diff --git a/poc-agent/src/tui.rs b/poc-agent/src/tui.rs
index 4d69034..6f6c90b 100644
--- a/poc-agent/src/tui.rs
+++ b/poc-agent/src/tui.rs
@@ -821,6 +821,7 @@ impl App {
         // Cursor position: scan the rendered buffer to find where the cursor should be.
         // This matches ratatui's actual word wrapping instead of trying to simulate it.
         let buffer = frame.buffer_mut();
+        let cursor_char = prompt.len() + self.cursor;  // Total chars from start (prompt + input)
         let mut char_count = 0usize;
         let mut cursor_x = input_area.x;
         let mut cursor_y = input_area.y;
@@ -833,7 +834,7 @@ impl App {
                     // Count visible characters (skip zero-width and empty)
                     if !symbol.is_empty() {
                         let width = symbol.width();
-                        if char_count + width > self.cursor {
+                        if char_count + width > cursor_char {
                             // Found the cursor position
                             cursor_x = x;
                             cursor_y = y;
diff --git a/scripts/provision-vllm.sh b/scripts/provision-vllm.sh
index ee35670..e5702ed 100755
--- a/scripts/provision-vllm.sh
+++ b/scripts/provision-vllm.sh
@@ -51,7 +51,7 @@ exec vllm serve "$MODEL" \
     --max-model-len "$MAX_MODEL_LEN" \
     --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION" \
     --enable-prefix-caching \
-    --tool-call-parser qwen3_xml \
+    --tool-call-parser hermes \
     --enable-auto-tool-choice \
     --reasoning-parser=qwen3 \
     --uvicorn-log-level warning