openclaw - 💡(How to fix) Fix [Bug]: Codex model catalog cold-start miss for gpt-5.3-codex after gateway restart

Bug Description

After a gateway restart, the Codex model catalog is not populated on first call. When the first session attempts to resolve openai/gpt-5.3-codex, the harness fails with:

Unknown model: openai-codex/gpt-5.3-codex

This causes an immediate fallback to the first configured fallback model (in our case microsoft-foundry/FW-MiniMax-M2.5). Sessions that hit this cold-start window become stuck on the fallback model until manually reset.

Reproduction

1. Restart the OpenClaw gateway
2. Before any other session resolves the Codex catalog, invoke a session that uses openai/gpt-5.3-codex
3. The session falls back to a non-Codex model with error: "model not found"

Observed Occurrences

• 2026-05-28: Stale plugin version (v2026.5.22) caused provider mismatch — fixed by plugin update (GH #87650)
• 2026-05-30: Plugin version WAS correct (v2026.5.28), but catalog was not warm after restart. Same symptom, different root cause.

Environment

• OpenClaw v2026.5.28
• @openclaw/codex plugin v2026.5.28
• Primary model: openai/gpt-5.3-codex
• 8 fallback models configured
• macOS arm64, Node 24.15.0

Expected Behavior

The Codex model catalog should be populated (or populated on-demand) when a session requests gpt-5.3-codex, without requiring a prior warmup call. If there is a cold-start delay, the model resolution should retry or wait rather than immediately falling back to a completely different provider.

Notes

• v2026.5.28 release notes mention fixing GPT-5.5 catalog resolution (#84362), but the same fix was not applied to gpt-5.3-codex.
• This affects any deployment that restarts the gateway and has user-facing sessions before the catalog warms up.
• Workaround: send a warmup message after every gateway restart, and check sessions.json for stale fallback entries.