Efficient tool calling, the Empryo Chrome extension, and one Logs view

Thank you, smakosh

smakosh (smakosh.com) sponsored Empryo's Apple Developer membership, which is what signing and notarizing the macOS app requires. Find smakosh at github.com/smakosh and x.com/smakosh. Thank you!

Efficient tool calling

• Tools you rarely need stay out of each request until the agent calls for one: computer use until it first drives, and orchestration, council, routines and the advisor outside their own mode. That is about 6,700 fewer tokens on every request.
• Looking up or suggesting a tool no longer changes the tool list, so the prompt cache keeps hitting through the session.
• The agent reaches any held-back tool in one step, in every mode, and Stop cancels it.
• Code scripts can call any tool in bulk: a batch of reads, searches or agent launches runs as one step instead of one call each.
• Code and explore scripts work on every model.
• Choose how tools load (auto, provider, Empryo or all) and switch code scripts on or off in the desktop Router or /agent-features in the TUI.

Computer use and the Empryo Chrome extension

• Good computer use on every model you run. With a TypeSafe key, the model plans and jev picks each click and keystroke, in about a third of a second and for a fraction of a cent per step.
• The Empryo Chrome extension is on the Chrome Web Store. It lets the agent work in your own Chrome tabs. Install it in one click from the desktop setup screen, or with /computer extension install.
• Choose who decides each step: jev on its own, jev together with the model, or the model alone. The desktop app has a picker for it.
• Computer use prefers the drivers that work best on your machine.
• Desktop apps can be driven through Cua.
• The permissions panel lists what you have allowed before, with Forget and Forget all.

What's new

• Logs brings errors, model calls and compactions into one place, with search and filters. Open it with /logs (or /errors) in the TUI, or from the desktop app. Open a row to see its details.
• The Model calls tab lists every request with its model, reasoning effort, tokens and timing.
• Each agent shows its reasoning effort next to its model, in the TUI, the desktop app and headless output.
• When a model runs out of context in the middle of a task, the conversation is compacted and the task keeps going.
• /router shows where each model choice comes from (this tab, the project or your global settings) and warns when another setting overrides the one you are editing.
• Compaction settings let you choose which model writes the summary and when it runs.
• The model picker can group models by subscription plan (Claude, ChatGPT, Grok…) and gateway models by vendor, each under its brand icon. On the desktop app every group is a folder you can collapse. Turn it on with Group at the bottom of the desktop picker or Ctrl+G in the TUI; the choice is shared between both.

Quality of life

• The desktop app starts lighter: other languages, icons and Morph blocks load the first time you need them.
• The TUI and the desktop app stay responsive with many background agents running.
• When the agent changes a setting for you, it checks the config file and tells you if the change did not stick.
• Turning on YOLO for a project you have not trusted points you to /trust first.
• When the status line runs out of room, the model, branch and changes move to the top bar.
• The cache chip shows how long the prompt cache has left. Its popover keeps to the countdown and the Warm cache now button; the details sit under About cache lifetime.
• The subscription chip in the desktop status bar opens your subscription accounts.
• Collapsing a turn's tools in the desktop chat is a single icon, with its label in the tooltip.
• The composer drops the Forge v2 indicator, since there is only one engine now.
• /compact tells you when it is queued behind a running reply.
• The agent inspector shows the full prompt, what each tool touched and the model's thinking.
• Durations read naturally in every language, from seconds up to years.
• Switching language in the TUI confirms in the language you picked.
• When the system blocks a notification, the desktop app tells you why.
• Typing a model name, such as an advisor tier, picks the right model. If the name matches several, you get the list.

Bug fixes

• The TUI follows a terminal resize that happens right after launch.
• Plan mode keeps every tool available and tells the agent to wait for your approval before changing files.
• Saving a setting for one project keeps your global settings.
• A broken project config.json is reported once, and Empryo keeps running.
• Refusals from a provider show up in the TUI, the desktop app and headless output.
• Conversations that used one provider's web search continue after you switch models.
• A reply that stalls mid-stream recovers on its own.
• A failed model switch on a desktop tab keeps the previous model and says why.
• A slow keychain is not mistaken for a missing key.
• Computer use works in your own Chrome tabs, not only in the browser Empryo manages.
• Scripts and screenshots in the built-in browser stop at a time limit.
• The hide button in the desktop model picker appears only where hiding takes effect.
• The computer-use options for who decides the next click keep their descriptions inside each card.
• Plus hundreds of smaller fixes across providers, computer use, Windows, Linux, containers and translations.