Files
DodoSSH/scripts
laurensandClaude Opus 5 597e5ae016
ci / build and test (push) Successful in 3m17s
ci / android head (push) Successful in 3m45s
ci / api image (push) Successful in 24s
ci / desktop nightly (push) Successful in 1m19s
Take the release credentials off the command line, where anyone could read them
Three scripts prompted for a secret with Read-Host -AsSecureString or read -s,
said in a comment that it therefore never lands in shell history, and then wrote
it straight into the arguments of the next process. The comment was true and was
answering the wrong question. A command line is not private: on Windows one call
to Get-CimInstance Win32_Process hands every argument of every process running as
this user to anything that asks, /proc/<pid>/cmdline is world-readable on Linux,
and process-creation auditing and EDR agents keep their own copy either way. The
values were the Android release signing key's passwords and the forge token —
between them the whole capability to ship an update that every installed copy
will fetch and run, which ADR 0011 rule 1 and ADR 0013 go to some length to keep
off a runner and which was then left in a process listing on a desk.

Both replacements are documented features of the tools rather than tricks.

The Android passwords use .NET for Android's env: prefix, which puts the name of
a variable on the command line and lets MSBuild read the value out of the
environment itself. Set immediately before the build and cleared in a finally, so
a failed signing does not leave them in the shell for whatever runs next, and
-nodeReuse:false so no persistent MSBuild worker outlives the call still holding
the inherited copy in its own environment.

That prefix is unsupported when AndroidPackageFormat is aab, and the failure is
silent — the package would be signed with the literal string
"env:DODOSSH_ANDROID_STOREPASS" as its password. So the format is read back and
the build stops if it is ever changed, in the same shape as the versionCode guard
a few lines above: a wrong release refused here rather than discovered later on
somebody's phone.

vpk documents VPK_TOKEN as --token's environment fallback, so the flag is simply
gone from both desktop scripts and the token never becomes an argument at all.
Unset afterwards on the failure path too, which on macOS needs the exit status
captured rather than a bare || stop_with.

release-macos.sh had the same fault and was missed when the other two were first
reported. Its Apple credentials were already right — a notarytool keychain profile
rather than --password — and only the forge token was exposed.

What this does not claim: an environment block is readable by a process holding
rights over this one, the same as its memory is. The comments say so rather than
implying the values are now safe. What is bought is real and narrower — off the
command line, out of every process listing, out of the audit trail, and out of any
log that echoes the invocation.

Both PowerShell files parse, all three shell scripts pass bash -n, the
AndroidPackageFormat probe evaluates and answers empty so the guard passes, and
dotnet build accepts -nodeReuse:false. The signing itself is not exercised here:
it needs the release keystore, which by design is not on this machine.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-01 14:31:02 +02:00
..