# Security Policy
`rclean` deletes user files. The trust model is the product, so we treat
security reports as first-class issues.
## Supported Versions
While the project is pre-1.0, only the latest `0.1.x` patch release
receives security fixes. Older `0.1.x` releases are not patched —
upgrade to the latest patch.
| 0.1.latest | :white_check_mark: |
| 0.1.x (older) | :x: |
| anything < 0.1 | :x: |
## Reporting a Vulnerability
**Please do not file a public GitHub issue for security vulnerabilities.**
Preferred channel: GitHub's [private vulnerability reporting](https://github.com/majiayu000/rclean/security/advisories/new).
This lets us coordinate a fix before details become public.
> **Maintainer setup note (one-time, takes ~30 seconds).** Private
> vulnerability reporting is opt-in per repository. If the link
> above returns 404 or "page not found", the feature has not been
> enabled yet. Maintainer action:
> *Repository → Settings → Code security → "Private vulnerability
> reporting" → Enable*.
### Fallback channels
If the private-reporting form is not yet available, in priority order:
1. Open a [GitHub Discussion](https://github.com/majiayu000/rclean/discussions)
with the title prefix `[SECURITY — DO NOT REPRODUCE PUBLICLY]`.
This is visible but signals intent. The maintainer will move it
private on first response.
2. Reach the maintainer through the contact on the
[@majiayu000 GitHub profile](https://github.com/majiayu000).
Filing a regular public GitHub issue is the **last** resort — only
do this if (1) and (2) have both been tried and failed, and even
then, **redact the reproduction steps** and request a private
channel in the issue body.
### What to include
A reproducible report is more valuable than a polished one. Useful
information:
- The version (`rclean --version`) and platform (Linux / macOS / Windows)
- A minimal directory layout that triggers the issue (`tree -L 2` is
fine; no need for real data)
- The command line invoked and the observed vs expected outcome
- Whether the issue requires a `.rcleanignore` / `.rclean.toml` /
ActionPlan file to trigger, and the contents of that file
### Response time
Best effort, pre-1.0:
- Acknowledge receipt within 7 days
- Triage and confirm or close within 14 days
- Coordinated public disclosure once a fix lands on `main`
If you don't hear back within those windows, ping the issue you filed
or reach out via the contact above.
## In Scope
Things the trust model promises and that we treat as security issues:
- **Symlink bypass** — a candidate that resolves to a symlink target
outside the scan root being deleted. (See closed issues #2, #3, #6.)
- **Broad-scope guard bypass** — `clean` operating against `/`, `$HOME`,
`/etc`, `/usr`, system roots, or `C:\` without `--allow-broad-root`.
- **ActionPlan tampering** — a JSON plan that promotes a blocked path
to `safe`, or names a path outside the original scan roots, being
executed without revalidation.
- **TOCTOU window** — the scanned candidate being swapped (e.g. to a
symlink to a system path) between scan and delete.
- **`.rcleanignore` / `.rclean.toml` injection** — a project-local
config file weakening built-in safety classification (e.g. user
rules producing `blocked`-safety candidates, or marker files giving
a candidate `safe` status when it shouldn't).
- **Root-boundary escape** — canonicalization that ends up outside the
declared scan root being treated as in-scope.
- **Dirty-git-as-safe bypass** — a candidate inside a dirty git
worktree being auto-selected by `clean --all` without
`--include-caution`.
- **Permanent-delete on TOCTOU symlink** — `clean --permanent`
following a symlink target swap.
### TOCTOU protection boundary
`clean` performs a final pre-delete validation for every selected
candidate. That validation rejects symlinks, junctions/reparse points,
hardlinked files, non-directories, protected user-data paths, and
protected runtime/system paths immediately before deletion. On the
supported Unix-like and Windows platforms, Rust's `remove_dir_all`
also uses platform deletion primitives that do not follow symlinks
during recursive permanent deletion.
This protection is scoped to trusted local developer workspaces. It is
not a lock over a contested filesystem, and it does not defend against
an attacker who already controls the host, the user's account, or a
shared directory being modified concurrently with the clean operation.
## Out of Scope
Things that aren't security issues even though they involve data loss:
- **User-initiated deletion of files the user wanted to keep but didn't
commit / back up.** rclean prints a plan, requires `--yes` or
interactive confirmation for `--all`, and writes to OS Trash by
default. If a user passes `--permanent --yes` and loses uncommitted
work, that's a user-input issue.
- **Permission denied / disk full / read-only filesystem during clean.**
These are operational errors, not policy failures.
- **Issues that require physical access or local root on the host.**
rclean does not defend against an attacker who already controls the
filesystem it runs on.
- **`cargo install` integrity.** Cargo + crates.io supply-chain trust
is owned by upstream, not by rclean. We do run `cargo audit` weekly
on our own dependency tree — see [`Audit` workflow](.github/workflows/audit.yml).
## Past Security Work
The threat model is shaped by issues already filed and closed against
the project:
- **#2** ActionPlan trusts safety field from JSON — blocked-bypass risk
- **#3** No deletion-time guard on interactive `--all` path (TOCTOU)
- **#6** No broad-scope guard for clean against /, $HOME, /usr, /etc
- **#7** Non-atomic ActionPlan write — SIGINT corrupts file
- **#8** Walkdir errors silently swallowed — under-reports project size
These shaped the safe / caution / blocked tiering, the revalidation
step in `clean --plan`, the broad-root guard, and the atomic plan
writer. New reports should reference these to clarify whether they
extend an existing class or are something new.