Skip to main content

AI review in LosOS Git

AI review on a pull request.

LosOS Reviewer reads the pull requests of a repository in LosOS Git and leaves a review, the way a code review bot does on a hosted forge. It runs on the box, with the same language model as Build with AI, so the code never leaves the box and the review costs nothing.

It is off for every repository until someone turns it on.

Turning it on​

Open the repository in LosOS Git, then Settings, and turn on AI review. Only the repository's administrators see the switch.

The switch can't be turned on while Build with AI is off on the box (losos.builder.enable), because that setting installs the model. A new box downloads the model once, the first time it is online. Until the download finishes, reviews wait in the queue and the pull request says so.

What it does​

When a pull request is opened or reopened, LosOS Reviewer reviews all of it. When new commits are pushed to it, it reviews what those commits changed since its last review. It posts one review per run, as the account LosOS Reviewer (losos-reviewer), marked as a bot:

  • an overview of what the pull request does;
  • a table of the changed files, each with one sentence on what changed;
  • comments on single lines, where it found a problem. When one changed line fixes it, the comment carries the line as a suggestion.

It only comments. It never approves, asks for changes, pushes or merges.

You answer its comments and resolve them like anyone else's. To have it look again, pick LosOS Reviewer under Reviewers on the pull request, even after it has reviewed once. Anyone who may ask for reviews on the pull request can do this.

While it works, the pull request's checks show a LosOS Reviewer row with its progress, for example "Reviewing: 2 of 5 files done", and the number of comments when it is finished.

What it can find​

Each changed file is read in pieces of up to 80 lines. For each piece, two things happen:

  • The box checks the new lines for a short list of mistakes it can be sure of: SQL put together from values, shell variables without quotes in commands that change files, keys and tokens written into the code, a shared default list in Python, a range one step past the end in Rust, == in JavaScript and TypeScript, a React effect without a dependency list, and a file name that no longer has its folders taken off.
  • The model is asked whether the new lines have a bug. A comment from the model is kept only if it names a line the pull request adds and quotes the code on that line. Everything else it says is dropped.

How much the model adds depends on its size. The default model, Qwen2.5-Coder 1.5B, fits a box with 4 GiB of memory. It describes changes well but rarely finds a bug by itself, so on such a box most comments come from the checks. Qwen2.5-Coder 7B finds bugs of its own and needs about 8 GiB. To use it, set losos.builder.model to qwen2.5-coder:7b on the Advanced pane. Build with AI then uses it too.

A review is no proof that the code is right. Read the code yourself before you merge.

What it leaves out​

  • Files tools write and nobody reads line by line: lock files, minified and generated code, pictures and other files that are not text.
  • Deleted files and lines that are only removed.
  • Anything beyond the first 24 pieces of a large pull request. The table says which files were left out.

The box stays usable​

Reviews run one at a time, one question at a time, in the order they were asked for. A new push replaces a review that hasn't finished for the same pull request, and turning the switch off drops the reviews still waiting. On two cores, a piece takes about 10 to 20 seconds, so a pull request of a few files is reviewed in a few minutes. A build in Build with AI and a review share the model, and each waits for the other's current question.

LosOS Git can reach only addresses on the box itself. The reviewer does not use the mesh's shared computers, because that would send your code to other people's boxes.