test(e2e): track WSL stub agent staging state for precise cleanup

Refactor `stageWslGoldenStubAgent` to track which artifacts it creates
during setup, then only remove those artifacts during cleanup. This
prevents the test from destructively removing pre-existing symlinks or
state from previous runs, improving test isolation and idempotency.

tests/e2e/golden-tab-bar-agent-launch.spec.ts

@@ -17,11 +17,7 @@ import { ensureTerminalVisible, waitForActiveWorktree, waitForSessionReady } fro
 import { getTerminalContent } from './helpers/terminal'
 import type { BuiltInWindowsTerminalShell } from '../../src/shared/windows-terminal-shell'
 
-// Covers the full tab-bar `+` agent launch chain — detection row, startup-plan
-// build, tab create, PTY spawn under the configured runtime, startup-command
-// injection. golden-agent-tui-launch.spec.ts covers one agent on the default
-// shell and never runs in the Windows golden lane, so a Windows-only break
-// anywhere in that chain ships as "clicking an agent does nothing".
+// Covers the Windows-only tab-bar launch path that the default-shell test misses.
 
 test.use({ launchEnv: getGoldenStubAgentLaunchEnv() })
 
@@ -80,10 +76,8 @@ test.describe('Windows runtimes', () => {
 
     const distro = await getFirstWslDistro(orcaPage)
     test.skip(!distro, 'No WSL distro is available on this Windows host')
-    test.skip(
-      !stageWslGoldenStubAgent(distro!),
-      'WSL distro would not accept the staged stub agent'
-    )
+    const stage = stageWslGoldenStubAgent(distro!)
+    test.skip(!stage, 'WSL distro would not accept the staged stub agent')
 
     try {
       // Why the runtime switch and not terminalWindowsShell: WSL is a project
@@ -97,7 +91,7 @@ test.describe('Windows runtimes', () => {
       // ran in WSL rather than on the Windows host.
       expect(await getTerminalContent(orcaPage)).toContain(GOLDEN_STUB_READY_MARKER)
     } finally {
-      removeWslGoldenStubAgent(distro!)
+      removeWslGoldenStubAgent(distro!, stage!)
     }
   })
 })

tests/e2e/helpers/golden-stub-agent.ts

@@ -7,8 +7,7 @@ import type { BuiltInWindowsTerminalShell } from '../../../src/shared/windows-te
 export const GOLDEN_STUB_READY_MARKER = 'GOLDEN_STUB_AGENT_READY'
 export const GOLDEN_STUB_EXIT_MARKER = 'GOLDEN_STUB_AGENT_EXITED'
 
-/** Agents the fixture dir shims onto PATH, so detection lists them and the
- *  tab-bar `+` menu renders a launch row for each. */
+/** Agents exposed by the fixture directory for tab-bar detection. */
 export const GOLDEN_STUB_AGENTS = [
   { id: 'codex', menuItemName: /^Codex(?:\s|$)/i },
   { id: 'claude', menuItemName: /^Claude(?:\s|$)/i }

tests/e2e/helpers/wsl-golden-stub-agent.ts

@@ -2,8 +2,7 @@ import { execFileSync } from 'node:child_process'
 import type { Page } from '@stablyai/playwright-test'
 import { buildWslExecArgs } from '../../../src/shared/wsl-login-shell-command'
 
-/** Distro path the stub is staged at. Only reachable from inside WSL, so the
- *  stub's marker in a pane is itself proof the pane executed in the distro. */
+/** A WSL-only path makes the stub marker proof that the pane ran in the distro. */
 const WSL_STUB_PATH = '/usr/local/bin/golden-stub-agent'
 const WSL_STUB_AGENT_LINK = '/usr/local/bin/codex'
 
@@ -12,18 +11,23 @@ const WSL_STUB_AGENT_LINK = '/usr/local/bin/codex'
 const STAGE_SCRIPT =
   `mkdir -p /usr/local/bin && ` +
   `printf '#!/bin/sh\\necho GOLDEN_STUB_AGENT_READY\\nexec sleep 3600\\n' > ${WSL_STUB_PATH} && ` +
-  `chmod 0755 ${WSL_STUB_PATH} && ` +
-  `ln -sf ${WSL_STUB_PATH} ${WSL_STUB_AGENT_LINK}`
+  `chmod 0755 ${WSL_STUB_PATH}`
 
-const UNSTAGE_SCRIPT = `rm -f ${WSL_STUB_AGENT_LINK} ${WSL_STUB_PATH}`
+const STAGE_CODEX_LINK_IF_MISSING_SCRIPT =
+  `if [ -e ${WSL_STUB_AGENT_LINK} ] || [ -L ${WSL_STUB_AGENT_LINK} ]; then ` +
+  `printf existing; else ln -s ${WSL_STUB_PATH} ${WSL_STUB_AGENT_LINK} && printf created; fi`
+
+const REMOVE_STUB_SCRIPT = `rm -f ${WSL_STUB_PATH}`
+const REMOVE_CODEX_LINK_AND_STUB_SCRIPT = `rm -f ${WSL_STUB_AGENT_LINK} ${WSL_STUB_PATH}`
 
 // Why --exec (via buildWslExecArgs): under `--`, wsl.exe expands `$name` in
 // every argument and silently rewrites the script.
-function runInWslAsRoot(distro: string, script: string): void {
-  execFileSync('wsl.exe', ['-u', 'root', ...buildWslExecArgs(distro, ['sh', '-c', script])], {
-    stdio: 'pipe',
-    windowsHide: true
-  })
+function runInWslAsRoot(distro: string, script: string): string {
+  return execFileSync(
+    'wsl.exe',
+    ['-u', 'root', ...buildWslExecArgs(distro, ['sh', '-c', script])],
+    { encoding: 'utf8', stdio: 'pipe', windowsHide: true }
+  )
 }
 
 export async function getFirstWslDistro(page: Page): Promise<string | null> {
@@ -34,20 +38,30 @@ export async function getFirstWslDistro(page: Page): Promise<string | null> {
   return wsl.available ? (wsl.distros[0] ?? null) : null
 }
 
-/** Returns false when the distro won't take the stub (no root interop, read-only
- *  rootfs); callers skip rather than fail, since that is a host limitation. */
-export function stageWslGoldenStubAgent(distro: string): boolean {
+export type WslGoldenStubAgentStage = {
+  createdCodexLink: boolean
+}
+
+/** Returns null when the distro cannot stage the stub. */
+export function stageWslGoldenStubAgent(distro: string): WslGoldenStubAgentStage | null {
   try {
     runInWslAsRoot(distro, STAGE_SCRIPT)
-    return true
+    return {
+      createdCodexLink:
+        runInWslAsRoot(distro, STAGE_CODEX_LINK_IF_MISSING_SCRIPT).trim() === 'created'
+    }
   } catch {
-    return false
+    removeWslGoldenStubAgent(distro, { createdCodexLink: false })
+    return null
   }
 }
 
-export function removeWslGoldenStubAgent(distro: string): void {
+export function removeWslGoldenStubAgent(distro: string, stage: WslGoldenStubAgentStage): void {
   try {
-    runInWslAsRoot(distro, UNSTAGE_SCRIPT)
+    runInWslAsRoot(
+      distro,
+      stage.createdCodexLink ? REMOVE_CODEX_LINK_AND_STUB_SCRIPT : REMOVE_STUB_SCRIPT
+    )
   } catch {
     // Best-effort cleanup; a leftover stub only affects this fixture's own name.
   }