Codex catalog refresh verification

Confirmed cause (2026-09-05)

Two read-only GETs to https://chatgpt.com/backend-api/codex/models used the same existing, unexpired Codex subscription session and identical auth headers. Only client_version changed:

Version	Status	Visible GPT-6 Astra
0.147.0 (previous plugin default)	200	absent
0.153.4 (npm @openai/codex stable at verification time)	200	gpt-6-astra, visibility list

This confirms version-gated catalog visibility for the tested account, not universal account access or successful model inference. No generation request, token refresh, credential write or user cache deletion was needed for this comparison. Official model documentation also describes a staged rollout; never synthesize access by adding a static row.

Changes

• By default the plugin automatically reads the stable version from public https://registry.npmjs.org/@openai%2fcodex/latest metadata. No subscription headers are sent and no code is downloaded or executed. Redirects are refused. Successful results are cached in memory for six hours, failures for five minutes. A 1.5-second deadline bounds lookup and response parsing; concurrent accounts share the request. Invalid/prerelease/regressed metadata is ignored. Failures retain the last known version or the verified 0.153.4 fallback.
• Optional plugin config codexClientVersion overrides automatic lookup, with validation; the version is sent only to the catalog endpoint. A manual model refresh also expires the version lookup cache. No CLI installation is needed.
• Settings → Subscriptions → Default reasoning effort now has Refresh model lists (all subscriptions). Unlike a normal read, it invalidates all account catalog caches, invalidates pool membership, and re-announces routes so the conversation model picker also re-queries. It does not clear usage, health, credentials or Copilot reasoning replay.
• Ordinary reads retain the five-minute discovery cache. Existing non-empty models.<provider> configurations still intentionally disable discovery.

Tests

Offline regressions exercise the version-gated GPT-6 response, explicit version overrides, malformed-version rejection before fetch, all-account cache refresh within the TTL, and real plugin RPC wiring (force validation, adapter cache invalidation and picker re-announcement). They do not claim live model execution or browser interaction coverage.

Automatic-version tests cover lookup coalescing, cache expiry, manual refresh, failure retry, last-good fallback, malformed/prerelease/regressed metadata, credential-free headers, stalled transports, late responses, and explicit configuration taking precedence over automatic lookup.