Problem

The global agents.defaults.thinkingDefault setting applies to all embedded agent runs, including the voice-call plugin's response generation. When set to "high", this causes two critical issues for voice calls:

1. Excessive latency: Extended thinking adds 10-30+ seconds before the first response token, making voice calls unusable (callers hang up before the model finishes thinking)
2. Behavioral changes: With thinking=high, the model sometimes calls the tts tool instead of returning plain text, even when the system prompt explicitly says "Never call the tts tool". The voice-call plugin expects plain text to route through ElevenLabs TTS — tool calls produce no audio output

Current Behavior

embedded run start: provider=anthropic model=claude-haiku-4-5 thinking=high messageChannel=voice
embedded run tool start: tool=tts  ← model calls tts despite system prompt prohibition
embedded run agent end: isError=false  ← "succeeds" but caller hears nothing

The resolveThinkingDefault() function returns the global default for all runs with no way to override per-agent:

function resolveThinkingDefault(params) {
  const configured = params.cfg.agents?.defaults?.thinkingDefault;
  if (configured) return configured;
  // ...
}

Expected Behavior

Allow per-agent thinkingDefault override in the agent list entries:

{
  agents: {
    defaults: { thinkingDefault: "high" },  // global default for coding agents
    list: [
      {
        id: "voice",
        thinkingDefault: "off",  // voice agent override — currently rejected as "Unrecognized key"
        model: { primary: "anthropic/claude-sonnet-4-6" }
      }
    ]
  }
}

Workaround Attempts

• Setting thinking or thinkingDefault on the agent entry → rejected by strict schema validation ("Unrecognized key")
• Switching to Haiku 4.5 (no thinking support) → thinking=high still sent, causes model to call tts tool
• Setting global thinkingDefault: "off" → fixes voice but degrades all other agents
• Resetting voice session → no effect

Environment

• OpenClaw v2026.2.24 (stable)
• macOS, LaunchAgent-managed gateway
• Voice pipeline: Twilio → OpenAI Realtime STT → LLM → ElevenLabs TTS

Related

• Discussion Reducing /v1/chat/completions latency for real-time voice agents #10588 (reducing /v1/chat/completions latency for real-time voice agents)