What it is
oss reads any repository through the GitHub API — no clone, any language, any
forge. That boundary is deliberate and it is why the tool generalises.
The cost is two questions it cannot answer alone:
- Does this actually run? Reading a diff cannot settle it.
- Have I worked this out before? Nothing keeps that for you.
Those are extensions. A runner executes something real; a memory remembers. Each is a directory containing one JSON file, run as a child process — so it can be bash, Python, Go, or a binary somebody handed you.
Nothing is required
Section titled “Nothing is required”| With | You get |
|---|---|
Just oss |
Facts from the API: diffs, commits, CI, review threads, conventions |
| + a runner | Execution against real applications |
| + a memory | Markdown filed by topic, and keyword search over it |
| + a local model | Verdicts, and search by meaning |
| + a cloud key | Escalation past the local budget |
The storing half needs no AI whatsoever — filing, the topic map and the digests are deterministic. Only finding by meaning needs a model.
The rule that does not bend
Section titled “The rule that does not bend”Reading a public repository is free to get wrong. Writing to one is not: a comment reaches every watcher and the mailing list the instant it is sent, and deleting it afterwards reaches neither.
So every outward write is refused by default, everywhere, and there is no setting that changes it. See Writing upstream.