Description

When the agent processes a request from a messaging channel (iMessage, Telegram, etc.) and makes multiple tool calls, any text generated between tool calls is sent to the source channel as a message. This leaks internal narration/thinking text to end users.

Steps to Reproduce

1. Send a message to the agent via any messaging channel (iMessage, Telegram, etc.)
2. The agent processes the request using multiple sequential tool calls
3. Text the agent generates between tool calls (narration, thinking, processing notes) gets sent as actual messages to the channel

Expected Behavior

Only the agent's final, intentional reply should be sent to the messaging channel. Internal narration between tool calls should be suppressed.

Actual Behavior

Internal processing text like:

• "Let me check which chat this is from."
• "The model has named ranges. Now I'll rebuild the analysis sheets with dynamic formulas."
• "Catbox is flaky. Let me just email Joe the files instead:"
• Error messages from failed tool calls

...all get sent as real messages to the user/group chat.

Impact

• Confusing for users who receive walls of internal processing text
• Leaks implementation details and thought process
• Particularly problematic in group chats where multiple non-technical users see the narration
• Affects ALL messaging channels (iMessage, Telegram, Slack, BlueBubbles, etc.) since the bug is upstream of the channel layer

Environment

• OpenClaw version: 2026.2.6-3
• OS: macOS (Darwin 25.2.0, arm64)
• Model: anthropic/claude-opus-4-6
• Channels affected: iMessage, Telegram (all channels)