Fix gpt-5.4 context length resolution for Codex

[zaycruz]

Summary

• map openai-codex and inferred OpenAI endpoints to the models.dev openai registry
• ignore suspicious small cached context values for OpenAI/Codex models when provider-aware metadata says the true context is much larger
• add regression tests for the stale gpt-5.4@https://chatgpt.com/backend-api/codex = 32000 cache case

Problem

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

I reproduced a bad persisted cache entry locally:
gpt-5.4@https://chatgpt.com/backend-api/codex: 32000

OpenAI's current model docs list gpt-5.4 at 1,050,000 context and 128,000 max output tokens, so the cached 32000 value is incorrect.

Testing

• pytest -q -o addopts='' tests/agent/test_models_dev.py tests/agent/test_model_metadata.py

[alt-glitch]

Fixes #5173 — stale gpt-5.4 context cache entry. Related to merged #15661 and #14935 (similar Codex context resolution bugs).

[alt-glitch]

Fixes #5173 — stale gpt-5.4 context cache entry. Related to merged #15661 and #14935 (similar Codex context resolution bugs).

[teknium1]

Thanks for identifying the stale 32K cache path; that specific gap is still present on current main.

Problems

• Current main already established that ChatGPT Codex OAuth must not use the direct OpenAI models.dev value: agent/model_metadata.py:1834-1852 sets gpt-5.4 to 272K, and agent/model_metadata.py:1912-1942 resolves from Codex /models before that fallback. The PR's new models.dev recovery would instead return 1.05M for this Codex route.
• The actual stale-32K bug remains: agent/model_metadata.py:2110-2118 only invalidates Codex cache values >= 400_000; the non-matching 32K entry reaches return cached at agent/model_metadata.py:2173-2178.

Suggested changes

• Adapt the stale-32K cache guard to current get_model_context_length() and fall through to the existing Codex resolver rather than models.dev.
• Add a regression test for a persisted gpt-5.4 32K Codex cache entry resolving to the existing 272K Codex value.

Automated hermes-sweeper review.

[harjothkhara]

Heads up: this looks like it landed on main via #68554 (commit 64702f8) — both prevent a stale Codex OAuth context cache from short-circuiting provider-aware resolution; main’s agent/model_metadata.py::_skip_persistent_context_cache() bypasses all openai-codex disk entries and re-resolves from the authenticated /models catalogue. Flagging in case you'd like to close or rebase.

[alt-glitch]

This was generated by AI during triage.

Related to merged #68554: current main now bypasses persistent openai-codex cache entries and resolves the authenticated Codex catalog, correcting the stale-cache failure. This PR still maps Codex to generic OpenAI models.dev metadata, so it is not the same mechanism and remains unsuitable without rebasing to the current resolver design.