October 25, 2026/ai /osswin

curt-format v1: one file, one conversation, kept whole

Spec v1 is frozen, [email protected] is on npm under MIT, and 27 conformance vectors run in CI. The format behind Curtis is now a standard with receipts — and zero outside consumers. Here's the pitch and the score.

Every AI app holds your conversations in its own schema, behind an export button that flattens them. Images downgraded, tool calls dropped, structure mangled. Your work product, stored in a shape you’ll never see, at the mercy of whoever built the export.

I keep writing that sentence. Porter was the answer for reading hostile exports. Curtis made it the storage format. As of this week the format itself is a standard you can build on without asking me for anything.

What shipped

.curt is one UTF-8 JSON file per conversation. Magic string, version, generator, extension list, then the conversation: id, title, timestamps, messages with roles and content. Base64 images ride in content, verbatim. Tool calls ride in content, verbatim. Anything your app adds that the spec doesn’t know about survives a re-save untouched — that’s normative, not a courtesy.

Three artifacts:

Spec v1, frozen. The core doesn’t move. New capability means a named extension in the registry — versioned, optional, rejected cleanly by readers that don’t know it. Never an edit to v1.
[email protected] on npm. TypeScript, zero dependencies, MIT. serializeCurt, parseCurt, curtFilename. Serialize, parse, round-trip — the whole install.
27 conformance vectors in CI. L1 CORE through L4 per-extension, plain fixtures. “Compatible” is a claim you can run, not an adjective on a landing page.

Two rules

One: the core is frozen. Not “stable”, not “mostly settled” — frozen. Expansion happens in the registry or at v2, never by editing v1 under people. Formats die by accretion: every “small” spec edit breaks a reader somewhere, and the readers find out last. Need something the core lacks? Add the field; unknown fields ride along on re-save. It catches on, it graduates to a named extension. If something genuinely can’t be an extension, that’s v2 — new version number, and every v1 archive on earth still parses.

Two: an extension goes draft → stable only when two independent implementations round-trip it. Six registry entries today, one author. Nothing is stable that I didn’t prove against myself — that’s the rule working, not an oversight. consensus-receipt stays draft until someone else implements it.

Both rules protect the reader, not my ego. Rigid where it protects you — the core — and fast where it serves you: need a field today, add it; need it named, open a registry PR and I’ll move on it same week. BDFL means decisions happen, not that the door is shut.

The score

Zero outside consumers. Six registry entries, one author. A spec, a library, a test suite, a site — all built in a vacuum, and I’m not going to dress that up. It’s the current score, it’s public in capabilities.json, and the first outside implementation sets the tone for everything after it.

If you build software where conversations matter — a chat app, an agent runtime, an archive tool — the move is an afternoon. npm install curt-format, round-trip a file, run the vectors. The adopt page has the path, the product page is here, and when you ship, you PR against the registry.

The format is free. The file is yours.