Who works on PRs from the community?

[fsmeier]

I see like more than 3500 open PRs, the Orca-Team creates PRs like very very fast by yourself.

Whats the procedure for external PRs after the AI Bots reviewed and somebody fixed all findings? I want to avoid that it will just die. Is there any process or some day a week where you look at open PRs and solve them?

All the best from Germany :)

[Pr1p]

Corrected, same day: my original comment conflated the project's own team with the community, and the corrected picture is worse, not better. Full explanation inside.

fsmeier's question has been sitting here unanswered for a few days, so instead of guessing I pulled the data. What I did: every PR created June 1 – Aug 31 — 13,012 of them — and for a random sample of 300 external + 81 internal, every comment, review, and merge/close datum. Snapshot Sep 28.

First, the correction, because it changes the headline numbers. I originally classified "external" by author_association. It turns out the four people who do almost all the merging — nwparker, brennanb2025, AmethystLiang, Jinwoo-H — show up as plain CONTRIBUTORs under that field: GitHub doesn't consider them org members, they just hold write access. I caught this a few hours later re-cutting the data by author. Once you separate them out, the "external" queue splits into two very different things:

• The team's own PRs: ~8,600, two-thirds of the whole cohort. Merged at 76%, median self-merge about an hour — they author and merge their own work, with bots reviewing it like everything else. This is the ship-daily machine, and it is genuinely fast.
• The actual community: 4,384 PRs. 17% merged. 51% still open. That's a census of the full cohort, not a sample estimate.

So the corrected answers to the questions in this thread:

Who reviews community PRs? Bots, and only bots. In my sample, 2 of 31 merged community PRs had any human comment or review at all. The human layer is four people pressing merge — and pressing it overwhelmingly for their own work or for a thin slice of the community.

How fast? For the community PRs that do get merged: median 43.7 hours, and only about a third inside a day. (The team's own: median ~1 hour.)

What happens to the rest? Silence. 42 of 42 still-open community PRs in my sample have never had a single human response. fsmeier's exact scenario — bot-reviewed, author came back and fixed the findings, still open, 30+ days, no human has ever said a word — that's 21 of my 98 community PRs. Half of the open ones did everything the process asked and got silence.

First-time contributors: still 0 for 31. 21 still open, 10 closed. Not typo fixes — the sample includes a 96-file feature.

And the trend is the line I'd underline. Community merge rate by month: June 50% → July 25% → August 8%. (August carries a maturity caveat; June→July doesn't.) Open share went 18% → 37% → 54%. As PR volume grew, the community's share of the merge budget collapsed.

Caveats: "community" here means anyone without write access (proxy: not one of the merger accounts, not MEMBER/OWNER/COLLABORATOR); the sample-based numbers carry ±9pp at n=98; the census numbers don't. Raw per-PR CSV and the updated scripts available on request — the first version of this comment had the wrong split, which is exactly why the scripts matter.

For what it's worth, the cheap fixes look the same as before, just sharper: any human signal on complete-but-stuck PRs, anything at all for first-timers, and some visible triage state. Though I might be missing context on why it's shaped this way — the throughput this team sustains is wild, and the community queue is apparently what gives.

(Disclosures, same as before: I'm a contributor with PRs in this queue myself — a couple in the stuck cohort; the pulls and analysis are scripted and reproducible; the prose was drafted with AI assistance and checked by me, including this correction.)

[bgrablin]

I concur with @fsmeier , and the metrics provided by @Pr1p reiterate that community PRs seem to fall into a black hole of staleness.

Isolated and forgotten contributor of #22561

[fsmeier]

cc @brennanb2025 @nwparker @Jinwoo-H

[nwparker]

Thanks for asking!

The Problem

It's a high noise to signal ratio
And the rate of incoming PRs is very high... and again... much noise to signal.
Just see the comment above from @Pr1p as a great example ☹️ -- sorry if this sounds harsh but reading through that is a chore. I know it's easy to pump out with AI but it's not helpful for us maintainers to get a wall of text like that. We need extremly high signal to noise.

As much as AI is amazing is does cause this problem.

With that being said, we read every single GH issues ever day (and it takes hours 🥹 ). And we are priortizing and addressing them.

Related to 2 above, every PR should be linked to an issue. If it's not, we ignore it. Sorry @bgrablin but that's part of what happened there. The PR tempalte calls this out but understandably not strongly enough. Improving that

It's generally hard to merge features from the community. As it's been said in my OSS projects, the cost isn't the PR but the ongoining maintenance of the PR

The solution

We're imporoving the task / PR filing process: see #23781

Plugin system is coming which will let users unblock themselves (but we need time to focus on it 🙏 )

[fsmeier]

Thank you for taking your time to have a look here! I will check out the discord later on

[bgrablin]

BLUF: I wouldn't have submitted a PR if I didn't love Orca and what you guys are doing! Keep up the great work!

That's fair on my issue #22561 , and thank you for addressing it, cherry picking it and merging your fix! #24064

I went straight to a solution for the problem I identified, as my normal workflow is to see if I can address the problem locally first, if I can, then submit my local changes via a PR. I should've opened the issue first. Noted for the future.

[nwparker]

Final note: If you ping the PR in our Discord it does get more eyes on it

[nwparker]

also if you @nwparker enough there's a decent chance you'll eventually get me

It's kind of like summoning the candyman

[Pr1p]

🫡 Understood — no more walls of text. Discord ping and candyman protocols duly noted for future reference 🍬