breakpilot-compliance/backend-compliance/compliance/api/agent_check/_agent_outputs.py

"""Run registered v3 specialist agents and surface their structured
AgentOutput per topic for the standardized result tabs.

Additive to the legacy B-wiring HTML (`_b18_wiring`): this does NOT
replace it — it puts a clean, typed `AgentOutput` into
`state["agent_outputs"][topic]`, which `_phase_f_persist` forwards into
the API result so the frontend can render a per-topic tab.

Phase 1 ships only impressum; the topic map extends to cookie / vendor /
… as those agents get wired (same contract, no code change here beyond
the map). Once the tabs are the source of truth, B18's v1 path retires.
"""

from __future__ import annotations

import logging

from compliance.services.specialist_agents import REGISTRY, AgentInput
from compliance.services.specialist_agents.impressum._classification import (
    scan_context_to_scope,
)

from ._sse import emit

logger = logging.getLogger(__name__)

# topic key (matches state["doc_texts"]) -> registered agent_id
_TOPIC_AGENTS: dict[str, str] = {
    "impressum": "impressum",
}

_MIN_TEXT = 100


def _derive_scope(profile_dict: dict) -> list[str]:
    """Business-scope aus dem erkannten Profil — identisch zu B18, damit
    der Tab denselben Scope sieht wie die bestehende Auswertung. Das
    Rechtsform-Gate kommt in einer späteren Phase (eigene Klassifizierung)."""
    scope: set[str] = set()
    if profile_dict.get("has_online_shop"):
        scope.add("ecommerce")
    if profile_dict.get("is_regulated_profession"):
        scope.add("regulated_profession")
    if profile_dict.get("industry") in ("insurance", "Finance", "finance"):
        scope.add("insurance")
    # §18 MStV — 3-stufig: Medienunternehmen (Verlag/Presse) = harte Pflicht;
    # nur Blog/News-Inhalte (has_editorial_content) = Graubereich → der Agent
    # wertet 'editorial_possible' als POSSIBLY_APPLICABLE (Pruef-Hinweis).
    if profile_dict.get("industry") == "media":
        scope.add("editorial")
    elif profile_dict.get("has_editorial_content"):
        scope.add("editorial_possible")
    return sorted(scope)


def doc_input_from_snapshot(snap: dict, doc_type: str) -> dict | None:
    """Baut den AgentInput für EINEN Doc-Type aus einem gespeicherten Snapshot
    (kein Re-Crawl). Pure + testbar: zieht den Text aus doc_entries, leitet den
    Scope aus scan_context + Profil ab (identisch zur Live-Auswertung) und nimmt
    site_label als company_name-Fallback. None, wenn kein/zu kurzer Text.
    """
    docs = snap.get("doc_entries") or []
    text = next((e.get("text") or e.get("content") or ""
                 for e in docs if e.get("doc_type") == doc_type), "")
    if len((text or "").strip()) < _MIN_TEXT:
        return None
    profile = snap.get("profile") or {}
    scope = sorted(
        set(scan_context_to_scope(snap.get("scan_context")))
        | set(_derive_scope(profile))
    )
    return {
        "doc_type": doc_type,
        "text": text,
        "business_scope": scope,
        "company_name": (profile.get("company_name") or snap.get("site_label") or ""),
        "origin_domain": snap.get("site_domain", ""),
    }


def impressum_input_from_snapshot(snap: dict) -> dict | None:
    """Rückwärtskompatibler Alias für den Impressum-Endpoint."""
    return doc_input_from_snapshot(snap, "impressum")


async def run_agent_outputs(state: dict) -> None:
    """Für jedes Topic mit registriertem v3-Agent + ausreichend Text:
    Agent laufen lassen, AgentOutput ablegen + als SSE topic-Event
    emittieren (Tab füllt sich progressiv)."""
    check_id = state.get("check_id", "")
    doc_texts = state.get("doc_texts") or {}
    profile_dict = state.get("profile_dict") or {}
    req = state.get("req")
    company_name = (
        (getattr(req, "company_name", None) or "")
        or (state.get("extracted_profile") or {}).get("company_name", "")
        or state.get("site_name", "")
    )
    origin_domain = (
        getattr(req, "origin_domain", None) or ""
    ) or state.get("domain", "")
    # Phase 3: die 8 Wizard-Felder (scan_context) sind der primäre
    # Scope-Treiber; das LLM-Profil ergänzt nur (v.a. regulated_profession,
    # das die 8 Felder nicht ausdrücken können).
    scan_context = getattr(req, "scan_context", None)
    scope = sorted(
        set(scan_context_to_scope(scan_context))
        | set(_derive_scope(profile_dict))
    )

    outputs: dict[str, dict] = state.get("agent_outputs") or {}
    for topic, agent_id in _TOPIC_AGENTS.items():
        text = (doc_texts.get(topic) or "").strip()
        if len(text) < _MIN_TEXT:
            continue
        agent = REGISTRY.get(agent_id)
        if agent is None:
            logger.warning("agent_outputs: agent '%s' not registered", agent_id)
            continue
        try:
            out = await agent.evaluate(AgentInput(
                doc_type=topic,
                text=text,
                business_scope=scope,
                company_name=company_name,
                origin_domain=origin_domain,
            ))
            outputs[topic] = out.model_dump(mode="json")
            emit(check_id, {"type": "topic", "topic": topic,
                            "output": outputs[topic]})
            logger.info(
                "agent_outputs[%s]: %d findings, confidence %.2f",
                topic, len(out.findings), out.confidence,
            )
        except Exception as e:  # noqa: BLE001 — best-effort, never break the run
            logger.warning("agent_outputs[%s] failed: %s", topic, e)

    if outputs:
        state["agent_outputs"] = outputs