diff --git a/src/lib/ai/coach/plans.ts b/src/lib/ai/coach/plans.ts
index 070335bfe..cade32d18 100644
--- a/src/lib/ai/coach/plans.ts
+++ b/src/lib/ai/coach/plans.ts
@@ -276,6 +276,9 @@ async function loadRecentTurns(
   for (const m of [...row.messages].reverse()) {
     const content = decryptOrNull(m.encryptedContent);
     if (content === null) continue;
+    // #781 — skip empty bodies: a cancelled-turn marker (an empty assistant
+    // row closing an aborted turn) carries nothing to extract from.
+    if (!content.trim()) continue;
     turns.push({ role: m.role, content });
   }
   return turns;
diff --git a/src/lib/rate-limit.ts b/src/lib/rate-limit.ts
index cad590f3b..b8dccc0a5 100644
--- a/src/lib/rate-limit.ts
+++ b/src/lib/rate-limit.ts
@@ -61,6 +61,12 @@ export async function checkRateLimit(
  * once `reset_at` has passed the counter is stale and a decrement would bleed
  * into the NEXT window's budget, so an expired row is left untouched. Never
  * drops below zero (concurrent refunds after a window flip are a no-op).
+ *
+ * Known slack: if the charge's window expires and a new request opens the
+ * NEXT window before a late refund lands, the refund decrements that fresh
+ * window instead — at most one extra slot there, only on failure paths that
+ * never reached the metered work. Accepted; carrying the charge's `reset_at`
+ * through every failure path is not worth one slot of slack.
  */
 export async function refundRateLimit(key: string): Promise<void> {
   await prisma.$executeRaw`