fix: fallback to config.llm.model when resolveAgentPrimaryModelRef returns undefined

Root cause: When generateReflectionText() is triggered via plugin hooks
(dc-channel--...), the plugin-scoped config lacks an 'agents' section.
resolveAgentPrimaryModelRef() returns undefined, causing
runEmbeddedPiAgent to receive provider=undefined, model=undefined.
This falls back to openai/gpt-5.4 (default), which has no API key,
resulting in a 56-second timeout before minimal fallback activation.

Fix: When resolveAgentPrimaryModelRef returns undefined, fall back to
config.llm.model (the MiniMax M2.1 configured in plugin's openclaw.json).
This ensures correct model is used for both agent:main (which gets full
config and resolveAgentPrimaryModelRef works) and dc-channel--... sessions
(which get plugin-scoped config).

Fixes the 56-second delay on /reset for dc-channel Discord sessions.

Author: jlin53882

index.ts

@@ -1154,8 +1154,9 @@ async function generateReflectionText(params: {
       onLog: onRetryLog,
       execute: async () => {
         const runEmbeddedPiAgent = await loadEmbeddedPiRunner();
-        const modelRef = resolveAgentPrimaryModelRef(params.cfg, params.agentId);
-        const { provider, model } = modelRef ? splitProviderModel(modelRef) : {};
+        const modelRef = resolveAgentPrimaryModelRef(params.cfg, params.agentId)
+          ?? (((params.cfg as Record<string, unknown>)?.llm as Record<string, unknown>)?.model as string | undefined);
+        const { provider, model } = modelRef ? splitProviderModel(modelRef) : { provider: void 0, model: void 0 };
         const embeddedTimeoutMs = Math.max(params.timeoutMs + 5000, 15000);
 
         return await withTimeout(