Bug: `CodeView` markdown remount scroll position jumping

[aagrawal05]

Prerequisites

• I have searched for duplicate or closed issues
• I have read the contributing guidelines

Describe the issue

When viewing diffs with richly rendered markdown, scrolling to bottom then scrolling back up, markdown stops rendering(?), the scroll position is lost jumping upwards to compensate for the change in scroll height, then markdown re-renders(?) — video shows this effect much more clearly.

Reproduction

Can reproduce in codiff v1.9.1 very regularly with large md/source changes

Preliminary AI audit of the DOM:
CodeView loses the scroll position when a tall asynchronous line annotation leaves the rendered window and remounts. In the reproduction, the annotation was initially measured at 1,395px, but its 180px loading placeholder was committed during remount. This reduced scrollHeight from 4,745px to 3,530px and moved scrollTop from 1,301px to 86px. The full height returned after loading, but the original scroll position did not.

Screen.Recording.2026-07-24.at.9.37.10.PM-github.mp4

What browser(s) are you seeing the problem on?

Electron/Chromium

What version of @pierre/diffs are you using?

@pierre/diffs@1.3.0-beta.9

[aagrawal05]

More details of the AI traced repro at #1025 (apologies for the slop issue).
Repro'd on v1.9.1 and checkout of v1.9.2 of Codiff

Let me know if you'd like the exact same diff from the video/trace to repro on, though it should be pretty easy to replicate. Was encountering it often on docs related changes in code-bases.

[amadeus]

Thanks for the new issue! In the process of getting edit released, so once that's out I'll poke around on this.

[amadeus]

Alright, so I spent some time today digging into this, and I have a couple things to share about it, but the core issue here I am pretty confident is something you guys need to fix on the codiffs side, and a fix here would be untenable. That said, I did find an opportunity to strengthen a few things on our end to help alleviate some small occasional scroll jitters, so ultimately, thanks for reporting! (see #1054)

Oki, so to dig into it a bit more.

I'm operating under the following assumption: you guys have a system where every time the markdown file is mounted, it kicks off an async rendering process which means it mounts as a placeholder loader, and then renders the full thing after some amount of time. Fundamentally, this is impossible to course correct for in a reasonable way when scrolling up on my end.

First some context:

Every time you render an annotation, we cache how tall it is, so the virtualizer can subsequently account for that. This is an optimization both for improving rendering performance, along with avoiding dom thrash from scroll fixing if it is needed.

Now, another piece that's important to understand is how we scroll anchor (aka scroll fix). The entire CodeView renderer computes which lines to render based on math it pre-computes, then attempts to verify when the lines mount. If anything differs, it updates anywhere where the math differs from our expectations, then scroll fixes if needed. So to put this into practice, before we render a new update content, we calculate a scroll anchor: basically the first full line or file's who's top is viewable in the viewport (basically how much it offset from the top of the viewport). Then we render our new batch of content, and reconcile those size changes, and ensure the anchor element is still visually in the same position before, if not, we scroll fix.

Why does any of this matter?

In a scenario where scrolling up, we need render lines before the lines we've already rendered. If we take the example in this bug, with the following assumptions -- it mounts with a smaller placeholder, then expands to full height:

• We are scrolling up
• We render the line with the massive annotation.
  • We assume the annotation is massive, and render that line, plus that annotation
  • During reconciliation, we find out the annotation is actually quite short now, so we immediately scroll fix and cache the new height
  • You kick off the async markdown render
• You're still scrolling up
  • Because we don't render a lot of over scroll and because the placeholder is short, the top of the file or a line of content comes into view
• Your markdown rendered response comes back and renders the new massive content
  • This kicks off a resize observer call, which grabs a scroll anchor and attempts to scroll fix and record the new annotation size
    • I hardened this a bit with [diffs] Harden CodeView resize handling #1054, before it would sometimes jitter, but it should be more synchronous now
  • Because the file or first line is in view, that means we just let the content push down with the top of the file or first line still in tact

My suggestion for you guys to fix this would be to be to hold onto your rendered markdown cache, so it mounts on subsequent renders immediately without an async delay (also, once my fix above ships, would also probably improve some scroll jitter issues). Scroll anchoring into dynamic elements is not something I think we should be supporting in diffs.

Anyways, sorry this was long winded. If there's anything I missed or some other specific repo that you can provide code examples for, I can re-open and attempt to dig into those