mirror of
https://github.com/heygen-com/hyperframes.git
synced 2026-11-19 19:36:04 +00:00
Real-AI-test ran capture against heygen.com twice and both runs died with:
✗ Capture failed: Unexpected token '^' at the "Extracting design tokens" stage.
And, after the first regex fix, again with:
⚠ Asset cataloging failed (no images will be downloaded):
SyntaxError: Unexpected token '^'
Root cause introduced by my commit 02f95f95 (broadening isLogo detection).
In two separate places I added a regex to detect "site-root href" anchors:
/^https?:\/\/[^/]+\/?$/.test(href)
The regex is correct when written in a plain `.ts` file. But the
capture pipeline ships both files as JavaScript strings INSIDE template
literals that get handed to `page.evaluate()`:
const EXTRACT_SCRIPT = `(() => { ... ${regex} ... })()`
const assets = await page.evaluate(`(() => { ... ${regex} ... })()`)
JavaScript template literals consume `\/` as just `/` (the backslash
is not preserved for any character that isn't a recognized escape).
So by the time the string reaches puppeteer's V8, the regex has become
/^https?:////[^/]+//?$/
— and `//` reads as a JS line comment, severing the regex at `[^` and
producing the "Unexpected token '^'" parse error.
The fix is to double-escape so the template literal yields the intended
single backslash:
/^https?:\\/\\/[^/]+\\/?$/.test(href)
^^ ^^ ^^
User applied the fix on the fly during their build to unstick capture
(I had to delete + re-run capture twice). This commit makes the fix
permanent in both locations.
Verified: smoke-tested via `node -e \`var s = `var test = /^https?:\\\\/\\\\/[^/]+\\\\/?$/.test('http://example.com/');`; eval(s);\`` — the template literal now emits the correct regex (`/^https?:\/\/[^/]+\/?$/`) and the test returns true.
Audited the other page.evaluate template literals across
animationCataloger.ts, contentExtractor.ts, designStyleExtractor.ts,
htmlExtractor.ts, mediaCapture.ts for the same pattern. None other
instances found — only my two from 02f95f95 had the bug.
Cost during the lived run: 3 full capture attempts before a usable
output, ~32 assets the first time vs ~100 after both fixes. Logged in
the factual error log from the heygen-showcase build.
Two files, one-character-times-six edit each. No semantic change.