"""Transport-only cron execution; output/delivery remain with the firing scheduler.""" import asyncio import hashlib import json import uuid class CronExecutionUnknown(RuntimeError): """The owner may have accepted this fire; never book it as failed or re-fire.""" def journal_path(job_id, request_id): from hermes_constants import get_hermes_home key = hashlib.sha256(json.dumps([job_id, request_id]).encode()).hexdigest() return get_hermes_home() / 'cron' / 'admissions' / (key + '.json') def _pause_identity(paused): """The exact pause this recovery uncertainty caused, so only that pause is ever lifted.""" return {'paused_at': paused.get('paused_at'), 'reason': paused.get('paused_reason')} if paused else True def replayable_completions(job_id, completions): """The recorded completion ids whose admission journal still exists: only journal recovery can replay a completion into ``mark_job_run``, so an id without one is a dead tombstone.""" return [eid for eid in completions or () if journal_path(job_id, eid).exists()] def pause_for_unknown(job_id, execution_id, reason): """The firer pauses the job for this fire's unverified admission and records that exact pause on the fire's journal, so the receipt that later proves the run terminal can lift it.""" from cron.jobs import pause_job from utils import atomic_json_write paused = pause_job(job_id, reason=reason) journal = journal_path(job_id, execution_id) if paused and journal.exists(): record = json.loads(journal.read_text(encoding='utf-8-sig')) record['pause_recorded'] = _pause_identity(paused) atomic_json_write(journal, record, mode=0o600, fsync_dir=True) def _lift_pause(job_id, pause): """A verified terminal receipt resolved the uncertainty that paused the schedule: resume it iff the job is still held by exactly that pause (an operator's own pause/resume is never touched).""" from cron.jobs import get_job, resume_job job = get_job(job_id) if isinstance(pause, dict) else None if (job is not None and job.get('state') == 'paused' and (job.get('paused_at'), job.get('paused_reason')) == (pause['paused_at'], pause['reason'])): try: resume_job(job_id) except ValueError: # a one-shot whose instant has passed stays paused for the operator pass def run_canonical_job(job, *, extra_prompt=None, cancel_event=None, execution_id=None): from gateway.session_cron import owner_for_home, operation from hermes_cli.gateway_client import connect_gateway from hermes_constants import get_hermes_home from utils import atomic_json_write request_id = execution_id or job.get('execution_id') if not request_id: request_id = job['execution_id'] = uuid.uuid4().hex params = {'job_id': job['id'], 'request_id': request_id, 'extra_prompt': extra_prompt} root = get_hermes_home() / 'cron' / 'admissions' journal = journal_path(params['job_id'], params['request_id']) record = {'params': params, 'receipt': None} if journal.exists(): record = json.loads(journal.read_text(encoding='utf-8-sig')) if record['params'] != params: raise CronExecutionUnknown('cron admission identity conflict') attempted = journal.exists() owner = owner_for_home(get_hermes_home()) def save(): root.mkdir(parents=True, exist_ok=True) atomic_json_write(journal, record, mode=0o600, fsync_dir=True) async def observe(call): nonlocal attempted attempted = True save() try: receipt = await call('submit', params) except Exception as exc: from hermes_cli.gateway_client import GatewayRPCError # A returned owner exception is definite; a transport disconnect is not. Even an # owner error after commit must retain the admitted identity, so verify its absence. if owner is not None or isinstance(exc, GatewayRPCError): state = await call('recover', params) if state['status'] == 'missing': error = f'{type(exc).__name__}: {exc}' result = [False, f'# Cron Job: {job["id"]} (FAILED)\n\n{error}\n', '', error] record['refusal'] = result record['job'] = dict(job) save() return tuple(result) if state['status'] not in {'queued', 'started', 'terminal'}: raise # Admitted before the answer was lost: observe it through the recovered receipt. receipt = {'session_id': state['session_id'], 'admission_id': state['admission_id']} else: raise record['receipt'] = receipt save() delay, progress = .1, None while True: if cancel_event is not None and cancel_event.is_set(): await call('cancel', receipt) state = await call('status', receipt) if state.get('progress_at') not in (None, progress): # The owner's agent is still working: refresh OUR ledger row (the owner's stamp is # fenced to the row's owner), once per owner stamp, so the stale-claim sweep measures # the run's silence. A wedged owner stops stamping and the sweep still fires. from cron.executions import touch_execution_progress progress = state['progress_at'] await asyncio.to_thread(touch_execution_progress, request_id) if state['status'] == 'terminal': job.update(state.get('job_flags') or {}) return tuple(state['result']) if state['status'] == 'unknown': raise CronExecutionUnknown('unknown_execution: cron admission was not replayed') await asyncio.sleep(delay) delay = min(delay * 2, 2.0) async def remote(): async with connect_gateway() as client: return await observe(lambda op, data: client.rpc('cron.' + op, **data)) try: if owner is not None: authority, loop = owner try: current = asyncio.get_running_loop() except RuntimeError: current = None if current is loop: raise RuntimeError('cron synchronous execution must run off the owner event loop') return asyncio.run_coroutine_threadsafe( observe(lambda op, data: operation(authority, op, data)), loop).result() return asyncio.run(remote()) except Exception as exc: if attempted: raise CronExecutionUnknown(f'Cron execution unverified; reconcile {journal}: {exc}') from exc error = f'{type(exc).__name__}: {exc}' return False, f'# Cron Job: {job["id"]} (FAILED)\n\n{error}\n', '', error def reconcile_pending(*, allow_connect=True): """Observe prepared fires; never submit missing or interrupted work. With allow_connect=False, a headless tick leaves durable receipts for the next live owner instead of ensuring or spawning a gateway. """ from gateway.session_cron import owner_for_home, operation from hermes_constants import get_hermes_home from hermes_cli.gateway_client import connect_gateway from cron.executions import FIRER_FINISHED_STATES, execution_owner_live, get_execution from cron.jobs import get_job, pause_job from cron.scheduler import (_RunDelivery, _FireOwnership, _apply_agent_failure_marker, _save_compose_deliver) from cron.scheduler_bookkeeping import fail_empty_response, finish_completed_run from utils import atomic_json_write import logging root = get_hermes_home() / 'cron' / 'admissions' for journal in sorted(root.glob('*.json')): try: record = json.loads(journal.read_text(encoding='utf-8-sig')) params = record['params'] if journal != journal_path(params['job_id'], params['request_id']): raise ValueError('cron journal identity conflict') if execution_owner_live(params['request_id'], params['job_id']): # Its firer (this process included) still owns the row and its tail: re-saving the # output or re-marking the job on every tick would duplicate its side effects. continue row = get_execution(params['request_id']) if row is not None and row['job_id'] == params['job_id'] and row['status'] in FIRER_FINISHED_STATES: # The firer booked this attempt's terminal outcome itself (ownership lost, shutdown # interrupt, a crashed tail): the receipt is not a result to save or deliver again. journal.unlink(missing_ok=True) continue if 'refusal' in record: state = {'status': 'terminal', 'result': record['refusal'], 'job': record['job']} else: owner = owner_for_home(get_hermes_home()) if owner is None and not allow_connect: continue async def observe(): if owner is not None: return await operation(owner[0], 'recover', params) async with connect_gateway() as client: return await client.rpc('cron.recover', **params) if owner is not None: state = asyncio.run_coroutine_threadsafe(observe(), owner[1]).result(timeout=20) else: state = asyncio.run(observe()) if state['status'] == 'missing': # The owner holds no admission for this fire: a definite non-admission, booked as # the same failed run the in-submit refusal books. Only ``unknown`` pauses. current = get_job(params['job_id']) if current is None: journal.unlink(missing_ok=True) continue error = 'Cron admission missing: the owner never admitted this fire' state = {'status': 'terminal', 'job': current, 'result': [False, f'# Cron Job: {params["job_id"]} (FAILED)\n\n{error}\n', '', error]} if state['status'] != 'terminal': if state['status'] == 'unknown' and not record.get('pause_recorded'): reason = 'Canonical cron unknown; no automatic re-execution' record['pause_recorded'] = _pause_identity(pause_job(params['job_id'], reason=reason)) atomic_json_write(journal, record, mode=0o600, fsync_dir=True) continue success, output, answer, error = state['result'] job = dict(state['job'], **(state.get('job_flags') or {})) job['execution_id'] = params['request_id'] # The same result verdicts as the ordinary tail: an agent-declared [CRON_FAILURE] and a # blank answer are failures, whichever process books them. success, error, declared = _apply_agent_failure_marker(job, success, error, answer) delivery = _RunDelivery(job, success, error, agent_declared=declared) # The canonical receipt supplies the result; a dead scheduler's fire claim does not # authorize replay and must not prevent finishing its durable bookkeeping. fence = _FireOwnership(dict(job, fire_claim=None)) _save_compose_deliver(delivery, fence, answer, output, adapters=None, loop=None, verbose=False, execution_token=None) fail_empty_response(delivery, answer) if finish_completed_run(delivery, None, params['request_id'], recovered=True): _lift_pause(params['job_id'], record.get('pause_recorded')) except Exception: logging.getLogger(__name__).warning('Cron receipt recovery deferred: %s', journal, exc_info=True)