diff --git a/apps/api/app/v1/chat/route.ts b/apps/api/app/v1/chat/route.ts
index 83fa290..7eeab59 100644
--- a/apps/api/app/v1/chat/route.ts
+++ b/apps/api/app/v1/chat/route.ts
@@ -113,7 +113,8 @@ export async function POST(req: Request): Promise<Response> {
   const cors = corsHeaders(req);
   // An in-process caller (the Telegram bridge, proven by the per-boot token) skips the
   // origin/key checks and rate-limits on the upstream user id it names in the bucket header.
-  const gate = await requestGate(req, { internal: isInternalCall(req) });
+  const internal = isInternalCall(req);
+  const gate = await requestGate(req, { internal });
   if (!gate.ok)
     return Response.json({ error: gate.reason }, { status: gate.status, headers: { ...cors, ...gate.headers } });
 
@@ -135,7 +136,7 @@ export async function POST(req: Request): Promise<Response> {
   // per-app breakdown. telegram is server-derived (only the in-process bridge holds the
   // internal token); browsers always send an Origin on a cross-site POST, so an Origin-less
   // call is server-to-server API traffic. Client-declared bits are analytics, not trust.
-  const channel: Channel = isInternalCall(req)
+  const channel: Channel = internal
     ? "telegram"
     : body.channel === "playground"
       ? "playground"
@@ -491,6 +492,13 @@ export async function POST(req: Request): Promise<Response> {
             }
           } finally {
             send({ type: "done" });
+            // The Telegram bridge invokes this handler from inside its own after() callback,
+            // and an after() registered there NEVER fires — which silently dropped usage
+            // rollups and transcripts for every bridge turn. Internal calls therefore persist
+            // INLINE before the stream closes: the bridge reads to EOF, so completion is
+            // guaranteed. Public calls keep after() below so their EOF isn't delayed by DB
+            // writes.
+            if (internal) await persistTurn();
             // close() also throws once the client has disconnected (already-closed controller);
             // swallow it like send() so an abort doesn't reject out of start().
             try {
@@ -507,7 +515,7 @@ export async function POST(req: Request): Promise<Response> {
   // After the streamed response finishes: flush spans, record spend, write the cache, roll up
   // usage, persist the turn. All best-effort and independent — allSettled so a failure in one
   // never skips the others (or affects the already-streamed response).
-  after(async () => {
+  async function persistTurn(): Promise<void> {
     await Promise.allSettled([
       flushTelemetry(),
       // Shared spend meter is best-effort; the in-memory per-instance meter still applies.
@@ -536,7 +544,8 @@ export async function POST(req: Request): Promise<Response> {
           )
         : Promise.resolve(),
     ]);
-  });
+  }
+  if (!internal) after(persistTurn);
 
   return new Response(stream, {
     headers: { ...cors, ...gate.headers, "content-type": "application/x-ndjson", "cache-control": "no-store" },