Two-stage classifier for Japanese tax accounting. Stage 1: keyword dictionary match (14 categories × ~50 keywords). Stage 2 (= deferred to Phase 1.B): Claude API fallback. Returns 勘定科目 + 税区分 + confidence.

7-rule exclusion check for Japanese accounting. Returns excluded:true if transaction should NOT be auto-journalized. Rules: 内容不明デビット / 借入金返済 / 社保税金 / 給与支払い / 投資 / ATM出金 / 公共料金.

Import CSV from 弥生会計/freee/MoneyForward or generic format. Auto-detects source format from headers. Runs each transaction through the full classification pipeline (Stage 0 exclusion → Stage 1+2 classification → confidence routing). Returns categorized results + review queue + Markdown report. 弥生 users: export 仕訳日記帳 as UTF-8 CSV.

Generate a monthly review report. Takes transaction data (from freee API or CSV import), classifies all transactions, detects anomalies, and produces a structured Markdown report with category breakdown, anomaly alerts, and review items. Designed for tax accountants to present to clients.

List transactions (取引) from freee API for the configured company. Useful for sync + classification dogfood.

List all companies (事業所) accessible by the configured freee OAuth token. Returns company IDs + names. Used for multi-company batch processing.

Cross-SaaS reconciliation (= freee ↔ MF). Currently freee-only mode (= MF connector pending Phase 1.B). Detects duplicate fingerprints within freee.

Check if a transaction already exists in freee (= by date + amount + memo prefix). Use BEFORE register to prevent double-posting.

取引先マスタ auto-creation in freee. Fuzzy match against existing partners; if new, create. Returns partner_id.

税理士修正フィードバック。誤分類を記録し、同パターンの取引が今後来たら修正後の勘定科目を自動適用。Linksee Memory caveat layer と同等（= 永続記憶、二度と同じ誤りをしない）。修正は全社共通 or 特定会社のみに適用可能。

過去の分類パターン・修正履歴を検索。取引の摘要/取引先から過去パターンを参照し、分類結果の根拠を確認。Memory stats（pattern hit率等）の確認にも使用。

Nightly batch pipeline. Processes ALL companies accessible by the token (= multi-company batch). Pipeline per company: fetch unprocessed (status=unsettled) → Stage 0 exclusion → Stage 1+2 classify → confidence routing (high=auto, medium=auto+log, low=human_review) → aggregate summary. Currently dry-run only (write-back pending Phase 1.B).