The multi-input component did not handle paste events for
comma-separated values. When users pasted emails like
'qa@example.com, test@example.com', the entire string was
inserted as-is, triggering validation errors.
The on-key-down handler already split text on commas/spaces
when typing, but paste events bypassed this logic.
Added an on-paste handler that:
- Detects if pasted text contains commas or whitespace
- Splits the text by commas and/or whitespace
- Validates each part individually
- Adds valid items to the items list
- Prevents default paste behavior
- Resets input state after processing
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* ♻️ Add mcp integration state management refactor (#10226)
* ♻️ Add mcp integration state management refactor
* 🐛 Fix access tokens do not appear
* ♻️ Refactor some names
* ♻️ Refactor token deletion
---------
Co-authored-by: Luis de Dios <luis.dedios@kaleidos.net>
* 🐛 Fix stale MCP token data after create/regenerate (#10280)
Fix the root cause in profile.cljs: remove the optimistic conj from
access-token-created and instead chain a fetch-access-tokens after the
create-access-token API call succeeds. This ensures all callers get a
fresh, server-consistent token list automatically.
Suggested-by: niwinz
Signed-off-by: kapilvus <kapil69265@gmail.com>
Co-authored-by: kapilvus <kapilvus@gmail.com>
* ✨ Remove non-recoverable mcp key warning from regenerated modal (#10298)
---------
Signed-off-by: kapilvus <kapil69265@gmail.com>
Co-authored-by: Luis de Dios <luis.dedios@kaleidos.net>
Co-authored-by: kapil971390 <kapil69265@gmail.com>
Co-authored-by: kapilvus <kapilvus@gmail.com>
Read PENPOT_TENANT env var (defaulting to "default") and embed it in
Redis Pub/Sub channel names as penpot.mcp.<tenant>.task.{req,res}.<id>.
This prevents cross-tenant interference when multiple environments share
a Redis instance, matching the backend convention
(e.g. penpot.rlimit.<tenant>.window.<name> in app.rpc.rlimit).
Co-authored-by: deepseek-v4-flash <deepseek-v4-flash@penpot.app>
* ✨ Batch multiple thumbnail deletions into a single RPC call
Replace the old per-object immediate thumbnail deletion with a
debounced batched approach. The frontend queues object-ids in state
and waits 200ms before sending a single RPC request with up to 200
object-ids. The backend deletes all matching thumbnails in one SQL
statement with a single RETURNING clause, then touches the affected
media objects.
This reduces RPC overhead when rapidly clearing thumbnails (e.g.
navigating pages) and makes deletions more efficient.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 📎 Fix missing issues
---------
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* ✨ Add MCP connection badge to the workspace toolbar
* ✨ Add MCP status button with single-tab connection control
* ♻️ Extract component for MCP indicator in the toolbar
* ♻️ Some improvements
---------
Co-authored-by: Luis de Dios <luis.dedios@kaleidos.net>
* 🐛 Fix race condition between MCP init and plugin runtime
Add promise-based synchronization to ensure MCP initialization waits
for plugin runtime to be ready before calling global.ɵloadPlugin.
- Add runtime-ready-promise in app.plugins that resolves when
init-plugins-runtime completes
- Add wait-for-runtime function for other modules to await readiness
- MCP init now waits for runtime via rx/from before starting plugin
- Add defensive guards in start-plugin!, load-plugin!, close-plugin!
to check if plugin APIs exist before calling
- Rename init-plugins-runtime! to init-plugins-runtime
Fixes: global.ɵloadPlugin is not a function error when MCP plugin
starts before async plugin runtime initialization completes.
* 📎 Add 'create-pr' opencode skill
The current Nginx example doesn't include the required Websocket settings to correctly proxy /mcp/ws, and leads to WebSocket connection error when trying to connect in a design. Adding this lines fixed the issue.
Signed-off-by: Justin Lin <30039756+lancatlin@users.noreply.github.com>
The rasterizer's create-image function was clearing image.src in its
Rx teardown cleanup. This caused the decoded pixel data to be discarded
before downstream operators (drawImage / createImageBitmap) could read
it, resulting in a browser NotReadableError.
Changes:
- Remove image.src = "" from cleanup; the image element will be
garbage collected naturally. Event handler nulling is kept to break
circular references.
- Add dimension validation in svg-get-adjusted-size to return nil for
zero/NaN dimensions instead of producing invalid sizes.
- Add fallback in svg-set-intrinsic-size! to use [max max] when SVG
dimensions can't be determined.
Error occurred in production (2.16.0-RC10) during thumbnail generation
in the workspace.
Signed-off-by: Andrey Antukh <niwi@niwi.nz>
* 🐛 Filter ignorable exceptions in error-boundary onError callback
The global uncaught-error-handler already skips NotFoundError/removeChild
and other harmless errors, but react-error-boundary's onError callback fires
independently of the window.onerror pipeline. This means the error boundary
was still logging these errors and setting last-exception, causing them to
continue appearing in error reports despite being non-actionable.
Add the is-ignorable-exception? check to the error-boundary* onError so
harmless errors are silently ignored, matching the behavior of the global
handler.
* 🐛 Fix dangerouslySetInnerHTML anti-pattern in context-notification
The previous code used dangerouslySetInnerHTML on the same element that
could also contain React children. This is a React anti-pattern that can
cause reconciliation mismatches and lead to removeChild DOMExceptions.
Refactor to use two separate element branches: one for raw HTML injection
and one for normal React children with links.
