Summary

After upgrading from OpenClaw 2026.2.14 to 2026.2.15, Telegram responses are delivered word-by-word ("tröpfelnd") with severe performance degradation. The first word arrives immediately, but subsequent output is extremely slow, making the system practically unusable for Telegram interactions.

Steps to Reproduce

1. Setup:

   • OpenClaw 2026.2.15 installed on Ubuntu 22.04 LTS (Hetzner vServer)
   • Telegram bot configured with standard settings
   • Default model: deepseek/deepseek-chat (cost optimization)
   • Vision model for screenshots: google/gemini-2.5-flash

2. Configuration:

   "channels": {
     "telegram": {
       "enabled": true,
       "dmPolicy": "pairing",
       "groupPolicy": "allowlist",
       "streamMode": "partial"
     }
   }

3. Reproduction:

   • Send any message to the Telegram bot
   • Observe response delivery pattern

Expected Behavior

• Responses should be delivered in reasonable chunks or as complete messages
• Performance should be similar to version 2026.2.14
• Streaming should be smooth and responsive

Actual Behavior

• Word-by-word delivery: Responses arrive one word at a time
• Severe latency: Significant delays between words
• Performance impact: Makes conversations extremely slow and frustrating
• Regression: Version 2026.2.14 worked perfectly

Technical Analysis

Identified Code Changes

Through git analysis, two commits appear relevant:

1. dddb1bc94 - "fix(telegram): fix streaming with extended thinking models overwriting previous messages"

   • Changes in src/telegram/bot-message-dispatch.ts
   • Adds logic to prevent error payloads from overwriting preview messages
   • Introduces hasStreamedMessage flag

2. c62b90a2b - "fix(telegram): stop block streaming from splitting messages when streamMode is off"

   • Changes logic for disableBlockStreaming
   • Modified condition: draftStream || streamMode === "off" ? true : undefined

Hypothesis

The changes intended to prevent message overwriting may have introduced aggressive buffering or changed the streaming behavior, causing the word-by-word delivery.

Impact and Severity

• Critical: Makes Telegram integration practically unusable
• Regression: Working functionality in 2026.2.14 broken in 2026.2.15
• User experience: Extremely poor, conversations become frustratingly slow

Workarounds Tested

1. Revert to 2026.2.14: ✅ Works perfectly
2. Change streamMode: Tested "partial", "block", "off" - ❌ No improvement
3. Adjust blockStreaming: ❌ No improvement

Environment

• OpenClaw version: 2026.2.15
• Operating system: Ubuntu 22.04.5 LTS
• Install method: npm
• Telegram configuration: Standard bot setup
• Models: deepseek/deepseek-chat (primary), google/gemini-2.5-flash (vision)

Additional Information

• The issue occurs with all message types (text responses, image analyses)
• No errors in logs, just extremely slow delivery
• Direct API calls to models work normally (not a model performance issue)
• Issue is specific to Telegram channel output

Request

Please investigate the streaming/buffering changes in 2026.2.15 and provide a fix to restore normal response delivery performance in Telegram.

Priority: High - this is a critical regression affecting core functionality.