🧠 Model Configuration

Configured Models

GLM-5.2 (Primary)

Thinking Modes

Kimi K2.6 (Fallback)

Failover Configuration

Configured in ~/.openclaw/openclaw.json:

"agents": {
  "defaults": {
    "model": {
      "primary": "glm/glm-5.2",
      "fallbacks": ["moonshot/kimi-k2.6"]
    },
    "imageModel": {
      "primary": "moonshot/kimi-k2.6"
    }
  }
}

When GLM returns a rate limit error, OpenClaw automatically retries on Kimi K2.6. If both fail, the run produces an error.

Model Disclosure in Discord

Rule: When responding in Discord, agents must include a model footer showing what model they used and whether they delegated to a harness:

📍 Model:  | Harness:  ()

Examples:

• 📍 Model: GLM-5.2 | Harness: self — plain self only when every line of the shipped code was written by the agent itself
• 📍 Model: GLM-5.2 | Harness: self (impl: grok-cli/grok-4.7) — a Grok CLI run wrote code that shipped (CLI default grok-4.7; use the model the run actually used)
• 📍 Model: GLM-5.2 | Harness: cursor (gpt-5.2-codex)
• 📍 Model: Kimi K2.6 | Harness: codex (claude-sonnet-4)

This gives Mike visibility into which models are doing the actual work.

Rate Limit Tips

• Pro Coding plan — rate limits are generous, thinking modes are safe to use
• Use /think high for per-session override, resets next session
• Spawn subagents to parallelize work
• Fallback to Kimi K2.6 happens automatically if GLM is unavailable

When to Use Thinking