That seems to work best (plan with the expensive model, cheap ones implement). The "Implemented feature XX, tests pass" summary drove me nuts too, so I make the planner record the decision and the why in my second brain app before anything gets built, and implementers report back against it.

Not really an orchestrator, but my second brain notes & tasks app Vist (https://usevist.dev) has a memories system built in. You connect it via MCP, and define the projects you want to keep track of yourself. Your agents automatically update the memory/project state when needed, and this works nicely for me in a similar setup with separate repos for the Vist app, the Desktop apps, public releases, and so on.