A graph-first document format with typed inline link facets, for agent-native systems.
Status: v0.1 draft (sections 0–6). Pre-implementation.
bendscript.com is being rebuilt as a protocol spec + reference framework, not a canvas/KAG SaaS. The previous direction (SvelteKit canvas editor + KAG server on Supabase) is archived under old_scrap/v1/ and is not being extended.
The new product is:
A JSON document format (bend: URI scheme, content-addressable ids, span-addressable graph edges)
A reference parser/validator/edge-extractor (@bendscript/core, not yet written)
An LLM round-trip test harness (the load-bearing claim of v0.1)
Five reserved vocabularies that integrate with the rest of the [&] portfolio
docs/spec/README.md — BendScript Protocol v0.1 draft (sections 0–6)
old_scrap/v1/ contains the entire previous product — SvelteKit app, Supabase migrations integration, AGENTS.md, prompts, the v1.0 canvas-and-KAG spec (old_scrap/v1/docs/spec/README.md), and all build configs. It is historical; do not extend it.
npm install
npm test # run all unit tests (vitest)
npm run typecheck # tsc --noEmitReference framework: @bendscript/core (TypeScript, Node 20+).
src/types.ts — Document, Block, Span, Edge, Mark types
src/parse.ts — parse(text) → Document, normalize(doc) (link-mark expansion), parseAndNormalize
src/validate.ts — structural validator per §2/§3 (throws ValidationError with JSON path)
src/canonicalize.ts — RFC 8785 (JCS) wrapper for canonical serialization
src/hash.ts — content-addressable document id (computeDocumentId) using sha2-256 + json multicodec → CIDv1 multibase string
src/expand-marks.ts — §2.5.1 link-mark → edge expansion (idempotent, dedup'd)
src/serialize.ts — pretty-print and canonical serializers
test/round-trip.test.ts exercises the §8.1.1 strict round-trip claim against test-corpus/v0.1/*.bend.json.
§8.1.1 parser round-trip — verified. 96 tests passing as of 2026-04-27 against the full 20-document corpus:
parse(serialize(parse(d))) === parse(d) holds for all 20 corpus documents (Unicode, ZWJ emoji, deeply nested lists, all five reserved vocabularies, supersedes/tombstones, untyped links)
Document id is stable across parse/serialize cycles
Editing span text changes the id; editing meta does NOT (§5.1 trade-off works as designed)
RFC 8785 canonicalization is field-order independent
Link-mark expansion is deterministic and idempotent
Untyped link marks (no predicate) do NOT emit edges (per §2.5.1)
§8 harness mechanics — verified with mock LLMs. No paid API calls yet:
Drift metric implemented (5 components per §8.4: block_structure, edge_preservation, span_coverage, mark_fidelity, vocabulary_integrity), zero-drift on identical docs
Identity / small-edit / lossy / garbage mock models exercise each failure mode
Median + P95 aggregation per §8.5
Threshold checking emits pass / fail verdict with failing thresholds enumerated
MockModel with lossy mode confirmed: harness correctly enforces spec invariant that link-mark edges always re-expand on parse, so dropping the edges array doesn't lose link-mark-derived edges
§8.1.2 LLM round-trip — not yet run against real models. Implementing a ModelClient against OpenRouter is the next step. Estimated cost: $5–20 per gating run (180 calls = 20 docs × 3 classes × 3 models).
§14.3.1 Graphonomous adoption — not yet implemented. This is the other gating commitment for v0.1 final.
Spec uses RFC discipline (MUST / SHOULD / MAY)
Canonical document form is JSON; text projection deferred to v0.2
.bend files are not hand-authored — produced by editors that emit JSON
No host coupling: the protocol must work in TypeScript, Elixir, Rust, plain HTML