Ion — engineering log

I read other people’s open source, find the bug the issue is actually about, and send the smallest fix that holds. This page is a log of those fixes and what actually broke in each one.

How I work

I don’t chase easy issues. I go for the ones where a maintainer already said “PR welcome” but nobody has done the reading yet: figure out the exact failure, confirm it against the real code, write a test that fails without the fix, then keep the change as small as the problem allows.

Half the value is knowing when not to send a patch. If a fix would break the project’s own examples, or the real problem is a maintainer-only release decision, the honest move is to say so on the issue and stop, not to ship something that gets closed.

Working with AI, openly

I use AI as part of how I work, and I disclose it on every pull request, naming the model that did the bulk of the work on that specific change. Here is what that means in practice.

How I route it:

I run a few models and pick by how hard the task is, not by brand loyalty. Claude (Opus for tricky logic and concurrency, Sonnet for the actual implementation, Haiku for search and reading), plus GPT, Grok, and GLM when a second opinion or a different strength helps.
Cheap, wide work (grepping a large codebase, reading unfamiliar files) goes to the fast, small models. The careful reasoning (root cause, race conditions, edge cases) goes to the strongest one I have.
When two models disagree about a root cause, that disagreement is useful. I dig until I can tell which one is right, in the code.

What I use it for:

Finding where a bug actually lives instead of guessing, then reproducing the failure and drafting the fix and the test that proves it.
Checking a repo’s contribution rules and AI policy before I touch anything, and following them.

What I check myself, every time:

That the bug reproduces and the fix genuinely resolves it, not just that tests are green.
That the change is the smallest one that works, and doesn’t quietly break the project’s examples or CI.
That the PR reads like a person wrote it, because a wall of generated prose wastes a maintainer’s time.

The pace looks automated from the outside, and the work is assisted. Every change is still read, tested, and submitted by me, and if a maintainer has questions about one, they come to me.

Recent fixes

restic/restic +91 -20

A race in restic init could orphan a key file

What broke: init wrote the key file before the config file, so two clients initializing the same repo at once could leave a key that no config ever pointed at. The fix reorders the two writes and moves key generation into the caller, so config lands first.

vectordotdev/vector +52 -3

S3 sink rejected by Cloudflare R2 with InvalidRequest

What broke: Vector sent its own Content-MD5 plus the AWS SDK’s newer checksum headers on every upload. Standard S3 tolerates both; R2 refuses more than one non-default checksum. The fix limits checksum calculation to when it’s actually required.

prettier/prettier +68 -0

import.meta silently dropped a whole script block

What broke: in a non-module <script>, Babel rejects import.meta, and the HTML embed logic swallows any embed that throws, so the entire block came out unformatted. The real fix was one line in the Babel error-recovery allowlist, not in the formatter.

zellij-org/zellij +108 -0

Can’t bind Ctrl+\ as a key

What broke: the key map only covered Ctrl+A..Z. Control chars from Ctrl with non-letters (\, ], ^, _) fell through with no modifier attached, so the bind never matched. Fixed at the call site, leaving the vendored code untouched.

rclone/rclone +19 -4

nfsmount passed Linux options to OpenBSD’s mount_nfs

What broke: the mount options were hardcoded in GNU syntax, so OpenBSD’s mount_nfs rejected mountport. Added a GOOS branch for OpenBSD’s syntax, matched the man page, and said plainly in the PR that a BSD user should confirm at runtime since I couldn’t.

sveltejs/kit +24 -14

A cookie-size limit that was a magic number

What broke: MAX_COOKIE_SIZE was 4129 and compared against the whole serialized cookie. The spec’s 4096 is for name and value only, so the fix aligns the value and checks the right thing. Also found the existing test was a no-op.

Colophon

One static HTML file, no build step, no trackers, no analytics. Set in the system old-style serif with a monospace for code.