YOLO mode, queued follow-ups, 22 sign-in lanes, and a prompt cache that stops crying wolf

Empryo 3.8.4-beta · Sep 17, 2026

Release artwork & highlights

3.8.4-betaSep 17, 2026current

Security & permissions

YOLO mode: One switch that lets the agent work without stopping to ask. Permission gates accumulate rather than reset, so turning it off puts every guard back exactly where it was. Auto mode now answers "ask" gates on its own instead of waiting for you.
The computer-use permission prompt names the task: It used to show a literal {reason} placeholder. It now says what the agent is about to do.
Your pointer is yours: Moving your own mouse no longer pauses a running task. Stopping the agent is something you do on purpose, in Empryo.

Chat & composer

Queue a follow-up: Press Alt+Q to park a whole prompt behind the running turn. It sends itself when the turn finishes. Enter still steers the turn you are in.
One home for mid-turn prose: The model's thinking-aloud text has its own row and its own toggle, separate from reasoning. No more reply text re-typing itself while the tool rail works.
The solo tab shows its name: With one tab open, the label centres and glides, and hovering reveals the tab id.

Sign-in & providers

22 sign-in lanes: Kilo Gateway (device code) and the Z.AI Coding Plan join Claude, ChatGPT, Copilot, Cursor, Devin, Grok, Gemini CLI, Antigravity, Kimi, Fire Pass, OpenRouter, Ollama Cloud and LLM Gateway. Alibaba's coding and token plans and Xiaomi MiMo take a pasted key, and Empryo finds your region on sign-in so you are never asked which one you bought.
Eight more providers by key: Z.AI, Zhipu, Alibaba (coding and token plans), Xiaomi, Kilo, SiliconFlow and Synthetic, each with live models, context windows and prices from the catalog.
Reasoning effort, everywhere it belongs: The max tier now reaches every per-provider setting, not just the composer control. A level you set is clamped to what your model actually publishes, so asking for more than a model takes no longer fails the turn.
Devin models work again: Empryo was asking every Devin model for 64,000 output tokens, which the upstream model refused outright on anything with a smaller ceiling. It now asks for what each model accepts, and sends Gemini the schema dialect it requires.
A new sign-in page: The page your browser lands on after signing in wears the mote, says what happened, and offers a button instead of closing itself behind your back.

Computer use

One ordered list of lanes: Empryo's own driver, then jev, then whatever Codex left on your disk. Drag to reorder. A disabled server no longer takes tools with it.
My Chrome in the browser panel: The panel follows whichever tab the agent is driving, and close_tab only ever closes a tab the agent opened.
Windows mote overlay: The glowing pointer that shows where the agent is working now draws on Windows too.
Dialogs answer in 250 ms: A page frozen by a JavaScript alert used to cost a 30-second timeout. Observing it now returns the dialog immediately.
Four checks, one setup surface: A switch, a headline and three destinations, the same on desktop and in the terminal.

Prompt cache

It stops crying wolf: Swapping the tool belt and swapping back reported three "cache broke" warnings inside one turn, on a tab that was 99% cached. Returning to a prefix Empryo already sent is now recognised as a revisit, not a break.
String content hashes honestly: A one-block string was compared as if it were 25 characters shorter every step, which reported an edit that never happened.

Fixes & improvements

Forge no longer forces a tool choice on a Claude lane, which could leave a turn without a final message.
A relative HOME can no longer leak into config and PATH: the desktop resolves your home directory through the OS account database.
OpenRouter's stateless sign-in callback is no longer rejected as a forgery.
Headless runs close their computer-use tabs when the turn ends.
The --events stream reports every prefix verdict the cache witness can give, instead of a copied list that went stale.
The design-system linter reads the website's stylesheets, so 1,415 real classes stopped being reported as dead.

Documentation

Reasoning effort has a page: The eight levels, which models take which, and every per-provider key.
Plain english throughout: The docs no longer say "belt", "rung", "blast radius" or "wall-clock", and the language count matches the grammars that actually ship (37).