What it actually does

Eight guarantees, in the order they matter when a payload lands on you. Internals are documented in SPEC.md.

Lossless by construction

Int64 order and entity IDs, and precise decimals, round-trip exactly — no float coercion anywhere in the pipeline.

Virtualized, chunked tree

Range rows keep a five-million-element array navigable; expansion state is restored after filtering.

Payload decoding

Zstd, Base64 (standard or URL-safe), and PostgreSQL bytea are sniffed and decoded locally, with a visible transformation trace.

Semantic compare

Identity-based array alignment by id, composite key, or auto-detection, with an editable alignment plan and per-row match labels.

Edit, undo, apply

Inline-edit primitives in the tree, or edit in a CodeMirror code view and apply with Ctrl/Cmd+S. Full undo/redo across edits.

Transport size inspector

Exact UTF-8 → Zstd → Base64 → envelope byte counts against editable budgets, so you know before the broker rejects it.

Nested JSON to linked tables

Every repeating array becomes its own sheet, joined to its parent by an id already in the data — never items/0/sku columns. The mapping you approve is a small file that produces the same columns next month, through the browser, the CLI or the MCP server.

Functions you keep

Name a script and press it over tomorrow's file, or run several as one report. Each learns which paths it reads, so a document missing them is named before the run rather than after an empty result — and the library exports as a playbook file carrying the questions, never the data.