cluster-knaster/core/src/knaster/system.py

"""Verdrahtung aller Komponenten (einmal pro Prozess, in Tests mit Fakes)."""

from __future__ import annotations

import json
from collections.abc import Callable
from dataclasses import dataclass
from datetime import datetime
from pathlib import Path

from .canary import CanaryProber
from .classifier import EmbeddingClassifier
from .embeddings import Embedder, LlamaCppReranker, OpenAIEmbedder, Reranker
from .executor import Executor
from .healer.drivers import CompositeInfra, InfraDriver, Notifier, PortainerDriver, ProxmoxDriver, WebhookNotifier
from .healer.engine import Healer, HealerConfig
from .lifecycle import JobRunner, Lifecycle, LifecycleConfig, LoggingJobRunner, WebhookJobRunner
from .llm import ChatClient, OpenAIChatClient
from .maintenance.actions import ActionExecutor
from .maintenance.agent import IntentMapper, LLMIntentMapper, MaintenanceAgent, MaintenanceConfig
from .metrics import Metrics
from .ontology import Ontology, OntologyFinding, build_snapshot, check, fetch_workflows
from .policy import Policy
from .registry import Registry
from .router import Router
from .runner import Runner
from .settings import Settings
from .skills import SkillLibrary
from .store import Store
from .util import now
from .verifier import Verifier


@dataclass
class System:
    settings: Settings
    store: Store
    registry: Registry
    library: SkillLibrary
    embedder: Embedder
    chat: ChatClient
    router: Router
    executor: Executor
    verifier: Verifier
    runner: Runner
    policy: Policy
    lifecycle: Lifecycle
    actions: ActionExecutor
    maintenance: MaintenanceAgent
    healer: Healer
    metrics: Metrics
    prober: CanaryProber


def _load_classifiers(directory: Path) -> dict[str, EmbeddingClassifier]:
    if not directory.is_dir():
        return {}
    return {p.stem: EmbeddingClassifier.load(p) for p in sorted(directory.glob("*.json"))}


def _optional(path: Path, loader: Callable, default):  # type: ignore[no-untyped-def]
    return loader(path) if path.exists() else default


def ontology_source_for(settings: Settings) -> Callable[[], list[dict]] | None:
    """Woher die n8n-Workflows für den Ontologie-Abgleich kommen: Datei (offline) oder n8n-API (nur lesend)."""
    if settings.ontology_snapshot:
        path = Path(settings.ontology_snapshot)

        def from_file() -> list[dict]:
            data = json.loads(path.read_text(encoding="utf-8"))
            return data.get("data", data) if isinstance(data, dict) else data

        return from_file
    if settings.n8n_host and settings.n8n_api_key:
        return lambda: fetch_workflows(settings.n8n_host, settings.n8n_api_key, timeout=settings.request_timeout)
    return None


def ontology_checker(settings: Settings, store: Store,
                     source: Callable[[], list[dict]] | None) -> Callable[[], list[OntologyFinding]] | None:
    source = source or ontology_source_for(settings)
    if source is None or not settings.ontology_path.exists():
        return None

    def run() -> list[OntologyFinding]:
        onto = Ontology.from_file(settings.ontology_path)  # bei jedem Lauf neu: Änderungen im Repo zählen sofort
        decided = {d["key"] for d in store.list_ontology_decisions()}
        return check(onto, build_snapshot(source()), decided=decided)

    return run


def build_system(
    settings: Settings,
    *,
    embedder: Embedder | None = None,
    chat: ChatClient | None = None,
    reranker: Reranker | None = None,
    infra: InfraDriver | None = None,
    notifier: Notifier | None = None,
    jobs: JobRunner | None = None,
    intent_mapper: IntentMapper | None = None,
    ontology_source: Callable[[], list[dict]] | None = None,
    clock: Callable[[], float] = now,
    wallclock: Callable[[], datetime] = datetime.now,
) -> System:
    if settings.db_url.startswith("sqlite:///") and not settings.db_url.startswith("sqlite:///:"):
        Path(settings.db_url.removeprefix("sqlite:///")).parent.mkdir(parents=True, exist_ok=True)
    store = Store(settings.db_url)
    registry = Registry.from_file(settings.site_dir / "registry.yaml", store)
    library = SkillLibrary.from_dir(settings.skills_dir, store)

    chat = chat or OpenAIChatClient(settings.litellm_base_url, settings.litellm_api_key, settings.request_timeout)
    embedder = embedder or OpenAIEmbedder(settings.litellm_base_url, settings.embed_model, settings.litellm_api_key)
    if reranker is None and settings.rerank_url:
        reranker = LlamaCppReranker(settings.rerank_url, settings.rerank_model)

    cfg = settings.config_dir
    policy = _optional(cfg / "policy.yaml", Policy.from_file, Policy())
    healer_cfg = _optional(cfg / "healer-rules.yaml", HealerConfig.from_file, HealerConfig())
    maint_cfg = _optional(cfg / "maintenance.yaml", MaintenanceConfig.from_file, MaintenanceConfig())
    life_cfg = _optional(cfg / "lifecycle.yaml", LifecycleConfig.from_file, LifecycleConfig())

    metrics = Metrics()
    router = Router(registry, library, embedder, reranker)
    executor = Executor(library, chat, embedder, _load_classifiers(cfg / "classifiers"))
    verifier = Verifier(registry, chat)
    runner = Runner(router, executor, verifier, library, store, metrics)

    if jobs is None:
        use_webhook = settings.job_webhook_url and not settings.dry_run
        jobs = WebhookJobRunner(settings.job_webhook_url, settings.webhook_auth) if use_webhook else LoggingJobRunner()
    lifecycle = Lifecycle(store, registry, jobs, life_cfg)

    if infra is None and not settings.dry_run:
        portainer = PortainerDriver(settings.portainer_host, settings.portainer_key) \
            if settings.portainer_host and settings.portainer_key else None
        proxmox = ProxmoxDriver(settings.proxmox_url, settings.proxmox_token_id, settings.proxmox_token_secret) \
            if settings.proxmox_url and settings.proxmox_token_id else None
        infra = CompositeInfra(portainer, proxmox)
    if notifier is None and settings.notify_webhook_url:
        notifier = WebhookNotifier(settings.notify_webhook_url, settings.webhook_auth)

    actions = ActionExecutor(store=store, registry=registry, library=library, lifecycle=lifecycle, infra=infra,
                             notifier=notifier, dry_run=settings.dry_run,
                             enabled_infra_actions=set(healer_cfg.enabled_actions), on_skills_changed=router.refresh)
    if intent_mapper is None and registry.has_slot("tier1-general"):
        intent_mapper = LLMIntentMapper(chat, registry.slot("tier1-general").spec.litellm_model)
    maintenance = MaintenanceAgent(store=store, registry=registry, library=library, index=router.index,
                                   policy=policy, actions=actions, lifecycle=lifecycle, config=maint_cfg,
                                   intent_mapper=intent_mapper,
                                   ontology_check=ontology_checker(settings, store, ontology_source),
                                   clock=clock, wallclock=wallclock)
    healer = Healer(healer_cfg, store, actions, policy, propose=maintenance.propose, clock=clock)
    prober = CanaryProber(registry, chat, embedder_for={settings.embed_model: embedder},
                          metrics_registry=metrics.registry)
    return System(settings, store, registry, library, embedder, chat, router, executor, verifier, runner, policy,
                  lifecycle, actions, maintenance, healer, metrics, prober)