OpenEForms

Open tools for European public procurement notices, in TypeScript.

eForms is the mandatory standard for procurement notices across the EU. Every working implementation of it is written in Java and addressed to institutions with an IT department. Germany's notice service publishes around 23,000 notices a month under a public-domain dedication, and in practice almost nobody reads them except commercial alert services that charge for access.

The obstacle is not the licence. It is the dependency.

Status

Early. @openeforms/core parses and normalises notices and is tested against a full month of production data. @openeforms/efx implements the EFX expression language and translates it to XPath. @openeforms/validate runs the SDK's Schematron rules. @openeforms/lineage groups notices into procedures and diffs corrigenda. Only the command-line interface is still to come.

Package	Purpose	State
@openeforms/core	Namespace-aware parser, normalised notice model, export ingest	working, tested
@openeforms/efx	EFX parser and XPath translator, no JVM	working, tested
@openeforms/validate	ISO Schematron validation, no JVM and no XSLT step	working, tested
@openeforms/lineage	Procedure index and notice-version diffing, with coverage accounting	working, tested
@openeforms/cli	Command-line ingest, validation and reporting	planned

Why it is built this way

Three design decisions, each forced by measurement rather than by reading the specification. The measurements are in a separate repository, openeforms-probe, which anyone can re-run.

Elements are addressed by namespace and local name, never by prefix. Seven distinct prefixes carry the same root elements in one month of German data: ns7:, ns8:, ns9:, cn:, can:, pin:, and no prefix at all. Prefix matching works on some notices and fails on others without raising anything, which is the worst kind of bug.

All three circulating customisation generations parse. eforms-de-2.1 is 59.2% of the feed, eforms-sdk-0.1 is 40.6%, eforms-sdk-1.0 is 0.2%. A reader targeting only the current German profile silently drops two fifths of the data.

Absent values stay absent. cbc:ContractFolderID, which links the notices of one procurement, is missing on 38.3% of notices and on 92.8% of those using the minimal profile. The model types it as optional and records the reason in diagnostics, so a caller can distinguish "not published" from "we failed to read it". Any lineage feature built on this data has to state its coverage rather than imply completeness.

A fourth decision lives in the ingest layer. The export returns one ETag per format, identical for every day and month, so replaying it as If-None-Match makes the server answer 304 with an empty body for data you have never fetched. fetchExport sends no conditional headers and treats an unexpected 304 as an error, never as an empty day.

Use

npm install @openeforms/core

import { parseNotice, keyOf } from "@openeforms/core";

const notice = parseNotice(await readFile("notice.xml"));

notice.kind;              // "ContractNotice"
notice.profile.family;    // "eforms-de"
notice.profile.minimal;   // false
notice.version;           // { raw: "01", number: 1, zeroPadded: true }
keyOf(notice);            // "abc@1" — same for a notice published as "1"
notice.contractFolderId;  // string | undefined, genuinely
notice.legalBasis;        // { raw: "VOB/A", code: "vob-a", scope: "below" }
notice.lots;              // [{ id, title, cpv, tenderDeadline }]
notice.diagnostics;       // what was absent or unrecognised, never thrown

Fetching is separate from parsing, so the library is usable against archives you already have:

import { fetchExport } from "@openeforms/core";

const archive = await fetchExport({ day: "2026-09-09" });  // eForms zip bytes

Develop

Requires Node 20 or newer. Node 22 runs the TypeScript sources directly.

npm install
npm test     # builds first, then runs every package's tests

npm test builds before running because @openeforms/lineage imports @openeforms/core's published entry point. Tests that need large fixtures skip themselves when those are absent, so a clean clone goes green without any downloads.

The corpus test checks the parser against a real monthly export and skips itself when the fixture is absent:

mkdir -p .fixtures
curl -o .fixtures/2026-08-eforms.zip \
  "https://oeffentlichevergabe.de/api/notice-exports?pubMonth=2026-08&format=eforms.zip"
unzip -q .fixtures/2026-08-eforms.zip -d .fixtures/2026-08
node --test packages/core/test/corpus.test.ts

It parses all 22,892 notices and asserts that the counts match those measured independently by the Python probe, which uses a different XML parser. Two implementations agreeing on that many production documents is the point; a disagreement means one of them is wrong.

Licence

MIT. The notice data is CC0 and belongs to its publisher.