seo-orchestrator (Layer 2)

Coordinates the whole audit in three phases: detect → dispatch → synthesize. The scripts are the acquisition layer; the model only judges. Read the compact stdout summaries and the JSON files they name — never paste HTML into the conversation. Nothing is written to the user's project; every artifact lives under ${CLAUDE_PLUGIN_DATA}/runs/<host>/<run-id>/.

1. Detect — one call

1. Resolve the target and the flags forwarded by the command (--pages, --max, --render, --ua, --vertical, --environment, --feed, --out; default --out "${CLAUDE_PLUGIN_DATA}/runs").
2. Run the whole deterministic pipeline in one command. audit.mjs acquires the target (it calls crawl.mjs for a URL, snapshot.mjs for --pages 1 or a local path, and skips acquisition entirely when handed a run directory), writes profile.json (platform / framework / cms plugins / hosting / environment / capabilities / write_targets / vertical_hints, plus a vertical guess), runs the checks registry, and calls report.mjs for a first deterministic-only report:

node "${CLAUDE_PLUGIN_ROOT}/scripts/audit.mjs" <url|path> \
  --out "${CLAUDE_PLUGIN_DATA}/runs" --checks deterministic --format json \
  [--pages N] [--max N] [--render static|auto|js] [--ua <preset>] \
  [--vertical <ids>] [--environment production|preview|staging|local] [--feed <path>]

Read only the JSON summary it prints: run_dir, mode, platform, vertical, coverage, tier, scores, findings, checks, report_json, report_md, findings_json, warnings. --render defaults to static here — pass --render auto when the target looks client-rendered. Exit 2 means the target could not be acquired; exit 3 only means a --fail-under / --fail-on-* gate tripped, which is a CI concern, not a failed audit. 3. Read <run_dir>/crawl.json (pages with slug/url/role, templates, sampling, warnings) and <run_dir>/profile.json. 4. Vertical: run seo-vertical-detect over the homepage parsed (<run_dir>/pages/<homepage-slug>.json) plus profile.vertical_hints. The script only guesses — profile.vertical.source is "inferred" unless --vertical was passed. If your reading adds or changes a vertical, re-run the deterministic pass over the same run (no re-crawl) so the conditional checks actually fire:

node "${CLAUDE_PLUGIN_ROOT}/scripts/audit.mjs" <run_dir> --checks deterministic --format json --vertical <ids>

references/routing.md maps vertical → conditional modules (M18 e-commerce, M19 local, M20 hreflang on multilingual). 5. The run now holds <run_dir>/findings.deterministic.json (the deterministic findings the agents must not re-emit) and <run_dir>/checks.json (checks run, errors, needs_api, manual_review, dropped).

2. Dispatch (parallel — one message, four Agent calls)

Spawn the four read-only specialists in one message so their verbose intermediate output stays isolated. Each prompt = the envelope block + that agent's module list. Agents never rely on ${…} substitution: pass absolute paths.

ENVELOPE
plugin_root: <absolute ${CLAUDE_PLUGIN_ROOT}>
run_dir: <absolute run dir>
pages: [{slug, url, role}, …]              # <run_dir>/pages/<slug>.json (+ <slug>.html, + <slug>.rendered.html when rendered)
site: robots=<run_dir>/site/robots.json  sitemaps=<run_dir>/site/sitemaps.json  discovery=<run_dir>/site/discovery.json
vertical: {primary: <v>, also: [<v>…], multilingual: <bool>}
platform: <run_dir>/profile.json           # one line: <platform>/<framework>, env=<kind>, head_owner=<…>
platform_cards: [<plugin_root>/references/platforms/<id>.md, …]   # omit when the platform is unknown
modules: [<M-ids for this agent>]
deterministic_findings: <run_dir>/findings.deterministic.json — do not re-emit these ids; add model-judged findings only
return: JSON array only — findings per schema/finding.schema.json, no prose

Subagents have no Write/Edit — the audit can never mutate files. If an agent returns prose around the array, keep only the array.

3. Synthesize

1. Save each returned array verbatim to <run_dir>/agents/<agent>.json via a Bash heredoc (mkdir -p "<run_dir>/agents" && cat > "<run_dir>/agents/<agent>.json" <<'EOF' … EOF) — into the plugin data dir, never the project. This is the one Bash call outside the node pattern; the mkdir -p / cat > entries in allowed-tools pre-approve it for paths under ${CLAUDE_PLUGIN_DATA}/runs/ (TO-VERIFY in the smoke — a heredoc body containing ;, | or && may still prompt).
2. node "${CLAUDE_PLUGIN_ROOT}/scripts/report.mjs" <run_dir> --merge "agents/*.json" --lang <en|es> [--vertical <ids>] [--environment <kind>] [--out-md <path>] — re-reads findings.deterministic.json, merges the agent arrays (dedupe by id + normalized location, most severe wins; needs_api / manual_review never override a scored status), drops schema-invalid findings into report.dropped_findings, scores per page and as a site rollup, and rewrites findings.json, report.json, report.md. Pass the same --vertical / --environment you used in step 1. --merge is repeatable and accepts a file, a directory, or a */? glob on the file name only.
3. Present from report.json (bands/interpretations per references/scoring-model.md):
   • Search SEO and AI Visibility — band + score + one-line interpretation, never blended; per-category table for each (value, weight, active).
   • Coverage: coverage.mode (full vs deterministic), tier reached, needs_api / manual_review counts.
   • Sampling table from crawl.json: templates discovered vs sampled, pages by role, every skip and its reason.
   • Platform profile line (platform/framework/plugins, environment; "expected on preview" notes when non-production).
   • Top actions sorted by impact ÷ effort: status, evidence, recommendation, fixability (auto/proposed/advisory), expected_impact.
   • The absolute report.md path, then offers: /claude-seo-ai:fix <target> (safe AUTO fixes, confirmed per change) and /claude-seo-ai:compare --baseline latest --against <target> after changes.

Degraded mode

If node is missing, or a skills-only install ships no scripts/ directory: say so first. Fall back to WebFetch summaries only; mark every header/status/render/robots-dependent finding needs_api; skip persistence; label the report "prompt-only mode — not comparable to a scripted run". Never present a WebFetch summary as raw HTML or as a measured status/header.