Since code review is the bottleneck now, fks2g helps developers decide how closely to review code.
Its for the devs who have already tried this method of reviewing code:
And for devs who have realized that this code review strategy leads to a finger pointing situation when bugs or bad architecture gets shipped:
To inform how closely to review a code change, the CLI collects:
- cosine similarity between file-name embeddings and configurable project text sources
- an LLM judgment about whether the closest files are likely to change soon based on source documents
- recent bug-fix commits classified by an LLM
- file change frequency from git history
- an LLM final risk assessment based on the collected evidence
Usage
OPENAI_API_KEY=<KEY> npx fks2g analyze --repo ../react --github-repo facebook/react --model gpt-5.5 $(git -C ../react show --name-only --format='' | tr '\n' ' ')
This analyzes the files from the most recent commit of the react repo if you've got it cloned on your machine.
Here's a run w/ gpt 5.5:
## packages/react-server-dom-webpack/src/server/ReactFlightDOMServerNode.js:medium
Risk reason: Rare historical changes and no concrete source-document signal, but recent bug-fix activity and broad RSC relevance make changes moderately risky.
Change frequency: Rare (12 historical changes)
Recent bug fixes: b91823e: [FlightReply] Don't drop FormData entries in `decodeReplyFromBusboy` (#36468)
Source signal: not likely (RSC-related issue is a broad desired behavior, while concrete bugs target other subsystems.)
Source references: [#36491 Bug:](https://github.com/facebook/react/issues/36491), [#36430 Bug: [19.2\] DEV-build logComponentRender throws SecurityError on cross-origin Window props](https://github.com/facebook/react/issues/36430), [#36440 Bug: ReactDOM.preloadModule crashes with repeated custom as value](https://github.com/facebook/react/issues/36440), [#36497 Bug:](https://github.com/facebook/react/issues/36497), [#36379 Bug: react-refresh leaks FiberRootNodes from secondary renderers](https://github.com/facebook/react/issues/36379)
If the CLI is run against a PR in a github action environment, it'll attempt to post the anaylsis as a comment on the PR.
Benchmarking risk ratings
Use benchmark with the same options as analyze, plus a commit range. For each commit in the range, fks2g runs a point-in-time risk analysis using only git change frequency, recent bug-fix counts, and GitHub issue context available as of that commit. It then checks the following 30 days of commits to see whether each assessed file changed or was touched by a bug-fix commit. Low, medium, and high risk ratings are scored as 0.33, 0.66, and 0.99 probabilities, and the command reports the mean squared error against the observed outcomes.
OPENAI_API_KEY=<KEY> npx fks2g benchmark --repo ../react --github-repo facebook/react --commit-range main~20..main --format markdown
You can also use --from-commit and --to-commit instead of --commit-range; --forecast-days defaults to 30. If no files are passed, each benchmark step assesses the files changed by that commit.
Docs
There's a --help flag.
Warning: 0.X software
This software is as ready for prime-time usage as its name suggests. I've only kicked the tires on it a bit with openai models. Theoretically, google, anthropic, and bedrock models are also supported.