Skip to content

Security: VeraTools/revera

Security

SECURITY.md

Security

Reporting

Report vulnerabilities privately through GitHub security advisories rather than public issues. Expect an acknowledgement within a week.

Model of trust

  • Configuration (revera.yaml) is trusted: in event mode, an in-checkout config is read from the PR base commit, never the PR head, and endpoint and credential-name trust checks run before any credential value is resolved. Configuration chooses providers, models and the names of credential environment variables. Credential values are read from the environment and are never logged, written to reports or included in review fingerprints.
  • Pull request content is untrusted. Models see it, but can only request read-only tools scoped to the checked-out head tree; they have no shell or write tools and no direct arbitrary network access. Revera makes configured provider, optional Vera and opt-in external web-search (web_search, off by default; queries go to Exa and require public event metadata or explicit allow_private: true) requests, accepts structured findings, and decides what is published.
  • Opt-in GitHub App fixes (docs/server.md#fixes-opt-in, off by default) edit a disposable head snapshot inside a separate Linux bubblewrap sandbox without a GitHub token; configured test commands run in their own sandbox without network or credentials, and Rust checks every patch before human-approved or explicitly configured draft-PR publication.
  • In the Action, the GitHub token needs contents: read and pull-requests: write to read pull request data and publish comments, and is used for nothing else; fork PRs are skipped by default because secrets are not available to them. In the self-hosted GitHub App, jobs receive short-lived per-repository installation tokens scoped to each job's required permissions.
  • Release binaries are published with SHA-256 checksums; scripts/install-revera.sh verifies them before install.

There aren't any published security advisories