Summary

When running onboarding with --auth-choice openai-codex, OpenClaw registers the Codex OAuth auth profile but leaves agents.defaults.model.primary unchanged (often still Anthropic). This creates a silent intent mismatch: users believe Codex is active, but runtime still targets the old primary model.

Why this matters

• During provider outages, users may think failover is broken while the system is actually still pinned to a different primary provider.
• TUI status can show the old primary model unless users manually inspect config.
• Recovery actions are delayed because behavior does not match onboarding intent.

Repro

1. Run onboarding with Codex auth choice:

   openclaw onboard --auth-choice openai-codex

2. Inspect config:

   openclaw config get agents.defaults.model.primary

3. Observe primary remains unchanged from previous value.

Expected

If onboarding auth choice implies a default provider/model selection, set agents.defaults.model.primary accordingly (or prompt explicitly and require confirmation).

Actual

Auth profile is added, but primary model remains old value.

Related issues

• Fallback does not escalate to different provider on overload errors #32533 (fallback/provider escalation under overload)
• 500/503 errors misclassified as rate_limit, triggering unnecessary cooldowns #22294 (error-classification/cooldown behavior)
• [Bug]: onboard --non-interactive silently mishandles --auth-choice in two ways #17191 (onboard auth-choice mismatch class)

Suggested fix

In onboarding flow, after successful provider auth setup:

1. Determine provider-appropriate default model alias
2. Set agents.defaults.model.primary (or ask for explicit opt-in)
3. Print a clear summary of resulting primary + fallbacks