# HCL JSON Syntax Specification This is the specification for the JSON serialization for hcl. HCL is a system for defining configuration languages for applications. The HCL information model is designed to support multiple concrete syntaxes for configuration, and this JSON-based format complements [the native syntax](../hclsyntax/spec.md) by being easy to machine-generate, whereas the native syntax is oriented towards human authoring and maintenance This syntax is defined in terms of JSON as defined in [RFC7159](https://tools.ietf.org/html/rfc7159). As such it inherits the JSON grammar as-is, and merely defines a specific methodology for interpreting JSON constructs into HCL structural elements and expressions. This mapping is defined such that valid JSON-serialized HCL input can be _produced_ using standard JSON implementations in various programming languages. _Parsing_ such JSON has some additional constraints not beyond what is normally supported by JSON parsers, so a specialized parser may be required that is able to: - Preserve the relative ordering of properties defined in an object. - Preserve multiple definitions of the same property name. - Preserve numeric values to the precision required by the number type in [the HCL syntax-agnostic information model](../spec.md). - Retain source location information for parsed tokens/constructs in order to produce good error messages. ## Structural Elements [The HCL syntax-agnostic information model](../spec.md) defines a _body_ as an abstract container for attribute definitions and child blocks. A body is represented in JSON as either a single JSON object or a JSON array of objects. Body processing is in terms of JSON object properties, visited in the order they appear in the input. Where a body is represented by a single JSON object, the properties of that object are visited in order. Where a body is represented by a JSON array, each of its elements are visited in order and each element has its properties visited in order. If any element of the array is not a JSON object then the input is erroneous. When a body is being processed in the _dynamic attributes_ mode, the allowance of a JSON array in the previous paragraph does not apply and instead a single JSON object is always required. As defined in the language-agnostic model, body processing is in terms of a schema which provides context for interpreting the body's content. For JSON bodies, the schema is crucial to allow differentiation of attribute definitions and block definitions, both of which are represented via object properties. The special property name `"//"`, when used in an object representing a HCL body, is parsed and ignored. A property with this name can be used to include human-readable comments. (This special property name is _not_ processed in this way for any _other_ HCL constructs that are represented as JSON objects.) ### Attributes Where the given schema describes an attribute with a given name, the object property with the matching name — if present — serves as the attribute's definition. When a body is being processed in the _dynamic attributes_ mode, each object property serves as an attribute definition for the attribute whose name matches the property name. The value of an attribute definition property is interpreted as an _expression_, as described in a later section. Given a schema that calls for an attribute named "foo", a JSON object like the following provides a definition for that attribute: ```json { "foo": "bar baz" } ``` ### Blocks Where the given schema describes a block with a given type name, each object property with the matching name serves as a definition of zero or more blocks of that type. Processing of child blocks is in terms of nested JSON objects and arrays. If the schema defines one or more _labels_ for the block type, a nested JSON object or JSON array of objects is required for each labelling level. These are flattened to a single ordered sequence of object properties using the same algorithm as for body content as defined above. Each object property serves as a label value at the corresponding level. After any labelling levels, the next nested value is either a JSON object representing a single block body, or a JSON array of JSON objects that each represent a single block body. Use of an array accommodates the definition of multiple blocks that have identical type and labels. Given a schema that calls for a block type named "foo" with no labels, the following JSON objects are all valid definitions of zero or more blocks of this type: ```json { "foo": { "child_attr": "baz" } } ``` ```json { "foo": [ { "child_attr": "baz" }, { "child_attr": "boz" } ] } ``` ```json { "foo": [] } ``` The first of these defines a single child block of type "foo". The second defines _two_ such blocks. The final example shows a degenerate definition of zero blocks, though generators should prefer to omit the property entirely in this scenario. Given a schema that calls for a block type named "foo" with _two_ labels, the extra label levels must be represented as objects or arrays of objects as in the following examples: ```json { "foo": { "bar": { "baz": { "child_attr": "baz" }, "boz": { "child_attr": "baz" } }, "boz": { "baz": { "child_attr": "baz" } } } } ``` ```json { "foo": { "bar": { "baz": { "child_attr": "baz" }, "boz": { "child_attr": "baz" } }, "boz": { "baz": [ { "child_attr": "baz" }, { "child_attr": "boz" } ] } } } ``` ```json { "foo": [ { "bar": { "baz": { "child_attr": "baz" }, "boz": { "child_attr": "baz" } } }, { "bar": { "baz": [ { "child_attr": "baz" }, { "child_attr": "boz" } ] } } ] } ``` ```json { "foo": { "bar": { "baz": { "child_attr": "baz" }, "boz": { "child_attr": "baz" } }, "bar": { "baz": [ { "child_attr": "baz" }, { "child_attr": "boz" } ] } } } ``` Arrays can be introduced at either the label definition or block body definition levels to define multiple definitions of the same block type or labels while preserving order. A JSON HCL parser _must_ support duplicate definitions of the same property name within a single object, preserving all of them and the relative ordering between them. The array-based forms are also required so that JSON HCL configurations can be produced with JSON producing libraries that are not able to preserve property definition order and multiple definitions of the same property. ## Expressions JSON lacks a native expression syntax, so the HCL JSON syntax instead defines a mapping for each of the JSON value types, including a special mapping for strings that allows optional use of arbitrary expressions. ### Objects When interpreted as an expression, a JSON object represents a value of a HCL object type. Each property of the JSON object represents an attribute of the HCL object type. The property name string given in the JSON input is interpreted as a string expression as described below, and its result is converted to string as defined by the syntax-agnostic information model. If such a conversion is not possible, an error is produced and evaluation fails. An instance of the constructed object type is then created, whose values are interpreted by again recursively applying the mapping rules defined in this section to each of the property values. If any evaluated property name strings produce null values, an error is produced and evaluation fails. If any produce _unknown_ values, the _entire object's_ result is an unknown value of the dynamic pseudo-type, signalling that the type of the object cannot be determined. It is an error to define the same property name multiple times within a single JSON object interpreted as an expression. In full expression mode, this constraint applies to the name expression results after conversion to string, rather than the raw string that may contain interpolation expressions. ### Arrays When interpreted as an expression, a JSON array represents a value of a HCL tuple type. Each element of the JSON array represents an element of the HCL tuple type. The tuple type is constructed by enumerating the JSON array elements, creating for each an element whose type is the result of recursively applying the expression mapping rules. Correspondence is preserved between the array element indices and the tuple element indices. An instance of the constructed tuple type is then created, whose values are interpreted by again recursively applying the mapping rules defined in this section. ### Numbers When interpreted as an expression, a JSON number represents a HCL number value. HCL numbers are arbitrary-precision decimal values, so a JSON HCL parser must be able to translate exactly the value given to a number of corresponding precision, within the constraints set by the HCL syntax-agnostic information model. In practice, off-the-shelf JSON serializers often do not support customizing the processing of numbers, and instead force processing as 32-bit or 64-bit floating point values. A _producer_ of JSON HCL that uses such a serializer can provide numeric values as JSON strings where they have precision too great for representation in the serializer's chosen numeric type in situations where the result will be converted to number (using the standard conversion rules) by a calling application. Alternatively, for expressions that are evaluated in full expression mode an embedded template interpolation can be used to faithfully represent a number, such as `"${1e150}"`, which will then be evaluated by the underlying HCL native syntax expression evaluator. ### Boolean Values The JSON boolean values `true` and `false`, when interpreted as expressions, represent the corresponding HCL boolean values. ### The Null Value The JSON value `null`, when interpreted as an expression, represents a HCL null value of the dynamic pseudo-type. ### Strings When interpreted as an expression, a JSON string may be interpreted in one of two ways depending on the evaluation mode. If evaluating in literal-only mode (as defined by the syntax-agnostic information model) the literal string is intepreted directly as a HCL string value, by directly using the exact sequence of unicode characters represented. Template interpolations and directives MUST NOT be processed in this mode, allowing any characters that appear as introduction sequences to pass through literally: ```json "Hello world! Template sequences like ${ are not intepreted here." ``` When evaluating in full expression mode (again, as defined by the syntax- agnostic information model) the literal string is instead interpreted as a _standalone template_ in the HCL Native Syntax. The expression evaluation result is then the direct result of evaluating that template with the current variable scope and function table. ```json "Hello, ${name}! Template sequences are interpreted in full expression mode." ``` In particular the _Template Interpolation Unwrapping_ requirement from the HCL native syntax specification must be implemented, allowing the use of single-interpolation templates to represent expressions that would not otherwise be representable in JSON, such as the following example where the result must be a number, rather than a string representation of a number: ```json "${ a + b }" ``` ## Static Analysis The HCL static analysis operations are implemented for JSON values that represent expressions, as described in the following sections. Due to the limited expressive power of the JSON syntax alone, use of these static analyses functions rather than normal expression evaluation is used as additional context for how a JSON value is to be interpreted, which means that static analyses can result in a different interpretation of a given expression than normal evaluation. ### Static List An expression interpreted as a static list must be a JSON array. Each of the values in the array is interpreted as an expression and returned. ### Static Map An expression interpreted as a static map must be a JSON object. Each of the key/value pairs in the object is presented as a pair of expressions. Since object property names are always strings, evaluating the key expression with a non-`nil` evaluation context will evaluate any template sequences given in the property name. ### Static Call An expression interpreted as a static call must be a string. The content of the string is interpreted as a native syntax expression (not a _template_, unlike normal evaluation) and then the static call analysis is delegated to that expression. If the original expression is not a string or its contents cannot be parsed as a native syntax expression then static call analysis is not supported. ### Static Traversal An expression interpreted as a static traversal must be a string. The content of the string is interpreted as a native syntax expression (not a _template_, unlike normal evaluation) and then static traversal analysis is delegated to that expression. If the original expression is not a string or its contents cannot be parsed as a native syntax expression then static call analysis is not supported.