diff --git a/LifeOS/install/skills/Evals/Graders/CodeBased/StaticAnalysis.ts b/LifeOS/install/skills/Evals/Graders/CodeBased/StaticAnalysis.ts
index 2ef1f59858..6fec77a99f 100755
--- a/LifeOS/install/skills/Evals/Graders/CodeBased/StaticAnalysis.ts
+++ b/LifeOS/install/skills/Evals/Graders/CodeBased/StaticAnalysis.ts
@@ -5,7 +5,9 @@
 
 import { BaseGrader, registerGrader, type GraderContext } from '../Base.ts';
 import type { GraderConfig, GraderResult, StaticAnalysisParams } from '../../Types/index.ts';
-import { $ } from 'bun';
+
+// Windows has no /bin/sh; cmd takes /c where POSIX shells take -c.
+const SHELL_PREFIX = process.platform === 'win32' ? ['cmd', '/c'] : ['/bin/sh', '-c'];
 
 export class StaticAnalysisGrader extends BaseGrader {
   type = 'static_analysis' as const;
@@ -26,14 +28,28 @@ export class StaticAnalysisGrader extends BaseGrader {
 
     for (const command of params.commands) {
       try {
-        const result = await $`cd ${workingDir} && ${command}`.quiet().nothrow();
+        // Run through a shell: the command is a user-authored string ("bun test",
+        // "eslint . --max-warnings 0"), and a tagged-template interpolation would
+        // escape the whole string into a single argv token, so nothing would run.
+        const proc = Bun.spawn([...SHELL_PREFIX, command], {
+          cwd: workingDir,
+          stdout: 'pipe',
+          stderr: 'pipe',
+        });
+        const [stdout, stderr, exitCode] = await Promise.all([
+          new Response(proc.stdout).text(),
+          new Response(proc.stderr).text(),
+          proc.exited,
+        ]);
 
-        const output = result.stdout.toString() + result.stderr.toString();
+        const output = stdout + stderr;
         const warnings = this.countIssues(output, 'warning');
         const errors = this.countIssues(output, 'error');
 
-        // Pass if no errors (and no warnings if fail_on_warning is set)
-        const passed = errors === 0 && (!params.fail_on_warning || warnings === 0);
+        // A non-zero exit is a failure on its own: a tool that reports its verdict
+        // only through its exit status (tsc, most test runners) prints nothing the
+        // keyword scan below would ever count.
+        const passed = exitCode === 0 && errors === 0 && (!params.fail_on_warning || warnings === 0);
 
         results.push({
           command,