Keychain reads no longer fall back to the AlchementaryLink service and
copy entries forward, and Windows Credential Manager entries use
keyring's own target naming instead of go-keyring's 'service:user'.
Credentials saved by Alchementary Link, or by earlier builds on Windows,
are no longer found and need re-entering.
SshConfig no longer has proxyPassword or totpSecret fields, so neither
can sit in clear in nexus.db or travel in an export. The proxy password
is a SecretKind (profile:<id>:proxy-password) edited like other saved
credentials and purged with the profile; the session manager hands both
it and the TOTP seed to the SSH client from the keychain/vault at
connect time.
This drops compatibility with Go-era profiles and .link/.nexus files
that carry these values inline: they are ignored on read and must be
re-entered (SEC-20261012-011).
Session-only credentials (kept for a whole session so reconnects don't
re-prompt), credentials pending until a connection succeeds, and the
credential prompt's text buffer are now Zeroizing and wiped when
dropped, including when a prompt is cancelled or abandoned
(SEC-20261012-007, partial).
Still open: the values handed to the protocol clients are plain Strings
for the duration of authentication, egui keeps its own copies of typed
text, and no mlock/PR_SET_DUMPABLE is applied.
- A JSON export without secrets now blanks login-script 'send' steps that
answer a password-like 'wait' prompt (SSH and Telnet), not just the
proxy password and inline TOTP seed (SEC-20261012-011, partial).
- Credential kinds were string literals in the session manager, profile
editor and secret store; they are now SecretKind with one ALL list, so
the profile-delete purge can't miss a kind (F19).
Moving proxyPassword out of the profile record is left open: it changes
.link/.nexus compatibility with the Go app and needs an owner decision.
Every shared lock was std::sync::Mutex with lock().unwrap() (~150 sites),
so a panic in any session's reader or callback poisoned the lock and the
next UI frame panicked too, dropping every open session. Locks are now
parking_lot::Mutex (no poisoning; already in the dependency tree), and a
panic hook reports worker panics as an error toast instead of the session
silently stopping (F3, F4).
A password typed at a 'saved' prompt was stored before the server
accepted it, so a typo was persisted and silently retried on every
reconnect. Typed values are now held per attempt and written once the
session connects; when the server refuses a saved credential the next
attempt asks again and the new value replaces it on success. A store
failure (keychain refused, vault locked) now shows a toast instead of
being swallowed, and reconnect prompts wait 5 minutes like the first
connect (SEC-20261012-015, F1, F2, F20).
Under the 'ask' policy a changed host key used to be a hard error. The
prompt now shows the stored and received fingerprints and offers Reject,
Accept Once (connect, keep the record) and Accept & Update (replace the
record; audited). Unknown keys gain an Accept Once option too. Strict and
accept-new still refuse changed keys without prompting. Known-host rows
written by the old app are confirmed removable by id (test).
The NickServ password (and any other "remember for this session" secret) was
resolved through a fresh per-connect cache, so every auto-reconnect prompted
again with a 60 s timeout and an unanswered prompt failed the reconnect.
Session-only secrets now live on the session, survive auto and explicit
reconnects, and IDENTIFY is sent on every registration as before.