diff --git a/scripts/scrum_review.sh b/scripts/scrum_review.sh
index 845fe1e..52bebe3 100755
--- a/scripts/scrum_review.sh
+++ b/scripts/scrum_review.sh
@@ -75,20 +75,27 @@ run_review() {
 \`\`\`diff
 $DIFF_CONTENT
 \`\`\`"
-  local body
-  body=$(jq -n --arg model "$model" --arg sys "$SYSTEM" --arg user "$user" \
-    '{model:$model, max_tokens:4096, messages:[{role:"system",content:$sys},{role:"user",content:$user}]}')
   printf "  %-6s %s ... " "$short" "$model"
   local t0=$SECONDS
   local status
-  # Pipe the body via stdin (`-d @-`) — large diffs (>~128KB) blow
-  # past the kernel's argv limit when passed via `--data <literal>`.
-  # Phase A+B was 128875 bytes and hit "Argument list too long" until
-  # this fix.
-  status=$(printf '%s' "$body" | curl -sS -o /tmp/scrum_resp.json -w '%{http_code}' --max-time 240 \
+
+  # Build the body via temp files — both jq's --arg AND curl's
+  # --data run into the kernel's argv limit (~128KB) when the diff
+  # is large. Voice-ai full bundle was 156K and hit it twice.
+  # Piping through files (and using --rawfile for jq) sidesteps both.
+  local body_file user_file sys_file
+  body_file=$(mktemp); user_file=$(mktemp); sys_file=$(mktemp)
+  printf '%s' "$user" > "$user_file"
+  printf '%s' "$SYSTEM" > "$sys_file"
+  jq -n --arg model "$model" --rawfile sys "$sys_file" --rawfile user "$user_file" \
+    '{model:$model, max_tokens:4096, messages:[{role:"system",content:$sys},{role:"user",content:$user}]}' \
+    > "$body_file"
+
+  status=$(curl -sS -o /tmp/scrum_resp.json -w '%{http_code}' --max-time 240 \
     -X POST "$GATEWAY/v1/chat" \
     -H 'Content-Type: application/json' \
-    --data-binary @-)
+    --data-binary "@$body_file")
+  rm -f "$body_file" "$user_file" "$sys_file"
   local elapsed=$((SECONDS - t0))
   if [ "$status" != "200" ]; then
     printf "✗ HTTP %s (%ds)\n" "$status" "$elapsed"