Security

QuickRun executes code from the repository you point it at, with your privileges, outside any sandbox. Everything below follows from that.

The confirmation gate

Every run shows the repository, the ref, the resolved commit and the exact list of commands — setup, tasks, stop, or the detected fallback — and waits for you to approve it. There is no setting that disables this, and a run triggered from the browser cannot skip it: the listener prepares the run and returns the plan, and a second, explicit confirmation starts it.

--yes on the command line skips the prompt, because you typed the command yourself. Nothing a web page can do reaches that flag.

Trusting a repository

Approving a repository stores a hash of those commands. If the config changes, the hash changes and you are asked again — so a repository cannot be trusted once and quietly turned into something else later. Trust is per repository, never per owner or per host, and revocable.

The hash deliberately ignores the repository, ref and commit: new commits to a trusted repository keep working, changed commands do not.

The listener

• Bound to 127.0.0.1 only, never to an address reachable from the network.
• Every endpoint that can start something requires the request to come from a browser extension, checked against the Origin header the browser attaches and a page cannot forge.
• /api/ping answers anyone on purpose: telling a page that QuickRun exists is its entire job. It reveals nothing else — no repository names, no paths, no run contents.
• https://github.com is deliberately not an accepted origin. A script running on GitHub itself cannot start a run.
• A caller with no Origin is not a browser — curl, or QuickRun's own CLI — and is allowed. Such a program already runs with your privileges and gains nothing by going through the daemon.
• There is no pairing token any more. It guarded against exactly what the origin check guards against, while costing every user a setup step.

What a link may carry

A quickrun:// link - what a README badge ends up following - is a string written by whoever wrote the page. So it is treated as one:

• Only repo, ref and pr survive. A command, a config, a token or a local path in the link is dropped without comment, because none of it was ever ours to accept.
• repo must be owner/name or an https:// URL. ssh://, file:// and git@host:owner/name are refused from a link - typing them on the CLI yourself is a choice, a link doing it is not.
• The link never starts anything. It opens QuickRun's own window at the repository, where the plan is prepared and the confirmation gate applies exactly as it does everywhere else.

Secrets

Values from password inputs are held in memory for the run, passed to child processes as environment variables, and never written to logs, run history or progress text. Access tokens are removed from every log line and error message before they are shown or stored.

Very short values are not redacted: blanket-replacing a one-character secret would mangle every log line it happened to appear in.

Auto-update

Updating is a code-execution channel and is treated as one:

• the asset is fetched only from the release's own github.com download URL over HTTPS
• its SHA-256 must match the SHA256SUMS published with the release; a mismatch aborts
• QuickRun only replaces its own binary when nothing else manages it — a package-manager install is told the upgrade command instead
• the update is applied on restart, never in the middle of a run
• --no-update disables checking entirely

What is not protected

• No sandbox. A trusted repository's commands run with your full privileges. Container isolation was considered and deferred: most repositories do not run in a container unmodified.
• Unsigned binaries. See install for what that means on each platform.
• The commands themselves. QuickRun shows you what will run. It cannot tell you whether it is safe. Only run repositories you would be willing to git clone && ./run.sh by hand.

Reporting a problem

Open an issue at github.com/fgilde/QuickRun. For anything you believe is exploitable, use GitHub's private vulnerability reporting on that repository rather than a public issue.