# Changelog All notable changes to the `prompt-area` package are documented here. This project adheres to [Semantic Versioning](https://semver.org/). ## 0.7.1 ### Fixed - **Editing a long prompt no longer slows down as the prompt grows.** Every keystroke walked the whole editor three times — once to read the segment model, once to measure the caret, and once more to put the caret back — so per-keystroke cost rose with the document and typing started to drag well before the prompt got unusual. The three walks are now one: the model scan resolves the caret offset in the same pass, and the caret is only re-placed when the decoration cycle actually moved the nodes it was anchored in (on a plain keystroke it leaves the caret exactly where the browser put it). The per-node checks on that pass also stopped reading `dataset` and `textContent`, which allocate, in favour of `hasAttribute` and `CharacterData.length`, which do not. Measured in Chromium on a production build, median handler time per key with the caret mid-document: | prompt size | typing | Enter | Backspace | | ----------- | ------------ | ------------ | ------------ | | 5k chars | 0.8 → 0.4 ms | 0.2 → 0.3 ms | 0.6 → 0.4 ms | | 50k chars | 2.5 → 0.7 ms | 1.0 → 0.8 ms | 2.0 → 0.8 ms | | 200k chars | 6.9 → 2.0 ms | 3.6 → 2.7 ms | 6.5 → 2.0 ms | Caret navigation (arrows, Home/End) was already off this path and is unchanged at ~0.1 ms. The remaining tail is not ours: a keystroke that merges two lines forces the engine to re-lay-out the whole contenteditable block, which costs ~12 ms at 200k characters both before and after this change. - The ordered-list renumber gate no longer scans every line on every keystroke. `hasOrderedListRun` short-circuits on a cheap regex when the prompt has no numbered line at all, and once a prompt is renumbered, a keystroke on a plain (non-list) line is proven not to have changed the list structure and skips the per-line scan entirely. Editing a list line, undo, drag and drop, or an external value still run the full check. ## 0.7.0 ### Changed - **BREAKING: the `cn` package replaces `clsx` + `tailwind-merge` as the peer dependency.** The class-merging helper is now re-exported from [`cn`](https://www.npmjs.com/package/cn), shadcn's compiled, zero-dependency engine with the same API and full parity with `clsx` + `tailwind-merge`. prompt-area still bundles zero runtime dependencies; `cn` is a peer so it dedupes with the copy your project already ships. npm, pnpm and bun install it automatically; on yarn, or in setups that skip peers, run `pnpm add cn`. `clsx` and `tailwind-merge` are no longer required by prompt-area — keep them only if something else in your project still imports them (`npx shadcn@latest migrate cn` switches a shadcn project over). - **BREAKING: Node 20+ is now required**, matching the `cn` package's engine range. Node 18 reached end of life in April 2025. ## 0.6.8 ### Fixed - **Enter no longer submits the prompt while an IME candidate is open.** Confirming a Japanese, Chinese or Korean candidate with Enter sent the prompt instead of committing the text, so the composition was lost and a half-finished message went out. Key handling is now skipped for the whole duration of a composition, including on legacy IMEs that report `keyCode` 229 without setting `isComposing`. - **An interrupted composition no longer disables the keyboard.** The composing flag was cleared only by `compositionend`, so a composition that never delivered one left the editor with no Enter, no undo/redo, no Escape and no chip-aware Backspace, and no way for the user to recover short of starting and finishing another composition. Blur now resets the flag: a blurred editor cannot still be composing. - **A completed composition is a single undo step.** Committing a candidate could push two entries, one at `compositionend` and another opened by the input event carrying the commit, so undoing one composition took two Ctrl+Z presses. Both browser orderings of `compositionend` and the trailing input now fold into one entry. - **`maxLength` is enforced when a composition completes.** An IME commit could carry the content past the cap. The committed value is now truncated the way typed input is. The `onChange` calls emitted DURING a composition are still uncapped, so a host forwarding them straight to a backend sees over-cap content until the composition ends. - **The caret at the start of chip-leading content lands before the chip.** Not IME specific, and it affects every caret path in the package. When the editor's first child was atomic (a chip, or a `
` on an empty line), offset 0 resolved to the position AFTER it, so `setCursorAtOffset` and `setSelectionAtOffsets` both placed the caret one position too far. Typing at the start of `@alice hello` produced `@aliceX hello` rather than `X@alice hello`, and a caret parked at offset 0 jumped past the chip on any controlled value update. ### Changed - **`onChange` no longer fires for input events that leave the content unchanged.** The prop is documented as "called when the content changes", but every input event notified, including the duplicate echo a browser emits after a composition. Hosts that counted keystrokes, drove a "user is typing" indicator, or set a dirty flag from `onChange` will see fewer calls. Calls emitted during a composition are unaffected and still fire on every event. - **Key handling is fully suspended during a composition.** Trigger dropdown navigation (arrows, Tab), Escape, undo/redo and the markdown shortcuts still ran mid-composition; they now wait for the commit, as Enter, Backspace and Delete already did. ## 0.6.7 ### Fixed - **The caret lands at the end of a pasted document.** Pasting several pages from Word or Google Docs left the cursor partway up the document — on an 11k-character contract, 1679 characters short of the end, roughly 85% in. The paste itself collapsed the caret correctly; the re-render that follows it (the pass that decorates the pasted text) preserved the caret as a child-node INDEX, and decoration replaces one text node with a text/``/text run, so the editor ends up holding more direct children than when the index was captured and the restore lands proportionally short. The drift scaled with the document, because a word-processor paste turns every bold heading into `**…**`. The caret is now preserved as a plain-text offset, which survives decoration. - **A pasted ordered list keeps the numbering its author wrote.** Ordered-list renumbering restarted every run at 1 — right for typing, where Tab-indenting restarts a sublist, but wrong for a paste, whose numbering was authored somewhere else. A contract's section 7 arrived as 1, and every section resuming after a paragraph of prose dropped back to 1 again. A pasted run now continues from its own first number; contiguity is still rebuilt, so a stale `7. 7. 7.` lands `7. 8. 9.` Typing is unchanged. - **Sublists emitted as a sibling of their parent item are no longer dropped.** Word, Outlook and Apple Notes put a nested list NEXT TO the `
  • ` it belongs to rather than inside it. The HTML→markdown converter skipped every child that was not an `
  • `, so every nested item vanished from such a paste — silent content loss. Both shapes are handled now, and a sibling sublist does not advance the parent list's counter. ### Changed - **A pasted Word list keeps the numbers Word rendered.** A list fragment whose markers read `3.` `4.` now pastes as `3. 4.` rather than being normalized to `1. 2.` (the behaviour 0.6.5 shipped). The numbers a word processor renders are the author's, and rewriting them is what corrupted real documents. ## 0.6.6 ### Fixed - **Typing stays fast when the editor holds pages of text.** Every keystroke used to do work proportional to the whole document: the caret offset was measured by deep-cloning everything before it (`Range.cloneContents`, two to three times per keystroke), the DOM was read and serialized to plain text up to four separate times, all decorations (bold/italic, URLs, bullets, indents, headings) were stripped and rebuilt across every line, autoGrow forced two full layout reflows, and a fresh empty suggestions array forced a React re-render even when nothing changed. On a multi-page document a single keystroke blocked the frame for 40–80 ms. Offsets are now computed by walking the live tree without cloning, one scan produces the segments and plain text for the whole keystroke, the decoration cycle runs only on the caret's `
    `-delimited line (decorations are line-local; anything ambiguous — foreign elements, literal newlines in text nodes, no collapsed caret — still takes the full pass), height syncs coalesce into one pre-paint `requestAnimationFrame`, and idle keystrokes no longer re-render. Measured at 600 seeded lines with markdown and autoGrow: median keystroke handler time dropped from 39.8 ms to 7.5 ms and input-to-frame latency from 45 ms to 8 ms, with p95 under one 60 Hz frame. - **Newlines insert without rebuilding the document.** Shift+Enter (and Enter with `submitOnEnter` off) re-rendered and re-decorated the entire editor to insert one `\n`, a ~37 ms stutter at 600 lines. When the caret is collapsed in a clean flat DOM and ordered-list renumbering is provably a no-op, the newline is now inserted surgically — split the caret's text node, insert one `
    `, re-decorate just that line — halving the cost; every other case keeps the full re-render. - **URLs keep their commas.** The URL matcher excluded commas outright, so a link like `https://en.wikipedia.org/wiki/Dominic_Johnson,_Baron_Johnson_of_Lainston` was cut short at the first comma. Commas now count as part of a URL, and a trailing one is trimmed along with the other trailing punctuation, in both the editor decoration pass and inline markdown parsing. ## 0.6.5 ### Fixed - **Pasting from Microsoft Word inserts the text instead of silently dropping it.** Word (macOS especially) puts a bitmap rendering of the copied selection on the clipboard next to its HTML; the image branch of the paste handler consumed it and returned before any text flavor was read, so with no `onImagePaste` consumer wired nothing happened at all. Office clipboards (detected via their `mso-`/`Mso`/ProgId/office-xmlns fingerprints) now let the text path win; the bitmap is only delivered to `onImagePaste` when the clipboard yields no text — e.g. an image or drawing object copied inside Word. Non-Office pastes are unchanged: a copied image still takes precedence over incidental `` HTML. - **Word lists convert to real markdown lists.** Word's clipboard HTML has no `