Architecture

Flipper is one Python application with a CLI coordinator (main.py) and a server-rendered FastAPI / Jinja web interface. Domain modules own parsing, pricing, inventory, eBay normalization and reconciliation, sales economics, and reports. Web routes reuse those services; templates contain no SQL or accounting formulas. CSS is static and requires no Node build.

SQLite under inventory/ is authoritative and versioned by ordered migrations. Attachments are controlled local files. External eBay data is normalized in memory and only supported explicit imports create minimized records. The dashboard's local pages do not depend on live eBay; the eBay Listings route deliberately isolates its live read.

Keep optional AI, market-data, Discord, and marketplace paths independently fallible. Preserve exact money, stable identity, transactionality, privacy boundaries, and existing valuation/scoring when working on unrelated features.

The ephemeral deals/ domain layers category/source normalization, DealOpportunity, exact expected economics, evidence/risk, and dimension-preserving comparison. The legacy PC parser and estimator adapt into that domain but retain their existing behavior. Deal analysis does not touch the inventory schema or realized-accounting services.

Working comparable research remains ephemeral. A bounded process-memory store keys records by a random browser cookie and source opportunity identity. An explicit save copies the evaluated aggregate into a schema-v11 durable research snapshot: queryable header fields plus a bounded, validated versioned JSON payload. Historical views use captured calculations and never depend on the working object or a live marketplace response. An optional inventory foreign key links history to a Q-number without making inventory depend on research or turning estimates into accounting facts.