Summary

Adds support for transaction splitting by surfacing the community fork's existing split methods as two new MCP tools. Splitting lets a single transaction be divided across multiple categories (e.g. a Target run that's part groceries, part household).

• get_transaction_splits(transaction_id) (read-only) — returns the current split legs, whether the transaction is split, and each leg's amount/category/merchant/notes.
• update_transaction_splits(transaction_id, splits) (write, idempotent) — full-replace semantics: the splits you pass become the complete set, replacing any existing ones. Pass an empty list to remove all splits and restore the single un-split transaction.

Inputs use a typed TransactionSplit model (amount + optional category_id, merchant_name, notes), translated into the API's camelCase splitData shape. Monarch validates that leg amounts sum to the parent total; its rejection payload is surfaced as a clear ValueError rather than swallowed. Both tools follow the existing patterns — api_call_with_retry, structured logging, 30s timeout, typed Pydantic structured output.

This brings the tool count from 19 → 21.

Example

Split a -100.00 transaction into groceries and household:

transaction_id="txn_123"
splits=[
    {"amount": -70.00, "category_id": "cat_groceries", "notes": "Food"},
    {"amount": -30.00, "category_id": "cat_household"},
]

Testing

• New tests/test_splits.py (8 cases): split retrieval, payload translation, merchant defaulting, split removal via empty list, and API-error surfacing.
• Added both tools to the failure-propagation matrix in test_tool_coverage.py.
• Updated the tool-count assertion (19 → 21) in test_structured_output.py.
• Refreshed CLAUDE.md tool/test counts and the fork capability inventory.

uv run python scripts/ci.py — ruff check, ruff format, mypy, and pytest all pass (206 passed, 5 skipped).