- Status
- Open for votes
- Cost
- L, more than a week
- Date
- Would become
- A four-step epic: spec and mock, shadow mode on the Hermes plugin, Jev routing behind a flag, an on-device fallback (YUI-41). Cards in the backlog, Chris picks.
- Votes
- Counting
Yui's call
Recommend, as a small step first. Build the spec and the mock, then run Jev in shadow mode and let real turns say whether it picks the right shape before a hint reaches anyone.
The problem
Today the agent's model reads a long guide on every turn and decides what the screen looks like. It picks one line or a deck, a card or a full screen, a map or none, a camera or none. It gets it wrong in ways you can see. A yes or no gets a deck. Chris, 2026-09-26: "eight screens of basically nothing just to tell me that I'm up-to-date". A status question gets a wall of text.
The native agents do it with hand-written patterns. "Plan my week" works. "Could you sort my week out" misses.
We found about twenty decisions like this on one turn. Most are small, closed questions on a short message. A big model is a slow, pricey way to answer them.
Who it is for
Anyone who talks to an agent in Yui and wants the answer in the right shape the first time. One line when a line is enough. A map when they asked where. The camera when the agent needs a photo.
How it works
Jev is a new kind of model from TypeSafe AI. It cannot write text. You give it a message and a list of typed questions. It answers each with a choice or a yes/no probability, plus how sure it is. Then the agent writes into the shape Jev picked.
This is one extra step on every call to an agent:
- The message arrives. The plugin cuts it to 400 characters and adds a few facts: the last reply shape, whether a photo came with it, the agent's tool names.
- One Jev call asks up to five questions at once: the reply shape (one line, yes/no, card, pages, full screen), how many things there are to read, whether it needs a map, whether it needs the camera, and which camera (plain, meal, document, barcode, tuner mic).
- Each answer has a confidence. Above the line (0.80 for a choice, 0.85 or 0.15 for a yes/no, to be tuned) it becomes one plain hint line. Under the line it says nothing.
- Under the line, or if Jev is down or slower than 600 ms, the agent reads the channel guide and decides, exactly as it does today. Jev never blocks a turn.
- On long replies only, a second call checks the draft against the shape it should have had. At first it only writes a log line.
- Later, the same questions go to more places: the photo intent on a meal turn, and the tool router that catches "could you sort my week out".
What is real and what is a claim. TypeSafe says: 70 to 500 ms, $0.042 per million input tokens, output free, and answers that are "calibrated", so 0.9 means right about nine times in ten. Those are TypeSafe's claims. We have not run a call. We have no key.
What we can work out from those claims:
| Per turn | Per 1,000 turns | |
|---|---|---|
| Shape call, about 800 tokens in | about $0.00003 | about 3 cents |
| Check call on a long reply, about 1,500 tokens | about $0.00006 | about 6 cents |
| Both, worst case | under $0.0001 | under 10 cents |
| Added wait, shape call | 70 to 500 ms claimed | runs beside the work the plugin already does |
Cost is not the limit. Privacy and trust are.
Pros
- Fixes the complaint Chris keeps making: a deck for a yes/no, a wall of text for a status.
- Fast and nearly free next to a big model, if TypeSafe's claims hold.
- A choice from a fixed list cannot be a broken screen. Jev can pick the wrong option, but not one outside the list.
- Never blocks: under the confidence line, Yui works as it does today.
- The log gives us labelled data. We learn how often the model picks the wrong shape.
- One place to tune reply shape, instead of a longer guide.
Cons
- The person's words leave the phone for a third party. Today they go only to their own agent.
- Text only. Jev cannot see a photo or a map, so camera and photo decisions come from the words.
- A days-old model, in early access, with a price that can change.
- One more moving part on every turn, and one more thing to watch.
- Calibration is TypeSafe's claim. Nobody outside has proven it.
Cost
L. The hint plumbing in the plugin is a few days. The real work is the eval set (about 100 labelled turns), the shadow run and the tuning. Steps, each a backlog card:
- Spec and playground mock. Write the questions and the hint line down, and mock the diagram and the hint in the playground. No key, no call. Card: SITE, "Jev layer spec and playground mock". S.
- Shadow mode on the Hermes plugin. Jev decides and logs. The model still decides. We compare on real turns and score Jev against the eval set. Needs the key and a yes on question 1 below. Card: YUI, "Jev shadow mode in the Hermes plugin". M.
- Jev routes for real, behind a flag. The hint goes into the turn, owner turns only, with a switch in Settings. The post check stays log only. Card: YUI, "Jev shape hint behind a flag". M.
- An on-device fallback. The easy calls (shape, camera) run on the phone with the small model from YUI-41, so no words leave the phone. Jev stays for the rest. Card: YUI-41 (exists), extended. L.
Chris picks which of these become work, and in what order.
Risks
- Private text goes to a third party. TypeSafe says it will not train on inputs. It keeps them "as long as reasonably necessary", with no fixed period. Mitigation: 400 characters, no thread, owner turns only, a switch in Settings.
- Shared agents carry someone else's words. Start with owner turns only.
- Wrong hints. A wrong hint could make a worse screen. Mitigation: the model can overrule, two clashing hints are dropped, and shadow mode measures before any hint reaches a person.
- Vendor risk. Early access, an outage at launch, price unknown. Every row falls back to today's behavior, so the loss is the gain, not the app.
- Claims not yet checked. Speed, price and calibration are TypeSafe's. We watch p95 latency and the wrong-shape rate in shadow mode, and stop if either is bad.
Open questions
- Can a person's message leave the phone for Jev? (a) Yes, owner turns only, 400 characters, behind a switch in Settings. (b) No, Jev only on our own test turns. (c) Only for agents the person marks. Yui would start with (a).
- Who makes the Jev account and key? It is a browser sign-up at console.typesafe.ai, so it is Chris's step. Nothing is built against a real key until then. The early-access waitlist was reported closed on 2026-09-27, but that is one third-party report.
- Where does the call run? On the Hermes plugin (one place, easy to log), in the app (works with any agent, the phone holds a key), or through a proxy on yuigui.com. Yui would start on the plugin.
- Groups: wait for the on-device model or use Jev? On-device keeps the words on the phone. Jev works on every phone. Yui would wait.
Credits and trail
- Proposed by
- Chris
Should Yui build this?
Two buttons. No login, no email.
Counting