← sources

Sandboxed Claude agents wrote 159 of my 188 PRs

source: reddit (reddit:ClaudeAI)published 2026-10-20 02:37ingested 2026-10-21 02:36

retained text

I'd been playing with Matt Pocock's [sandcastle](https://github.com/mattpocock/sandcastle), which gives you the pieces to build your own agent loops. I liked it, but I kept wanting something with fewer choices: one loop, already decided, that I could point at a repo and walk away from.

So I built **bunker**. You put a `bunker:queued` label on a GitHub issue, and it:

- runs Claude Code on it in its own git worktree and Docker container
- has a second Claude session review the change and send it back if it's not right
- runs the repo's tests in the sandbox with no network
- opens a PR and turns on GitHub auto-merge, so your required CI check decides whether it lands

Anything touching CI, its own config or agent instructions waits for a human. Only maintainers can queue work.

It's been building itself. About 160 of its first 190 PRs were its own, and an issue costs around $1.30 of Sonnet time on average. This week I pointed it at a few of my other repos and found a pile of wrong assumptions on bunker's side. For example, fine-grained GitHub tokens can't read check runs at all, so its "fix the red CI" step was blind until I changed how it reads CI.

It's v0.11 (and moving towards v1), Claude only, Linux/macOS with Docker. MIT.

Repo: https://github.com/dbggr/bunker · Docs: https://bunker.true.engineer

If you try it, I'd really like to hear where it breaks.

document fetched 2026-10-21 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xagj87/sandboxed_claude_agents_wrote_159_of_my_188_prs