Summary

openclaw models list reports a significantly smaller context window for openai-codex/gpt-5.2 (~266k/272k) than expected (400k). This affects model selection / planning and appears to cap the usable context for the Codex OAuth provider.

What I expected

• openai-codex/gpt-5.2 should reflect a 400k context window (as documented for GPT-5.2).
• If the Codex OAuth backend truly has a smaller limit, the CLI should clearly distinguish that this is a provider/backend-specific cap (and ideally link to where the value comes from), rather than silently presenting it as the model’s context window.

What I see instead

• openclaw models list shows:

  • openai-codex/gpt-5.2 → 266k
  • openrouter/x-ai/grok-4.1-fast → 1953k
  • kimi-coding/k2p5 → 256k

• openclaw models list --json returns (excerpt):

  • openai-codex/gpt-5.2 → contextWindow: 272000
  • openrouter/x-ai/grok-4.1-fast → contextWindow: 2000000
  • kimi-coding/k2p5 → contextWindow: 262144

• Config override is applied and visible via:

  • openclaw config get models shows contextWindow: 400000 for openai-codex/gpt-5.2

Yet models list continues to report ~266k/272k for openai-codex/gpt-5.2.

Steps to reproduce

1. On a fresh install, authenticate the Codex provider:
   • openclaw models auth login --provider openai-codex
2. Set default model to openai-codex/gpt-5.2 (or add as fallback).
3. Run:
   • openclaw models list
   • openclaw models list --json
   • openclaw config get models

Environment

• OpenClaw: 2026.2.1
• OS: Ubuntu Server (headless)
• Models involved:
  • openai-codex/gpt-5.2
  • openrouter/x-ai/grok-4.1-fast
  • kimi-coding/k2p5

Notes / hypothesis

It looks like the model registry is applying a provider-specific context window cap for the Codex OAuth backend (272000), which overrides the configured contextWindow and is presented in models list. If this is intentional (Codex OAuth has a smaller max context), it would be helpful to:

• document the Codex OAuth context limit explicitly, and
• surface the source (provider cap vs configured cap) in models list / models status.

If it is not intentional, then the configured contextWindow should be respected (or at least models list should reflect the configured value when models.mode=replace).