Summary

In OpenClaw 2026.4.27, openai-codex/gpt-5.5 appears to support xhigh at the provider level, but gateway sessions reject or downgrade xhigh to high. The shared thinking resolver behaves as if the openai-codex provider profile is not present in the active gateway plugin registry.

This affects normal gateway sessions, not just CLI help text or display output.

Environment

• OpenClaw: 2026.4.27
• Runtime: Docker gateway
• Platform: Linux arm64 / Raspberry Pi
• Model: openai-codex/gpt-5.5
• Config has thinkingDefault: "xhigh"
• plugins.entries.openai.enabled: true

Current behavior

A normal gateway-backed agent turn with --thinking xhigh fails validation before falling back locally:

GatewayClientRequestError: Error: Thinking level "xhigh" is not supported for openai-codex/gpt-5.5. Use one of: off, minimal, low, medium, high.

A live sessions.list result for a gpt-5.5 / openai-codex session reports:

{
  "modelProvider": "openai-codex",
  "model": "gpt-5.5",
  "thinkingDefault": "xhigh",
  "thinkingLevel": "high",
  "thinkingLevels": ["off", "minimal", "low", "medium", "high"]
}

So thinkingDefault remains xhigh, but the effective level is silently/remapped to high, and the advertised supported levels omit xhigh.

Expected behavior

For openai-codex/gpt-5.5, gateway validation and session metadata should include and accept xhigh, matching the OpenAI Codex provider's own thinking profile.

Expected levels:

["off", "minimal", "low", "medium", "high", "xhigh"]

Reproduction

1. Configure an agent/session to use openai-codex/gpt-5.5.
2. Set thinkingDefault to xhigh, or run an agent turn with --thinking xhigh.
3. Observe that the gateway rejects xhigh or stores the session as thinkingLevel: "high".
4. Query session metadata via sessions.list and observe that thinkingLevels omits xhigh.

I also tried the native config path:

openclaw plugins enable openai
# restart/recreate gateway

That did not change the behavior. The config already had plugins.entries.openai.enabled: true; the enable command only changed metadata.

Local diagnostic evidence

The bundled OpenAI Codex provider itself advertises xhigh for gpt-5.5 when instantiated directly. In the installed bundle, the provider has an xhigh model list including gpt-5.5, and its resolveThinkingProfile({ modelId: "gpt-5.5" }) returns levels including xhigh.

However, the shared thinking resolver only includes xhigh if the active plugin registry exposes the provider profile. A direct Node check showed this split:

direct provider profile: includes xhigh
shared resolver without registry provider: off,minimal,low,medium,high
shared resolver with provider injected into active registry: off,minimal,low,medium,high,xhigh

Gateway startup logs also showed the startup plugin list did not include openai even though the OpenAI plugin is enabled in config:

http server listening (8 plugins: acpx, bonjour, browser, device-pair, memory-core, phone-control, talk-voice, telegram; ...)

The bundled OpenAI plugin manifest has:

"activation": {
  "onStartup": false
}

Suspected cause

The gateway thinking validation/session metadata path calls the shared thinking resolver before the openai provider plugin has been activated/imported into globalThis[Symbol.for("openclaw.pluginRegistryState")].activeRegistry.providers.

Because the provider-specific resolveThinkingProfile is not visible in that registry, the shared resolver falls back to the base thinking levels and omits xhigh.

Suggested fix direction

The fix likely belongs in the shared thinking/provider-profile plumbing rather than individual callers:

• Ensure provider plugins needed for configured/default models are activated before gateway thinking validation and session metadata resolution, or
• Make the thinking resolver use the canonical provider lookup/runtime provider registry instead of only the currently active startup registry, or
• Add a provider-triggered activation path for thinking profile resolution.

A regression test should cover the gateway path, not only direct provider instantiation:

• listThinkingLevelOptions("openai-codex", "gpt-5.5") includes xhigh in a gateway/session-validation context.
• A gateway agent turn with --thinking xhigh --model openai-codex/gpt-5.5 is accepted.

Impact

Users who configure xhigh for OpenAI Codex GPT-5.x models get silently downgraded to high in gateway sessions, or receive a gateway validation error for explicit --thinking xhigh turns.