# `saveDir` / `filename` → filesystem destination: the adversarial table

`veloxd` is the only process that turns an untrusted string into a place bytes get
written. `capture.offer` means that string can originate from a web page, and
`download.add` over the Unix socket is reachable by any same-UID process. CLAUDE.md §4
("paths are canonicalized and checked against allowed roots before any write") and the M1
DoD ("no path traversal in `saveDir` … → `-32011`") make this a security boundary, not a
formatting nicety.

This table is written **before** `fs/safepath.cpp`, the way EXT did for `shouldCapture`.
Every row is a test in `daemon/tests/safepath_test.cpp`.

Roots for the examples: `allowedRoots = ["/home/u/Downloads", "/data/dl"]`, already
`realpath`-resolved and stored canonical at load time. `$HOME = /home/u`.

| # | Input (`saveDir`, `filename`) | Attack | Required outcome |
|---|---|---|---|
| A1 | `/home/u/Downloads/../.ssh`, `authorized_keys` | `..` climbs out of the root | `-32011`, `data.path` = the input `saveDir`. No dir created. |
| A2 | `/home/u/Downloads/a/b/../../../etc`, `x` | `..` chain escaping after descending | `-32011`. |
| A3 | `/etc`, `cron.d-payload` | absolute path, simply outside every root | `-32011`. |
| A4 | `/home/u/Downloads-evil`, `x` | prefix-match confusion with `/home/u/Downloads` | `-32011` — containment is component-wise, not `starts_with`. |
| A5 | `/home/u/Downloads`, `../.bashrc` | `..` in the **leaf**, not the dir | leaf rejected → `-32011` (or `InvalidParams`); a leaf is one component, never a path. |
| A6 | `/home/u/Downloads`, `sub/dir/file` | `/` in the leaf | leaf rejected — `filename` names a file, not a subpath. |
| A7 | `/home/u/Downloads/link-out` where `link-out` → `/etc` (pre-existing symlink) | symlink component points outside a root | `realpath` resolves it to `/etc`; `-32011`. |
| A8 | `/home/u/Downloads/goodsub`, `iso.img` — but between our check and CORE's `open`, `goodsub` is swapped for a symlink to `/etc` | **TOCTOU** on a directory component | Defense: resolve + create with `openat`/`mkdirat` from an `O_NOFOLLOW|O_DIRECTORY` fd walk, then `realpath` the final dir **again** and re-assert containment. A component that is a symlink at walk time → `-32011`. |
| A9 | `/home/u/Downloads`, `file<NUL>.iso` (`0x00` in the leaf) | NUL truncation — the write path sees `file`, logs/UI see more; CORE's fuzzer hit exactly this via `Content-Disposition` | NUL and every `< 0x20` byte and `0x7F` stripped from the leaf before use (mirrors `core/src/net/content_disposition.cpp` `sanitize_leaf`). If the leaf is empty after stripping → reject. |
| A10 | `/home/u/Downloads`, `"\r\nSet-Cookie: x".iso` | CR/LF injection into logs / downstream | control bytes stripped as A9. |
| A11 | `/home/u/Downloads`, `.` / `..` / `` (empty) | degenerate leaf | rejected. |
| A12 | `/home/u/Downloads`, `con` / `aux` / `nul` | Windows device names | **allowed** on Linux — we are not Windows; do not over-reject. (Noted so a future "harden" pass doesn't add it thinking it was missed.) |
| A13 | `/home/u/Downloads`, `<260 chars>` | overlong leaf, `ENAMETOOLONG` at `open` | leaf capped at 255 **bytes of UTF-8**, never splitting a codepoint (docs/04 §2). |
| A14 | `/home/u/Downloads/<260 chars>/x`, `y` | overlong directory component | `mkdirat` / `realpath` returns `ENAMETOOLONG` → mapped `-32011`, not a crash. |
| A15 | `saveDir` empty / null | no destination given | caller substitutes `saveTo.defaultDir`; `resolve_target` itself rejects an empty dir rather than defaulting silently. |
| A16 | root `/home/u/Downloads` is itself a symlink to `/mnt/big/dl` | a symlinked root | `canonicalize_root` `realpath`s every configured root at load; the stored root is `/mnt/big/dl`, and a `saveDir` resolving there passes. A `saveDir` of the literal `/home/u/Downloads/x` also passes because it `realpath`s to `/mnt/big/dl/x`. |
| A17 | `/home/u/Downloads` exists as a **file**, not a directory | destination is not a directory | `-32011` (`not_a_dir`), no write attempt. |
| A18 | `/home/u/Downloads/新しい/フォルダ`, `映画.mkv` | non-ASCII, legitimate | **succeeds** — UTF-8 is fine; only control bytes and the structural checks apply. |
| A19 | `/home/u/Downloads/./sub/.`, `x` | redundant `.` segments, no escape | normalized away; **succeeds** at `/home/u/Downloads/sub`. |
| A20 | `/home/u/Downloads`, `  trailing-spaces.iso  ` / `dots...` | trailing space/dot (Windows-hostile, and confuses "same file" checks) | trimmed: leading/trailing whitespace and trailing dots removed before use. Empty after trim → reject. |
| A21 | relative `saveDir` (`Downloads/x`, `./x`, `x`) | a relative path has no well-defined base and invites cwd games | rejected — `resolve_target` requires an absolute `saveDir`. The GUI/CLI resolve against the default dir before calling. |

## Implementation (`fs/safepath.cpp`, as built)

1. **Sanitize the leaf first**, in isolation: strip `[0x00,0x20) ∪ {0x7F}`, trim
   whitespace, strip trailing dots and spaces, reject `.`/`..`/empty/`contains '/'`, cap
   255 UTF-8 bytes on a codepoint boundary. (A5, A6, A9–A13, A20)
2. **Require `saveDir` absolute; reject any `..` component lexically.** A legitimate
   client never sends `..`; a web-origin path with `..` is an attack, so it does not even
   reach `realpath`. (A1, A2, A21)
3. **If the directory already exists:** `realpath(saveDir)` — this follows every symlink,
   so a symlinked root or component resolves to where it *really* points — then assert the
   resolved path is inside a canonical root, component-wise (`d == root || d starts with
   root + "/"`). A symlink that escapes is caught here (A7); one that stays inside passes
   (A16). Open the resolved dir `O_PATH|O_DIRECTORY` for the leaf check. (A3, A4, A7, A16,
   A17, A19)
4. **If a tail is missing (`mkdir -p` case):** find the deepest existing ancestor,
   `realpath` + root-check *that*, then create the missing components through an
   `openat/mkdirat` walk with `O_NOFOLLOW|O_DIRECTORY` from the ancestor's fd — the tail
   has no symlinks because it had no entries; a race that plants one trips `ELOOP` →
   `-32011`. Then re-derive the final dir's path from its fd (`/proc/self/fd/N`) and
   re-assert containment. (A8 for the created tail, A14)
5. **Best-effort leaf check:** `fstatat(dir_fd, leaf, AT_SYMLINK_NOFOLLOW)` — refuse if it
   is already a symlink. This narrows, but does not close, the create-after-check race on
   the leaf: a symlink planted *after* this `fstatat` and *before* CORE opens the file is
   still followed. Closing it needs CORE to open with `O_NOFOLLOW` (plus `O_EXCL` on a
   fresh download). **Verified 2026-09-10: it does not yet** —
   `core/src/io/sparse_file.cpp:77` is `O_WRONLY | O_CREAT | O_CLOEXEC`. The flag change
   has been raised with CORE; until it lands this race is open, see the residual below.
6. **Every failure is `-32011`, `data.path` = the *original* `saveDir`** — never the
   resolved path, which would leak where the roots actually live. The one exception is a
   `filename` that violates the schema's own `maxLength`, which is `-32602` at the param
   layer before this code runs.

### Residual — currently OPEN, tracked

Two TOCTOU gaps this code does not close on its own:

1. **An existing intermediate directory** swapped for an out-of-root symlink between our
   `realpath` (step 3) and the write. Step 3 trusts `realpath` for the pre-existing
   prefix; a full `O_NOFOLLOW` chase would reject the legitimate symlinked directories
   A16 requires us to allow.
2. **The leaf** swapped for a symlink between our `fstatat` (step 5) and CORE's `open`.

Both are closed by CORE opening the file `O_NOFOLLOW` (and, for a fresh download,
`O_EXCL`). **As verified on 2026-09-10 that is not yet the case** —
`core/src/io/sparse_file.cpp:77` opens `O_WRONLY | O_CREAT | O_CLOEXEC`. The flag change has
been raised with CORE; when it lands, update step 5 and this paragraph and re-verify the
flags at that line.

What limits the exposure *today*: the download directory lives under `~/.local/share` /
`~/Downloads`, both `0700` — an attacker planting a symlink there already has write access
to the user's account. The unqualified "you are covered" version of this claim does not
hold until the CORE change lands; this is exactly the boundary where a reader stops
checking, so it is spelled out.

The post-M1 hardening for gap 1 is a per-step "resolve one component, re-validate the
running path against the roots" chase (systemd's `chase_symlinks` shape).