AI Assist

AI Assist writes commit messages and pull request descriptions from what is on the machine: the staged patch, or the commits and diff a branch adds over its base. It uses an account of your own, ChatGPT or Grok, linked once in Settings. Exeora runs no model of its own and bills nothing for it.

Linking an account#

Each provider is linked one of two ways. Link subscription shows a code and a page on the provider's site; sign in there, enter the code, and the provider hands the gateway a token for your plan. Use an API key takes a key from the provider's platform, checks it with the provider, and keeps it. Both are stored encrypted, the same way as the token GitHub gives, and are deleted when you unlink.

The subscription flows reuse the sign-in the providers built for their own command-line tools. Neither OpenAI nor xAI documents them for a product like this, and either may change or close them without notice; if that happens, an API key still works, and the card says so. Requests made through a subscription are governed by the provider's terms for your account, not by anything Exeora can promise. A self-hosted gateway can leave the subscription flows off and offer keys only.

Provider, model and instructions#

With more than one provider linked, one is the default and the button beside a field offers the other for that time. Each operation, commit messages and pull requests, can name its own provider and model, and carry instructions of its own: conventions to follow, a language to write in. They are added to what Exeora asks for, which is a short, imperative subject and a body that says why.

Using it#

• Commit messages. The button beside the message in Source Control, once something is staged. The answer lands in the field only if the field is as it was when you asked; words you typed meanwhile stay, and a note says the generated text was set aside.
• Pull requests. The button beside the title in the Pull Request view writes the title and the description from the commits and patch against the chosen base.
• Ship changes. The whole trip from a working tree to an open pull request, described with pull requests.

A generation can be stopped with the same button. It has a minute to answer; a provider that does not, or that refuses the token, is reported in words, and a token the provider no longer accepts asks you to link again.

What leaves the machine#

What the assistant reads is read from the machine that holds the workspace, passed through the gateway to the provider you chose, and kept nowhere: the branch, the staged files with their line counts and the staged patch, or the commits with their messages, the changed files and the patch against the base, each cut at a fixed size. Every generation appears in the audit stream as ai.commit_message or ai.pull_request, with the project and workspace and never the content. The provider keeps what its own terms say it keeps.