v1.5.1 beta

What's new

The app checks the agent's work itself

Until now a step of the agent's plan closed on the agent's word unless it had run tests. Now the app looks at the result before a step closes, for every kind of task:

• Code: the project's tests, build or linter must have run after the last change, and passed. A lot of new code in a project with tests asks for tests too. A server the agent started must answer; one that answers with an error keeps the step open.
• Documents: a document is read back after it was changed, or the app shows the agent the document as it now is (pages, words, headings, the start). At the end of the plan a separate check weighs the documents against what you asked for; it sees only your request and the files.
• Files: what a command created, moved, changed and removed is noted; files that were removed are named, so nothing disappears unnoticed.
• Sites: a page of your project is opened apart from the agent's browser at phone and desktop width, in the light and the dark theme, with its console errors.
• Research: links the agent cites must have been opened, and quotes must be found word for word on the pages it opened, in the documents and in the final answer.
• At the top of the chat each step shows what it was closed on; a step closed only on the agent's word has a hollow tick.

Memory that even small models keep

• After every reply the chat's own model looks at the exchange apart and keeps the facts in the chat's private record (with this computer's time) and in the memory shared by all chats: who you are, how you like answers, budgets, dates, decisions. Small models such as GPT-6 Luna used to save almost nothing by themselves.
• Facts come from the exchange alone: what you asked to forget does not come back, and a request that came to nothing is not written down.

First start

• Welcome cards after the opening: what the agent can do for different people, with the mini chat showing it at work.
• The app speaks your computer's language on the very first start (English, Russian, Italian or French; English for any other). If you have used the app before, nothing changes.

Good to know

• One small extra request goes to your model after each reply to keep the facts; it is skipped for short messages such as "ok" or "thanks".
• The new checks were tested on live models (GPT-6 Luna and DeepSeek): the whole set of the app's real-model checks was run, and every failure was compared with 1.5.0.
• Everything listed under "What does not work, or works differently" in v1.4.5 still holds.

Downloads

• Windows: OpenGhost-1.5.1-Setup.exe (installs over the previous version; chats and settings stay; from 1.5.0 the Update button brings it)
• macOS: OpenGhost-1.5.1-mac.dmg (Apple silicon and Intel)
• Linux: OpenGhost-1.5.1-linux.tar.gz