openclaw - 💡(How to fix) Fix claude-tmux harness: sourceVisibleReplies delivery silently drops responses in direct chats [1 comments, 2 participants]

Description

When an OpenClaw agent is switched to an Anthropic model (Opus, Sonnet, Haiku) via the UI model selector, any message sent produces a "thinking" indicator but the response never arrives — the message effectively disappears.

The backend processes the request correctly. The failure is in delivery — the claude-tmux harness has deliveryDefaults: { sourceVisibleReplies: "message_tool" } which is intended to route responses through the message tool, but the harness never actually delivers the response back to the source channel.

Steps to Reproduce

1. Open the OpenClaw web UI or Telegram bot
2. Switch from the default model to an Anthropic model (e.g. Opus 4.7) via the model selector
3. Send any message
4. The thinking indicator appears briefly, then the message disappears with no response

Root Cause

In ~/.openclaw/plugins/claude-tmux/harness.js line ~376:

deliveryDefaults: { sourceVisibleReplies: "message_tool" }

This setting instructs the harness to deliver replies via the message tool, but the harness code does not implement that delivery path for direct chat sessions. The response is silently dropped.

Expected Behavior

When an Anthropic model is selected and a message is sent, the response should be delivered back to the source channel (Telegram/web UI).

Proposed Fix

deliveryDefaults: { sourceVisibleReplies: "automatic" }

This ensures responses are delivered inline without relying on the unimplemented message_tool delivery path.

Environment

• OpenClaw: 2026.5.18
• Plugin: claude-tmux
• OS: Linux 6.17.x