tinyoauth 0.1.0

First CRAN release. Consolidates the 0.0.1.x development cycle: a minimal OAuth 2.0 client (client-credentials and authorization-code grants with token refresh), an oauth_import_httr() bridge for legacy .httr-oauth caches, a no-listener manual mode that auto-engages on remote/headless sessions, and an OpenAI Codex device-login helper. Contributions to the Codex route from Sounkou Mahamane Toure.

tinyoauth 0.0.1.6

oauth_token_authcode() now defaults manual = NA, which auto-detects a remote/headless session (SSH, RStudio Server, or unix with no display) and uses the manual paste flow instead of hanging on a loopback listener the redirect can never reach. A first-time login over SSH (e.g. through a pipeline that calls tinytuber::yt_oauth()) now shows the paste prompt rather than blocking. Pass manual = TRUE/FALSE to force either mode.

oauth_token_authcode() now prints “Authorization complete.” after a successful exchange (the manual mode previously finished silently).
Clearer manual = TRUE instructions: the copy-the-address-bar step now includes the Ctrl+L / Ctrl+C (Cmd on macOS) shortcut, and a short reminder lives in the input prompt itself – the line still on screen when you return from the browser.

oauth_token_authcode(manual = TRUE) adds a no-listener mode for remote/headless use: it prints step-by-step instructions, you approve in a browser anywhere, then paste the redirected 127.0.0.1 address (or the bare code) back at the prompt. Uses the same registered loopback redirect_uri, so it is not the deprecated out-of-band flow. Flows through wrappers that pass ... to it (e.g. tinytuber::yt_oauth(..., manual = TRUE)).

Fixed a crash in oauth_token_openai_codex() device login when OpenAI’s poll endpoint returns its error as a nested object ({code, message}) rather than a string — the classifier compared a length>1 value and errored on the first poll. The error field is now reduced to a single string before matching.

oauth_token_openai_codex() device polling now keeps polling on HTTP 403/404 (OpenAI’s “still pending” responses) instead of erroring, while still stopping immediately on hard denials. It also finalizes the token on every path, so a valid cached token missing account_id has it derived from the access-token JWT; with login = FALSE, a token whose account id can’t be determined returns NULL.

Added an OpenAI Codex (ChatGPT) device-login route: oauth_token_openai_codex(), openai_codex_client(), and openai_codex_account_id(). Tokens are cached and refreshed through the existing oauth_cache_path() / oauth_refresh() machinery. Adapted from Sounkou Mahamane Toure’s llm.api PR #20, which proved the device flow; tinyoauth adds the token persistence and refresh.
Added oauth_jwt_payload(), a generic (signature-free) decode of a JWT payload for reading claims from tokens you already trust.

Initial release: a minimal OAuth 2.0 client (client-credentials and authorization-code grants with token refresh) built on curl and jsonlite.