Onsite audit

On-page audit (read-only)

1. Read profile if a brain repo exists (optional - this skill also works bare).
2. Public checks per URL (WebFetch): title (length, keyword presence), meta description (present, and flagged when over ~155 chars - Google truncates around there), H1 count, heading structure, canonical, robots meta, image alts, internal links out, JSON-LD present/valid, answer-capsule presence in the first 200 words, server-rendered content check.
3. Page essentials dimension - runs for each audited page, and site-wide where a check says so. Every check emits the standard falsifiable signal line (observation | why it matters | "we are wrong if..." | leading indicator) and, where a fix is actionable, feeds onsite-propose. Evidence discipline throughout (see plugin/docs/evidence.md): state what was measured versus what was inferred, and never promise a third-party outcome (rankings, citations) as the result of a fix.
   1. Author entity (E-E-A-T): the author has a visible on-page bio, sameAs links, and - for AI-attributed content - a named human reviewer note on the page; the author schema node carries description + sameAs. A thin entity -> P2 signal + a gated proposal. The remediation is site-level: the CMS adapter can write the author profile description where capabilities()['author_profile_fields'] is true (the WordPress adapter's update_user writes it over REST), plus the schema settings; any step beyond the adapter's capabilities() ends partially-applied per skills/onsite-apply, with the human step named.
   2. Answer capsule: a 40-60 word standalone answer sits above the first H2. Missing -> P2 signal + a content-refresh proposal.
   3. In-content images: zero images in a 500+ word explainer -> P3 signal referencing the image-brief path (skills/ce-image writes <slug>-image-brief.md). Image findings - this gap and step 2's missing alts - now feed image-fix proposals (the image-fix action type in skills/onsite-propose, executed gated by skills/onsite-apply).
   4. Meta description length: lives in step 2's meta checks (extended there with the ~155-char flag) - listed here for completeness, never duplicated as a second check.
   5. Social image shape: og:image is square while twitter:card is summary_large_image -> P3 signal + proposal (site-level asset); if no landscape asset exists, route via the image-brief fallback above.
   6. Publisher schema shape: the publisher node missing logo/url, or mixing Person and Organization types -> P3 signal for the entity-schema lane (entity-schema-engineer).
   7. Sitemap membership (site-wide): any published post older than 1 hour absent from the sitemap -> P2 signal ("likely regeneration lag or cache; if the SEO plugin's cache purge needs admin, this ends partially-applied").
4. Link health and internal-link graph dimension - site-wide, one capped crawl per run. Own-site rule, stated because it matters: this crawl touches ONLY the profile's own site. ADR-0006's no-scraping decision bans scraping third parties; your own property is yours to crawl, and onsite.linkgraph.crawl enforces it - it never follows an off-host link. Run crawl(base_url, stdlib_fetch, max_pages=100, seed_urls=<the profile sitemap's URLs>) (seeding with the sitemap is what lets a page nothing links to enter the graph at all; no sitemap means orphan detection is skipped and noted, never guessed), then analyze(graph). Checks, each in the standard falsifiable signal form:
   1. Broken internal links (analyze()['broken']): a linked target answering 4xx or unreachable -> P2 signal naming the source page, the target, and the status. From the clearest case, ONE gated fix proposal per run via onsite-propose: rewrite the source link to the correct target, or a redirect where the target moved (the redirect action type in skills/onsite-apply). Links resolving through a 3xx hop (redirect_chains) ride along as evidence in the same signal lane; rewriting the link to the final target is the usual fix.
   2. Orphan pages (analyze()['orphans']): in the sitemap seed but zero inbound internal links -> P3 signal recommending internal links from the top hubs (analyze()['hubs'] names them). Where an orphan also sits on the latest weekly striking-distance list (skills/hoo-weekly), note the linkage in the signal - the two findings share one fix.
   3. Shallow striking-distance pages: fewer than 2 inbound links AND on the latest weekly striking-distance list -> P2 signal. This is the highest-leverage internal-link play: the query already sits at position 4-15 and links from the hubs are the one lever fully in our hands. Falsifiability line, verbatim in the signal: "we are wrong if 2+ new hub links do not move the query's position within 28 days."
5. If CMS credentials exist: pull the post via the CMS adapter (WordPress today) with get_post + get_rendered_head for the rendered truth; list the SEO meta field values (RankMath fields on WordPress).
6. Launch technical-seo-auditor for site-level context when auditing > 3 URLs.
7. Output: per-URL scorecard table + prioritized issue list. File signals for P0/P1 issues, plus the page-essentials signals from step 3 and the link-health signals from step 4, if a brain repo exists. Propose nothing here; that is onsite-propose's job (the one broken-link fix in step 4 included - the audit only names it).

At each stage boundary, append a one-line progress marker with a UTC timestamp to the run report file before starting the stage - headless runs are watched by tailing that file, not a terminal.