File size: 3,524 Bytes
c4e3582 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 | """Run Harbor with each task's verifier mounted separately and read-only."""
from harbor.trial.trial import Trial, ServiceVolumeConfig
from harbor.cli.main import app
from harbor.agents.terminus_2.terminus_2 import Terminus2
from harbor.models.trajectories import Step
from contextvars import ContextVar
from datetime import datetime, timezone
import json
import litellm
original_mounts = Trial._verifier_env_mounts
def verified_mounts(self, env_config):
mounts = original_mounts(self, env_config)
mounts.append(ServiceVolumeConfig(type='bind',
source=str(self.task.paths.tests_dir.resolve()),
target='/tests', read_only=True))
return mounts
Trial._verifier_env_mounts = verified_mounts
# Preserve and terminate provider refusals, rather than repeatedly asking the
# same blocked prompt until the eight-hour timeout. No prompts are rewritten,
# no fallback model is selected, and no model answer is synthesized.
last_refusal = ContextVar('last_refusal', default=None)
original_completion = litellm.acompletion
original_interaction = Terminus2._handle_llm_interaction
original_loop = Terminus2._run_agent_loop
class ProviderRefusal(Exception):
pass
async def completion_with_refusal_record(*args, **kwargs):
last_refusal.set(None)
response = await original_completion(*args, **kwargs)
choice = response['choices'][0]
message = choice['message']
if choice.get('finish_reason') == 'content_filter':
last_refusal.set(dict(model=response.get('model'), id=response.get('id'),
finish_reason='content_filter',
refusal=message.get('refusal'),
content=message.get('content') or '',
observed_at=datetime.now(timezone.utc).isoformat()))
return response
async def interaction_with_refusal_record(self, *args, **kwargs):
result = await original_interaction(self, *args, **kwargs)
refusal = last_refusal.get()
if refusal:
model = refusal['model'] or self._model_name.removeprefix('openrouter/')
self._last_response_model_name = model
self._trajectory_steps.append(Step(step_id=len(self._trajectory_steps) + 1,
timestamp=refusal['observed_at'], source='agent',
model_name=model, message=refusal['content'],
extra={'provider_refusal': refusal}))
(self.logs_dir / 'provider-refusal.json').write_text(json.dumps(refusal, indent=2) + '\n')
raise ProviderRefusal('Provider returned content_filter; no replacement answer generated')
return result
async def loop_with_refusal_stop(self, *args, **kwargs):
try:
return await original_loop(self, *args, **kwargs)
except ProviderRefusal:
# Normal task completion lets the unchanged separate grader score the
# actual artifact (normally no answer, hence zero).
return None
litellm.acompletion = completion_with_refusal_record
Terminus2._handle_llm_interaction = interaction_with_refusal_record
Terminus2._run_agent_loop = loop_with_refusal_stop
if __name__ == '__main__':
# Hub downloads may omit executable bits; the verifier mount is read-only.
from pathlib import Path
for script in (Path(__file__).resolve().parent / 'tasks').rglob('*.sh'):
script.chmod(script.stat().st_mode | 0o111)
app()
|