v12.7.3 - Drop-In
A release about being answered. A local provider that would not talk, tools that reported more than they did, and a call you can now interrupt in the middle of a word.
Fixed
- LM Studio answers again. The provider card said no API key was needed and every message came back asking for one. The guard in front of the chat kept its own list of the providers that need none, and had never been told about LM Studio. That list is now written once and asked from everywhere.
- The LM Studio card's Refresh Models button exists. Its own hint told you to press it; the button was never drawn, and the model list behind it was refused for want of a key it never needed. A running LM Studio now shows up in the model picker, in model search, on the Agents page and in the provider grid, and an agent pinned to it stays pinned.
- A group chat can be pointed at your own OpenAI-compatible endpoint, and the group-chat skill runs on a local model with no key at all.
- A failed Windows update leaves a trail. Every step of the update path is written down with a timestamp, and the last lines of it ride along in Settings > Advanced > Diagnostics. The app also stops quitting on a timer while the installer is still starting: it waits for the installer to actually be there, and says so when it is not.
- A failed send is explained the way a failed test is: the mailbox that refuses your password says which credential it wants and where to get it, on every mail action rather than only when you press Test.
- A mailbox that cannot be opened no longer answers under the name you asked for, and a list that was shortened says that it was shortened. Emptying the trash reports what it deleted, not what it found.
- A calendar that could not be read says so instead of shortening your week, and a week with no calendar connected no longer reads as a free one.
- Speech, images and video stop claiming more than they did. A voice that fell through to a free fallback says which provider refused and why; text that was too long to speak says it was cut; a video job that failed says it failed instead of going quiet after "started"; and a provider that answers 401, 402 or 429 is named as your account rather than as a fault in the app.
- There was a second copy of the speech cascade, and the two had drifted far enough that the same settings produced two different voices. There is one now.
- A file too large for the phone is refused before the connection breaks. The ceiling was set by an estimate that counted one round of encoding; it is now the measured number, written down once for both file paths.
Changed
- Talking over Call Mode interrupts her mid-word. Call Mode used to stop the audio only after the sentence had finished. It now listens through playback with the same calibrated listener the Iris window uses, and hands the turn back to you as soon as you start speaking. Where the microphone cannot be opened for it, the old behaviour stays and the screen says why.
- Iris uses your name the way a person does: once at the start of a conversation, and after that only when she is actually addressing you.
- Qwen 3.7 gets a profile of its own: the thinking-mode sampling the vendor publishes instead of the generic Qwen profile's non-thinking numbers, and a long-thinking budget so a silent stretch is not cut off as a stalled stream.
- The guide has its first screenshots, and the LM Studio and KoboldCpp setup pages have caught up with the provider grid that replaced the old dropdown.
- The acknowledgment that fills the silence follows where your endpoint actually points, not what it is labelled. A custom endpoint on the internet is treated as being on the internet; one that cannot be read at all stays silent, as before.