Paul Stack

I haven't written code in months (probably for the best). Five of us build swamp, none of us write code manually. The job got bigger, not smaller. The code was never the job — the understanding was. New post up: stack72.dev/the-code-was-n…

An agent can open a PR in thirty seconds that takes an hour to review. Multiply that by every repo with an issue tracker and you have the current state of open source. We banned external code contributions and here's why. stack72.dev/the-community-…

@ccollicutt Delighted to hear that it’s given you some ideas :) thanks for the kind words

@stack72 This is really great stuff - I am reading through all of this, I think you all are really onto something here, thanks for the posts! I'm working to bring this adversarial thinking into my projects!

4,236 unit tests passed yet the binary was broken. A separate UAT suite running against the real compiled artifact caught it in four minutes. Three recent bugs, same pattern: all unit tests green, bug at an integration seam, no user ever saw it. stack72.dev/the-gate-betwe…

@fiunchinho great question! I found that by having bash and then gh pr * in the ask, I wanted to be specific that gh pr checks was allowed - it's just about trying to be very specific over what's in the ask

@stack72 Thanks for the great post! one question: isn't `Bash(gh pr checks:*)` already covered by just `Bash`? why allow both?

I sorted every agent tool into two buckets: can git checkout undo it, or does it touch something external? Eliminated 90% of permission prompts. Kept the ones that matter like gh pr create. stack72.dev/agent-trust-is…

@sudomateo we wrote some libraries to help us out for this - github.com/systeminit/swa…

@stack72 This is perfect! Thank you kindly!

@sudomateo Some of the extension tests we have that are like testing the extensions we create, can have mocking to asset behaviour against any API - you can see an example here github.com/systeminit/swa…

@stack72 Sweet. That's what I figured when reading but it wasn't explicitly spelled out so I wanted to confirm.

@sudomateo we do actually make those requests - in this specific set of tests (that ensure that things work correctly) there are no mocks. They are making real api calls to whatever we need to test. So swamp itself is actually running. commands

@stack72 That's basically what I'm asking. I'm unfamiliar with the subcommands that swamp offers but if there's something like swamp create foo that requires reaching out to some foo API, how does that work?

@sudomateo what type of dependencies?

@stack72 How do commands that require actual dependencies work during testing? Are you mocking anything or actually setting up the dependencies?

Agents open PRs in our repo. So the diff is untrusted input flowing into models with merge authority. That's a recipe for disaster Multiple AI reviews, each scoped to a different concern. Minimal tool sets and a security review that audits changes to the CI pipeline itself — because the reviewers are defined in workflow files too. stack72.dev/anatomy-of-a-s…

@jedgar You too buddy!!

@stack72 @SwiftOnSecurity Hey Paul, it's been a minute! Hope you're well. :)

Sad news. My friend Simon who did the original bird for Twitter, Octocat for GitHub and Redid Sammy The Shark for DigitalOcean has passed away, he was 56. He was a great guy. What a loss.

Vibe coding works great for ten PRs*. Then you do it two hundred more times and the codebase quietly stops making sense. The fix isn't better prompts. It's giving agents the same structural context a senior engineer carries in their head. Build the machine. The vibes don't scale. stack72.dev/the-vibes-dont… * The number may vary depending on the weather and the timezone ;)

The thing that decides whether an AI agent helps you or talks past you is context — the knowledge it loads before it does anything In swamp we package that as skills, and we test them in CI the same way we test code. Tessl for quality, promptfoo for routing, weekly regression across four models stack72.dev/skills-are-con…

@rhoml yeah and the pressure on GitHub infra at this team means it's going to get worse

@stack72 Man GHA is such drag… the worst part is that many are still moving out of worst tools like teamcity into gha just to find that they have to leave gha 😂

GitHub issues aren't a good fit for an agentic process, so we stopped using them. Not because GitHub is bad, but because the workflow we actually run every day doesn't belong in a comment thread. Plan versions, adversarial findings, and feedback rounds all became comments, and it got hard to tell what had actually been decided. So we moved the lifecycle into swamp itself — state machine, versioned plans, checks that won't let the agent skip ahead. stack72.dev/the-lifecycle-…

@rhoml yeah the first phase was issues - now we are planning GHA next :)

@stack72 Issues and PRs are pretty chaotic these days

That little voice behind engineers head that says "don't trust AI" is holding a whole industry back. Remember there was a time where people didn't trust inteliJ or any IDE auto-complete capabilities, this is no different.

@neuroserve They say it's unsupported but works fine at this time

@stack72 Is sqlite still supported with current Ghost versions?

Deployed Ghost on DigitalOcean with swamp — load balancer, Litestream SQLite replication, Certbot, DNSimple, all of it. The whole stack written by Claude using swamp as the guardrails. One workflow run to deploy, one to update. stack72.dev/deploying-ghos…