On the WebGL renderer, a sync from a main component re-measures every
auto-grow text it changes on the page. The sync runs on every edit to
a main component, so an edit that cannot change a text size, such as
opacity, shadow or text color, still re-measures and re-lays out every
synced text.
The sync now re-measures a text only when the change sets an attribute
that can change its size: :grow-type, :width, :height,
:layout-item-h-sizing, :layout-item-v-sizing, or :content that differs
in more than its fills. A text color change reaches copies as a
:content change, so fills are removed from both contents before
comparing. Sync emits operations only for attributes that changed, so
the operation attributes are enough. Moves (x and y only) do not
re-measure.
Tests: frontend-tests.data.workspace-sync-text-wasm-test adds cases for
a fill change and an opacity change on the main text (no re-measure),
and for a font size change and a width change on an auto-height text
(re-measure).
Relates to #12721
AI-assisted-by: claude-opus-5-5
On the SVG renderer, opening a page measures auto-grow text copies
with overridden content and writes back their size and position data
when they drifted. The write went through the normal text commits,
which stack onto the user's last undo entry. So the first Ctrl+Z after
opening a page also reverted the heal.
A component sync also set the position data of a copy to nil, even
for a fill change on the main. Such a copy skipped the check and went
through the normal path, with undo.
Heal requests now take :save-undo? false. resize-text and
update-position-data accept optional opts, and a request with
:save-undo? false marks its text as a heal for the current batch. A
later user request for the same text in the batch removes that mark.
The batch commit splits in two. User texts commit as before. Heal
texts commit with :save-undo? false and :skip-component-sync? true,
so each copy heals on its own page open instead of syncing.
The heal resize asks for the parents' reflow with :save-undo? false.
The layout buffer lays out heal updates apart from user ones and
passes the flag to apply-modifiers, so the reflow records no undo
step either.
The :check phase also heals a copy without position data, with no
undo. When the size drifted, the text moves to a new :settle phase.
The measurement that the heal resize causes also commits without undo.
Known limit: when the heal's reflow moves the text (for example,
re-centering it in a flex parent), the renderer measures the text
again with no phase. That position-data commit still stacks onto the
last undo entry. Left for a follow-up.
Tests: frontend-tests.data.workspace-heal-undo-test (SVG store) covers
a heal resize with no undo, no component sync and a reflow without
undo; a heal and a user change in one batch; a user request after a
heal of the same text; and the layout buffer keeping a heal's reflow
apart. frontend-tests.ui.viewport-texts-html-test covers the options
each phase sends: drifted copy, matching copy, copy without position
data, unmeasurable text, settle, and no phase.
Relates to #12721
Relates to #13065
AI-assisted-by: claude-opus-5-5
On the WebGL renderer, the workspace resizes auto-grow texts whose
stored size no longer matches their WASM layout. This runs when a page
opens and when fonts finish loading. These resizes are not user
actions, but each one recorded an undo step. So the first Ctrl+Z after
opening a page undid the resize instead of the user's last edit, and
the copy went back to the wrong size.
The listener for the stale-text event now asks for the resize with
:save-undo? false. The flag passes through the resize debounce to the
commit, which hands it to apply-wasm-modifiers, so the shape update
records no undo step.
Resizes are batched in a debounce window. A batch skips the undo step
only when every request in it asked for that. A user resize that lands
in the same window keeps its undo entry. Batch options are now merged
for every request, including those without options, so this check
works.
Tests: frontend-tests.data.workspace-heal-undo-test covers a heal
resize that commits without an undo step, a resize without the flag
that still records one, a mixed batch that keeps the undo step in
either order, and a batch of heals that skips it.
Relates to #12721
Relates to #13065
AI-assisted-by: claude-opus-5-5
On the SVG renderer, a text is measured only when its text
properties change or when it has no position-data. A component copy
with overridden text content keeps its own position-data. If its
main changes while the page is closed, the copy takes the main's
size, and nothing measures it when the page opens. It keeps the
main's height, so its flex siblings overlap the extra lines.
When a page opens, each auto-width or auto-height text with
overridden content on a copy starts in a :check phase and gets one
hidden HTML measurement. In :check, the result is written (size and
position-data) only if the size differs by more than 1px on the axis
the text grows on. Otherwise nothing is written, though the text
modifier is still cleaned. The phase is then dropped, so later
measurements work as before. A copy with no position-data skips the
check and is measured and committed as usual.
The new dwt/text-size-drifted? holds the 1px rule. It ignores an axis
that fills its layout parent, because the layout sets that size and
the check would otherwise find drift on every open. The 1px tolerance
matches the WebGL page-load check. SVG stores whole pixels, rounded
up, so smaller gaps are only rounding. A finer tolerance would
rewrite the same texts each time a page opened.
Each content-overridden auto-grow copy costs one extra measurement
per page open.
Tests: frontend-tests.data.workspace-texts-test covers
text-size-drifted? for sub-pixel and exact 1px gaps, the growing axis
only, and fill axes. frontend-tests.ui.viewport-texts-html-test covers
which copies start in :check, the drifted, matching and unmeasurable
cases, and the 1px commit when no phase is set.
Known limit: the heal commits like any SVG measurement, so it adds to
the user's last undo entry.
Relates to #12721
Relates to #13065
With the WebGL renderer, a component copy synced while its page was
closed takes the main's size for its auto-width or auto-height text.
Nothing measured it again when the page opened. The only check of stored
text sizes against the WASM layout ran after font loads, and only for
texts that waited for a font. Once fonts were loaded, which is true for
every page after the first in a session, the stale copy kept the main's
size. Its flex siblings then overlapped the extra lines.
Page loads now check every auto-grow text they bring into WASM. This
covers the sync path for small pages and the chunked async path for
pages with 100 or more shapes. A stale text goes through the existing
stale-text event, so the workspace resizes it. Texts still waiting for a
font skip the check, since the font check handles them when the font
arrives. So no text is resized twice.
The page check uses a 1px tolerance, and the font check keeps 0.1px. The
classic renderer stores whole pixels and WASM stores fractions, so a
finer tolerance would rewrite the same texts each time a page opened.
Shape updates made while editing do not run the check, so typing does
not resize texts.
Tests: new frontend-tests.render-wasm.stale-text-selrects-test covers
small and large page loads, sub-pixel gaps, fixed texts, texts waiting
for a font, and shape updates while editing.
Note: set-objects also serves thumbnails, the viewer and the export
worker, where nothing listens to the event. There the check costs one
cached size read per auto-grow text, the same cost the font check already
had for texts that waited for a font.
Relates to #12721
Relates to #13065
Create an issue and assign its Issue Type in one command instead of the
manual create-then-mutate flow. The type name is case-insensitive and the
command prints the issue number, URL, title and type as JSON.
Add unit tests for the full flow, labels, invalid types, gh failures and
the CLI output, and update the gh, creating-issues and create-issue skill
documentation.
AI-assisted-by: hy4-flash-free
Hosts need a vendor-agnostic way to react to named audit events in the
browser (for example SaaS config.js). Penpot exposes
penpotAddEventCallback / penpotEventCallbacks, loads an early stub before
config.js, and invokes matching callbacks after make-event, without
coupling to any analytics SDK.
Closes#13091
AI-assisted-by: Composer
The analysis of a package reads the metadata of every media and
thumbnail of its container to attribute the shared objects to each
file. It read them all at once, one promise per entry, and the
container of a big file holds thousands of them: every read opens its
own decompression stream, and Chromium fails the whole analysis under
that load while Firefox keeps up. The failure is masked by zip.js with
a generic "Cannot close a ERRORED writable stream".
The references now go through an observable, one entry at a time, and a
reference that cannot be read leaves its file without a size instead of
failing the package: a number that misses one of the objects of the
file is worse than no number.
Closes#13022
AI-assisted-by: deepseek-v4.1-flash
Add the disable-admin-console flag to the backend devenv PENPOT_FLAGS.
The nginx entrypoint checks this flag to decide whether to proxy the
/admin-console path; without it, devenv configured a proxy to an
admin-console service that does not run there. This aligns backend
routing with the frontend, where the admin console is off by default
in devenv.
AI-assisted-by: hy4-flash-free
* 🐛 Keep fill copies laid out after a component sync in the WebGL render
Lay out other pages in CLJS and reflow nested layouts that keep their size.
Closes#13015
AI-assisted-by: claude-opus-5-5
* ⚡ Skip the nested layout pass during WebGL drag previews
Gap, padding and margin drags keep the cost they had before the fix.
Relates to #13015
AI-assisted-by: claude-opus-5-5
Plugins and the MCP could not learn which Penpot instance they run on,
so agents guessed the domain when they wrote links to files.
Add penpot.location with the instance base URL (baseUrl), the URL of
the current screen (href) and its query params (params). It works in
the dashboard and the workspace and requires content:read. Params that
carry secrets (share-id, the MCP authorize params, and any key ending
in token, secret or password) are left out, and href is rebuilt from
the remaining params so it never carries one.
The MCP instructions tell agents to use href instead of building links.
The plugin API test suite checks location against the open file, and
its API surface is regenerated from the current types.
Closes#13013
AI-assisted-by: claude-opus-5-5
* 🔧 Test unique layer names in the file builder
The builder read used names from :unames while track-used-name wrote
them to ::unames, so two boards without a name both ended up as
"Board". #12794 fixed the key. These tests cover it: two unnamed boards
get "Board" and "Board-1", and two names that clean to the same path
get "-1" on the second one. Both fail with the old key.
AI-assisted-by: claude-opus-5-5
* 🔧 Test that component root names count as used names
#12794 also made add-component register the renamed main instance name.
A shape added later with the same name must get "-1". The test fails
without that track-used-name call.
AI-assisted-by: claude-opus-5-5
* 🔧 Test per-page unique names and anonymous components
The builder tracks used names per page, so the same name on two pages
stays unchanged. A component without a name is called "anonymous".
Each test fails without the code it covers.
AI-assisted-by: claude-opus-5-5
Promote Clojure design rules to a required core memory and simplify
the discovery guide. Correct test output redirection and RPC routes.
AI-assisted-by: gpt-6.1-terra
Protect staged changes when rewriting commit messages and close plans
after verified commits. Clarify checks and remove the unused Taiga skill.
AI-assisted-by: gpt-6.1-terra
Make review criteria proportional to risk and project conventions.
Keep code reviews read-only, use existing test evidence and stable IDs.
AI-assisted-by: gpt-6.1-terra
Apply clear review improvements to the plan and resolve consequential
decisions through the walkthrough before updating the plan metadata.
AI-assisted-by: gpt-6.1-terra
Include remedies, tests and decisions in review findings. Preserve review
context in plans and offer to resume the reviewer after implementation.
Keep follow-up verification optional and report deferred findings clearly.
AI-assisted-by: gpt-6.1-terra
Agent skills are invoked directly with @-mentions, so the slash-command
wrappers that exposed them as /commands are no longer needed. Update
.agents/README.md and every skill that references another skill to use the
@ syntax, and delete the six .opencode/commands/ dispatcher files.
AI-assisted-by: hy4-flash-free
Measure the coverage of the backend, common, frontend and render-wasm
test suites in CI, and show it on the workflow run pages and in one
"Coverage Report" comment on each pull request.
- Backend and common run kaocha with the cloverage plugin through a new
`:coverage` alias, which writes lcov, HTML and an EMMA report. The PR
report shows lines and Clojure forms (as statements); cloverage
measures no functions or branches. Plain local runs are unchanged.
- Frontend runs the compiled test bundle with V8 coverage
(`pnpm run test:coverage`, `frontend/scripts/coverage.js`) and maps it
back to the ClojureScript sources through the source maps of the test
build. UI namespaces (`app/main/ui`) are left out.
- Render-wasm runs its native Rust tests under cargo-llvm-cov when
`RENDER_WASM_COVERAGE` is set. The devenv image lacks the tool, so the
workflow downloads a pinned release and checks its sha256.
- `scripts/coverage-summary.mjs` renders an lcov file as a markdown table
by source area for `$GITHUB_STEP_SUMMARY`; each workflow also uploads
the HTML report as an artifact.
- `scripts/coverage-comment.mjs` keeps a single PR comment, laid out like
the Vitest Coverage Report Action, with a section per project and the
coverage of the changed files. Each job updates its own section; the
data of every project lives in a hidden block of the comment, and a
job writes again when a concurrent one dropped its data.
- Lines have targets that only colour the report: 70% backend and
common, 50% frontend, 40% render-wasm.
- The comment step only runs for PRs from branches of this repo: fork
PRs get a read-only token and keep the summary on the run page.
AI-assisted-by: claude-opus-5-5
Resolve the UI CSS source and destination from the package directory so
Vite works when launched outside that directory. Add a regression test and
quiet the bundle's rsync output.
AI-assisted-by: gpt-6-luna
* 🐛 Fix shared files 500 on shapes with invalid child ids
Old files can list in :shapes ids that are not uuids or that point to
no shape. Migration 0033 walked them and hit the uuid assert in
get-shape, which throws an AssertionError. The shared files list only
caught Exception, so that Error broke the whole list with a 500.
Make migration 0033 skip those ids, and catch Throwable when loading
each shared file, so a broken file is skipped and logged.
Closes#12343
AI-assisted-by: claude-opus-5-5
* 🐛 Repair invalid child ids in migration 0033 v2
Skipping invalid child ids in migration 0033 was not enough: the ids
stayed in :shapes, so the file still failed schema validation and
could not load. Files that already applied the migration would also
never run a changed version of it.
The migration now repairs each container before the naming walk:
string ids are parsed with uuid/coerce, nil and ids that point to no
shape are dropped, the duplicate that 0032 adds when a child is
listed as a string is removed, and a :shapes list becomes a vector.
Only the shapes that change are re-assoc'ed, so a healthy ObjectsMap
is not marked as modified and its page is not rewritten.
It ships as 0033-normalize-main-instance-names-v2 at the same
position in available-migrations, so files that recorded the old id
run it again. That position also makes it parse strings before
0035-fix-dangling-children drops the ids it cannot find.
The shared files listing restores the interrupt flag when the cause
of a skipped file is an InterruptedException.
Tests cover children outside the root walk, legacy components, the
objects-map wrapper of the unrealized save path, a list turned into a
vector, a shape without a :shapes key, a file that applied the old
id, and the end-to-end load of a broken shared file.
Relates to #12343
AI-assisted-by: claude-opus-5-5
AI-assisted-by: inkling-2-small-free
When a peer restored an autosaved version, the notice said "restored an
earlier version" instead of naming it by its date and time.
The modal wrapper passes the props map to the component without
camelCasing its keys, so the `created-at` prop reached the component as
"created-at" while it read `createdAt`, and the date was always nil.
Named versions were fine because `label` is one word. The prop is now
`date`, and a test renders the registered modal the way the wrapper
does.
Closes#13037
AI-assisted-by: claude-opus-5-5
* ⬆️ Bump the JDK to 27.0.1.1 on devenv and the backend image
The devenv image compiled the backend bundle with JDK 27 while the
production image ran JDK 26. The runtime starts with --enable-preview,
so a runtime line behind the compiler line is a silent trap rather than
a working setup.
Both pins now use Zulu 27.0.1.1 (zulu27.30.201-ca), the current GA of
the line, with the same sums on both architectures. The jlink module
list in the backend image still resolves on 27, so it does not change.
Sums checked by downloading both tarballs; the jlink JRE built from
them starts and reports 27.0.1.1.
AI-assisted-by: deepseek-v4.1-flash
* 🔥 Remove the unused gitpod setup
No workflow built docker/gitpod and its AdoptOpenJDK 16 install no
longer matches the devenv. The README badge linked to a setup that no
longer exists, so it goes with it.
The historical changelog entry stays: CHANGES.md is generated from the
milestones.
AI-assisted-by: deepseek-v4.1-flash
* ⬆️ Bring the docker/ updates from develop
Move the devenv and the production images to what develop already
carries:
- Base images to the current dhi.io digests.
- Node 24.22.3 and pnpm 12.21.0 in the devenv image and in every
production Dockerfile that ships pnpm.
- Rust 1.99.0 and Serena 2.1.0, with SERENA_UPDATE_VERSION raised to
match and a note that it must never be lower than the image pin.
- tini as the exporter ENTRYPOINT.
Devenv additions: PENPOT_DEVENV=true so the frontend compiles with the
admin console disabled by default, xauth and libvulkan1 plus the GPU
compose overlay, the MCP OAuth metadata route in nginx, and clipboard
support in tmux.
The JDK stays at 27.0.1.1 on both pins.
AI-assisted-by: deepseek-v4.1-flash
* ⬆️ Finish the Node 24.22.3 bump outside docker/
The docker/ import brought Node 24.22.3 into the devenv and the image
tags while .nvmrc and the two plugins-deploy workflows stayed on
24.22.2.
Node stays on the 24.x line on every site: the plugins-deploy runtime
pins follow .nvmrc, even though develop moved them to node@26.13.1 in
99db795966.
AI-assisted-by: deepseek-v4.1-flash
* 🐛 Validate the shapes of a page added by an update-file change
The backend validates every shape an `:add-obj` or a `:mod-obj` change
writes, but an `:add-page` change carries a whole page and its shapes
reached the file unchecked. Duplicate page and the undo of a page
deletion send that page as the client holds it, so a tab that loaded a
file before a migration ran could write the old format back into it.
A tab opened before #12006 deployed holds camelCase svg attr keys such
as `:fillRule`, and the closed `:svg-attrs` schema rejects them. The
file saves, and the next read that migrates it fails the file data
check, so the whole file stops opening (#13107).
`:add-page` now validates each shape of the page on the backend, as
`:add-obj` does for one.
AI-assisted-by: claude-opus-5-5
* 🐛 Rerun the svg attr normalization on files that recorded 0035
Files that recorded migration 0035 never run it again, so camelCase
svg attr keys written after it ran stay in the file. A tab that loaded
a file before 0035 ran could save such keys through an `:add-page`
change, and the closed `:svg-attrs` schema then rejects the whole file
on its next migrating read (#13107).
The svg attr normalization now runs under the id
`0037-fix-shape-svg-attrs-and-defs`, and the 0035 id leaves
`available-migrations`, as 0021 did when 0035 replaced it. Every file
that recorded 0035 migrates again. Kebab input passes through
unchanged, so the rerun touches only the keys written after 0035.
AI-assisted-by: claude-opus-5-5
* ♻️ Move the 0037 migration tests to their own namespace
The svg-attr migration was renamed to 0037, but its tests stayed in
the 0035 test namespace under migration-0035-* names. Move the whole
section to files_migrations_0037_test.cljc, register the new namespace
in the test runner, and rename its tests to match the migration id.
Replace the doseq over the two withdrawn ids with two self-contained
deftests, one for 0021 and one for 0035. No production code changes.
AI-assisted-by: hy4-flash-free
* 🐛 Keep the log watch installed until the rejection record arrives
The logger swaps the record on an executor thread, after the call that
logged it returns. The test removed the watch in the `finally` around
`update-file` and only then polled for the record, so under load the
record landed on nobody and the test read a nil `:attempts`.
AI-assisted-by: deepseek-v4.1-flash
* 🐛 Count only update-file calls in the save-test transport log
The mocked transport is installed with `set!` and outlives the fixture: a
timer another module left pending can still reach it mid-test, as a
`:delete-file-object-thumbnails` call did in CI. Filter to the only
command persistence sends, `update-file`, so a stray call neither fails
the emptiness assertion nor releases a `wait-for` on a save.
AI-assisted-by: deepseek-v4.1-flash
---------
Co-authored-by: Andrey Antukh <niwi@niwi.nz>
worker_processes auto starts one worker per host CPU, ignoring the
container CPU limit, and since worker_connections went from 2048 to
65535 each worker reserves about 30 MB at startup. A frontend limited
to a few CPUs on a 64-core node needs close to 2 GB before serving a
request and gets OOM killed.
Default the worker count to the CPUs the container can use, its cpuset
and its cgroup CPU quota rounded up, when they are fewer than the CPUs
of the host, and keep auto otherwise. Make both settings configurable
through PENPOT_HTTP_SERVER_WORKER_PROCESSES and
PENPOT_HTTP_SERVER_WORKER_CONNECTIONS, and refuse to start on values
that are not valid.
Refs #13103
Signed-off-by: David Barragán Merino <david.barragan@kaleidos.net>
Files written by old Penpot versions hold maps where an svg
attribute value should be, such as a `:filter` that once read
`url(#filter_2)`. Since #12006 the `:svg-attrs` schema is closed and
admits scalars only, and migration 0035 kept those maps, so the file
data check at the end of every migration rejected the file. Such a
file can be neither imported nor opened. Two built-in templates,
Avataaars and Tutorial for beginners, are among them.
`attrs->kebab-props` now drops a value that is not a string, a number,
or a boolean, at the top level and inside `:style`. Migration 0035 and
every birth path go through it, so stored attrs match the schema.
Of the 143 files in penpot/penpot-files, 4 fail to import at
`:svg-attrs :filter` before this change, and all 4 import after it.
AI-assisted-by: claude-opus-5-5
check_signed_off_by was never registered in the validator list, so it never
ran, and no test exercised it. The DCO sign-off applies to external
contributors through CONTRIBUTING.md, not to every commit, so the local
checker has nothing to enforce there.
Drop the function and the DCO mention in the script docstring.
AI-assisted-by: deepseek-v4.1-flash
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
code-review-criteria and review-code carried weight that changed no
decision: a 16-row rationalizations table, red flags that repeated the axes,
and three sections written for the author of a change rather than for the
reviewer. 338 lines become 251.
Drop that layer, dedupe DRY/KISS/YAGNI between the core principles and the
axes, merge the two instruction lists of the flow, and compress the GitHub
comment format.
The reviewer no longer re-runs the suite or a build: the author's run and CI
already provide that evidence, and repeating it only made reviews longer. It
runs at most the single test file that settles a specific doubt, and states
which evidence it relied on.
AI-assisted-by: deepseek-v4.1-flash
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
The checker rejected any Co-authored-by trailer, which also blocked a real
person. The trailer is meant to credit a human co-author, and git and
GitHub use it for that; only AI providers, agents and bots are unwanted.
Match the trailer value against a blocklist of providers and agents plus
bot identities, and point to the dedicated AI-assisted-by trailer for AI
participation. Co-authored-by joins the trailer exemption of the
80-character body rule.
Update the tests, CONTRIBUTING.md, the create-commit skill and the commit
memory to the new rule.
AI-assisted-by: deepseek-v4.1-flash
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 🐛 Make execute_code wait for save before success
* 🐛 Address review feedback on waitForSave
Reject with an Error carrying a `code`: save-timeout, rolled-back,
save-retrying, or the failure code (save-failed by default). Lower the
default timeout from 90s to 60s, leaving room for the 30s layout wait
and the code under the MCP tool timeout.
When persistence is already in error at the start of the wait, keep the
stored failure code, and tell a failure Penpot keeps retrying apart from
one that needs the file reopened.
Share timeout validation with waitForLayoutUpdate through
app.plugins.utils, and make reject-not-valid reject with an Error, which
also fixes the string rejections of connectLibrary, restore, remove,
pin, export, findVersions and saveVersion.
The MCP execute_code handler now waits for the save only when the
workspace is ready, like the layout wait, so dashboard calls no longer
fail on a save error from an earlier file.
Move the waitForSave tests to frontend-tests.plugins, add tests for the
error codes, the retrying state and a failure during the wait, and add
plugin API suite and MCP handler tests. Update the docs, CHANGELOG and
the generated API files.
AI-assisted-by: claude-opus-5-5
worker_processes auto starts one worker per host CPU, ignoring the
container CPU limit, and since worker_connections went from 2048 to
65535 each worker reserves about 30 MB at startup. A frontend limited
to a few CPUs on a 64-core node needs close to 2 GB before serving a
request and gets OOM killed.
Default the worker count to the CPUs the container can use, its cpuset
and its cgroup CPU quota rounded up, when they are fewer than the CPUs
of the host, and keep auto otherwise. Make both settings configurable
through PENPOT_HTTP_SERVER_WORKER_PROCESSES and
PENPOT_HTTP_SERVER_WORKER_CONNECTIONS, and refuse to start on values
that are not valid.
Refs #13103
Signed-off-by: David Barragán Merino <david.barragan@kaleidos.net>
* 🐛 Show the error page for critical WASM errors
Since #8651 the :wasm-error handler showed the error page for :panic
only, so a :wasm-critical error (a WASM trap with no error code) fell
through to a toast and the user kept working on a broken renderer.
Only the codes in recoverable-wasm-codes now show a toast: :non-blocking
from Rust, plus :wasm-reload-context-failure and :viewer-canvas-failed,
raised by the frontend. Any other code shows the error page. This covers
errors from the render loop, event handlers, promises and store events;
errors thrown inside React effects still reach the error boundary.
The h/call docstring now describes what it does: it always throws.
Closes#12990
AI-assisted-by: claude-opus-5-5
* ✨ Add a guard for React effects that call WASM
React sends any error thrown in an effect, or in its cleanup, to the
error boundary at the app root, which replaces the whole workspace with
the error page. That holds even for recoverable WASM errors, which
should only show a toast.
errors/wrap-wasm-effect runs an effect fn and the cleanup fn it
returns. A recoverable WASM error goes to on-error (toast); any other
error is rethrown, so it still reaches the boundary. The macro
with-wasm-effect wraps mf/with-effect with it.
The macro lives in its own namespace, app.main.ui.hooks.wasm-effect:
app.main.ui.hooks cannot require app.main.errors, because the renderer
loads it and that makes a circular dependency.
AI-assisted-by: claude-opus-5-5
* 🐛 Show a toast for recoverable WASM errors in the viewport
A recoverable WASM error thrown inside a viewport effect reached the
React error boundary, which replaced the whole workspace with the
error page.
Every viewport effect that calls WASM in its body or cleanup now uses
with-wasm-effect, so a recoverable error shows a toast and the
workspace stays; panics and critical errors still show the error page.
Effects that call WASM only from a listener, ts/asap or ts/raf keep
mf/with-effect: those errors never reach React.
AI-assisted-by: claude-opus-5-5
* 🐛 Show a toast for recoverable WASM errors in the text editor
The WASM text editor calls WASM from two React effects: on mount (select
all, render) and in the first caret blink. A recoverable error there
reached the error boundary and replaced the workspace with the error
page.
Both effects now use with-wasm-effect. The mount effect also guards its
select-all and render step on its own: it opens an undo transaction
with no timeout first, and without that inner guard a recoverable error
would stop the effect before it returned the cleanup that commits the
transaction and disposes the WASM editor.
AI-assisted-by: claude-opus-5-5