Crate agentcarousel 

Evaluate agents and skills from YAML/TOML fixtures, run cases with mocks or live backends, persist runs to SQLite, and export evidence for reporting or registry upload.

Audience

• CLI users — run the agentcarousel or agc binary for validate, test, eval, report, bundle, publish, export, and trust-check.
• Library embedders — use runner to execute fixtures programmatically, fixtures to load them, evaluators for scoring, reporters for output, and core for shared types (Run, Case, FixtureFile, …).

Crate layout

Module	Role
cli	Clap-based CLI; cli::run is the process entrypoint.
core	Serializable models, errors, and judge provider helpers.
runner	Async execution: runner::run_fixtures, runner::run_eval.
evaluators	Rules, golden, process, and LLM judge evaluators.
fixtures	Load and validate fixtures; fixtures::MockEngine for stubbed tool/LLM responses.
reporters	Terminal, JSON, JUnit, history persistence, and run diffs.

Quick start (CLI)

agentcarousel validate path/to/fixture.yaml
agentcarousel test path/to/fixture.yaml --offline true

Install and full options are described in the repository README and on docs.rs for this crate version.

Library quick start

Typical flow: load fixtures with fixtures::load_fixture, build runner::RunnerConfig or runner::EvalConfig, then call runner::run_fixtures or runner::run_eval inside a tokio runtime. See runner for configuration fields.

Re-exports

pub use cli::*;

pub use core::*;

pub use evaluators::*;

pub use fixtures::*;

pub use reporters::*;

pub use runner::*;

Modules

cli

Command-line interface built with clap: parse flags, load merged configuration from TOML and environment, and dispatch to validate / test / eval / report / init / bundle / publish / export / trust-check.

core

Shared domain types for fixtures, runs, traces, and metrics, plus CoreError and judge_provider helpers for LLM-backed judges.

evaluators

Pluggable evaluators that score a finished crate::CaseResult against fixture rubrics or external references: RulesEvaluator, GoldenEvaluator, ProcessEvaluator, JudgeEvaluator, and the Evaluator trait.

fixtures

Fixture I/O and validation: load YAML/TOML into crate::FixtureFile, validate against the bundled JSON Schema, and resolve tool/LLM responses via MockEngine.

reporters

Human-readable and machine-readable output: terminal tables, JSON, JUnit, persisted history (SQLite), and diff_runs / print_diff for comparing two runs.

runner

Async test and eval execution: expand fixtures into Case rows, apply mocks or live generation, optionally run evaluators (rules / golden / process / judge), and produce a Run.