"""AI Spending dashboard — read-only aggregation domain.

PR #253 (Phase 2) rewrote this domain to a clean 4-bucket model. The
domain stitches four upstream data sources into a single finance-
facing view, with no dedup ladder required because each bucket is
backed by exactly one source table:

* SaaS AI seats — ``software_monthly_bills`` for closed periods,
  live preview from ``software_seat_allocations`` × plan price for
  the current period. Filtered to ``asset_software_catalog.category='ai'``.
* Bailian token spend — ``bailian_usage_daily`` (Aliyun BSS-derived;
  per-API-key breakdown unavailable — BSS only exposes account-level).
* Azure AI Foundry — ``cmdb_cost_facts`` filtered to Azure integrations
  whose ``service`` matches the Azure entry of the
  ``ai_spending.service_whitelist`` platform setting.
* AWS Bedrock — ``cmdb_cost_facts`` filtered to AWS integrations whose
  ``service`` matches the AWS whitelist entry.

A future ``vertex`` bucket renders when GCP AI facts are present.

Bailian's LLM is hooked up as the optional "AI summary" generator —
the domain calls qwen-max-latest with the period aggregates and
caches the bilingual JSON result in ``ai_spending_summaries``. PII
(user emails / display names) is replaced with stable hashes before
the prompt leaves the org.

Read-mostly. The writes are:
* ``ai_spending_summaries`` rows (LLM-generated; auto + admin manual)
* ``platform_settings`` rows under the ``ai_spending.*`` namespace
  (admin settings update + the daily regenerate-quota counters under
  ``ai_spending.regenerate_counter.<user>.<utc-date>``)
* ``audit_logs`` rows for every state-changing endpoint above

Cross-domain reads are tolerated under the
"reporting/coordinator-class service" carve-out documented in
PLATFORM.md — the dashboard reads from asset / cmdb / bailian /
identity / apikey, but those domains never read from ai_spending.
"""
from __future__ import annotations

from app.domains.ai_spending import models as _models  # noqa: F401

__all__ = ["_models"]