* 🐛 Log errors readably in MCP execute_code console output
The console captured during execute_code rendered every object argument
with JSON.stringify. An Error has no enumerable fields, so logging a
caught error printed "{}" and the agent lost the message. A circular
object or a BigInt made JSON.stringify throw, so a console.log call
aborted the executed code with an unrelated serialization error.
Format console arguments in a small pure module: errors by their string
form (name and message), other objects as indented JSON, falling back to
the string form when JSON cannot serialize the value.
AI-assisted-by: claude-opus-5.5
Signed-off-by: breken-ai <312387581+breken-ai@users.noreply.github.com>
* ♻️ Keep console formatting within ExecuteCodeTaskConsole
---------
Signed-off-by: breken-ai <312387581+breken-ai@users.noreply.github.com>
Co-authored-by: breken-ai <312387581+breken-ai@users.noreply.github.com>
Link the MCP plugin to plugins/libs/plugin-types instead of the
published 1.5.0 package, which lacks PenpotMgmt. The repo types
declare the penpotMgmt global, so the inline import in index.d.ts
goes away and the plugin checks against the API it runs with.
Type Context.management as PenpotMgmt without on/off, so new
management methods need no extra listing.
Mark the background plugin mock in load-plugin.spec.ts as
NonNullable: createPlugin can return undefined, and the build
type-check failed on it.
Penpot only exposes penpotMgmt to plugins running with global
scope. Without it, the MCP plugin failed on load when registering
its workspacechange listener.
When penpotMgmt is missing, the plugin now builds the workspace
metadata from the open file and skips the listener. execute_code
receives penpotMgmt as null and waits for layout updates as before.
The tool description and instructions tell agents that only the
open file is available when penpotMgmt is null.
Global plugin lifecycle:
- Add optional manifest scope "global" for plugins that remain active
across dashboard and workspace navigation within one browser tab.
Existing plugins retain workspace scope by default.
- Preserve global sandboxes when a workspace ends or another plugin
loads, including when that other plugin fails to initialize.
- Avoid duplicate instances of an already-running global plugin.
- Close global plugins on explicit closure or logout, with idempotent
cleanup of listeners, timers and UI resources. Global scope does not
imply automatic startup or persistence across a browser reload.
Plugin API and navigation:
- Keep a permanently live penpot facade whose getters follow the current
workspace, while user, theme and UI facilities remain available on
the dashboard.
- Expose penpotMgmt to global plugins with workspace metadata,
none/loading/ready status and workspacechange subscriptions.
- Add awaitable openFile(fileId, {teamId}) navigation in the current tab,
defaulting to the current team and resolving only once the requested
file and its active page are ready.
- Reject invalid identifiers, failed or superseded navigation and waits
exceeding 30 seconds; release navigation watchers on completion.
- Guard workspace-bound operations while no workspace is ready and
return null from root, currentFile and currentPage in that state.
- Require content:read for the initial management API. Callers must
reacquire file, page and shape references after changing files.
MCP lifecycle and connection controls:
- Make MCP a global plugin and initialize its integrated controller per
authenticated application session instead of per workspace.
- Load the enabled integration with a valid token on the dashboard as
well as in a file, retaining its existing hidden-iframe implementation.
- Keep enabling and connecting separate: an enabled plugin starts idle,
and each tab connects explicitly through dashboard or workspace
controls.
- Preserve the iframe, WebSocket, heartbeat and reconnect machinery
during navigation; disconnect on request and close on disable/logout.
MCP sessions and execution:
- Derive the short MCP session ID solely from currentUser.sessionId,
excluding the file ID. Retain the SHA-256-based, 50-bit Base32 encoding
to produce a stable, copyable 10-character identifier.
- Allow sessions without an open file and update workspace metadata over
the existing connection without replacing its routing identity.
- Refresh context after the connection handshake so initialization does
not leave stale file metadata.
- Expose penpotMgmt to execute_code, preserve execution storage across
navigation and skip layout waits when no workspace is ready.
- Update API types, tool guidance, documentation and the changelog.
Validation:
- Cover authenticated startup, workspace survival, logout cleanup,
readiness, navigation failure, duplicate loads and session routing.
- Verify that short IDs remain stable across files and reconnects and
that disconnect cancels pending asynchronous ID initialization.
- Pass 16 focused frontend tests, 73 runtime tests, 30 MCP plugin tests,
builds, TypeScript checks and frontend/runtime lint.
- Pass all 73 MCP server tests with --test-force-exit; the normal runner
still stays alive after assertions finish.
- Verify a live dashboard-to-file-to-dashboard round trip without a
document or iframe reload, preserving facade identity and storage.
AI-assisted-by: gpt-6
* 🎉 Add multi-file session routing to the MCP server
Session routing:
- Track each user's Penpot connections by session ID.
- Dispatch explicit sessions directly; otherwise discover the sole session.
- Return a retryable error when discovery fails or is incomplete.
Plugin and tools:
- Initialize connections with a fresh session ID and file metadata.
- Share the optional sessionId schema and place it last in tool inputs.
Validation:
- 84 tests, server type checking, and MCP formatting checks pass.
- Disable test logging to avoid the logging worker shutdown hang.
AI-assisted-by: gpt-6
* 🎉 Support independent MCP connections
- Require explicit connection intent for each integrated workspace.
- Derive short session IDs from the Penpot app instance and file.
- Show and copy session IDs in the MCP menu and standalone plugin UI.
- Stop stale callbacks and reconnect attempts after disconnect.
Validate with frontend and plugin tests, type checking, live multi-tab
checks, and standalone UI checks with a simulated connection. Server
tests pass with --test-force-exit; the normal runner can hang on shutdown.
AI-assisted-by: gpt-6
* 📚 Document multi-file MCP sessions
- Explain independent connections, session selection, and copying IDs.
- Describe session ID lifetimes for integrated and standalone plugins.
- Correct browser focus, connection, and recovery guidance.
Validate with the documentation site build and diff checks.
AI-assisted-by: gpt-6
* ✨ Change behavior on session ID duplication
* ✨ Open MCP toolbar menu on hover and match menu style
---------
Co-authored-by: alonso.torres <alonso.torres@kaleidos.net>
Co-authored-by: elhombretecla <delacruzgarciajuan@gmail.com>
Extract the WebSocket close-code decision into a pure ReconnectPolicy
module and cover it with node:test. The plugin now names the 1008
policy-violation code instead of using a magic number, and the terminal
behavior of a 1008 close is pinned by a regression test so a future
refactor cannot reintroduce the reconnect storm.
Document the contract in mem:mcp/core: 1008 is terminal, it comes from a
duplicate user token or a missing token in multi-user mode, and recovery
is manual via "Connect here".
Closes#11510
AI-assisted-by: deepseek-v4.1-flash
The MCP server closes duplicate plugin connections with 1008 (policy
violation). The plugin treated every close as retryable: shouldReconnect
stayed true, scheduleReconnect re-armed, and each brief onopen reset the
backoff - looping at ~1s until the browser froze. Treat 1008 closes as
terminal and stay disconnected until the user reconnects explicitly.
* MCP server: Update instruction manual to point to method instead of sleep
* MCP plugin: Call waitForLayoutUpdate prior to every code execution
* Composable test suite: Use waitForLayoutUpdate instead of 200ms delay in tests
* ✨ Bound the size of plugin task responses
When using the integrated remote MCP server, bound response size.
All responses are passed to LLMs, which themselves impose bounds.
This is a measure to bound memory usage in the centrally provided
MCP server.
GitHub #9493
* ✨ Bound parallelism in ExportShapeTool
Use an integer semaphore to bound parallel requests to this
memory-intensive tool, thus bounding memory usage.
GitHub #9493
* ✨ Add (manual) integration test script for ExportShapeTool parallelism
Add dependency tsx to facilitate executions.
GitHub #9493
* ✨ Make number of parallel export requests configurable in ExportShapeTool
Use env var PENPOT_MCP_EXPORT_SHAPE_MAX_PARALLEL_REQUESTS to configure
the maximum number of requests in multi-user mode (default 0, no limit).
Resolves#9420 (critical memory usage issue in PROD deployment)
When the plugin's ExecuteCodeTaskHandler returns a Uint8Array (e.g. from penpotUtils.exportImage),
JSON.stringify previously serialized it as an object with numeric string keys,
causing ~10x payload expansion and large peak heap usage on the server side.
The plugin now wraps a top-level Uint8Array result in a tagged envelope
{ __type: "base64", data: <base64> }, and ImageContent.byteData decodes this envelope
on the server. The legacy numeric-keyed-object path is retained as a fallback for
compatibility with older plugin builds.
This is a temporary workaround for penpot/penpot-mcp#27.
It adds a wait time before exports via the export_shape tool to account
for asynchronous updates in Penpot, increasing the likelihood of exporting
the fully updated state.
If the MCP version (as given in mcp/package.json) does not match
the Penpot version (as given by penpot.version), display a warning
message in the plugin UI.
This is important for users running the local MCP server, as it
is a common failure mode to combine the MCP server with an
incompatible Penpot version.
* ✨ Add core changes for mcp server
* ✨ Changes to plugins-runtime to add mcp extensions
* ✨ Changes to MCP plugin
* ✨ Changes post-review and ci fixes
This matches the behaviour of findShape, more closely aligning with
the LLM's expectations (given the lack of concrete information in
the instructions)
