openclaw - 💡(How to fix) Fix Document/support a Pro-plan path for gpt-5.5-pro and retired Spark via Codex CLI/app-server [1 comments, 2 participants]

Summary

OpenClaw currently documents and supports the normal Codex subscription path around openai/gpt-5.5, but there is no clear supported path for operators with a ChatGPT Pro subscription who want to use Pro-tier or special Codex models such as:

• gpt-5.5-pro
• gpt-5.4-pro
• retired/hidden Spark-style refs such as gpt-5.3-codex-spark

Existing issue threads show these refs can be visible in OpenClaw catalogs or tests, yet fail at runtime through ChatGPT/Codex OAuth with backend errors such as:

The 'gpt-5.5-pro' model is not supported when using Codex with a ChatGPT account.

Spark is now explicitly suppressed as no longer exposed by OpenAI/Codex catalogs.

What I found

Related issues and PRs:

• #79158: gpt-5.4-pro and gpt-5.5-pro rejected by Codex ChatGPT OAuth; closed after docs clarification that availability depends on bundled Codex harness version and selected account.
• #79362: duplicate user report from a ChatGPT Pro ($200/mo) subscription; same backend rejection for gpt-5.5-pro and gpt-5.4-pro.
• #81151: notes gpt-5.5-pro in fallback chains can still be rejected for ChatGPT accounts.
• #79582: documents Codex harness model availability and says users should inspect the live harness/account catalog.
• #83519: fixes provider-owner loading so OpenClaw can reach the Codex backend for openai-codex/gpt-5.5-pro, but its real-behavior proof still stops at the expected account/model support rejection.
• Spark-related historical issues: #15602, #31519, #44924, #46734.

The v2026.5.18 Codex harness docs list the current bundled app-server model snapshot as:

• gpt-5.5
• gpt-5.4
• gpt-5.4-mini
• gpt-5.3-codex
• gpt-5.2
• hidden codex-auto-review

They do not list gpt-5.5-pro.

Problem

For Pro subscribers, the current state is confusing:

1. Some OpenClaw catalog/config surfaces still include gpt-5.5-pro / gpt-5.4-pro.
2. Runtime can route far enough to the Codex backend, then the backend rejects the model for ChatGPT-account OAuth.
3. The docs correctly say availability depends on the bundled Codex harness and selected account, but they do not answer whether there is any supported OpenClaw path to use ChatGPT Pro-only models.
4. It is unclear whether the Codex CLI can access these models directly when OpenClaw cannot, and if so whether OpenClaw should expose a codex cli fallback/adapter path distinct from the current app-server/Responses route.
5. Spark/Cerebras-style models have historical references and aliases, but no current operator guidance beyond suppression.

Requested discussion / feature

Please clarify or implement one of these paths:

Option A: Document as unsupported

If gpt-5.5-pro, gpt-5.4-pro, and Spark are not available through OpenClaw's Codex OAuth/app-server runtime, document this directly in the OpenAI/Codex provider docs and model picker guidance.

Suggested behavior:

• Hide or mark these refs as "listed but not runnable with ChatGPT/Codex OAuth" for affected accounts.
• Ensure model pickers/fallback chains do not select them by default.
• Surface a clear diagnostic: "Your current Codex harness/account catalog does not expose this model; use /codex models or openclaw models status --probe."

Option B: Add account-aware live probing

If some ChatGPT Pro / Business / Enterprise accounts can use gpt-5.5-pro, OpenClaw should detect that per account rather than globally exposing or suppressing the model.

Suggested behavior:

• Add a Codex app-server viability probe for a specific model/profile.
• Cache/account-scope the result.
• Use the result in model picker, fallback selection, and status output.

Option C: Support a Codex CLI bridge

If the standalone Codex CLI can access Pro-only models that the OpenClaw app-server/Codex Responses path cannot, add a supported runtime route or adapter for CLI-backed model calls.

Questions to answer:

• Can OpenClaw safely invoke codex --model gpt-5.5-pro or equivalent for agent turns?
• Can it preserve tools, streaming, workspace context, and session state well enough to be usable?
• Should this be a separate runtime from the current Codex app-server harness?

Acceptance criteria

• Docs clearly state whether gpt-5.5-pro, gpt-5.4-pro, and Spark are supported through OpenClaw for ChatGPT Pro users.
• openclaw models list, /codex models, model picker, and fallback behavior no longer imply a model is runnable when the current account/harness cannot run it.
• If a supported path exists, provide exact config/CLI steps and a probe command to verify it.
• If no supported path exists, keep these models out of default fallbacks and provide a concise error that points users to openai/gpt-5.5 or a documented alternative.

Environment context

This was observed while investigating OpenClaw 2026.5.12 through 2026.5.18 behavior and current GitHub issue history. The issue is not tied to one local account; it is about the public operator path for Pro-plan model availability and Codex runtime selection.