openclaw - 💡(How to fix) Fix MiniMax-only configuration still invokes claude-cli/codex-cli and can hang TUI/model execution [2 comments, 3 participants]

Summary

With OpenClaw 2026.4.26 configured to use only minimax/MiniMax-M2.7, the TUI / agent execution path still invokes claude-cli and later codex-cli. The run becomes extremely slow or hangs, and related OpenClaw processes can consume high CPU.

Environment

• OpenClaw: 2026.4.26
• macOS: Darwin 25.4.0 arm64
• Node.js: v24.11.1
• Package manager path: pnpm global install
• Configured default model: minimax/MiniMax-M2.7
• openclaw models status --json reports:
  • defaultModel: minimax/MiniMax-M2.7
  • resolvedDefault: minimax/MiniMax-M2.7
  • fallbacks: []
  • allowed: [minimax/MiniMax-M2.7]

Expected Behavior

When the default model and allowed model list are MiniMax-only, OpenClaw should not invoke claude-cli, OpenAI embedded models, or codex-cli unless those are explicitly configured as fallbacks or enabled for that run.

Actual Behavior

The runtime logs show OpenClaw reading Claude CLI credentials and invoking Claude/Codex paths despite MiniMax-only model configuration:

[agents/auth-profiles] read anthropic credentials from claude cli keychain
[agent/cli-backend] cli exec: provider=claude-cli model=opus promptChars=505
[agent/cli-backend] claude live session start: provider=claude-cli model=claude-opus-4-7
[agent/cli-backend] claude live session turn failed: provider=claude-cli model=claude-opus-4-7 durationMs=92374 error=FailoverError
[agent/embedded] embedded_run_failover_decision ... provider=openai model=gpt-5.5 ... fallbackConfigured=false timedOut=true
[agent/cli-backend] cli exec: provider=codex-cli model=gpt-5.5
[agent/cli-backend] cli watchdog timeout: provider=codex-cli model=gpt-5.5 noOutputTimeoutMs=9000

A minimal local inference test also hangs:

openclaw infer model run --local --model minimax/MiniMax-M2.7 --prompt ping --json

During that test, openclaw-infer reached high CPU usage and did not return promptly.

Additional Observations

• openclaw models status --json shows MiniMax as the only configured/allowed model.
• ~/.openclaw/agents/main/agent/models.json contains a codex provider entry as well as MiniMax, even though the status command reports only MiniMax as allowed.
• The issue appears to be in internal planner / cli-backend / failover routing, not in the user-selected default model itself.
• Gateway startup also loads many bundled plugins and Feishu/channel components even for seemingly lightweight commands, but the main bug here is model routing violating the MiniMax-only expectation.

Impact

The terminal UI becomes very slow or appears stuck after input. OpenClaw can spawn long-running or high-CPU processes, requiring manual cleanup with pkill/kill -9.