Suggest image tags from entail.dev: client and admin endpoint (SONA-220, PR 1)

AI_POLICY.md

@@ -24,6 +24,14 @@ second AI model reads the first one's work, and CodeRabbit, a third-party
 service, reviews the diff. A human approves every merge and clicks every
 deploy. There is no review team here, just one maintainer and a set of tools.
 
+The running site makes one AI call of its own. When the operator asks for tag
+suggestions on a piece of artwork, the site sends entail.dev, an image
+classifier, a public link to the artwork's source post or to the picture in
+that post. In normal operation, nothing a visitor does reaches an AI service.
+The one exception is the diagnostic case the /ai page describes: when the
+maintainer is debugging, logs that can include visitor IP addresses and page
+URLs may pass through the development tools.
+
 ## Rules the agents work under
 
 These rules predate this file, and every agent session on this project works

UPDATING.md

@@ -48,6 +48,31 @@ you sync — or to catch up on what shipped since you last did — read the
 the merged changes since the previous one. `git log --oneline <last-tag>..upstream/main`
 after a fetch gives the same view from a clone.
 
+## Read before upgrading: tag suggestions call entail.dev (SONA-220)
+
+This release adds an admin-only tag suggestion lookup for artwork. Only you can
+start it, and nothing runs on its own. When you do, the site sends a public
+link to entail.dev, an image classifier, and shows you the tags it returns. For a
+Bluesky post that link is the post itself. For an X post the site first asks X's
+own API which picture the post carries, then sends the picture link X hands back.
+Nothing from either reply is stored; the suggestions are yours to accept or drop.
+
+The built-in privacy policy and the `/ai` page describe this call already. If
+you pasted your own privacy text or your own `/ai` text in Settings, **neither
+was updated**, and your pages will not mention services your site now contacts.
+In your privacy policy's list of feature integrations, replace your X and
+Bluesky entry with the first line below and add the second:
+
+- `X (formerly Twitter) and Bluesky (fetching the profile pictures shown on this site; X is also asked which picture a post carries)`
+- `entail.dev (an image classifier that suggests tags for artwork from its source post or the picture in it)`
+
+And add this to your `/ai` text: "The site calls an AI service in one place. When
+the site owner asks for tag suggestions on a piece of artwork, the site sends a
+public link to entail.dev, an image classifier. That link points at the artwork's
+source post, or at the picture in that post. For a post on X, the site first asks
+X's own service which picture the post carries. Only the site owner can start
+that, so nothing you do is sent to an AI service as you browse."
+
 ## One-time backfill: sticker animation flags (SONA-123)
 
 The release that adds the per-sticker download-format menu also adds a

docs/architecture.md

@@ -45,7 +45,8 @@ graph TB
         Resend[✉️ Resend]
         Turnstile[🧩 Cloudflare Turnstile]
         ConsFYI[📅 cons.fyi]
-        Avatars[🖼️ Bluesky + X — profile picture sources]
+        Avatars[🖼️ Bluesky + X — profile pictures, tweet media]
+        Entail[🏷️ entail.dev — image tag classifier]
         UT[☁️ UploadThing — optional]
     end
 
@@ -91,14 +92,15 @@ graph TB
     Auth -->|reset email| Resend
     RateLimit --> Turnstile
     Public -->|convention dates| ConsFYI
+    API -->|tag suggestions for a source post| Entail
 
     RegClient -->|search / pull / submit| RegWorker
     RegWorker --> RegD1
 
     CI --> Deploy
     Deploy -->|wrangler pages deploy| Hooks
     CronWF -->|POST /api/cron/* with CRON_SECRET| API
-    API -->|fetch profile pictures to re-host| Avatars
+    API -->|fetch profile pictures to re-host, resolve tweet media| Avatars
     Admin -->|fetch profile pictures to re-host| Avatars
     Release -.->|pull tagged releases| Forks
 ```
@@ -133,6 +135,11 @@ graph TB
   zone rather than the reader's or UTC.
 - Telegram, FurTrack, FuzzySearch, Resend, and Turnstile are optional
   integrations, keyed off secrets or settings (see `wrangler.toml.example` for the full list).
+- entail.dev needs no key or secret. The app calls it only when an operator
+  asks for tag suggestions on an image whose source post is on Bluesky or X,
+  and never on a render path or a schedule. An X post takes one extra hop:
+  X's own API resolves the post to its image, and that image URL is what
+  entail.dev classifies.
 - GitHub Actions is part of the runtime, not just delivery: the scheduled
   workflows (`sticker-resync` daily 06:00 UTC, `artist-sync` 06:30,
   `avatar-refresh` 07:00, `cleanup-orphans` weekly, `backfill-animated`

src/lib/ai-disclosure.test.ts

@@ -33,7 +33,14 @@ describe('defaultAiDisclosure', () => {
 	});
 
 	it('states the runtime boundary and the dev-time access plainly', () => {
-		expect(all).toMatch(/never calls an AI service, so nothing you do is sent to one as you browse/);
+		// SONA-220 added the one runtime call; the disclosure names it and keeps
+		// the browsing-time claim scoped so it stays true.
+		expect(all).toMatch(/calls an AI service in one place/);
+		expect(all).toContain('entail.dev');
+		// Both paths: the post link (Bluesky) and the picture link X hands back.
+		expect(all).toMatch(/points at the artwork's source post, or at the picture in that post/);
+		expect(all).toMatch(/the site first asks X's own service which picture the post carries/);
+		expect(all).toMatch(/Only the site owner can start that, so nothing you do is sent to an AI service as you browse/);
 		expect(all).toMatch(/logs and database/);
 		expect(all).toContain('CodeRabbit');
 		// Honesty about what dev-time log access can expose: no "your data never

src/lib/ai-disclosure.ts

@@ -2,8 +2,9 @@
 //
 // Wording approved by the operator 2026-08-12 after a peer review that set the
 // register: short, disclosure-only, no persuasion. Facts only — every claim
-// here is backed by the codebase (no runtime AI calls anywhere), the repo
-// policy (AI_POLICY.md), or the operator's own practice. An owner whose
+// here is backed by the codebase (the one runtime AI call is the operator's
+// tag-suggestion lookup against entail.dev, SONA-220), the repo policy
+// (AI_POLICY.md), or the operator's own practice. An owner whose
 // practice differs (or who wants their own words) overrides the text via
 // Settings, or turns the page off entirely with the aiPageEnabled toggle.
 //
@@ -70,7 +71,7 @@ export function defaultAiDisclosure(): AiDisclosure {
 			},
 			{
 				lead: 'Your data.',
-				body: "The running site never calls an AI service, so nothing you do is sent to one as you browse. When the software is being worked on, the developer's tools can read this site's logs and database, as any developer's could, and those logs can include visitors' IP addresses and the pages they requested. Code goes to Anthropic and to CodeRabbit, a review service. Model training is switched off on the accounts used, and CodeRabbit states that the data from its reviews is never used for training. The privacy policy has the details."
+				body: "The site calls an AI service in one place. When the site owner asks for tag suggestions on a piece of artwork, the site sends a public link to entail.dev, an image classifier. That link points at the artwork's source post, or at the picture in that post. For a post on X, the site first asks X's own service which picture the post carries. Only the site owner can start that, so nothing you do is sent to an AI service as you browse. When the software is being worked on, the developer's tools can read this site's logs and database, as any developer's could, and those logs can include visitors' IP addresses and the pages they requested. Code goes to Anthropic and to CodeRabbit, a review service. Model training is switched off on the accounts used, and CodeRabbit states that the data from its reviews is never used for training. The privacy policy has the details."
 			},
 			{
 				lead: 'The model.',

src/lib/legal.test.ts

@@ -104,7 +104,7 @@ describe('defaultPrivacyPolicy', () => {
 		// The runtime boundary, honestly scoped: nothing browsing-time goes to the
 		// tools, but shared diagnostic logs can carry request data — both halves
 		// must stay, or the paragraph overclaims again.
-		expect(text).toMatch(/nothing you do here is sent to them as you browse/);
+		expect(text).toMatch(/nothing you do here is sent to those tools as you browse/);
 		expect(text).toMatch(/can contain request data such as IP addresses/);
 	});
 
@@ -180,7 +180,10 @@ describe('defaultPrivacyPolicy', () => {
 		expect(text).toContain('cons.fyi');
 		// The integrations list reads exhaustive, so it must actually be: every
 		// remote service a feature calls out to is named (SONA-167 round 1).
-		expect(text).toContain('Bluesky');
+		// Picture lookup is an X-only ask; Bluesky posts go to entail.dev whole.
+		expect(text).toMatch(/X \(formerly Twitter\) and Bluesky \(fetching the profile pictures shown on this site; X is also asked which picture a post carries\)/);
+		// SONA-220: the tag-suggestion lookup sends a post or picture URL to entail.dev.
+		expect(text).toMatch(/entail.dev \(an image classifier that suggests tags for artwork from its source post or the picture in it\)/);
 		expect(text).toContain('FurTrack');
 		expect(text).toContain('FuzzySearch');
 		// The lookup sends a copy of the image, and an unpublished one too; naming
@@ -297,8 +300,8 @@ describe('LEGAL_DEFAULTS_UPDATED tracks the default text', () => {
 	// privacy page would show a "Last updated" line older than its own text.
 	// Deliberately two assertions, not a diff — the point is to force the date
 	// bump, not to review the prose.
-	const RECORDED_TEXT_HASH = '3a1187cb2970888de509560a9e0c028b3340d5ba39f393da83e0c943b407a780';
-	const RECORDED_UPDATED = '2026-09-14';
+	const RECORDED_TEXT_HASH = 'b2c656867b76ec288a85b230ccce8279bdff6cd963edea4e90e1b31f0c7351fa';
+	const RECORDED_UPDATED = '2026-09-15';
 
 	function defaultsText(): string {
 		// Fixed opts so the hash depends on the prose alone, not the caller. Both

src/lib/legal.ts

@@ -27,7 +27,7 @@ export interface LegalSection {
 // on every fork by construction (a build/deploy date would falsely advance on a
 // redeploy that didn't touch the text). Bump this whenever you edit
 // defaultPrivacyPolicy or defaultTerms.
-export const LEGAL_DEFAULTS_UPDATED = '2026-09-14';
+export const LEGAL_DEFAULTS_UPDATED = '2026-09-15';
 
 /**
  * Resolve the "Last updated" date to show on a legal page from a *stable* source
@@ -150,13 +150,13 @@ export function defaultPrivacyPolicy(opts: LegalOptions): LegalSection[] {
 				// paragraph with the toggle would delete a real processor
 				// disclosure. Only the vendor NAMES follow the affirmation, since
 				// those are the part a declining owner has not stood behind.
-				'Sites running this software are typically built and maintained with AI development tools, which do not run as part of the site itself, so nothing you do here is sent to them as you browse. When the site owner or their developer is diagnosing a problem, the operational data they share with development or code-review tools can include server logs and database records, and those logs can contain request data such as IP addresses, page URLs, and browser user-agent strings.',
+				'Sites running this software are typically built and maintained with AI development tools, which do not run as part of the site itself, so nothing you do here is sent to those tools as you browse. When the site owner or their developer is diagnosing a problem, the operational data they share with development or code-review tools can include server logs and database records, and those logs can contain request data such as IP addresses, page URLs, and browser user-agent strings.',
 				...(opts.aiToolsDisclosed === false
 					? []
 					: [
 							"For this site those tools are Anthropic's Claude, which writes and debugs code under the developer's direction, and CodeRabbit, a code review service that reads proposed changes."
 						]),
-				"For specific features the site also talks to Cloudflare Turnstile (bot protection on the sign-in page), Telegram (importing sticker packs), cons.fyi (convention listings), X (formerly Twitter) and Bluesky (fetching the profile pictures shown on this site), FurTrack (importing fursuit photos), FuzzySearch (finding where a piece was posted; when the site owner asks for a lookup, a copy of that image is sent, including images not published on this site), and the shared artist registry (syncing artist credits; the registry receives this site's name and hostname as part of the sync). The site contacts these services to run the feature; they are not used to track visitors."
+				"For specific features the site also talks to Cloudflare Turnstile (bot protection on the sign-in page), Telegram (importing sticker packs), cons.fyi (convention listings), X (formerly Twitter) and Bluesky (fetching the profile pictures shown on this site; X is also asked which picture a post carries), entail.dev (an image classifier that suggests tags for artwork from its source post or the picture in it), FurTrack (importing fursuit photos), FuzzySearch (finding where a piece was posted; when the site owner asks for a lookup, a copy of that image is sent, including images not published on this site), and the shared artist registry (syncing artist credits; the registry receives this site's name and hostname as part of the sync). The site contacts these services to run the feature; they are not used to track visitors."
 			]
 		},
 		{