mcp-scraper 0.3.44 → 0.3.46

package/dist/{chunk-NONBSIES.js → chunk-SS5YBZVI.js}

@@ -1,6 +1,6 @@
 import {
   PACKAGE_VERSION
-} from "./chunk-3Q2NSHSR.js";
+} from "./chunk-WT6OT2T3.js";
 import {
   browserServiceProfileName,
   browserServiceProfileSaveChanges,
@@ -91,9 +91,13 @@ function sanitizeHarvestResult(result) {
 }
 
 // src/mcp/server-instructions.ts
-var SERVER_INSTRUCTIONS = `
+function serverInstructions(savesReportsLocally) {
+  const reportLine = savesReportsLocally ? "All report-producing tools also save a full Markdown report to disk." : "On this hosted endpoint, small reports return inline; large reports are stored as artifacts \u2014 read them back with report_artifact_read.";
+  return `
 # MCP Scraper
 
+${reportLine}
+
 Scrapes and analyzes web, search, maps, social, and site data. **Pick a tool by NAME from the map
 below, then load its schema before calling.** Where a tool needs a value another tool produces, the
 seam is noted so you can chain them.
@@ -158,25 +162,41 @@ Multi-step orchestrations \u2014 prefer these over hand-chaining primitives when
 - Stand up a rank-tracking blueprint (database + cron + ingestion plan) -> **rank_tracker_workflow**.
 - Find which sources an AI answer cites for a query (AEO) -> **query_fanout_workflow** (open a browser
   session on chatgpt.com or claude.ai FIRST, then run it against that session).
+- Not sure which workflow fits a goal -> **workflow_suggest**; to just see what's available -> **workflow_list**.
+- Hosted/stepwise chain: **workflow_run** starts and returns \`runId\` (+ \`nextStep\` when multi-leg) ->
+  **workflow_step** with that \`runId\` advances one leg at a time until \`done\` -> **workflow_status**
+  re-opens a run to check state or recover artifact ids -> **workflow_artifact_read** pulls one artifact's
+  full content by \`runId\` + \`artifactId\`.
+- Browser replay recording: **browser_replay_start** begins it -> **browser_replay_mark** (while
+  recording) locates a DOM target and returns a ready-to-use timed annotation -> **browser_replay_stop**
+  ends it -> feed collected annotations to **browser_replay_annotate**, or just
+  **browser_replay_download** the plain MP4. **browser_list_replays** recovers replay ids.
+- Video breakdown is async: **video_frame_analysis** takes a DIRECT media file URL (resolve
+  YouTube/Facebook/Instagram page URLs to one first, e.g. via \`facebook_page_intel\`'s \`videoUrl\`) and
+  returns a \`runId\` immediately -> poll **video_frame_analysis_status** with that \`runId\` until \`status\`
+  is \`done\`.
 
 ## Notes
 - Bulk / full-site crawls: call \`extract_site\` with \`rotateProxies:true\` for blocked or rate-limited
   sites. It returns a saved folder/artifact plus a summary, not the full content inline.
 - Browser sessions and media transcription cost credits and take longer \u2014 prefer the cheapest tool that
   answers the question (plain search/extract before browser agents).
-- Large results are saved to disk or an artifact and returned as a summary plus a path; read the path for
-  full detail rather than expecting the whole payload inline.
+- Large results are saved to disk or an artifact and returned as a summary plus a path or artifactId;
+  read it back for full detail rather than expecting the whole payload inline.
 `.trim();
+}
+var SERVER_INSTRUCTIONS = serverInstructions(true);
 
 // src/mcp/paa-mcp-server.ts
 import { McpServer, ResourceTemplate } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { readdirSync, readFileSync, statSync } from "fs";
-import { basename, join as join2 } from "path";
+import { basename, join as join3 } from "path";
+import { createHash } from "crypto";
 
 // src/mcp/mcp-response-formatter.ts
-import { mkdirSync, writeFileSync } from "fs";
-import { homedir } from "os";
-import { join } from "path";
+import { mkdirSync as mkdirSync2, writeFileSync as writeFileSync2 } from "fs";
+import { homedir as homedir2 } from "os";
+import { join as join2 } from "path";
 
 // src/mcp/workflow-catalog.ts
 var WORKFLOW_RECIPES = [
@@ -713,7 +733,129 @@ function renderImageSection(audit) {
   return lines.join("\n");
 }
 
+// src/mcp/report-artifact-offload.ts
+import { randomBytes } from "crypto";
+
+// src/api/blob-store.ts
+import { mkdirSync, writeFileSync } from "fs";
+import { homedir } from "os";
+import { join, dirname } from "path";
+function byteLength(data) {
+  return Buffer.isBuffer(data) ? data.length : Buffer.byteLength(data);
+}
+var LocalBlobStore = class {
+  constructor(baseDir) {
+    this.baseDir = baseDir;
+  }
+  baseDir;
+  kind = "local";
+  async put(key, data, contentType = "application/octet-stream") {
+    const path = join(this.baseDir, "blobs", key);
+    mkdirSync(dirname(path), { recursive: true });
+    writeFileSync(path, data);
+    return { key, url: `file://${path}`, bytes: byteLength(data), contentType };
+  }
+  async get(key) {
+    const path = join(this.baseDir, "blobs", key);
+    try {
+      const { readFileSync: readFileSync2 } = await import("fs");
+      return readFileSync2(path);
+    } catch {
+      return null;
+    }
+  }
+};
+function localBaseDir() {
+  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join(homedir(), "Downloads", "mcp-scraper");
+}
+var cached = null;
+function getBlobStore() {
+  if (cached) return cached;
+  if (process.env.BLOB_READ_WRITE_TOKEN) {
+    cached = new VercelBlobStore(process.env.BLOB_READ_WRITE_TOKEN);
+  } else {
+    cached = new LocalBlobStore(localBaseDir());
+  }
+  return cached;
+}
+var VercelBlobStore = class {
+  constructor(token) {
+    this.token = token;
+  }
+  token;
+  kind = "vercel-blob";
+  async put(key, data, contentType = "application/octet-stream") {
+    const { put } = await import("@vercel/blob");
+    const body = Buffer.isBuffer(data) ? data : Buffer.from(data);
+    const result = await put(key, body, {
+      access: "public",
+      token: this.token,
+      contentType,
+      addRandomSuffix: true
+    });
+    return { key, url: result.url, bytes: byteLength(data), contentType };
+  }
+  async get(key) {
+    try {
+      const { list } = await import("@vercel/blob");
+      const res = await list({ prefix: key, token: this.token, limit: 1 });
+      const match = res.blobs.find((b) => b.pathname === key || b.pathname.startsWith(key));
+      if (!match) return null;
+      const resp = await fetch(match.url);
+      if (!resp.ok) return null;
+      return Buffer.from(await resp.arrayBuffer());
+    } catch {
+      return null;
+    }
+  }
+};
+
+// src/mcp/report-artifact-offload.ts
+var REPORT_BLOB_TTL_MS = 24 * 60 * 60 * 1e3;
+var REPORT_BLOB_PREFIX = "mcp-reports/";
+var PREVIEW_CHARS = 2e3;
+var ARTIFACT_OFFLOAD_ENABLED = process.env.MCP_SCRAPER_ARTIFACT_OFFLOAD !== "false";
+async function offloadReport(toolName, ownerId, report) {
+  const timestamp = Date.now();
+  const random = randomBytes(6).toString("hex");
+  const key = `${REPORT_BLOB_PREFIX}${ownerId}/${toolName}/${timestamp}-${random}.md`;
+  const stored = await getBlobStore().put(key, report, "text/markdown");
+  return {
+    artifactId: stored.key,
+    bytes: stored.bytes,
+    expiresAt: new Date(timestamp + REPORT_BLOB_TTL_MS).toISOString(),
+    preview: report.slice(0, PREVIEW_CHARS)
+  };
+}
+function artifactOwnerId(artifactId) {
+  if (!artifactId.startsWith(REPORT_BLOB_PREFIX)) return null;
+  const rest = artifactId.slice(REPORT_BLOB_PREFIX.length);
+  const segment = rest.split("/")[0];
+  return segment || null;
+}
+async function readArtifactWindow(artifactId, offset, maxBytes) {
+  const buf = await getBlobStore().get(artifactId);
+  if (!buf) return null;
+  const totalBytes = buf.length;
+  const start = Math.max(0, offset);
+  const end = Math.min(totalBytes, start + maxBytes);
+  const slice = buf.subarray(start, end);
+  const nextOffset = end < totalBytes ? end : null;
+  return { text: slice.toString("utf8"), totalBytes, nextOffset };
+}
+function summaryEnvelope(executiveSummary, offloaded) {
+  return [
+    executiveSummary.trim(),
+    "",
+    "--- Full report stored as artifact ---",
+    `artifactId: ${offloaded.artifactId} \xB7 ${offloaded.bytes} bytes \xB7 expires ${offloaded.expiresAt}`,
+    "Read it with report_artifact_read (supports offset/maxBytes windowing)."
+  ].join("\n");
+}
+
 // src/mcp/mcp-response-formatter.ts
+var INLINE_BUDGET_BYTES = Number(process.env.MCP_SCRAPER_INLINE_BUDGET ?? 5e4);
+var STRUCTURED_ARRAY_CAP = 25;
 var reportSavingEnabled = true;
 function configureReportSaving(enabled) {
   reportSavingEnabled = enabled;
@@ -731,16 +873,16 @@ function reportTitle(full) {
   return title?.replace(/^#\s+/, "").trim() || "MCP Scraper Report";
 }
 function outputBaseDir() {
-  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join(homedir(), "Downloads", "mcp-scraper");
+  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join2(homedir2(), "Downloads", "mcp-scraper");
 }
 function saveFullReport(full) {
   if (!reportSavingEnabled || process.env.MCP_SCRAPER_SAVE_REPORTS === "false") return null;
   const outDir = outputBaseDir();
   try {
-    mkdirSync(outDir, { recursive: true });
+    mkdirSync2(outDir, { recursive: true });
     const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
-    const file = join(outDir, `${stamp}-${slugifyReportName(reportTitle(full))}.md`);
-    writeFileSync(file, full, "utf8");
+    const file = join2(outDir, `${stamp}-${slugifyReportName(reportTitle(full))}.md`);
+    writeFileSync2(file, full, "utf8");
     return file;
   } catch {
     return null;
@@ -755,9 +897,9 @@ function saveBulkSite(siteUrl, pages, seo, imageAudit) {
   if (!reportSavingActive()) return null;
   try {
     const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
-    const dir = join(outputBaseDir(), `extract-${slugifyReportName(siteUrl)}-${stamp}`);
-    const pagesDir = join(dir, "pages");
-    mkdirSync(pagesDir, { recursive: true });
+    const dir = join2(outputBaseDir(), `extract-${slugifyReportName(siteUrl)}-${stamp}`);
+    const pagesDir = join2(dir, "pages");
+    mkdirSync2(pagesDir, { recursive: true });
     const indexRows = pages.map((p, i) => {
       const num = String(i + 1).padStart(4, "0");
       const slug = slugifyReportName(p.url.replace(/^https?:\/\//, "")).slice(0, 60) || "page";
@@ -771,7 +913,7 @@ function saveBulkSite(siteUrl, pages, seo, imageAudit) {
         "",
         body || "_(no content extracted)_"
       ].filter(Boolean).join("\n");
-      writeFileSync(join(pagesDir, fname), content, "utf8");
+      writeFileSync2(join2(pagesDir, fname), content, "utf8");
       return `| ${i + 1} | ${cell(p.title ?? "Untitled")} | ${p.url} | pages/${fname} |`;
     });
     const dataFilesSection = seo ? [
@@ -803,40 +945,40 @@ function saveBulkSite(siteUrl, pages, seo, imageAudit) {
 |---|-------|-----|------|
 ${indexRows.join("\n")}`
     ].filter(Boolean).join("\n");
-    const indexFile = join(dir, "index.md");
-    writeFileSync(indexFile, index, "utf8");
+    const indexFile = join2(dir, "index.md");
+    writeFileSync2(indexFile, index, "utf8");
     let seoFiles;
     if (seo) {
       const toJsonl = (rows) => rows.map((r) => JSON.stringify(r)).join("\n");
-      const pagesJsonl = join(dir, "pages.jsonl");
-      const linksJsonl = join(dir, "links.jsonl");
-      const metricsJsonl = join(dir, "link-metrics.jsonl");
-      const issuesFile = join(dir, "issues.json");
-      const reportFile = join(dir, "report.md");
-      writeFileSync(pagesJsonl, toJsonl(seo.pageRows), "utf8");
-      writeFileSync(linksJsonl, toJsonl(seo.edges), "utf8");
-      writeFileSync(metricsJsonl, toJsonl(seo.metrics), "utf8");
-      writeFileSync(issuesFile, JSON.stringify(seo.issues, null, 2), "utf8");
-      writeFileSync(reportFile, seo.reportMd + (imageAudit ? `
+      const pagesJsonl = join2(dir, "pages.jsonl");
+      const linksJsonl = join2(dir, "links.jsonl");
+      const metricsJsonl = join2(dir, "link-metrics.jsonl");
+      const issuesFile = join2(dir, "issues.json");
+      const reportFile = join2(dir, "report.md");
+      writeFileSync2(pagesJsonl, toJsonl(seo.pageRows), "utf8");
+      writeFileSync2(linksJsonl, toJsonl(seo.edges), "utf8");
+      writeFileSync2(metricsJsonl, toJsonl(seo.metrics), "utf8");
+      writeFileSync2(issuesFile, JSON.stringify(seo.issues, null, 2), "utf8");
+      writeFileSync2(reportFile, seo.reportMd + (imageAudit ? `
 
 ${renderImageSection(imageAudit)}` : ""), "utf8");
-      const linkReportFile = join(dir, "link-report.md");
-      const linksSummaryFile = join(dir, "links-summary.json");
-      const externalDomainsFile = join(dir, "external-domains.json");
-      writeFileSync(linkReportFile, renderLinkReport(seo.linkReport), "utf8");
-      writeFileSync(linksSummaryFile, JSON.stringify(seo.linkReport.summary, null, 2), "utf8");
-      writeFileSync(externalDomainsFile, JSON.stringify(seo.linkReport.externalDomains, null, 2), "utf8");
+      const linkReportFile = join2(dir, "link-report.md");
+      const linksSummaryFile = join2(dir, "links-summary.json");
+      const externalDomainsFile = join2(dir, "external-domains.json");
+      writeFileSync2(linkReportFile, renderLinkReport(seo.linkReport), "utf8");
+      writeFileSync2(linksSummaryFile, JSON.stringify(seo.linkReport.summary, null, 2), "utf8");
+      writeFileSync2(externalDomainsFile, JSON.stringify(seo.linkReport.externalDomains, null, 2), "utf8");
       seoFiles = [pagesJsonl, linksJsonl, metricsJsonl, issuesFile, reportFile, linkReportFile, linksSummaryFile, externalDomainsFile];
       if (imageAudit) {
-        const imagesJsonl = join(dir, "images.jsonl");
-        const imagesSummary = join(dir, "images-summary.json");
-        writeFileSync(imagesJsonl, imageAudit.rows.map((r) => JSON.stringify(r)).join("\n"), "utf8");
-        writeFileSync(imagesSummary, JSON.stringify(imageAudit.summary, null, 2), "utf8");
+        const imagesJsonl = join2(dir, "images.jsonl");
+        const imagesSummary = join2(dir, "images-summary.json");
+        writeFileSync2(imagesJsonl, imageAudit.rows.map((r) => JSON.stringify(r)).join("\n"), "utf8");
+        writeFileSync2(imagesSummary, JSON.stringify(imageAudit.summary, null, 2), "utf8");
         seoFiles.push(imagesJsonl, imagesSummary);
       }
       if (seo.branding) {
-        const brandingFile = join(dir, "branding.json");
-        writeFileSync(brandingFile, JSON.stringify(seo.branding, null, 2), "utf8");
+        const brandingFile = join2(dir, "branding.json");
+        writeFileSync2(brandingFile, JSON.stringify(seo.branding, null, 2), "utf8");
         seoFiles.push(brandingFile);
       }
     }
@@ -849,12 +991,12 @@ function saveUrlInventory(siteUrl, urls) {
   if (!reportSavingActive()) return null;
   try {
     const outDir = outputBaseDir();
-    mkdirSync(outDir, { recursive: true });
+    mkdirSync2(outDir, { recursive: true });
     const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
-    const file = join(outDir, `${stamp}-urlmap-${slugifyReportName(siteUrl.replace(/^https?:\/\//, ""))}.csv`);
+    const file = join2(outDir, `${stamp}-urlmap-${slugifyReportName(siteUrl.replace(/^https?:\/\//, ""))}.csv`);
     const csv = (v) => /[",\n]/.test(v) ? `"${v.replace(/"/g, '""')}"` : v;
     const rows = ["url,status", ...urls.map((u) => `${csv(u.url)},${u.status ?? ""}`)];
-    writeFileSync(file, rows.join("\n"), "utf8");
+    writeFileSync2(file, rows.join("\n"), "utf8");
     return file;
   } catch {
     return null;
@@ -863,12 +1005,12 @@ function saveUrlInventory(siteUrl, urls) {
 function persistScreenshotLocally(base64, url) {
   if (!reportSavingEnabled || process.env.MCP_SCRAPER_SAVE_REPORTS === "false") return null;
   try {
-    const dir = join(outputBaseDir(), "screenshots");
-    mkdirSync(dir, { recursive: true });
+    const dir = join2(outputBaseDir(), "screenshots");
+    mkdirSync2(dir, { recursive: true });
     const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
     const slug = url.replace(/^https?:\/\//, "").replace(/[^a-z0-9]+/gi, "-").replace(/^-+|-+$/g, "").slice(0, 60);
-    const filePath = join(dir, `${stamp}-${slug}.png`);
-    writeFileSync(filePath, Buffer.from(base64, "base64"));
+    const filePath = join2(dir, `${stamp}-${slug}.png`);
+    writeFileSync2(filePath, Buffer.from(base64, "base64"));
     return filePath;
   } catch {
     return null;
@@ -881,6 +1023,20 @@ function oneBlock(content, diskContent) {
 \u{1F4C4} Saved: \`${filePath}\`` : content;
   return { content: [{ type: "text", text }] };
 }
+async function maybeOffload(toolName, ctx, fullText, summaryText, structuredContent) {
+  if (!ctx?.hosted || !ARTIFACT_OFFLOAD_ENABLED) return null;
+  const bytes = Buffer.byteLength(fullText);
+  if (bytes <= INLINE_BUDGET_BYTES) return null;
+  const offloaded = await offloadReport(toolName, ctx.ownerId, fullText);
+  return {
+    content: [{ type: "text", text: summaryEnvelope(summaryText, offloaded) }],
+    structuredContent: { ...structuredContent, artifact: offloaded }
+  };
+}
+function capArray(items, cap) {
+  if (items.length <= cap) return { items };
+  return { items: items.slice(0, cap), truncatedCount: items.length - cap };
+}
 function workflowRecipeTable(recipes) {
   if (!recipes.length) return "";
   return [
@@ -1232,7 +1388,7 @@ ${headingSection}${kpoSection}${brandingSection}${bodySectionFull}${memSection}$
   }
   return { ...textResult, structuredContent };
 }
-function formatMapSiteUrls(raw, input) {
+async function formatMapSiteUrls(raw, input, ctx) {
   const parsed = parseData(raw);
   if ("error" in parsed) return { content: [{ type: "text", text: parsed.error }], isError: true };
   const d = parsed.data;
@@ -1273,20 +1429,28 @@ ${broken.map((u) => `- ${u.url} (${u.status})`).join("\n")}` : "",
 - Extract content from all pages: use \`extract_site\`
 - Scrape a single page: use \`extract_url\``
   ].filter(Boolean).join("\n");
-  return {
-    ...oneBlock(full),
-    structuredContent: {
-      startUrl: d.startUrl ?? input.url,
-      totalFound: d.totalFound ?? urls.length,
-      truncated: d.truncated === true,
-      okCount: ok.length,
-      redirectCount: redirects.length,
-      brokenCount: broken.length,
-      inventoryFile: inventoryFile ?? null,
-      urls: urls.map((u) => ({ url: u.url, status: u.status ?? null })),
-      durationMs: d.durationMs ?? 0
-    }
+  const structuredContent = {
+    startUrl: d.startUrl ?? input.url,
+    totalFound: d.totalFound ?? urls.length,
+    truncated: d.truncated === true,
+    okCount: ok.length,
+    redirectCount: redirects.length,
+    brokenCount: broken.length,
+    inventoryFile: inventoryFile ?? null,
+    urls: urls.map((u) => ({ url: u.url, status: u.status ?? null })),
+    durationMs: d.durationMs ?? 0
   };
+  const summary = `# URL Map: ${input.url}
+**${d.totalFound} URLs** \xB7 ${(d.durationMs / 1e3).toFixed(1)}s
+
+## Summary
+- 2xx: ${ok.length}
+- 3xx: ${redirects.length}
+- 4xx+: ${broken.length}`;
+  const capped = capArray(structuredContent.urls, STRUCTURED_ARRAY_CAP);
+  const offloaded = await maybeOffload("map_site_urls", ctx, full, summary, { ...structuredContent, urls: capped.items, truncatedCount: capped.truncatedCount });
+  if (offloaded) return offloaded;
+  return { ...oneBlock(full), structuredContent };
 }
 function buildSeoExport(siteUrl, pages, branding) {
   const { edges, metrics } = buildLinkGraph(pages, siteUrl);
@@ -1306,7 +1470,7 @@ function buildSeoExport(siteUrl, pages, branding) {
     branding: branding ?? void 0
   };
 }
-function formatExtractSite(raw, input) {
+async function formatExtractSite(raw, input, ctx) {
   const parsed = parseData(raw);
   if ("error" in parsed) return { content: [{ type: "text", text: parsed.error }], isError: true };
   const d = parsed.data;
@@ -1341,6 +1505,36 @@ function formatExtractSite(raw, input) {
       schemaTypes: schemaTypesOf(p)
     })));
     const preview = pages.slice(0, BULK_PAGE_THRESHOLD).map(pageRow).join("\n");
+    const summaryHeader = [
+      `# Site Extract: ${input.url}`,
+      durationLine,
+      `
+## Pages (first ${Math.min(BULK_PAGE_THRESHOLD, pages.length)} of ${pages.length})
+| # | Title | URL | Schema |
+|---|-------|-----|--------|
+${preview}`
+    ].join("\n");
+    if (!bulk && ctx?.hosted) {
+      const pageDetails2 = pages.map((p, i) => {
+        const body = (p.bodyMarkdown ?? "").trim();
+        return [
+          `
+## ${i + 1}. ${p.title ?? "Untitled"}`,
+          `- **URL:** ${p.url}`,
+          p.metaDescription ? `- **Description:** ${p.metaDescription}` : "",
+          body ? `
+${body}` : "_(no content extracted)_"
+        ].filter(Boolean).join("\n");
+      }).join("\n");
+      const fullForOffload = `${summaryHeader}
+
+---
+# Full Page Content
+${pageDetails2}${tips}`;
+      const capped = capArray(structuredContent.pages, STRUCTURED_ARRAY_CAP);
+      const offloaded = await maybeOffload("extract_site", ctx, fullForOffload, `${summaryHeader}${tips}`, { ...structuredContent, pages: capped.items, truncatedCount: capped.truncatedCount });
+      if (offloaded) return offloaded;
+    }
     const location = bulk ? `
 ## \u{1F4C1} Bulk scrape saved
 - **Folder:** \`${bulk.dir}\` \u2190 all scraped page content is here
@@ -1393,7 +1587,7 @@ ${body}` : "_(no content extracted)_"
 ${pageDetails}${tips}`;
   return { ...oneBlock(full, diskReport), structuredContent };
 }
-async function formatAuditSite(raw, input) {
+async function formatAuditSite(raw, input, ctx) {
   const parsed = parseData(raw);
   if ("error" in parsed) return { content: [{ type: "text", text: parsed.error }], isError: true };
   const d = parsed.data;
@@ -1424,6 +1618,24 @@ async function formatAuditSite(raw, input) {
     images: { unique: img.unique, totalBytes: img.totalBytes, over100kb: img.over100kb, legacyFormat: img.legacyFormat },
     links: { internal: lr.internal.totalLinks, external: lr.external.totalLinks, orphans: lr.internal.orphans, brokenInternal: lr.internal.brokenInternal, externalDomains: lr.external.uniqueDomains }
   };
+  const summary = [
+    `# Technical SEO Audit: ${input.url}`,
+    durationLine,
+    `
+## Top issues
+${topIssues || "_none found_"}`,
+    `
+## Links
+${linkLine}`,
+    `
+## Images
+${imgLine}`
+  ].join("\n");
+  if (!bulk && ctx?.hosted) {
+    const fullForOffload = [summary, "\n---\n# Full Audit Report", seo.reportMd, renderImageSection(imageAudit), renderLinkReport(seo.linkReport)].join("\n");
+    const offloaded = await maybeOffload("audit_site", ctx, fullForOffload, summary, structuredContent);
+    if (offloaded) return offloaded;
+  }
   const location = bulk ? `
 ## \u{1F4C1} Technical audit saved
 - **Folder:** \`${bulk.dir}\` \u2190 full audit + page content here
@@ -2194,7 +2406,7 @@ ${rows}`,
     }
   };
 }
-function formatDirectoryWorkflow(raw, input) {
+async function formatDirectoryWorkflow(raw, input, ctx) {
   const parsed = parseData(raw);
   if ("error" in parsed) return { content: [{ type: "text", text: parsed.error }], isError: true };
   const d = parsed.data;
@@ -2246,26 +2458,29 @@ ${businessRows}` : null,
     durationMs != null ? `
 *Completed in ${(durationMs / 1e3).toFixed(1)}s*` : null
   ].filter(Boolean).join("\n");
-  return {
-    ...oneBlock(full),
-    structuredContent: {
-      query: d.query,
-      state: d.state,
-      minPopulation: d.minPopulation,
-      populationYear: d.populationYear,
-      maxResultsPerCity: d.maxResultsPerCity,
-      concurrency: d.concurrency,
-      censusSourceUrl: d.censusSourceUrl,
-      usZipsSourcePath: d.usZipsSourcePath ?? null,
-      warnings,
-      extractedAt: d.extractedAt,
-      selectedCityCount: d.selectedCityCount,
-      totalResultCount,
-      csvPath,
-      cities,
-      durationMs: durationMs ?? 0
-    }
+  const structuredContent = {
+    query: d.query,
+    state: d.state,
+    minPopulation: d.minPopulation,
+    populationYear: d.populationYear,
+    maxResultsPerCity: d.maxResultsPerCity,
+    concurrency: d.concurrency,
+    censusSourceUrl: d.censusSourceUrl,
+    usZipsSourcePath: d.usZipsSourcePath ?? null,
+    warnings,
+    extractedAt: d.extractedAt,
+    selectedCityCount: d.selectedCityCount,
+    totalResultCount,
+    csvPath,
+    cities,
+    durationMs: durationMs ?? 0
   };
+  const summary = `# Directory Workflow: ${input.query}
+**Markets:** ${cities.length} \xB7 **Maps results:** ${totalResultCount} \xB7 **State:** ${d.state ?? input.state ?? "US"}`;
+  const capped = capArray(cities, STRUCTURED_ARRAY_CAP);
+  const offloaded = await maybeOffload("directory_workflow", ctx, full, summary, { ...structuredContent, cities: capped.items, truncatedCount: capped.truncatedCount });
+  if (offloaded) return offloaded;
+  return { ...oneBlock(full), structuredContent };
 }
 function formatMapsPlaceIntel(raw, input) {
   const parsed = parseData(raw);
@@ -2786,28 +3001,36 @@ ${rows}` : ""
   };
 }
 
+// src/mcp/output-schema-registry.ts
+var ADVERTISE_OUTPUT_SCHEMAS = process.env.MCP_SCRAPER_ADVERTISE_OUTPUT_SCHEMAS === "true";
+var OUTPUT_SCHEMAS = {};
+function recordOutputSchema(name, schema) {
+  OUTPUT_SCHEMAS[name] = schema;
+  return ADVERTISE_OUTPUT_SCHEMAS ? schema : void 0;
+}
+
 // src/mcp/mcp-tool-schemas.ts
 import { z } from "zod";
 var HarvestPaaInputSchema = {
-  query: z.string().min(1).describe('The search query. For localized results, KEEP the place in the query text \u2014 if the user says "best hvac company in Denver CO", use query="best hvac company Denver CO" (and also set location="Denver, CO"). City-in-query is what localizes results reliably; do not strip it out.'),
-  location: z.string().optional().describe('City, region, or country for geo signals (uule + location evidence), e.g. "Denver, CO", "Tokyo, Japan". Set it alongside a city-in-query wording; on its own it does NOT reliably localize results.'),
-  maxQuestions: z.number().int().min(1).max(200).default(30).describe("Number of PAA questions to extract. Default 30. Maximum 200. Use 10 for quick probes, 30 for normal research, 100-200 when the user asks for everything/full/deep research. Larger harvests get a longer server time budget (151-200 questions \u2192 up to 280s). Credits are charged by extracted question; unused request hold is refunded."),
-  gl: z.string().length(2).default("us").describe("Google country code inferred from location or user language. Examples: United States us, United Kingdom gb, Japan jp, Canada ca, Australia au."),
-  hl: z.string().default("en").describe("Google interface/content language inferred from the user request. Use en unless the user asks for another language or locale."),
-  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use desktop by default; use mobile only when the user asks for mobile rankings."),
-  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress \u2014 reliable). Do NOT set "location" just because the user named a city; localized results come from city-in-query wording. "location" forces a residential geo-IP for rank-tracking fidelity only, is frequently CAPTCHA-blocked, and should be used only when the user explicitly demands as-seen-from-the-city results and accepts failures.'),
-  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location" (see its warning); ignored otherwise.'),
-  debug: z.boolean().default(false).describe("Include sanitized browser/session/location diagnostics in the response. Use true when debugging localization, CAPTCHA, or proxy behavior.")
+  query: z.string().min(1).describe('The search query. KEEP the place in the query text for localized results (e.g. "best hvac company Denver CO") and also set location \u2014 city-in-query is what localizes reliably.'),
+  location: z.string().optional().describe('City, region, or country for geo signals, e.g. "Denver, CO". Set alongside city-in-query wording; alone it does NOT reliably localize.'),
+  maxQuestions: z.number().int().min(1).max(200).default(30).describe("PAA questions to extract. Default 30, maximum 200. Use 10 for quick probes, 100-200 for deep research. Billed per extracted question; unused hold refunded."),
+  gl: z.string().length(2).default("us").describe("Google country code inferred from location or user language."),
+  hl: z.string().default("en").describe("Google interface/content language inferred from the user request."),
+  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use mobile only for mobile rankings."),
+  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress). Do NOT set "location" just because a city was named \u2014 that comes from city-in-query wording. "location" forces residential geo-IP for rank-tracking fidelity, is frequently CAPTCHA-blocked, and accepts failures.'),
+  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location".'),
+  debug: z.boolean().default(false).describe("Include sanitized diagnostics for debugging localization, CAPTCHA, or proxy behavior.")
 };
 var ExtractUrlInputSchema = {
-  url: z.string().url().describe("Public http/https URL to extract. Use this when the user provides one specific page URL."),
-  screenshot: z.boolean().default(false).describe("Also capture a full-page screenshot of the URL. Saved to ~/Downloads/mcp-scraper/screenshots/ and returned inline. Use when the user asks to see or capture the page visually."),
-  screenshotDevice: z.enum(["desktop", "mobile"]).default("desktop").describe("Viewport for screenshot. desktop = 1440\xD7900. mobile = 390\xD7844. Default desktop."),
-  extractBranding: z.boolean().default(false).describe("Extract brand colors, fonts, logo, and favicon using a rendered browser session. Returns colorScheme (light/dark), colors (primary/accent/background/text/heading as hex), fonts (heading/body family names), and assets (logo URL, favicon URL). Use when the user asks about brand colors, site theme, or brand assets."),
-  downloadMedia: z.boolean().default(false).describe("Extract and download all page media (images, video, audio) to ~/Downloads/mcp-scraper/media/. Ad networks, tracking pixels, and noise URLs are filtered automatically. Use when the user asks to download or harvest assets from a page."),
+  url: z.string().url().describe("Public http/https URL to extract."),
+  screenshot: z.boolean().default(false).describe("Capture a full-page screenshot, saved to ~/Downloads/mcp-scraper/screenshots/ and returned inline."),
+  screenshotDevice: z.enum(["desktop", "mobile"]).default("desktop").describe("Viewport for screenshot. desktop = 1440\xD7900, mobile = 390\xD7844."),
+  extractBranding: z.boolean().default(false).describe("Extract brand colors, fonts, logo, and favicon via a rendered browser session."),
+  downloadMedia: z.boolean().default(false).describe("Extract and download page media (images/video/audio) to ~/Downloads/mcp-scraper/media/. Ad/tracking noise is filtered automatically."),
   mediaTypes: z.array(z.enum(["image", "video", "audio"])).default(["image", "video", "audio"]).describe("Which media types to download. Default all three."),
-  allowLocal: z.boolean().default(false).describe("Allow localhost and private-network URLs. For local development only."),
-  depositToVault: z.boolean().default(false).describe("Also save the full page content into the user's MCP Memory Library vault, server-side. The complete body is stored and embedded for semantic recall WITHOUT passing through this conversation \u2014 the tool result only confirms the saved note. Use when the user asks to save/remember/store a page, or to build a knowledge base from scrapes."),
+  allowLocal: z.boolean().default(false).describe("Allow localhost and private-network URLs. Local development only."),
+  depositToVault: z.boolean().default(false).describe("Save the full page content into the user's MCP Memory vault server-side, embedded for semantic recall \u2014 the full body is NOT returned to chat."),
   vaultName: z.string().trim().min(1).max(120).optional().describe("Optional vault to deposit into. Defaults to the user's personal vault.")
 };
 var MapSiteUrlsInputSchema = {
@@ -2815,34 +3038,34 @@ var MapSiteUrlsInputSchema = {
   maxUrls: z.number().int().min(1).max(1e4).optional().describe("Maximum URLs to discover. Use 100 for normal maps, up to 10000 for a full inventory. Large maps (over 500 URLs) write the complete inventory to a local file and return only a summary plus the file path instead of the full list inline.")
 };
 var ExtractSiteInputSchema = {
-  url: z.string().url().describe("Public website URL or domain to crawl for page CONTENT across multiple pages (map + scrape). Use when the user wants the content/text/markdown of a site's pages. For a technical SEO audit (issues, link graph, indexability, headings, image weights) use audit_site instead \u2014 extract_site returns content only, not analysis."),
-  maxPages: z.number().int().min(1).max(1e4).optional().describe("Maximum pages to extract. Use 50 for a normal crawl, up to 10000 for a full-site bulk scrape. Bulk crawls (over 25 pages) switch to folder mode: every page is saved as its own Markdown file in a local folder and the response returns only a summary plus the folder path, so the full content never floods the context window."),
-  rotateProxies: z.boolean().optional().describe("Route page fetches through rotating residential proxies in headful browsers to defeat rate-limiting and bot blocks (403/429). Discovers URLs from the sitemap, fetches in batches with a fresh proxy per batch, retries failures on a new proxy, and automatically parallelizes across the account's concurrency slots. Slower and pricier than the default fetch path, so use only when a site blocks normal crawling."),
-  rotateProxyEvery: z.number().int().min(1).max(100).optional().describe("When rotateProxies is on, how many pages to fetch per browser/proxy before discarding it and minting a fresh one. Default 30. Lower values rotate IPs more aggressively against strict rate limits."),
-  formats: z.array(z.enum(["markdown", "links", "json", "images", "branding"])).optional().describe("Which data formats to include in the per-page output. markdown=page content as Markdown, links=internal/external link graph, json=structured data/schema, images=per-page image URLs (links only, no downloads), branding=site-level logo/colors/fonts captured once from the homepage (requires a browser, adds time). markdown, links, json, and images are always captured cheaply from the HTML; branding is the only opt-in capture. Defaults to markdown+links when omitted.")
+  url: z.string().url().describe("Public website URL or domain to crawl for page CONTENT (map + scrape). For a technical SEO audit use audit_site instead \u2014 this returns content only, not analysis."),
+  maxPages: z.number().int().min(1).max(1e4).optional().describe("Maximum pages to extract. Bulk crawls (over 25 pages) switch to folder mode: each page saved as its own Markdown file, with a summary plus folder path returned instead of inlining content."),
+  rotateProxies: z.boolean().optional().describe("Route page fetches through rotating residential proxies to defeat rate-limiting and bot blocks (403/429). Slower and pricier \u2014 use only when a site blocks normal crawling."),
+  rotateProxyEvery: z.number().int().min(1).max(100).optional().describe("When rotateProxies is on, pages fetched per proxy before rotating. Default 30."),
+  formats: z.array(z.enum(["markdown", "links", "json", "images", "branding"])).optional().describe("Per-page output formats: markdown, links, json, images are captured cheaply from HTML; branding (site-level logo/colors/fonts) requires a browser and adds time. Defaults to markdown+links.")
 };
 var AuditSiteInputSchema = {
-  url: z.string().url().describe("Public website URL or domain to run a full technical SEO audit on. Use when the user asks for a technical audit, SEO audit, site health check, or a Screaming-Frog-style crawl \u2014 i.e. they want ANALYSIS (issues, internal link graph, indexability, heading breakdown, image sizes/formats), not just page content. For plain content scraping use extract_site instead."),
-  maxPages: z.number().int().min(1).max(1e4).optional().describe("Maximum pages to crawl and audit. Use 50 for a normal audit, up to 10000 for a full-site audit. The audit always writes a folder of analysis files (report.md, issues.json, pages.jsonl, links.jsonl, link-metrics.jsonl, images.jsonl) plus per-page content, and returns a headline summary plus the folder path."),
-  rotateProxies: z.boolean().optional().describe("Route page fetches through rotating residential proxies to defeat rate-limiting and bot blocks (403/429). Slower and pricier than the default fetch path, so use only when a site blocks normal crawling."),
-  rotateProxyEvery: z.number().int().min(1).max(100).optional().describe("When rotateProxies is on, how many pages to fetch per browser/proxy before discarding it and minting a fresh one. Default 30.")
+  url: z.string().url().describe("Public website URL or domain for a full technical SEO audit (issues, link graph, indexability, headings, images). For plain content use extract_site instead."),
+  maxPages: z.number().int().min(1).max(1e4).optional().describe("Maximum pages to crawl and audit. Always writes a folder of analysis files plus per-page content, returning a summary plus the folder path."),
+  rotateProxies: z.boolean().optional().describe("Route page fetches through rotating residential proxies to defeat rate-limiting and bot blocks. Slower/pricier \u2014 use only when a site blocks normal crawling."),
+  rotateProxyEvery: z.number().int().min(1).max(100).optional().describe("When rotateProxies is on, pages fetched per proxy before rotating. Default 30.")
 };
 var YoutubeHarvestInputSchema = {
   mode: z.enum(["search", "channel"]).describe("Use search for topic/keyword requests. Use channel when the user provides @handle, channel ID, or channel URL."),
   query: z.string().optional().describe("Required when mode is search. The YouTube search topic in the user\u2019s words."),
   channelHandle: z.string().optional().describe("YouTube channel handle, channel ID, or URL. Examples: @mkbhd, UC..., https://youtube.com/@mkbhd."),
-  maxVideos: z.number().int().min(1).max(500).default(50).describe("Number of videos to return. Default 50, maximum 500. Use 10-25 for quick topic discovery, 50 for normal channel/search harvests, and larger values only when the user asks for full channel/history because large responses consume more context.")
+  maxVideos: z.number().int().min(1).max(500).default(50).describe("Number of videos to return. Default 50, maximum 500.")
 };
 var YoutubeTranscribeInputSchema = {
   videoId: z.string().min(1).optional().describe("YouTube video ID, e.g. dQw4w9WgXcQ. Use only an ID returned by youtube_harvest or visible in a YouTube URL; do not invent one."),
-  url: z.string().url().optional().describe("Full YouTube URL, e.g. https://www.youtube.com/watch?v=dQw4w9WgXcQ or https://youtu.be/dQw4w9WgXcQ. Use this when the user pasted a URL instead of an ID. Provide videoId or url.")
+  url: z.string().url().optional().describe("Full YouTube URL. Use when the user pasted a URL instead of an ID. Provide videoId or url.")
 };
 var FacebookPageIntelInputSchema = {
-  pageId: z.string().optional().describe("Facebook advertiser/page ID. Use only a pageId returned by facebook_ad_search/facebook_page_intel or copied from Facebook Ad Library; do not construct one yourself."),
-  libraryId: z.string().optional().describe("Facebook Ad Library archive ID for a known ad or advertiser sample. Use a libraryId returned by facebook_ad_search, or a libraryId/adArchiveId visible in Ad Library."),
+  pageId: z.string().optional().describe("Facebook advertiser/page ID. Use only a value returned by facebook_ad_search or copied from Ad Library."),
+  libraryId: z.string().optional().describe("Facebook Ad Library archive ID. Use a value returned by facebook_ad_search, or a libraryId/adArchiveId visible in Ad Library."),
   query: z.string().optional().describe("Advertiser or brand name when pageId/libraryId is not known. One of pageId, libraryId, or query is required."),
-  maxAds: z.number().int().min(1).max(200).default(50).describe("Maximum ads to inspect. Default 50, maximum 200. Prefer 25-50 for focused advertiser scans; use 100-200 only when the user asks for a broad ad archive sweep."),
-  country: z.string().length(2).default("US").describe("Two-letter Ad Library country code. Default US. Examples: US, CA, GB, AU. Infer from the user request when they name a country.")
+  maxAds: z.number().int().min(1).max(200).default(50).describe("Maximum ads to inspect. Default 50, maximum 200."),
+  country: z.string().length(2).default("US").describe("Two-letter Ad Library country code. Default US.")
 };
 var FacebookAdSearchInputSchema = {
   query: z.string().min(1).describe("Advertiser, brand, competitor, niche, or keyword to search in Facebook Ad Library."),
@@ -2850,7 +3073,7 @@ var FacebookAdSearchInputSchema = {
   maxResults: z.number().int().min(1).max(20).default(10).describe("Maximum advertisers to return. Default 10, maximum 20. Prefer tighter search terms over maxing this out.")
 };
 var RedditThreadInputSchema = {
-  url: z.string().min(1).describe("A reddit.com thread/post URL (www, old, or new Reddit, or a redd.it link). The service fetches it via old.reddit through a residential proxy and returns the post plus its comment tree."),
+  url: z.string().min(1).describe("A reddit.com thread/post URL (www, old, new Reddit, or redd.it)."),
   maxComments: z.number().int().min(1).max(2e3).optional().describe("Optional cap on comments returned. Omit to return all captured comments.")
 };
 var VideoFrameAnalysisInputSchema = {
@@ -2864,14 +3087,14 @@ var VideoFrameAnalysisStatusInputSchema = {
   runId: z.string().min(1).describe("The runId returned by video_frame_analysis.")
 };
 var FacebookAdTranscribeInputSchema = {
-  videoUrl: z.string().url().describe("Direct Facebook CDN video URL from a facebook_page_intel ad result. Do not pass a public Facebook reel/post/share URL here; use facebook_video_transcribe for organic Facebook URLs.")
+  videoUrl: z.string().url().describe("Direct Facebook CDN video URL from facebook_page_intel. Do not pass a public post/reel/share URL \u2014 use facebook_video_transcribe for those.")
 };
 var FacebookVideoTranscribeInputSchema = {
-  url: z.string().url().describe("Organic Facebook reel, video, watch, post, or share URL from facebook.com, m.facebook.com, or fb.watch. The tool renders the page, extracts the best matching public Facebook CDN MP4 URL, then transcribes it. Use this when the user pastes a normal Facebook video page URL and asks for the transcript or downloadable MP4."),
+  url: z.string().url().describe("Organic Facebook reel/video/watch/post/share URL from facebook.com, m.facebook.com, or fb.watch."),
   quality: z.enum(["best", "hd", "sd"]).default("best").describe("Preferred progressive MP4 quality. Use best by default; hd prefers the highest HD progressive URL; sd forces the SD URL.")
 };
 var GoogleAdsSearchInputSchema = {
-  query: z.string().min(1).describe("A domain (e.g. getviktor.com) or advertiser/brand name to look up in the Google Ads Transparency Center. Domains resolve to the advertisers running ads that point to that site."),
+  query: z.string().min(1).describe("A domain (e.g. getviktor.com) or advertiser/brand name to look up in Google Ads Transparency Center."),
   region: z.string().length(2).default("US").describe("Two-letter region code for where the ads are shown. Default US. Examples: US, CA, GB, AU."),
   maxResults: z.number().int().min(1).max(20).default(10).describe("Maximum advertisers to return. Default 10, maximum 20.")
 };
@@ -2882,78 +3105,84 @@ var GoogleAdsPageIntelInputSchema = {
   maxAds: z.number().int().min(1).max(200).default(50).describe("Maximum creatives to inspect and hydrate. Default 50, maximum 200. Prefer 25-50 for focused scans.")
 };
 var GoogleAdsTranscribeInputSchema = {
-  videoUrl: z.string().url().describe("Direct googlevideo.com playback URL from a google_ads_page_intel video ad result (the videoUrl field). For YouTube-hosted ads use youtube_transcribe with the returned youtubeVideoId instead.")
+  videoUrl: z.string().url().describe("Direct googlevideo.com playback URL from google_ads_page_intel. For YouTube-hosted ads use youtube_transcribe instead.")
 };
 var InstagramProfileContentInputSchema = {
   handle: z.string().min(1).optional().describe("Instagram handle, with or without @. Provide handle or url."),
-  url: z.string().url().optional().describe("Instagram profile URL, e.g. https://www.instagram.com/nasaartemis/. Provide handle or url."),
-  profile: z.string().min(1).optional().describe("Optional saved hosted browser profile name to load authenticated Instagram access. If omitted, the server uses its configured default profile when present."),
-  saveProfileChanges: z.boolean().optional().describe("Whether to save browser changes back to the hosted profile. Leave unset unless intentionally updating the saved login session."),
-  maxItems: z.number().int().min(1).max(2e3).default(50).describe("Maximum profile grid post/reel/tv URLs to collect. Default 50, maximum 2000. Use higher values only when the user asks for a fuller archive."),
-  maxScrolls: z.number().int().min(0).max(250).default(10).describe("Maximum pagination scroll attempts. Default 10, maximum 250. Increase for long profiles when Instagram continues loading more grid links."),
-  scrollDelayMs: z.number().int().min(250).max(5e3).default(1200).describe("Delay after each pagination scroll before collecting newly loaded links. Default 1200ms. Increase to 2000-3000ms when Instagram loads slowly."),
-  stableScrollLimit: z.number().int().min(1).max(10).default(4).describe("Stop after this many consecutive scrolls with no new links or scroll progress. Default 4.")
+  url: z.string().url().optional().describe("Instagram profile URL. Provide handle or url."),
+  profile: z.string().min(1).optional().describe("Optional saved hosted browser profile name for authenticated Instagram access."),
+  saveProfileChanges: z.boolean().optional().describe("Save browser changes back to the hosted profile. Leave unset unless intentionally updating the saved login."),
+  maxItems: z.number().int().min(1).max(2e3).default(50).describe("Maximum grid URLs to collect. Default 50, maximum 2000."),
+  maxScrolls: z.number().int().min(0).max(250).default(10).describe("Maximum pagination scroll attempts. Default 10, maximum 250."),
+  scrollDelayMs: z.number().int().min(250).max(5e3).default(1200).describe("Delay after each scroll before collecting new links. Default 1200ms."),
+  stableScrollLimit: z.number().int().min(1).max(10).default(4).describe("Stop after this many consecutive scrolls with no new links.")
 };
 var InstagramMediaDownloadInputSchema = {
-  url: z.string().url().describe("Instagram post, reel, or tv URL, e.g. https://www.instagram.com/reel/SHORTCODE/. The tool renders the page, extracts text, image metadata, and Instagram CDN media tracks."),
-  profile: z.string().min(1).optional().describe("Optional saved hosted browser profile name to load authenticated Instagram access. If omitted, the server uses its configured default profile when present."),
-  saveProfileChanges: z.boolean().optional().describe("Whether to save browser changes back to the hosted profile. Leave unset unless intentionally updating the saved login session."),
-  mediaTypes: z.array(z.enum(["image", "video", "audio"])).default(["image", "video", "audio"]).describe("Which media types to download when downloadMedia is true. Reels commonly expose separate video-only and audio-only MP4 tracks."),
-  downloadMedia: z.boolean().default(true).describe("Download extracted text/media files to the MCP Scraper output directory when the API server can write files. Always returns extracted media URLs even when false."),
-  downloadAllTracks: z.boolean().default(false).describe("Download every captured Instagram MP4 track instead of only the selected best video and audio tracks. Use false by default to avoid duplicate bitrates."),
-  includeTranscript: z.boolean().default(false).describe("Transcribe the selected audio track when available. This adds transcription cost and may take longer."),
-  mux: z.boolean().default(true).describe("When video and audio tracks are downloaded separately, try to mux them into a single MP4 if ffmpeg is available. Returns separate tracks when muxing is unavailable.")
+  url: z.string().url().describe("Instagram post, reel, or tv URL, e.g. https://www.instagram.com/reel/SHORTCODE/."),
+  profile: z.string().min(1).optional().describe("Optional saved hosted browser profile name for authenticated Instagram access."),
+  saveProfileChanges: z.boolean().optional().describe("Save browser changes back to the hosted profile. Leave unset unless intentionally updating the saved login."),
+  mediaTypes: z.array(z.enum(["image", "video", "audio"])).default(["image", "video", "audio"]).describe("Which media types to download when downloadMedia is true."),
+  downloadMedia: z.boolean().default(true).describe("Download extracted text/media files to the output directory. Media URLs are always returned even when false."),
+  downloadAllTracks: z.boolean().default(false).describe("Download every captured MP4 track instead of only the best video/audio pair."),
+  includeTranscript: z.boolean().default(false).describe("Transcribe the selected audio track. Adds transcription cost and time."),
+  mux: z.boolean().default(true).describe("Mux separately downloaded video/audio tracks into one MP4 if ffmpeg is available.")
 };
 var MapsPlaceIntelInputSchema = {
-  businessName: z.string().min(1).describe('Business name only. If user says "Elite Roofing Denver CO", use businessName="Elite Roofing" and location="Denver, CO".'),
-  location: z.string().min(1).describe('City/region/country where the business should be searched, e.g. "Denver, CO". Infer from the user request when possible.'),
+  businessName: z.string().min(1).describe('Business name only, e.g. "Elite Roofing" (not "Elite Roofing Denver CO" \u2014 put the city in location).'),
+  location: z.string().min(1).describe('City/region/country where the business should be searched, e.g. "Denver, CO".'),
   gl: z.string().length(2).default("us").describe("Google country code inferred from location."),
   hl: z.string().length(2).default("en").describe("Language inferred from user request."),
-  includeReviews: z.boolean().default(false).describe("Whether to fetch individual review cards. Use true when the user asks for reviews, customer pain, complaints, praise, themes, or review evidence."),
-  maxReviews: z.number().int().min(1).max(500).default(50).describe("Max review cards to return when includeReviews is true. Default 50, maximum 500. Use 50 for normal review analysis and larger values only for deep review mining.")
+  includeReviews: z.boolean().default(false).describe("Fetch individual review cards \u2014 for reviews, customer pain, complaints, or praise themes."),
+  maxReviews: z.number().int().min(1).max(500).default(50).describe("Max review cards when includeReviews is true. Default 50, maximum 500.")
 };
 var MapsSearchInputSchema = {
-  query: z.string().min(1).describe('Business category, niche, keyword, or search term. If the user says "roofers in Denver CO", use query="roofers" and location="Denver, CO". Do not put the location here when it can be separated.'),
-  location: z.string().optional().describe('City, region, country, or service area for the Maps search, e.g. "Denver, CO". Infer from the user request when present.'),
+  query: z.string().min(1).describe('Business category, niche, or search term, e.g. "roofers". Do not include location here \u2014 use location instead.'),
+  location: z.string().optional().describe('City, region, country, or service area, e.g. "Denver, CO".'),
   gl: z.string().length(2).default("us").describe("Google country code inferred from location."),
   hl: z.string().length(2).default("en").describe("Language inferred from user request."),
-  maxResults: z.number().int().min(1).max(50).default(10).describe("Number of Google Maps business/profile candidates to return. Default 10. Maximum 50. Use 10 unless the user asks for more."),
-  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_MAPS_PROXY_MODE).describe("Proxy targeting mode. Maps defaults to location so local market searches get city/state residential proxy targeting and rotation. Use configured to force the service proxy without city/ZIP targeting, and none only for local direct-network debugging."),
-  proxyZip: z.string().regex(/^\d{5}$/).optional().describe("Optional US ZIP override for residential location proxy targeting. Use when the user gives a specific ZIP or city-center ZIP."),
-  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics when debugging Maps localization, CAPTCHA, or proxy behavior.")
+  maxResults: z.number().int().min(1).max(50).default(10).describe("Number of candidates to return. Default 10, maximum 50."),
+  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_MAPS_PROXY_MODE).describe("Defaults to location (city/state residential proxy targeting). configured forces the service proxy without city/ZIP targeting; none is local debugging only."),
+  proxyZip: z.string().regex(/^\d{5}$/).optional().describe("Optional US ZIP override for residential proxy targeting."),
+  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics.")
 };
 var DirectoryWorkflowInputSchema = {
-  query: z.string().min(1).describe("Business category, niche, or keyword to search on Google Maps for every selected market, e.g. roofers, dentists, med spas. Do not include the city here."),
-  state: z.string().min(2).default("TN").describe("US state abbreviation or state name used to select Census places, e.g. TN or Tennessee."),
-  minPopulation: z.number().int().min(0).default(1e5).describe('Minimum Census place population for market selection. Use 100000 for "cities above 100k population".'),
-  populationYear: z.number().int().min(2020).max(2025).default(2025).describe("Census population estimate year from the 2020-2025 Population Estimates Program city/place dataset."),
-  maxCities: z.number().int().min(1).max(100).default(25).describe("Maximum number of markets to process after sorting by population descending."),
-  maxResultsPerCity: z.number().int().min(1).max(50).default(50).describe("Google Maps business/profile candidates to collect for each city. Maximum 50."),
-  concurrency: z.number().int().min(1).max(5).default(5).describe("How many city Maps searches to run in parallel. Use 5 for broad directory batches unless debugging."),
-  includeZipGroups: z.boolean().default(true).describe("Attach ZIP groups from a configured US ZIPS CSV when available. Set MCP_SCRAPER_USZIPS_CSV_PATH on the API server or pass usZipsCsvPath in local/test mode."),
-  usZipsCsvPath: z.string().optional().describe("Local/test-only path to a US ZIPS CSV with state_abbr, zipcode, county, city columns, such as Lead Magician tools/analytics/data/uszips.csv. Deployed APIs should use MCP_SCRAPER_USZIPS_CSV_PATH instead."),
-  saveCsv: z.boolean().default(true).describe("Save a directory-ready CSV to the MCP Scraper output directory and return its path. CSV rows include source_location, result_position, business_name, review_stars, review_count, category, address, phone, hours_status, website_url, directions_url, place_url, CID fields, population, and ZIP groups."),
-  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_MAPS_PROXY_MODE).describe("Proxy targeting mode for every city Maps search. Maps workflows default to location so each city can use city/state or ZIP-group residential proxy targeting. Use configured to force the service proxy without city/ZIP targeting, and none only for local direct-network debugging."),
-  proxyZip: z.string().regex(/^\d{5}$/).optional().describe("Optional ZIP override for proxy targeting. Normally omit it so each city can use its ZIP group or city/state location."),
-  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics in each Maps browser session when supported.")
+  query: z.string().min(1).describe("Business category, niche, or keyword to search on Google Maps for every market. Do not include the city."),
+  state: z.string().min(2).default("TN").describe("US state abbreviation or name used to select Census places, e.g. TN."),
+  minPopulation: z.number().int().min(0).default(1e5).describe("Minimum Census place population for market selection."),
+  populationYear: z.number().int().min(2020).max(2025).default(2025).describe("Census population estimate year (2020-2025 Population Estimates Program)."),
+  maxCities: z.number().int().min(1).max(100).default(25).describe("Maximum markets to process after sorting by population descending."),
+  maxResultsPerCity: z.number().int().min(1).max(50).default(50).describe("Google Maps candidates to collect per city."),
+  concurrency: z.number().int().min(1).max(5).default(5).describe("City Maps searches to run in parallel."),
+  includeZipGroups: z.boolean().default(true).describe("Attach ZIP groups from a configured US ZIPS CSV when available (MCP_SCRAPER_USZIPS_CSV_PATH or usZipsCsvPath)."),
+  usZipsCsvPath: z.string().optional().describe("Local/test-only path to a US ZIPS CSV (state_abbr, zipcode, county, city columns). Deployed APIs should use MCP_SCRAPER_USZIPS_CSV_PATH instead. For ZIP enrichment, set MCP_SCRAPER_USZIPS_CSV_PATH on the server, or pass this in local/test mode."),
+  saveCsv: z.boolean().default(true).describe("Save a directory-ready CSV of results to the MCP Scraper output directory and return its path."),
+  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_MAPS_PROXY_MODE).describe("Proxy targeting per city search. Defaults to location (city/ZIP-group residential targeting); configured forces the service proxy; none is local debugging only."),
+  proxyZip: z.string().regex(/^\d{5}$/).optional().describe("Optional ZIP override for proxy targeting; normally omitted."),
+  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics.")
 };
+var ArtifactPointerOutputSchema = z.object({
+  artifactId: z.string(),
+  bytes: z.number().int().min(0),
+  expiresAt: z.string(),
+  preview: z.string()
+});
 var RankTrackerModeSchema = z.enum(["maps", "organic", "ai_overview", "paa"]);
 var RankTrackerBlueprintInputSchema = {
   projectName: z.string().min(1).optional().describe("Optional name for the rank tracker project, client, or campaign."),
-  targetDomain: z.string().min(1).optional().describe("Primary domain to track in organic results, AI Overview citations, and PAA sources, e.g. example.com."),
-  targetBusinessName: z.string().min(1).optional().describe("Primary Google Business Profile or brand/business name to match in Maps results. Required for reliable Maps rank tracking."),
-  trackingModes: z.array(RankTrackerModeSchema).min(1).max(4).default(["maps", "organic", "ai_overview", "paa"]).describe("Rank tracker surfaces to build. maps uses directory_workflow/maps_search. organic uses search_serp. ai_overview uses search_serp/harvest_paa. paa uses harvest_paa source presence."),
-  keywords: z.array(z.string().min(1)).max(200).default([]).describe("Seed keywords or service queries to track. Leave empty when the downstream AI should create the keyword table from user input."),
-  locations: z.array(z.string().min(1)).max(100).default([]).describe("Markets, cities, ZIPs, or service areas to track. Use city/state strings like Denver, CO for localized SERP and Maps checks."),
+  targetDomain: z.string().min(1).optional().describe("Primary domain to track in organic results, AI Overview citations, and PAA sources."),
+  targetBusinessName: z.string().min(1).optional().describe("Primary Google Business Profile/brand name to match in Maps results. Required for Maps tracking."),
+  trackingModes: z.array(RankTrackerModeSchema).min(1).max(4).default(["maps", "organic", "ai_overview", "paa"]).describe("Rank tracker surfaces to build: maps, organic, ai_overview, paa."),
+  keywords: z.array(z.string().min(1)).max(200).default([]).describe("Seed keywords or service queries to track. Leave empty to derive from user input downstream."),
+  locations: z.array(z.string().min(1)).max(100).default([]).describe('Markets, cities, ZIPs, or service areas to track, e.g. "Denver, CO".'),
   competitors: z.array(z.string().min(1)).max(100).default([]).describe("Optional competitor domains or business names to persist as comparison targets."),
-  database: z.enum(["postgres", "neon", "supabase", "sqlite", "mysql"]).default("postgres").describe("Database family the downstream AI should target when generating migrations."),
-  scheduleCadence: z.enum(["daily", "weekly", "monthly", "custom"]).default("weekly").describe("Default recurring rank check cadence for the generated cron/heartbeat plan."),
+  database: z.enum(["postgres", "neon", "supabase", "sqlite", "mysql"]).default("postgres").describe("Database family to target when generating migrations."),
+  scheduleCadence: z.enum(["daily", "weekly", "monthly", "custom"]).default("weekly").describe("Default recurring rank check cadence."),
   customCron: z.string().min(1).optional().describe("Cron expression to use when scheduleCadence is custom."),
-  timezone: z.string().min(1).default("UTC").describe("IANA timezone for scheduled rank checks, e.g. America/Denver."),
-  includeCron: z.boolean().default(true).describe("Include a cron or heartbeat worker plan. Keep true for production rank trackers."),
-  includeDashboard: z.boolean().default(true).describe("Include dashboard/reporting requirements in the generated prompt."),
-  includeAlerts: z.boolean().default(true).describe("Include alert rules for rank movement and SERP feature gains/losses."),
-  notes: z.string().max(4e3).optional().describe("Extra product, client, stack, or hosting requirements to include in the implementation prompt.")
+  timezone: z.string().min(1).default("UTC").describe("IANA timezone for scheduled rank checks."),
+  includeCron: z.boolean().default(true).describe("Include a cron/heartbeat worker plan."),
+  includeDashboard: z.boolean().default(true).describe("Include dashboard/reporting requirements."),
+  includeAlerts: z.boolean().default(true).describe("Include alert rules for rank movement and SERP feature changes."),
+  notes: z.string().max(4e3).optional().describe("Extra product, client, stack, or hosting requirements.")
 };
 var NullableString = z.string().nullable();
 var MapsSearchAttemptOutput = z.object({
@@ -3050,7 +3279,9 @@ var DirectoryWorkflowOutputSchema = {
     attempts: z.array(MapsSearchAttemptOutput),
     results: z.array(DirectoryMapsBusinessOutput)
   })),
-  durationMs: z.number().int().min(0)
+  durationMs: z.number().int().min(0),
+  truncatedCount: z.number().int().min(0).optional(),
+  artifact: ArtifactPointerOutputSchema.optional()
 };
 var RankTrackerToolPlanOutput = z.object({
   tool: z.string(),
@@ -3103,7 +3334,7 @@ var EntityIdsOutput = z.object({
     kgId: z.string().nullable(),
     cid: z.string().nullable(),
     gcid: z.string().nullable()
-  })).describe("Each entity actually named on the page, with whichever of its kgId/cid/gcid were found \u2014 the linked view. kgIds/cids/gcids below are the same IDs as flat deduplicated lists, without the name link, kept for backward compatibility."),
+  })).describe("Entities named on the page with their kgId/cid/gcid. Flat lists below are the same IDs deduplicated, kept for backward compatibility."),
   kgIds: z.array(z.string()),
   cids: z.array(z.string()),
   gcids: z.array(z.string())
@@ -3170,7 +3401,9 @@ var ExtractSiteOutputSchema = {
     title: NullableString,
     schemaTypes: z.array(z.string())
   })),
-  durationMs: z.number().min(0)
+  durationMs: z.number().min(0),
+  truncatedCount: z.number().int().min(0).optional(),
+  artifact: ArtifactPointerOutputSchema.optional()
 };
 var AuditSiteOutputSchema = {
   url: z.string(),
@@ -3190,7 +3423,8 @@ var AuditSiteOutputSchema = {
     orphans: z.number().int().min(0),
     brokenInternal: z.number().int().min(0),
     externalDomains: z.number().int().min(0)
-  })
+  }),
+  artifact: ArtifactPointerOutputSchema.optional()
 };
 var MapsPlaceIntelOutputSchema = {
   name: z.string(),
@@ -3262,7 +3496,9 @@ var MapSiteUrlsOutputSchema = {
     url: z.string(),
     status: z.number().int().nullable()
   })),
-  durationMs: z.number().min(0)
+  durationMs: z.number().min(0),
+  truncatedCount: z.number().int().min(0).optional(),
+  artifact: ArtifactPointerOutputSchema.optional()
 };
 var YoutubeHarvestOutputSchema = {
   mode: z.string(),
@@ -3570,7 +3806,7 @@ var WorkflowIdSchema = z.enum([
   "ai-overview-language"
 ]);
 var WorkflowListInputSchema = {
-  includeRecipes: z.boolean().default(true).describe("Include high-level AI-facing recipes such as market analysis, ICP research, forum/review research, brand design briefings, CRO audits, positioning briefs, content gaps, and AI search visibility audits.")
+  includeRecipes: z.boolean().default(true).describe("Include high-level AI-facing recipes (market analysis, ICP research, CRO audits, content gaps, etc).")
 };
 var WorkflowSuggestInputSchema = {
   goal: z.string().min(1).describe('The user goal or job to route, e.g. "market analysis for roofers in Tennessee", "ICP research for med spas", "CRO audit for this URL", or "brand design briefing".'),
@@ -3583,12 +3819,12 @@ var WorkflowSuggestInputSchema = {
   maxSuggestions: z.number().int().min(1).max(8).default(3).describe("Number of matching workflow recipes to return.")
 };
 var WorkflowRunInputSchema = {
-  workflowId: WorkflowIdSchema.describe("Workflow to run: directory, get-leads, agent-packet, local-competitive-audit, map-comparison, serp-comparison, paa-expansion-brief, ai-overview-language. Use only these values; call workflow_list or workflow_suggest first when unsure."),
-  input: z.record(z.unknown()).default({}).describe("Workflow-specific input object. Examples: get-leads uses {query, location, maxResults?, enrichWebsites?, hydrateReviewCounts?}; agent-packet uses {keyword, domain?, location?, maxQuestions?}; local-competitive-audit uses {query, state, minPopulation?, maxCities?, maxResultsPerCity?, hydrateTop?, maxReviews?}; serp-comparison uses {keyword, domain?, url?, location?, extractTop?}."),
+  workflowId: WorkflowIdSchema.describe("Workflow to run. Call workflow_list or workflow_suggest first when unsure."),
+  input: z.record(z.unknown()).default({}).describe("Workflow-specific input object; shape depends on workflowId. Call workflow_list or workflow_suggest to see required fields."),
   webhookUrl: z.string().url().optional().describe("Optional HTTPS webhook to receive the completed hosted workflow run event.")
 };
 var WorkflowStepInputSchema = {
-  runId: z.string().min(1).describe("Workflow run id returned by workflow_run, workflow_step, or workflow_status. Use only a returned runId; do not construct one yourself. Advances the run by exactly one step (one logical leg, e.g. one live harvest).")
+  runId: z.string().min(1).describe("Workflow run id returned by workflow_run/workflow_step/workflow_status. Advances the run by exactly one step.")
 };
 var WorkflowStatusInputSchema = {
   runId: z.string().min(1).describe("Workflow run id returned by workflow_run, workflow_step, or workflow_status. Use only a returned runId; do not construct one yourself.")
@@ -3596,7 +3832,7 @@ var WorkflowStatusInputSchema = {
 var WorkflowArtifactReadInputSchema = {
   runId: z.string().min(1).describe("Workflow run id returned by workflow_run, workflow_step, or workflow_status. Use only a returned runId; do not construct one yourself."),
   artifactId: z.string().min(1).describe("Artifact id from the run artifact list returned by workflow_run, workflow_step, or workflow_status. Use only a returned artifactId; do not construct one yourself."),
-  maxBytes: z.number().int().min(1e3).max(1e6).default(2e5).describe("Maximum bytes of artifact text to return inline. Use lower values for large CSV/JSON artifacts; call again with the downloadUrl if needed outside MCP.")
+  maxBytes: z.number().int().min(1e3).max(1e6).default(2e5).describe("Maximum bytes of artifact text to return inline.")
 };
 var WorkflowRecipeOutput = z.object({
   id: z.string(),
@@ -3655,28 +3891,28 @@ var WorkflowArtifactReadOutputSchema = {
   text: z.string()
 };
 var SearchSerpInputSchema = {
-  query: z.string().min(1).describe('The search query. For localized results, KEEP the place in the query text \u2014 "best dentist in Brooklyn NY" stays query="best dentist Brooklyn NY" (and also set location="Brooklyn, NY"). City-in-query is what localizes results reliably; do not strip it out.'),
-  location: z.string().optional().describe("City, region, or country for geo signals (uule + location evidence). Set it alongside a city-in-query wording; on its own it does NOT reliably localize results."),
+  query: z.string().min(1).describe('The search query. KEEP the place in the query text for localized results (e.g. "best dentist Brooklyn NY") and also set location \u2014 city-in-query is what localizes reliably.'),
+  location: z.string().optional().describe("City, region, or country for geo signals. Set alongside city-in-query wording; alone it does NOT reliably localize."),
   gl: z.string().length(2).default("us").describe("Google country code inferred from location or user language."),
   hl: z.string().default("en").describe("Google interface/content language inferred from user request."),
-  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use desktop by default; use mobile only when the user asks for mobile rankings."),
-  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress \u2014 reliable). Do NOT set "location" just because the user named a city; localized results come from city-in-query wording. "location" forces a residential geo-IP for rank-tracking fidelity only, is frequently CAPTCHA-blocked, and should be used only when the user explicitly demands as-seen-from-the-city results and accepts failures.'),
-  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location" (see its warning); ignored otherwise.'),
-  debug: z.boolean().default(false).describe("Include sanitized browser/session/location diagnostics in the response. Use true when debugging localization, CAPTCHA, or proxy behavior."),
-  pages: z.number().int().min(1).max(2).default(1).describe("Number of result pages to fetch (1\u20132)")
+  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use mobile only for mobile rankings."),
+  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress). Do NOT set "location" just because a city was named \u2014 that comes from city-in-query wording. "location" forces residential geo-IP for rank-tracking fidelity, is frequently CAPTCHA-blocked, and accepts failures.'),
+  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location".'),
+  debug: z.boolean().default(false).describe("Include sanitized diagnostics for debugging localization, CAPTCHA, or proxy behavior."),
+  pages: z.number().int().min(1).max(2).default(1).describe("Number of result pages to fetch (1\u20132).")
 };
 var CaptureSerpSnapshotInputSchema = {
-  query: z.string().min(1).describe('Search query to capture as a structured SERP Intelligence snapshot. For localized captures, KEEP the place in the query text (e.g. "botox clinic austin tx") and also set location; city-in-query is what localizes results reliably.'),
-  location: z.string().optional().describe("City, region, country, or service area used for localized Google results. MCP Scraper records location evidence; UULE alone is not proof of localization."),
-  gl: z.string().length(2).default("us").describe("Google country code inferred from the requested market, e.g. us, gb, ca, au."),
+  query: z.string().min(1).describe('Search query to capture. KEEP the place in the query text for localized captures (e.g. "botox clinic austin tx") and also set location.'),
+  location: z.string().optional().describe("City, region, country, or service area for localized Google results."),
+  gl: z.string().length(2).default("us").describe("Google country code inferred from the requested market."),
   hl: z.string().default("en").describe("Google interface/content language inferred from the user request."),
-  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use mobile only when the user asks for mobile rankings or mobile SERP evidence."),
-  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress \u2014 reliable). Do NOT set "location" just because a city was named; localized captures come from city-in-query wording. "location" forces a residential geo-IP (rank-tracking fidelity only), is frequently CAPTCHA-blocked, and should be used only when the user explicitly needs as-seen-from-the-city evidence and accepts failures.'),
-  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location" (see its warning); ignored otherwise.'),
-  pages: z.number().int().min(1).max(2).default(1).describe("Number of Google result pages to capture. Use 1 normally and 2 only when the user needs deeper ranking evidence."),
-  debug: z.boolean().default(false).describe("Include sanitized browser, proxy, and location diagnostics. Use true when debugging localization, CAPTCHA, proxy selection, or capture reliability."),
-  includePageSnapshots: z.boolean().default(false).describe("Also capture ranking-page snapshots for selected SERP URLs through the same product capture path."),
-  pageSnapshotLimit: z.number().int().min(0).max(10).default(0).describe("Maximum ranking-page snapshots to capture when includePageSnapshots is true. Use 0 when only SERP evidence is needed.")
+  device: z.enum(["desktop", "mobile"]).default("desktop").describe("SERP device context. Use mobile only for mobile rankings/evidence."),
+  proxyMode: z.enum(["location", "configured", "none"]).default(DEFAULT_PROXY_MODE).describe('Leave unset (clean egress). Do NOT set "location" just because a city was named \u2014 that comes from city-in-query wording. "location" forces residential geo-IP, is frequently CAPTCHA-blocked, and accepts failures.'),
+  proxyZip: z.string().regex(/^\d{5}$/).optional().describe('US ZIP for residential geo-IP targeting. Only meaningful with proxyMode "location".'),
+  pages: z.number().int().min(1).max(2).default(1).describe("Google result pages to capture. Use 2 only for deeper ranking evidence."),
+  debug: z.boolean().default(false).describe("Include sanitized browser/proxy/location diagnostics."),
+  includePageSnapshots: z.boolean().default(false).describe("Also capture ranking-page snapshots for selected SERP URLs."),
+  pageSnapshotLimit: z.number().int().min(0).max(10).default(0).describe("Maximum ranking-page snapshots when includePageSnapshots is true.")
 };
 var ScreenshotInputSchema = {
   url: z.string().url().describe("URL to capture as a full-page screenshot. Use http or https. Pass allowLocal: true to capture localhost or private-network URLs during development."),
@@ -3684,15 +3920,26 @@ var ScreenshotInputSchema = {
   allowLocal: z.boolean().default(false).describe("Allow localhost and private-network URLs (127.x, 192.168.x, 10.x, etc.). For local development only \u2014 not for production use.")
 };
 var CaptureSerpPageSnapshotsInputSchema = {
-  urls: z.array(z.string().url()).min(1).max(25).describe("Public HTTP/HTTPS URLs to capture as SERP Intelligence page snapshots. Do not pass localhost, private IPs, file URLs, or internal admin URLs."),
+  urls: z.array(z.string().url()).min(1).max(25).describe("Public HTTP/HTTPS URLs to capture. Do not pass localhost, private IPs, file URLs, or internal admin URLs."),
   targets: z.array(z.object({
     url: z.string().url().describe("Public HTTP/HTTPS URL to capture."),
-    sourceKind: z.enum(["organic", "ai_citation", "local_pack_website", "configured_target", "site_subject"]).default("configured_target").describe("Why this page is being captured for SERP Intelligence evidence."),
+    sourceKind: z.enum(["organic", "ai_citation", "local_pack_website", "configured_target", "site_subject"]).default("configured_target").describe("Why this page is being captured."),
     sourcePosition: z.number().int().min(1).optional().describe("Ranking or citation position when the page came from SERP evidence.")
-  }).strict()).min(1).max(25).optional().describe("Structured page snapshot targets. Use this instead of urls when source kind or position should be preserved."),
-  maxConcurrency: z.number().int().min(1).max(5).default(2).describe("Parallel page captures. Use 2 normally; higher values can increase proxy/browser pressure."),
-  timeoutMs: z.number().int().min(1e3).max(6e4).default(15e3).describe("Per-page capture timeout in milliseconds. Increase for slow pages; timeout artifacts are returned as structured capture failures."),
-  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics for page snapshot debugging. Use true for capture, network, or proxy troubleshooting.")
+  }).strict()).min(1).max(25).optional().describe("Structured targets. Use instead of urls when source kind or position should be preserved."),
+  maxConcurrency: z.number().int().min(1).max(5).default(2).describe("Parallel page captures."),
+  timeoutMs: z.number().int().min(1e3).max(6e4).default(15e3).describe("Per-page capture timeout in milliseconds; timeouts return as structured capture failures."),
+  debug: z.boolean().default(false).describe("Include sanitized browser/proxy diagnostics.")
+};
+var ReportArtifactReadInputSchema = {
+  artifactId: z.string().min(1).describe("Artifact id returned inline by a tool whose result was too large to inline. Use only a returned artifactId; do not construct one yourself."),
+  offset: z.number().int().min(0).default(0).describe("Byte offset to start reading from. Pass the previous call's nextOffset to continue."),
+  maxBytes: z.number().int().min(1e3).max(1e5).default(2e4).describe("Maximum bytes of artifact text to return in this window.")
+};
+var ReportArtifactReadOutputSchema = {
+  artifactId: z.string(),
+  text: z.string(),
+  totalBytes: z.number().int().min(0),
+  nextOffset: z.number().int().min(0).nullable()
 };
 
 // src/mcp/rank-tracker-blueprint.ts
@@ -4019,6 +4266,9 @@ function buildRankTrackerBlueprint(input) {
 }
 
 // src/mcp/paa-mcp-server.ts
+function hashOwnerId(callerKey) {
+  return createHash("sha256").update(callerKey).digest("hex").slice(0, 24);
+}
 function liveWebToolAnnotations(title) {
   return {
     title,
@@ -4031,16 +4281,16 @@ function liveWebToolAnnotations(title) {
 function registerSerpIntelligenceCaptureTools(server, executor) {
   server.registerTool("capture_serp_snapshot", {
     title: "SERP Intelligence Snapshot",
-    description: "Capture a structured SERP Intelligence Google snapshot through POST /serp-intelligence/capture, the same product capture path used by Phoenix. Split query from location and infer gl/hl. Routing is automatic \u2014 leave proxyMode unset. Set debug true when investigating location evidence, proxy behavior, CAPTCHA, or capture reliability.",
+    description: "Capture a structured SERP Intelligence snapshot of a Google query \u2014 the persistent evidence format used by rank-tracking and comparison pipelines. Split query from location; leave proxyMode unset.",
     inputSchema: CaptureSerpSnapshotInputSchema,
-    outputSchema: CaptureSerpSnapshotOutputSchema,
+    outputSchema: recordOutputSchema("capture_serp_snapshot", CaptureSerpSnapshotOutputSchema),
     annotations: liveWebToolAnnotations("SERP Intelligence Snapshot")
   }, async (input) => formatCaptureSerpSnapshot(await executor.captureSerpSnapshot(input), input));
   server.registerTool("capture_serp_page_snapshots", {
     title: "SERP Intelligence Page Snapshots",
-    description: "Capture public ranking-page evidence through POST /serp-intelligence/page-snapshots, the same product page snapshot path used by Phoenix. Provide urls for simple captures or targets when preserving organic, AI citation, local-pack, configured target, or site-subject source metadata. Private IPs, localhost, file URLs, and internal URLs are rejected by the service. Use timeoutMs for slow pages and debug true for sanitized proxy/browser diagnostics.",
+    description: "Capture public ranking pages as SERP Intelligence page snapshots \u2014 persistent page evidence linked to a captured SERP. Provide urls, or targets to preserve source metadata. Private IPs, localhost, file, and internal URLs are rejected.",
     inputSchema: CaptureSerpPageSnapshotsInputSchema,
-    outputSchema: CaptureSerpPageSnapshotsOutputSchema,
+    outputSchema: recordOutputSchema("capture_serp_page_snapshots", CaptureSerpPageSnapshotsOutputSchema),
     annotations: liveWebToolAnnotations("SERP Intelligence Page Snapshots")
   }, async (input) => formatCaptureSerpPageSnapshots(await executor.captureSerpPageSnapshots(input), input));
 }
@@ -4056,7 +4306,7 @@ function localPlanningToolAnnotations(title) {
 function listSavedReports() {
   try {
     const dir = outputBaseDir();
-    return readdirSync(dir).filter((f) => f.endsWith(".md")).map((f) => ({ filename: f, mtimeMs: statSync(join2(dir, f)).mtimeMs })).sort((a, b) => b.mtimeMs - a.mtimeMs).slice(0, 100);
+    return readdirSync(dir).filter((f) => f.endsWith(".md")).map((f) => ({ filename: f, mtimeMs: statSync(join3(dir, f)).mtimeMs })).sort((a, b) => b.mtimeMs - a.mtimeMs).slice(0, 100);
   } catch {
     return [];
   }
@@ -4082,236 +4332,257 @@ function registerSavedReportResources(server) {
       const requested = Array.isArray(variables.filename) ? variables.filename[0] : variables.filename;
       const filename = basename(decodeURIComponent(String(requested ?? "")));
       if (!filename.endsWith(".md")) throw new Error("Only saved .md reports can be read");
-      const text = readFileSync(join2(outputBaseDir(), filename), "utf8");
+      const text = readFileSync(join3(outputBaseDir(), filename), "utf8");
       return { contents: [{ uri: uri.href, mimeType: "text/markdown", text }] };
     }
   );
 }
 function buildPaaExtractorMcpServer(executor, options = {}) {
-  const server = new McpServer({ name: "mcp-scraper", version: PACKAGE_VERSION }, { instructions: SERVER_INSTRUCTIONS });
+  const server = new McpServer({ name: "mcp-scraper", version: PACKAGE_VERSION }, { instructions: serverInstructions(options.savesReportsLocally !== false) });
   registerPaaExtractorMcpTools(server, executor, options);
   return server;
 }
 function registerPaaExtractorMcpTools(server, executor, options = {}) {
   const savesReports = options.savesReportsLocally !== false;
-  const reportNote = savesReports ? " Saves a full Markdown report to disk." : " Reports are returned inline; no files are saved on this hosted endpoint.";
-  const withReportNote = (description) => `${description}${reportNote}`;
+  const fileBehavior = (local, hosted) => savesReports ? local : hosted;
+  const ownerId = options.ownerId ?? "local";
+  const ctx = { hosted: !savesReports, ownerId };
   if (savesReports) registerSavedReportResources(server);
   server.registerTool("harvest_paa", {
     title: "Google PAA + SERP Harvest",
-    description: withReportNote('Best default tool for Google search research. Extracts People Also Ask questions plus answers/source URLs, organic SERP, local pack when present, entity IDs (CID/GCID/KG MID), and AI Overview. Infer the user language: split topic from location (e.g. "best hvac company in Denver CO" => query "best hvac company", location "Denver, CO", gl "us", hl "en"). Routing is automatic \u2014 leave proxyMode unset. If a deep harvest exceeds the time budget it returns the questions gathered so far (partial), billed per returned question. Use maxQuestions 30 normally, 100-200 for "full", "deep", "all", or comprehensive research. Deep harvests above 100 questions can run for several minutes with no interim progress \u2014 warn the user before starting one and keep maxQuestions at or below 100 unless they explicitly want a deep harvest. Credits are charged by extracted question; unused request hold is refunded.'),
+    description: "Best default tool for Google search research: People Also Ask questions with answers/sources, organic SERP, local pack, entity IDs, and AI Overview. Split topic from location; leave proxyMode unset. Warn the user before maxQuestions above 100 \u2014 deep harvests can run several minutes with no interim progress, billed per extracted question.",
     inputSchema: HarvestPaaInputSchema,
-    outputSchema: HarvestPaaOutputSchema,
+    outputSchema: recordOutputSchema("harvest_paa", HarvestPaaOutputSchema),
     annotations: liveWebToolAnnotations("Google PAA + SERP Harvest")
   }, async (input) => formatHarvestPaa(await executor.harvestPaa(input), input));
   server.registerTool("search_serp", {
     title: "Google SERP Lookup",
-    description: withReportNote("Fast Google SERP lookup without PAA expansion. Use when the user asks for rankings, organic results, local pack, quick SERP, or positions. Split topic from location and infer gl/hl from the user request. Routing is automatic \u2014 leave proxyMode unset; set location only to force local-pack targeting."),
+    description: "Fast Google SERP lookup without PAA expansion \u2014 rankings, organic results, local pack, positions. Split topic from location; leave proxyMode unset.",
     inputSchema: SearchSerpInputSchema,
-    outputSchema: SearchSerpOutputSchema,
+    outputSchema: recordOutputSchema("search_serp", SearchSerpOutputSchema),
     annotations: liveWebToolAnnotations("Google SERP Lookup")
   }, async (input) => formatSearchSerp(await executor.searchSerp(input), input));
   server.registerTool("extract_url", {
     title: "Single URL Extract",
-    description: withReportNote("Extract structured data from one public URL when the user provides one page, asks to inspect/scrape a page, or needs page content, schema, headings, metadata, screenshots, branding, or media assets. Returns structured page fields plus artifact handles for saved reports/screenshots/media when requested. Set depositToVault:true to also save the full page into the user's MCP Memory vault server-side (the full body is NOT returned to chat). Use map_site_urls before extracting a site inventory; use extract_site for multi-page crawling."),
+    description: "Extract structured data from one public URL: content, schema, headings, metadata, screenshots, branding, or media assets. Set depositToVault:true to save the full page into the user's MCP Memory vault server-side (not returned to chat).",
     inputSchema: ExtractUrlInputSchema,
-    outputSchema: ExtractUrlOutputSchema,
+    outputSchema: recordOutputSchema("extract_url", ExtractUrlOutputSchema),
     annotations: liveWebToolAnnotations("Single URL Extract")
   }, async (input) => formatExtractUrl(await executor.extractUrl(input), input));
   server.registerTool("map_site_urls", {
     title: "Site URL Map",
-    description: withReportNote("Map/crawl a public website when the user asks for a sitemap, URL inventory, broken-link scan, redirect scan, or crawl planning. Returns internal URLs with HTTP status and truncation metadata. Supports up to maxUrls=10000; maps over 500 URLs write the complete inventory to a local CSV file and return a summary plus the file path instead of the full list inline. Use this before extract_site when choosing pages for an audit; use extract_url for one known page."),
+    description: `Map/crawl a public website for a sitemap, URL inventory, or broken-link scan. Returns internal URLs with HTTP status; ${fileBehavior("maps over 500 URLs are written to a local CSV file instead of inlined.", "large results are stored as a retrievable artifact \u2014 you get an inline summary plus an artifactId for report_artifact_read.")}`,
     inputSchema: MapSiteUrlsInputSchema,
-    outputSchema: MapSiteUrlsOutputSchema,
+    outputSchema: recordOutputSchema("map_site_urls", MapSiteUrlsOutputSchema),
     annotations: liveWebToolAnnotations("Site URL Map")
-  }, async (input) => formatMapSiteUrls(await executor.mapSiteUrls(input), input));
+  }, async (input) => formatMapSiteUrls(await executor.mapSiteUrls(input), input, ctx));
   server.registerTool("extract_site", {
     title: "Multi-Page Site Content Crawl",
-    description: withReportNote("Crawl a public website and return the page CONTENT across multiple pages (map + scrape). Returns per-page titles and Markdown content. Supports up to maxPages=10000; bulk crawls over 25 pages switch to folder mode: each page is saved as its own Markdown file in a local folder and the response returns a summary plus the folder path instead of inlining all content. This tool returns content only \u2014 for a technical SEO audit (issues, internal link graph, indexability, heading breakdown, image sizes/formats) use audit_site. Use map_site_urls first when URL selection matters; use extract_url for one page."),
+    description: `Crawl a public website and return page CONTENT (Markdown) across multiple pages. ${fileBehavior("Bulk crawls over 25 pages are saved as per-page Markdown files in a local folder instead of inlined.", "Large results are stored as a retrievable artifact \u2014 you get an inline summary plus an artifactId for report_artifact_read.")} Content only \u2014 for a technical SEO audit use audit_site instead.`,
     inputSchema: ExtractSiteInputSchema,
-    outputSchema: ExtractSiteOutputSchema,
+    outputSchema: recordOutputSchema("extract_site", ExtractSiteOutputSchema),
     annotations: liveWebToolAnnotations("Multi-Page Site Content Crawl")
-  }, async (input) => formatExtractSite(await executor.extractSite(input), input));
+  }, async (input) => formatExtractSite(await executor.extractSite(input), input, ctx));
   server.registerTool("audit_site", {
     title: "Technical SEO Audit",
-    description: withReportNote("Run a full technical SEO audit (Screaming-Frog-style) on a public website. Crawls up to maxPages and produces a complete on-page + crawl analysis: per-page title/meta lengths, heading breakdown, indexability and canonical status, schema, an internal link graph (inlinks/orphans/crawl-depth), a link analysis (top internal pages by inlinks, top external domains by link count, broken/orphan counts), an issues report, and an image audit (byte size, format, over-100KB and legacy-format flags). Writes everything to a local folder (report.md, issues.json, pages.jsonl, links.jsonl, link-metrics.jsonl, link-report.md, links-summary.json, external-domains.json, images.jsonl) plus per-page content, and returns a headline summary plus the folder path. Use this when the user wants an audit or analysis; use extract_site when they just want page content."),
+    description: `Run a full technical SEO audit (Screaming-Frog-style) on a public website: on-page issues, internal link graph, indexability, heading/image analysis. ${fileBehavior("Writes a folder of analysis files plus per-page content, and returns a summary plus the folder path.", "Large results are stored as a retrievable artifact \u2014 you get an inline summary plus an artifactId for report_artifact_read.")} Use extract_site instead for plain page content.`,
     inputSchema: AuditSiteInputSchema,
-    outputSchema: AuditSiteOutputSchema,
+    outputSchema: recordOutputSchema("audit_site", AuditSiteOutputSchema),
     annotations: liveWebToolAnnotations("Technical SEO Audit")
-  }, async (input) => formatAuditSite(await executor.auditSite(input), input));
+  }, async (input) => formatAuditSite(await executor.auditSite(input), input, ctx));
   server.registerTool("youtube_harvest", {
     title: "YouTube Video Harvest",
-    description: withReportNote('Harvest YouTube video metadata when the user wants to find videos by topic, inspect a channel library, compare video angles, or get videoIds for later transcription. Use mode "search" for keyword/topic requests and mode "channel" for @handles, channel IDs, or channel URLs. Returns titles, views, durations, URLs, and videoIds for follow-up transcription. Use youtube_transcribe after selecting one video.'),
+    description: 'Harvest YouTube video metadata by topic search or channel library. Use mode "search" for keyword/topic requests, mode "channel" for @handles/channel IDs/URLs. Returns titles, views, durations, and videoIds.',
     inputSchema: YoutubeHarvestInputSchema,
-    outputSchema: YoutubeHarvestOutputSchema,
+    outputSchema: recordOutputSchema("youtube_harvest", YoutubeHarvestOutputSchema),
     annotations: liveWebToolAnnotations("YouTube Video Harvest")
   }, async (input) => formatYoutubeHarvest(await executor.youtubeHarvest(input), input));
   server.registerTool("youtube_transcribe", {
     title: "YouTube Transcription",
-    description: withReportNote("Fetch and transcribe captions from a YouTube video. Use this when the user asks what was said in a YouTube video, wants claims/offers/lessons extracted from a video, or provides a YouTube URL for transcript work. Returns full transcript, timestamped chunks, word count, and resolvedInputs. Pass videoId from youtube_harvest results, or pass url when the user pasted a YouTube URL. Use youtube_harvest first when you need to discover videos by topic/channel."),
+    description: "Fetch and transcribe captions from a YouTube video. Pass videoId from youtube_harvest, or a url the user pasted. Returns full transcript, timestamped chunks, and word count.",
     inputSchema: YoutubeTranscribeInputSchema,
-    outputSchema: YoutubeTranscribeOutputSchema,
+    outputSchema: recordOutputSchema("youtube_transcribe", YoutubeTranscribeOutputSchema),
     annotations: liveWebToolAnnotations("YouTube Transcription")
   }, async (input) => formatYoutubeTranscribe(await executor.youtubeTranscribe(input), input));
   server.registerTool("facebook_page_intel", {
     title: "Facebook Advertiser Ad Intel",
-    description: withReportNote("Harvest ads from a Facebook advertiser when the user wants current ad copy, creative angles, CTAs, video URLs, or competitive ad intelligence for one brand/page. Returns ad copy, headlines, CTAs, creative type, status, landing URLs, and direct ad video URLs ready for facebook_ad_transcribe. Accepts pageId, libraryId, or a brand/advertiser name as query. Use facebook_ad_search first when the advertiser handle is unknown. For normal public Facebook reels/posts/watch/share URLs, use facebook_video_transcribe instead."),
+    description: "Harvest ads from a Facebook advertiser: copy, creative angles, CTAs, and direct video URLs ready for facebook_ad_transcribe. Accepts pageId, libraryId, or a brand/advertiser name.",
     inputSchema: FacebookPageIntelInputSchema,
-    outputSchema: FacebookPageIntelOutputSchema,
+    outputSchema: recordOutputSchema("facebook_page_intel", FacebookPageIntelOutputSchema),
     annotations: liveWebToolAnnotations("Facebook Advertiser Ad Intel")
   }, async (input) => formatFacebookPageIntel(await executor.facebookPageIntel(input), input));
   server.registerTool("facebook_ad_search", {
     title: "Facebook Ad Library Search",
-    description: withReportNote("Search Facebook Ad Library when the user wants to find advertisers by brand, competitor, niche, or keyword. Returns advertisers with ad counts and library IDs. Use this to discover competitor/page handles, then pass libraryId or pageId to facebook_page_intel for ad details."),
+    description: "Search Facebook Ad Library to find advertisers by brand, competitor, niche, or keyword. Returns advertisers with ad counts and library IDs.",
     inputSchema: FacebookAdSearchInputSchema,
-    outputSchema: FacebookAdSearchOutputSchema,
+    outputSchema: recordOutputSchema("facebook_ad_search", FacebookAdSearchOutputSchema),
     annotations: liveWebToolAnnotations("Facebook Ad Library Search")
   }, async (input) => formatFacebookAdSearch(await executor.facebookAdSearch(input), input));
   server.registerTool("reddit_thread", {
     title: "Reddit Thread + Comments",
-    description: withReportNote("Capture a Reddit post and its comment tree from any reddit.com thread URL. Use this when the user wants Reddit comments, opinions, discussion, or audience voice from a specific thread (find threads first with search_serp, which surfaces reddit.com/comments links). Returns the post title, author, score, body, and a list of comments with author, score, nesting depth, and text. Handles Reddit's bot protection automatically (fetches via old.reddit through a residential proxy with retries); pass maxComments to cap the list."),
+    description: "Capture a Reddit post and its comment tree from a reddit.com thread URL \u2014 comments, opinions, audience voice. Handles Reddit's bot protection automatically; pass maxComments to cap the list.",
     inputSchema: RedditThreadInputSchema,
-    outputSchema: RedditThreadOutputSchema,
+    outputSchema: recordOutputSchema("reddit_thread", RedditThreadOutputSchema),
     annotations: liveWebToolAnnotations("Reddit Thread + Comments")
   }, async (input) => formatRedditThread(await executor.redditThread(input), input));
   server.registerTool("video_frame_analysis", {
     title: "Video Breakdown (frame-by-frame + transcript)",
-    description: `Produce a deep, structured breakdown of a video: what happens on screen, what is said, how it is paced, and how to make one like it. Use this when the user wants to analyze, break down, study, reverse-engineer, or "figure out how they made" a video (an ad, a YouTube video, a reel, a competitor's content). It samples up to 120 frames spread across the whole video (up to 30 minutes), runs vision AI on every frame, transcribes the audio, and returns a report with: Summary, Pacing & Energy, Words-per-Minute & delivery, Topic Outline, Key Points, Hook Analysis, Visual Style, and a step-by-step "How to Replicate This Video" recipe \u2014 all quality-checked. Pass a DIRECT video file URL (.mp4/.webm/.mov); for a YouTube/Facebook/Instagram link, resolve it to a direct media URL first (e.g. facebook_page_intel returns a videoUrl). This runs asynchronously and costs a flat $1 (refunded if it fails): it returns a runId immediately \u2014 then call video_frame_analysis_status with that runId every few seconds until status is "done" to get the finished report.`,
+    description: "Produce a deep frame-by-frame + transcript breakdown of a video \u2014 pacing, hook, visual style, and how to replicate it. Pass a DIRECT video file URL (.mp4/.webm/.mov), not a page URL. Runs asynchronously and costs a flat $1 (refunded on failure): returns a runId immediately; poll video_frame_analysis_status until done.",
     inputSchema: VideoFrameAnalysisInputSchema,
-    outputSchema: VideoFrameAnalysisOutputSchema,
+    outputSchema: recordOutputSchema("video_frame_analysis", VideoFrameAnalysisOutputSchema),
     annotations: { title: "Video Breakdown", readOnlyHint: false, destructiveHint: false, idempotentHint: false, openWorldHint: true }
   }, async (input) => executor.videoFrameAnalysis(input));
   server.registerTool("video_frame_analysis_status", {
     title: "Video Breakdown Status",
-    description: 'Check the progress of a video breakdown started with video_frame_analysis, using the runId it returned. Poll this every few seconds. While running it reports the current phase and how many frames have been analyzed; when status is "done" it returns the full breakdown report (Markdown) and the memory vault path where it was saved. Free to call. Stop polling when status is "done" or "failed".',
+    description: 'Check progress of a video breakdown started with video_frame_analysis, using its runId. Free to call. When status is "done" it returns the full report and vault path; stop polling on "done" or "failed".',
     inputSchema: VideoFrameAnalysisStatusInputSchema,
-    outputSchema: VideoFrameAnalysisStatusOutputSchema,
+    outputSchema: recordOutputSchema("video_frame_analysis_status", VideoFrameAnalysisStatusOutputSchema),
     annotations: { title: "Video Breakdown Status", readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: false }
   }, async (input) => executor.videoFrameAnalysisStatus(input));
   server.registerTool("facebook_ad_transcribe", {
     title: "Facebook Ad Transcription",
-    description: "Transcribe audio from a Facebook ad video CDN URL. Use this when facebook_page_intel returned a direct videoUrl and the user asks what the ad says, what claims it makes, or wants ad-message extraction. Returns full transcript, timestamped chunks, word count, and resolvedInputs. Use only with the direct videoUrl value from facebook_page_intel results; do not pass public Facebook post/reel/share URLs. Use facebook_video_transcribe for organic Facebook URLs, and facebook_page_intel first when you only have a brand/page/ad library handle.",
+    description: "Transcribe audio from a Facebook ad video CDN URL returned by facebook_page_intel. Use only that direct videoUrl value \u2014 do not pass public Facebook post/reel/share URLs (use facebook_video_transcribe for those).",
     inputSchema: FacebookAdTranscribeInputSchema,
-    outputSchema: FacebookAdTranscribeOutputSchema,
+    outputSchema: recordOutputSchema("facebook_ad_transcribe", FacebookAdTranscribeOutputSchema),
     annotations: liveWebToolAnnotations("Facebook Ad Transcription")
   }, async (input) => formatFacebookAdTranscribe(await executor.facebookAdTranscribe(input), input));
   server.registerTool("google_ads_search", {
     title: "Google Ads Transparency Search",
-    description: withReportNote("Search the Google Ads Transparency Center to find advertisers running Google ads (Search, Display, YouTube) by domain or brand/advertiser name. Pass a domain like getviktor.com to see every advertiser account pointing ads at that site, or an advertiser name. Returns advertisers with their advertiser ID and an approximate ad count. Use this to discover an advertiserId, then pass it (or the domain) to google_ads_page_intel for the actual ad creatives."),
+    description: "Search the Google Ads Transparency Center to find advertisers by domain or brand name. Returns advertisers with advertiser ID and approximate ad count, to pass to google_ads_page_intel.",
     inputSchema: GoogleAdsSearchInputSchema,
-    outputSchema: GoogleAdsSearchOutputSchema,
+    outputSchema: recordOutputSchema("google_ads_search", GoogleAdsSearchOutputSchema),
     annotations: liveWebToolAnnotations("Google Ads Transparency Search")
   }, async (input) => formatGoogleAdsSearch(await executor.googleAdsSearch(input), input));
   server.registerTool("google_ads_page_intel", {
     title: "Google Ads Advertiser Intel",
-    description: withReportNote("Harvest an advertiser's ads from the Google Ads Transparency Center when the user wants a competitor's Google/YouTube ad creatives, copy angles, image URLs, or video ads. Accepts an advertiserId (from google_ads_search) or a domain directly. Returns each creative's format (text/image/video), image URLs, landing info, a detail URL, and \u2014 for video ads \u2014 the YouTube video ID and/or a direct video URL ready for transcription. Use google_ads_search first when you only have a brand name. Transcribe video ads with google_ads_transcribe (video URL) or youtube_transcribe (YouTube ID)."),
+    description: "Harvest an advertiser's ad creatives from the Google Ads Transparency Center: format, image URLs, and \u2014 for video ads \u2014 a YouTube video ID or direct video URL. Accepts advertiserId (from google_ads_search) or a domain.",
     inputSchema: GoogleAdsPageIntelInputSchema,
-    outputSchema: GoogleAdsPageIntelOutputSchema,
+    outputSchema: recordOutputSchema("google_ads_page_intel", GoogleAdsPageIntelOutputSchema),
     annotations: liveWebToolAnnotations("Google Ads Advertiser Intel")
   }, async (input) => formatGoogleAdsPageIntel(await executor.googleAdsPageIntel(input), input));
   server.registerTool("google_ads_transcribe", {
     title: "Google Ad Video Transcription",
-    description: "Transcribe audio from a Google video ad. Use this when google_ads_page_intel returned a direct videoUrl (a googlevideo.com playback URL) for a video creative and the user asks what the ad says or wants its claims/messaging. Returns full transcript, timestamped chunks, and word count. For YouTube-hosted ads, pass the returned youtubeVideoId to youtube_transcribe instead; run google_ads_page_intel first when you only have an advertiser/domain.",
+    description: "Transcribe audio from a Google video ad's direct videoUrl (a googlevideo.com playback URL) returned by google_ads_page_intel. For YouTube-hosted ads, use youtube_transcribe with the returned youtubeVideoId instead.",
     inputSchema: GoogleAdsTranscribeInputSchema,
-    outputSchema: GoogleAdsTranscribeOutputSchema,
+    outputSchema: recordOutputSchema("google_ads_transcribe", GoogleAdsTranscribeOutputSchema),
     annotations: liveWebToolAnnotations("Google Ad Video Transcription")
   }, async (input) => formatGoogleAdsTranscribe(await executor.googleAdsTranscribe(input), input));
   server.registerTool("facebook_video_transcribe", {
     title: "Facebook Organic Video Transcription",
-    description: withReportNote("Transcribe audio from an organic Facebook reel, video, watch, post, or share URL, including fb.watch links. Use this when the user pastes a normal Facebook video page URL and wants the transcript or downloadable MP4. Renders the Facebook page in a browser, selects the best matching public Facebook CDN MP4 URL from page state, then returns sourceUrl, resolved pageUrl, videoId, ownerName, selectedQuality, bitrate, videoDurationSec, extracted MP4 URL, full transcript, and timestamped chunks."),
+    description: "Transcribe audio from an organic Facebook reel/video/post/share URL (including fb.watch). Renders the page, selects the best public CDN MP4, and returns the transcript plus resolved video metadata.",
     inputSchema: FacebookVideoTranscribeInputSchema,
-    outputSchema: FacebookVideoTranscribeOutputSchema,
+    outputSchema: recordOutputSchema("facebook_video_transcribe", FacebookVideoTranscribeOutputSchema),
     annotations: liveWebToolAnnotations("Facebook Organic Video Transcription")
   }, async (input) => formatFacebookVideoTranscribe(await executor.facebookVideoTranscribe(input), input));
   server.registerTool("instagram_profile_content", {
     title: "Instagram Profile Content Discovery",
-    description: withReportNote("Discover Instagram profile grid content links for a handle or profile URL. Use this when the user wants a person or brand account content inventory before selecting posts/reels to download. Returns profile stats, collected post/reel/tv URLs, shortcodes, type counts, browser details, pagination attempts, stop reason, and limitations."),
+    description: "Discover Instagram profile grid content links (posts/reels/tv) for a handle or profile URL, for later selection with instagram_media_download. Returns profile stats and collected URLs.",
     inputSchema: InstagramProfileContentInputSchema,
-    outputSchema: InstagramProfileContentOutputSchema,
+    outputSchema: recordOutputSchema("instagram_profile_content", InstagramProfileContentOutputSchema),
     annotations: liveWebToolAnnotations("Instagram Profile Content Discovery")
   }, async (input) => formatInstagramProfileContent(await executor.instagramProfileContent(input), input));
   server.registerTool("instagram_media_download", {
     title: "Instagram Post/Reel Media Download",
-    description: withReportNote("Extract and download media from one Instagram post, reel, or tv URL. Use after instagram_profile_content or when the user gives a specific Instagram URL and wants the image, caption/text, reel audio/video tracks, optional muxed MP4, or optional transcript. Reels commonly expose separate video-only and audio-only MP4 tracks; this tool selects the best video and audio tracks and attempts muxing when ffmpeg is available."),
+    description: "Extract and download media from one Instagram post, reel, or tv URL \u2014 image, caption, video/audio tracks, optional muxed MP4, or transcript. Selects the best video/audio track pair and muxes when ffmpeg is available.",
     inputSchema: InstagramMediaDownloadInputSchema,
-    outputSchema: InstagramMediaDownloadOutputSchema,
+    outputSchema: recordOutputSchema("instagram_media_download", InstagramMediaDownloadOutputSchema),
     annotations: liveWebToolAnnotations("Instagram Post/Reel Media Download")
   }, async (input) => formatInstagramMediaDownload(await executor.instagramMediaDownload(input), input));
   server.registerTool("maps_place_intel", {
     title: "Google Maps Business Profile Details",
-    description: withReportNote('Extract Google Maps business intelligence for one known/named business: rating, review count, category, address, phone, website, hours, booking URL, review histogram, review topics, about attributes, entity IDs, and optional review cards. Do not use this for category searches, local market prospect lists, or requests for multiple GMB/GBP profiles; use maps_search first for those. Split business name from location (e.g. "Elite Roofing Denver CO" => businessName "Elite Roofing", location "Denver, CO"). Pass includeReviews true when the user asks for reviews/customer pain.'),
+    description: "Extract Google Maps business intelligence for one known/named business: rating, reviews, category, address, phone, hours, entity IDs. Not for category searches or multi-business prospect lists \u2014 use maps_search for those. Split business name from location.",
     inputSchema: MapsPlaceIntelInputSchema,
-    outputSchema: MapsPlaceIntelOutputSchema,
+    outputSchema: recordOutputSchema("maps_place_intel", MapsPlaceIntelOutputSchema),
     annotations: liveWebToolAnnotations("Google Maps Business Profile Details")
   }, async (input) => formatMapsPlaceIntel(await executor.mapsPlaceIntel(input), input));
   server.registerTool("maps_search", {
     title: "Google Maps Business Search",
-    description: withReportNote('Search Google Maps for multiple businesses/profiles by category, niche, keyword, or local market. Use this when the user asks for several Google Business Profiles, GMBs, GBPs, leads, prospects, competitors, or "more than the 3-pack." Routing is automatic and location-aware \u2014 leave proxyMode unset; pass proxyZip only to force a specific ZIP. Returns up to 50 candidates with names, place URLs, CIDs when available, ratings, review counts, profile metadata, and sanitized attempt telemetry. Default maxResults is 10; maximum is 50. Use maps_place_intel afterward only when a selected business needs full details and reviews.'),
+    description: "Search Google Maps for multiple businesses by category, niche, or local market \u2014 leads, prospects, competitors, or beyond the 3-pack. Leave proxyMode unset. Returns up to 50 candidates (default 10) with names, place URLs, CIDs, ratings.",
     inputSchema: MapsSearchInputSchema,
-    outputSchema: MapsSearchOutputSchema,
+    outputSchema: recordOutputSchema("maps_search", MapsSearchOutputSchema),
     annotations: liveWebToolAnnotations("Google Maps Business Search")
   }, async (input) => formatMapsSearch(await executor.mapsSearch(input), input));
   server.registerTool("directory_workflow", {
     title: "Directory Workflow: Markets + Maps",
-    description: withReportNote('Build directory/prospecting datasets by selecting US city markets from the free Census Population Estimates city/place dataset, optionally joining configured US ZIPS/Lead Magician ZIP groups, then running Google Maps business searches for each city in parallel. Use this when the user wants "all cities over 100k population in a state", "build a directory CSV", "find markets then get Maps data", or similar location-database + Maps workflows. Set minPopulation, state, query, maxResultsPerCity, and concurrency. Routing is automatic per city. Saved CSV rows include source_location, result_position, business_name, review_stars, review_count, category, address, phone, hours_status, website_url, directions_url, place_url, cid, cid_decimal, city population, and ZIP groups. Structured city results include sanitized attempt telemetry. Use maps_place_intel only when a selected profile needs deeper review topics, profile review count confirmation, or review cards. For local Lead Magician ZIP enrichment, set MCP_SCRAPER_USZIPS_CSV_PATH on the API server or pass usZipsCsvPath only in local/test mode.'),
+    description: `Build directory/prospecting datasets: selects US city markets from Census population data, optionally joins configured ZIP groups, then runs Google Maps business searches per city in parallel. Use for "all cities over 100k population in a state" or market+Maps workflows. ${fileBehavior("Saves a CSV of results per city.", "Large results are stored as a retrievable artifact \u2014 you get an inline summary plus an artifactId for report_artifact_read.")}`,
     inputSchema: DirectoryWorkflowInputSchema,
-    outputSchema: DirectoryWorkflowOutputSchema,
+    outputSchema: recordOutputSchema("directory_workflow", DirectoryWorkflowOutputSchema),
     annotations: liveWebToolAnnotations("Directory Workflow: Markets + Maps")
-  }, async (input) => formatDirectoryWorkflow(await executor.directoryWorkflow(input), input));
+  }, async (input) => formatDirectoryWorkflow(await executor.directoryWorkflow(input), input, ctx));
   server.registerTool("workflow_list", {
     title: "Workflow Catalog",
-    description: "List MCP Scraper higher-level workflows and AI-facing recipes. Use this when the user asks what MCP Scraper can do beyond single tools, or asks for market analysis, ICP research, forum/review acquisition, full brand design briefings, CRO audits, competitive positioning, content gap briefs, or AI search visibility audits. Returns runnable workflow ids plus recipe guidance for the best tool chain.",
+    description: "List MCP Scraper higher-level workflows and recipes \u2014 market analysis, ICP research, CRO audits, competitive positioning, content gap briefs, AI search visibility, and more. Returns runnable workflow ids plus tool-chain guidance.",
     inputSchema: WorkflowListInputSchema,
-    outputSchema: WorkflowListOutputSchema,
+    outputSchema: recordOutputSchema("workflow_list", WorkflowListOutputSchema),
     annotations: localPlanningToolAnnotations("Workflow Catalog")
   }, async (input) => formatWorkflowList(await executor.workflowList(input), input));
   server.registerTool("workflow_suggest", {
     title: "Workflow Intent Router",
-    description: 'Route a high-level business or research goal to the right MCP Scraper workflow/tool chain. Use before running work when the user says things like "do market analysis", "research my ICP", "acquire forum and review language", "build a brand design brief", "run a CRO audit", "compare competitors", "make a content gap brief", or "audit AI search visibility". This tool does not spend credits; it tells the AI exactly which workflow/tool chain to run next.',
+    description: "Route a high-level business/research goal (market analysis, ICP research, CRO audit, competitor comparison, content gap brief, AI search visibility, etc) to the right MCP Scraper workflow/tool chain. Free; tells you what to run next.",
     inputSchema: WorkflowSuggestInputSchema,
-    outputSchema: WorkflowSuggestOutputSchema,
+    outputSchema: recordOutputSchema("workflow_suggest", WorkflowSuggestOutputSchema),
     annotations: localPlanningToolAnnotations("Workflow Intent Router")
   }, async (input) => formatWorkflowSuggest(input));
   server.registerTool("workflow_run", {
     title: "Run Workflow",
-    description: withReportNote("Start a higher-level MCP Scraper workflow. Use after workflow_suggest or workflow_list. Runnable workflow ids: directory, agent-packet, local-competitive-audit, map-comparison, serp-comparison, paa-expansion-brief, ai-overview-language. This is the main MCP tool for market analysis, ICP evidence packets, local competitive audits, Maps/SERP comparisons, content gap briefs, and AI Overview language guidance. Stepwise workflows (e.g. agent-packet) run ONE leg per call and return runId, the step output, and nextStep \u2014 when nextStep is present, call workflow_step with the runId to run the next leg, and keep calling it until done is true. This keeps each call short instead of one long blocking run, so report each step result to the user as it arrives."),
+    description: "Start a higher-level MCP Scraper workflow (directory, agent-packet, local-competitive-audit, map-comparison, serp-comparison, paa-expansion-brief, ai-overview-language). Use after workflow_suggest or workflow_list. Stepwise workflows return runId + nextStep \u2014 call workflow_step with the runId until done is true.",
     inputSchema: WorkflowRunInputSchema,
-    outputSchema: WorkflowRunOutputSchema,
+    outputSchema: recordOutputSchema("workflow_run", WorkflowRunOutputSchema),
     annotations: liveWebToolAnnotations("Run Workflow")
   }, async (input) => formatWorkflowRun(await executor.workflowRun(input), input));
   server.registerTool("workflow_step", {
     title: "Advance Workflow Step",
-    description: withReportNote("Run the next leg of a stepwise MCP Scraper workflow started with workflow_run. Pass the runId. Each call executes exactly one logical step (typically one live harvest), persists that step's artifacts, and returns the step output plus nextStep. Keep calling workflow_step with the same runId until done is true, reporting each step result to the user as it lands. Use this instead of waiting on one long workflow call \u2014 it avoids client timeouts on long multi-step jobs."),
+    description: "Run the next leg of a stepwise workflow started with workflow_run. Pass the runId; each call executes one step and returns its output plus nextStep. Keep calling with the same runId until done is true.",
     inputSchema: WorkflowStepInputSchema,
-    outputSchema: WorkflowStepOutputSchema,
+    outputSchema: recordOutputSchema("workflow_step", WorkflowStepOutputSchema),
     annotations: liveWebToolAnnotations("Advance Workflow Step")
   }, async (input) => formatWorkflowStep(await executor.workflowStep(input), input));
   server.registerTool("workflow_status", {
     title: "Workflow Status",
-    description: "Fetch a hosted workflow run by id and list its current status and artifacts. Use when a workflow may still be running, when the model needs to re-open a run, inspect artifact ids, recover from a long workflow, or decide whether to call workflow_step or workflow_artifact_read next. Use only a runId returned by workflow_run/workflow_step/workflow_status; do not construct one yourself.",
+    description: "Fetch a hosted workflow run by id and list its current status and artifacts, to re-open a run or recover artifact ids. Use only a runId returned by workflow_run/workflow_step/workflow_status.",
     inputSchema: WorkflowStatusInputSchema,
-    outputSchema: WorkflowStatusOutputSchema,
+    outputSchema: recordOutputSchema("workflow_status", WorkflowStatusOutputSchema),
     annotations: liveWebToolAnnotations("Workflow Status")
   }, async (input) => formatWorkflowStatus(await executor.workflowStatus(input), input));
   server.registerTool("workflow_artifact_read", {
     title: "Read Workflow Artifact",
-    description: "Read a workflow artifact back into MCP context by run id and artifact id. Use this before writing final deliverables so the answer is grounded in generated evidence.json, CSVs, Markdown briefs, reports, and task files instead of memory. Use workflow_status first when artifact ids are unknown. Use only artifactId values returned by workflow_run/workflow_step/workflow_status; do not construct one yourself. Use maxBytes to limit large CSV/JSON artifacts.",
+    description: "Read a workflow artifact back into context by run id and artifact id, so final deliverables are grounded in generated evidence rather than memory. Use workflow_status first when artifact ids are unknown. Use maxBytes to limit large artifacts.",
     inputSchema: WorkflowArtifactReadInputSchema,
-    outputSchema: WorkflowArtifactReadOutputSchema,
+    outputSchema: recordOutputSchema("workflow_artifact_read", WorkflowArtifactReadOutputSchema),
     annotations: liveWebToolAnnotations("Read Workflow Artifact")
   }, async (input) => formatWorkflowArtifactRead(await executor.workflowArtifactRead(input), input));
+  server.registerTool("report_artifact_read", {
+    title: "Read Report Artifact",
+    description: "Read back a stored report artifact by artifactId (returned by any tool whose result was too large to inline). Windowed: pass offset/maxBytes and keep reading until nextOffset is null.",
+    inputSchema: ReportArtifactReadInputSchema,
+    outputSchema: recordOutputSchema("report_artifact_read", ReportArtifactReadOutputSchema),
+    annotations: liveWebToolAnnotations("Read Report Artifact")
+  }, async (input) => {
+    const owner = artifactOwnerId(input.artifactId);
+    if (!owner || owner !== ownerId) {
+      return { content: [{ type: "text", text: "Artifact not found or not owned by this caller." }], isError: true };
+    }
+    const window = await readArtifactWindow(input.artifactId, input.offset ?? 0, input.maxBytes ?? 2e4);
+    if (!window) {
+      return { content: [{ type: "text", text: "Artifact not found or expired." }], isError: true };
+    }
+    return {
+      content: [{ type: "text", text: window.text }],
+      structuredContent: { artifactId: input.artifactId, text: window.text, totalBytes: window.totalBytes, nextOffset: window.nextOffset }
+    };
+  });
   server.registerTool("rank_tracker_workflow", {
     title: "Rank Tracker Blueprint Builder",
-    description: "Generate a build-ready database, cron/heartbeat, ingestion, metrics, and AI implementation prompt for a rank tracker powered by MCP Scraper. Supports Maps rankings through directory_workflow/maps_search, organic rankings through search_serp, AI Overview citation tracking, and People Also Ask source presence tracking. This tool is local planning only; it does not call the web or spend credits.",
+    description: "Generate a build-ready database schema, cron plan, and implementation prompt for a rank tracker powered by MCP Scraper (Maps, organic, AI Overview, or PAA tracking). Local planning only \u2014 does not call the web or spend credits.",
     inputSchema: RankTrackerBlueprintInputSchema,
-    outputSchema: RankTrackerBlueprintOutputSchema,
+    outputSchema: recordOutputSchema("rank_tracker_workflow", RankTrackerBlueprintOutputSchema),
     annotations: localPlanningToolAnnotations("Rank Tracker Blueprint Builder")
   }, async (input) => buildRankTrackerBlueprint(input));
   server.registerTool("credits_info", {
     title: "MCP Scraper Credits & Costs",
-    description: "Answer questions about MCP Scraper credits, usage limits, and concurrency upgrades: current credit balance, what a specific tool/action costs, the full cost table, current concurrency limit, the extra-slot price, billing URL, and the terminal checkout command. Does not expose payment methods or credit card information.",
+    description: "Answer questions about MCP Scraper credits, usage limits, and concurrency upgrades \u2014 balance, tool costs, concurrency limits, billing URL. Does not expose payment methods or card information.",
     inputSchema: CreditsInfoInputSchema,
-    outputSchema: CreditsInfoOutputSchema,
+    outputSchema: recordOutputSchema("credits_info", CreditsInfoOutputSchema),
     annotations: {
       title: "MCP Scraper Credits & Costs",
       readOnlyHint: true,
@@ -4605,30 +4876,30 @@ var HttpMcpToolExecutor = class {
 
 // src/mcp/browser-agent-mcp-server.ts
 import { McpServer as McpServer2 } from "@modelcontextprotocol/sdk/server/mcp.js";
-import { mkdirSync as mkdirSync3, writeFileSync as writeFileSync3 } from "fs";
-import { homedir as homedir3 } from "os";
-import { join as join5 } from "path";
+import { mkdirSync as mkdirSync4, writeFileSync as writeFileSync4 } from "fs";
+import { homedir as homedir4 } from "os";
+import { join as join6 } from "path";
 
 // src/services/fanout/export.ts
-import { mkdirSync as mkdirSync2, writeFileSync as writeFileSync2 } from "fs";
-import { homedir as homedir2 } from "os";
-import { join as join3 } from "path";
+import { mkdirSync as mkdirSync3, writeFileSync as writeFileSync3 } from "fs";
+import { homedir as homedir3 } from "os";
+import { join as join4 } from "path";
 import Papa from "papaparse";
 function outputBaseDir2() {
-  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join3(homedir2(), "Downloads", "mcp-scraper");
+  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join4(homedir3(), "Downloads", "mcp-scraper");
 }
 function safe(value) {
   return value.replace(/[^a-zA-Z0-9._-]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 80) || "capture";
 }
 function writeTable(path, rows, delimiter) {
-  writeFileSync2(path, Papa.unparse(rows.length ? rows : [{}], { delimiter }));
+  writeFileSync3(path, Papa.unparse(rows.length ? rows : [{}], { delimiter }));
 }
 function exportFanout(enriched) {
   const stamp = safe(enriched.capturedAt.replace(/[:.]/g, "-"));
   const outputDir = outputBaseDir2();
-  const relativeDir = join3("fanout", `${stamp}-${safe(enriched.platform)}`);
-  const dir = join3(outputDir, relativeDir);
-  mkdirSync2(dir, { recursive: true });
+  const relativeDir = join4("fanout", `${stamp}-${safe(enriched.platform)}`);
+  const dir = join4(outputDir, relativeDir);
+  mkdirSync3(dir, { recursive: true });
   const queryRows = enriched.queries.map((q, i) => ({ index: i + 1, query: q }));
   const citationRows = enriched.aggregates.citationOrder.map((c) => {
     const src = enriched.citedUrls.find((s) => s.url === c.url);
@@ -4641,25 +4912,25 @@ function exportFanout(enriched) {
   const relativePaths = {
     relativeTo: "MCP_SCRAPER_OUTPUT_DIR or ~/Downloads/mcp-scraper",
     dir: relativeDir,
-    json: join3(relativeDir, "fanout.json"),
-    queriesCsv: join3(relativeDir, "queries.csv"),
-    queriesTsv: join3(relativeDir, "queries.tsv"),
-    citationsCsv: join3(relativeDir, "citations.csv"),
-    sourcesCsv: join3(relativeDir, "sources.csv"),
-    browsedOnlyCsv: join3(relativeDir, "browsed-only.csv"),
-    snippetsCsv: join3(relativeDir, "snippets.csv"),
-    domainsCsv: join3(relativeDir, "domains.csv"),
-    report: join3(relativeDir, "report.html")
+    json: join4(relativeDir, "fanout.json"),
+    queriesCsv: join4(relativeDir, "queries.csv"),
+    queriesTsv: join4(relativeDir, "queries.tsv"),
+    citationsCsv: join4(relativeDir, "citations.csv"),
+    sourcesCsv: join4(relativeDir, "sources.csv"),
+    browsedOnlyCsv: join4(relativeDir, "browsed-only.csv"),
+    snippetsCsv: join4(relativeDir, "snippets.csv"),
+    domainsCsv: join4(relativeDir, "domains.csv"),
+    report: join4(relativeDir, "report.html")
   };
-  writeFileSync2(join3(outputDir, relativePaths.json), JSON.stringify(enriched, null, 2));
-  writeTable(join3(outputDir, relativePaths.queriesCsv), queryRows, ",");
-  writeTable(join3(outputDir, relativePaths.queriesTsv), queryRows, "	");
-  writeTable(join3(outputDir, relativePaths.citationsCsv), citationRows, ",");
-  writeTable(join3(outputDir, relativePaths.sourcesCsv), sourceRows, ",");
-  writeTable(join3(outputDir, relativePaths.browsedOnlyCsv), browsedOnlyRows, ",");
-  writeTable(join3(outputDir, relativePaths.snippetsCsv), snippetRows, ",");
-  writeTable(join3(outputDir, relativePaths.domainsCsv), domainRows, ",");
-  writeFileSync2(join3(outputDir, relativePaths.report), renderReportHtml(enriched));
+  writeFileSync3(join4(outputDir, relativePaths.json), JSON.stringify(enriched, null, 2));
+  writeTable(join4(outputDir, relativePaths.queriesCsv), queryRows, ",");
+  writeTable(join4(outputDir, relativePaths.queriesTsv), queryRows, "	");
+  writeTable(join4(outputDir, relativePaths.citationsCsv), citationRows, ",");
+  writeTable(join4(outputDir, relativePaths.sourcesCsv), sourceRows, ",");
+  writeTable(join4(outputDir, relativePaths.browsedOnlyCsv), browsedOnlyRows, ",");
+  writeTable(join4(outputDir, relativePaths.snippetsCsv), snippetRows, ",");
+  writeTable(join4(outputDir, relativePaths.domainsCsv), domainRows, ",");
+  writeFileSync3(join4(outputDir, relativePaths.report), renderReportHtml(enriched));
   return relativePaths;
 }
 function esc(s) {
@@ -4756,18 +5027,18 @@ var BrowserOpenInputSchema = {
   label: z2.string().optional().describe("Optional human label for this session, shown in the watch console."),
   url: z2.string().url().optional().describe("Optional URL to navigate to immediately after opening."),
   profile: z2.string().optional().describe("Optional saved hosted profile name to load a logged-in session for a site."),
-  save_profile_changes: z2.boolean().optional().describe("Persist cookies and browser storage back to the named profile when the session is closed. Use this for profile setup or intentional auth refreshes; avoid parallel sessions writing to the same profile."),
-  timeout_seconds: z2.number().int().min(60).max(259200).optional().describe("How long the direct no-proxy browser session may live before auto-termination. Defaults to 600. The browser idles into a zero-cost standby between actions, so a longer timeout is cheap.")
+  save_profile_changes: z2.boolean().optional().describe("Persist cookies/storage back to the named profile on close. Avoid parallel sessions writing to the same profile."),
+  timeout_seconds: z2.number().int().min(60).max(259200).optional().describe("Session lifetime before auto-termination. Defaults to 600.")
 };
 var BrowserProfileConnectInputSchema = {
-  email: z2.string().optional().describe("Account email for the login being saved. Used to derive a stable, reusable profile name AND recorded as a note of who is connected. Does NOT import existing cookies \u2014 the user signs in fresh via the returned watch_url."),
-  profile: z2.string().optional().describe("Profile to add this login to. Omit to derive one from email. A single profile holds MANY site logins \u2014 pass the same profile name again with a different domain to stack another account (e.g. add claude.ai to a profile that already has chatgpt.com)."),
-  domain: z2.string().optional().describe("Site to log into, e.g. chatgpt.com, claude.ai, reddit.com. Defaults to chatgpt.com. Each domain is a separate login inside the profile."),
-  login_url: z2.string().url().optional().describe("Login page for the domain. Defaults to https://<domain>/ (https://chatgpt.com/ when domain is omitted)."),
-  url: z2.string().url().optional().describe("Deprecated alias for login_url. Use login_url."),
-  note: z2.string().optional().describe('Free-text note describing this login, e.g. "ChatGPT Plus, work account". Surfaced by browser_profile_list so you can see what the profile holds.'),
+  email: z2.string().optional().describe("Account email for the login. Derives a stable profile name and is recorded as a note. Does NOT import existing cookies \u2014 the user signs in fresh."),
+  profile: z2.string().optional().describe("Profile to add this login to. Omit to derive from email. A single profile holds MANY logins \u2014 pass the same name with a different domain to stack accounts."),
+  domain: z2.string().optional().describe("Site to log into, e.g. chatgpt.com, claude.ai, reddit.com. Defaults to chatgpt.com."),
+  login_url: z2.string().url().optional().describe("Login page for the domain. Defaults to https://<domain>/."),
+  url: z2.string().url().optional().describe("Deprecated alias for login_url."),
+  note: z2.string().optional().describe("Free-text note describing this login. Surfaced by browser_profile_list."),
   label: z2.string().optional().describe("Optional human label for this sign-in setup session."),
-  timeout_seconds: z2.number().int().min(60).max(259200).optional().describe("How long the sign-in setup session may live before auto-termination. Defaults to 600.")
+  timeout_seconds: z2.number().int().min(60).max(259200).optional().describe("Sign-in session lifetime before auto-termination. Defaults to 600.")
 };
 var BrowserProfileListInputSchema = {
   profile: z2.string().optional().describe("Profile whose saved logins to list. Omit to derive from email."),
@@ -4776,7 +5047,7 @@ var BrowserProfileListInputSchema = {
   connection_id: z2.string().optional().describe("A specific login connection id returned by browser_profile_connect, to poll just that one.")
 };
 var BrowserSessionInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself.")
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions.")
 };
 var BrowserLocateTargetSchema = z2.object({
   name: z2.string().optional().describe("Optional label for this target, echoed in the result."),
@@ -4788,94 +5059,94 @@ var BrowserLocateTargetSchema = z2.object({
   message: "target requires selector or text"
 });
 var BrowserLocateInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   targets: z2.array(BrowserLocateTargetSchema).min(1).max(20).describe("DOM targets to locate in the current viewport. Use selectors for exact elements, or text for visible text ranges.")
 };
 var BrowserGotoInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   url: z2.string().url().describe("URL to navigate the browser to.")
 };
 var BrowserClickInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   x: z2.number().describe("X coordinate to click, in screenshot pixels. Use only coordinates from the latest browser_screenshot, browser_read, or browser_locate result; do not guess."),
   y: z2.number().describe("Y coordinate to click, in screenshot pixels. Use only coordinates from the latest browser_screenshot, browser_read, or browser_locate result; do not guess."),
   button: z2.enum(["left", "right", "middle"]).default("left").describe("Mouse button."),
   num_clicks: z2.number().int().min(1).max(3).optional().describe("Number of clicks, e.g. 2 for double-click.")
 };
 var BrowserTypeInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   text: z2.string().describe("Text to type at the current focus. Click a field first to focus it."),
   delay: z2.number().int().min(0).max(500).optional().describe("Optional per-keystroke delay in ms for human-like typing.")
 };
 var BrowserScrollInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   delta_y: z2.number().default(5).describe("Vertical scroll in wheel units. Positive scrolls down, negative up."),
   delta_x: z2.number().default(0).describe("Horizontal scroll in wheel units."),
   x: z2.number().optional().describe("X position to scroll at. Defaults to screen center."),
   y: z2.number().optional().describe("Y position to scroll at. Defaults to screen center.")
 };
 var BrowserPressInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
   keys: z2.array(z2.string()).min(1).describe('Keys or combinations to press, e.g. ["Return"], ["Ctrl+a"], ["Ctrl+Shift+Tab"].')
 };
 var BrowserReplayStopInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
-  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays. Use only a returned replay_id; do not construct one yourself.")
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
+  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays.")
 };
 var BrowserReplayDownloadInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
-  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays. Use only a returned replay_id; do not construct one yourself."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
+  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays."),
   filename: z2.string().optional().describe("Optional local MP4 filename. Defaults to a timestamped replay filename.")
 };
 var BrowserReplayAnnotationSchema = z2.object({
-  type: z2.enum(["box", "circle", "underline", "arrow", "label"]).default("box").describe("Annotation style to draw over the replay."),
-  start_seconds: z2.number().min(0).default(0).describe("When the annotation should appear in the replay."),
-  end_seconds: z2.number().min(0).optional().describe("When the annotation should disappear. Defaults to two seconds after start_seconds."),
-  left: z2.number().optional().describe("Target left edge in browser screenshot pixels. Use element.left from browser_screenshot/browser_read."),
-  top: z2.number().optional().describe("Target top edge in browser screenshot pixels. Use element.top from browser_screenshot/browser_read."),
-  width: z2.number().positive().optional().describe("Target width in browser screenshot pixels. Use element.width from browser_screenshot/browser_read."),
-  height: z2.number().positive().optional().describe("Target height in browser screenshot pixels. Use element.height from browser_screenshot/browser_read."),
-  x: z2.number().optional().describe("Point target x coordinate in browser screenshot pixels when no box is available."),
-  y: z2.number().optional().describe("Point target y coordinate in browser screenshot pixels when no box is available."),
-  from_x: z2.number().optional().describe("Arrow start x coordinate in browser screenshot pixels. Defaults near the target."),
-  from_y: z2.number().optional().describe("Arrow start y coordinate in browser screenshot pixels. Defaults near the target."),
-  to_x: z2.number().optional().describe("Arrow target x coordinate in browser screenshot pixels. Defaults to the target box center."),
-  to_y: z2.number().optional().describe("Arrow target y coordinate in browser screenshot pixels. Defaults to the target box center."),
-  label: z2.string().max(120).optional().describe("Optional text callout to render near the annotation."),
-  color: z2.string().regex(/^#?[0-9a-fA-F]{6}$/).optional().describe("Annotation color as hex, for example #ff3b30."),
+  type: z2.enum(["box", "circle", "underline", "arrow", "label"]).default("box").describe("Annotation style."),
+  start_seconds: z2.number().min(0).default(0).describe("When the annotation should appear."),
+  end_seconds: z2.number().min(0).optional().describe("When it disappears. Defaults to 2s after start_seconds."),
+  left: z2.number().optional().describe("Target left edge in screenshot pixels (element.left)."),
+  top: z2.number().optional().describe("Target top edge in screenshot pixels (element.top)."),
+  width: z2.number().positive().optional().describe("Target width in screenshot pixels (element.width)."),
+  height: z2.number().positive().optional().describe("Target height in screenshot pixels (element.height)."),
+  x: z2.number().optional().describe("Point target x coordinate when no box is available."),
+  y: z2.number().optional().describe("Point target y coordinate when no box is available."),
+  from_x: z2.number().optional().describe("Arrow start x coordinate. Defaults near the target."),
+  from_y: z2.number().optional().describe("Arrow start y coordinate. Defaults near the target."),
+  to_x: z2.number().optional().describe("Arrow end x coordinate. Defaults to the target box center."),
+  to_y: z2.number().optional().describe("Arrow end y coordinate. Defaults to the target box center."),
+  label: z2.string().max(120).optional().describe("Optional text callout."),
+  color: z2.string().regex(/^#?[0-9a-fA-F]{6}$/).optional().describe("Annotation color as hex, e.g. #ff3b30."),
   thickness: z2.number().min(1).max(24).optional().describe("Stroke thickness in pixels. Defaults to 5.")
 });
 var BrowserReplayMarkInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself. A replay must already be recording."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. A replay must already be recording."),
   target: BrowserLocateTargetSchema.describe("The exact DOM element or text range to mark in the current viewport."),
-  type: z2.enum(["box", "circle", "underline", "arrow"]).default("box").describe("Annotation style to generate. Labels are included on the returned annotation when label is provided."),
+  type: z2.enum(["box", "circle", "underline", "arrow"]).default("box").describe("Annotation style to generate."),
   label: z2.string().max(120).optional().describe("Optional callout text to render near the target."),
-  color: z2.string().regex(/^#?[0-9a-fA-F]{6}$/).optional().describe("Annotation color as hex, for example #ff3b30."),
+  color: z2.string().regex(/^#?[0-9a-fA-F]{6}$/).optional().describe("Annotation color as hex, e.g. #ff3b30."),
   thickness: z2.number().min(1).max(24).optional().describe("Stroke thickness in pixels. Defaults to 5."),
   padding: z2.number().min(0).max(80).default(8).describe("Pixels to expand the DOM bounds so the highlight does not touch the text edge."),
-  start_offset_seconds: z2.number().min(-5).max(10).default(-0.25).describe("Offset from the current replay time. Negative values make the callout appear just before the mark action is captured."),
-  duration_seconds: z2.number().min(0.5).max(30).default(4).describe("How long the generated annotation should remain visible.")
+  start_offset_seconds: z2.number().min(-5).max(10).default(-0.25).describe("Offset from the current replay time; negative appears just before the mark action."),
+  duration_seconds: z2.number().min(0.5).max(30).default(4).describe("How long the annotation should remain visible.")
 };
 var BrowserReplayAnnotateInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. Use only a returned session_id; do not construct one yourself."),
-  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays. Use only a returned replay_id; do not construct one yourself."),
-  annotations: z2.array(BrowserReplayAnnotationSchema).min(1).max(50).describe("Timed overlay annotations to render on the replay. Prefer annotations returned by browser_replay_mark; otherwise use exact DOM bounds from browser_locate/browser_screenshot/browser_read."),
-  filename: z2.string().optional().describe("Optional output MP4 filename. Defaults to a timestamped annotated replay filename."),
+  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions."),
+  replay_id: z2.string().describe("The replay id returned by browser_replay_start or browser_list_replays."),
+  annotations: z2.array(BrowserReplayAnnotationSchema).min(1).max(50).describe("Timed overlay annotations. Prefer ones from browser_replay_mark; otherwise use exact DOM bounds from browser_locate."),
+  filename: z2.string().optional().describe("Optional output MP4 filename. Defaults to a timestamped filename."),
   source_width: z2.number().positive().optional().describe("Width of the screenshot coordinate space used for annotations. Defaults to the replay video width."),
-  source_height: z2.number().positive().optional().describe("Height of the coordinate space used for annotations. When this is a page viewport height smaller than the replay video height, the annotator infers the browser chrome top offset."),
-  source_left_offset: z2.number().min(0).optional().describe("Optional explicit X offset from annotation coordinates to replay video coordinates. Usually omitted."),
-  source_top_offset: z2.number().min(0).optional().describe("Optional explicit Y offset from annotation coordinates to replay video coordinates. Usually omitted because browser chrome offset is inferred when source_height is a page viewport height.")
+  source_height: z2.number().positive().optional().describe("Height of the annotation coordinate space; if smaller than the replay video height, the browser chrome offset is inferred."),
+  source_left_offset: z2.number().min(0).optional().describe("Explicit X offset from annotation to replay video coordinates. Usually omitted."),
+  source_top_offset: z2.number().min(0).optional().describe("Explicit Y offset from annotation to replay video coordinates. Usually omitted.")
 };
 var BrowserListInputSchema = {
   include_closed: z2.boolean().default(false).describe("Include closed sessions in the list.")
 };
 var BrowserCaptureFanoutInputSchema = {
-  session_id: z2.string().describe("The session id returned by browser_open or browser_list_sessions. The session must be on chatgpt.com or claude.ai and logged in through a saved hosted profile. Use only a returned session_id; do not construct one yourself."),
-  prompt: z2.string().optional().describe("Optional prompt to type into the chat composer and submit before capturing. Omit to passively capture the fan-out of a prompt the user just ran in the visible browser. The prompt must trigger web search for there to be a fan-out."),
-  wait_ms: z2.number().int().min(0).max(18e4).optional().describe("How long to wait for the answer stream to finish before parsing. Defaults to 90000 when a prompt is sent, 8000 for passive capture. Capture ends as soon as the answer stream finishes."),
-  first_party_domain: z2.string().optional().describe("The brand/site being researched, for example example.com. Sources on this domain are tagged First-party/vendor. Set this when the prompt is about a specific business."),
-  reset: z2.boolean().default(false).describe("Clear any previously buffered stream for this page before capturing. Use when re-running to avoid mixing turns."),
-  export: z2.boolean().default(false).describe("When true, write JSON, CSV, TSV exports (queries, citations, all sources, browsed-only sources, snippets, domains) and a self-contained HTML report to MCP_SCRAPER_OUTPUT_DIR/fanout, returning paths relative to MCP_SCRAPER_OUTPUT_DIR.")
+  session_id: z2.string().describe("Session id from browser_open. Must be on chatgpt.com or claude.ai, logged in via a saved hosted profile."),
+  prompt: z2.string().optional().describe("Optional prompt to type and submit before capturing. Omit to passively capture a prompt the user just ran. Must trigger web search to produce a fan-out."),
+  wait_ms: z2.number().int().min(0).max(18e4).optional().describe("How long to wait for the answer stream to finish. Defaults to 90000 when a prompt is sent, 8000 for passive capture."),
+  first_party_domain: z2.string().optional().describe("The brand/site being researched, e.g. example.com \u2014 sources on this domain are tagged First-party/vendor."),
+  reset: z2.boolean().default(false).describe("Clear any previously buffered stream for this page before capturing."),
+  export: z2.boolean().default(false).describe("Write JSON/CSV/TSV/HTML exports to MCP_SCRAPER_OUTPUT_DIR/fanout, returning relative paths.")
 };
 var FanoutSourceOutput = z2.object({
   url: z2.string(),
@@ -4899,10 +5170,10 @@ var BrowserCaptureFanoutOutputSchema = {
     title: z2.string(),
     rounds: z2.number().int().min(0)
   }),
-  queries: z2.array(z2.string()).describe("Every web-search sub-query the model issued, in capture order. Classify these by funnel stage and type yourself from the text."),
-  browsed_urls: z2.array(FanoutSourceOutput).describe("Every researched URL (cited and browsed-only), cited first."),
-  cited_urls: z2.array(FanoutSourceOutput).describe("The subset of researched URLs actually cited in the final answer."),
-  browsed_only: z2.array(FanoutSourceOutput).describe("Researched URLs the model pulled but did not cite."),
+  queries: z2.array(z2.string()).describe("Every web-search sub-query issued, in capture order."),
+  browsed_urls: z2.array(FanoutSourceOutput).describe("Every researched URL, cited first."),
+  cited_urls: z2.array(FanoutSourceOutput).describe("Researched URLs cited in the final answer."),
+  browsed_only: z2.array(FanoutSourceOutput).describe("Researched URLs pulled but not cited."),
   snippets: z2.array(z2.object({ url: z2.string(), domain: z2.string(), title: z2.string(), text: z2.string() })),
   counts: z2.object({
     subQueries: z2.number().int().min(0),
@@ -5086,7 +5357,7 @@ var BrowserListSessionsOutputSchema = {
 import { execFile } from "child_process";
 import { mkdtemp, rm, stat, writeFile } from "fs/promises";
 import { tmpdir } from "os";
-import { join as join4 } from "path";
+import { join as join5 } from "path";
 import { promisify } from "util";
 var execFileAsync = promisify(execFile);
 function finiteNumber(value) {
@@ -5309,8 +5580,8 @@ function ffmpegFilterPath(path) {
 async function annotateReplayVideo(inputFilePath, outputFilePath, options) {
   if (!options.annotations.length) throw new Error("annotations must include at least one item");
   const size = await videoSize(inputFilePath);
-  const tmp = await mkdtemp(join4(tmpdir(), "mcp-scraper-ass-"));
-  const assPath = join4(tmp, "annotations.ass");
+  const tmp = await mkdtemp(join5(tmpdir(), "mcp-scraper-ass-"));
+  const assPath = join5(tmp, "annotations.ass");
   try {
     await writeFile(assPath, buildAssSubtitle(options, size), "utf8");
     await execFileAsync("ffmpeg", [
@@ -5380,7 +5651,7 @@ function actionResult(tool, sessionId, ok, data, nextRecommendedTool = "browser_
   });
 }
 function outputBaseDir3() {
-  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join5(homedir3(), "Downloads", "mcp-scraper");
+  return process.env.MCP_SCRAPER_OUTPUT_DIR?.trim() || join6(homedir4(), "Downloads", "mcp-scraper");
 }
 function safeFilePart(value) {
   return value.replace(/[^a-zA-Z0-9._-]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 120) || "replay";
@@ -5389,7 +5660,7 @@ function replayFilePath(sessionId, replayId, filename) {
   const requested = filename?.trim();
   const stamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "-");
   const name = requested ? safeFilePart(requested).replace(/\.mp4$/i, "") : `${stamp}-${safeFilePart(sessionId)}-${safeFilePart(replayId)}`;
-  return join5(outputBaseDir3(), "browser-replays", `${name}.mp4`);
+  return join6(outputBaseDir3(), "browser-replays", `${name}.mp4`);
 }
 function slugPart(value) {
   const trimmed = value?.trim().toLowerCase();
@@ -5463,8 +5734,8 @@ function registerBrowserAgentMcpTools(server, opts) {
       }
       const bytes = Buffer.from(await res.arrayBuffer());
       const filePath = replayFilePath(sessionId, replayId, filename);
-      mkdirSync3(join5(outputBaseDir3(), "browser-replays"), { recursive: true });
-      writeFileSync3(filePath, bytes);
+      mkdirSync4(join6(outputBaseDir3(), "browser-replays"), { recursive: true });
+      writeFileSync4(filePath, bytes);
       return {
         ok: true,
         data: {
@@ -5505,9 +5776,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_profile_connect",
     {
       title: "Save a Site Login to a Profile",
-      description: `Save a logged-in browser session so this server can reuse a site the user is signed into (ChatGPT, Claude, Reddit, or any account-gated site). Returns an mcpscraper.dev watch_url \u2014 give it to the user, they sign in once, and the cookies are saved to a named profile for later browser_open calls. ONE profile holds MANY logins: call this again with the SAME profile and a DIFFERENT domain to stack another account (e.g. add claude.ai to a profile that already has chatgpt.com). Pass note to record what each login is. Side effects: opens a short-lived sign-in session and persists cookies to the profile; it does NOT import the user's existing browser cookies (they log in fresh). NOT for one-off scraping of public pages (use extract_url) and NOT a way to drive the browser (use browser_open). Sample asks: "save my ChatGPT login", "connect my Reddit account", "add my Claude login to the same profile". After the user signs in, poll browser_profile_list until the login reads AUTHENTICATED, then browser_open with the profile.`,
+      description: "Save a logged-in browser session so this server can reuse a site the user is signed into (ChatGPT, Claude, Reddit, any account-gated site). Returns a watch_url; the user signs in fresh (existing browser cookies are NOT imported) and cookies save to a named profile. ONE profile holds MANY logins \u2014 call again with the same profile and a different domain to stack another account. NOT for one-off scraping (use extract_url) or driving the browser (use browser_open). After sign-in, poll browser_profile_list until AUTHENTICATED, then browser_open with the profile.",
       inputSchema: BrowserProfileConnectInputSchema,
-      outputSchema: BrowserProfileConnectOutputSchema,
+      outputSchema: recordOutputSchema("browser_profile_connect", BrowserProfileConnectOutputSchema),
       annotations: annotations("Save a Site Login to a Profile")
     },
     async (input) => {
@@ -5563,9 +5834,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_profile_list",
     {
       title: "List Saved Logins in a Profile",
-      description: `List every site login saved in a profile, each with its current auth status (NEEDS_AUTH / AUTHENTICATED), account email, and note. Use it to (1) see what a profile is connected to before opening a session, and (2) poll a just-saved login until it flips to AUTHENTICATED. Read-only, no cost. Pass profile (or email to derive it); narrow with domain or connection_id to poll a single login. Sample asks: "what's connected in my profile", "is my ChatGPT login ready yet", "which accounts are saved". Pairs with browser_profile_connect (which adds logins) and browser_open (which uses them).`,
+      description: "List every site login saved in a profile with its auth status (NEEDS_AUTH/AUTHENTICATED), email, and note. Use to check what's connected, or to poll a just-saved login until AUTHENTICATED. Read-only, no cost. Pass profile (or email to derive it); narrow with domain or connection_id.",
       inputSchema: BrowserProfileListInputSchema,
-      outputSchema: BrowserProfileListOutputSchema,
+      outputSchema: recordOutputSchema("browser_profile_list", BrowserProfileListOutputSchema),
       annotations: annotations("List Saved Logins in a Profile", true)
     },
     async (input) => {
@@ -5597,9 +5868,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_open",
     {
       title: "Open Browser Session",
-      description: "Open a direct no-proxy hosted browser session you can drive. Pass a saved profile name to load a session that is already logged into the sites saved in that profile (one profile can hold several logins). To set up or refresh a login first, use browser_profile_connect, then poll browser_profile_list until it reads AUTHENTICATED. Returns a session_id used by all other browser_* tools.",
+      description: "Open a direct no-proxy hosted browser session you can drive. Pass a saved profile name to load a session already logged into that profile's sites (set one up first with browser_profile_connect). Returns a session_id used by all other browser_* tools.",
       inputSchema: BrowserOpenInputSchema,
-      outputSchema: BrowserOpenOutputSchema,
+      outputSchema: recordOutputSchema("browser_open", BrowserOpenOutputSchema),
       annotations: annotations("Open Browser Session")
     },
     async (input) => {
@@ -5631,9 +5902,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_screenshot",
     {
       title: "See Page (Screenshot + Elements)",
-      description: "Capture what the browser currently shows. Returns a screenshot image PLUS a text snapshot listing interactive elements with their center x,y coordinates, the page url and title, and visible text. This is your primary way to perceive the page. Click elements by their listed x,y. If a Cloudflare/CAPTCHA challenge is visible, wait and screenshot again rather than clicking it.",
+      description: "Capture what the browser currently shows: a screenshot plus a text snapshot of interactive elements with x,y coordinates, page url/title, and visible text. Primary way to perceive the page; click elements by their listed x,y. If a Cloudflare/CAPTCHA challenge is visible, wait and screenshot again rather than clicking it.",
       inputSchema: BrowserSessionInputSchema,
-      outputSchema: BrowserScreenshotOutputSchema,
+      outputSchema: recordOutputSchema("browser_screenshot", BrowserScreenshotOutputSchema),
       annotations: annotations("See Page", true)
     },
     async (input) => {
@@ -5663,9 +5934,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_read",
     {
       title: "Read Page Text + Elements",
-      description: "Return the page url, title, visible text, and the list of interactive elements (with x,y) without an image. Cheaper than browser_screenshot when you only need to read content or find a target element to click.",
+      description: "Return the page url, title, visible text, and interactive elements (with x,y) without an image. Cheaper than browser_screenshot when you only need to read content or find a click target.",
       inputSchema: BrowserSessionInputSchema,
-      outputSchema: BrowserReadOutputSchema,
+      outputSchema: recordOutputSchema("browser_read", BrowserReadOutputSchema),
       annotations: annotations("Read Page", true)
     },
     async (input) => {
@@ -5687,9 +5958,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_locate",
     {
       title: "Locate DOM Targets",
-      description: "Locate exact visible DOM elements or text ranges in the current browser viewport and return left/top/width/height bounds in screenshot pixels. Use this before drawing annotations or when a callout must literally circle, box, underline, or point to a real element. Prefer CSS selectors for exact UI elements; use text when selector is unknown. When a replay is actively recording, the result includes replay_elapsed_seconds for timing.",
+      description: "Locate exact visible DOM elements or text ranges and return left/top/width/height bounds in screenshot pixels. Use before drawing annotations that must circle, box, underline, or point to a real element. Prefer CSS selectors; use text when selector is unknown.",
       inputSchema: BrowserLocateInputSchema,
-      outputSchema: BrowserLocateOutputSchema,
+      outputSchema: recordOutputSchema("browser_locate", BrowserLocateOutputSchema),
       annotations: annotations("Locate DOM Targets", true)
     },
     async (input) => {
@@ -5712,9 +5983,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_goto",
     {
       title: "Navigate To URL",
-      description: "Navigate an existing browser session to a URL when the user asks you to go to another page or continue browsing in the same session. Use browser_open first if no session exists. Follow with browser_screenshot to see redirects, login walls, CAPTCHA/Cloudflare state, or the loaded page.",
+      description: "Navigate an existing browser session to a URL. Use browser_open first if no session exists; follow with browser_screenshot to see the loaded page.",
       inputSchema: BrowserGotoInputSchema,
-      outputSchema: BrowserActionOutputSchema,
+      outputSchema: recordOutputSchema("browser_goto", BrowserActionOutputSchema),
       annotations: annotations("Navigate To URL")
     },
     async (input) => {
@@ -5726,9 +5997,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_click",
     {
       title: "Click",
-      description: "Click a visible page target using screenshot pixel coordinates. Use this when the user asks you to press a button, open a menu, choose a result, or activate a visible UI element. Use x/y only from the latest browser_screenshot, browser_read, or browser_locate result; do not guess coordinates. Follow with browser_screenshot when the click may change the page.",
+      description: "Click a visible page target using screenshot pixel coordinates. Use x/y only from the latest browser_screenshot, browser_read, or browser_locate result; do not guess coordinates.",
       inputSchema: BrowserClickInputSchema,
-      outputSchema: BrowserActionOutputSchema,
+      outputSchema: recordOutputSchema("browser_click", BrowserActionOutputSchema),
       annotations: annotations("Click")
     },
     async (input) => {
@@ -5745,9 +6016,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_type",
     {
       title: "Type Text",
-      description: 'Type text into the currently focused browser field. Use this when the user asks you to fill a search box, form field, login field, or editable text area. Click or Tab to the field first if focus is uncertain. Use browser_press with ["Return"] to submit, and browser_screenshot afterward when page state matters.',
+      description: 'Type text into the currently focused browser field. Click or Tab to the field first if focus is uncertain. Use browser_press with ["Return"] to submit.',
       inputSchema: BrowserTypeInputSchema,
-      outputSchema: BrowserActionOutputSchema,
+      outputSchema: recordOutputSchema("browser_type", BrowserActionOutputSchema),
       annotations: annotations("Type Text")
     },
     async (input) => {
@@ -5759,9 +6030,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_scroll",
     {
       title: "Scroll",
-      description: "Scroll the page to reveal more content before reading, clicking, or locating elements. Positive delta_y scrolls down; negative delta_y scrolls up. Follow with browser_screenshot or browser_read to inspect newly revealed content.",
+      description: "Scroll the page to reveal more content. Positive delta_y scrolls down; negative scrolls up.",
       inputSchema: BrowserScrollInputSchema,
-      outputSchema: BrowserActionOutputSchema,
+      outputSchema: recordOutputSchema("browser_scroll", BrowserActionOutputSchema),
       annotations: annotations("Scroll")
     },
     async (input) => {
@@ -5778,9 +6049,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_press",
     {
       title: "Press Keys",
-      description: 'Press keyboard keys or combinations in the active browser session. Use this for submit, Escape, Tab navigation, select-all, or keyboard shortcuts. Examples: ["Return"], ["Escape"], ["Tab"], ["Ctrl+a"], ["Ctrl+Shift+Tab"]. Use browser_type for text entry and browser_screenshot after keypresses that may change the page.',
+      description: "Press keyboard keys or combinations in the active browser session \u2014 submit, Escape, Tab navigation, select-all, or shortcuts. Use browser_type for text entry.",
       inputSchema: BrowserPressInputSchema,
-      outputSchema: BrowserActionOutputSchema,
+      outputSchema: recordOutputSchema("browser_press", BrowserActionOutputSchema),
       annotations: annotations("Press Keys")
     },
     async (input) => {
@@ -5792,9 +6063,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_replay_start",
     {
       title: "Start Recording",
-      description: "Start recording an MP4 replay of the session. Returns replay_id, view_url when available, and a download_url. Use to capture a task for later review; stop with browser_replay_stop.",
+      description: "Start recording an MP4 replay of the session. Returns replay_id and a download_url. Stop with browser_replay_stop.",
       inputSchema: BrowserSessionInputSchema,
-      outputSchema: BrowserReplayStartOutputSchema,
+      outputSchema: recordOutputSchema("browser_replay_start", BrowserReplayStartOutputSchema),
       annotations: annotations("Start Recording")
     },
     async (input) => {
@@ -5815,9 +6086,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_replay_stop",
     {
       title: "Stop Recording",
-      description: "Stop a replay recording and expose the final view_url and download_url. Use browser_replay_download to save the MP4 to the configured output directory.",
+      description: "Stop a replay recording and expose its final view_url/download_url. Use browser_replay_download to save the MP4.",
       inputSchema: BrowserReplayStopInputSchema,
-      outputSchema: BrowserReplayStopOutputSchema,
+      outputSchema: recordOutputSchema("browser_replay_stop", BrowserReplayStopOutputSchema),
       annotations: annotations("Stop Recording")
     },
     async (input) => {
@@ -5838,9 +6109,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_list_replays",
     {
       title: "List Replay Videos",
-      description: "List replay recordings for a browser session, including final view_url and authenticated download_url values when available.",
+      description: "List replay recordings for a browser session, including view_url and download_url when available.",
       inputSchema: BrowserSessionInputSchema,
-      outputSchema: BrowserListReplaysOutputSchema,
+      outputSchema: recordOutputSchema("browser_list_replays", BrowserListReplaysOutputSchema),
       annotations: annotations("List Replay Videos", true)
     },
     async (input) => {
@@ -5860,9 +6131,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_replay_download",
     {
       title: "Download Replay MP4",
-      description: "Download a replay recording through MCP Scraper and save the MP4 under MCP_SCRAPER_OUTPUT_DIR/browser-replays. Use after browser_replay_stop or browser_list_replays.",
+      description: opts.savesReportsLocally === false ? "Download a replay recording. Returns the download_url; fetch it directly (nothing is saved on this hosted endpoint). Use after browser_replay_stop or browser_list_replays." : "Download a replay recording and save the MP4 under MCP_SCRAPER_OUTPUT_DIR/browser-replays. Use after browser_replay_stop or browser_list_replays.",
       inputSchema: BrowserReplayDownloadInputSchema,
-      outputSchema: BrowserReplayDownloadOutputSchema,
+      outputSchema: recordOutputSchema("browser_replay_download", BrowserReplayDownloadOutputSchema),
       annotations: annotations("Download Replay MP4")
     },
     async (input) => {
@@ -5884,9 +6155,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_replay_mark",
     {
       title: "Mark Replay Annotation",
-      description: "While a replay is actively recording, locate one exact DOM target and return a ready-to-use annotation object with DOM bounds and replay-relative timing. Use this instead of guessing start_seconds or drawing rough rectangles. Workflow: start browser_replay_start, navigate until the target is visible and stable, call browser_replay_mark for each callout, then stop the replay and pass the returned annotations to browser_replay_annotate.",
+      description: "While a replay is actively recording, locate one exact DOM target and return a ready-to-use annotation with DOM bounds and replay-relative timing, instead of guessing start_seconds or rectangles. Pass the returned annotations to browser_replay_annotate after stopping the replay.",
       inputSchema: BrowserReplayMarkInputSchema,
-      outputSchema: BrowserReplayMarkOutputSchema,
+      outputSchema: recordOutputSchema("browser_replay_mark", BrowserReplayMarkOutputSchema),
       annotations: annotations("Mark Replay Annotation", true)
     },
     async (input) => {
@@ -5934,9 +6205,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_replay_annotate",
     {
       title: "Annotate Replay MP4",
-      description: "Download a browser replay MP4, render visual annotations over it, and save a new annotated MP4 under MCP_SCRAPER_OUTPUT_DIR/browser-replays. Use this after browser_replay_stop when the user wants proof videos with circles, boxes, arrows, underlines, or labels. For accurate timing and placement, prefer annotations returned by browser_replay_mark while the replay is recording; otherwise use exact left/top/width/height bounds from browser_locate. If the replay video size differs from the screenshot coordinate space, pass source_width and source_height.",
+      description: "Download a browser replay MP4, render visual annotations (circles/boxes/arrows/labels) over it, and save a new annotated MP4. Prefer annotations from browser_replay_mark for accurate timing; otherwise use exact bounds from browser_locate. Pass source_width/source_height if the replay video size differs from the screenshot coordinate space.",
       inputSchema: BrowserReplayAnnotateInputSchema,
-      outputSchema: BrowserReplayAnnotateOutputSchema,
+      outputSchema: recordOutputSchema("browser_replay_annotate", BrowserReplayAnnotateOutputSchema),
       annotations: annotations("Annotate Replay MP4")
     },
     async (input) => {
@@ -5946,7 +6217,7 @@ function registerBrowserAgentMcpTools(server, opts) {
       try {
         const sourcePath = String(downloaded.data.file_path);
         const outputPath = annotatedReplayFilePath(input.session_id, input.replay_id, input.filename);
-        mkdirSync3(join5(outputBaseDir3(), "browser-replays"), { recursive: true });
+        mkdirSync4(join6(outputBaseDir3(), "browser-replays"), { recursive: true });
         const result = await annotateReplayVideo(sourcePath, outputPath, {
           annotations: input.annotations,
           sourceWidth: input.source_width,
@@ -5976,9 +6247,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_close",
     {
       title: "Close Browser Session",
-      description: "Close and release a browser session when the task is done, when the user asks to stop the browser, or when active browser billing should end. Use browser_list_sessions first if you need to recover the session_id.",
+      description: "Close and release a browser session when the task is done, to end active browser billing. Use browser_list_sessions first to recover a session_id.",
       inputSchema: BrowserSessionInputSchema,
-      outputSchema: BrowserCloseOutputSchema,
+      outputSchema: recordOutputSchema("browser_close", BrowserCloseOutputSchema),
       annotations: { title: "Close Browser Session", readOnlyHint: false, destructiveHint: true, idempotentHint: true, openWorldHint: false }
     },
     async (input) => {
@@ -5997,9 +6268,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "browser_list_sessions",
     {
       title: "List Browser Sessions",
-      description: "List browser sessions and their status, with a watch_url for each. Use this to recover a session_id, find an active browser, or decide which session to close. Use browser_open to create a new session.",
+      description: "List browser sessions and their status, with a watch_url for each. Use to recover a session_id or decide which session to close.",
       inputSchema: BrowserListInputSchema,
-      outputSchema: BrowserListSessionsOutputSchema,
+      outputSchema: recordOutputSchema("browser_list_sessions", BrowserListSessionsOutputSchema),
       annotations: annotations("List Browser Sessions", true)
     },
     async (input) => {
@@ -6019,9 +6290,9 @@ function registerBrowserAgentMcpTools(server, opts) {
     "query_fanout_workflow",
     {
       title: "Capture AI Search Fan-Out",
-      description: `Capture the query fan-out behind a ChatGPT or Claude web-search answer for Answer Engine Optimization (AEO): the exact sub-queries the model issued, every researched URL split into cited vs browsed-only (with citation frequency and snippet), each source tagged by category (First-party/vendor, News/media, Reddit, Social/video, Encyclopedia, Review site, Docs, Blog), plus top sourced sites and citation order. Set export=true for durable JSON, CSV, TSV, and HTML artifacts; export paths are returned relative to MCP_SCRAPER_OUTPUT_DIR. This returns the raw structured data; YOU analyze it \u2014 classify each sub-query by funnel stage (Problem-aware, Solution-aware, Decision-aware, Retention) and type (BoFu, Branded, Comparison, How-to, Search operator), name the brands the model researched, and give AEO insights. Sample asks: "capture the fan-out for this ChatGPT answer", "what did Claude search and cite for best CRM", "show the sub-queries and sources behind this AI answer". WRITE NOTE: passing prompt submits a real message in the user's logged-in account (a new conversation turn) \u2014 only send when the user wants that; omit prompt to capture a prompt the user just ran. Setup, if not already connected or if the profile disconnected: call browser_profile_connect, give the user the watch_url, let them complete login, then poll browser_profile_list until status is AUTHENTICATED. After that, browser_open the saved direct no-proxy profile and go to chatgpt.com or claude.ai. Fan-out is captured only as it streams, so the session must be open when the prompt runs. NOT for Google AI Overview citations \u2014 use harvest_paa for those; this tool is ChatGPT and Claude only.`,
+      description: "Capture the query fan-out behind a ChatGPT or Claude web-search answer for AEO: sub-queries issued, every researched URL split into cited vs browsed-only, and top sourced sites. Returns raw structured data for you to classify and analyze. Set export=true for JSON/CSV/TSV/HTML artifacts. WRITE NOTE: passing prompt submits a real message in the user's logged-in account \u2014 only send when the user wants that; omit it to capture a prompt the user just ran. The session must already be open on chatgpt.com or claude.ai (see browser_profile_connect) while the prompt streams. NOT for Google AI Overview \u2014 use harvest_paa for that.",
       inputSchema: BrowserCaptureFanoutInputSchema,
-      outputSchema: BrowserCaptureFanoutOutputSchema,
+      outputSchema: recordOutputSchema("query_fanout_workflow", BrowserCaptureFanoutOutputSchema),
       annotations: annotations("Capture AI Search Fan-Out")
     },
     async (input) => {
@@ -6084,9 +6355,11 @@ export {
   renderIssueReport,
   auditImages,
   renderImageSection,
+  getBlobStore,
   configureReportSaving,
   outputBaseDir,
   SERVER_INSTRUCTIONS,
+  hashOwnerId,
   registerSerpIntelligenceCaptureTools,
   buildPaaExtractorMcpServer,
   registerPaaExtractorMcpTools,
@@ -6094,4 +6367,4 @@ export {
   exportFanout,
   registerBrowserAgentMcpTools
 };
-//# sourceMappingURL=chunk-NONBSIES.js.map
+//# sourceMappingURL=chunk-SS5YBZVI.js.map