Summary

When an agent runs multiple tool calls in a single turn, intermediate narration text gets sent as separate Telegram messages even though it's part of the same response. Results in a flood of near-identical messages.

Steps to reproduce

1. Agent connected to Telegram channel (direct chat)
2. Agent performs a multi-step task requiring several sequential tool calls with narration between them
3. Each narration block between tool calls is sent as a separate Telegram message instead of being batched into the final response

Expected behavior

Only the final composed reply is sent to Telegram. Intermediate reasoning/narration between tool calls should not be delivered as separate messages.

Actual behavior

Every text block between tool calls is sent as an individual Telegram message, resulting in 8+ near-identical messages flooding the chat (see screenshot — repeated variations of "The API key format might need different auth").

OpenClaw version

2026.2.16

Operating system

Linux 6.12.69+deb13-cloud-amd64 (Debian, Hetzner VPS)

Install method

git

Logs, screenshots, and evidence

Impact and severity

Moderate — clutters the chat, confusing for the human, makes the agent look broken. Happens whenever multi-step tool work includes narration.

Additional information