🥮 mooncakes.io
    Sign in

    json-schema

    Compile-once JSON Schema 2020-12 validator for MoonBit: exact decimal arithmetic, precise pattern semantics, no silent schema errors, no hidden I/O.

    json-schema
    validation
    jsonschema
    schema
    Download zip
    Author
    Version
    0.1.5
    License
    Apache-2.0
    Last updated
    12 hours ago
    Downloads
    2K

    #colmugx/json-schema

    A compile-once JSON Schema validator for MoonBit, built around precise 2020-12 validation scenarios rather than permissive approximations.

    Status: development milestone, not a complete 2020-12 implementation. Static and dynamic references, resource identifiers and anchors, and unevaluatedProperties/unevaluatedItems are supported. Cross-resource references use an explicit preloaded registry: the compiler and validator never fetch URIs. Custom dialects can combine recognized standard vocabularies and enforce their metaschemas; unknown required vocabularies fail compilation. This is not a general plugin API for arbitrary vocabulary execution. Current profile boundaries and reproduced evidence are maintained in the conformance contract.

    #Use it for

    • Closed tool/request argument objects (additionalProperties: false).
    • Structured result validation, tuples, array membership and dependencies.
    • Exact decimal multipleOf, integer checks and numeric bounds.
    • Unicode code-point string lengths and an explicitly supported pattern profile.
    • Reusing one owned compiled schema for many instances.

    The module is maintained as an independent repository and has no dependency on MCP. Run the commands below from this module's root, including when it is checked out as a workspace member. Runners use moon run instead of assuming a fixed binary output path inside an enclosing workspace.

    #Quick start

    Import "colmugx/json-schema" in your package. Parse numbers through this library when their original decimal values matter:

    ///|
    let document = @json-schema.parse(
      "{\"type\":\"object\",\"required\":[\"amount\"],\"properties\":{\"amount\":{\"type\":\"number\",\"multipleOf\":0.1}},\"additionalProperties\":false}",
    )

    ///|
    let schema = @json-schema.Schema::compile(document)

    ///|
    let instance = @json-schema.parse("{\"amount\":0.3}")

    ///|
    let valid = schema.valid(instance)

    ///|
    let failures = schema.faults(instance)

    These operations can raise errors; propagate or catch them at your application's boundary. A malformed/unsupported schema is not a valid compiled schema. A malformed numeric node or exhausted traversal budget is not an ordinary false validation result.

    #Choose the result you need

    • schema.valid(instance) short-circuits assertion failures and does not build a diagnostic array.
    • schema.faults(instance) returns failures with instance_path, schema_path, keyword and message. Paths use JSON Pointer escaping.
    • Both methods check the instance's numeric representations and traversal budget. Failed speculative anyOf, oneOf, not and if branches do not leak errors into a successful result.

    #Preload referenced resources

    Pass external documents as a URI-to-schema map. base_uri is the root document's retrieval URI; $id establishes canonical resource identities and resolves relative references. An embedded $id starts a new resource, with its own anchors. Registry keys and base_uri must be absolute URIs; a relative $id requires an absolute base. Resource namespaces are registered when referenced, so unrelated legacy-dialect documents can coexist in the registry without being compiled.

    ///|
    test "preloaded resource references" {
      let resources : Map[String, Json] = Map::from_array([
        (
          "https://example.com/schemas/positive",
          { "type": "integer", "minimum": 1 },
        ),
      ])
      let schema = @json-schema.Schema::compile(
        { "$ref": "positive" },
        base_uri="https://example.com/schemas/request",
        resources~,
      )
      assert_true(schema.valid(Json::number(2)))
      assert_false(schema.valid(Json::number(0)))
    }

    Missing resources or anchors and conflicting identities are compile errors, not network requests or ignored constraints. The compiled schema owns its plan; mutating the registry or source documents afterwards does not change validation. $dynamicRef uses the evaluation's dynamic resource scope rather than always following a fixed target. unevaluated* observes successful evaluation marks, including marks from references and all successful anyOf branches.

    Faults retain document-relative schema_path JSON Pointers. schema_uri, when available, identifies the source document containing that pointer, so failures from different documents can be distinguished without changing pointer syntax.

    #Preserve number precision

    A floating-point value cannot recover digits lost before validation. This library's parser keeps number lexemes in Json::Number's repr field; numeric predicates use integer coefficients and decimal scales, not an epsilon or a rounded decimal quotient. Core-parsed or programmatically constructed numbers without a lexeme use the available finite Double representation instead.

    Parsing and arithmetic support are separate: the parser retains valid JSON number literals even when their exponent exceeds the evaluator's supported range. Exact arithmetic currently bounds the exponent and decimal scale to ±100000; schema compilation or evaluation outside that range raises a typed error rather than rounding a value or returning an ordinary validation failure. This lets protocol callers safely round-trip such literals without claiming that the validator can execute constraints over every possible decimal scale.

    Do not construct inconsistent Json::number(value, repr=...) nodes: numeric representations are checked, not silently discarded.

    #Supported now

    The compiler/evaluator covers boolean schemas, type, const, enum, numeric bounds and multipleOf, string lengths and patterns, object properties/names/ dependencies, 2020-12 prefixItems/items, array counts/contains/uniqueItems, allOf/anyOf/oneOf/not/if/then/else, static $ref (including $defs, forward references, recursive schemas and preloaded external resources), $id, $anchor, $dynamicAnchor/$dynamicRef, and unevaluatedProperties/unevaluatedItems.

    Patterns use moonbitlang/regexp@0.3.6 with a small explicit compatibility layer. The adapter preserves ECMA shorthand, whitespace/dot handling, General_Category properties and leading lookahead chains. Backreferences currently support only the exact ^(body)\\1$ profile; other unsupported syntax raises a typed error, never a silent non-match. See the compatibility profile and upstream review checklist.

    format and content keywords are annotations, not format/content assertions. Unknown extension keywords are annotations. Known malformed supported keywords are compile errors. An unimplemented known semantic keyword is an explicit unsupported error, not an ignored assertion.

    #Evidence and development

    The completion-roadmap baseline was reproduced on actual native, JS and Wasm execution: 1301/1301 standard cases, with 0 mismatches, 0 unsupported and 0 rejected. Optional annotation-mode results were 547/1036: 480 mismatches, 6 unsupported and 3 rejected. This baseline is not a complete-draft claim and is not silently relabeled as evidence for later implementation changes.

    Use the conformance contract for the current evidence entry point, exact provenance, profile distinctions and case-level gap accounting. Unsupported/rejected cases remain in their original denominator. Historical benchmark-results/conformance.json and conformance-dialects.json are historical snapshots, not automatically the newest checkout's acceptance result.

    Run these commands from this independent module's root:

    moon check --target native --deny-warn --output-json
moon test --target native --deny-warn --output-json
moon check --target js --deny-warn --output-json
moon test --target js --deny-warn --output-json
moon check --target wasm --deny-warn --output-json
moonx benchmarks/run-campaign.mbtx

    The campaign downloads the official corpus, runs every official case, and measures 11 scenarios in two execution environments, each with five fresh processes (110 measurements total):

    • moonx: validator and timing loop execute inside its linear-memory Wasm sandbox, providing a portable-runtime performance baseline.
    • moon run --target native --release: validator and timing loop execute in the device's compiled native binary, providing its native performance baseline.

    The two environments use the same workloads and counters, but publish separate median/min/max tables. A Wasm orchestrator launching native children is not itself a Wasm validator benchmark. See the runner guide for the same-source entry generation and build-layout differences.

    It saves the corpus, provenance, detailed outcomes and logs in benchmark-results/. Move or remove the previous campaign's logs/ and runs/ before starting another; existing evidence is not overwritten silently.

    For a shorter infrastructure check, use moonx benchmarks/run-campaign.mbtx smoke. For case-level TDD, see the runner guide. Ordinary official-suite gaps do not fail CI; compilation, runtime/harness failures and invalid performance results do. Unit tests cover library-specific contracts rather than duplicating the full official corpus.

    The independent benchmark workflow runs on pushes to main and manual dispatch from this independent repository. Standalone .mbtx orchestration uses moonx; the native package runners can be invoked with moon run.

    #Performance policy

    Correctness comes first. Compilation hoists schema parsing and regular-expression compilation out of repeated validation; object checks share a property walk. The implementation retains numeric/depth preflight. uniqueItems uses exact pairwise equality generally, with specialized set-based paths for eligible homogeneous constructed primitive arrays. Precision-bearing and structured values retain their exact comparison semantics. It does not claim to be the fastest validator. Performance claims require reproducible benchmarks, representative valid/invalid cases and the same semantics on both sides.

    #License

    Apache-2.0. Vendored conformance data, when present, retains its upstream license and pinned-source attribution.

    CompileError

    source
    pub(all) suberror CompileError {
      InvalidKeyword(path~ : String, reason~ : String)
      UnknownDialect(String)
      UnsupportedVocabulary(dialect~ : String, vocabulary~ : String)
      UnsupportedKeyword(path~ : String, keyword~ : String)
      UnsupportedPattern(path~ : String, pattern~ : String, reason~ : String)
      DepthLimit(Int)
      NumericBudget(path~ : String)
    } derive(
    Debug    )

    EvaluationError

    source
    pub(all) suberror EvaluationError {
      DepthLimit(Int)
      InvalidNumber(path~ : String)
    } derive(
    Debug    )

    NumericError

    source
    pub(all) suberror NumericError {
      InvalidNumber
    } derive(
    Debug    )

    ParseError

    source
    pub(all) suberror ParseError {
      InvalidJson(position~ : Int, reason~ : String)
      DepthLimit(Int)
    } derive(
    Debug    )

    PatternError

    source
    pub(all) suberror PatternError {
      NotSupported(String)
    } derive(
    Debug    )

    CompiledPattern

    source
    pub struct CompiledPattern {
      // private fields
    }

    Opaque compiled pattern. No external regexp type leaks through this API.

    ExactNumber

    source
    pub enum ExactNumber {
      Repr(coeff~ : 
    BigInt    , scale~ : Int)
      Double(Double)
    }

    A number taken from a JSON tree, kept exactly.

    Repr numbers carry the literal as written ("0.3", "1e10"); Double numbers were constructed in code and only the IEEE value remains. The representation stays private: literal validation is exact, while Double validation uses the finite double's shortest decimal rendering. Original digits discarded by another parser cannot be recovered.

    ExactNumber::compare

    source
    fn ExactNumber::compare(self : ExactNumber, other : ExactNumber) -> Int raise NumericError

    ExactNumber::equal

    source
    fn ExactNumber::equal(self : ExactNumber, other : ExactNumber) -> Bool raise NumericError

    Mathematical equality, including differently spelled decimal numbers.

    ExactNumber::from_double

    source
    fn ExactNumber::from_double(v : Double) -> ExactNumber

    Interpret a Json number node's literal exactly when available.

    ExactNumber::from_literal

    source
    fn ExactNumber::from_literal(text : String) -> ExactNumber?

    Interpret a decimal literal (JSON number grammar) exactly.

    ExactNumber::is_integer

    source
    fn ExactNumber::is_integer(self : ExactNumber) -> Bool raise NumericError

    Decimal integer predicate; no rounding or floating-point tolerance.

    ExactNumber::multiple_of

    source
    fn ExactNumber::multiple_of(self : ExactNumber, divisor : ExactNumber) -> Bool raise NumericError

    Exact divisibility of aligned integer coefficients, never rounded division.

    ExactNumber::of_json

    source
    fn ExactNumber::of_json(v : Json) -> ExactNumber?

    A Json number node read as exactly as the source allows: parsed trees carry only the double (@json.parse does not keep literal text on this toolchain — verified by probe), constructed trees may carry a repr.

    ExactNumber::sign

    source
    fn ExactNumber::sign(self : ExactNumber) -> Int raise NumericError

    Fault

    source
    pub(all) struct Fault {
      instance_path : String
      schema_path : String
      keyword : String
      message : String
      schema_uri : String?
    } derive(Eq, 
    Debug    )

    One assertion failure. schema_path is a JSON Pointer relative to the schema document that owns the failing schema; schema_uri identifies that document (None when the document was supplied without any URI), so schema_uri + "#" + schema_path dereferences the failing position.

    Fault::equal

    source
    fn Fault::equal(Fault, Fault) -> Bool

    Fault::not_equal

    source
    fn Fault::not_equal(x : Fault, y : Fault) -> Bool

    Fault::to_repr

    source

    PatternSupport

    source
    pub(all) enum PatternSupport {
      Supported(CompiledPattern)
      Unsupported(reason~ : String)
    }

    Schema

    source
    pub struct Schema {
      // private fields
    }

    An owned, precompiled 2020-12 validation plan. Compilation does not mutate the source document; constants and enumerations are copied into the plan. After compilation the schema depends only on the plan: later mutations to the documents or the registry map are not observed.

    Schema::compile

    source
    fn Schema::compile(document : Json, base_uri? : String, resources? : Map[String, Json], max_depth? : Int) -> Schema raise CompileError

    Compile a schema once, optionally together with a registry of further schema documents (resources, keyed by the URI each document was supplied under) and the document's own base_uri. Reference resolution follows RFC 3986 against $id-established resource identities; the registry is consulted in memory only — compilation never performs I/O. Unknown extension keywords are annotations; known malformed keywords, unresolvable references, unknown anchors and duplicate identities fail compilation. The returned schema owns its plan: later mutations to the documents or to the registry map are not observed.

    Schema::faults

    source
    fn Schema::faults(self : Schema, value : Json, max_depth? : Int) -> Array[Fault] raise

    Collect assertion failures. Paths are RFC 6901 JSON Pointers; failed speculative anyOf/oneOf/not/if branches do not leak their diagnostics.

    Schema::valid

    source
    fn Schema::valid(self : Schema, value : Json, max_depth? : Int) -> Bool raise

    Validate without allocating a diagnostic array. Invalid JSON numbers and exhausted traversal budgets are errors, not ordinary validation failures.

    compile_pattern

    source
    fn compile_pattern(pattern : String) -> PatternSupport

    Unanchored search semantics. Patterns are passed to the regexp engine after only the rewrites the engine lacks (see translate): the ECMA dot line-terminator class, \xHH and \cX escapes, and General_Category property-name normalization. Backreferences, named groups, scoped flag groups and shorthand classes use the engine's own ECMA semantics, and dialect-invalid syntax surfaces as a typed Unsupported result. A leading anchored positive/negative lookahead chain is evaluated with separate compiled expressions; other lookarounds are explicit unsupported errors.

    parse

    source
    fn parse(source : StringView, max_nesting_depth? : Int) -> Json raise ParseError

    Parse JSON while retaining every numeric literal in Json::Number's repr. Duplicate object keys use the last value. Offsets are UTF-16 code units.

    pattern_matches

    source
    fn pattern_matches(support : PatternSupport, text : String) -> Bool raise PatternError

    Unsupported patterns cannot silently turn into a false validation result.

    Source Files

    Powered by MoonBit

    Site sourceReport issuePackagesBuild queueSkillsStatistics

    © 2026 mooncakes.io