schemalint

Static analysis for JSON Schema compatibility with LLM structured-output providers.

---

Lint your JSON Schemas against provider capability profiles before you send them to the LLM API. Catches unsupported keywords, structural violations, and restricted values at build time instead of at runtime.

schemalint check --profile openai.so.2026-04-30 schema.json

Supported Providers

Provider	Profile ID
OpenAI Structured Outputs	openai.so.2026-04-30
Anthropic Structured Outputs	anthropic.so.2026-04-30

Check multiple profiles simultaneously — each produces its own tagged diagnostics:

schemalint check \
  --profile openai.so.2026-04-30 \
  --profile anthropic.so.2026-04-30 \
  schema.json

Install

cargo install schemalint
# or build from source
cargo build --workspace --release

Features

Rule engine

• Class A keyword rules — auto-generated from the provider profile TOML. Forbid, warn, or allow any JSON Schema keyword per provider.
• Class B structural rules — configurable limits on nesting depth, total properties, enum cardinality, string-length budgets, root type requirements, additionalProperties: false, and required-field completeness.
• Semantic rules — provider-aware heuristics that catch structural anti-patterns:
  • Empty objects with additionalProperties: false
  • additionalProperties specified as an object schema instead of false
  • anyOf over object-typed branches
  • allOf combined with $ref (Anthropic-specific)
• Value restrictions — restrict keyword values to provider-approved sets (e.g., format limited to date-time, email, uuid).

Multi-profile orchestration

• Run N profiles in a single invocation with --profile repeated.
• Each diagnostic is tagged with its originating profile.
• Profiles are deduplicated by name.
• Built-in profile IDs resolve without filesystem paths.

Output formats

Format	Flag	Use case
Human-readable	--format human (default in TTY)	Terminal inspection
JSON	--format json	Scripting, programmatic consumers
SARIF v2.1.0	--format sarif	GitHub Code Scanning, Azure DevOps
GitHub Actions	--format gha	CI workflow annotations (::error / ::warning)
JUnit XML	--format junit	GitLab, Jenkins, CircleCI test reports

JSON-RPC server

Long-running server mode over stdin/stdout with line-delimited JSON-RPC 2.0:

schemalint server

{"jsonrpc":"2.0","method":"check","params":{"schema":{...},"profiles":["openai.so.2026-04-30"],"format":"json"},"id":1}
{"jsonrpc":"2.0","method":"shutdown","id":2}

• Persistent disk cache — normalized schemas survive process restarts via ~/.cache/schemalint-<pid>/.
• Hardened — 10 MB payload limit, 100k node limit, path traversal protection, broken pipe detection.

Performance

Measured on Apple M3:

Scenario	Time
Single 200-property schema	< 1 ms
500 schemas, cold start	< 50 ms
Incremental (cache hit)	< 5 ms

• Content-hash caching — identical schemas across files are normalized once.
• Rayon parallelism — file processing scales across all cores.
• Arena-allocated IR — NodeId(u32) indexing with zero pointer indirection.

Usage

# Check a single schema
schemalint check --profile openai.so.2026-04-30 schema.json

# Check all .json files in a directory
schemalint check --profile openai.so.2026-04-30 ./schemas/

# Output structured JSON
schemalint check --profile openai.so.2026-04-30 --format json schema.json

# CI-friendly output
schemalint check --profile openai.so.2026-04-30 --format gha schema.json

# Multi-profile check
schemalint check \
  --profile openai.so.2026-04-30 \
  --profile anthropic.so.2026-04-30 \
  --format sarif \
  ./schemas/

# Write output to file
schemalint check --profile openai.so.2026-04-30 --output results.json schema.json

# Start JSON-RPC server
schemalint server

Output

Human

error[OAI-K-allOf]: keyword 'allOf' is not supported by openai.so.2026-04-30
   --> schema.json
     |
     = profile: openai.so.2026-04-30
     = schema path: /
     = see: https://schemalint.dev/rules/OAI-K-allOf

1 issue found (1 error, 0 warnings) across 1 schema

SARIF

{
  "version": "2.1.0",
  "runs": [{
    "tool": { "driver": { "name": "schemalint", "rules": [...] } },
    "results": [{
      "ruleId": "OAI-K-allOf",
      "level": "error",
      "message": { "text": "keyword 'allOf' is not supported..." },
      "locations": [{ "physicalLocation": { "artifactLocation": { "uri": "schema.json" } } }]
    }]
  }]
}

GitHub Actions

::error file=schema.json,title=OAI-K-allOf::keyword 'allOf' is not supported...

JUnit

<testsuites>
  <testsuite name="schema.json" tests="1" failures="1" skipped="0" errors="0" time="0">
    <testcase name="OAI-K-allOf - keyword 'allOf'...">
      <failure type="error" message="keyword 'allOf'...">keyword 'allOf'...</failure>
    </testcase>
  </testsuite>
</testsuites>

Exit codes

Code	Meaning
0	No errors (warnings alone are OK)
1	At least one error, or a fatal parse/IO error
2	I/O error writing output file

Architecture

crates/
├── schemalint/              # Core engine + CLI
│   ├── src/cli/             # Args, file discovery, server, output formatters
│   ├── src/ir/              # Arena-allocated IR (Node, NodeId, Arena)
│   ├── src/normalize/       # Normalizer pipeline (dialect, $ref resolution, desugar)
│   ├── src/profile/         # TOML profile loader
│   ├── src/rules/           # Rule trait, registry, Class A/B + semantic rules
│   ├── tests/               # 128 tests across 12 test files
│   └── benches/             # Criterion benchmarks
└── schemalint-profiles/     # Bundled provider profiles (zero deps)
    └── profiles/
        ├── openai.so.2026-04-30.toml
        └── anthropic.so.2026-04-30.toml

Contribute

cargo build --workspace
cargo test --workspace
cargo clippy --workspace -- -D warnings
cargo fmt --all -- --check
cargo bench --no-run --workspace

Pre-commit hooks via lefthook run fmt, clippy, and tests on staged .rs files.

MSRV: 1.80.

License

MIT OR Apache-2.0