You already let AI review your contributors' pull requests and draft fixes for your issues, and you still re-check every line it gives you.
I maintain Apache SkyWalking BanyanDB, the database behind Apache SkyWalking, and it gets a steady stream of contributor pull requests and issues. I started using AI for both: a bot to review incoming pull requests, a coding agent to fix issues. The time I hoped to save went into checking its work. A review comment could be right or could be noise, and I could not tell without reading the code myself. An agent would say an issue was fixed, and I had to find out on my own whether a test had ever failed before its change, or whether it had quietly changed things it was not asked to touch. I was doing the review twice. What I was missing was AI output that comes with proof I can check in a minute, instead of claims I have to redo.
GraphLoom runs coding agents (Claude Code, Codex or Pi, with your own logins, on your machine) through a fixed process, and it only hands you a result once the evidence for it has been checked.
Pull request review you can act on. pr-review reads an incoming pull request and first asks whether the change is needed at all: is the problem real on the base branch, and does the fix fit your project's design principles? Then three reviewers look at correctness, maintainability and security. Every finding is checked by a second, independent reviewer on a different agent before it is drafted; a finding it refutes is dismissed, not posted as a finding. You get one review comment whose first line is the verdict, with each finding pointing at the lines it is about.
Issue fixes you can merge on evidence. issue-fix takes an issue and plans the fix. It writes a regression test and proves the test fails on the current code, makes the fix, proves the same test now passes, and has an independent reviewer check the change. You get a local branch with a pull-request description, not a claim that it is done. If the issue leaves a product decision open, the run stops and asks you instead of guessing.
Runs you can leave alone. Each run works in its own Git worktree with fixed budgets for agent calls, jobs and time. When something fails, a recovery agent investigates and retries within those limits. graphloom watch tells you about progress, a status every 15 minutes, and anything that needs you.
What it does not do: it does not push branches or post reviews unless you enable it, it does not approve or merge pull requests, and it does not run as a hosted service.
Both recordings are real runs, shortened: the commands and their output are quoted from the run's state and the pull requests it produced (long lines are cut with "…"; the times on the right are real).
pr-review on apache/skywalking-banyandb#1460: one run, the stop because posting was not yet enabled, the resume, and the review it posted ("Changes requested: 2 open", two inline comments).
issue-fix on apache/skywalking#14169 (BanyanDB): a 9 minute 52 second run, the regression test failing on the base (RED), passing with the fix (GREEN), an independent review, and the branch that became pull request #1459.
With reduced-motion enabled the recordings show their final frame.
- Install. Releases bundle their own Node. You need Linux, an authenticated
gh, and a logged-in Claude Code, Codex or Pi. The installer checks the release's signature against the signers file from the release tag.Thegh release download v0.2.0 -R hanahmily/graphloom -p install.sh curl -fsSLO https://raw.githubusercontent.com/hanahmily/graphloom/v0.2.0/release/allowed_signers sh install.sh --require-signature --allowed-signers allowed_signers graphloom doctor
graphloomcommand is installed in~/.local/bin; put that directory on yourPATH. Later versions arrive throughgraphloom upgradeand a daily automatic check. - Start a review or a fix. In a working directory of your own, link the graph and create the task:
mkdir -p plugins && ln -s "$(graphloom version --root)/examples/plugins/pr-review" plugins/pr-review graphloom task new pr-review <owner/repo#N> --repo <your clone> --profile <profile.json> graphloom run reviews/<repo>-<N>/task.yaml
<profile.json>is a path, or a name looked up inprofiles/beside you and then in the bundle; onlybanyandbships forpr-review, so for another project copy it and adjust it. For an issue, write atask.yamlwithgraph: issue-fixandrequest: "Fix <owner/repo#N>."(seeexamples/issue-fix-graphloom-task.yaml) and aconfig:profile copied fromplugins/issue-fix/profiles/. - See the result.
graphloom watch <run>follows it. The review is in the run's worktree under.graphloom/review/(review.html, and the draft indraft.md;graphloom inspect <run>prints the path); the fix is a local branchfix/<issue>.
Or drive all of this from Claude Code, Codex or opencode with the GraphLoom skills: graphloom skills install, then ask it to run issue-fix or pr-review.
Does it post to my pull requests by itself? Only if you give it a token for writing (github.writeTokenEnv in the runner config). Without one, the run stops with the draft ready for you to read.
Will it approve pull requests? No. It posts a comment review; approving stays yours.
How it works inside: docs/overview.md.
GraphLoom is licensed under the Apache License 2.0; see NOTICE.