the impulse was to build. i listed three apps i needed and the agent fired subagents at all of them at once. the course correction arrived in one sentence: kill the subagents and read first. so everything stopped, and reading what was already on disk was the actual news. four android repos, all doing the same assistant job.
the current one is minimal on purpose, typed input only, the server url stored as a preference with zero network code behind it, media and overlay and a router as the real working parts. the previous generation app has the richest plumbing, an api client, a wake word, a widget, and got superseded a day ago. another one is an accessibility driver with hardened loop guards and a green test suite. and the oldest was replaced twice over. none of them is exactly the app i described, but one is closest and the other three are proven code libraries wearing app-shaped folders.
so the decision became an archive, not a build. keep the current one, port the loop guards and the accessibility bits into it, keep the superseded one installed as a reference until parity, archive the rest. the website audit was the same story. the logo fetcher service already has everything on my wishlist, random fetch, vector search, a color tester, a company check, an api. the only real gaps were in the wakeword studio, multi word wake words and prosody samples.
the rule going forward: read before delegating, count the overlaps before building, and when four things do one job, one becomes the app and the rest become references.