Release v10.4.0

• (@GermanBluefox) The endpoint of an AI request comes out of the configuration and no longer out of the request. An address in the message used to win over the stored one, which was the comfortable way to put a proxy in front of OpenAI - and at the same time the way to have this instance carry the stored key, as the Authorization of that request, to an address of somebody else's choosing. Whoever may send a message to this instance could do that. A proxy belongs in the configuration now, where it is entered once by whoever may configure the adapter. The editor never sent an address, so nothing changes for it; the Test buttons of the settings dialog keep trying what stands in the form, with one limit: an address out of the form counts only together with a key out of the form - one's own key to one's own endpoint gives nothing away. In the credential-store mode the saved address is used, so an endpoint of one's own has to be saved before it can be tested there
• (@GermanBluefox) Only an entry that was stored as an AI credential is read from the central credential store. The store holds the secrets of the whole system - a database password, the login of a camera - and a request for "the key of this provider" could name any one of them
• (@GermanBluefox) The execute message - which runs a script here, with the whole ioBroker API behind it - checks who asked for it. A message never said on whose behalf it came, so there was nothing to check, and everyone who may send one to this instance could have code executed. Where the js-controller names the user (7.2.5 and newer, which put the user of a socket connection into the message), that user needs the right that a command on the host needs. Where it does not name one - an older controller, a script, an adapter - nothing changes
• (@GermanBluefox) The question before a debug session ("The script will be stopped and must be activated manually after debugging") started the debugging even when it was answered with "Cancel" (#2382)
• (@GermanBluefox) Stopping a debug session could end the debugged process with "uncaught exception: write EPIPE": the inspector process exited at once, although the process it debugs writes its whole log through its pipes and was still shutting down. The inspector now stops that process first and waits till it is really gone, a broken pipe cannot end a debugged script anymore, and the reset of the connection to the debugged process is not reported as an internal error of the inspector (#2382)
• (@GermanBluefox) The AI chat stayed empty when a request took longer than 30 seconds: the answer is pushed as an instance message now instead of waiting for the socket callback
• (@GermanBluefox) An AI answer without content now says why (stop reason, token counts) instead of leaving an empty chat bubble, and every failed request is logged
• (@GermanBluefox) An answer cut off at the output limit is marked as incomplete
• (@GermanBluefox) Sorted the scripts in the @ list of the AI chat by their path
• (@GermanBluefox) The AI chat can be maximized to the whole editor area (#2389)
• (@GermanBluefox) The AI chat shows a spinner for the whole request, not only until the first tool call
• (@GermanBluefox) Fixed an accepted AI suggestion not reaching the visible editor
• (@GermanBluefox) Fixed the splitters jumping back to their previous width while being dragged
• (@GermanBluefox) Blockly: fixed a parameter of a "JavaScript function" with its default name making the script unsavable
• (@GermanBluefox) Added a script history: saved versions can be listed, compared and restored in the editor. "Saved versions per script" (30 by default, 0 switches it off)
• (@GermanBluefox) Added a "Script history" tab with the used storage per script and buttons to delete single, orphaned or all histories
• (@GermanBluefox) Moved the credentials table to its own tab
• (@GermanBluefox) Mirror: a changed file no longer resets enabled and engine of the script (#2396)