[codex] fix model cache invalidation on config reload

[lost9999]

Summary

Fix /api/models returning stale model metadata after config.yaml changes outside the WebUI default-model endpoint.

Root Cause

get_available_models() detects a config.yaml mtime change and calls reload_config(), but the first mtime check happens before the fast-path cache logic. After reload_config() updates _cfg_mtime, the later cache invalidation path no longer sees a config change, so the 24h in-memory _available_models_cache can keep serving old default_model and model groups.

Changes

• Invalidate the available-models cache immediately after the early config mtime reload.
• Replace the existing mtime cache test with a real temporary config.yaml edit so it catches stale in-memory cache reuse.

Validation

• HERMES_HOME=/home/lost9999/.hermes HERMES_WEBUI_STATE_DIR=/home/lost9999/.hermes/webui /home/lost9999/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_ttl_cache.py -q

Refs #1240.

[Michaelyklam]

Reconciled this against current origin/master while working the #1240 model/provider cleanup slice.

Findings:

• The cache-invalidation part here is already superseded on origin/master: api/config.py now stamps /api/models disk cache entries with _source_fingerprint, _schema_version, and _webui_version, and rejects both memory and disk caches when the current config.yaml or auth.json fingerprint changes. See api/config.py::_models_cache_source_fingerprint(), _is_loadable_disk_cache(), _get_fresh_memory_models_cache(), and _save_models_cache_to_disk().
• Current regression coverage exists in tests/test_issue1699_model_cache_source_fingerprint.py for auth/config source drift and tests/test_issue1633_models_cache_version_stamp.py for disk cache version/schema invalidation.
• I found one still-useful behavior in this PR that was not fully present on origin/master: routing model IDs listed only under custom_providers[].models back to the named custom provider. I extracted that as a fresh Michael-owned narrow PR: fix: route custom provider models dict selections #1752.

Verification from the fresh branch:

• Targeted model/cache suites: 51 passed in 48.98s
• Full isolated suite: 4597 passed, 2 skipped, 3 xpassed, 1 warning, 8 subtests passed in 378.87s

Recommendation: close this conflicting PR in favor of the already-merged cache work plus #1752 for the remaining custom-provider routing slice.

[nesquena-hermes]

Closing — superseded on master

Thanks @lost9999 — closing per @Michaelyklam's reconciliation comment from May 06.

The cache-invalidation slice from this PR is already on master: api/config.py now stamps /api/models disk cache entries with _source_fingerprint, _schema_version, and a TTL, and the on-config-reload invalidation path runs through the centralized invalidate_models_cache() helper that set_provider_key and remove_provider_key both call. That landed across the v0.50.282 / v0.50.288 / v0.51.x release cycle.

The branch here was draft + DIRTY for several weeks while the master-side cleanup absorbed the same fix shape. Rather than ask you to rebase a draft against the now-superseded code, closing this out so the queue stays clean. Appreciate getting it on our radar early — the canonical fix on master is structurally similar to your approach.

If you spot a residual cache-invalidation gap that the master-side invalidate_models_cache() plumbing doesn't cover, please open a fresh issue with a repro — happy to triage.