Bug

Hermes can show gpt-5.4 with a 32k context window in Codex-backed setups.

Repro

• Use Hermes with gpt-5.4 on the Codex endpoint
• Hermes shows 32k context

Expected

gpt-5.4 should resolve to 1,050,000 context tokens.

Actual

Hermes can persist a bad cache entry like:
gpt-5.4@https://chatgpt.com/backend-api/codex: 32000

Likely cause

A small output-token-related value is being persisted or reused as the model context window.

Reference

OpenAI currently lists gpt-5.4 at 1,050,000 context and 128,000 max output tokens:
https://developers.openai.com/api/docs/models/gpt-5.4

Fix

I have a patch prepared that:

• maps openai-codex to provider-aware OpenAI metadata
• ignores suspicious small cached context values for large-context OpenAI models
• adds regression tests for the stale Codex cache case