A self-improving RLM agent for coding workflows and long-running autonomous tasks.
Rust
#!/usr/bin/env python3
"""Focused checks for the binary verifier's evidence parsing and assertions."""
import copy
import json
import unittest
from check_model_facts import MARKER, MODELS, latest_probe, verify_probe
def probe(model, count):
return {'kernel_probe_count': count,
'model_info': {'provider': 'fixture', 'id': model['id'], 'input': model['input']},
'compact_status': {'context_window': model['contextWindow'], 'scheduled': False,
'tokens': 11, 'percent': 0.1}}
class VerifierTests(unittest.TestCase):
def test_latest_tool_result_wins_over_retained_history_and_tool_arguments(self):
old, new = probe(MODELS[0], 1), probe(MODELS[1], 2)
body = {'messages': [
{'role': 'tool', 'tool_call_id': 'old', 'content': MARKER + json.dumps(old)},
{'role': 'assistant', 'content': MARKER + json.dumps(old)},
{'role': 'tool', 'tool_call_id': 'new', 'content': [{'type': 'text', 'text': MARKER + json.dumps(new)}]}]}
self.assertEqual(latest_probe(body, 'new'), new)
def test_historical_result_cannot_satisfy_missing_current_call(self):
body = {'messages': [{'role': 'tool', 'tool_call_id': 'old',
'content': MARKER + json.dumps(probe(MODELS[0], 1))}]}
with self.assertRaises(AssertionError):
latest_probe(body, 'new')
def test_stale_live_model_and_window_are_rejected(self):
live = probe(MODELS[1], 2)
verify_probe(live, MODELS[1], 2)
for member in ('model_info', 'compact_status'):
stale = copy.deepcopy(live)
stale[member] = probe(MODELS[0], 2)[member]
with self.assertRaises(AssertionError):
verify_probe(stale, MODELS[1], 2)
def test_rebuilt_kernel_is_rejected(self):
with self.assertRaises(AssertionError):
verify_probe(probe(MODELS[1], 1), MODELS[1], 2)
if __name__ == '__main__':
unittest.main()