Report vulnerabilities privately to alex@deblasis.net (or open a GitHub security advisory on this repository). Do not open a public issue for something exploitable.
You will get an acknowledgement within 5 days, and a fix or a stated assessment as soon as practical afterwards. This is a one-maintainer project; that is the honest SLA.
incoda is a local developer tool. It runs commands you already had the
authority to run, as your own user, and it coordinates them through files in a
per-user state directory on a local filesystem.
The security boundary is the OS user. What is in scope:
- A queue key escaping the state directory (key validation refuses path
separators,
.., and Windows device names rather than sanitising them). - Correctness failures that make two heavy jobs run when the tool said they
would not: the locking probe in
incoda doctorexists to catch filesystems that do not enforce locks. - Anything that would make a less privileged process able to influence a more privileged one through the state directory.
- The answer channel typing at a job it should not reach:
incoda answerwrites into a pty job's console input, so it is double-gated — the owner gate attach shares (lane.MayAttach, checked by the command and again by the run against the live ticket) and a blocked gate (the run delivers only while the job's own OSC 7501 record saysblockedwith a kind). The typed text never echoes into the job's log, and a refusal carries its reason back through the same request file.
Out of scope: denial of service by a local user against their own machine
(they have kill), and multi-user isolation on a shared host. State is
per-user by design; two OS users on one machine get separate lanes and are
expected not to trust each other.