Trust
Security model
Who can read what, how code is kept apart, and what runs where.
Who can read what
Everything belongs to an organization, and only its members see it. Within one, each person has a role:
| Role | Can |
|---|---|
| Viewer | Read the repositories, search, trace and ask. |
| Member | Also start fixes and changes. |
| Admin | Also connect repositories, invite people, and manage integrations, API keys and usage. |
| Owner | Also make owners, and delete the organization. |
Approving a fix or change takes write access to the repository on GitHub, checked at the moment of approving, never remembered. A high-risk plan needs two approvers.
Keeping organizations apart
Each organization's rows are kept apart by the database itself (Postgres row-level security), not only by the application: a query made for one organization can't read another's.
Reading your code
Private repositories are read through the repo-sensei GitHub App, on the repositories its installation was given; it reads contents and metadata. Signing in with GitHub gives no access to repositories.
To index and answer, parts of the code are sent to the AI providers repo-sensei uses: Anthropic (answers, summaries, plans and patches) and Voyage AI (embeddings and reranking).
Running code
Fixes and changes run your tests in a sandbox: a container per run with no internet access (at most a private network to a database container the tests need). Only installing dependencies reaches the package registries, through a proxy that allows nothing else. A patch counts only if the tests pass by their own report, and patches that edit tests, skip them or fake their results are refused.
Signing in, keys and tokens
- Browser sessions are cookies the page's scripts can't read, with protection against cross-site requests. Settings → Sessions lists yours and signs any of them out.
- API keys belong to an organization, carry only the scopes chosen for them (
read:repo,ask), are kept only as a hash, and are shown once. - Editors and the command line sign in with OAuth (approved on a page here) and get tokens for that purpose only, which you can revoke.
- Integration secrets (GitHub, Sentry and Slack credentials) are encrypted at rest.
Spending
Agent runs stop at the organization's daily spending cap, and summaries at its monthly budget. Admins see the spend on Settings → Usage.
The audit log
Changes to members, roles, repositories, keys, integrations and fixes are recorded, with who made them and when, in Settings → Security. The log is append-only.
Deleting
- An organization an owner deletes is hidden at once and can be restored by its owners for 7 days; then it is purged with everything in it.
- An account is deleted from Settings → Profile: its sessions end, its GitHub tokens are deleted, and its name, login, email and avatar are removed. It can't be deleted while it is the only owner of an organization.
Public pages
Repositories on Explore are public: their pages and short answers about them need no account. Each question there passes a check that a person is asking, and has per-address limits.
Reporting a problem
Found a security issue? Write to security@reposensei.com, with what you saw and how to see it again. Please don't post it publicly before it is fixed.