test(regression): preserve cancellation evidence across Box retirement

tools/mock_llm_scenarios.json

@@ -303,7 +303,7 @@
           {
             "name": "exec",
             "arguments": {
-              "command": "test ! -e /tmp/langbot-{mock_scope}-ready && test ! -e /tmp/langbot-{mock_scope}-leak && printf SETUP_READY > /tmp/langbot-{mock_scope}-ready && cat /tmp/langbot-{mock_scope}-ready"
+              "command": "test ! -e /workspace/langbot-{mock_scope}-ready && test ! -e /workspace/langbot-{mock_scope}-leak && printf SETUP_READY > /workspace/langbot-{mock_scope}-ready && cat /workspace/langbot-{mock_scope}-ready"
             }
           }
         ]
@@ -314,7 +314,7 @@
           {
             "name": "exec",
             "arguments": {
-              "command": "sleep 8; printf LATE_SIDE_EFFECT > /tmp/langbot-{mock_scope}-leak",
+              "command": "sleep 8; printf LATE_SIDE_EFFECT > /workspace/langbot-{mock_scope}-leak",
               "timeout_sec": 20
             }
           }
@@ -333,7 +333,7 @@
           {
             "name": "exec",
             "arguments": {
-              "command": "sleep 9; if test ! -f /tmp/langbot-{mock_scope}-ready; then printf INTERRUPT_SETUP_MISSING; exit 41; fi; if test -e /tmp/langbot-{mock_scope}-leak; then printf INTERRUPT_LATE_SIDE_EFFECT; exit 42; fi; rm /tmp/langbot-{mock_scope}-ready && printf INTERRUPT_CLEANUP_OK",
+              "command": "sleep 9; if test ! -f /workspace/langbot-{mock_scope}-ready; then printf INTERRUPT_SETUP_MISSING; exit 41; fi; if test -e /workspace/langbot-{mock_scope}-leak; then printf INTERRUPT_LATE_SIDE_EFFECT; exit 42; fi; rm /workspace/langbot-{mock_scope}-ready && printf INTERRUPT_CLEANUP_OK",
               "timeout_sec": 15
             }
           }

tools/regression_execution_cases.py

@@ -128,6 +128,7 @@ def _stream(suite, agent, scope, row, *, interrupt=None, box_session_id=None, fr
                     'command', ''
                 ).startswith('sleep 8;'):
                     row['box_session_id'] = checked(row, 'exec_active', lambda: _running_box(suite, box_session_id))
+                    row['box_before_interrupt'] = _sessions(suite).get(row['box_session_id'])
                     if interrupt == 'cancel':
                         break  # Closing this public debug stream cancels its run.
             if frame.get('kind') in {'completed', 'error'}:
@@ -299,13 +300,15 @@ def run_case(suite, case, row):
                 'exec_stopped',
                 lambda: _poll(
                     suite,
-                    lambda: _sessions(suite).get(row['box_session_id'], {}).get('status') == 'idle',
+                    lambda: _sessions(suite).get(row['box_session_id'], {}).get('status') in (None, 'idle'),
                     'Interrupted Box exec remains active',
                 ),
             )
+            row['box_after_interrupt'] = _sessions(suite).get(row['box_session_id'])
             config = _agent_config(suite, box_key, 'fault_interrupt_check_model', box=True)
             _data(suite, 'PUT', f'/api/v1/agents/{agent}', {'config': config})
             recovery = _stream(suite, agent, scope, row, frame_key='recovery_frames')
+            row['box_after_recovery'] = _sessions(suite).get(row['box_session_id'])
             calls = _tool_events([frame['data'] for frame in recovery if frame.get('kind') == 'result'], 'exec')
             row['recovery_exec_results'] = calls
             try:

tools/test_regression_execution_cases.py

@@ -47,7 +47,15 @@ class ExecutionCasesTests(unittest.TestCase):
         )
 
     def run_fixture(
-        self, scenario, *, tool_ok=True, fault=True, primary_status=None, recovery_status='completed', delete_ok=True
+        self,
+        scenario,
+        *,
+        tool_ok=True,
+        fault=True,
+        primary_status=None,
+        recovery_status='completed',
+        delete_ok=True,
+        recycled=False,
     ):
         expected = {
             'execution.cancel_cleanup': 'cancelled',
@@ -92,6 +100,8 @@ class ExecutionCasesTests(unittest.TestCase):
             if method == 'GET' and path == '/api/v1/agents/agent':
                 return 404, {'code': -1}
             if path == '/api/v1/box/sessions':
+                if recycled and len(runs) == 1:
+                    return 200, {'code': 0, 'data': []}
                 return 200, {'code': 0, 'data': [{'session_id': box_session_id, 'status': 'idle'}]}
             if path.endswith('/events'):
                 run_id = path.split('/')[-2]
@@ -195,10 +205,27 @@ class ExecutionCasesTests(unittest.TestCase):
         self.assertEqual(self.row['recovery_exec_results'][0]['result']['exit_code'], 42)
         self.assertTrue(self.deleted)
 
+    def test_recycled_box_requires_the_same_persistent_file_check(self):
+        for scenario in ('execution.cancel_cleanup', 'execution.timeout_cleanup'):
+            with self.subTest(scenario=scenario):
+                with sqlite3.connect(self.ledger) as db:
+                    db.execute('DELETE FROM agent_run')
+                result = self.run_fixture(scenario, recycled=True)
+                self.assertEqual(result['result'], 'FAULT_CASE_OK')
+                self.assertIsNone(self.row['box_after_interrupt'])
+                self.assertEqual(self.row['assertions']['no_late_side_effect'], 'passed')
+                with sqlite3.connect(self.ledger) as db:
+                    db.execute('DELETE FROM agent_run')
+                with self.assertRaisesRegex(ValueError, 'INTERRUPT_LATE_SIDE_EFFECT'):
+                    self.run_fixture(scenario, recycled=True, tool_ok=False, recovery_status='failed')
+                self.assertEqual(self.row['assertions']['no_late_side_effect'], 'failed')
+
     def test_recovery_command_distinguishes_missing_setup_and_late_side_effect(self):
         scenarios = json.loads(Path(cases.__file__).with_name('mock_llm_scenarios.json').read_text())
         command = scenarios['mock-local-interrupt-check']['steps'][0]['tool_calls'][0]['arguments']['command']
-        command = command.removeprefix('sleep 9; ').replace('/tmp/langbot-{mock_scope}', str(self.root / 'fixture'))
+        command = command.removeprefix('sleep 9; ').replace(
+            '/workspace/langbot-{mock_scope}', str(self.root / 'fixture')
+        )
         ready, leaked = self.root / 'fixture-ready', self.root / 'fixture-leak'
         for setup, leak, status, output in (
             (False, False, 41, 'INTERRUPT_SETUP_MISSING'),
@@ -253,7 +280,10 @@ class ExecutionCasesTests(unittest.TestCase):
         try:
             self.assertEqual(cases._stream(self.suite, 'agent', 'scope', {})[-1]['kind'], 'completed')
             row = {}
-            with patch.object(cases, '_running_box', return_value='box'):
+            with (
+                patch.object(cases, '_running_box', return_value='box'),
+                patch.object(cases, '_sessions', return_value={'box': {'status': 'running'}}),
+            ):
                 cancelled = cases._stream(self.suite, 'agent', 'scope', row, interrupt='cancel')
             self.assertEqual(len(cancelled), 1)
             self.assertEqual(row['box_session_id'], 'box')
@@ -292,7 +322,7 @@ class ExecutionCasesTests(unittest.TestCase):
                 for scope in ('run-a', 'run-b'):
                     call = request(scope)['choices'][0]['message']['tool_calls'][0]
                     command = json.loads(call['function']['arguments'])['command']
-                    self.assertIn(f'/tmp/langbot-{scope}-ready', command)
+                    self.assertIn(f'/workspace/langbot-{scope}-ready', command)
                     self.assertNotIn('{mock_scope}', command)
                     commands.append(command)
                 self.assertNotEqual(commands[0], commands[1])