Summary

When I send messages in Japanese, the assistant replies sometimes include Chinese characters, and occasionally Russian text, even though I explicitly forbid Chinese and Russian and require Japanese-only output in SOUL.md / AGENTS.md / IDENTITY.md.

Steps to reproduce

1. Authenticate with MiniMax Coding Plan via OAuth.
2. Select model: MiniMax M 2.5.
3. Add language rules to SOUL.md / AGENTS.md / IDENTITY.md, e.g.:
   • "Always respond in Japanese."
   • "Never use Chinese."
   • "Never use Russian."
4. Restart the gateway: openclaw gateway restart
5. Start a fresh session: /new
6. Send normal Japanese messages (questions and coding requests).

Expected behavior

If I send messages in Japanese, the assistant should respond in Japanese only (no Chinese/Russian characters).

Actual behavior

Even when I send messages in Japanese, the assistant sometimes returns replies containing Chinese characters, and sometimes Russian text is mixed in as well.

OpenClaw version

2026.2.14

Operating system

macOS 15.7.3

Install method

No response

Logs, screenshots, and evidence

Impact and severity

This makes OpenClaw unusable for Japanese coding assistance and documentation, because the output language cannot be controlled reliably.

Workarounds tried

• Added "Japanese only / never use Chinese / never use Russian" rules to SOUL.md / AGENTS.md / IDENTITY.md
• openclaw gateway restart
• /new to start a fresh session
  No improvement.

Question / Help wanted

Is there a supported way to strictly enforce Japanese-only output when using MiniMax M 2.5 via Coding Plan OAuth?
If strict enforcement is not possible, what is the recommended OpenClaw-side approach (e.g., output validation + automatic retry) to prevent Chinese/Russian from appearing in the final response?