erpnext/specs/accounts_refactor_spec.md

Accounts / Controller Refactor — Spec

Motivation

Move ERPNext away from the deep AccountsController → SellingController/BuyingController → SalesInvoice inheritance chain and the monolithic sales_invoice.py / god-object accounts_controller.py toward composition: per-doctype services/ plus shared module-level accounts/services/. Goal is testability, readability, and factoring shared domain logic out so Sales/Purchase voucher logic is not duplicated.

Target structure

erpnext
├── controllers
│   └── transaction_controller.py        # thin lifecycle base, delegates to services
├── accounts
│   ├── general_ledger.py                # the SINK (unchanged): post / merge / round-off / reverse
│   ├── services
│   │   ├── base_gl_composer.py          # BaseGLComposer — shared GL helpers
│   │   ├── gl_validator.py              # list-level validation (functions, stateless)
│   │   ├── advances.py
│   │   ├── taxes.py
│   │   └── budget.py
│   └── doctype
│       └── sales_invoice
│           ├── sales_invoice.py         # thin: delegates to services
│           ├── services
│           │   ├── gl_composer.py       # SalesInvoiceGLComposer(BaseGLComposer)
│           │   ├── pos.py
│           │   ├── loyalty.py
│           │   ├── status.py
│           │   ├── inter_company.py
│           │   ├── fixed_assets.py
│           │   └── timesheet_billing.py
│           ├── mapper.py
│           └── api.py

GL layer — frozen design

Pipeline:

SalesInvoiceGLComposer.compose()  →  gl_entries  →  gl_validator.validate(gl_entries)  →  general_ledger.make_gl_entries()

Role	Location	Form	Responsibility
Composer (base)	accounts/services/base_gl_composer.py → BaseGLComposer	class (stateful, holds self.doc)	shared row factory + common entries
Composer (doctype)	sales_invoice/services/gl_composer.py → SalesInvoiceGLComposer(BaseGLComposer)	class	voucher-specific rows via .compose()
Validator	accounts/services/gl_validator.py	module functions (stateless)	assert the finished gl_entries list is legal to post
Sink	accounts/general_ledger.py (unchanged)	module functions	merge / round-off / post / reverse

Naming decisions (frozen)

• Chose compose over make/build — the sink already owns the verb make (make_gl_entries); compose avoids a two-makers collision.
• base_ prefix on the shared/abstract file; the concrete subclass carries the specific name, no prefix.
• Rejected: gl_map (it's a list, not a map — but it's an entrenched public param; rename to gl_entries later as its own deprecation pass), gl_processor (redundant with general_ledger.py), gl_entries.py (collides with the gl_entry doctype + the ubiquitous local var), ledger_builder (clashes with stock/payment ledger), builder/maker (generic; "maker" collides with make_gl_entries).

Bucketing accounts_controller.py

• Base composer (BaseGLComposer): get_gl_dict, get_value_in_transaction_currency, make_discount_gl_entries (+ get_amount_and_base_amount, get_tax_amounts), make_precision_loss_gl_entry, make_exchange_gain_loss_journal (+ gain_loss_journal_already_booked), set_transaction_currency_and_rate_in_gl_map. Regional hooks update_gl_dict_with_regional_fields / ..._app_based_fields stay free functions called inside get_gl_dict.
• Advances service: set_advances, get_advance_entries, clear_unallocated_advances, validate_advance_entries, set_advance_gain_or_loss, calculate_total_advance_from_ledger, set_total_advance_paid, set_advance_payment_status, delink_advance_entries, create_advance_and_reconcile, get_advance_payment_doctypes, _remove_advance_payment_ledger_entries, module funcs get_advance_journal_entries / get_advance_payment_entries.
• Validator (from general_ledger.py): validate_disabled_accounts, validate_accounting_period, validate_cwip_accounts, check_freezing_date, validate_against_pcv, validate_allowed_dimensions. (Moved in Phase 1.)
  • Balance trio stays in general_ledger.py for now (revised during Phase 1): get_debit_credit_difference / get_debit_credit_allowance / raise_debit_credit_not_equal_error. get_debit_credit_difference mutates entries (rounds debit/credit in place) and the trio is interleaved with process_debit_credit_difference → make_round_off_gle (the round-off repair run before and after balancing). It is not a standalone pre-post gate, so it can't move into a pure validate(gl_entries) without changing behavior. It travels with round-off when that moves compose-side (see below).
  • Stays in compose (do NOT move to validator): process_debit_credit_difference / make_round_off_gle — these repair balance by appending a round-off entry (mutation), not validation.
  • Stays in composer (not validator): row-level checks (right account for a row, dimension applicability) — validator only validates the finished list.
• Leave in controller: validate_company_in_accounting_dimension, validate_company (dimension validation, not GL).

Phases

Each phase is behavior-preserving, one draft PR, gated by the Phase-0 snapshot suite + bench run-tests --site test-site-ai.

Phase 0 — Safety net (first, mandatory)

Characterization tests snapshotting gl_entries output for representative transactions (SI/PI with taxes, multi-currency, advances, discounts, round-off, POS). Every later phase passes iff snapshots are byte-identical.

Phase 1 — Extract gl_validator.py (lowest risk) — DONE

Moved the 6 pure list-level validators to erpnext/accounts/services/gl_validator.py; general_ledger.py imports and calls them at the existing call sites (no behavior change). A consolidated gl_validator.validate(gl_entries) facade is deferred — the current checks run at different points (make_gl_entries / save_entries per-entry / make_reverse_gl_entries), so collapsing them into one call would alter ordering. Verified: all 12 Phase-0 snapshots byte-identical.

Phase 2 — Pilot composer on Sales Invoice only — DONE

Added BaseGLComposer (minimal: holds self.doc) and SalesInvoiceGLComposer. SI's get_gl_entries is a thin shim delegating to SalesInvoiceGLComposer(self).compose(). All 11 SI-specific row builders (make_customer/tax/item/internal_transfer/pos/loyalty/write_off/rounding GL entries, stock_delivered_but_not_billed, get_gl_entries_for_fixed_asset, get_gle_for_change_amount) moved onto the composer and operate on self.doc. The super().get_gl_entries() stock-expense call became super(SalesInvoice, doc).get_gl_entries() (MRO-faithful). Bucket-A shared helpers (get_gl_dict, make_discount_gl_entries, make_precision_loss_gl_entry, set_transaction_currency_and_rate_in_gl_map, get_tax_amounts, get_amount_and_base_amount) stay on the controller — they're still called via self.doc and only lift to BaseGLComposer once all doctypes use composers (can't move while other doctypes inherit them). Verified: 12 snapshots + 10 existing SI tests (perpetual super(), POS change, write-off, returns, fixed-asset disposal/regain, internal transfer, loyalty) all green.

Phase 3 — Second doctype: Purchase Invoice (base earns its shape) — IN PROGRESS

Added PurchaseInvoiceGLComposer (scaffolding: compose() = the moved get_gl_entries orchestration; PI.get_gl_entries is a thin shim). Decision after comparing SI and PI: keep BaseGLComposer minimal (self.doc + abstract compose). The two flows differ too much to share a template — different step order (SI tax→item, PI item→tax), different builders (SI: discount/loyalty/POS/SDBNB; PI: tax-withholding/payment/purchase-expense), and a per-doctype make_regional_gl_entries. Forcing a template would be hook-heavy and risk behavior changes. Revisit base-lifting only when a 3rd+ doctype reveals a real common shape. Verified: 12 snapshots + 6 existing PI GL tests (perpetual inventory, non-stock, return, update_stock, tax withholding, provisional) green. (PI row-builder method migration onto the composer, mirroring SI, still pending.)

Phase 4 — Roll out composer to remaining GL-posting doctypes

Payment Entry, Journal Entry, Delivery Note, Stock Entry, etc. Mechanical now; one PR per doctype (or small batches), each snapshot-gated.

Phase 5 — Extract advances.py

Move the advances cluster. After composers, because advances cross-calls the exchange-gain/loss helper now on BaseGLComposer.

Phase 6 — Extract remaining domain services from accounts_controller

taxes.py, budget.py, etc. Shrink accounts_controller to a thin lifecycle base that delegates.

Phase 7 — Split the rest of the sales_invoice.py monolith

Non-GL doctype services: pos.py, loyalty.py, status.py, inter_company.py, fixed_assets.py, timesheet_billing.py. Independent of GL work; can run parallel to 5–6.

Phase 8 — Collapse the inheritance chain

Flatten SellingController / BuyingController layers that are now pass-through. Last, because only safe once the delegated-to services exist.

Dependencies: 1→2→3 sequential; 4 and 7 can parallelize once 3 lands; 8 always last.

Cross-cutting rules

• Public signatures stay stable — keep the gl_map= param and make_gl_entries intact. The gl_map → gl_entries rename is its own deprecation pass, deferred to the end (or excluded).
• Composers are classes (stateful, per-document); sink and validator are stateless module functions.
• Every phase: behavior-preserving, snapshot + bench run-tests --site test-site-ai green before merge, draft PR.