Interface ToolCandidate

A candidate tool submitted for creation review by the EmergentJudge.

Contains all the information the judge needs to evaluate safety, correctness, determinism, and bounded execution: the tool's identity and schemas, its source code or composition spec, the sandbox API allowlist, and the results of any test runs.

interface ToolCandidate {
    name: string;
    description: string;
    inputSchema: JSONSchemaObject;
    outputSchema?: JSONSchemaObject;
    source: string;
    implementationMode: "sandbox" | "compose";
    allowlist?: string[];
    testResults: {
        input: unknown;
        output: unknown;
        success: boolean;
        error?: string;
    }[];
}

Properties

name description inputSchema outputSchema? source implementationMode allowlist? testResults

Properties

name

name: string

Machine-readable name for the proposed tool.

description

description: string

Natural language description of the tool's purpose and behaviour.

inputSchema

inputSchema: JSONSchemaObject

JSON Schema defining the tool's input arguments.

Optional outputSchema

outputSchema?: JSONSchemaObject

JSON Schema defining the tool's expected output shape (optional).

source

source: string

The tool's source code (for sandbox mode) or serialized composition spec (for compose mode). Included in the LLM prompt for security auditing.

implementationMode

implementationMode: "sandbox" | "compose"

Whether the tool was built by composing existing tools ('compose') or by running arbitrary code in a sandbox ('sandbox').

Optional allowlist

allowlist?: string[]

Sandbox API allowlist — the set of APIs the tool's code is permitted to invoke. Only relevant for 'sandbox' mode tools. Used by the safety auditor to verify that the code does not exceed its declared API surface.

testResults

testResults: {
    input: unknown;
    output: unknown;
    success: boolean;
    error?: string;
}[]

Results of test runs executed against the candidate tool before review. Each entry contains the input, output, success flag, and optional error. The judge uses these to assess correctness and determinism.

Type declaration

  • input: unknown
  • output: unknown
  • success: boolean
  • Optional error?: string

Settings

Member Visibility

Theme

On This Page

namedescriptioninputSchemaoutputSchemasourceimplementationModeallowlisttestResults