openclaw - 💡(How to fix) Fix [Bug]: Internal Reasoning Markers Leaking to Chat Output [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

During an agent session with thinking=medium, internal reasoning markers that should be suppressed were rendered into the visible chat output. The user observed text like:

[Krilling... 🛠️ Exec: run which gogcli → find…~ -> show first 3 lines → run true 🛠️ Exec 🛠️ Exec: run which gogcli → find…~ -> show first 3 lines → run true 🛠️ Exec: completed; command run…~ -> show first 3 lines → run true ]

This text appeared inline in the Telegram chat while the agent was processing a multi-step task (building the Sunday tech brief). The markers were visible to the end user and had to be manually dismissed/copied.

Steps to reproduce

1. Start a session with thinking=medium on ollama/kimi-k2.6:cloud
2. Trigger a multi-step agent workflow requiring several tool calls in sequence
3. Observe that internal planning/processing annotations occasionally appear in the chat output

Expected behavior

All internal reasoning, planning, and execution annotations should be completely hidden from the user-facing message stream.

Actual behavior

Internal "Krilling..." headers and 🛠️ tool execution labels leaked into the rendered assistant message.

OpenClaw version

2026.5.7

Operating system

Ubuntu 26 LTS

Install method

NPM Global

Model

Model: ollama/kimi-k2.6:cloud

Provider / routing chain

Openclaw gateway

Additional provider/model setup details

Environment: • OpenClaw version: 2026.5.7 (eeef486) • Model: ollama/kimi-k2.6:cloud • Runtime: OpenClaw Pi Default • Thinking mode: medium • Surface: Telegram direct chat • Host: ROSECLAW (Linux, local Ollama)

Logs, screenshots, and evidence

Krilling...
🛠️ Exec: run which gogcli → find…~ -> show first 3 lines → run true
🛠️ Exec
🛠️ Exec: run which gogcli → find…~ -> show first 3 lines → run true
🛠️ Exec: completed; command run…~ -> show first 3 lines → run true

Impact and severity

• Breaks user experience by exposing internal implementation details • Creates confusion about whether the agent is "talking to itself" or the user • May leak tool names or paths that shouldn't be visible

Additional information

Possible Cause: The thinking=medium reasoning stream may not be fully decoupled from the final message assembly pipeline when the model generates tool calls in rapid succession.