distribea-mcp 1.2.0 → 1.3.1

package/README.md

@@ -58,4 +58,27 @@ A [Distribea](https://distribea.com) subscription is required.
 - Your project files never leave your machine except the specific images you
   ask to edit.
 
+## Privacy Policy
+
+This local connector contains **no API keys and no secrets**. To generate an
+image it sends to the hosted Distribea engine only what is needed for the
+request:
+
+- **What is sent:** your image brief (the subject/prompt you ask for), a short
+  text excerpt of the page being worked on (used to infer the site's style and
+  context), a hashed project identifier, and any image file you explicitly ask
+  to edit. **Your other project files stay on your machine** — they are never
+  read or uploaded.
+- **What is stored, tied to your account:** your locked visual style, recurring
+  characters/products/avatars, the images generated for you (hosted on
+  Distribea's CDN), and your credit/billing usage.
+- **Third parties:** images are produced through Distribea's hosted AI
+  providers solely to fulfil your request. Your data is **never sold**.
+- **Authentication:** calls are authenticated with your personal key
+  (`dmcp_…`), issued and revocable at
+  <https://distribea.com/account/mcp>.
+- **Retention & contact:** data is kept for the life of your account; full
+  policy, retention details and contact at
+  <https://distribea.com/legal/privacy-policy>.
+
 © Distribea, <https://distribea.com>

package/index.mjs

@@ -16,7 +16,7 @@
 // ---------------------------------------------------------------------------
 
 import { createHash } from "node:crypto";
-import { readFileSync, readdirSync, statSync } from "node:fs";
+import { readdirSync, readFileSync, statSync } from "node:fs";
 import { mkdir, readFile, writeFile } from "node:fs/promises";
 import { createRequire } from "node:module";
 import {
@@ -229,7 +229,75 @@ const SKIP_DIRS = new Set([
 const PLACEHOLDER_SRC_RE =
 	/placehold\.co|via\.placeholder\.com|placekitten|picsum\.photos|dummyimage\.com|loremflickr\.com|placeimg\.com|fakeimg\.pl|images\.unsplash\.com|source\.unsplash\.com|images\.pexels\.com|cdn\.pixabay\.com|placeholder/i;
 const IMG_TAG_RE = /<(?:img|Image)\b[\s\S]*?>/g;
-const HEADING_RE = /<h[1-3][^>]*>([\s\S]*?)<\/h[1-3]>/g;
+const HEADING_SCAN_RE = /<h[1-4][^>]*>([\s\S]*?)<\/h[1-4]>/gi;
+const NEXT_HEADING_RE = /<h[1-4][^>]*>/i;
+// Frontières de bloc : on ne relie JAMAIS une image au titre d'une autre section.
+const BLOCK_BOUNDARY_RE =
+	/<\/?(?:section|article|header|footer|main|nav|aside)\b[^>]*>/gi;
+
+// Le « bloc » qui contient l'image : entre la frontière de section juste avant
+// et juste après. Page mal codée (aucune balise de section) → fenêtre bornée
+// pour ne pas aller chercher un titre à l'autre bout du fichier.
+function sectionBoundsForImg(content, imgStart, imgEnd) {
+	let lo = 0;
+	let hi = content.length;
+	for (const b of content.matchAll(BLOCK_BOUNDARY_RE)) {
+		if (b.index < imgStart) {
+			lo = b.index + b[0].length;
+		} else if (b.index >= imgEnd) {
+			hi = b.index;
+			break;
+		}
+	}
+	if (lo === 0 && hi === content.length) {
+		lo = Math.max(0, imgStart - 1200);
+		hi = Math.min(content.length, imgEnd + 1200);
+	}
+	return { lo, hi };
+}
+
+// Titre + description du MÊME bloc que l'image : on prend le titre le plus
+// proche, qu'il soit AU-DESSUS ou EN DESSOUS (cartes image-en-haut/titre-dessous,
+// blocs 2 colonnes…), sans déborder sur la carte voisine. Aucun titre trouvé →
+// on retombe sur le texte juste avant l'image.
+function headingAndContextForImg(content, imgStart, imgEnd) {
+	const { lo, hi } = sectionBoundsForImg(content, imgStart, imgEnd);
+	const scope = content.slice(lo, hi);
+	const relStart = imgStart - lo;
+	const relEnd = imgEnd - lo;
+	let best = null;
+	let bestDist = Number.POSITIVE_INFINITY;
+	for (const h of scope.matchAll(HEADING_SCAN_RE)) {
+		const hStart = h.index;
+		const hEnd = h.index + h[0].length;
+		let dist = 0;
+		if (hEnd <= relStart) {
+			dist = relStart - hEnd;
+		} else if (hStart >= relEnd) {
+			dist = hStart - relEnd;
+		}
+		if (dist < bestDist) {
+			bestDist = dist;
+			best = { text: stripMarkup(h[1]), hEnd };
+		}
+	}
+	if (!best) {
+		return {
+			heading: "",
+			context: stripMarkup(
+				content.slice(Math.max(0, imgStart - 300), imgStart)
+			).slice(-250),
+		};
+	}
+	// Description = le texte qui suit CE titre, borné au titre suivant (pour ne
+	// pas avaler la carte d'après).
+	const afterHead = scope.slice(best.hEnd);
+	const nextHead = afterHead.search(NEXT_HEADING_RE);
+	const desc = stripMarkup(
+		nextHead >= 0 ? afterHead.slice(0, nextHead) : afterHead
+	).slice(0, 280);
+	return { heading: best.text, context: desc };
+}
 
 function walkFiles(dir, out = []) {
 	for (const name of readdirSync(dir)) {
@@ -309,15 +377,18 @@ function scanFileForSlots(file, content) {
 				h = Number(dimM[2]);
 			}
 		}
-		const before = content.slice(Math.max(0, m.index - 700), m.index);
-		const headM = [...before.matchAll(HEADING_RE)].pop();
+		const { heading, context } = headingAndContextForImg(
+			content,
+			m.index,
+			m.index + tag.length
+		);
 		slots.push({
 			file,
 			tag,
 			src: srcM[2],
 			alt: altM?.[2] ?? "",
-			heading: headM ? stripMarkup(headM[1]) : "",
-			context: stripMarkup(before).slice(-250),
+			heading,
+			context,
 			// Le prénom de l'auteur d'un avis est presque toujours SOUS sa photo.
 			after: stripMarkup(
 				content.slice(m.index + tag.length, m.index + tag.length + 500)
@@ -820,6 +891,130 @@ async function runGenerateImage(args, progress) {
 		.join("\n");
 }
 
+// --- blog cover ---------------------------------------------------------------
+const ARTICLE_UA =
+	"Mozilla/5.0 (compatible; DistribeaImages/1.0; +https://distribea.com)";
+
+// Résout le contenu de l'article : fichier local > URL publique > texte collé.
+async function resolveArticle(projectDir, args) {
+	const file = args.article_file ?? args.article_path;
+	if (file) {
+		const p = resolveIn(projectDir, String(file));
+		let raw;
+		try {
+			raw = await readFile(p, "utf8");
+		} catch (e) {
+			throw new Error(`Article introuvable : ${p} (${e.message})`);
+		}
+		return { title: "", text: stripMarkup(raw).slice(0, 8000) };
+	}
+	if (args.article_url) {
+		const u = String(args.article_url).trim();
+		if (!/^https?:\/\//i.test(u)) {
+			throw new Error("article_url doit commencer par http:// ou https://");
+		}
+		let res;
+		try {
+			res = await fetch(u, {
+				headers: { "user-agent": ARTICLE_UA },
+				signal: AbortSignal.timeout(20_000),
+				redirect: "follow",
+			});
+		} catch (e) {
+			throw new Error(`Lecture de l'article impossible (${u}) : ${e.message}`);
+		}
+		if (!res.ok) {
+			throw new Error(
+				`Lecture de l'article impossible (HTTP ${res.status}) — ${u}`
+			);
+		}
+		const html = await res.text();
+		const title = oneLine(
+			html.match(/<title[^>]*>([\s\S]*?)<\/title>/i)?.[1] ?? ""
+		);
+		const stripped = html
+			.replace(/<script[\s\S]*?<\/script>/gi, " ")
+			.replace(/<style[\s\S]*?<\/style>/gi, " ");
+		return { title, text: stripMarkup(stripped).slice(0, 8000) };
+	}
+	const t = String(args.article_text ?? args.article ?? "").trim();
+	return { title: "", text: t.slice(0, 8000) };
+}
+
+async function runBlogCover(args, progress) {
+	const projectDir = resolveIn(
+		process.cwd(),
+		args.project_dir ?? process.cwd()
+	);
+	const resolved = await resolveArticle(projectDir, args);
+	const title = oneLine(args.title ?? "") || resolved.title || "";
+	const text = resolved.text;
+	if (!(title || text)) {
+		throw new Error(
+			"Donne l'article : article_text (texte collé), article_url (lien public) ou article_file (fichier local)."
+		);
+	}
+	const extra = Math.max(
+		0,
+		Math.min(5, Math.round(Number(args.illustrations ?? 0)))
+	);
+	const count = 1 + extra;
+	// Cover de blog : 16:9 ("wide") par défaut, standard. portrait/square au choix.
+	const orientation = ["wide", "landscape", "portrait", "square"].includes(
+		args.orientation
+	)
+		? args.orientation
+		: "wide";
+	const saveDir = args.save_dir
+		? resolveIn(projectDir, args.save_dir)
+		: join(projectDir, "public", "images");
+
+	progress?.(
+		`🎨 Cover de blog (${count} image${count > 1 ? "s" : ""}, ≈ ${IMAGE_CREDITS_HINT * count} crédits, 30-60 s)…`,
+		0,
+		1
+	);
+
+	const out = await engine("blog_cover", projectDir, {
+		title,
+		text,
+		count,
+		orientation,
+		character: args.character,
+		product: args.product,
+		pages_excerpt: pagesExcerpt(projectDir),
+		client_ref: "blog_cover",
+	});
+
+	const slugBase = slugify(title || text.slice(0, 60) || "article");
+	const images = Array.isArray(out.images) ? out.images : [];
+	const saved = await mapPool(images, 4, async (im, i) => {
+		const fileName =
+			im.role === "cover"
+				? `blog-${slugBase}-cover.webp`
+				: `blog-${slugBase}-${i}.webp`;
+		const outPath = join(saveDir, fileName);
+		await saveUrl(im.cdn_url, outPath);
+		return { ...im, fileName, outPath };
+	});
+
+	const lines = [
+		out.style_inferred ? STYLE_INFERRED_NOTE : "",
+		`Cover de blog générée ✔ (${out.credits} crédits)`,
+	];
+	for (const im of saved) {
+		lines.push(
+			"",
+			im.role === "cover" ? "🖼️ COVER" : "🖼️ illustration",
+			`file: ${im.outPath}`,
+			`size: ${im.width}×${im.height} — ${Math.round(im.bytes / 1024)} KB (WebP)`,
+			`alt: ${im.alt}`,
+			`<img src="/images/${im.fileName}" alt="${String(im.alt).replace(/"/g, "&quot;")}" width="${im.width}" height="${im.height}" loading="${im.role === "cover" ? "eager" : "lazy"}" />`
+		);
+	}
+	return lines.filter(Boolean).join("\n");
+}
+
 const OUT_FORMAT_BY_EXT = {
 	".png": "png",
 	".jpg": "jpeg",
@@ -1449,13 +1644,16 @@ function scanRebrandCandidates(projectDir) {
 				continue;
 			}
 			const altM = tag.match(/\balt\s*=\s*\{?\s*(["'])([\s\S]*?)\1/);
-			const before = content.slice(Math.max(0, m.index - 700), m.index);
-			const headM = [...before.matchAll(HEADING_RE)].pop();
+			const { heading } = headingAndContextForImg(
+				content,
+				m.index,
+				m.index + tag.length
+			);
 			const entry = seen.get(resolved) ?? {
 				path: resolved,
 				usedIn: new Set(),
 				alt: altM?.[2] ?? "",
-				heading: headM ? stripMarkup(headM[1]) : "",
+				heading,
 			};
 			entry.usedIn.add(relative(projectDir, file));
 			seen.set(resolved, entry);
@@ -1499,9 +1697,13 @@ async function runRebrandImages(args, progress) {
 		c.bytes = st.size;
 		candidates.push(c);
 	}
+	const eligibleBeyondCap = Math.max(0, candidates.length - max);
 	candidates.splice(max);
 
 	const notes = [
+		eligibleBeyondCap
+			? `+${eligibleBeyondCap} autre(s) image(s) au-delà de la limite (max_images=${max}) — monte max_images (jusqu'à 20) ou relance pour continuer.`
+			: "",
 		placeholders
 			? `${placeholders} placeholder(s)/stock détecté(s) aussi → make_images (sans rebrand) s'en charge.`
 			: "",
@@ -1673,7 +1875,7 @@ const TOOLS = [
 		title: "Generate one on-brand website image",
 		annotations: {
 			readOnlyHint: false,
-			destructiveHint: false,
+			destructiveHint: true,
 			openWorldHint: true,
 		},
 		description:
@@ -1720,6 +1922,70 @@ const TOOLS = [
 			required: ["subject"],
 		},
 	},
+	{
+		name: "blog_cover",
+		title: "Cover image (and optional illustrations) for a blog article",
+		annotations: {
+			readOnlyHint: false,
+			destructiveHint: true,
+			openWorldHint: true,
+		},
+		description:
+			"AUTONOMOUS TRIGGER — call this when an EDITORIAL ARTICLE still LACKS a cover/hero image (a brand-new post, or one left with a placeholder/stock image), even if the user never mentions an image. GUARDRAIL against noise: fire it ONCE per article, only when the cover is missing; if the article already has a real cover, leave it — do NOT regenerate; a pure text/typo/SEO edit is NOT a trigger; never re-run a successful call or 'improve' a cover on your own — only on the user's explicit request (each call bills the subscriber). Recognise the article by MEANING, not by tags: any standalone piece written to be READ — a headline + a body of prose on one topic (blog post, news, guide, tutorial, case study, magazine/journal piece) — in ANY language, on ANY stack (markdown/MDX, a headless CMS like WordPress/Sanity/Contentful, or a custom DB-driven route). Folders/routes (/blog, /posts, /articles, /news, /journal, /actualités, /magazine, /guides), markdown/MDX frontmatter, or an <article> element are only NON-EXHAUSTIVE hints — none is required: if it simply reads like an article, it qualifies (a marketing/landing/product page does NOT — that keeps make_images). You already hold the article you are working on → pass its own title + text as article_text (or article_url for a public link, or article_file for a local md/mdx/html/txt file). It generates the COVER, driven by the ARTICLE'S OWN specific subject (never a generic trade photo), kept visually coherent with the site's locked style (palette, light, medium), adapted to the site's country (driving side, architecture, currency), with NO recurring face/logo/product stamped on by default (every article gets a DIFFERENT scene — no 'same baker's face on every post'). Default = the cover ONLY at 16:9; set illustrations:N to also get N in-article images. Delivers an optimised WebP + ALT, ready to host. To DELIBERATELY feature a locked character/product, pass character/product. If the site has no locked style it infers one from the project; if it still lacks the trade/brand it tells you what to provide.",
+		inputSchema: {
+			type: "object",
+			properties: {
+				article_text: {
+					type: "string",
+					description: "The article's text (title + body), pasted directly",
+				},
+				article_url: {
+					type: "string",
+					description: "Public URL of the article — its text is read for you",
+				},
+				article_file: {
+					type: "string",
+					description:
+						"Local file to read (md/mdx/html/txt…), relative to the project",
+				},
+				title: {
+					type: "string",
+					description:
+						"Optional article title (overrides the one read from the source)",
+				},
+				illustrations: {
+					type: "number",
+					description:
+						"Extra in-article images beyond the cover (default 0, max 5)",
+				},
+				orientation: {
+					type: "string",
+					enum: ["wide", "portrait", "square"],
+					description:
+						"Cover shape — default 'wide' (16:9, the standard blog-cover ratio). Use portrait/square only if the layout needs it.",
+				},
+				character: {
+					type: "string",
+					description:
+						"Optional: name/role of a locked character to deliberately feature",
+				},
+				product: {
+					type: "string",
+					description:
+						"Optional: name of a locked product to deliberately feature",
+				},
+				save_dir: {
+					type: "string",
+					description:
+						"Directory to save the WebP into (default <project>/public/images)",
+				},
+				project_dir: {
+					type: "string",
+					description: "Project path (default: current directory)",
+				},
+			},
+		},
+	},
 	{
 		name: "edit_image",
 		title: "Retouch an image: edit, redo, cutout, upscale, extend",
@@ -1769,7 +2035,7 @@ const TOOLS = [
 			openWorldHint: true,
 		},
 		description:
-			"The site's art direction in one tool — action: 'setup' (lock the style from a short brief and/or the user's existing site_url; run FIRST on a new project, or let make_images infer it), 'refine' (plain-language feedback — « plus chaleureux » — becomes PERMANENT for every future image), 'lock_image' (« j'adore CELLE-LÀ, fais les autres pareil » — the approved image becomes the permanent style reference). Action inferred if omitted: image_path→lock_image, feedback→refine, otherwise setup. Free (optional moodboard billed as 1 image).",
+			"The site's art direction in one tool — action: 'setup' (lock the style from a short brief and/or the user's existing site_url; run FIRST on a new project, or let make_images infer it), 'refine' (plain-language feedback — « plus chaleureux », OR a CORRECTION of a wrong/off-topic depiction such as « un VTC est un chauffeur privé, jamais un taxi » or « montre toujours le pro au travail, pas un client » — the rule is RECORDED PERMANENTLY in the site's scene rules and obeyed by every future image, blog_cover included, so the same mistake is never reproduced), 'lock_image' (« j'adore CELLE-LÀ, fais les autres pareil » — the approved image becomes the permanent style reference). Action inferred if omitted: image_path→lock_image, feedback→refine, otherwise setup. Free (optional moodboard billed as 1 image).",
 		inputSchema: {
 			type: "object",
 			properties: {
@@ -1863,7 +2129,7 @@ const TOOLS = [
 			"Lock a recurring character, product or place (identical everywhere)",
 		annotations: {
 			readOnlyHint: false,
-			destructiveHint: false,
+			destructiveHint: true,
 			openWorldHint: true,
 		},
 		description:
@@ -1957,6 +2223,7 @@ const TOOLS = [
 const TOOL_RUNNERS = {
 	make_images: runMakeImages,
 	generate_image: runGenerateImage,
+	blog_cover: runBlogCover,
 	edit_image: runEditImage,
 	site_style: runSiteStyle,
 	brand_pack: runBrandPack,
@@ -1964,6 +2231,8 @@ const TOOL_RUNNERS = {
 	finish_images: runFinishImages,
 	pack_status: runPackStatus,
 	// Anciens noms — gardés en coulisse (compat tests / vieux clients).
+	blog_image: (a, p) => runBlogCover(a, p),
+	blog: (a, p) => runBlogCover(a, p),
 	setup_style: (a, p) => runSiteStyle({ ...a, action: "setup" }, p),
 	refine_style: (a, p) => runSiteStyle({ ...a, action: "refine" }, p),
 	lock_style_image: (a, p) => runSiteStyle({ ...a, action: "lock_image" }, p),
@@ -2054,8 +2323,10 @@ THE EASY PATH (prefer it): write the page with <img src="https://placehold.co/12
 Also:
 - User provides a person's photo ("c'est lui le professeur") → create_reference (kind character, photo_path) FIRST, so make_images/generate_image reuse that exact face in every scene. A product that must stay identical in every shot → create_reference (kind product). Never paste the raw photo into the page when a styled scene would serve better.
 - User gives style feedback ("plus chaleureux") → site_style (action refine): the change becomes permanent. User approves ONE image and wants the rest the SAME ("j'adore celle-là, fais les autres pareil") → site_style (action lock_image) with that file.
+- User points out the image got the SUBJECT/trade/role WRONG or off-topic (e.g. "a VTC is a private chauffeur, not a taxi", "show the baker working, not a customer", wrong product) → site_style (action refine) with that correction: it is recorded as a permanent scene rule and obeyed by EVERY future image (blog_cover included), so the same off-topic is never reproduced. Do this the moment a correction is given — never just regenerate and hope.
 - EXISTING site whose real images look cheap/stock/off-brand → make_images with rebrand:true: the first call lists every replaceable image for free, then apply:true rebrands them ALL in place (code untouched, originals kept as *.original).
 - REVIEW / TESTIMONIAL sections: just put a placeholder <img> next to each review and run make_images — the avatars come out as ultra-real casual smartphone selfies (UGC look: car, bedroom, living room…), NOT brand photos. Same reviewer name = same face everywhere on the site; a face is NEVER reused on another site. Never ship a review section without these avatars.
+- An EDITORIAL ARTICLE that still LACKS a cover/hero image → use blog_cover (NOT make_images, NOT placeholder stock). GUARDRAIL against noise: ONE cover per article, only when it is missing — if the article already has a real cover, leave it (never regenerate); a pure text/typo/SEO edit is NOT a trigger; never re-run on your own initiative (each call bills). Recognise it by MEANING, not by tags: any standalone piece written to be READ — a headline plus a body of prose on one topic (blog post, news, guide, tutorial, case study, magazine/journal piece) — in ANY language and on ANY stack (markdown/MDX, a headless CMS like WordPress/Sanity/Contentful, or a custom DB-driven route). Structural cues are only NON-EXHAUSTIVE hints, none of them required: folders/routes such as /blog /posts /articles /news /journal /actualités /magazine /guides, markdown/MDX frontmatter, or an <article> element — but if it simply READS like an article, it qualifies even with none of these. (It is the article body, not a marketing/landing/product page — those keep make_images.) You are the one writing/editing it, so you ALREADY have its text: pass its own title + body as article_text (or article_url / article_file). blog_cover reads the article, illustrates its SPECIFIC subject (not a generic trade photo), matches the site's locked style, adapts to the site's country (driving side, architecture…), and returns a 16:9 WebP + ALT. Default = the cover only; pass illustrations:N for in-article images. No style locked yet? it infers one from the project; if it still lacks the trade/brand it tells you what to provide — give it, don't guess.
 - One-off image → generate_image. Retouch/redo/cutout/upscale/widen an existing one → edit_image with the right action.
 - Finish a page properly: brand_pack (logo + favicon + og:image in ONE call) then finish_images (ALT auto-fixed + WebP optimisation, free).
 COST DISCIPLINE — NEVER iterate on your own. ONE make_images call dresses a page: the job is then DONE. Do NOT regenerate, redo, retouch or "improve" an image on your own initiative, do NOT call tools in a loop, do NOT re-run a call that succeeded — every generation bills the subscriber's credits and burns their tokens. Retouch or regenerate ONLY when the USER explicitly asks for it.

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "distribea-mcp",
-  "version": "1.2.0",
-  "description": "Distribea MCP, on-brand website imagery (style-locked, recurring characters, UGC review avatars) generated by the hosted Distribea engine. Requires a Distribea subscription key.",
+  "version": "1.3.1",
+  "description": "Distribea MCP, on-brand website imagery (style-locked, recurring characters, UGC review avatars, blog covers) generated by the hosted Distribea engine. Requires a Distribea subscription key.",
   "type": "module",
   "bin": {
     "distribea-mcp": "index.mjs"