[PAMulligan]

Follow-up to #55 (already closed) and #157. Two small bugs in conversation export that shipped in 1.17.0. They were found by the re-review of #157's final fix wave, after that wave had closed, so they went out unapplied. The feature is opt-in, so no embed is affected unless a site has switched it on.

What

1. A crafted line could hide the rest of a rendered transcript (09f45fc). Message text gets a backslash in front of a line-leading < so an unterminated HTML block cannot swallow later messages. Autolinks such as <https://example.com> and <help@example.com> are exempt, because escaping them shows the brackets and links the wrong place. That exemption accepted any text before the @, so <!--a@b>, <?a@b> and <![CDATA[a@b> passed as addresses, stayed unescaped, and opened a block that never closes. The address must now start with a letter or digit. Real autolinks are still left alone. Only reachable by a deliberately crafted message, and only against the author's own export.

2. copyText could reject, which it promises never to do (03a76a1, 7a3bf81). The execCommand fallback removed its textarea and restored focus in a finally without guarding either call, and copyText called the fallback outside its own try. Each cleanup step is now guarded, so a copy that worked still reports true and a failed removal does not also cost the visitor their focus. The whole call is wrapped too, so "never rejects" is structural rather than something each statement has to be audited for. The temporary textarea also leaves the tab order, so if removal ever failed it could not linger as a keyboard stop holding the transcript.

3. A comment I wrote gave the wrong reason the fix is safe (e639d3b). It said !, ? and [ "begin exactly the blocks that span blank lines". [ begins nothing, and <script/<pre/<style/<textarea blocks also span blank lines and start with a letter. They stay escaped for a structural reason: the tag name must be followed by whitespace, > or the end of the line, and neither : nor @ is any of those. The comment and the design doc now say that, and what would break it. The design doc had never recorded the autolink exemption at all. No behaviour change.

How it was verified

Test-first throughout: each new test was seen failing against the 1.17.0 code for the stated reason before its fix.

Check	Result
Widget tests	679 passing (673 on main)
ESLint, Prettier, tsc, TypeDoc	clean (the 2 react-refresh warnings on ChatWidget.tsx are on main too)
Build, size gate	pass, no limit touched
Playwright, chromium-desktop	20/20

Beyond the unit tests:

• The symptom, against a real parser. For each crafted line, the message after it now survives rendering by micromark, and the legitimate autolinks still render as links.
• The execCommand fallback in real Chrome, for the first time. Until now it had only ever run in jsdom against a mocked execCommand. With navigator.clipboard removed to force the fallback: the copy command ran once on the focused textarea, the "Copied to clipboard" status appeared, the textarea was removed, focus returned to the menu trigger, and the transcript was on the system clipboard.
• An independent review before opening this PR. It ran about 124,000 generated lines through micromark looking for any that slip past the new rule and still open a block that spans blank lines, and found none. It confirmed no legitimate autolink regressed and that the regex stays linear on hostile input (400,000 characters in under 3 ms per match). Its one substantive finding was item 3 above.

Notes

• Merging cuts 1.17.1, and the release workflow rebuilds cdn/claudius.iife.js. Until then, sites on the @1 tag keep the 1.17.0 build. Since the feature is off by default, nothing is exposed unless a site has enabled it.
• Still open from feat(widget): conversation export (Markdown, JSON, copy to clipboard) #157, and unchanged by this PR: the Safari check. Safari does not focus a button when it is clicked, and the fix for that in feat(widget): conversation export (Markdown, JSON, copy to clipboard) #157 could only be tested by replaying the event order. To check on a Mac or iPhone with the option on: press More options, then press it again. The menu should close.

🤖 Generated with Claude Code

[github-actions]

size-limit report 📦

Path	Size
claudius.iife.js (raw)	217.75 KB (+0.04% 🔺)
claudius.iife.js (gzip)	70.27 KB (+0.04% 🔺)
claudius.iife.js (brotli)	61.38 KB (-0.05% 🔽)
claudius.js (raw)	112.75 KB (+0.12% 🔺)
claudius.js (gzip)	31.13 KB (+0.07% 🔺)
claudius.js (brotli)	26.6 KB (+0.17% 🔺)
claudius.css (raw)	24.03 KB (0%)
claudius.css (gzip)	5.06 KB (0%)
claudius.css (brotli)	4.45 KB (0%)