🔌 Bring your own agent CLI, for real this time#
Settings said Azynote spawns your command for everything routed to CLI. It was true for four capabilities out of six: OnePagers and Smart Actions ignored the field entirely and ran claude no matter what you typed. Point the command at Codex, opencode, Cortex, or anything else you run, and it now drives every capability, including which model they use. One field, one command, no hidden behaviour underneath it.
🧪 Try your command before trusting it with a real session#
There is a new Test command button in Settings → AI → CLI. It runs your command once with a throwaway prompt and shows you the exact command it spawned, what came back, and how long it took. That argv line is the point: a flag with one dash too many is invisible in a text field and obvious the moment you see the tokens. Nothing is started, nothing is saved, and it takes about a second.
🔍 Failures say what actually happened#
A CLI could run, exit cleanly, return a one-line refusal, and Azynote would write a OnePager with just a title on it and report success. Now the command and its reply are in the log on every run, a model that answers without the right structure fails loudly with its own words in the error, and a clean exit with nothing in it counts as a failure rather than an empty page.
🧠 Free the model you want, keep the one you need#
The Local AI indicator already listed both models with their own memory figures and only let you release both at once. Each line is now the button that frees that model. Hand back the 14 GB chat model and keep the search model, which reloads in about a second anyway. The setting is now called Keep the chat model ready, because the chat model is all it ever covered.
💾 The local model stops quietly overshooting your memory#
llama-server was started with no cache budget, so it took its own default of 8 GB on top of the model. On a 16 GB Mac that turned a 6.3 GB model into a process that could reach 14.5 GB, and Settings was quoting you the smaller number while you decided whether you could afford to keep it loaded. The cache is now budgeted to what your machine actually has left, and the figure you see is the figure you get.
📝 OnePagers stop paying for the same work twice#
Improve a transcript by hand, generate a OnePager, and it announced "Improving transcript 1/1" and ran a second pass anyway. It was not re-improving your transcript, it was improving a hidden snapshot kept for speaker attribution, overwriting the artifact it exists to preserve. That snapshot is now left alone, and every progress line reports a decision already made instead of one it is about to reconsider.
🐛 Notable fixes#
4 fixes
- Long transcripts finish improving instead of timing out and silently falling back to the uncorrected version.
- New OnePagers reliably reach search. The local search model was rejecting its own chunks, so a OnePager could be written and never indexed.
- An unconfigured provider is no longer painted red. Not set up is not the same as broken, and only a key we checked and found invalid gets the error colour now.
- Moving between screens no longer logs errors. Leaving a screen was being reported as a failure to load it.
Happy meetings! 🎙️