Hi, I'm running OpenClaw 2026.2.24 on a Vultr VPS with Telegram as my main channel.

My auth is ChatGPT Pro via OAuth (openai-codex provider).

The problem: GPT-4o works perfectly when using openai-completions API, but every GPT-5+ model fails immediately with:

Failed to extract accountId from token

Models tested:

• gpt-4o → ✅ Works (via openai-completions)
• gpt-5 → ❌ Failed to extract accountId from token
• gpt-5-mini → ❌ Same error
• gpt-5.2 → ❌ Same error
• gpt-4.1 → ❌ Same error

What I've tried:

• Updated OpenClaw from 2026.2.23 to 2026.2.24
• Re-authenticated via openclaw agents add main (OAuth flow)
• Changed API endpoint from openai-responses to openai-completions in models.json (this is what fixed GPT-4o)
• Added GPT-5+ models explicitly in models.json with "api": "openai-completions"
• Cleared cooldowns and error counts in auth-profiles.json
• Verified the model definitions are correct in both openclaw.json and models.json

Background: The original issue was that gpt-5.3-codex started failing with HTTP 401: Missing scopes: api.responses.write. My OAuth token only has scopes: openid, profile, email, offline_access — it's missing api.responses.write which the Responses API now requires.

Switching to openai-completions fixed GPT-4o, but OpenClaw seems to internally force a different code path for GPT-5+ models that still fails.

Questions:

• Is there a way to force GPT-5+ models to use the Chat Completions API path with OAuth?
• Is there a plan to update the OAuth flow to request the api.responses.write scope?
• Any workaround to use GPT-5.2 or GPT-4.1 with a ChatGPT Pro subscription?

Relevant config:

• Provider: openai-codex with OAuth
• API set to: openai-completions
• Auth profile type: oauth
• OpenClaw version: 2026.2.24 (df9a474)
• OS: Ubuntu on Vultr VPS