This fixes a bug in the caching of received message headers at various places in the code which are later used for reply or iopub messages. The existing code on main branch assumes that the shell processes a single request at a time so that the cached header is valid until that request is replied to, but now of course there can be multiple subshells in different threads executing code concurrently.

The solution is to cache a different header per subshell, or to be more precise a different header per thread. This is accomplished using a ContextVar which is essentially thread-local storage that also supports asynchronous execution. We are already using a ContextVar at

I found this whilst debugging use of ipympl in jupyterlab, trying to create multiple matplotlib plots using the option of one subshell per comm-target (the default option). The best way to view this bug is to use kernelspy as it shows that in this scenerio some messages (e.g. some idle status messages) have the wrong parent header and are therefore displayed in the wrong section of the kernelspy output.

All new parent_header attributes are private (leading underscore). Where they were previously public I have added a public parent_header getter to extract the appropriate value from the ContextVar so that downstream libraries that have been reading the parent_header in the main shelll should have the same functionality as before.