mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-11-09 16:02:36 +01:00
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work > - People write task requests and chat messages in the Markdown editor, and the stored text becomes the agent's prompt > - MDXEditor supports `==highlight==` marks, and its exporter escapes every `=` in prose to protect that syntax > - The editor therefore stores `a=b` as `a\=b` and `x=>y` as `x\=>y` > - A command or code snippet in a chat message then reaches the agent as invalid code > - This pull request removes the `=` escape when the escape cannot change how the text parses, as #19350 did for intraword underscores > - The benefit is that commands, assignments and arrows reach agents exactly as the person typed them ## Linked Issues or Issue Description Refs #19350 **What happened?** A chat evaluation sent a `node -e` command through the chat composer. The stored comment contained `const fs\=require(...)` and `()\=>`. Every `=` had a backslash. The agent received invalid JavaScript, and an exact-text check on the stored message failed. **Expected behavior** Text the person types is stored as typed. A `=` that cannot start or end a highlight needs no escape. **Steps to reproduce** 1. Type `a=b` or `x => y` in any Markdown editor field, such as the chat composer. 2. Read the stored value through the API. 3. Before this change it is `a\=b` or `x \=> y`. ## What Changed - New `ui/src/lib/equals-markdown.ts`: `unescapeEqualsSigns` removes `\=` escapes from prose only when they cannot change the parse: - any run of escaped `=` whose length is not exactly two (a highlight sequence is exactly two `=`); - a run of exactly two with whitespace or the text edge on both sides (such a run can neither open nor close a highlight). - Other two-runs, such as `a\=\=b` or `\=\=text\=\=`, stay escaped. Escaped backslashes (`\\=`) stay as they are. - `ui/src/lib/underscore-markdown.ts`: the code-range walk is now the shared `rewriteMarkdownProse` helper. Code spans, code blocks and raw HTML stay byte-identical for both rewrites. Underscore behavior does not change. - `MarkdownEditor.tsx`: `toStoredMarkdown` applies the new step after the underscore step, on the exporter's exact text. - Tests: unit cases for commands, `===`, spaced `==`, quotes and lists, highlight-capable runs, code spans and blocks, raw HTML and escaped backslashes. One editor test for a pasted command with a code block. ## Verification - `npx vitest run src/components/MarkdownEditor.test.tsx src/lib src/components/task-chat` in `ui/`: 280 files, 3388 tests passed. - Negative control: with the new step removed from `toStoredMarkdown`, the new editor test fails. - Parser check: I parsed the unescaped forms with `micromark-extension-highlight-mark` 1.2.0, the version MDXEditor uses. `a=b and c=d`, `x === y`, `a == b == c`, `()=>x` and `x!==y` produce zero highlight nodes. The controls `==a==`, `a ==b== c` and `say ==this== now` produce one highlight node each. - `tsc -p ui/tsconfig.json`: no errors in the changed files. The remaining errors are in files this change does not touch. ## Risks - Low risk. The step only removes an escape when the highlight tokenizer cannot use that `=`. Text the parser cannot place is returned unchanged. - Some `==` in code-like prose (for example `x!==y`) stays escaped. This is intentional: such a run can open a highlight. > For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and discuss it in `#dev` before opening the PR. Feature PRs that overlap with planned core work may need to be redirected — check the roadmap first. See `CONTRIBUTING.md`. ## Model Used - Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, with tool use and code execution in Claude Code. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass - [x] I have added or updated tests where applicable - [x] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [ ] All Paperclip CI gates are green - [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
56 lines
2.6 KiB
TypeScript
56 lines
2.6 KiB
TypeScript
import { isEscaped, rewriteMarkdownProse } from "./underscore-markdown";
|
|
|
|
/**
|
|
* MDXEditor's `==highlight==` extension (`mdast-util-highlight-mark`) marks
|
|
* every `=` in prose as unsafe, so the exporter stores `a=b` as `a\=b` and
|
|
* `() => x` as `()\=> x`. Stored text feeds agent prompts, so a pasted command
|
|
* or snippet reaches the agent with backslashes the person never typed.
|
|
*
|
|
* A highlight opens or closes only on a run of exactly two `=`, and a two-run
|
|
* with whitespace on both sides can do neither. A run that starts a line can
|
|
* underline a heading, so it keeps its escape. Other escapes cannot change how
|
|
* the text parses, so they are removed. Code
|
|
* spans, code blocks and raw HTML stay byte-identical, as do escaped
|
|
* backslashes (`\\=`).
|
|
*/
|
|
const ESCAPED_EQUALS_RUN_RE = /(?:\\=)+/g;
|
|
|
|
function unescapeProse(text: string): string {
|
|
// One forward pass: matches arrive in order, so track whether only
|
|
// indentation or quote markers precede the current offset on its line.
|
|
let scanned = 0;
|
|
let lineStartOnly = true;
|
|
return text.replace(ESCAPED_EQUALS_RUN_RE, (run: string, offset: number, whole: string) => {
|
|
for (; scanned < offset; scanned += 1) {
|
|
const char = whole[scanned]!;
|
|
if (char === "\n") lineStartOnly = true;
|
|
// The exporter writes a leading `>` as `\>`; it is still a quote marker.
|
|
else if (lineStartOnly && char === "\\" && whole[scanned + 1] === ">") scanned += 1;
|
|
else if (char !== ">" && !/\s/u.test(char)) lineStartOnly = false;
|
|
}
|
|
const atLineStart = lineStartOnly;
|
|
scanned = offset + run.length;
|
|
lineStartOnly = false;
|
|
if (isEscaped(whole, offset)) return run;
|
|
// At the start of a line (after indentation or quote markers) a run of `=`
|
|
// can underline the previous line as a heading, so its escape stays. This
|
|
// includes the start of a prose chunk, whose preceding text is unknown.
|
|
if (atLineStart) return run;
|
|
const length = run.length / 2;
|
|
const before = whole[offset - 1];
|
|
const after = whole[offset + run.length];
|
|
// A neighbouring literal `=` would join this run into a different length.
|
|
if (before === "=" || after === "=") return run;
|
|
// A chunk edge borders a code span or block, not whitespace, so it never
|
|
// counts as space around a two-run.
|
|
const spaced = (char: string | undefined) => char !== undefined && /\s/u.test(char);
|
|
if (length === 2 && !(spaced(before) && spaced(after))) return run;
|
|
return "=".repeat(length);
|
|
});
|
|
}
|
|
|
|
export function unescapeEqualsSigns(markdown: string): string {
|
|
if (!markdown.includes("\\=")) return markdown;
|
|
return rewriteMarkdownProse(markdown, unescapeProse);
|
|
}
|