跳到正文
DEV Community· John Alexander Kamau·· 8 小時前AI 評分54

一個月向開源專案提交 22 個 PR:John Alexander Kamau 復盤合併與未合併原因

22 pull requests to open-source projects in a month: what got merged, what didn't, and why

AI 導讀

John Alexander Kamau 在 2026 年 10 月 15 日至 11 月 13 日期間,向開源專案提交 22 個 PR,其中 11 個合併、4 個仍在審核、7 個未合併。

正文

Between 15 October and 13 November 2026 I opened 22 pull requests to open-source projects I don't own. Eleven were merged, four are still in review and seven were closed without merging.

That's a 50% merge rate overall, or 61% of the ones where a maintainer has made a decision. The median merged PR took about four days from opening to merge. The fastest took about 40 minutes; the slowest, almost nine days.

Stacked bar: of 22 pull requests, 11 merged, 4 in review and 7 closed

This post covers what those PRs fixed, what the maintainers changed before merging, and why the other seven didn't land. The closed ones taught me as much as the merged ones.

The merged eleven

Project Stars What the fix did
Besu, the Ethereum client (Java) 1.9k Three fixes, below
rust-analyzer (Rust) 17k Made VS Code link Rust errors to files again with rustc's unicode output
Quarto (TypeScript) 6.2k Made the dismiss button on website announcement bars usable from the keyboard
Ratatui (Rust) 23k Tests that stop halfwidth katakana sound marks from being counted twice
Thumb-Key (Kotlin) 1.6k Two fixes: layout search, and a "hide letters" key for custom layouts
Authlib (Python) 5.5k Three documentation PRs answering open questions from the issue tracker

Star counts as of 18 November 2026.

Seven of the eleven are code fixes. Three are documentation and one is tests only.

Besu: three small bugs, three tests

Besu is a Java Ethereum client. I sent three fixes the same day:

  • miner_getExtraData dropped leading zero bytes. Set the extra data to 0x000102 and it came back as 0x102, so the value didn't round-trip.
  • admin_logsRemoveCache treated the safe and finalized block tags as the chain head instead of resolving them.
  • debug_standardTraceBlockToFile ignored enableReturnData: false, because the value was hard-coded to true.

Each came with a test that fails without the fix. The maintainer merged all three within about nine days. For the third, the maintainer also opened a docs PR so the documentation matched the code.

rust-analyzer: one changed line

My smallest merged PR. rustc can print error locations with unicode box drawing, like ╭▸ src/main.rs:2:5 instead of --> src/main.rs:2:5. rust-analyzer's VS Code problem matcher, the regex that turns compiler output into clickable errors, didn't know about ╭▸. So the arrow became part of the file name and VS Code couldn't find the file.

The fix was one line in the regex. This one had no automated test: the problem matcher lives in the extension's package.json, and I checked the new pattern against rustc's own unicode output samples instead. The maintainer said as much when merging: it couldn't easily be tested, but the change was right.

Quarto: accessibility, tested in three browsers

On Quarto websites, the "X" that dismisses an announcement bar couldn't be reached or pressed with a keyboard. I made it a real button and added a Playwright test that runs in Chromium, Firefox and WebKit.

The maintainer added one commit, to keep the icon at the top when the bar wraps onto several lines, and noted that the tests fail without the change in all three browsers. It merged the next day.

What the maintainers changed

The PR I learned the most from is the one where my fix didn't ship.

In Ratatui, halfwidth katakana sound marks (゙ and ゚) were counted twice when measuring text width with one version of the unicode-width crate. I wrote a guarded workaround that worked on both 0.2.2 and 0.2.3, plus tests.

The maintainer took a simpler route: raise the minimum unicode-width version to 0.2.3, where the bug doesn't exist, and delete the need for a workaround. The maintainer kept my tests, rebased them onto main and merged them as the guard against the bug coming back.

Fewer lines, less to maintain, and the right call. The lesson for me: before writing compatibility code, ask which versions actually need supporting.

In the other merges, maintainers changed little: in Besu they merged main into my branches and asked for a changelog entry in the right section; in Thumb-Key and Authlib the PRs merged as sent.

The seven that didn't merge

  • pandas (50k stars): closed automatically the same day. The issue was already assigned to another contributor, and pandas allows one contributor per issue. I hadn't checked.
  • Thumb-Key: I closed one myself because another PR had already landed the same fix.
  • reqwest (12k stars), twice: I closed the first myself after realising I hadn't followed the PR template, which asks for a description written in your own words. The maintainer closed the second, a docs change, saying the docs were already clear. Fair enough: it's their project.
  • ParadeDB: one PR was closed without a comment. I've asked why and haven't heard back yet.
  • REA, twice: both were superseded by the maintainer's own commits. One of those commits adapted my change and credits me as co-author, so the code landed even though the PR didn't.

None of these were about code quality. Most came down to things I could have checked first: who already claimed the issue, whether someone else was on it, and what the project asks of a PR.

Still in review

  • Besu: a larger change that rejects block overrides and unknown parameters on five JSON-RPC methods, now through several review rounds.
  • Qwen Code (30k stars): inline session rename with an F2 shortcut. A maintainer has called it merge-ready.
  • ParadeDB: support for constant output columns in an aggregate scan. The reviewer said it "mostly looks great", with one blocker that I've since fixed.
  • Ratatui: removing APIs that were deprecated back in version 0.26.

Where AI fits in

I use AI throughout this work, the way I use any good tool. It reads unfamiliar codebases with me, drafts and runs tests, checks a fix against the reproduction, and reviews my diff cold, seeing only the issue and the change, before anything goes out. That turns a day of finding my way around a new Java or Rust codebase into an hour or two.

It doesn't make the decisions. I choose the issues, decide whether a fix is right for the project, read every line before it's sent, and answer the reviewers myself. If a maintainer asks why a line is there, I can explain it. Projects that don't accept AI-assisted contributions get nothing from me, and where a PR template asks for a description in your own words, that's what it gets.

How I work on a PR now

After a month of this, these are the rules I follow:

  1. Every fix comes with a test that fails without it. This earns more trust than any description. Where a test isn't practical, like the rust-analyzer regex, I say so in the PR and show how I checked it.
  2. Keep the diff as small as the bug. One of my open PRs grew several times over during review. Small PRs get reviewed; big ones wait.
  3. Read the contribution rules first. CONTRIBUTING, the PR template, the CLA, and the project's policy on AI-assisted contributions. Some projects don't accept them at all, and I don't send them anything.
  4. Check who's already working on it. Look for assignees, linked PRs and recent comments before writing code.
  5. Leave product decisions to the maintainers. When the docs and the code disagreed in Besu, I fixed the code and asked which behaviour they wanted documented, rather than deciding for them.
  6. Respect review bandwidth. After my first few Besu PRs, the maintainer asked me to get those through review before opening new ones. That was fair, and all three merged within a week. I now keep at most two PRs waiting per project.

Why I'm doing this

Two reasons. Reading and fixing code in mature projects across Java, Rust, TypeScript, Kotlin and Python makes me a better engineer faster than anything I build alone. And every merged PR is a small, public, checkable piece of work, reviewed by people who don't know me and have no reason to be kind.

If you maintain an open-source project, what makes a first-time contributor's PR easy for you to merge?

來源:DEV Community · dev.to