gem-hunter
主要言語: MoonBit
Runtime JSON schema validation for MoonBit, inspired by Zod/Pydantic. Designed for LLM Tool Calling: all-error collection, Strip mode hallucination defense, schema-to-prompt, JSON Schema export, MoonBit struct code generation, WASM-ready.
| Document | Description |
|---|---|
| Design Document | Important! Core architecture, design decisions, and future directions |
| API Reference | Detailed API documentation |
| CLI Reference | Command-line usage |
| Benchmark | Performance comparison with other validation libraries |
| Examples | Practical usage examples |
| Exporters | moon_zod exporters documentation |
| Importers | moon_zod importers documentation |
moon_zod is a runtime Schema intermediate representation (IR) — a validation contract layer decoupled from input sources and output targets. It provides runtime JSON schema validation with a fluent chainable API, designed primarily for LLM Tool Calling, and serves as a cross-format schema interoperability bridge. See DESIGN.md for details.
A Schema IR core providing runtime validation, multi-source import, multi-format export, and LLM hallucination defense — closing the loop from JSON Schema to prompt generation.
moon add Betterlol/moon_zod
Or add to moon.mod:
import { "Betterlol/moon_zod", }
let schema = @moon_zod.object({ "name": @moon_zod.string().min(2).max(50), "age": @moon_zod.number().int().min(0).max(150), "email": @moon_zod.string().email(), }) match schema.parse(input_json) { Ok(valid) => { println("Valid") println(@debug.to_string(valid)) } Err(errors) => { println("Invalid") println(errors.length().to_string()) for e in errors { println(e.to_string()) } } }
parse_stream() validates one JSON document as text arrives. Feed it arbitrary
LLM token chunks synchronously; it can reject malformed JSON early. Keep feeding
until the token source ends, then call finish() to confirm success and run
schema validation once. Trailing content is rejected across chunk boundaries.
let parser = @moon_zod.object({ "name": @moon_zod.string().min(2) }).parse_stream() parser.feed("{\"na") // StreamPending parser.feed("me\":\"Ada\"}") // StreamPending: still accepts trailing whitespace match parser.finish() { StreamValid(value) => println("valid: " + @debug.to_string(value)) StreamInvalidSchema(errors) => // report validation errors StreamInvalidJson(message) => // report malformed JSON StreamPending => () }
Input defaults to at most 1,048,576 UTF-16 code units and 128 nesting levels.
Set max_length and max_depth on parse_stream() to override these limits.
The parser buffers one document and releases its source text when JSON parsing
completes; successful schema validation is reported only by finish().
Zero-code CLI validation:
# Infer schema from sample, validate data moon run cmd/validate -- '{"name":"Alice","age":30}' '{"name":"Bob","age":25}' # PASS # Batch validation with JSON Lines moon run cmd/validate -- '{"name":"Alice"}' '{"name":"Bob"}\n{"name":"Eve"}' # Results: 2 passed, 0 failed
| Feature | moon_zod | Typical Validation |
|---|---|---|
| Error collection | Collects all errors in one pass | Most libs fail-fast on first error |
| Hallucination defense | Default Strip mode silently removes unknown fields | Would pass through hallucinated data |
| Named schema export | schema_to_prompt_named() generates modular TypeScript interfaces with type name references |
Inline expansion + duplication |
| JSON Schema export | to_json_schema() generates standard schema for LLM API |
Manual schema maintenance |
| Path precision | Every error includes exact field path (users[0].profile.age) |
Often just a flat message |
| Wasm-ready | Mutable path stack — zero heap allocation on success path | String-heavy allocation per parse |
In LLM Tool Calling, the model often produces multiple errors at once and hallucinates extra fields. moon_zod collects every error in a single pass (so you can send them all back for self-correction), and strips unknown fields by default (no silent data corruption from hallucinated keys).
moon_zod/
├── core/ # Core schema validation library
│ ├── types.mbt # ValidationError, SchemaResult, core types
│ ├── schema.mbt # Schema struct, parse dispatch, path stack
│ ├── string.mbt # string() factory + validators (trim, to_lower, to_upper)
│ ├── number.mbt # number() factory + validators
│ ├── boolean.mbt # boolean() factory
│ ├── null.mbt # null() factory
│ ├── bigint.mbt # bigint() factory
│ ├── any_unknown.mbt # any() / unknown() pass-through schemas
│ ├── array.mbt # array() factory + parse_array
│ ├── tuple.mbt # tuple() factory + parse_tuple
│ ├── object.mbt # object() + modes (strip/passthrough/strict), pick/omit/partial/extend/merge
│ ├── optional.mbt # optional() factory
│ ├── default.mbt # default() factory
│ ├── enum.mbt # enum_values() factory
│ ├── literal.mbt # literal() factory (constant values)
│ ├── union.mbt # union() factory
│ ├── intersection.mbt # intersection() / intersect()
│ ├── refine.mbt # refine() for custom validation
│ ├── transform.mbt # transform() for data transformation
│ ├── preprocess.mbt # preprocess() for input preprocessing
│ ├── shared_utils.mbt # Common utilities (unwrap_schema, peel_optional, etc.)
│ ├── constraint_extractor.mbt # Extract constraint info from rules
│ └── moon_zod_wbtest.mbt # White-box tests (path stack invariants)
│
├── combinators/ # Schema combinator utilities
│ └── schema_combinators.mbt # Schema composition helpers
│
├── async/ # Optional async extensions
│ ├── schema.mbt # Async post-parse refinements
│ ├── issue.mbt # Async issue values and factory
│ ├── jsonl.mbt # Backpressure-aware JSON Lines validation
│ ├── sse.mbt # SSE framing and metadata
│ ├── sse_json.mbt # Per-frame JSON validation
│ ├── sse_json_stream.mbt # Assembled JSON and async refinements
│ ├── line_reader.mbt # Bounded UTF-8 chunk reading
│ └── stream_limits.mbt # Line/frame buffering budgets
│
├── exporters/ # Code/schema export tools
│ ├── prompt.mbt # schema_to_prompt() / schema_to_prompt_named()
│ ├── prompt_renderer.mbt # Trait-based prompt rendering
│ ├── json_schema.mbt # to_json_schema() / to_json_schema_named()
│ ├── json_schema_renderer.mbt # Trait-based JSON Schema rendering
│ ├── moonbit_struct.mbt # schema_to_moonbit_struct() + static to_schema() generation
│ └── schema_exporter.mbt # Shared exporter utilities
│
├── importers/ # Schema import tools
│ └── from_json_schema.mbt # json_schema_to_moon_zod() — reverse JSON Schema → moon_zod code generation
│
├── tests/ # Test suite (479 tests)
│ ├── test_string.mbt # string() validator tests (trim, to_lower, to_upper, nonempty)
│ ├── test_number.mbt # number() validator tests
│ ├── test_boolean_null.mbt # boolean/null tests
│ ├── test_object.mbt # object() mode + pick/omit/partial/extend/merge tests
│ ├── test_array.mbt # array() + nonempty tests
│ ├── test_tuple.mbt # tuple() tests
│ ├── test_combinators.mbt # union/literal/optional/default/brand/bigint tests
│ ├── test_any_unknown_preprocess.mbt # any/unknown/preprocess tests
│ ├── test_transform_refine.mbt # transform/refine tests
│ ├── test_json_schema.mbt # JSON Schema export + $defs/$ref tests
│ ├── test_json_schema_fixes.mbt # exclusiveMin/Max semantics + enum edge cases
│ ├── test_moonbit_struct.mbt # MoonBit struct generation tests
│ ├── test_prompt.mbt # Prompt generation tests
│ ├── test_prompt_named.mbt # Named schema export tests
│ ├── test_custom_message.mbt # Custom error message tests
│ ├── test_errors.mbt # Error collection tests
│ ├── test_schema_to_code.mbt # Code generation tests
│ └── reexporter.mbt # Test re-exports
│
├── cmd/ # CLI tools
│ ├── main/ # Benchmark runner (performance baselines)
│ ├── wasm/ # WebAssembly cross-language benchmark
│ ├── json2schema/ # JSON → moon_zod schema code generator + JSON Schema reverse importer
│ ├── gen-struct/ # JSON Schema → MoonBit struct generator
│ └── validate/ # JSON schema validator (infer-then-validate)
│
└── examples/ # LLM agent demonstrations
├── gen-struct/ # JSON Schema → MoonBit struct generator demo
├── json2schema/ # JSON → moon_zod schema code generation
├── mock/ # Mock agent demonstrations
│ ├── llm_agent/ # Basic LLM tool calling example
│ └── educational_agent/ # Multi-round self-correction demo
├── multiple_schemas/ # Handling multiple schemas
├── real_llm_agent/ # Real LLM integration (with API fallback to mock)
├── resources/ # Sample data files (JSON, JSON Schema)
├── schema2json/ # Schema → JSON Schema export demo
├── schema2prompt/ # Schema → prompt generation showcase
├── shared_schemas/ # Shared schema definitions (library package)
└── validate_cli/ # CLI validation demo
# Testing & Building moon test # Run all tests (479 total, 0 warnings) moon build # Build the library moon check # Type check (0 errors, 0 warnings) moon info && moon fmt # Update interface + format # CLI Tools moon run cmd/main # Run performance benchmarks moon run cmd/json2schema -- '{"hello":"world"}' # JSON → moon_zod schema code moon run cmd/json2schema -- --from-json-schema '<{...}>' # JSON Schema → moon_zod code moon run cmd/json2schema -- --from-json-schema '<{...}>' --verbose # with debug output moon run cmd/gen-struct -- --schema '<{...}>' # JSON Schema → MoonBit structs moon run cmd/validate -- '{"name":"Alice"}' '{"name":"Bob"}' # Validate JSON # Examples moon run examples/mock/llm_agent # Basic LLM tool calling demo moon run examples/mock/educational_agent # Multi-round self-correction demo moon run examples/real_llm_agent -- product prompt # Real LLM with mock fallback moon run examples/real_llm_agent -- product validate # Validate with real API moon run examples/multiple_schemas # Multiple schema handling moon run examples/schema2json -- product schema # Schema → JSON Schema export moon run examples/schema2prompt -- product schema # Schema → prompt generation showcase moon run examples/json2schema # JSON → moon_zod schema code gen
string(), number(), boolean(), null(), bigint()object(Map), array(Schema), tuple([Schema...]), union(Array[Schema]), intersection(Array[Schema]), enum_values(Array[String]), literal(Json)any() and unknown() accept any JSON value (semantic distinction).min(n), .max(n), .nonempty(), .trim(), .to_lower(), .to_upper(), .email() (full RFC validation), .url() (full structure), .regex(pattern) (regular expression match), .startsWith(), .endsWith(), .includes(), .uuid(), .cuid(), .ulid(), .datetime(), .ip()/.ipv4()/.ipv6(), .length(n).int(), .positive(), .negative(), .multipleOf(), .finite(), .safe(), .min(), .max().strip() (default, removes unknown fields), .passthrough() (keeps unknown fields), .strict() (rejects unknown fields).pick(keys), .omit(keys), .partial(), .extend_with(Map), .merge(Schema).optional() and .default(value) with correct rule chaining through wrappers.transform(fn) validates then transforms; preprocess(fn, schema) transforms then validates.refine(check, message), .intersect(other) for explicit intersection.name(text) for named exports, .describe(text) for LLM prompts, .brand(text) for nominal typingmsg? parameter on all validators, .message(text) override method, type-level required_error / invalid_type_errorusers[0].profile.age)Betterlol/moon_zod/async adds single/multi-issue async refinements, JSON Lines, and single/multi-frame SSE JSON validation while keeping the core API synchronous. See the async guideschema_to_prompt(schema) — inline TypeScript-interface with constraint commentsschema_to_prompt_named(schema, include_names?) — modular interfaces with topological sorting and type name referencesto_json_schema(schema) — standard JSON Schema with full constraint annotationsto_json_schema_skeleton(schema) — lightweight skeleton (structure only)to_json_schema_named(schema, include_names?) — separate $defs and $ref referencesjson_schema_to_moon_zod(json_schema) — generate moon_zod source code from standard JSON Schema$defs, $ref, constraints, format validation, enumschema_to_moonbit_struct(schema) — recursively generate MoonBit struct/enum definitions for every object/enum schemaschema_to_moonbit_struct_full(schema) — generate definitions plus static Type::to_schema() functionsmoonbitlang/regexp for regex validationwebsite: extension.