The last piece of the #1687 follow-up. BuildZipDownloadJob allows itself
an hour, every shipped topology runs exactly one worker, and everything
shares the default queue -- so one large archive delayed every
notification email queued behind it. The size cap and the
one-build-per-person rule bounded that in July; they did not remove it.
onQueue('zips') in the constructor rather than at the dispatch site, so a
second caller cannot forget it. Both images grow a worker for it:
compose.yaml gains worker-zips, supervisord gains [program:queue-zips],
and the existing worker in each narrows to --queue=default. --tries=1
there matches the job, which records its own failure rather than being
retried.
The part that needs care is the manual install. A worker whose command
still says plain `queue:work` consumes `default` only, so it would send
email happily and never finish a single zip, with nothing in any log
saying why. INSTALL.md's unit now reads --queue=default,zips -- one
worker watching both, which is right for most installations -- and says
what happens if you leave it off, with the two-worker split offered for
anyone who would rather keep the two kinds of work apart. CHANGELOG
carries it as an upgrade note, since it is something to do rather than
something that was done.
Verified in the dev stack rather than only in a test: dispatched a build
and watched worker-zips take it while the default worker stayed idle.
A full parallel run failed once and passed on retry while I was doing the
#1703 follow-up. A flake is worse than a steady failure: it trains you to
re-run rather than look, and it quietly weakens every green run reported
beside it.
Upload parts are real files under storage_path('app/uploads-tmp/{session_id}'),
not a faked disk. Every parallel worker gets its own database, so session
ids restart at 1 in each of them, and two workers writing parts land in
the same directory. On top of that ChunkedUploadsTest's afterEach deleted
the whole tree rather than its own share, for everybody. Six test files
write parts, so this was reachable without anything I added.
The same collision exists inside one worker: RefreshDatabase rolls back,
so ids restart at 1 for every test, and a run that died before its
cleanup leaves parts sitting under the id the next test is about to
claim.
LocalPartStore now reads its root from config, defaulting to exactly
where it always was -- an installation with UPLOAD_PARTS_PATH unset
behaves identically. Tests\TestCase points it at a per-worker directory
and empties that directory per test, which closes the cross-worker, the
cross-run and the intra-worker versions together. ChunkedUploadsTest's
cleanup and its two directory assertions read the configured root rather
than the hardcoded path, so they can no longer reach into a neighbour.
Verified with eight consecutive parallel runs, green, and by watching the
per-worker directories appear separately (w1, w2, w4 … w14) rather than
one shared tree. The isolation itself cannot be asserted from inside a
single test; what a test can pin is the mechanism it rests on, so one
does: parts go where the configured root says.
Follow-up to #1703, which made `exists:folders,id` mean what its ten
readers already assumed. Two things it named and deliberately left.
**The chunked upload is two requests.** store()'s rule only ever sees the
first: POST /uploads records the resolved folder on the UploadSession and
complete() reads it back from the session rather than from the caller, so
deleting the folder while the bytes are in flight still files the
assembled file into it -- the same orphan state #1703 removes, reached by
a door a validation rule cannot watch. complete() now re-resolves through
Folder::query() and files at the root when the folder has gone.
Root rather than a refusal, because the two moments cost different
things. At store() nothing has been sent, so refusing is free and honest,
which is the call #1703 made. Here the bytes are already uploaded, and
discarding somebody's finished transfer over a folder that vanished
underneath them is the harsher of the two surprises. The file lands
somewhere they can see it and move it.
**The refusal now explains itself.** "The selected folder id is invalid"
says nothing when the answer is that the folder has been deleted -- and
that is the usual way to meet this rule, since a live id picked from a
list is how anybody gets here. It matters most on the chunked path, the
one place #1703 makes a previously-working request fail. A small
ValidationRule object carries the message, which keeps the single
definition Rules::folderId() exists for: a messages() array would have to
be repeated at all ten call sites, and rules meaning different things in
ten places is what went wrong in the first place.
One note for whoever writes the next test here. Upload parts live in
storage_path('app/uploads-tmp/{session_id}'), which is a real shared
directory rather than a faked disk, and each parallel worker's database
restarts session ids at 1 -- so two files writing parts on two workers
collide, and ChunkedUploadsTest's afterEach deletes the whole tree for
everybody. Six test files write parts today. These two cases live in
ChunkedUploadsTest rather than beside the rest of their subject so this
change does not add a seventh racer; the underlying isolation problem
predates it and is worth its own fix.
#1698 moved the library boundary into FileCommentPolicy, where it
belongs, and said plainly what that cost: the moderation screen went
from 65 queries to 465 for a client-scoped moderator with five assigned
clients. Measured here, those numbers are exactly right.
The cost is not in asking. It is that StaffLibraryScope::files() rebuilds
its query every time, and building one runs four immediate lookups per
assigned client — the client's group ids, the same ids again inside
Folder::sharedFolderIds(), that method's own assignment lookup, and the
shared-folder get() in Folder::scopeVisibleToClient(). None of them
depend on the query being built. Gate resolves a fresh policy for every
check, so a listing paid for all of it once per row.
The built query is now memoised per user and handed back as a clone,
since every caller adds to it, and the scope is registered as `scoped`
rather than transient so the memo survives a request. Scoped rather than
a singleton on purpose: a long-lived queue worker keeps singletons
between jobs, and a library query built from one job's data has no
business answering the next one's question.
That is 465 queries down to 60 on the same page — below the 65 it cost
before #1698, because the memo also helps the callers that were already
asking repeatedly. FileVersions::sharedAudience(), which runs the same
helper twice per candidate while resolving notification recipients, gets
it for free.
So the answer to the question #1698 left open is neither of the two it
offered. can_delete stays a real question asked of the policy; nothing
restates the boundary; and the page is faster than it was before the
fix. Three tests: one user's query never answers another's, one caller's
constraints never follow the next, and the moderation screen does not
ask once per row.
Folder uses SoftDeletes. The `exists` rule runs against the table, so a
folder in the trash passes it -- while every resolution that follows goes
through Folder::query(), which honours the soft delete and finds nothing.
Ten rules across five controllers rely on that check, and each one reads
it as "this folder exists".
Two of them then wrote the id anyway. Api\FilesController::store()
resolves the folder, hands the null to Folder::uploadableBy(), is told
yes -- correctly, that is the rule for a root upload -- and passes
$validated['folder_id'] to the write. FilesController::store() is the
same shape once #1694 gives it the guard. FilesController::update() and
its API twin write it straight through with nothing in between.
The result is a live file inside a deleted folder, which is a state
nothing else in the application produces: FolderService::delete() deletes
every file in the subtree along with it. The row is reachable by id, in
search and over the API, and missing from the listing its uploader would
look in.
Rules::folderId() makes the check mean what its readers assume, once,
where the reasoning can be written down -- the same argument slug() makes
for itself one method above. Every site takes it, so the file cannot end
up with two spellings of the same rule and no way to tell which is the
safe one.
What changes, path by path:
- POST /files, POST /api/v1/files, PATCH /files/{file} and
PATCH /api/v1/files/{file} refuse a folder in the trash instead of
writing its id. This is the fix.
- POST /uploads used to accept it and quietly file the upload at the
root -- its guard and its write already agreed, on null. It now says
so instead, which is what the other upload paths do.
- files/{file}/move, files/bulk-edit, folders, folders/{folder}/move
and the portal's my-folders already refused, through
StaffLibraryScope::folders() or Folder::scopeVisibleToClient(), both
of which drop trashed rows. They still refuse; the answer is now 422
naming folder_id rather than a bare 404. Those two guards are asking
a different question -- "is this folder yours" -- and they keep
asking it.
No live folder id behaves differently anywhere, and the root (a null
folder_id) is untouched.
The published API document is unchanged: `exists` renders the same either
way. Regenerated with php artisan scramble:export and byte-identical.
Follow-up to #1687, which made a zip build report failure honestly. Four
things it passed near, none of them regressions it introduced.
A zip has never had a size limit — only a cap of 10,000 files, which
bounds nothing that costs anything. Ten thousand spreadsheets zip in
seconds; two hundred videos is an hour of stream-copying and an archive
that fills the disk. Bytes are what a build actually costs, so the new
Settings → Downloads screen caps the total size instead, at 2 GB out of
the box. It is a setting rather than a constant because the safe figure
depends on free disk, on whether sources live on a remote disk, and on
the plan a hosted tenant is on — the file count stays fixed, since it is
a foot-gun rail and not a knob anybody needs. The controller measures
the selection at request time and names both numbers when it refuses;
the job measures again, because it re-derives the selection at run time
and a folder can grow while the job waits in the queue.
Every shipped topology runs exactly one queue worker, and everything
shares the default queue, so raising the job timeout to an hour handed
any signed-in person an hour of everyone else's notification mail. There
is now one build in progress per requester and a named throttle bucket
on the endpoint, which had neither. A pending row older than an hour is
treated as abandoned rather than in progress, so a worker killed hard
enough to skip failed() cannot lock somebody out for good. Giving zip
builds their own queue is the structural fix and wants its own change:
it touches compose, supervisord and the systemd unit in INSTALL.md, and
an install that upgrades without changing its worker command would stop
building zips silently.
zip_downloads.requested_by cascades on delete, so removing a user takes
their rows with it and strands every archive they built — invisible to a
purge that walks rows, and to OrphanFileScanner, which skips zips/ on
purpose. The purge now also sweeps files in zips/ that no row explains,
after a day's grace so a build in progress is never taken out from under
itself.
Two smaller things while in here. A build that failed because every file
had already hit its download limit said only that nothing was available,
and dropped the skipped list — the same distinction the store guard goes
out of its way to draw at request time. And a failed close() now logs
libzip's reason, which the @ silencing had been discarding: "the disk is
full" and "the source vanished" are different problems for whoever has
to fix one, while the requester still sees a message with no server
paths in it.
BuildZipDownloadJob deferred every write to ZipArchive::close() but then
marked the row STATUS_READY regardless of the result:
- close() returns false when a source file was deleted between addFile()
and close() (a concurrent staff delete runs FileDiskCleanup at once) or
the disk filled up; the row went ready over an archive libzip never
wrote, and the download controller X-Accel-served a path that isn't there.
- An archive that ended up with no entries (every selected file removed or
its allowance spent before the queued job ran) is written as no file at
all by libzip, yet close() still returns true — again marked ready.
Check both the close() return and the added-entry count, and fail the row
(deleting any partial archive) when either says nothing was written.
The job also had no $tries/$timeout/failed(): a build of up to MAX_FILES
sources runs past the worker's default 60s timeout, and the kill skips the
catch, stranding the row as PENDING while the frontend polls forever. Give
it room, run it once, and add a failed() backstop that fails a row still
pending (leaving an already-resolved one alone).
Finally, purge leftover zips/{id}.zip* by row id: a killed build leaves a
partial archive and libzip temp file with no path recorded, so the path
field alone never cleaned them up.
The orphan scanner skips derived artifacts by path prefix, but the list
was a hard-coded ['thumbnails/', 'zips/'] that never learned about
'previews/'. ImageRendition::Preview caches under previews/ (and
previews/external/) on the local files disk, so every cached preview was
reported as an orphan: offered for import on the orphans screen, and
deleted by the purge command once past the grace period. An imported
preview also became a File row pointing at a path the rendition cache
owns -- destroyed the moment its source file was deleted or the cache
was flushed.
Derive the rendition prefixes from ImageRendition::cases() rather than
repeating them, so a future rendition can't be forgotten here the way
previews were; 'zips/' (the download-bundle job's) stays as it was.
Setting folder_id through update() is the same privileged reparent as
move() and bulkUpdate(), but only those two verified the target folder
was inside the caller's library (StaffLibraryScope::folders). update()
validated it only for existence, so a client-scoped staff member could
reparent an in-scope file into a folder shared with a client they are
not assigned to -- which File::scopeVisibleToClient then exposes to that
client, sidestepping the boundary the sharing endpoints enforce
(guardAssignable), and likewise into a public folder without
upload_public.
Apply the same scope->folders()->findOrFail() guard on both the web and
API update(), but only when folder_id actually changes, so re-saving a
file that already sits in an out-of-scope folder (reachable via a direct
client share) still works.
store() checks Setting::MaxFileSizeMb against the size the client declares
when it opens the session, and complete() re-checks the storage quota
against the real assembled byte count -- but nothing re-checked the size
limit itself. A client that declared a one-byte upload and then streamed
gigabytes of parts passed store()'s check and was never stopped, so the
configured limit (which store() applies to everyone, staff included) did
not hold for the resumable path that real uploads use.
Re-check the assembled byte count against MaxFileSizeMb in complete(),
cleaning up the assembled bytes and the session exactly as the quota
branch already does.
Two bugs a green suite could not find, both from pointing the
application at a real Google Cloud Storage bucket.
The adapter attaches a legacy per-object ACL to every write, and a
bucket with uniform bucket-level access — which our own setup
instructions require, and which Google recommends — refuses it:
"Cannot insert legacy ACL for an object when uniform bucket-level access
is enabled". So the default configuration could not write to the
recommended bucket. The library ships
UniformBucketLevelAccessVisibility for exactly this, and nothing is
lost by never setting an ACL: every object here is private and every
read is a signed URL.
The second is worse and was never about Google. Both file disks are
configured 'throw' => false, so a refused write returns false rather
than raising, and LocalPartStore ignored the return. The upload reported
success, the File row was written, and the bytes were nowhere — the
listing showed a file whose download could never work. An expired S3
credential did the same thing. It now checks, and the controller already
turns that into a validation error rather than a 500, so the person
uploading is told.
Verified against a live bucket with a key scoped to
roles/storage.objectAdmin: the probe lists, writes land, reads
round-trip byte for byte, and a signed URL comes back 200 carrying
"Informe año.pdf" intact through both the ASCII and RFC 8187 forms of
Content-Disposition.
Two routes still assumed every file sits on local disk, which stopped
being true the moment external storage was switched on. A share link
answered with X-Accel-Redirect whatever the file's disk said, pointing
nginx at a path it has nothing behind; a public listing built a
thumbnail from Storage::disk('files')->path(), which for an externally
stored file is a path nobody ever wrote. Both fail only for installs
using S3, and only on those two routes, so the same file downloading
correctly from the file manager made the share link look like the
broken thing rather than where the file lives.
Neither is a new rule. FileDownloadController and
FileThumbnailController already did it right, which is the actual
finding: the knowledge was sitting in a private method on one class and
inline in another, so the next caller could not inherit it and did not.
Both are now objects with one job.
StoredFileResponse replaces InlineFileResponse and grows an
attachment() alongside inline(), since the two differ only by
disposition. LocalSourceFile takes a closure rather than returning a
path: the version that returned one also left the caller to unlink it,
and both of those are exactly the mistakes made here.
The regression tests fail against the previous controllers — checked in
both directions rather than assumed.
A download is not served by PHP. PHP authorizes it and hands the web
server the path with X-Accel-Redirect, so the web server has to open a
file PHP wrote. Where those are different users — cPanel and Plesk
commonly arrange it that way — it cannot: uploads land 0600 inside a 0700
directory, and traversing 0700 means being its owner. Nothing else on the
site shows a symptom. Uploading works, the library lists everything, and
only downloads fail, as ERR_INVALID_RESPONSE in the browser and
`open() ... failed (13: Permission denied)` in the web server's log.
FILES_WEB_SERVER_READABLE writes uploads 0644/0755 instead. Opt-in and
spread into the disk configuration rather than switched by a ternary, so
an install that does not set it keeps byte-for-byte the configuration it
had: the relaxed modes are readable by every account on the machine,
which is the wrong trade wherever the web server and PHP are one user, as
in the image and on most self-administered servers.
The two halves are not enforced alike, which is the part worth knowing.
`visibility` has Flysystem chmod each file after writing it, so 0644
holds under any umask. A directory is created by mkdir(), which masks its
mode argument, so 0755 is a ceiling: a pool at umask 0077 still produces
0700 and still cannot be traversed. That cannot be fixed from config, so
INSTALL.md carries it — how to tell the two users apart, the one-time
chmod for files already on disk, and the pool setting for the umask.
FilePermissionsTest asserts all three modes, umask cases included, since
the asymmetry is invisible from the configuration.
Reported by @denkfabrik-li (#1668), who diagnosed it and verified the
remedy on the affected host.
The two things staff most often want to know about a file — who
downloaded it, who looked at it — were answerable only by reading the
whole activity log past everything else that had happened to it, or by
going back to the library list for the details panel.
The file's own page now has a Downloads & previews tab: the twenty most
recent times it was taken or looked at, each with who did it and the
address it went to, over a running count of both. Below them, two
buttons open the file's full history already filtered — one to every
download, one to every preview — so the narrow question is one click
and the whole log is still one click further.
Which filter value stands for "every download" is decided server-side
and travels with the payload, because it is a fact about the log's
vocabulary: downloads are three actions and share a group, previews are
one action and are filtered by name. The history page now also keeps
whatever filter it was sent with visible in its dropdown even at a
count of zero, so a button cannot land somebody on an empty table above
a select that has gone blank.
v1 could preview four kinds of file in a modal — images, video, audio and
PDF. v2 previewed only images, and not by decision: preview shipped as part
of the image *thumbnail* work (1c68aa1), so "previewable" quietly became a
synonym for "GD can decode it". FileThumbnailController::preview() gated on
ThumbnailGenerator::SUPPORTED_MIME_TYPES, the frontend mirrored the same
four types, and the dialog was a hardcoded <img>.
Rather than widen that list — it drives pathFor(), extensionFor(),
generate() and FileDiskCleanup, and a video reaching getimagesize() is a
500 — this separates the two questions. PreviewKind now answers "may these
bytes be served inline, and what element renders them?", while
ThumbnailGenerator keeps answering the narrower "can this app decode it
itself?", which is what renditions, the cache and the watermark hook
actually depend on. Image delegates to it so the two cannot drift.
The allowlist stays a security boundary: mime_type is sniffed from the
bytes, so text/html and image/svg+xml remain excluded, and PreviewKind is
deliberately narrower than "formats a browser might cope with" — no
quicktime, avi or matroska, because an embedded player for those shows a
black rectangle. Those still download exactly as before.
docs/security-audit-2026-08-05.md finding 1 recorded that adding
application/pdf "should be a conscious decision". This is that decision,
and three things were measured rather than assumed:
- An <iframe sandbox> cannot be used. Chrome refuses to run its PDF viewer
in a sandboxed frame at all (ERR_BLOCKED_BY_CLIENT, with or without
allow-same-origin) — the attribute removes the feature, it does not
harden it.
- nginx's `Content-Security-Policy: sandbox; default-src 'none'` on
/protected-files/ does work (a <video> frame lands in an opaque origin),
but Chrome exempts its PDF viewer from it, so it is not what protects
the PDF case.
- What does is the allowlist plus the browser's own PDF sandbox, where PDF
JavaScript has no DOM and no cookies.
Range requests were verified end to end: 206 with a correct Content-Range,
a byte-perfect file reassembled from three ranges, and a real browser
seeking to 10s of a 20s clip. nginx drops the upstream Content-Length on
the X-Accel path, so there is no collision.
Two settings, both defaulting on so no installation loses what it has:
clients_can_preview_files and public_listing_preview_enabled. Staff are
never gated. The anonymous side needed a route of its own — there was no
public preview endpoint — with its own throttle bucket, since a bare
throttle: shares one counter across that whole block.
A preview now logs at most one FilePreviewed per viewer per file per five
minutes: a <video> turns one deliberate act into a long tail of Range
requests, and a row each would bury the log.
Also fixes a layout bug the tests could never catch. A portal file row was
flex justify-between with three children — name, comment trigger, download
— so the middle one settled wherever the name happened to end and the
comment icon sat at a different place on every row. The name block now
takes the slack and every action lives in one trailing group, with the
comment trigger in a fixed-width slot so the icons form a column. And
because half the previewable files have no thumbnail to click — a PDF, an
mp3 and an mp4 all render as a generic icon — every row gains an explicit
PreviewAction beside DownloadAction, matching whatever style that theme
gives its download control.
The grouped filter had a second member, "All previews", built on a
public-preview action that does not exist: previewing is recorded one
way today, so its own option already answers "who previewed this?" in
full. Static analysis caught the reference; the group would have been
unreachable even if it had compiled, since a group with a single
present member is deliberately not offered.
Previews get a group here the day a second way to preview a file is
recorded separately, and the test now pins the single-member case on a
file whose log holds one flavour of download.
A file's history was only reachable from the library list, through the
details panel's Activity tab — so anyone who arrived at the file from a
link, a search or a notification had to go back and find the row they
came from to ask what had happened to it.
The file's own page now carries an Activity tab of its own, next to
General and Sharing: the twenty most recent entries, fetched only if the
tab is opened, and a link to the full history. It is behind the same
view_actions_log permission as everywhere else.
That full history is now filterable, which is the point of sending
somebody to it. The action list is built from the file's own log rather
than from the eighty-odd actions the software can record — all but a
handful of which can never apply to a file — and each option carries its
count. Downloads are three separate actions on purpose (a signed-in
recipient, a public link, the public group listing), so "All downloads"
asks that question once instead of three times; the group only appears
when the file's log actually holds more than one of its members.
Narrowing by who acted and by date range works the same as it does on
the main activity log, the reader's own calendar day included.
Delete a folder called Test and you could never have a folder called Test
again. The deletion worked, the folder left the screen, and the name went
with it — permanently, with an error that named a collision against a row
the interface will not show you and offered nothing to do about it.
Files and groups had it too. All three carry a unique index on slug and
all three soft-delete, so the trashed row sat in the index holding a name
nothing could reach. A public one failed outright at the validator, which
checks the table and therefore sees rows the screen does not. A private
one failed more quietly: the derived slug stepped around the trashed row
into report-2, then report-3, once per deletion, climbing forever.
The reservation was deliberate — a trashed row's slug was kept so that
restoring it could not land on somebody else's URL. But nothing in this
application restores anything. There is no restore() call, no route, no
screen; File's own comment says as much. Soft deletes are here so rows can
outlive their delete for foreign keys, the activity log and the erasure
grace period, never so they can come back. The slug was being held for a
page that could not return, and route binding already 404s the trashed row
in the meantime.
So deleting now hands the slug back, and the database is what makes that a
rewrite rather than a gentler lookup: teaching the collision checks to skip
trashed rows would leave two rows holding "report", which the unique index
rejects whatever the application thinks. The slug moves to report__deleted-42
instead. Underscores are the whole trick — Str::slug() turns them into
hyphens and Rules::slug() refuses them outright, so no derived slug and no
hand-typed one can ever land on a vacated one. That is a guarantee about
the character class rather than a hope about collisions.
The format lives in VacatedSlug rather than on the trait because the
migration needs it too and a trait constant cannot be reached through the
trait's own name — the first version of this was a fatal error waiting for
whoever ran migrations. The migration matters as much as the hook: without
it the fix only helps installations that have never deleted anything, and
every name already buried stays buried.
The collision checks still count trashed rows. It costs nothing and keeps
them honest about what the index will accept if a row is ever soft-deleted
by something that bypasses model events.
previous_file_id had this same bug and was fixed this same way, in
File::detachOnDelete — a trashed row holding its predecessor's unique slot
so the chain could never be re-linked. This is that fix, for the other four
unique indexes' worth of the same mistake. users.email is the one left, and
is deliberately not in here: an email address is a login identity rather
than a URL handle, and freeing it silently is the wrong answer.
Fixes#1645
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Client file sharing, rebuilt from the ground up: a private area per
client, resumable uploads, folders, groups and categories, sharing with
expiry dates and download limits, comments, file versions, an activity
log, a REST API, and sixteen languages.
This repository begins here. ProjectSend 2 was developed privately, and
that development history is not published — the previous generation
remains available, with its own history, at projectsend/legacy.
Free software under the GNU General Public License v2, or (at your
option) any later version.
Ignacio Nelson