Description

When using a model with extended thinking (e.g. Claude Opus) and Telegram streaming (new in v2026.2.15), text output after a thinking block overwrites the previous message instead of creating a new one.

Steps to Reproduce

1. Configure OpenClaw with Telegram channel and streaming enabled (default in v2026.2.15)
2. Switch model to one with extended thinking (e.g. anthropic/claude-opus-4-6)
3. Send a message that triggers: text output → tool call → thinking → more text output
4. Observe in Telegram

Expected Behavior

After a tool call + thinking block, the new text output should be sent as a new Telegram message, preserving the previous message.

Actual Behavior

The new text output edits/overwrites the previous Telegram message. The user loses the first part of the response before they can read it.

Workaround

• View the full response in the OpenClaw desktop app (all text visible there)
• Disable reasoning (/reasoning toggle)
• Use a model without extended thinking (e.g. Claude Sonnet)

Environment

• OpenClaw: v2026.2.15 (upgraded from v2026.2.14)
• Channel: Telegram (DM)
• Model: anthropic/claude-opus-4-6 (thinking: low)
• OS: macOS (ARM64)