Report privately through GitHub's security advisories rather than in a public issue.
Please include the version (package.json), your DSH and Node versions, and the
smallest reproduction you can manage. If the report involves a command being
executed, say so explicitly in the title — this plugin sits in the path of every
bash call, so that class of report is triaged first.
Response times. Acknowledgement within 7 days; an assessment (affected versions, whether a fix is warranted) within 30 days. These are targets for a volunteer-maintained project, not a contract.
In scope:
- A command that runs something other than what the model asked for, beyond the documented rewrite.
- Argument injection through a rewritten command.
- Escaping the harness's sandbox or approval stack by any means other than the documented fallback to the original command.
- Reading or writing files outside what the harness permits.
Out of scope:
- The behaviour of
rtkitself. Report those to https://github.com/rtk-ai/rtk. - The behaviour of the harness. Report those to the DSH maintainers.
- Anything that requires an already-compromised host, or a malicious
rtkExecutablethe user configured themselves. - Compaction altering text. That is the feature, it is bounded by the documented
settings, and
readCompactionis off by default so file reads stay exact.
dsh-rtk replaces exec.arguments for the duration of one dispatch. The harness
documents tools/execute wrappers as permitted to change only exec.signal, so
this is an unsupported seam taken knowingly — see "Caveat: the rewriting seam" in
README.md. It is disclosed rather than fixed because no supported
alternative exists. If a future harness closes it, rewriting stops working and
every command runs unrewritten; there is no path by which it silently runs the
wrong thing.