[nedtwigg]

When one change lands on many stories, like a header change that touches every page, you currently have to review each story separately. This finds the changes that repeat and lets you accept each one once. It's fully client-side, with no SQL or stored-format changes.

What it does

• Finds repeated changes: the batch page groups hunks by the size of each side, then by the SHA-256 of each side's pixels. Any group in at least 3 undecided stories is offered as a delta.
• Lets you look first: one delta is open at a time. Previous and Next go through its stories, each with a link, the crop, and the whole before and after images with every place it has the delta outlined. Whole images matter because a delta's pixels are the same everywhere, so only where it landed can differ.
• Accepting shrinks the diff: the delta is drawn onto each golden in this browser, so the story page shows only the hunks that are left. It says the golden is redrawn and that the one on record hasn't changed. Accepted deltas are kept in localStorage, and each can be undone.
• Reaching zero approves: an undecided image with no hunks left goes through the existing POST /batches/:id/decisions, after the usual bulk confirmation. Undoing a delta doesn't take back an approval.

The specs are updated: ui.md and its rationale, access.md, and glossary.md. The page's CSP now allows blob: images for the redrawn goldens.

Trying it on main

On main's real batch of 124 changed stories, Selfie offers 28 deltas. Accepting them all approves 83 stories, leaves 35 with a smaller diff, and leaves 6 with no delta to accept.

Testing

• pnpm test passes: ui 96, store 621 (92 skipped), render 148, cli 512.
• cli/test/deltas.test.ts runs the real differ over every combination of hunks, and shows that redrawing some leaves exactly the rest.
• All 83 finished images were compared pixel by pixel with their current image in a browser, through canvas and PNG, and match.
• Previous, Next and "Look through" were driven on the real batch. At 1000px wide, none of the 54 crops overflow.
• The new stories render without failures.

Known limitations

• Rendering noise is still offered as deltas: 23 of those 28 offers are faint edge differences you can't see, such as 1/255 on a rounded corner. They appear to come from Chromium's raster tiles, because a 15px shift moves edges across tile seams. Fixing this is left for later.
• Stats are approximate on a redrawn story: spacing inside changed hunks isn't counted, and the largest channel difference is hidden. Exact numbers need per-hunk stats in the stored diff format.
• Accepting all 28 takes 12 confirmation dialogs: there's no "Accept all" yet.
• Buttons only: Previous and Next have no keyboard shortcuts.

🤖 Generated with Claude Code