fixes

src/accessLive.ts

@@ -1,4 +1,5 @@
 import type { AccessMessage, AccessRequest } from "./access.ts";
+import type { Report } from "./reports.ts";
 
 /**
  * Live access-request updates over WebSocket (replaces HTTP polling).
@@ -8,14 +9,19 @@ import type { AccessMessage, AccessRequest } from "./access.ts";
  *   UUID, exactly like GET /api/access/ticket/:id. Receives this ticket's
  *   messages + status changes.
  * - admin socket   `/api/admin/ws?token=…` — admin Bearer (query param
- *   because browsers can't set WS headers). Receives every access event,
- *   driving the AdminCenter badge + triage list.
+ *   because browsers can't set WS headers). Receives every access event
+ *   plus content-report events, driving the AdminCenter badges + lists.
+ *
+ * Every event carries its full payload: clients apply it straight to state
+ * (no HTTP refetch on event — the socket IS the data, not a hint to poll).
  */
 
 export type AccessLiveEvent =
   | { type: "access_message"; request_id: string; message: AccessMessage }
   | { type: "access_status"; request_id: string; request: AccessRequest }
-  | { type: "access_created"; request_id: string; request: AccessRequest };
+  | { type: "access_created"; request_id: string; request: AccessRequest; message: AccessMessage; message_count: number }
+  | { type: "report_created"; report: Report }
+  | { type: "report_status"; report: Report };
 
 export interface AccessLiveSubscriber {
   /** Ticket to follow, or null for the admin firehose. */
@@ -37,8 +43,15 @@ export function subscribeAccessLive(s: AccessLiveSubscriber): () => void {
 export function broadcastAccessLive(evt: AccessLiveEvent): void {
   if (subscribers.size === 0) return;
   const data = JSON.stringify(evt);
+  // Ticket-scoped events route to that ticket's socket; report events and
+  // everything else fan out to the admin firehose only.
+  const ticketId = "request_id" in evt ? evt.request_id : null;
   for (const s of [...subscribers]) {
-    if (s.ticketId !== evt.request_id && !(s.isAdmin && s.ticketId === null)) continue;
+    if (s.isAdmin && s.ticketId === null) {
+      // Admin firehose: every event.
+    } else if (s.ticketId === null || s.ticketId !== ticketId) {
+      continue;
+    }
     try {
       if (!s.send(data)) subscribers.delete(s);
     } catch {

src/index.ts

@@ -560,8 +560,10 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
       const created = await createAccessRequest({ username, email, message });
       const firstMsg = await addAccessMessage(created.id, "user", message);
       clearAttempts(ip);
-      broadcastAccessLive({ type: "access_created", request_id: created.id, request: created });
-      broadcastAccessLive({ type: "access_message", request_id: created.id, message: firstMsg });
+      // One event carries the whole creation (ticket + its first message):
+      // subscribers apply it directly, no HTTP refetch. No ticket socket
+      // can exist for this id yet, so nothing needs the message separately.
+      broadcastAccessLive({ type: "access_created", request_id: created.id, request: created, message: firstMsg, message_count: 1 });
       // Ticket page: /review/<uuid> (unguessable, same pattern as share links).
       const reviewUrl = `${config.appUrl}/review/${created.id}`;
       sendAccessReceivedEmail(email, username, reviewUrl).catch(() => {});
@@ -1093,6 +1095,7 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
             clientPrompt,
             clientModel,
           });
+          broadcastAccessLive({ type: "report_created", report });
           return json({ report }, 201);
         } catch (e) {
           if ((e as Error).message === "TOO_MANY_OPEN") return json({ error: "too many open reports" }, 429);
@@ -1122,7 +1125,9 @@ const server = Bun.serve<{ ticketId: string | null; isAdmin: boolean }>({
           }
           const adminNote = cleanOptionalStr(body.adminNote, 1000);
           if (adminNote === null) return json({ error: "adminNote: max 1000 chars" }, 400);
-          return json({ report: await setReportStatus(target.id, status, adminNote) });
+          const updated = await setReportStatus(target.id, status, adminNote);
+          if (updated) broadcastAccessLive({ type: "report_status", report: updated });
+          return json({ report: updated });
         }
       }
 

src/tools/ocr.ts

@@ -77,8 +77,7 @@ export class OcrTool implements PlatformTool {
   async transcribe(image: Uint8Array): Promise<{ text: string; truncated: boolean }> {
     if (!(await this.isAvailable())) throw new Error((await this.unavailableReason()) ?? "ocr unavailable");
     const t0 = Date.now();
-    const maxChars = config.ocrMaxChars;
-    let proc;
+    const maxChars = config.ocrMaxChars;    let proc;
     try {
       proc = Bun.spawn([tessBin(), "stdin", "stdout", "-l", config.ocrLang, "--psm", "3"], {
         stdin: image,
@@ -104,7 +103,7 @@ export class OcrTool implements PlatformTool {
       if (code !== 0) {
         throw new Error(`ocr failed (exit ${code}): ${errText.trim().slice(0, 300)}`);
       }
-      let text = out.replace(/\f/g, "").trim();
+      let text = sanitizeOcrText(out);
       let truncated = false;
       if (text.length > maxChars) {
         text = text.slice(0, maxChars);
@@ -118,4 +117,50 @@ export class OcrTool implements PlatformTool {
   }
 }
 
+/**
+ * Strip unrecognizable output from the OCR engine while preserving every
+ * legitimate character (accents, CJK, punctuation, emoji…).
+ *
+ * Tesseract emits raw control bytes, Unicode private-use / unassigned /
+ * surrogate code points and U+FFFD replacement chars when glyphs are
+ * unreadable, plus form-feeds between pages and \r line endings. Those
+ * poison the model prompt and the attachment preview, so they are removed
+ * here. Visible-but-dubious glyphs (|, ~, …) are kept: they may be real
+ * content and only the user can judge them in the review step.
+ */
+export function sanitizeOcrText(raw: string): string {
+  // Normalize line endings first; form-feeds are page breaks, not content.
+  let text = raw.replace(/\r\n?/g, "\n").replace(/\f/g, "\n");
+  // Strip BOMs / zero-width no-break spaces anywhere (keep ZWJ + VS16 so
+  // emoji sequences survive).
+  text = text.replace(/[\uFEFF\u200B\u200C\u2060\u180E]/g, "");
+  let out = "";
+  for (const ch of text) {
+    if (ch === "\n" || ch === "\t" || ch === "\u200D" || ch === "\uFE0F") {
+      out += ch;
+      continue;
+    }
+    const cp = ch.codePointAt(0)!;
+    // U+FFFD (replacement char) and U+FFFE/FFFF are the engine's markers
+    // for unreadable glyphs — never real content.
+    if (cp === 0xfffd || cp === 0xfffe || cp === 0xffff) continue;
+    // Keep letters, marks, numbers, punctuation, symbols and spaces.
+    // Drops Cc/Cf/Cs/Co/Cn (controls, formats, surrogates, private-use,
+    // unassigned) and Zl/Zp separators.
+    if (/^[\p{L}\p{M}\p{N}\p{P}\p{S}\p{Zs}]$/u.test(ch)) {
+      out += ch;
+    }
+    // Anything else is unrecognizable — drop it.
+  }
+  // Tidy whitespace without touching indentation: trim line ends, collapse
+  // 3+ blank lines, drop leading/trailing blank lines.
+  out = out
+    .split("\n")
+    .map((line) => line.replace(/[ \t\u00A0]+$/g, ""))
+    .join("\n")
+    .replace(/\n{3,}/g, "\n\n")
+    .trim();
+  return out;
+}
+
 export { OCR_MAX_BYTES };