/seo-meta-onpage

name: seo-meta-onpage
description: Audit and generate the document head — title and meta description length against the shared bands (title pass 30-60 chars, recommended 50-60; description pass 70-160, recommended 150-160), viewport, charset, <html lang>, and robots-meta sanity — flagging missing/duplicate/out-of-band values and producing length-bounded, keyword-aware replacements. Module M7. Feeds the Search SEO score.
allowed-tools: Read, Grep, Glob, WebFetch, Bash

<!-- Portions Copyright (c) 2026 Enrique Rocha (@soyenriquerocha) - see LICENSE for original authors · tododeia.com · id er-d03758e1 -->

seo-meta-onpage (M7)

Head hygiene is cheap, deterministic, and the first thing every crawler reads. Reference: `references/schema-tier1.md` for the structured-data layer that sits alongside the head (owned by M5).

Inputs

Work from the PageSnapshot named in your dispatch envelope: read `parsed` from `<run_dir>/pages/<slug>.json` (`title`, `metas[]`, `robots_meta[]`, `canonicals[]`, `word_count`) plus `headers` (`content-language`, `x-robots-tag`); Grep `pages/<slug>.html` for verbatim evidence; site artifacts live in `<run_dir>/site/{robots.json,sitemaps.json,discovery.json}`. Deterministic findings already emitted by `audit.mjs` are listed in `<run_dir>/findings.deterministic.json` — do not re-emit those ids; add model-judged findings only. If invoked directly with a URL/path and no snapshot exists, first run `node "${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.mjs" <target> --out "${CLAUDE_PLUGIN_DATA}/runs"` and use the printed snapshot path.

Audits

Working from the PageSnapshot (`parsed_rendered` when `render.used` is not `none`, else `parsed`): 1. **Title** — exactly one `<title>`. Length is judged against the **shared bands** in `scripts/lib/bands.mjs` (the single source; `parse-html.mjs`, the on-page checks and this skill all read it): **pass 30-60 characters**, **recommended 50-60**. Inside 50-60 the verdict is `recommended`; inside 30-60 but outside that, `pass`; below 30 `short`, above 60 `long` — both are `warn`, never `fail`. **Fail** only when the title is missing or empty, or duplicated site-wide across distinct URLs. Not all caps; not a pure keyword list. 2. **Meta description** — present, judged against the same shared bands: **pass 70-160 characters**, **recommended 150-160**. Outside the pass band is `warn`, never `fail` — Google rewrites descriptions freely. Not duplicated site-wide; describes this page, not boilerplate. 3. **Viewport** — `<meta name="viewport" content="width=device-width, initial-scale=1">` present (fail if missing — mobile rendering/indexing depends on it). 4. **Charset** — `<meta charset="utf-8">` present and in the first 1024 bytes of `<head>`. 5. **Language** — `<html lang="…">` present and a valid BCP-47 tag. 6. **Robots meta** — sanity-check `<meta name="robots">`: flag an accidental `noindex`/`nofollow` on a page meant to rank. Full indexability/canonical logic (including the canonical tag) is owned by **seo-indexability (M2)** — here only **note presence** of `<link rel="canonical">`, do not adjudicate it.

Fixes

• **AUTO** — add missing `viewport`, `charset` (utf-8), and `<html lang>` (inferred from `Content-Language` header or page text; leave a TODO if ambiguous). These are deterministic, additive, verifiable writes → diff for `fix`.
• **PROPOSED** — generate/trim a `<title>` and `meta description` from `<h1>`, lead paragraph, and primary topic: length-bounded, keyword-aware but **not** stuffed, brand suffix only if the site uses one. Each is a draft requiring per-item accept (humans own messaging).
• **ADVISORY** — removing an intentional `noindex` is never auto-written; surface it and let the user decide.
• **Never fabricate** a title, description, or `lang` — when content is too thin to derive a value, leave a clearly-marked `TODO:` placeholder for the user to fill.

Verification

• Method `dom_assert`: `node "${CLAUDE_PLUGIN_ROOT}/scripts/parse-html.mjs" --snapshot <pages/<slug>.json>` (or `--url <u>` / `--file <path>`) — returns `title` (value/length/ok), `meta_description` (value/length/ok), `robots_meta`, `canonical`, and `head_hygiene` (viewport/charset/lang) so each head element's presence and length is asserted directly against the snapshot.
• Duplicate-title/description checks need the full crawl set; when only a single page is available the cross-URL assertion is reported `needs_api`, **never** a false `pass`.

Findings

Findings conform to `schema/finding.schema.json`. **Severity policy**: 5 is reserved for catastrophic, eligibility-killing facts at site/template scope; 4 major · 3 moderate · 2 minor · 1 cosmetic · 0 informational — and only an `established` severity-5 `fail` in an active category can cap a score (`references/scoring-model.md`). Head hygiene never reaches 5. Examples:

• `M7.title.missing` — no `<title>` (fail, severity 3, `fixable: proposed`, axis `search`, confidence `established`).
• `M7.title.length_out_of_band` — title 78 chars, outside the 30-60 pass band and likely truncated in the SERP (warn, severity 3, `fixable: proposed`, axis `search`, confidence `directional`). Quote the length and the band that produced the verdict.
• `M7.description.length_out_of_band` — description outside the 70-160 pass band (warn, severity 2, `fixable: proposed`, axis `search`, confidence `directional`).
• `M7.viewport.missing` — no responsive viewport (fail, severity 3, `fixable: auto`, axis `search`, confidence `established`).
• `M7.title.duplicate_across_urls` — the same `<title>` is served from several sampled URLs (warn, severity 3, `fixable: proposed`, axis `search`, confidence `directional`). One finding per duplicate cluster, quoting the shared title and the URLs.
• `M7.description.duplicate_across_urls` — the same meta description is served from several sampled URLs (warn, severity 2, `fixable: proposed`, axis `search`, confidence `directional`).

Both duplicate checks are cross-URL: with fewer than two