Please do not open a public issue for security problems. Use GitHub's private reporting instead: Report a vulnerability.

Include what you found, how to reproduce it, and the version and OS. You will get an acknowledgement within a few days, and credit in the fix's release notes unless you prefer otherwise.

Only the latest release is supported with security fixes.

Knowing the design helps judge what is a vulnerability:

• Yardsort launches programs you configured — coding agents and your shell — in folders you chose, with your login shell's environment. An agent can do whatever its own permission model allows; that is the agent's security boundary, not Yardsort's.
• It stores no credentials, has no account and sends no telemetry. Agents use their own logins. Its only network request is the update check: a download of latest.json from this repository's GitHub releases, shortly after start and once a day (it can be switched off).
• Updates are signed. An update is installed only if its signature verifies against the public key built into the app, and only after the user asks for it.
• Programs are started with an argument list, never through a shell, so prompts and settings cannot inject commands.
• When browsing or diffing a workspace, the webview names the workspace by id and gives a relative path; the core resolves it and refuses anything that leaves the workspace folder.

Reports about breaking those guarantees — path traversal out of a workspace, command injection through a prompt or a harness field, a webview escape, anything that would let an unsigned or downgraded update install — are exactly what we want to hear about.