Beacon

Reconciles the two release lines. The 1.2.x releases were tagged off the internal-links branch while main carried the GEO/LLM work; both are now on main. Sites on 1.1.x gain the internal-links feature, sites on 1.2.x gain the GEO/LLM feature set, and both gain the performance work below.

Added

• Internal links and the GEO/LLM feature set are now available in a single release — see the 1.2.0 and 1.1.0 entries for their details.

Changed

• Beacon no longer queries on every front-end request: the settings, schema, AI-bot and site-settings rows it reads on each page are cached across requests and dropped by tag when the underlying record is written.
• beacon_render_cache.contentKey is now NOT NULL (empty string for a type's master document), which lets writes become a single upsert and closes a duplicate-row window — both databases treated NULL keys as distinct in the unique index, so master rows were never covered by it.
• GEO rescoring targets are coalesced per request and processed in batches of 100 rather than one queue job per element.
• 404s no longer pay for a short-link lookup on sites that have no short links.

Fixed

• The AI provider key is no longer written to the on-disk cache in the clear; it is encrypted with the app security key.
• Cache-tag invalidation now also fires after the element-save transaction commits, so a concurrent request can no longer re-cache a pre-write value against the bumped tag — which could leave short links unresolvable or settings stale indefinitely.
• A value built while a write lands mid-rebuild is no longer stored as fresh.
• Cached record payloads carry a 5-minute ceiling, bounding divergence on multi-server deployments where tag invalidation only reaches one node.
• Buffered GEO rescoring targets survive a request that dies before its after-request event, and drain after every queue job rather than waiting for a long-lived worker to exit.
• A failing element in a batched GEO rescore no longer abandons the rest of the batch.
• The Sitemap Health widget labels an unchunked site's sitemap (single) again instead of rendering a blank cell.

Fixed

• PHPStan: use Http::request() for JSON detection in RequiresLinksEnabledTrait.
• PHPStan: add Action<Controller> generic to LinksController::beforeAction().

Added

• Internal links feature: TF-IDF keyword index, optional AI-embedding suggestions, CP reports (overview, orphans, link map, link detail, click depth, broken links, anchor text, external links), entry-editor sidebar, dashboard widget, craft.beacon.links.* Twig API, and console commands (beacon/link-index, link-snapshot, link-report, link-audit).
• Master Enable internal links toggle on Links → Settings — disables indexing, sidebar, reports, Twig helpers, and console (except link-index/clear).
• Per-suggestion Highlight buttons in the entry sidebar, mutually exclusive with highlight-all.

Fixed

• SSRF guard for IPv6-only hosts in broken-link audit (#9).
• Protocol-relative URL classification in link parser (#10).
• AI content controller error handling (#11).
• Rate limiter TTL reset for active editors (#12).
• robots.txt cache invalidation after AI-usage policy changes (#13).
• Citation/competitor substring false positives (#14).
• Avg-links KPI denominator consistency (#15).
• URI resolution when site URL substring repeats (#16).
• Shared index pipeline via IndexesEntries trait (#17).
• Atomic Db::upsert in suggestion recordInteraction() (#18).

Added

• GEO/LLM feature set: AI-assisted content generation, AI-visibility citation tracking across benchmark prompts, licensing/usage signals, token-aware llms-full.txt, and entity linking.
• Read/write MCP tool surface (beacon.*) for redirects, 404s, llms.txt, entry SEO/meta, and GEO scores, gated by Beacon permissions.

Security

• MCP entry read tools (get_entry_seo, get_geo_score) now enforce element-level canView authorization in addition to the dashboard permission.
• set_entry_meta requires an authenticated token user with canSave on the target entry and declares an explicit edit permission.
• AI-visibility benchmark prompts are now site-scoped on save and delete, preventing cross-site edits/deletions.

Changed

• Capped AI-visibility probes per run at 100 to stay within queue-job time limits.