Types & Schemas

Models answer in text. If an answer is entirely JSON, optionally inside a ```json fence, Halo parses it into an object, so gates and later stages can read fields like context.data.review.score. A schema declares the shape a pipeline must return.

Schemas

schema Address { city: String, zip: String }

schema UserProfile {
  id: String,
  age: Integer,
  tags: Array<String>,
  isPremium: Boolean,
  address: Address,
}

agent Extractor [openai/gpt-4o-mini, temp: 0]
  system "Extract a user profile. Reply with JSON only."

pipeline ExtractProfile(text: String) -> UserProfile:
  text |> Extractor.extract()

Field types:

Type	Accepts
String	a string
Number	a finite number
Integer	a whole number
Boolean	true / false
Array, List, Array<T>	a list (every item checked against T)
Object, JSON	an object
Any	anything
another schema's name	an object matching that schema

Every field is required. Extra fields are allowed. The validator warns about unknown type names.

Enforcement

When a pipeline is declared -> SchemaName, the runtime checks the output once the pipeline completes. A string output is parsed as JSON first. If the output doesn't match, the run throws a SchemaValidationError (code SCHEMA_VALIDATION) with the problem, for example missing field "isPremium". If it matches, the output is returned parsed.

@retry does not retry schema failures. To have the model fix bad JSON, add a Require(JSON) policy to the agent together with @correct(n). That checks the answer at the agent call and asks the model again (see Directives).

Halted and paused runs aren't checked. A return type that isn't a declared schema, like Result, isn't checked either.

Parameters

Parameter types (text: String, items: List, user: UserProfile) document the pipeline. They are not checked at runtime. The CLI requires that every parameter is given.

There are no Asset or Embedding types. Multimodal inputs are not supported.