diff --git a/src/bytecode.rs b/src/bytecode.rs
index 3672dcd..ee869b7 100644
--- a/src/bytecode.rs
+++ b/src/bytecode.rs
@@ -25,7 +25,7 @@ use debug2::Formatter;
 pub struct FnBlock<'bc> {
     /// The bytecode of the function
     pub code: Vec<'bc, Instr>,
-    /// The sizes of the stack required by the function at each instruction. This is only used
+    /// The sizes of the stack required by the function after the instruction at the same index. This is only used
     /// during compilation to calculate local variable offsets.
     pub stack_sizes: Vec<'bc, usize>,
     /// The corresponding source code location of each instruction. This is debuginfo and only
diff --git a/src/compile.rs b/src/compile.rs
index 0bbb79e..af2b5e4 100644
--- a/src/compile.rs
+++ b/src/compile.rs
@@ -213,11 +213,7 @@ impl<'bc, 'gc> Compiler<'bc, 'gc> {
 
         self.compile_block(ast_block)?;
 
-        self.push_instr(
-            Instr::ShrinkStack(self.current_stack_size() - pre_loop_stack_size),
-            StackChange::None,
-            span,
-        );
+        self.shrink_stack(pre_loop_stack_size, span);
 
         let jmp_offset = self.back_jmp_offset(first_stmt_idx);
         self.push_instr(Instr::Jmp(jmp_offset), StackChange::None, span);
@@ -247,11 +243,7 @@ impl<'bc, 'gc> Compiler<'bc, 'gc> {
 
         self.compile_block(&while_stmt.body)?;
 
-        self.push_instr(
-            Instr::ShrinkStack(self.current_stack_size() - pre_loop_stack_size),
-            StackChange::None,
-            while_stmt.span,
-        );
+        self.shrink_stack(pre_loop_stack_size, while_stmt.span);
         let jmp_offset = self.back_jmp_offset(cond_index);
         self.push_instr(Instr::Jmp(jmp_offset), StackChange::None, while_stmt.span);
 
@@ -370,6 +362,20 @@ impl<'bc, 'gc> Compiler<'bc, 'gc> {
         todo!()
     }
 
+    fn shrink_stack(&mut self, jmp_target_stack_size: usize, span: Span) {
+        let amount = self.current_stack_size() - jmp_target_stack_size;
+
+        if amount == 0 {
+            return;
+        }
+
+        self.push_instr(
+            Instr::ShrinkStack(amount),
+            StackChange::ShrinkN(amount),
+            span,
+        );
+    }
+
     fn end_loop(&mut self) {
         let breaks = self.breaks.remove(&self.loop_nesting);
         if let Some(breaks) = breaks {
@@ -418,7 +424,7 @@ impl<'bc, 'gc> Compiler<'bc, 'gc> {
     fn push_instr(&mut self, instr: Instr, stack_change: StackChange, span: Span) -> usize {
         let block = &mut self.blocks[self.current_block];
         let stack_top = block.stack_sizes.last().copied().unwrap_or(0);
-        let new_stack_top = stack_top as isize + stack_change as isize;
+        let new_stack_top = stack_top as isize + stack_change.as_isize();
         assert!(new_stack_top >= 0, "instruction popped stack below 0");
         let new_stack_top = new_stack_top as usize;
 
@@ -434,9 +440,20 @@ impl<'bc, 'gc> Compiler<'bc, 'gc> {
 }
 
 #[derive(Debug, Copy, Clone)]
-#[repr(i8)]
 enum StackChange {
-    Shrink = -1,
-    None = 0,
-    Grow = 1,
+    Shrink,
+    None,
+    Grow,
+    ShrinkN(usize),
+}
+
+impl StackChange {
+    fn as_isize(&self) -> isize {
+        match self {
+            StackChange::Shrink => -1,
+            StackChange::None => 0,
+            StackChange::Grow => 1,
+            StackChange::ShrinkN(n) => -(*n as isize),
+        }
+    }
 }
diff --git a/src/lex.rs b/src/lex.rs
index a22f739..04e205c 100644
--- a/src/lex.rs
+++ b/src/lex.rs
@@ -276,22 +276,28 @@ impl<'code, 'gc> Iterator for Lexer<'code, 'gc> {
                 }
                 char => {
                     if char.is_ascii_digit() {
+                        let mut num_buffer = String::from(char); // we need to ignore `_`
                         let mut had_dot = false;
                         let end = loop {
                             // peek here because the character signaling the end should not be consumed
-                            match self.code.peek() {
+                            match self.code.peek().copied() {
                                 Some((_, '.')) if !had_dot => {
                                     let _ = self.code.next();
+                                    num_buffer.push('.');
                                     had_dot = true;
                                 }
+                                Some((_, '_')) => {
+                                    let _ = self.code.next();
+                                }
                                 Some((_, next_char)) if next_char.is_ascii_digit() => {
                                     let _ = self.code.next();
+                                    num_buffer.push(next_char);
                                 }
-                                Some((end, _)) => break *end,
+                                Some((end, _)) => break end,
                                 None => break self.src.len(), // reached EOF, so parse this number
                             }
                         };
-                        let number_str = &self.src[start..end];
+                        let number_str = &num_buffer;
                         let span = Span::start_end(start, end);
                         let number = number_str.parse::<f64>();
                         break match number {
@@ -449,6 +455,16 @@ pls :) o(*￣▽￣*)ブ
         lex_test("3 . . 2 . . 1 . . 0");
     }
 
+    #[test]
+    fn underscore_number() {
+        lex_test("1_000_000");
+    }
+
+    #[test]
+    fn trailing_underscore_number() {
+        lex_test("1_00_");
+    }
+
     #[test]
     fn larger_numbers() {
         lex_test("123456789, 123456789.1234, 64785903");
diff --git a/src/lib.rs b/src/lib.rs
index dbf6504..b85f553 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -33,10 +33,15 @@ type HashSet<T> = rustc_hash::FxHashSet<T>;
 
 pub struct Config<'io> {
     pub debug: bool,
+    pub step: bool,
     pub stdout: &'io mut dyn Write,
 }
 
 pub fn run_program(program: &str, cfg: &mut Config) {
+    if cfg.debug {
+        eprintln!("Config: debug: {}, step: {}", cfg.debug, cfg.step);
+    }
+
     let ast_alloc = Bump::new();
 
     // SAFETY: I will try to 🥺
@@ -73,7 +78,7 @@ fn process_ast(program: &str, ast: &Program, mut runtime: RtAlloc, cfg: &mut Con
                 }
             }
 
-            let result = vm::execute(&code, runtime, cfg.stdout);
+            let result = vm::execute(code, runtime, cfg);
             if let Err(result) = result {
                 eprintln!("error: {}", result);
             }
diff --git a/src/main.rs b/src/main.rs
index 5e58915..dfe7190 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -9,13 +9,14 @@ fn main() {
 
         let mut cfg = Config {
             debug: false,
+            step: false,
             stdout: &mut stdout,
         };
 
         for arg in args {
             match &*arg {
                 "--debug" => cfg.debug = true,
-                "--shut-up-clippy" => println!("yeah shut up pls"), // please do
+                "--step" => cfg.step = true,
                 _ => {}
             }
         }
diff --git a/src/snapshots/dilaria__lex__test__trailing_underscore_number.snap b/src/snapshots/dilaria__lex__test__trailing_underscore_number.snap
new file mode 100644
index 0000000..6869b58
--- /dev/null
+++ b/src/snapshots/dilaria__lex__test__trailing_underscore_number.snap
@@ -0,0 +1,11 @@
+---
+source: src/lex.rs
+assertion_line: 375
+expression: tokens
+
+---
+[
+    Number(
+        100.0,
+    ),
+]
diff --git a/src/snapshots/dilaria__lex__test__underscore_number.snap b/src/snapshots/dilaria__lex__test__underscore_number.snap
new file mode 100644
index 0000000..ce20b0f
--- /dev/null
+++ b/src/snapshots/dilaria__lex__test__underscore_number.snap
@@ -0,0 +1,11 @@
+---
+source: src/lex.rs
+assertion_line: 375
+expression: tokens
+
+---
+[
+    Number(
+        1000000.0,
+    ),
+]
diff --git a/src/vm.rs b/src/vm.rs
index 02dd450..645dc04 100644
--- a/src/vm.rs
+++ b/src/vm.rs
@@ -1,7 +1,8 @@
 use crate::bytecode::{FnBlock, Instr};
 use crate::gc::{Object, RtAlloc, Symbol};
+use crate::Config;
 use std::fmt::{Debug, Display, Formatter};
-use std::io::Write;
+use std::io::{Read, Write};
 
 type VmError = &'static str;
 type VmResult = Result<(), VmError>;
@@ -9,7 +10,7 @@ type VmResult = Result<(), VmError>;
 pub fn execute<'bc>(
     bytecode: &'bc [FnBlock<'bc>],
     alloc: RtAlloc,
-    stdout: &mut dyn Write,
+    cfg: &mut Config,
 ) -> Result<(), VmError> {
     let mut vm = Vm {
         _blocks: bytecode,
@@ -17,7 +18,8 @@ pub fn execute<'bc>(
         pc: 0,
         stack: Vec::with_capacity(1024 << 5),
         _alloc: alloc,
-        stdout,
+        stdout: cfg.stdout,
+        step: cfg.step,
     };
 
     vm.execute_function()
@@ -48,9 +50,11 @@ struct Vm<'bc, 'io> {
     _blocks: &'bc [FnBlock<'bc>],
     current: &'bc FnBlock<'bc>,
     _alloc: RtAlloc,
+    /// Index of the instruction currently being executed
     pc: usize,
     stack: Vec<Value>,
     stdout: &'io mut dyn Write,
+    step: bool,
 }
 
 impl<'bc> Vm<'bc, '_> {
@@ -69,11 +73,15 @@ impl<'bc> Vm<'bc, '_> {
     }
 
     fn dispatch_instr(&mut self, instr: Instr) -> VmResult {
+        if self.step {
+            self.step_debug();
+        }
+
         match instr {
             Instr::Nop => {}
             Instr::Store(index) => {
                 let val = self.stack.pop().unwrap();
-                self.stack.insert(index, val);
+                self.stack[index] = val;
             }
             Instr::Load(index) => self.stack.push(self.stack[index]),
             Instr::PushVal(value) => self.stack.push(value),
@@ -163,7 +171,7 @@ impl<'bc> Vm<'bc, '_> {
             }
             Instr::Jmp(pos) => self.pc = (self.pc as isize + pos) as usize,
             Instr::ShrinkStack(size) => {
-                assert!(self.stack.len() > size);
+                assert!(self.stack.len() >= size);
                 let new_len = self.stack.len() - size;
                 // SAFETY: We only ever shrink the vec, and we don't overflow. Value is copy so no leaks as a bonus
                 unsafe { self.stack.set_len(new_len) }
@@ -188,6 +196,22 @@ impl<'bc> Vm<'bc, '_> {
     fn type_error(&self) -> VmError {
         "bad type"
     }
+
+    fn step_debug(&self) {
+        let current_instr = &self.current.code[self.pc];
+        let curr_stack_size = self.stack.len();
+        let expected_stack_size = &self.current.stack_sizes[self.pc];
+
+        eprintln!(
+            "Current Instruction: {:?}
+Current Stack size: {}
+Expected Stack size after instruction: {}",
+            current_instr, curr_stack_size, expected_stack_size
+        );
+
+        let mut buf = [0; 64];
+        let _ = std::io::stdin().read(&mut buf);
+    }
 }
 
 impl Display for Value {
diff --git a/test.dil b/test.dil
index 791f9b0..7a64ee2 100644
--- a/test.dil
+++ b/test.dil
@@ -1,9 +1,5 @@
 let i = 0;
 
 while i < 1000 {
-    print "lol";
     i = i + 1;
-}
-
-
-print "hi";
\ No newline at end of file
+}
\ No newline at end of file
diff --git a/tests/common.rs b/tests/common.rs
index 5f01285..55d82c3 100644
--- a/tests/common.rs
+++ b/tests/common.rs
@@ -15,6 +15,7 @@ pub fn _run_test(code: &str) -> String {
     let mut stdout = Vec::<u8>::new();
     let mut cfg = dilaria::Config {
         debug: false,
+        step: false,
         stdout: &mut stdout,
     };