name: Bug report
about: Report a problem or unexpected behavior in Clawdbot.
title: "[Bug]: Reasoning/thinking content leaking to Discord"
labels: bug

Summary

Internal reasoning/thinking content from the AI is being transmitted directly to Discord messages, visible to the user. This occurs with both reasoning models (kimi-k2.5:cloud) and non-reasoning models (glm-4.7).

Steps to reproduce

1. Start any conversation with Clawdbot via Discord
2. Send any message that requires the AI to process
3. Observe the response - internal reasoning/planning is visible before the final output
4. Issue persists across:
   • Model switches (kimi → GLM → kimi)
   • Gateway restarts
   • Session resets (/new, /reset)
   • Context compaction

Expected behavior

Only the final output (content within final tags) should be sent to Discord. Internal reasoning should be completely filtered and never visible to the user.

Actual behavior

The AI's internal monologue/reasoning is being broadcast directly to Discord. For example, the user can see messages like:

"The user is saying... I should... Let me check..."

This appears before or instead of the actual response.

Environment

• Clawdbot version: 2026.1.30 (76b5208)
• OS: Ubuntu (Linux 6.8.0-90-generic)
• Node: v22.22.0
• Install method: npm
• Primary model: ollama/kimi-k2.5:cloud
• Also affected: zai/glm-4.7 (non-reasoning model)

Logs or screenshots

Evidence from OpenClaw logs:

All runs show thinking=low parameter, even for non-reasoning models:

{"subsystem":"agent/embedded", "1":"embedded run start: ... model=kimi-k2.5:cloud thinking=low ..."}
{"subsystem":"agent/embedded", "1":"embedded run start: ... model=glm-4.7 thinking=low ..."}

Model configuration:

From openclaw.json:

• kimi-k2.5:cloud: has "reasoning": true
• glm-4.7: does NOT have reasoning field (non-reasoning model)

Yet BOTH are being run with thinking=low.

Session memory files:

Multiple session files created documenting this issue:

• memory/2026-02-01-reasoning-leak.md
• memory/2026-02-01-context-overflow.md

Additional context

The issue appears to have two components:

1. BUG fix: add @lid format support and allowFrom wildcard handling #1: thinking=low is being set on ALL model runs, including non-reasoning models like GLM that shouldn't have thinking enabled.

2. BUG Login fails with 'WebSocket Error (socket hang up)' ECONNRESET #2: When models return reasoning content (in the reasoning field), OpenClaw is not properly filtering it out before sending to the channel. The reasoning content is being transmitted alongside or instead of the final output.

This is a critical privacy/usability issue as the AI's internal planning is exposed to users, breaking the expected behavior of the system.