openclaw - 💡(How to fix) Fix [bug] #82744 fix incomplete: /think xhigh still rejected on canonical openai/* (only openai-codex catalog patched)

Thanks for the quick turnaround on #82744 (PR #82761 in v2026.5.16-beta.4). The bundled openai-codex catalog now exposes xhigh, and agents.defaults.thinkingDefault: "xhigh" continues to apply at config-time. However, the runtime /think directive is still rejected on canonical openai/* model refs that get routed to Codex via auth.order.openai Codex-first.

Setup

• Version: 2026.5.16-beta.4
• Config:

  "agents.defaults.model.primary": "openai/gpt-5.5",
  "agents.defaults.models": {
    "openai/gpt-5.5": { "agentRuntime": { "id": "pi" } },
    "openai/gpt-5.4": { "agentRuntime": { "id": "pi" } }
  },
  "agents.defaults.thinkingDefault": "xhigh",
  "auth.order.openai": ["openai-codex:<email>", "openai:default"]

• /status confirms: Current thinking level: xhigh ✅ (config default applies)
• Actual routing per ↪️ Fallback line: openai/gpt-5.5 → openai-codex/gpt-5.5 ✅ (auth.order Codex-first works as of #82605)

Reproducer

1. Run /think (interactive picker).
2. Expected: full reasoning options matching the catalog now declared for openai-codex/gpt-5.5 (i.e. at least off / minimal / low / medium / high / xhigh / adaptive).
3. Actual:

   Current thinking level: xhigh.
   Choose level for /think.
   Options: default, off.

Why this surfaces

The fix in PR #82761 added reasoning metadata to the openai-codex bundled catalog. The runtime /think directive capability check appears to resolve against the configured model id (openai/gpt-5.5) at the agent-defaults level, not against the post-auth.order routed Codex model id. The canonical openai catalog still does not declare xhigh for gpt-5.5 / gpt-5.4, so the picker only offers default, off.

thinkingDefault: xhigh (config-time) bypasses this directive-handler capability check, which is why the current level correctly reads xhigh — the bug only blocks runtime /think changes, not config-time defaults.

Suggested fix (two options)

1. Mirror the metadata into the canonical openai catalog for gpt-5.5 / gpt-5.4 (and probably gpt-5.4-mini), since auth.order.openai-routed PI runs land on the same chatgpt.com backend models. Cleanest for users who configure canonical model ids and rely on auth-order routing.
2. Have the /think directive capability check follow auth.order.* routing when present, so it resolves against the effective Codex catalog rather than the literal configured model id.

Workaround

agents.defaults.thinkingDefault: "xhigh" in config works (verified end-to-end). Runtime /think directive is the only path still affected. Acceptable for users who pin a single level; blocks per-turn thinking-level changes for the canonical-model + auth-order setup.

Related

• Original: #82744
• Closing PR: #82761
• Auth-order Codex-first routing fix that made canonical openai/* viable: #82605 / #82521