June
June: I submitted 316 AI-generated PRs to open source
Last verified:
What is June?
June is a tool that Speedrunning Open Source is a live case study and pipeline built around the idea of using AI‑driven bots—called clankers—to submit large volumes of pull requests across open‑source repositories, then observing how maintainers respond. The project documents how projects react to AI‑generated contributions, what signals they use to detect low‑effort or robot‑like PRs, and how they filter out noise. The core tool behind this is a pipeline called sweep that automates bug fixes, PR generation, and submission to GitHub while also learning from the patterns of rejections and merges.
One of the key outputs of this project is a GitHub Action script that maintainers can plug into their repos to automatically flag and, in some cases, close or ban AI‑like PRs. The script encodes real‑world heuristics gathered from observed rejections, such as em‑dash patterns, low‑content descriptions, wrong branches, missing tests, and excessive velocity across an org. The idea is to give projects a cheap, self‑hosted defense that can be run in CI and requires only a small configuration file and optionally an LLM key for richer checks.
The tool is aimed at two audiences: open‑source maintainers who want to protect their attention from AI‑generated spam while still accepting genuine contributions, and AI‑assisted contributors who want to understand how their workflows look to humans and how to avoid being lumped in with clankers. The methodology also explicitly frames the problem as ‘anti‑slop, not anti‑AI’: the goal is to raise the quality bar for all contributions, regardless of whether they are written by humans or models, and to systematize the filtering that maintainers are already doing manually.
June pricing
Pricing model: Freemium
The tooling and methodology described in Speedrunning Open Source are available as open‑source scripts and pipelines, with no separate paid plans or proprietary tiers. The GitHub Action filter script is free to use and can be run in any GitHub repository’s CI, optionally enhanced with an LLM API key if the maintainer wants description‑quality checks. The associated pipeline (sweep) is also open source and can be self‑hosted; there is no mention of a hosted or service‑based pricing model, so the project currently operates as a no‑cost, MIT‑style endeavor.
June pros
- Generates real bug‑fix PRs tied to existing issues
- Produces large‑scale data on how maintainers react to AI PRs
- Provides concrete heuristics for detecting low‑effort contributions
- Delivers a compact GitHub Action script (250 lines) that can be dropped into any repo
- Runs purely in CI without external hosted services
- Supports optional LLM scoring for description depth and tone
- Offers org‑wide three‑strike bans to reduce cross‑repo spam
- Paces submissions with a drip queue to avoid flooding single orgs
- Teaches evasion tactics like lineup matching and em‑dash stripping
- Captures behavioral patterns (velocity, resubmission, description style) that are hard to fake
- Makes maintainer heuristics explicit and automatable
- Serves as a learning resource for contributors on how to write better PRs
- Helps maintainers avoid spending hours closing AI‑style PRs by hand
- Encourages structured fix patterns such as tests‑touched whenever source files change
- Shortens the feedback loop between AI‑driven pipelines and real‑world review behavior
June cons
- Can be perceived as aggressive or spammy by maintainers
- May temporarily block or ban associated accounts on multiple projects
- Focuses on GitHub‑specific workflows instead of broader ecosystems
- Relies heavily on manual heuristics that may need updating as norms evolve
- Does not prevent human‑written low‑effort PRs, only patterns observed so far
- Requires some tooling setup and understanding of GitHub Actions
- Honeypot / impossibility‑issue ideas are not yet shipped or tested at scale
- No built‑in UI or dashboard for non‑technical maintainers
Frequently asked questions about June
What is 'Speedrunning Open Source' actually measuring?
It measures how open‑source maintainers respond to AI‑driven pull requests submitted at scale, including how quickly they close, whether they comment, and which behavioral or textual signals they use to flag contributions as low‑effort or AI‑like. The project tracks closure rates, merge rates, and the exact wording of rejection comments across many repos to reverse‑engineer the heuristics that maintainers are using silently.
What is the 'sweep' pipeline?
The sweep pipeline is the underlying automation that finds issues, generates fixes with AI models, and submits PRs to GitHub repositories. It is the engine used to run the live experiments described in the post, and it learns from each rejection and merge to refine patterns such as description style, pacing, and test coverage so that future submissions look more like high‑quality human‑driven PRs.
How does the GitHub Action filter work?
The GitHub Action script runs in CI and checks each PR for a set of signals such as em dashes in titles, very short descriptions, missing tests when source files change, wrong base branches, high velocity across an org, and known AI‑policy or 'AI‑sloppiness' indicators. If warnings fire and the author has a low track record or no prior merges, it can post a warning or even close and ban the PR, depending on configuration.
Does the project require an LLM API key?
An LLM API key is optional. The core action can run on bash and grep alone, using structural heuristics. However, if maintainers want the script to judge description depth or tone, they can wire in an LLM key to evaluate whether a PR explanation reads as high‑effort or not.
How does the three‑strike ban mechanism operate?
The three‑strike ban is org‑wide: if an author has three PRs closed with a gate comment in the same GitHub organization, the script flags that user for an automatic ban across all repos under that org. This prevents the same account from repeatedly spamming the same maintainer group.
What are the 'honeypot' issues mentioned in the post?
The honeypot concept involves creating intentionally impossible‑to‑fix issues that a clanker would blindly try to solve, such as a fix that contradicts an architectural invariant or breaks existing tests by design. When a bot submits a PR against such an issue, it can be instantly filtered or banned, but humans are expected to recognize the impossibility and walk away.
Who is this project designed for?
The project is designed for open‑source maintainers who want to protect their time from AI‑driven spam and for AI‑assisted contributors who want to understand how their workflows are perceived and how to avoid being labeled as low‑effort. It also serves as a case study for the broader community on how AI‑PRs and human‑driven filters coevolve.
Is the project anti‑AI or anti‑slop?
The project explicitly frames itself as anti‑slop, not anti‑AI. It aims to filter out low‑effort, noisy contributions regardless of whether they are written by humans or models, and to raise the quality bar for all PRs. The heuristics focus on signs of low effort such as sparse descriptions, missing tests, and high velocity, rather than trying to prove AI authorship.
Can I run this pipeline on my own accounts?
Yes. The author describes how to run the sweep pipeline on your own GitHub identity, using the same prompts and setup that powered the experiments. Running it yourself will propagate the same patterns and heuristics, so it is recommended primarily for actually fixing bugs and learning from the feedback, not for vanity metrics.
Are there any costs to using the tools?
The action script and pipeline documentation are free and open source. The only potential cost is if you choose to hook in an LLM API key for richer analysis of PR descriptions, in which case you pay the model provider’s standard usage fees. There is no mention of a hosted or commercial product layer with its own pricing.