fix: decode quoted benchmark shell commands

scripts/agent-benchmark/driver.mjs

@@ -45,6 +45,7 @@ import {
   benchmarkCcache,
   ccacheMeasurements,
   runnerToolOutput,
+  topLevelShellCommand,
 } from './run-guards.mjs';
 
 const launchCrashVariant = 'launch-crash';
@@ -1628,12 +1629,6 @@ function commandEvidence(meta, eventsPath, runDir) {
   return { commands, activities, completedEvents, invalidReasons };
 }
 
-function topLevelShellCommand(command) {
-  const trimmed = String(command ?? '').trim();
-  const match = trimmed.match(/^\/bin\/(?:zsh|bash|sh) -lc (["'])([\s\S]*)\1$/);
-  return (match?.[2] ?? trimmed).trim();
-}
-
 function agentDeviceOpenCommand(meta, appAlive) {
   const prefix = agentDeviceCommand(meta, 'open com.appandflow.trailhead --foreground');
   if (!meta.deviceTargetingRequired) return prefix;

scripts/agent-benchmark/run-guards.mjs

@@ -60,10 +60,13 @@ export function benchmarkTarget(config, selection) {
   return { key, machine: config.machine, ...value };
 }
 
-function topLevelShellCommand(command) {
+export function topLevelShellCommand(command) {
   const trimmed = String(command ?? '').trim();
   const match = trimmed.match(/^\/bin\/(?:zsh|bash|sh) -lc (["'])([\s\S]*)\1$/);
-  return (match?.[2] ?? trimmed).trim();
+  if (!match) return trimmed;
+  const source =
+    match[1] === '"' ? match[2].replace(/\\(["\\$`\n])/g, (_, char) => (char === '\n' ? '' : char)) : match[2];
+  return source.trim();
 }
 
 export function shellCommandSegments(command) {

scripts/agent-benchmark/run-guards.test.mjs

@@ -6,6 +6,7 @@ import {
   benchmarkTiming,
   parseBenchmarkTargets,
   shellCommandSegments,
+  topLevelShellCommand,
   stimShellProvenanceInvalidReasons,
   benchmarkCcache,
   assertAndroidDoctorClean,
@@ -211,6 +212,25 @@ describe('benchmark run guards', () => {
       'echo "a && b"',
       'stim guide agent',
     ]);
+    const search = 'rg -n "run:android|agent-device|emulator" .';
+    const body = `${search} | sed -n '1,180p'`;
+    const wrapped = `/bin/zsh -lc ${JSON.stringify(body)}`;
+    expect(shellCommandSegments(wrapped)).toEqual([search, "sed -n '1,180p'"]);
+    expect(agentDeviceIsolationInvalidReasons([{ command: wrapped }], 'env expected agent-device ')).toEqual([]);
+  });
+
+  it('decodes one shell quoting layer while preserving proof command boundaries and literal backslashes', () => {
+    const proof =
+      'env AGENT_DEVICE_STATE_DIR=/tmp/bench AGENT_DEVICE_SESSION=run agent-device wait text "Offline maps"';
+    expect(topLevelShellCommand(`/bin/zsh -lc ${JSON.stringify(proof)}`)).toBe(proof);
+    expect(topLevelShellCommand(`/bin/zsh -lc ${JSON.stringify(`${proof}; agent-device close`)}`)).not.toBe(proof);
+    const literal = "rg '\\d+\\s' file";
+    expect(topLevelShellCommand(`/bin/zsh -lc ${JSON.stringify(literal)}`)).toBe(literal);
+    expect(topLevelShellCommand('/bin/zsh -lc "echo \\$VALUE"')).toBe('echo $VALUE');
+    expect(topLevelShellCommand('/bin/zsh -lc "echo \\q"')).toBe('echo \\q');
+    const quotedNewline = "printf '%s' 'before" + '\\'.repeat(2) + "\nafter'";
+    expect(topLevelShellCommand(`/bin/zsh -lc "${quotedNewline}"`)).toBe("printf '%s' 'before\\\nafter'");
+    expect(topLevelShellCommand('/bin/zsh -lc "echo before\\\nafter"')).toBe('echo beforeafter');
   });
 
   it('rejects setup recovery inside the timer', () => {