Store encoded media in RAM, upload a shape-sized display texture
eagerly, and only promote the native full texture when zoom needs
it so oversized fillImages do not saturate integrated GPUs.
Avoid an offscreen buffer per Fill::Image during tile walks: only use
save_layer when a shape image filter is present; axis-aligned rects and
frames without corner radii also skip the redundant container clip.
Present visible tiles via ViewportReady so zoom settle turns sharp
without waiting on the interest ring, and paint at atlas slot size so
DPR 2 does not rasterize 1024 only to downscale into 512 slots.
Keep the shape-to-tile grid zoom-only while scaling GPU paint by DPR
(capped at 1024px). Fix fractional browser zoom and pan, cap atlas slots
with eviction instead of panic, and split paint vs view scale with
explicit paint/atlas/screen tile sizes.
* 🐛 Fix editor v3 quitting when changing typography options
* 🎉 Apply text styles to collapsed caret
* 🐛 Fix not persisting the new selrect
* 🐛 Fix selrect not being recomputed on caret style changes
* 🐛 Fix quitting the editor when changing typography on empty texts
Split the integration suite into four shards running two Playwright
workers each. Median wall time for the job drops from ~40 min to an
expected ~15 min; the build job is unchanged at ~4 min.
Shard reports are merged into a single HTML report, and the merged
run is summarised in the job step summary: totals, failed specs and
flaky specs ranked by retry count.
Chromium is installed into a shared volume so shards do not
re-download it. `workflow_dispatch` allows running the suite manually
against an arbitrary ref, with configurable shard layout and workers.
PRs targeting `staging` keep running serially while the current
release stabilizes. The exception is marked TEMPORARY and removed in
a follow-up.
Enable Playwright's JSON reporter alongside `list` and publish a
summary of flaky tests to the job step summary. The JSON report is
kept as an artifact for 30 days so flakiness rates can be aggregated
over time.
CI already runs with `retries: 2`, so unstable tests have been passing
silently on retry. This only surfaces what the suite already absorbs;
no test behaviour changes.
The reporter in `frontend/scripts/test-e2e` becomes overridable via
`PLAYWRIGHT_REPORTER` so the local developer default stays untouched.
Enable Playwright's JSON reporter alongside `list` and publish a
summary of flaky tests to the job step summary. The JSON report is
kept as an artifact for 30 days so flakiness rates can be aggregated
over time.
CI already runs with `retries: 2`, so unstable tests have been passing
silently on retry. This only surfaces what the suite already absorbs;
no test behaviour changes.
The reporter in `frontend/scripts/test-e2e` becomes overridable via
`PLAYWRIGHT_REPORTER` so the local developer default stays untouched.
The comments layer lives in the viewport overlays, which are absolutely
positioned above the canvas, and the container itself carries a high
z-index. A comment bubble panned into the ruler bars therefore painted
on top of them, covering the ticks and numbers.
Clip the comments container to the area outside the ruler bars while
the rulers are visible, the same thing the `clip-handlers` clip path
already does so the selection handlers stay off the rulers. Clipping
only the comments container leaves the text editing overlay, which
shares the viewport overlays, untouched.
Fixes#11163.
Signed-off-by: Filip Sajdak <filip.sajdak@siili.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The generated /etc/nginx/overrides/server.d/mcp-locations.conf used a
plain proxy_pass target (e.g. `proxy_pass http://penpot-mcp:4402;`)
where $PENPOT_MCP_URI/$PENPOT_MCP_URI_WS are shell variables substituted
once by envsubst in nginx-entrypoint.sh at container startup, not nginx
variables. nginx resolves a literal proxy_pass hostname once when the
config loads and never re-checks it, so the existing
`resolver 127.0.0.11 valid=10s;` directive in
overrides/http.d/resolvers.conf has no effect on these three locations
- it only applies to nginx variables evaluated per-request.
In multi-container deployments where the penpot-mcp container restarts
or is recreated independently of penpot-frontend (image update, OOM,
orchestrator reschedule), it gets a new IP from Docker's/the
orchestrator's DNS, and the frontend's nginx keeps forwarding to the
old, now-dead address until penpot-frontend itself is restarted. This
surfaces to users as `wss://<host>/mcp/ws` failing to connect from the
browser after enabling the MCP plugin, with
`connect() failed (111: Connection refused)` in the frontend's nginx
logs.
Route each location through a `set $var ...; proxy_pass $var;` pair so
proxy_pass evaluates a real nginx variable, letting the pre-existing
resolver directive re-resolve penpot-mcp within its 10s TTL instead of
caching the address for the container's lifetime.
For /mcp/stream and /mcp/sse, the set value also appends
$is_args$args explicitly: when proxy_pass targets a variable AND that
variable's value includes a URI/path component, nginx does not
automatically forward the original request's query string the way it
does for a static proxy_pass target - it must be appended by hand, or
the userToken query parameter used for multi-user authentication is
silently dropped before reaching the MCP server. /mcp/ws has no path
component in its target so it isn't affected by this and needed no
such change.
Verified locally: force-recreated the penpot-mcp container onto a
different IP while leaving penpot-frontend untouched; the /mcp/ws
WebSocket upgrade kept returning 101 Switching Protocols throughout,
both immediately and after the resolver's TTL window. Separately
verified /mcp/stream: a POST with ?userToken=... now shows up
server-side as userTokenFp=<redacted first 8 chars> instead of <none>,
and an actual MCP client (Claude Code) using this proxy can now call
authenticated tools like execute_code successfully.
Signed-off-by: Jules LaPrairie <jules@lucidbox.ca>
The design sidebar named the same "mixed values" concept with two
different translation keys. Most sections use settings.multiple, while
the blur options and the design system numeric input used
labels.mixed-values.
Both read "Mixed" in English, so the split is invisible in the default
locale, but labels.mixed-values has no translation at all in 16 locales
and a different wording in 8 more. Where it is missing the string falls
back to the default language, so those controls rendered the English
word next to sections showing the localized one; where both exist, a
single sidebar named the same concept two ways (fr "Divers" against
"Melange", ru "Smeshanyy" against "Smeshat").
Point the two outliers at settings.multiple, the key the rest of the
sidebar already uses and the one translated in every locale that ships
a translation for it.
Fixes#11148.
Signed-off-by: Filip Sajdak <filip.sajdak@siili.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Add `project-id` to the guard condition in `use-plugin-register`'s layout effect so the plugin "Try out" flow waits until projects have loaded.
Previously, only `plugin-url` was checked, which allowed the fetch to fire
before projects were available, sending a nil `project-id` and causing a 400
validation error from the backend.
AI-assisted-by: mimo-v2.5-pro
* 🐛 Declare the shape attributes stored files carry
`schema:shape-attrs` is the shape model as *declared*, and it has fallen
behind the `Shape` record. Three record fields are absent from it:
`rotation`, `flip-x` and `flip-y` are therefore present on every shape
that exists and declared nowhere. `rotation` is already named twice in
this namespace, in `allowed-shape-attrs`, and once in
`app.common.types.shape.attrs/editable-attrs`, so the schema is
demonstrably the odd one out rather than the data being unusual.
Nothing complains, because the maps are open: an undeclared key
validates fine. What breaks is everything that reads the model *from the
schema* rather than from a live value, such as the generative tests'
shape generator, the generated OpenAPI surface, and any consumer
reflecting over `schema:shape-attrs`.
Whether an entry is optional, nilable, or both is decided by the record
rather than by taste. `app.common.record/defrecord` cannot remove a base
field: its `without` assocs nil and its `containsKey` answers true
whatever the field holds, on both platforms. So a `Shape` base field is
always present, and nil is how that field says "unset". Every other key
lives in the `$extmap`, disappears on dissoc, and is dropped by
`setup-shape` when a caller passes nil. Base fields are therefore
nilable, and the rest are optional.
Declared here, measured over a 305-shape corpus:
- `rotation`, `flip-x` and `flip-y`, record fields present on every
shape, nilable for the reason above: `make-minimal-shape` gives the
two flip fields no default, so they are nil on all 305. Optional as
well, unlike the geometry below, because `schema:shape-generic-attrs`
has a second job: `check-shape-generic-attrs` validates partial update
payloads with it, such as the `{:blocked true}` that
`app.main.data.workspace/update-shape` passes, and a required key here
would reject every such payload.
- `hide-in-viewer`, moved out of `schema:frame-attrs`, because circles,
rects and texts carry it too, 197 shapes.
- `svg-attrs`, `svg-defs`, `svg-transform` and `svg-viewbox`, the SVG
provenance an import leaves behind, 101 shapes and 63 for the
transform. Typed `:map` rather than more precisely on purpose: legacy
files hold `svg-transform` as a plain `{:a … :f}` map rather than a
`::gmt/matrix` record, and `svg-viewbox` as either a `::grc/rect`
record or a plain map, so a tighter schema would reject files that are
otherwise valid.
- `use-for-thumbnail` on frames. The model has long had it:
`app.common.files.migrations` renames `:use-for-thumbnail?` to it and
`app.common.logic.libraries` reads it. This schema had not declared
it.
- `rx` and `ry` on rects and circles, the legacy radii SVG import parses
off the element and migration 0003 assocs as `0`. Superseded by `r1`
to `r4`, but stored files carry them.
- `content` on svg-raw. `shapes-builder/create-raw-svg` sets it and
`allowed-svg-attrs` names it. Typed `[:or :map :string]`, because a
bare text node arrives as the string itself: `<text>hi</text>` becomes
one svg-raw for the element and another for `"hi"`, and
`shapes-builder/parse-svg-element` carries a FIXME about exactly that.
`schema:nilable-geom-attrs` is new, for bool and path. Those two are the
only shape types whose geometry can be nil: `make-minimal-shape` gives
`x`, `y`, `width` and `height` a default for every other type and skips
those two, whose extent their content and `selrect` imply instead. The
four keys stay required, as they already are in the other seven
branches, and only the nil is new.
**Do not make the analogous change to `ctf/schema:file`.** That map
carries `:backend`, `:comment-thread-seqn` and `:ignore-sync-until`,
none of which the schema declares, and declaring them breaks saving:
`app.binfile.common/update-file!` derives its UPDATE column list from a
file map's keys, and the `file` table has no `backend` column, it being
synthesized on read. Measured at 185 failures, mostly `rpc-file-test`.
Whether a schema serving as both read description and write contract is
itself a defect is a real design question, and a separate one. The
`check-shape-generic-attrs` case above is a second instance of it.
Adding entries changes what `shape-generator` produces, so generative
tests begin exercising code paths with these attributes present. That is
where a problem would surface. With this applied the common suite is
1142 tests and 24702 assertions on the Clojure side, 992 tests and 24017
assertions on the ClojureScript side, no failures on either.
AI-assisted-by: mixed models
* ✨ Align shape generator with declared schema and add key-presence test
shape-generator now selects geometry attrs per-type: nilable-geom-attrs
for bool/path, shape-geom-attrs for everything else, and always merges
them. This removes the dead attrs2 generation for bool/path and the
implicit dependency on create-shape adding nil defaults for missing
base record fields.
The new shape-generator-key-presence test asserts that generated shapes
carry the required keys: rotation, flip-x, flip-y on all shapes and x,
y, width, height on bool/path, even when nilable.
AI-assisted-by: longcat-2.0-free
* 🐛 Sample 200 shapes in the key-presence test, not 10
`sg/sample` hands its options to `malli.generator/sample`, which reads
`:size`. `:num` is test.check's option. It is correct for the
`smt/check!` call directly above, where it came from, but `sg/sample`
ignores it and falls back to its default of 10.
Ten samples leave the bool and path assertions vacuous about one run in
fourteen. Simulated over 200 draws of 10, 14 contained no bool and no
path at all, and the median draw held 2. Those four assertions defend
exactly the keys this branch made required, so a run that skips them
silently is the one case worth not missing.
The assertion count shows the arithmetic. The test contributed 42 with
`:num`, which is 10 shapes times 3 keys plus 3 bool-or-path shapes times
4 keys, and contributes 756 with `:size`. The common suite goes from
1143 tests and 24744 assertions to 1143 tests and 25458 assertions, no
failures either way.
AI-assisted-by: mixed models
---------
Co-authored-by: Andrey Antukh <niwi@niwi.nz>
When a profile is deleted, only the current session was being
invalidated. Other active sessions on different devices remained
functional until the background cleanup task completed.
Add session/invalidate-all helper that deletes all sessions for
a profile by profile_id, and call it from delete-profile before
the response transform. This ensures immediate access revocation
across all devices when an account is deleted.
Closes#11114
AI-assisted-by: qwen3.7-plus
The fetch-manifest function previously had no timeout, causing the
plugin installation flow to hang indefinitely if the server accepted
the connection but never completed the response.
Added a 15-second timeout using rx/timeout to abort the request
automatically.
Closes#11119
AI-assisted-by: qwen3.7-plus
Replace the placeholder rlimit.edn with a real per-endpoint
configuration covering auth, SSRF, search, email, media and project
operations. The previous file only had a commented-out example, so
all limits fell back to the 200k/h default window.
Also propagate the evaluated `now` timestamp into both bucket and
window result maps, so consumers (e.g. soft-mode reports) can know
exactly when the limit was checked.
AI-assisted-by: minimax-m3
Share link IDs function as capability secrets — anyone possessing
the ID can read a file without authentication. The previous UUIDv8
scheme is predictable (56 bits fixed per process + 48-bit timestamp).
Changed to uuid/random (UUIDv4) for genuine unpredictability.
Closes#11116
AI-assisted-by: qwen3.7-plus
Replace standard '=' operator with MessageDigest/isEqual to prevent
timing attacks on shared key authentication middleware.
Closes#11121
AI-assisted-by: qwen3.7-plus
* 🐛 Add permission checks to WebSocket subscription handlers
Check file and team read permissions before allowing WebSocket
subscriptions to prevent resource enumeration via presence
notifications.
AI-assisted-by: mimo-v2.5-pro
* 🐛 Fix random backend test failure
