0042 — Guided Anthropic sign-in

Status: accepted.

Reuse the pinned native Claude CLI's auth login process through a private PTY owned by the runtime. Do not register this PTY with TerminalManager. The runtime extracts only an allowed HTTPS OAuth authorization URL; it discards its bounded transient output before accepting a code and never retains echoed input.

Existing account operations gain account.submit and account.verify. They carry the current attempt ID; submission additionally carries one bounded code. Status returns stage, expiry, safe failure category, and the pending URL to the initiating member only. All mutating account operations require the personal owner or a shared/Team/background administrator. The original authority is renewed during login and verification; replacement, expiry and revocation stop the old process. Sign-in counts as active maintenance work.

Successful login loads the native catalog and performs one tools-disabled CLI request against an empty workspace. No automatic retry occurs. Only explicit verification retry performs another request. Persist safe health metadata in the existing native metadata table, bound to the connection generation. No control-plane schema migration is needed. Authentication failures persist until new sign-in and verification succeed. Historical failed turns are projected into safe reconnect messages rather than replayed or rewritten.

Codes and authorization URLs are not audit data. Provider errors become stable categories; never render arbitrary provider error output. Keep OpenAI device login and its existing terminal capability separate from the Anthropic path.

Validate synthetic attempts, member/admin authorization, stale callbacks, credential generation changes, native protocol compatibility, and actual pinned CLI URL extraction on the target architecture. A live owner sign-in and reply remain required product acceptance; never substitute health checks for them.