Refolk
November 16, 2026·9 min read

GitHub's PR Caps Trust Merges Per Repo. Only 129 Profiles Say Maintainer

GitHub now exempts contributors from PR caps by merged PRs per repo. Here's how recruiters can source on that signal instead of green squares and opened PRs.

github sourcing recruitershow to evaluate github contributions candidatesfind open source contributors to hireai generated pull requests githubgithub pull request limits
GitHub's PR Caps Trust Merges Per Repo. Only 129 Profiles Say Maintainer

If you source engineers on GitHub, the contribution graph and the opened-PR count are now the least reliable numbers on a profile, because coding agents inflate them faster than anything a maintainer has to approve. The signal that still holds is the one maintainers use themselves: merged PRs per repo, in repos the candidate doesn't own.

On Oct 28, 2026, GitHub let maintainers write that trust rule into repo settings. A contributor now skips the pull request cap if their account is old enough or if they've had enough PRs merged in that specific repository. That second rule is the most useful sourcing filter GitHub has shipped in years.

What did GitHub change on Oct 28, 2026?

GitHub now lets maintainers exempt contributors from pull request limits based on account age or on the number of PRs merged in that repository. A contributor who meets at least one enabled rule gets past the cap.

GitHub pull request limits are repo-level moderation settings that restrict or cap pull requests from people without write access. They arrived in three steps:

  1. February 2026: maintainers could disable pull requests entirely or restrict them to collaborators.
  2. June 17, 2026: GitHub added a limit on how many PRs a user without write permissions can have open at once in a repo. That same month, GitHub's Community PM said bypass rules would follow, "based on trust signals like account age, number of merged PRs in the repository, and organization status," with issue assignment as an additional signal.
  3. Oct 28, 2026: the exemptions shipped. Maintainers can "exempt contributors who have had at least the number of pull requests you set merged in that repository." The settings are under Moderation > Interaction limits.

Some projects were already doing this by hand. Apache Arrow ran a workflow that closes PRs from contributors without repository access once they have more than 3 open at the same time. An Apache infrastructure ticket now proposes setting the new exemptions in config, using 365 days of account age and 2 merged PRs as an example. It's a proposal, not an adopted policy, but it gives you a concrete threshold to start from.

The detail that matters for sourcing is scope. GitHub's docs say a contributor's merged pull requests "are counted separately for each repository." A merge is one maintainer's decision about one codebase, and GitHub doesn't carry it over to other repos. Recruiters shouldn't either.

Why are contribution graphs worth less every month?

Contribution graphs track commits and pushes, and coding agents inflate those faster than anything else. From September 2025 to September 2026, commits on GitHub grew 6x while merged pull requests grew 3.9x.

6x
Growth in GitHub commits, Sep 2025 to Sep 2026
Merged pull requests grew 3.9x over the same period, so raw activity outran merges by about 1.5x.

Every activity layer grew faster than merges:

Metric, Sep 2025 to Sep 2026Growth
Commits6x
Actions jobsabout 4.8x
Pushesabout 4.5x
Merged pull requestsabout 3.9x

The gap comes from cost. An agent can produce commits, pushes and Actions jobs at almost no cost to the person running it. A merge in someone else's repo still needs a human with write access to read the change and approve it. The more output agents generate, the less each green square says about the person.

Octoverse 2026 shows how fast that changed. AI agents started about 17% of pull requests in March and more than 40% by September, roughly 110 million agent-started PRs in September alone. Opened-PR counts track output and nothing else. In April 2026, one engineer running an autonomous coding agent submitted more than 130 pull requests across multiple repositories in 72 hours. A recruiter filtering on "PRs opened" would have ranked that account near the top.

Why does a global merged-PR count still mislead?

A total merged-PR count mixes self-merges in a candidate's own repos with real endorsements from outside maintainers, and agents have inflated the first kind. Only merges in repos the candidate doesn't control work as references.

The Octoverse numbers show the effect. By September, 90% of agent-started pull requests merged within two weeks, up from 83% in February. PRs without a detected agent stayed around 86%. Much of that agent volume is developers merging their own agents' output in repositories they control, where no outside maintainer has to approve anything.

GitHub's exemption sidesteps this by counting merges per repository. A PR you merged into your own side project took one click. A PR merged into a repository you don't control means an outside maintainer read the change and accepted it.

A merge in your own repo is a habit. A merge in someone else's repo is a reference.

How to evaluate GitHub contributions candidates have made

Rank each signal by how hard it is to fake. Merged PRs in repos the candidate doesn't own go at the top. Contribution graphs and opened-PR totals go at the bottom.

SignalWhat it actually measuresInflated by agents?Use it as
Contribution graphCommits and pushesHeavily (commits up 6x)Ignore
Opened PR countOutput volumeHeavily (130 PRs in 72 hours)Ignore
Global merged PRsMerges anywhere, including own reposYes (agent PRs merge at 90%)Context only
Merged PRs per external repoA maintainer's acceptanceLeastPrimary filter
Fast-closed, unmerged PRsRejected drive-bysPattern of slopNegative filter

A workflow that holds up when agents are involved:

  1. Pick the repos the role depends on. If you're hiring for a data platform team, that could be Apache Arrow and two or three projects next to it, not "anyone active on GitHub."
  2. Exclude repos the candidate controls. Drop their personal account and any org where they hold write access. Those merges don't count as endorsements.
  3. Count merged PRs per repo, not in total. Use a floor in the same range as the Apache proposal's 2 merged PRs. Three merges in one demanding repo tell you more than thirty spread across throwaway projects.
  4. Check the spread over time. A few merges per repo spread across months fit the pattern the exemption rules reward. A burst of PRs in a single week doesn't.
  5. Run the negative filter. Look for unmerged PRs closed within days across many unrelated repos (more on this below).
  6. Read one review thread. Open a merged PR and see how the candidate handled pushback from the maintainer. That tells you more than any count.

The weak points are steps 2 and 3. Doing this by hand means listing every PR a handle has opened, removing their own repos, grouping merges by repository and checking timestamps, and then working out who the handle belongs to. For one candidate it takes an afternoon. For a shortlist of twenty it takes most of a sprint, and the role stays open the whole time.

Why can't profile headlines find open source maintainers?

Profile headlines can't find open source maintainers because almost none of them write it there. In Refolk's index of professional profiles, only 129 US profiles and 49 German profiles have "open source maintainer" in the headline.

129
US profiles with "open source maintainer" in the headline
Germany has 49. Many maintainers carry generic titles like "Software Engineer" or "Principal Engineer".

The wider "open source contributor" phrase does a bit better, at 698 US profiles, still a small share of the people who actually contribute.

SegmentCountSource
"Open source maintainer" in headline, US129Refolk's index
"Open source maintainer" in headline, Germany49Refolk's index
"Open source contributor" in headline, US698Refolk's index
US vs. Germany self-described maintainers2.6xDerived (129/49)
US contributors per self-described maintainer5.4:1Derived (698/129)

These are headline keyword matches, so they undercount real maintainers by a lot. That undercount is the point. The people who do label themselves mostly work at vendors and foundations. In the US sample they're at Okta/Auth0, Docker, Microsoft, AWS, CNCF, Dynatrace and Octopus Deploy, with titles like "VP OSS and Kubernetes" and "Lead Maintainer AboutCode and Scancode." In Germany they're at Datadog, IBM (Linux/KVM on Z), HubSpot and YaCy.net, clustered in Karlsruhe, Berlin and Munich. Search the headline and you get the people whose employer already knows what they do. Everyone else who contributes doesn't show up.

So the search has to run in the opposite direction from the usual one. Start from the merge history in the repos you care about, then work out who each handle is, rather than starting from a profile and hoping it links to a GitHub account. That's the search Refolk is built for: you describe the contributor in plain English and get back named people instead of a list of handles.

Use time-to-close as a cheap negative filter

Pull requests that maintainers flag as low-quality AI work are rarely merged and are typically closed within 3 days. A trail of fast-closed, unmerged PRs across many repos is the cheapest red flag to check.

That pattern comes from an arXiv study of PRs that maintainers flagged as slop: they are mostly additive, rarely merged, and typically closed within 3 days. The authors disclosed AI use in only 14% of cases, so you can't count on candidates to label agent-generated work. Look at the pattern, not at disclosures.

Maintainers are dealing with a lot of it. In a Linux Foundation survey led by Hilary Carter, SVP of Research, maintainers estimated that a quarter of incoming pull requests were low-quality "AI slop." On large projects the estimate rises to 31%. Daniel Stenberg's curl project canceled its bug bounty programs because they became magnets for low-effort AI submissions, and it's now moving low- and medium-severity flaws public.

PR limits exist to protect maintainer attention. A candidate whose history shows dozens of PRs opened and quickly closed across unrelated repos has already been turned away by several maintainers. You'd be the next one to review them.

The signals to check:

  • Many unrelated repos, few merges: activity spread wide with almost nothing accepted.
  • Closed in days, not weeks: matches the 3-day pattern for flagged slop.
  • Additive-only diffs: changes that mostly add code and rarely touch existing logic.
  • Bursts: dozens of PRs in a single short window, the 130-in-72-hours shape.

FAQ

Do GitHub's pull request limits apply to every repository?

No. They're moderation settings each maintainer chooses to configure under Moderation > Interaction limits. The Oct 28, 2026 exemptions are optional too: a maintainer can enable an account-age rule, a merged-PR rule, or both, and a contributor gets past the cap by meeting at least one. Outsiders can't see how a given repo is configured, which is another reason to read merge history directly.

Should I avoid candidates who use AI coding agents?

No. By September, 90% of agent-started pull requests merged within two weeks, so agent use on its own says nothing bad about a candidate. What matters is where the merges land. A candidate who uses agents and still gets repeated merges in repositories run by other people is showing the judgment you want to hire for. The warning sign is volume with no outside merges.

How many merged PRs should count as a trusted contributor?

There's no universal number, and GitHub leaves it to each maintainer. The Apache infrastructure proposal uses 2 merged PRs and 365 days of account age as an example, which is a reasonable floor for sourcing. Always apply it per repository and only to repos the candidate doesn't control. Two merges in a demanding project beat dozens in projects the candidate owns.

Can recruiters search GitHub bypass lists directly?

No. Exemption rules are moderation settings outsiders can't see, so nobody outside Apache Arrow can tell who is exempt there. What you can search is the public evidence those lists are built from: merged PRs per repository, account age, and issue assignment, which GitHub's Community PM named as an additional trust signal in June. Set expectations with hiring managers accordingly. Promise "people Arrow maintainers have merged code from at least twice over several months," not "people on the Arrow bypass list." You can back up that claim, and it likely describes much the same group.

The data behind this

Try it on the search you came here for

Stop building boolean strings. Just describe the person.

Type one sentence. I plan the search, read GitHub, public LinkedIn and Crunchbase records, and the open web as it is right now, and hand back a ranked list with the reason next to every name.

  1. 01Describe them

    One plain sentence. Role, city, stack, stage, whatever matters to you.

  2. 02I read the web live

    GitHub, public LinkedIn and Crunchbase records, the open web. Not a database that went stale last quarter.

  3. 03You read the shortlist

    Ranked, with the reasoning under every name. Open a profile, ask a follow-up, narrow it down.

  • No boolean, no filters, no seat to buy. One box.
  • Read at search time, so a profile updated yesterday counts today.
  • Every step visible as it runs, every name with its reason.

500 free credits on sign-up. No card, no demo call. See real searches.

Read next