just-dna-format

The declarative schema + integrity contract for just-dna annotation modules: both the authored input DSL and the compiled manifest.json. Dependency-light (Pydantic + cryptography, for Ed25519 signature verification) so it can be shared as the single source of truth by:

• just-dna-compiler — the reference transform (spec → parquet artifact + manifest);
• just-dna-pipelines — emits manifest.json when it compiles a module;
• just-dna-marketplace — indexes, serves, and verifies those manifests;
• just-dna-agents — authoring support (renders reference.authoring_reference()).

Keeping the contract in one small package prevents the sides from drifting.

What's here

Usage

from just_dna_format.manifest import ModuleManifest, read_manifest
from just_dna_format.integrity import build_artifact, verify_manifest

# Compiler side: hash outputs and record the digest.
artifact = build_artifact(output_dir, ["weights.parquet", "annotations.parquet", "studies.parquet"])

# Downloader side: verify before installing (raises IntegrityError on any mismatch).
manifest = read_manifest(module_dir / "manifest.json")
verify_manifest(module_dir, manifest)                            # a registry install
verify_manifest(module_dir, manifest, require_marketplace=False)  # a locally-compiled module

Those two calls are different policies, not a strict and a lax mode — pick per install route rather than picking one. The default additionally requires compiled_by == "marketplace-server", so it rejects any module compiled anywhere else, this project's own compiler included (it leaves compiled_by null on purpose). Both check every file hash and the digest in full. compiled_by is an unsigned string written by whoever compiled the module, so it proves nothing on its own. The check that carries weight is a signature verified against a key you pinned in advance: verify_signature(digest, signature, trusted_public_key=...).

All hashes are SHA-256, lowercase hex, prefixed sha256:. The artifact.digest is a Merkle-style root over the canonical file listing — verifying it verifies the whole set and is the version's byte identity: these bytes, from this compiler. The content identity is content_signature, over the authored rows and independent of the reference that resolved them, so a rebuild that moves the digest while leaving the signature still means the data did not change.

Develop

uv sync
uv run pytest -vvv