern to strip that context for WhatsApp, or should I be using a separate workspace/agent setup for this?
@fleet cairn