Ship lighter isolated runtimes, richer Books and Mastery flows,
Reading/session polish, and explicit model API capabilities.
Release notes: assets/releases/ver1-6-4.md
Replaces the oauth_cli_kit indirection with DeepTutor's own PKCE loopback
sign-in, an account-scoped live model catalog, and a Settings card that exposes
only the OAuth surface. Credentials live in `<user-root>/private/openai-codex/`
with owner-only permissions, atomic writes, and a generation guard that stops a
late refresh from resurrecting a signed-out session; `~/.codex` is never read or
modified. A CLI path shares the same service.
A Codex token authorizes one person's ChatGPT plan, so credentials anchor to the
signing-in account's own root and the managed profile is marked `owner_bound`:
`redacted_model_access` drops it for every caller, and it is not offered as an
assignable model. Signing in publishes the profile but only activates it when no
LLM is configured yet, so it never repoints a deployment that already has one.
Also from review: reschedule the status poll so one dropped response cannot
strand the card on "waiting"; defer the catalog reload when the operator has
unsaved edits instead of discarding them; log the real cause behind the generic
provider error; stop promising a retry when the token refresh actually failed;
bind both loopback stacks since `localhost` may resolve to either; and drop the
provider metadata fields nothing consumed.
Bingxi Zhao (Frank)