From 7bab3dc9bf174dcb8b13c9f368596135929ed537 Mon Sep 17 00:00:00 2001
From: bowlerjim <jcrews@gmail.com>
Date: Sun, 20 Sep 2026 23:30:06 -0400
Subject: [PATCH] fix(Evals): run static_analysis commands through a shell and
 gate on exit code (#2184)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

`await $`cd ${workingDir} && ${command}`` escapes the interpolation into a
single argv token, so a command like `bun --version` was executed as one
program of that literal name. The shell answered `command not found`, which
contains no `error`/`failed` keyword, so countIssues() returned 0 and
`passed = errors === 0` was true. Every static_analysis grader scored 1.0
regardless of what it was asked to run — a lint or typecheck suite that
could not go red.

Spawn `/bin/sh -c <command>` (`cmd /c` on win32) and fail on a non-zero
exit as well as on counted errors, so a tool that reports only through its
exit status is graded correctly.

Verified: commands `bun --version` -> 1/1 passed, `echo hello && exit 1` ->
0/1 passed, `echo all good` -> 1/1 passed. Before the change all three
scored 1.0.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PuS77R8a6Lhp7jK6dnwdTh
---
 .../Evals/Graders/CodeBased/StaticAnalysis.ts | 26 +++++++++++++++----
 1 file changed, 21 insertions(+), 5 deletions(-)

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,