The panel could show the Spark and it could not launch the session I wanted, because a second model server was up on the box and the add form did not offer it. I had the launch built, then looked at the form and changed it. I pick the folder and the title, then where to open it, including a plain terminal when the model does not have the tools yet. A launch that pretends the model can do the job is worse than a terminal I can see.
The form lists models from what the box writes down. It knew one server and not the other, so the second server was a single hardcoded name. I had the collector ask the second server and write the answer in the same file. The panel still makes no call of its own.
I also want a local session to be able to ask for an end-to-end test without holding a subscription account. It would hand over the address, the code, and the cases it already has, and the test hub would run them and write a report back. That is a plan. The first sketch was a shared folder. I have not settled the handoff.
On the egress page, 26 shell commands were skipped because they were too long to classify. That cannot stand. A long command is not a way past the check. I put that next, ahead of the prettier views.
Live agents on the dashboard are the other open piece. I used to want every agent to report in through a hook. I do not need that now. The tools write their own records. What I want is who is doing what, light enough to leave on, and off when the panel is not in front of me. Function first. A chart that looks busy and lies is the failure I already had with the lock file.
I also ran another pass on the voice skill. More of my own meeting transcripts, only the ones that label me as the speaker. Notes that do not label the speaker stayed out, because I cannot tell my sentences from anyone else's. The skill should sound like the measurement. It should not sound like a summary of me.