matchblox runs commands on your machine on your behalf, so security reports come first.
Use GitHub private vulnerability reporting.
Do not open a public issue. Include the version (matchblox version), the OS,
how it ran (on the host, over SSH, remote mode), and the steps to reproduce.
We answer within 3 working days.
In scope:
- Anything that runs a step the person did not see, or a destructive step without its typed confirm.
- A guard that does not stop a step when its fact no longer holds.
- A shell string built from data, or text from a pane or a model that becomes a command.
- Secrets that reach a log or the console.
- The service socket, the hook command (
matchblox hook), the ssh forced command, and the install script. - A release asset that does not match its checksum or provenance.
matchblox is pre-1.0. Only the latest release receives fixes.
- matchblox runs as you, at the lowest CPU priority, and opens no port.
- Every step is an argv list, never a shell string, and is shown before it runs.
- A destructive step needs
xand then a typedy. Its guard (the same process and start time, a worktree still clean and unused, a session still idle) is checked again just before it runs. - Steps run with a time limit, no terminal prompts, and SSH in batch mode.
- The service listens on a Unix socket in a directory that only you can open. The console draws only what the service sends.
- Setup edits
~/.claude/settings.jsonand the tmux config only after it shows the exact change, and keeps a backup. - Remote mode uses your own SSH. A connected host gets a key that can start only the console's stream: no shell, no forwarding, a strict host-key check.
- matchblox sends no pane text to a model.