* ♻️ Handle fetch-error gracefully with toast instead of full-page error
Network-level failures (lost connectivity, DNS failure, etc.) on RPC
calls were propagating as :internal/:fetch-error to the global error
handler, which replaced the entire UI with a full-page error screen.
Now the :internal handler distinguishes :fetch-error from other internal
errors and shows a non-intrusive toast notification instead, allowing
the user to continue working.
* ✨ Add automatic retry with backoff for idempotent RPC requests
Idempotent (GET) RPC requests are now automatically retried up to 3
times with exponential back-off (1s, 2s, 4s) when a transient error
occurs. Retryable errors include: network-level failures
(:fetch-error), 502 Bad Gateway, 503 Service Unavailable, and browser
offline (status 0).
Mutation (POST) requests are never retried to avoid unintended
side-effects. Non-transient errors (4xx client errors, auth errors,
validation errors) propagate immediately without retry.
* ♻️ Make retry helpers public with configurable parameters
Make retryable-error? and with-retry public functions, and replace
private constants with a default-retry-config map. with-retry now
accepts an optional config map (:max-retries, :base-delay-ms) enabling
callers and tests to customize retry behavior.
* ✨ Add tests for RPC retry mechanism
Comprehensive tests for the retry helpers in app.main.repo:
- retryable-error? predicate: covers all retryable types (fetch-error,
bad-gateway, service-unavailable, offline) and non-retryable types
(validation, authentication, authorization, plain errors)
- with-retry observable wrapper: verifies immediate success, recovery
after transient failures, max-retries exhaustion, no retry for
non-retryable errors, fetch-error retry, custom config, and mixed
error scenarios
* ♻️ Introduce :network error type for fetch-level failures
Replace the awkward {:type :internal :code :fetch-error} combination
with a proper {:type :network} type in app.util.http/fetch. This makes
the error taxonomy self-explanatory and removes the special-case branch
in the :internal handler.
Consequences:
- http.cljs: emit {:type :network} instead of {:type :internal :code :fetch-error}
- errors.cljs: add a dedicated ptk/handle-error :network method (toast);
restore :internal handler to its original unconditional full-page error form
- repo.cljs: simplify retryable-types and retryable-error? — :network
replaces the former :internal special-case, no code check needed
- repo_test.cljs: update tests to use {:type :network}
* 📚 Add comment explaining the use of bit-shift-left
* ⬆️ Update opencode and copilot deps
* 🐛 Decouple workspace-content from workspace-local to reduce scroll re-renders
Move workspace-local subscription from workspace-content* (parent) into
viewport* and viewport-classic* (children). workspace-content* now only
subscribes to the new workspace-vport derived atom, which changes only on
window resize — not on every scroll event. This prevents the sidebar,
palette and other workspace-content children from re-rendering on scroll.
* 🐛 Throttle wheel events to one state update per animation frame
Accumulate wheel event deltas in a mutable ref and flush them via
requestAnimationFrame, so that multiple wheel events between frames
produce a single state mutation instead of one per event. This prevents
the cascade of synchronous React re-renders (via useSyncExternalStore)
that can exceed the maximum update depth on rapid scrolling.
Both panning (scroll) and zoom (ctrl/mod+wheel) are throttled. Scroll
deltas are summed additively; zoom scales are compounded multiplicatively
with the latest cursor point used as the zoom center.
* ♻️ Extract schedule-zoom! and schedule-scroll! from on-mouse-wheel
* ♻️ Avoid zoom dep on on-mouse-wheel by using a ref
* 🐛 Fix crash in apply-text-modifier with nil selrect or modifier
Guard apply-text-modifier against nil text-modifier and nil selrect
to prevent the 'invalid arguments (on pointer constructor)' error
thrown by gpt/point when called with an invalid map.
- In text-wrapper: only call apply-text-modifier when text-modifier is
not nil (avoids unnecessary processing)
- In apply-text-modifier: handle nil text-modifier by returning shape
unchanged; guard selrect access before calling gpt/point
* 📚 Add tests for apply-text-modifier in workspace texts
Add exhaustive unit tests covering all paths of apply-text-modifier:
- nil modifier returns shape unchanged (identity)
- modifier with no recognised keys leaves shape unchanged
- :width / :height modifiers resize shape correctly
- nil :width / :height keys are skipped
- both dimensions applied simultaneously
- :position-data is set and nil-guarded
- position-data coordinates translated by delta on resize
- shape with nil selrect + nil modifier does not throw
- position-data-only modifier on shape without selrect is safe
- selrect origin preserved when no dimension changes
- result always carries required shape keys
* 🐛 Fix zero-dimension selrect crash in change-dimensions-modifiers
When a text shape is decoded from the server via map->Rect (which
bypasses make-rect's 0.01 minimum enforcement), its selrect can have
width or height of exactly 0. change-dimensions-modifiers and
change-size were dividing by these values, producing Infinity scale
factors that propagated through the transform pipeline until
calculate-selrect / center->rect returned nil, causing gpt/point to
throw 'invalid arguments (on pointer constructor)'.
Fix: before computing scale factors, guard sr-width / sr-height (and
old-width / old-height in change-size) against zero/negative and
non-finite values. When degenerate, fall back to the shape's own
top-level :width/:height so the denominator and proportion-lock base
remain consistent.
Also simplify apply-text-modifier's delta calculation now that the
transform pipeline is guaranteed to produce a valid selrect, and
update the test suite to test the exact degenerate-selrect scenario
that triggered the original crash.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* ♻️ Simplify change-dimensions-modifiers internal logic
- Remove the intermediate 'size' map ({:width sr-width :height sr-height})
that was built only to be assoc'd and immediately destructured back into
width/height; compute both values directly instead.
- Replace the double-negated condition 'if-not (and (not ignore-lock?) …)'
with a clear positive 'locked?' binding, and flatten the three-branch
if-not/if tree into two independent if expressions keyed on 'attr'.
- Call safe-size-rect once and reuse its result for both the fallback
sizes and the scale computation, eliminating a redundant call.
- Access :transform and :transform-inverse via direct map lookup rather
than destructuring in the function signature, consistent with how the
rest of the let-block reads shape keys.
- Clean up change-size to use the same destructuring style as the updated
function ({sr-width :width sr-height :height}).
- Fix typo in comment: 'havig' -> 'having'.
* ✨ Add tests for change-size and change-dimensions-modifiers
Cover the main behavioural contract of both functions:
change-size:
- Scales both axes to the requested target dimensions.
- Sets the resize origin to the shape's top-left point.
- Nil width/height each fall back to the current dimension (scale 1 on
that axis); both nil produces an identity resize that is optimised away.
- Propagates the shape's transform and transform-inverse matrices into the
resulting GeometricOperation.
change-dimensions-modifiers:
- Changing :width without proportion-lock only scales the x-axis (y
scale stays 1), and vice-versa for :height.
- With proportion-lock enabled, changing :width adjusts height via the
inverse proportion, and changing :height adjusts width via the
proportion.
- ignore-lock? true bypasses proportion-lock regardless of shape state.
- Values below 0.01 are clamped to 0.01 before computing the scale.
- End-to-end: applying the returned modifiers via gsh/transform-shape
yields the expected selrect dimensions.
* ✨ Harden safe-size-rect with additional fallbacks
The previous implementation could still return an invalid rect in several
edge cases. The new version tries four sources in order, accepting each
only if it passes a dedicated safe-size-rect? predicate:
1. :selrect – used when width and height are finite, positive
and within [-max-safe-int, max-safe-int].
2. points->rect – computed from the shape corner points; subject to
the same predicate.
3. Top-level shape fields (:x :y :width :height) – present on all rect,
frame, image, and component shape types.
4. grc/empty-rect – a 0,0 0.01×0.01 unit rect used as last resort so
callers always receive a usable, non-crashing value.
The out-of-range check (> max-safe-int) is new: it rejects coordinates
that pass d/num? (finite) but exceed the platform integer boundary defined
in app.common.schema, which previously slipped through undetected.
Tests cover all four fallback paths, including the NaN, zero-dimension,
and max-safe-int overflow cases.
* ⚡ Optimise safe-size-rect for ClojureScript performance
- Replace (when (some? rect) ...) with (and ^boolean (some? rect) ...)
to keep the entire predicate as a single boolean expression without
introducing an implicit conditional branch.
- Replace keyword access (:width rect) / (:height rect) with
dm/get-prop calls, consistent with the hot-path style used throughout
the rest of the namespace.
- Add ^boolean type hints to every sub-expression of the and chain in
safe-size-rect? (d/num?, pos?, <=) so the ClojureScript compiler emits
raw JS boolean operations instead of boxing the results through
cljs.core/truth_.
- Replace (when (safe-size-rect? ...) value) in safe-size-rect with
(and ^boolean (safe-size-rect? ...) value), avoiding an extra
conditional and keeping the or fallback chain free of allocated
intermediate objects.
* ✨ Use safe-size-rect in apply-text-modifier delta-move computation
safe-size-rect was already used inside change-dimensions-modifiers to
guard the resize scale computation. However, apply-text-modifier in
texts.cljs was still reading (:selrect shape) and (:selrect new-shape)
directly to build the delta-move vector via gpt/point.
gpt/point raises "invalid arguments (on pointer constructor)" when
given a nil value or a map with non-finite :x/:y, which can happen when
a shape's selrect is missing or degenerate (e.g. decoded from the server
via map->Rect, bypassing make-rect's 0.01 floor).
Changes:
- Promote safe-size-rect from defn- to defn in app.common.types.modifiers
so it can be reused by consumers outside the namespace.
- Replace the two raw (:selrect …) accesses in the delta-move computation
with (ctm/safe-size-rect …), which always returns a valid, finite rect
through the established four-step fallback chain.
- Add two frontend tests covering the delta-move path with a fully
degenerate (zero-dimension) selrect, ensuring neither a bare
position-data modifier nor a combined width+position-data modifier
throws.
* ♻️ Ensure all test shapes are proper Shape records in modifiers-test
All shapes in safe-size-rect-fallbacks tests now start from a proper
Shape record built by cts/setup-shape (via make-shape) instead of plain
hash-maps. Each test that mutates geometry fields (selrect, points,
width, height) does so via assoc on the already-initialised record,
which preserves the correct type while isolating the field under test.
A (cts/shape? shape) assertion is added to each fallback test to make
the type guarantee explicit and guard against regressions.
The unused shape-with-selrect helper (which built a bare map) is
removed.
* 🔥 Remove dead code and tighten visibility in app.common.types.modifiers
Dead functions removed (zero callers across the entire codebase):
- modifiers->transform-old: superseded by modifiers->transform; only
ever appeared in a commented-out dev/bench.cljs entry.
- change-recursive-property: no callers anywhere.
- move-parent-modifiers, resize-parent-modifiers: convenience wrappers
for the parent-geometry builder functions; never called.
- remove-children-modifiers, add-children-modifiers,
scale-content-modifiers: single-op convenience builders; never called.
- select-structure: projection helper; only referenced by
select-child-geometry-modifiers which is itself dead.
- select-child-geometry-modifiers: no callers anywhere.
Functions narrowed from defn to defn- (used only within this namespace):
- valid-vector?: assertion helper called only by move/resize builders.
- increase-order: called only by add-modifiers.
- transform-move!, transform-resize!, transform-rotate!, transform!:
steps of the modifiers->transform pipeline.
- modifiers->transform1: immediate helper for modifiers->transform; the
doc-string describing it as 'multiplatform' was also removed since it
is an implementation detail.
- transform-text-node, transform-paragraph-node: leaf helpers for
scale-text-content.
- update-text-content, scale-text-content, apply-scale-content: internal
scale-content pipeline; all called only by apply-modifier.
- remove-children-set: called only by apply-modifier.
- select-structure: demoted to defn- rather than deleted because it is
still called by select-child-structre-modifiers, which has external
callers.
* ✨ Add more tests for modifiers
---------
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Include request URI and status in frontend handle-response error data,
and add request path/context to backend IOException handler logs and
response body. Previously these errors had no identifying information
about which endpoint or request caused the failure.
When AbortController.abort(reason) is called with a custom reason (a
ClojureScript ExceptionInfo), modern browsers (Chrome 98+, Firefox 97+)
reject the fetch promise with that reason object directly instead of with
the canonical DOMException{name:'AbortError'}. The ExceptionInfo has
.name === 'Error', so both the p/catch guard and is-ignorable-exception?
failed to recognise it as an abort, letting it surface to users as an
error toast.
Fix by calling .abort() without a reason so the browser always produces
a native DOMException whose .name is 'AbortError', which is correctly
handled by all existing guards.
Also add a defense-in-depth check in is-ignorable-exception? that
filters errors whose message matches the 'fetch to \'' prefix, guarding
against any future re-introduction of a custom abort reason.
Co-authored-by: Penpot Dev <dev@penpot.app>
* 🐛 Fix TypeError when token error map lacks :error/fn key
Guard against missing :error/fn in token form control resolve streams.
When schema validation errors are produced they may not carry an
:error/fn key; calling nil as a function caused a TypeError crash.
Apply an if-let guard at all 7 affected sites across input.cljs,
color_input.cljs and fonts_combobox.cljs, falling back to :message
or returning the error map unchanged.
* ♻️ Extract token error helpers and add unit tests
Extract resolve-error-message and resolve-error-assoc-message helpers
into errors.cljs, replacing the seven duplicated inline lambdas in
input.cljs, color_input.cljs and fonts_combobox.cljs with named
function references. Add frontend-tests.tokens.token-errors-test
covering both helpers for the normal path (:error/fn present) and the
fallback path (schema-validation errors that lack :error/fn).
Signed-off-by: Penpot Dev <dev@penpot.app>
---------
Signed-off-by: Penpot Dev <dev@penpot.app>
* ✨ Improve error handling and exception formatting
- Enhance exception formatting with visual separators and cause chaining
- Add new handler for :internal error type
- Refine error types: change assertion-related errors to :assertion type
- Improve error messages and hints consistency
- Clean up error handling in zip utilities and HTTP modules
* 🐛 Properly handle AbortError on fetch request unsubscription
When a fetch request in-flight is cancelled due to RxJS unsubscription
(e.g. navigating away from the workspace while thumbnail loads are
pending), the AbortController.abort() call triggers a catch handler
that previously relied solely on a @unsubscribed? flag to suppress the
error.
This was unreliable: nested observables spawned inside rx/mapcat (such
as datauri->blob-uri conversions within get-file-object-thumbnails)
could abort independently, with their own AbortController instances,
meaning the outer unsubscribed? flag was never set and the AbortError
propagated as an unhandled exception.
Add an explicit AbortError name check as a disjunctive condition so
that abort errors originating from any observable in the chain are
suppressed at the source, regardless of subscription state.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 🐛 Fix dissoc error when detaching stroke color from library
The detach-value function in color-row was only passing index to
on-detach, but the stroke's on-color-detach handler expects both
index and color arguments. This caused a protocol error when trying
to dissoc from a number instead of a map.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 🐛 Fix crash when detaching color asset from stroke
The color_row detach-value callback calls on-detach with (index, color),
but stroke_row's local on-color-detach wrapper only took a single argument
(fn [color] ...), so it received index as color and passed it to
stroke.cljs which then called (dissoc index :ref-id :ref-file), crashing
with 'No protocol method IMap.-dissoc defined for type number'.
Fix the wrapper to accept (fn [_ color] ...) so it correctly ignores the
index passed by color_row (it already has index in the closure) and
forwards the actual color map to the parent handler.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
---------
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
When pasting an image (with no text content) into the text editor,
Draft.js calls handlePastedText with null/empty text. The previous fix
guarded splitTextIntoTextBlocks against null, but insertText still
attempted to build a fragment from an empty block array, causing
Modifier.replaceWithFragment to crash with 'Cannot read properties of
undefined (reading getLength)'.
Fix insertText to return the original state unchanged when there are no
text blocks to insert. Also guard handle-pasted-text in the ClojureScript
editor to skip the insert-text call entirely when text is nil or empty.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The splitTextIntoTextBlocks function in @penpot/draft-js called
.split() on the text parameter without a null check. When pasting
content without text data (e.g., images only), Draft.js passes null
to handlePastedText, causing a TypeError.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The save-path-content function only converted content to PathData when
there was a trailing :move-to command. When there was no trailing
:move-to, the content from get-path was stored as-is, which could be
a plain vector if the shape was already a :path type with non-PathData
content. This caused segment/get-points to fail with 'can't access
property "get", cache is undefined' when the with-cache macro tried
to access the cache field on a non-PathData object.
The fix ensures content is always converted to PathData via path/content
before being stored in the state.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The get-parent-with-data function traverses the DOM using parentElement
to find an ancestor with a specific data-* attribute. When the current
node is a non-Element DOM node (e.g. Document node reached from event
handlers on window), accessing .-dataset returns undefined, causing
obj/in? to throw "right-hand side of 'in' should be an object".
This adds a nodeType check to skip non-Element nodes during traversal
and continue up the parent chain.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
When the browser denies clipboard read permission (NotAllowedError),
the unhandled exception handler was showing a generic 'Something wrong
has happened' toast. This change adds proper error handling for
clipboard permission errors in paste operations and shows a
user-friendly warning message instead.
Changes:
- Add error handling in paste-from-clipboard for NotAllowedError
- Improve error handling in paste-selected-props to detect permission errors
- Mark clipboard NotAllowedError as ignorable in the uncaught error handler
to prevent duplicate generic error toasts
- Add translation key for clipboard permission denied message
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
PostHog recorder throws errors like 'Cannot assign to read only property
'assert' of object' which are unrelated to the application and should be
ignored to prevent noise in error reporting.
Add a nil guard before subscribing to the stream in the use-stream
hook. When a nil/undefined stream is passed (e.g., from a conditional
expression or timing edge case during React rendering), the subscribe
call on undefined causes a TypeError. The guard ensures we only
subscribe when the stream is defined.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The with-cache macro in impl.cljc assumed the target was always a
PathData instance (which has a cache field). When content was a plain
vector, (.-cache content) returned undefined in JS, causing:
TypeError: Cannot read properties of undefined (reading 'get')
Fix:
- path/get-points (app.common.types.path) is now the canonical safe
entry point: converts non-PathData content via impl/path-data and
handles nil safely before delegating to segment/get-points
- segment/get-points remains a low-level function that expects a
PathData instance (no defensive logic at that level)
- streams.cljs: replace direct call to path.segm/get-points with
path/get-points so the safe conversion path is always used
- with-cache macro: guards against nil/undefined cache, falling back
to direct evaluation for non-PathData targets
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
- Change the default for the newWindow param from true to false, so
openPage() navigates in the same tab instead of opening a new one
- Accept a UUID string as the page argument in addition to a Page object,
avoiding the need to call penpot.getPage(uuid) first
- Add validation error when an invalid page argument is passed
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
When clipboard items have types that don't match the allowed types
list, the filtering results in an empty array. Calling getType with
undefined throws a NotFoundError. This change adds a check for null/undefined
types and filters them from the result.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Guard against transit-decoded clipboard content that is not a map
before calling assoc, which caused a runtime crash ('No protocol
method IAssociative.-assoc defined for type number').
Also route :copied-props paste data to paste-transit-props instead
of incorrectly sending it to paste-transit-shapes.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Identify and silence "signal is aborted without reason" errors by:
- Providing an explicit reason to AbortController when subscriptions are disposed.
- Updating the global error handler to ignore AbortError exceptions.
- Ensuring unhandled rejections use the ignorable exception filter.
The root cause was RxJS disposal calling .abort() without a reason, combined
with the on-unhandled-rejection handler missing the ignorable error filter.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Guard get-option fallback with (when (seq options) ...) to avoid
"No item 0 in vector of length 0" when options is an empty vector.
Also guard the selected-option memo in select* to mirror the same
pattern already present in combobox*.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Replace int? with number? in on-change handlers for layout item margins,
min/max sizes, and layer opacity. Using int? caused float values like 8.5
to fall into the design token branch, calling (first 8.5) and crashing.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 🐛 Fix error when creating guides without frame
The error 'Cannot read properties of undefined (reading
$cljs$core$IFn$_invoke$arity$0$)' occurred when creating a new
guide. It is probably a race condition because it is not reproducible
from the user point of view.
The cause is mainly because of use incorrect jsx handler :& where :>
should be used. This caused that some props pased with incorrect casing
and the relevant callback props received as nil on the component and
on the use-guide hook.
The fix is simple: use correct jsx handler
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 💄 Add cosmetic changes to viewport guides components
---------
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Two related issues that could cause crashes during fast navigation
in the dashboard:
1. grid.cljs: On drag-start, a temporary counter element is appended
to the file card node for the drag ghost image, then scheduled for
removal via requestAnimationFrame. If the user navigates away before
the RAF fires, React unmounts the section and removes the card node
from the DOM. When the RAF fires, item-el.removeChild(counter-el)
throws because counter-el is no longer a child. Fixed by guarding
the removal with dom/child?.
2. sidebar.cljs: Keyboard navigation handlers used ts/schedule-on-idle
(requestIdleCallback with a 30s timeout) to focus the newly rendered
section title after navigation. This left a very wide window for the
callback to fire against a stale DOM after a subsequent navigation.
Additionally, the idle callbacks were incorrectly passed as arguments
to st/emit! (which ignores non-event values), making the scheduling
an accidental side effect. Fixed by replacing all occurrences with
ts/schedule (setTimeout 0), which is sufficient to defer past the
current render cycle, and moving the calls outside st/emit!.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The previous implementation passed document.body directly as the
React portal containerInfo. During unmount, React's commit phase
(commitUnmountFiberChildrenRecursively, case 4) sets the current
container to containerInfo and then calls container.removeChild()
for every DOM node inside the portal tree.
When two concurrent state updates are processed — e.g. navigating
away from a dashboard section while a file-menu portal is open —
React could attempt document.body.removeChild(node) twice for the
same node, the second time throwing:
NotFoundError: Failed to execute 'removeChild' on 'Node':
The node to be removed is not a child of this node.
The fix allocates a dedicated <div> container per portal instance
via mf/use-memo. The container is appended to body on mount and
removed in the effect cleanup. React then owns an exclusive
containerInfo and its unmount path never races with another
portal or the modal container (which also targets document.body).
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Setting horizontalSizing/verticalSizing on a FlexLayoutProxy was
dispatching update-layout-child instead of update-layout, so the
frame's auto-sizing (hug content) was never triggered even though
the getter read back the value correctly.
Also restricts accepted values to #{:fix :auto} (matching shape.cljs)
since frames cannot use :fill, and fixes a copy-paste error that
reported :horizontalPadding instead of :horizontalSizing in error messages.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
Use nil-safe path/get-points wrapper (some-> based) instead of
direct path.segment/get-points calls in edition.cljs to prevent
'Cannot read properties of undefined (reading get)' crash.
Add nil-safety test to verify path/get-points returns nil without
throwing when content is nil.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
