v1.5.0 beta

What's new

An agent for long, serious work

• Every chat has a memory of its own. A private record the agent keeps by itself: decisions, agreements, things not to do, facts it found. No other chat can read it, not even one working in the same folder; the chat's own mini chat can. The app stamps each entry with the date and time.
• Simpler, sturdier compaction. A new summary is built from the last summary and the messages after it, not from the whole chat again. Its size follows the model instead of a fixed 8,000 tokens, and the two latest exchanges are kept word for word.
• The agent's plan. A task of several parts is laid out in steps, shown in a black island at the top of the chat, like on an iPhone. The agent cannot finish while steps are left open; a long proof under a step opens with a click.
• Reconnecting. A short provider failure (an overloaded server, a network blink) no longer ends the work: the app asks again by itself, with a quiet "Reconnecting" beside the ghost.
• Agents in one folder stay out of each other's way. Builds and git take turns, a file one agent is changing is guarded from the other, servers get different ports.
• Git only when it belongs. Outside an existing repository the agent leaves git alone, and it commits only when asked.
• Pictures while it works. The agent sends what matters as it goes, without stopping: a product photo from the built-in browser at full size, part of the screen or all of it, a picture by its address, an image file.
• Pictures in a chat with a password are encrypted on disk with that chat's own key.

See what the agent is doing

• One line beside the ghost: "Reading…", "Running…", "Editing…". The next step rolls in over the last like a drum, a wave of light runs along the words, and steps never pile up in the chat.
• Code edits open for a moment under the line: "+3 −2", the old lines turning over to the new ones, then the card folds softly away.
• Approvals appear the way the agent's questions do: the message field turns into the card.

Providers

• Instant for an OpenAI-compatible service: answers without thinking. The idea came from #35 by @lohrynn, thank you.

Around the app

• An Update button by the globe: the new version downloads by itself, and you see how much has arrived.
• Narrow windows: the sidebar and the browser step aside for the chat and open over it; drag the sidebar's edge to resize it (it remembers), or all the way to fold it like Finder; the plan's island shrinks to a dot.
• The model menu: the model's name sits in the button, flies into the list and back; the whole window blurs, without a jerk at the end.
• Two-line chat rows: the title, then the time and the start of the last message, with the chat's latest photo on the right.
• The usage panel is made of the same glass as the settings, 220 px wide; neither turns clear over the sidebar any more.
• The composer: labels line up, the round buttons never squeeze, a long model name in the quick chat ends with "…".
• Copy and paste with the mouse: Copy in the menu over selected words, and a right-click menu (Cut, Copy, Paste, Select all) in every field, API keys included.
• Buttons, labels and panels no longer get selected like a web page; messages and fields still do.
• The browser: a site wider than the panel fits it whole; the tab's spinner turns only while the page loads.
• Diagrams: no stray asterisks of bold text in labels.

Good to know

• This is the first version with the Update button. From 1.4.9 or older, install 1.5.0 by hand once; later versions come through the button.
• Everything listed under "What does not work, or works differently" in v1.4.5 still holds.

Downloads

• Windows: OpenGhost-1.5.0-Setup.exe (installs over the previous version; chats and settings stay)
• macOS: OpenGhost-1.5.0-mac.dmg (Apple silicon and Intel)
• Linux: OpenGhost-1.5.0-linux.tar.gz