diff --git a/crates/oxide_cli/src/main.rs b/crates/oxide_cli/src/main.rs
index 54e1c0c3..22884690 100644
--- a/crates/oxide_cli/src/main.rs
+++ b/crates/oxide_cli/src/main.rs
@@ -46,11 +46,17 @@ struct Cli {
 enum Commands {
     Eval {
         code: String,
+        /// 逐指令 trace：每条指令向 stderr 写一行 pc + opcode + 操作数。
+        #[arg(long)]
+        trace: bool,
     },
     Run {
         file: String,
         #[arg(long, default_value = "1")]
         repeat: u64,
+        /// 逐指令 trace：每条指令向 stderr 写一行 pc + opcode + 操作数。
+        #[arg(long)]
+        trace: bool,
     },
     Compile {
         #[arg(short = 'e')]
@@ -87,16 +93,16 @@ fn main() -> ExitCode {
     let cli = Cli::parse();
 
     match cli.command {
-        Some(Commands::Eval { code }) => {
+        Some(Commands::Eval { code, trace }) => {
             let kernel = make_kernel(cli.verbose, cli.quiet);
             let pool = make_pool(&kernel);
-            eval(&code, &kernel, &pool)
+            eval(&code, &kernel, &pool, trace)
         }
-        Some(Commands::Run { file, repeat }) => {
+        Some(Commands::Run { file, repeat, trace }) => {
             let kernel = make_kernel(cli.verbose, cli.quiet);
             let pool = make_pool(&kernel);
             for n in 0..repeat {
-                let code = run(&file, &kernel, &pool);
+                let code = run(&file, &kernel, &pool, trace);
                 if code != ExitCode::SUCCESS && code != ExitCode::FAILURE {
                     return code;
                 }
@@ -155,7 +161,7 @@ fn make_pool(kernel: &Arc<KernelCore>) -> Arc<VmPool> {
     VmPool::new(Arc::clone(kernel), kernel.config.min_pool_size, kernel.config.max_pool_size)
 }
 
-fn eval(code: &str, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>) -> ExitCode {
+fn eval(code: &str, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>, trace: bool) -> ExitCode {
     let allocator = Allocator::default();
     let program = match oxide_parser::parse(&allocator, code) {
         Ok(p) => p,
@@ -180,6 +186,7 @@ fn eval(code: &str, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>) -> ExitCode {
     };
 
     let mut guard = pool.spawn();
+    guard.vm_mut().set_instruction_trace(trace);
     match guard.vm_mut().run(&module) {
         Ok(result) => {
             format_result(guard.vm(), kernel.perm_interner().as_ref(), kernel.shape_forge().as_ref(), result);
@@ -284,9 +291,9 @@ fn format_array(
     format!("[{}]", items.join(", "))
 }
 
-fn run(file: &str, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>) -> ExitCode {
+fn run(file: &str, kernel: &Arc<KernelCore>, pool: &Arc<VmPool>, trace: bool) -> ExitCode {
     match fs::read_to_string(file) {
-        Ok(source) => eval(&source, kernel, pool),
+        Ok(source) => eval(&source, kernel, pool, trace),
         Err(err) => {
             kernel_error!("cannot read {}: {}", file, err);
             eprintln!("{}", Red.paint(format!("Cannot read {file}: {err}")));
diff --git a/crates/oxide_cli/tests/cli_smoke.rs b/crates/oxide_cli/tests/cli_smoke.rs
index 282bca71..56f5692f 100644
--- a/crates/oxide_cli/tests/cli_smoke.rs
+++ b/crates/oxide_cli/tests/cli_smoke.rs
@@ -47,3 +47,34 @@ fn bench_help() {
     let stdout = String::from_utf8_lossy(&output.stdout);
     assert!(stdout.contains("--iterations"), "bench --help should list --iterations flag");
 }
+
+#[test]
+fn eval_trace_flag_pc_lines() {
+    let output = Command::new(env!("CARGO_BIN_EXE_oxide"))
+        .args(["eval", "--trace", "1 + 2"])
+        .output()
+        .expect("failed to run oxide eval --trace");
+
+    assert!(output.status.success(), "eval --trace should exit 0");
+    // trace 行只走 stderr，stdout 保持纯结果。
+    let stdout = String::from_utf8_lossy(&output.stdout);
+    assert_eq!(stdout.trim(), "3", "--trace 不应污染 stdout: {stdout}");
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    assert!(stderr.lines().all(|l| l.starts_with("pc=")), "trace 行应全部为 pc 行: {stderr}");
+    assert!(stderr.contains("LOAD_CONST"), "trace 应含 LOAD_CONST: {stderr}");
+    assert!(stderr.contains("ADD"), "trace 应含 ADD: {stderr}");
+    assert!(stderr.contains("HALT"), "trace 应含 HALT: {stderr}");
+}
+
+#[test]
+fn eval_no_trace_flag_no_pc_lines() {
+    let output = Command::new(env!("CARGO_BIN_EXE_oxide"))
+        .args(["eval", "1 + 2"])
+        .output()
+        .expect("failed to run oxide eval");
+
+    assert!(output.status.success(), "eval should exit 0");
+    // 默认路径零 trace 输出。
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    assert!(!stderr.contains("pc="), "默认路径不应有 trace 行: {stderr}");
+}
diff --git a/crates/oxide_vm/src/vm/dispatch_loop.rs b/crates/oxide_vm/src/vm/dispatch_loop.rs
index 183b8bfb..add49f4c 100644
--- a/crates/oxide_vm/src/vm/dispatch_loop.rs
+++ b/crates/oxide_vm/src/vm/dispatch_loop.rs
@@ -202,12 +202,19 @@ impl Vm {
                 return Err("program counter out of bounds".into());
             }
 
-            let instr = self.bytecode[self.pc];
+            let pc = self.pc;
+            let instr = self.bytecode[pc];
             let op = opcode::opcode(instr);
             let rd = opcode::rd(instr) as usize;
             let a = opcode::a(instr) as usize;
             let b = opcode::b(instr) as usize;
-            self.pc += 1;
+            self.pc = pc + 1;
+
+            // 逐指令 trace：运行时开关短路（默认关闭，零输出零分配）；
+            // pc 为字节码 word 序号，与反汇编 offset 列同单位。
+            if self.trace_instructions {
+                eprintln!("pc={pc:4} {op:>16} rd={rd} a={a} b={b}");
+            }
 
             match op {
                 OpCode::NOP => {}
diff --git a/crates/oxide_vm/src/vm/mod.rs b/crates/oxide_vm/src/vm/mod.rs
index dcd1798a..c1d7cf66 100644
--- a/crates/oxide_vm/src/vm/mod.rs
+++ b/crates/oxide_vm/src/vm/mod.rs
@@ -350,6 +350,10 @@ pub struct Vm {
     pub(crate) saved_flat_id_stack: Vec<u32>,
     /// 帧切换时暂存调用方表代际的栈（与 saved_flat_id_stack 同步 push/pop）。
     pub(crate) saved_table_gen_stack: Vec<u32>,
+    /// 逐指令 trace 开关：开启时 dispatch 主循环对每条指令向 stderr 写一行
+    /// `pc + opcode + 操作数`。运行时开关，旁路 tracing 级别上限（release
+    /// 构建可用）；默认关闭，关闭时热路径仅一次 bool 比较，零输出零分配。
+    pub trace_instructions: bool,
 }
 
 impl Drop for Vm {
@@ -413,6 +417,12 @@ impl Vm {
         self.regs[idx as usize] = val;
     }
 
+    /// 设置逐指令 trace 开关（开启后 dispatch 主循环逐指令向 stderr 写
+    /// `pc + opcode + 操作数` 行，pc 为字节码 word 序号，与反汇编 offset 同单位）。
+    pub fn set_instruction_trace(&mut self, on: bool) {
+        self.trace_instructions = on;
+    }
+
     /// 只读访问 VM 的 epoch arena。
     pub fn epoch(&self) -> &Epoch {
         &self.epoch
diff --git a/crates/oxide_vm/src/vm_support.rs b/crates/oxide_vm/src/vm_support.rs
index 4745e673..c035b0fe 100644
--- a/crates/oxide_vm/src/vm_support.rs
+++ b/crates/oxide_vm/src/vm_support.rs
@@ -157,6 +157,7 @@ impl Vm {
             active_table_gen: 0,
             saved_flat_id_stack: Vec::new(),
             saved_table_gen_stack: Vec::new(),
+            trace_instructions: false,
         };
         vm.init_generator_intrinsics();
         vm.init_promise_intrinsics();
@@ -291,6 +292,7 @@ impl Vm {
             active_table_gen: 0,
             saved_flat_id_stack: Vec::new(),
             saved_table_gen_stack: Vec::new(),
+            trace_instructions: false,
         };
         vm.init_generator_intrinsics();
         vm.init_promise_intrinsics();