v0.0.185
Release notes
Seven new rules this time, with a soft spot for the things that quietly go wrong as a site grows: titles chopped mid-word by a character limit, pages left behind on an old template, a palette that sprawled into a dozen almost identical blues, one anchor text pointing three different ways, a 404 page that answers 200, a home page that does not load at all, and the WebMCP tools a browsing agent can call. Every report now says how much JavaScript a site ships and who sent it, squirrel report --group filters by rule group, and script-heavy pages audit faster.
Added
- When your balance makes
squirrel auditrun a smaller audit, the CLI prints your balance, what the next audit needs, and a top-up or upgrade link, at most once a day per account (squirrelscan/repo#1118). - New rule
content/truncation-artifacts(squirrelscan/repo#1360): flags a<title>, meta description or body excerpt that was cut mid-word by a character limit. A text counts as cut when it is the start of a longer text on the page (the H1,og:title,og:descriptionor the body) and the longer text carries on inside a word, when an ellipsis is glued to a partial word at a round CMS limit (60, 70, 155, 160 and so on), or when a title or description of exactly that length ends on a dangling word such astheorand. A title or description cut fails; body copy ending in an orphaned ellipsis with no read more link only warns. An ellipsis after a whole word, a quotation and noindex pages are left alone. - New rule
integrity/chrome-outlierflags pages left on an outdated or broken site template: pages in small template clusters and singletons on a site where one template covers at least 70% of pages (10 pages minimum) that lack that template's navigation, its footer, all of its stylesheets or all of its CSS and script hosts. A page that only adds a route-level stylesheet or a body class, or keeps the site stylesheet while lacking a post stylesheet, an image CDN or a preconnect host, is not reported. It reads the stored template fingerprint, skips pages thatintegrity/template-discontinuityalready reports, and warns on a few stragglers or fails when they reach 10% of the crawl. See integrity/chrome-outlier. (squirrelscan/repo#1367) - New rule
links/anchor-target-conflict(squirrelscan/repo#1372): reports one anchor text that points at several different pages across the site (the pages that deviate from the leading destination are listed), and one page linked with four or more unrelated anchor texts. Generic anchors such asread more, navigation links and the homepage are left out, it needs at least 10 pages, and an anchor is only reported when one destination takes at least 70% of the pages using it. See links/anchor-target-conflict. - New rule
content/color-sprawl: converts every color literal (hex,rgb(),hsl(),oklch(), named) in<style>blocks andstyle=attributes to OKLab and clusters by perceptual distance. It warns when more colors are visually indistinguishable from another thanmax_redundant(default 6), failing at three times that, and warns for a color that only one page uses inline and no linked stylesheet the audit read uses or defines. Literals in--custom-propertydefinitions are not counted, so a token-driven design system stays clean. Near-duplicates are counted from inline styles only. Up to 20 same-site linked stylesheets are read (ctx.site.stylesheets) to define the palette, so a brand color a stylesheet uses site-wide is not an outlier; every finding says how many stylesheets were read, or that none were. A site with no inline color data is skipped, not passed. A crawl under 10 pages is skipped. Local rule, no account, no credits. The per-page signal joins the site-dom-signals record, so each site's first audit after the upgrade runs fresh once. See content/color-sprawl. (squirrelscan/repo#1377) - New probing rule
crawl/custom-404(squirrelscan/repo#1386): on anactiveoraggressiverun it sends oneGETto a random path that cannot exist and judges the answer. A 2xx is a soft 404 (fail), a redirect is a warning, and a real 404 or 410 is checked for a stock server page, a route back into the site andnoindex. Belowactive, offline, or when no 404 response was observed it reports skipped, never pass. It extendscrawl/soft-404, which still judges the pages the crawl fetched. - New rule
crawl/home-page-unreachable: when the audited URL does not load because its DNS did not resolve, the TLS handshake failed, the connection was refused or dropped, the server timed out or it answered 5xx, the audit now reports an error finding with the cause, so a down site shows up in findings and alerts instead of only as a failed run. A 4xx answer, a robots.txt refusal and a bot wall are not reported by this rule. When the audit has no result for the home page (it was not fetched, or failed for a reason squirrel could not classify), the check is listed as not checked and an earlier finding stays open. See crawl/home-page-unreachable. (squirrelscan/repo#2035) - A JavaScript shipped line in every report format. It gives the script bytes the audit measured, the framework runtime's share when one is recognised by its chunk path (Next.js, Nuxt, Astro or SvelteKit), the third-party share, the scripts not measured, and the three pages that load the most script. The JSON report has a top-level
javascriptobject (schema 1.28), and the LLM and XML reports a<javascript>element. The line is left out when no script was measured, never affects the score, and survives a latersquirrel report. See JavaScript shipped (squirrelscan/repo#3397). - New rule
ax/webmcp(squirrelscan/repo#396): reports declarative WebMCP tools, the<form toolname tooldescription>declarations a browsing agent can call, and flags invalid ones (an empty or malformedtoolname, a missingtooldescription, duplicate names,toolnameon a non-form, named-parameter problems). It also lists forms that are not tools, notes inlinenavigator.modelContextregistration (detected, not validated) and notes a visible CAPTCHA widget on a tool form, recommending an invisible or agent-tolerant check while keeping the protectionsecurity/form-captchaasks for. It is informational and never lowers the score; the accessibility and layout-shift checks stay in the existinga11y/*andperf/cls-hintsrules, which it links to. See ax/webmcp. That makes 343 rules. - On rendered pages, the XHR and fetch requests a page makes now feed endpoint discovery as
rendercandidates, under the same same-origin filter and caps. Cross-origin requests are recorded but never probed (squirrelscan/repo#4188). - Filter by rule group (
seo,performance,security,ai).squirrel report --group securityshows only that group's issues; it takes a comma-separated list and combines with--category, so an issue must match both.rules.enableandrules.disablein config, and--rule-includeand--rule-exclude, acceptgroup:<id>, which expands to every category in the group (group:performancecoversperf/*). An unknown group is an error that lists the valid ones. See squirrel report and config (squirrelscan/repo#639).
Changed
- A re-audit within 24 hours reads the site's scripts from the local cache instead of requesting them again, keeping each script's content type, compression, caching headers and response time;
--refreshor--no-incrementalturns the cache off, and a script that redirects or sends aSourceMapheader is never cached, so its redirect and source map findings stay accurate on every run (squirrelscan/repo#1415). - Script-heavy pages audit faster (squirrelscan/repo#1864): each inline script's text is read once per rule.
security/leaked-secretsno longer scans every inline script a second time inside the page's HTML, which takes its per-page cost on a 1 MB storefront page from about 32 to 17 ms. A value in an inline script is reported only atinline-script(a public Shopify token on a Shopify page used to show ashtml), andmeta.occurrencesno longer lists a duplicatehtmlentry for it.a11y/skip-linkandperf/js-librariesno longer serialise or search the whole page for what they already read; their results are unchanged. schema/breadcrumbno longer fails a page because of one link: only a breadcrumb trail named in the markup (anaria-label,classoridcontaining breadcrumb, outside the site header and footer, and not a page-state class such ashas-breadcrumbsor a page wrapper) counts, a link the page classifier's model alone placed in the breadcrumbs section never does, and the trail needs at least two links. Because the trail is read from the markup, the check now also runs with--no-classifier, and a link in a markup-named trail now carries thebreadcrumbssection (squirrelscan/repo#3365).- With the page classifier on,
content/keyword-stuffingleaves words in the header, navigation, footer, sidebar and other site chrome out of its density count, and checks footer link blocks on their own (keyword-stuffing-footer). With the classifier off nothing changes (squirrelscan/repo#3544).
Fixed
crawl/sitemap-coverageandintegrity/orphan-pagetreat a sitemap whose read state was not recorded as unknown, not complete. Coverage skips its missing-page check and says why, and orphan-page no longer escalates a page to likely injected on that sitemap's word (#1869).- The CLI's bot protection warning now appears only when the crawl was actually refused, and it says how many URLs were refused. A site whose protection let the crawl through no longer gets the warning (squirrelscan/repo#2237).
- A link the cloud link checker answers with a bare 401, 406 or 999 is listed as unverifiable in
waf-blocked-external-links, withmeta.unverifiedReasonblocked, instead of warning as possibly broken. The answer is no longer written to the link cache, and late-bound link verdicts carry the same answer. A bare 403 or 405 from the cloud checker's final browser check is kept as sent (squirrelscan/repo#2730). schema/expected-typeno longer adds an info finding to every page when no page classifier ran (in the CLI,--no-classifierorclassifier.enabled = false). Each page now shows a skippedNot checked: no page classifier ran on this pageline instead, and an earlierexpected-typefinding on the page stays open rather than resolving, because the page was not checked (squirrelscan/repo#3364).- A report rebuilt with
squirrel reportnow carries therenderblock (pages rendered and fetched as plain HTML) in its JSON andllmoutput, and it is the block the audit wrote, fallback reason included. It was missing, so a rendered audit read as one that rendered nothing. Audits stored before this release get an approximate block rebuilt from the stored pages, without the fallback reason (squirrelscan/repo#4853).
Docs
- Rule totals say 340+. The README, the docs and the npm and plugin descriptions state the rule count rounded down to the nearest 10 (342 enabled rules today) instead of 300+. (squirrelscan/repo#4856)
Update to this version
squirrelscan will auto-update, or run this command to update now.