Description

Description

When using Kimi K2.6 (moonshotai/kimi-k2.6) via NVIDIA Endpoints and the
agent attempts a tool call that fails (e.g. web search without BRAVE_API_KEY),
the model's internal reasoning/chain-of-thought text leaks into the user-
visible output. The user sees paragraphs of "But wait, the user might be
testing..." internal deliberation instead of a clean response.

This only happens with complex prompts that trigger reasoning. Simple
prompts (math, factual Q&A) do not leak.

Device:        MacBook Pro (Apple M4 Pro, 48 GB)
OS:            macOS 26.0.1
NemoClaw:      v0.0.36
OpenClaw:      2026.4.24
Model:         moonshotai/kimi-k2.6
Provider:      nvidia-prod (NVIDIA Endpoints)
Plugin:        kimi-inference-compat v0.1.0 (auto-installed)

1. Onboard with Kimi K2.6: NEMOCLAW_MODEL=moonshotai/kimi-k2.6
2. Do NOT configure BRAVE_API_KEY
3. Run: openclaw agent -m "Search the web: What is the latest NVIDIA GPU?"
4. Observe the output

Agent responds with a clean answer, noting that web search is unavailable.
Internal reasoning tokens are filtered before display.

Output includes internal chain-of-thought:
"But wait, the user might be testing the web search functionality..."
"Actually, re-reading: this means they want to know..."
"I need to be careful. The user might be asking about..."

Multiple paragraphs of reasoning appear before the actual answer.
PR #3046 (support reasoning models in OpenClaw harness) may need
to filter thinking tokens from the kimi-inference-compat stream.

Bug Details

[NVB#6154911]