Bug Description

The voice-call plugin's responseModel config option is ignored at runtime. When responseModel is set to a fast model (e.g., openai/gpt-4o-mini or openai/gpt-4.1-nano) but the global agent default model is different (e.g., anthropic/claude-opus-4-6), the embedded PI runner throws LiveSessionModelSwitchError and the voice response is never generated.

Steps to Reproduce

1. Set global default model to anthropic/claude-opus-4-6 in openclaw.json
2. Configure voice-call plugin with responseModel: "openai/gpt-4o-mini"
3. Enable streaming (STT + TTS)
4. Make an outbound call in conversation mode
5. Caller speaks — STT transcribes successfully
6. Response generation fails with:

[voice-call] Response generation failed: LiveSessionModelSwitchError: Live session model switch requested: anthropic/claude-opus-4-6

Expected Behavior

The voice-call plugin should use the model specified in responseModel config for generating voice responses, regardless of the global default model. Voice conversations need fast models (gpt-4o-mini, gpt-4.1-nano) for acceptable latency — forcing claude-opus adds 3-5 seconds per turn, making phone calls unusable.

Actual Behavior

resolveLiveSessionModelSelection() in live-model-switch.ts resolves the global default model (claude-opus), detects a mismatch with the responseModel passed to runEmbeddedPiAgent(), and throws LiveSessionModelSwitchError. The response is never generated, and the caller hears silence.

Root Cause Analysis

In response-generator.ts:219-224, the responseModel is correctly resolved:

const modelRef = voiceConfig.responseModel || `${agentRuntime.defaults.provider}/${agentRuntime.defaults.model}`;

But runEmbeddedPiAgent() internally calls resolveLiveSessionModelSelection() which reads from the session store or global defaults — not from the passed provider/model parameters. When it detects a mismatch, it aborts with LiveSessionModelSwitchError.

Workaround

Manually pin the session's providerOverride and modelOverride in the session store before calling runEmbeddedPiAgent():

if (voiceConfig.responseModel && sessionEntry) {
  sessionEntry.providerOverride = provider;
  sessionEntry.modelOverride = model;
  await agentRuntime.session.saveSessionStore(storePath, sessionStore);
}

Environment

• OpenClaw version: 2026.3.28 (f9b1079)
• macOS (Apple Silicon)
• Voice-call plugin with Twilio provider
• Streaming enabled (OpenAI Realtime STT + ElevenLabs TTS)

Impact

This bug makes voice calls practically unusable when the global default model is a high-latency model like claude-opus. Users cannot use fast models for voice responses even though the responseModel config exists for exactly this purpose.

Suggested priority: p1 — voice-call is non-functional for users with non-OpenAI default models.