Bug Description

When using ollama/kimi-k2.6:cloud as the model, assistant responses leak raw tool call blocks and chain-of-thought text to the user in Telegram.

What Happened

During a conversation about researching a gymnastics club for a child, the assistant (powered by kimi-k2.6:cloud) produced a response that included:

1. Chain-of-thought text in the user-visible text block: "Sending an email to inquire about gymnastics for Sophia. Let me draft it in German..."
2. All 10 tool_call blocks (web_search, web_fetch calls) were rendered as visible output in Telegram

The user saw the entire internal reasoning and tool call flow instead of just the final response.

Evidence

Session transcript shows the problematic message at line 7960:

• role: assistant
• model: kimi-k2.6:cloud
• provider: ollama
• Content contains 1 text block + 10 toolCall blocks
• The text block starts with internal reasoning ("Sending an email to inquire...") before the actual user-facing content

Session ID: 773b51bd-8ba4-41d0-92ec-25cf57ca22e5

Expected Behavior

• Tool calls should be executed by OpenClaw and their results incorporated into the response
• Chain-of-thought reasoning should NOT be included in the text block sent to the user
• Only the final, polished response should be delivered to the Telegram channel

Environment

• OpenClaw 2026.4.15
• Model: ollama/kimi-k2.6:cloud (via Ollama Cloud)
• Channel: Telegram
• The same conversation with glm-5.1:cloud did NOT exhibit this behavior

Additional Context

The model kimi-k2.6:cloud was added to the config on 2026-04-23. The tool call leaking only appeared when this model was active. After switching back to glm-5.1:cloud, the behavior returned to normal.

This may be related to how Kimi K2.6 structures its tool call output differently from other models, or how OpenClaw parses the response from the Ollama Cloud provider for this specific model.