fix(setup): fail closed when setup start is unverified

src/shared/setup-agent-sequencing-env.ts

@@ -4,3 +4,8 @@
 export const SETUP_AGENT_SEQUENCE_STARTUP_COMMAND_ENV = 'ORCA_SEQUENCED_STARTUP_COMMAND'
 export const SETUP_AGENT_SEQUENCE_STARTUP_SCRIPT_ENV = 'ORCA_SEQUENCED_STARTUP_SCRIPT'
 export const SETUP_AGENT_SEQUENCE_SETUP_SCRIPT_ENV = 'ORCA_SEQUENCED_SETUP_SCRIPT'
+
+/** Sentinel the setup half writes before running, shared by both platform gates. */
+export function setupStartedPath(markerPath: string): string {
+  return `${markerPath}.started`
+}

src/shared/setup-agent-sequencing-posix-gate.ts

@@ -1,11 +1,11 @@
-import { SETUP_AGENT_SEQUENCE_SETUP_SCRIPT_ENV, SETUP_AGENT_SEQUENCE_STARTUP_COMMAND_ENV } from './setup-agent-sequencing-env'
+import {
+  SETUP_AGENT_SEQUENCE_SETUP_SCRIPT_ENV,
+  SETUP_AGENT_SEQUENCE_STARTUP_COMMAND_ENV,
+  setupStartedPath
+} from './setup-agent-sequencing-env'
 
 /** Shell fragments for the POSIX half of the setup-to-agent gate: the setup launch that records
  *  an outcome, and the agent launch that waits for one. */
-export function setupStartedPath(markerPath: string): string {
-  return `${markerPath}.started`
-}
-
 export function buildPosixSetupCommand(bareRunnerCommand: string): string {
   const script = [
     `if [ -n "\${${SETUP_AGENT_SEQUENCE_SETUP_SCRIPT_ENV}:-}" ]; then`,
@@ -17,7 +17,11 @@ export function buildPosixSetupCommand(bareRunnerCommand: string): string {
   return `bash -lc ${quotePosixArg(script)}`
 }
 
-export function buildPosixSetupScript(setupCommand: string, markerPath: string, nonce: string): string {
+export function buildPosixSetupScript(
+  setupCommand: string,
+  markerPath: string,
+  nonce: string
+): string {
   const marker = quotePosixArg(markerPath)
   const tmp = quotePosixArg(`${markerPath}.tmp`)
   const started = quotePosixArg(setupStartedPath(markerPath))
@@ -79,11 +83,10 @@ export function buildPosixStartupScript(
     'echo "Setup started; waiting for it to finish." >&2;',
     'fi;',
     'if [ "$setup_started" = "0" ] && [ "$SECONDS" -ge "$start_deadline" ]; then',
-    // Why: nothing reported starting, so we cannot claim setup failed — only that this terminal
-    // has no way to observe it. Starting the agent unsequenced beats a terminal that waits out
-    // the whole bound for an outcome no one is going to record.
-    `echo "Setup never reported starting within ${grace}s, so this terminal cannot tell whether it ran. Starting the agent without waiting for setup." >&2;`,
-    `${launchAgent}`,
+    // Why: no start evidence is not success evidence. Stop here rather than allowing an
+    // unsequenced agent to run against an environment whose setup outcome is unknown.
+    `echo "Setup never reported starting within ${grace}s; the agent was not started because setup could not be verified. Open the Setup tab and retry once setup is running." >&2;`,
+    'exit 125;',
     'fi;',
     'if [ "$SECONDS" -ge "$deadline" ]; then',
     `echo "Timed out waiting for setup before starting agent. Waited ${timeout}s without a result; the agent was not started. Open the Setup tab for its output." >&2;`,
@@ -154,4 +157,3 @@ function quotePosixArg(value: string): string {
   }
   return `'${value.replace(/'/g, `'\\''`)}'`
 }
-

src/shared/setup-agent-sequencing-windows-gate.ts

@@ -1,6 +1,8 @@
 import { encodePowerShellCommand } from './powershell-command-encoding'
-import { SETUP_AGENT_SEQUENCE_STARTUP_COMMAND_ENV } from './setup-agent-sequencing-env'
-import { setupStartedPath } from './setup-agent-sequencing-posix-gate'
+import {
+  SETUP_AGENT_SEQUENCE_STARTUP_COMMAND_ENV,
+  setupStartedPath
+} from './setup-agent-sequencing-env'
 
 /** Shell fragments for the native-Windows half of the setup-to-agent gate. Both sides run under
  *  PowerShell so the marker write and the bounded poll stay parseable without a batch label loop. */
@@ -97,8 +99,8 @@ export function buildWindowsStartupCommand(
     '    [Console]::Error.WriteLine("Setup started; waiting for it to finish.")',
     '  }',
     '  if (-not $setupStarted -and (Get-Date) -ge $startDeadline) {',
-    `    [Console]::Error.WriteLine("Setup never reported starting within ${grace}s, so this terminal cannot tell whether it ran. Starting the agent without waiting for setup.")`,
-    '    & $launchAgent',
+    `    [Console]::Error.WriteLine("Setup never reported starting within ${grace}s; the agent was not started because setup could not be verified. Open the Setup tab and retry once setup is running.")`,
+    '    exit 125',
     '  }',
     '  if ((Get-Date) -ge $deadline) {',
     `    [Console]::Error.WriteLine("Timed out waiting for setup before starting agent. Waited ${timeout}s without a result; the agent was not started. Open the Setup tab for its output.")`,
@@ -123,4 +125,3 @@ function encodePowerShellInvocation(script: string): string {
 function quotePowerShellString(value: string): string {
   return `'${value.replace(/'/g, "''")}'`
 }
-

src/shared/setup-agent-sequencing.test.ts

@@ -776,7 +776,7 @@ describe('setup outcome recording', () => {
   )
 
   it.skipIf(process.platform === 'win32')(
-    'starts the agent unsequenced when setup is never started at all',
+    'fails closed when setup is never started at all',
     async () => {
       const tempDir = makeTempDir()
       const runnerScriptPath = join(tempDir, 'setup-runner.sh')
@@ -801,10 +801,13 @@ describe('setup outcome recording', () => {
         })
       )
 
-      expect(startupExit.code).toBe(0)
-      expect(startupExit.stderr).toContain('Setup never reported starting within 1s')
-      expect(readFileSync(logPath, 'utf8')).toBe('agent-start\n')
-    }
+      expect(startupExit.code).toBe(125)
+      expect(startupExit.stderr).toContain(
+        'Setup never reported starting within 1s; the agent was not started'
+      )
+      expect(readIfExists(logPath)).toBe('')
+    },
+    15_000
   )
 
   it.skipIf(process.platform === 'win32')(
@@ -836,7 +839,8 @@ describe('setup outcome recording', () => {
         })
       )
       // The wrapped launch record was dropped on the way to the setup terminal, so the gated
-      // script is absent from its env; the fallback must still run setup.
+      // script is absent from its env. Setup may still run, but the agent must fail closed because
+      // no authoritative setup result was recorded.
       const setupExit = await waitForExit(
         spawn('bash', ['-lc', commands.setupCommand], { stdio: 'pipe', env: { ...process.env } })
       )
@@ -844,10 +848,13 @@ describe('setup outcome recording', () => {
 
       expect(setupExit.code).toBe(0)
       expect(readFileSync(setupLogPath, 'utf8')).toBe('setup-ran\n')
-      expect(startupExit.code).toBe(0)
-      expect(startupExit.stderr).toContain('Setup never reported starting within 2s')
-      expect(readFileSync(logPath, 'utf8')).toBe('agent-start\n')
-    }
+      expect(startupExit.code).toBe(125)
+      expect(startupExit.stderr).toContain(
+        'Setup never reported starting within 2s; the agent was not started'
+      )
+      expect(readIfExists(logPath)).toBe('')
+    },
+    15_000
   )
 
   it.skipIf(process.platform === 'win32')(

src/shared/setup-agent-sequencing.ts

@@ -29,9 +29,8 @@ import {
 const DEFAULT_WAIT_TIMEOUT_SECONDS = 30 * 60
 // Why: the setup terminal is spawned in the same host operation as the agent terminal and writes
 // its start sentinel before running a single line of the script, so anything past a slow shell
-// profile plus an SSH round trip means nobody is going to run setup at all. Expiring does not
-// fail the launch — it starts the agent unsequenced and says so — so a false positive costs a
-// warning line, never a dead terminal.
+// profile plus an SSH round trip means nobody is going to run setup at all. Expiring fails closed
+// with a clear retry message; an unknown setup outcome must never authorize an agent launch.
 const SETUP_START_GRACE_SECONDS = 45
 // Why: a silent terminal reads as a hang, so the wait reports itself on a human interval.
 const WAIT_PROGRESS_INTERVAL_SECONDS = 15