Skip to main content

Changelog

[Unreleased]​

Changes in this section are merged to main but not yet included in a tagged release.

Fixed​

  • Canonicalization preserves list literals. canonicalize → parse now round-trips list literals as lists, not as quoted strings. Previously, any list literal containing a comma or an inner double-quote (e.g. [a,b,c] or ["alpha","beta"]) was re-quoted by the canonicalizer, converting it from a list into a string after the next parse. The affected canonical fixture (plan) has been regenerated with the corrected representation.

  • SDIF AI $-column cells preserve string type through expansion. After expand_ai_doc() strips the $ suffix from column names and records their indices in Table.quoted_columns, the JSON decoder now honours that set when decoding table cells. Previously, numeric-looking strings such as HTTP status codes ("200", "404") were coerced to integers during SDIF AI → JSON conversion.

  • expand_ai_doc() added to sdif.ai. A new function that expands .sdif.ai aliases and returns a Document without invoking the canonicalizer. Callers that need semantic equivalence (not canonical form) should use expand_ai_doc() instead of sdif_from_ai(). The latter now delegates to expand_ai_doc() before canonicalizing.

Benchmark​

  • SDIF AI round-trip fidelity reaches 100% across all 20 benchmark documents after the above fixes. Canonical SDIF also remains at 100%.

1.0.0 — 2026-05-22​

Stable v1 contract​

  • Freezes @sdif 1.0 and @sdif.ai 1.0 as the supported v1 document directives.
  • Defines canonical-syntax-v1 as the reproducible source of canonical hashes.
  • Treats .sdif.ai as a derived, reversible projection of canonical SDIF.
  • Ships the Python parser, canonicalizer, validator, CLI, conformance fixtures, and shared specification.

Benchmark tracks​

  • Token efficiency, context packing, round-trip fidelity, delta compactness, retrieval accuracy, and semantic quality.
  • Evidence written to results/<track>/ on success; failed runs remain under tmp/<track>/ for diagnosis.

Ecosystem split​

  • Benchmark ownership moved to sdif-benchmarks.
  • Tree-sitter grammar ownership moved to tree-sitter-sdif.
  • Core repository focused on specification, reference implementation, conformance fixtures, and documentation.