verdict-code

An interactive opencode-like CLI assistant that demonstrates the verdict Rust agent framework.

Every user message is processed through a real verdict Pipeline — with AgentSteps, Guards, Verdicts, registered Tools, and Skills.

Features

• 3-step pipeline per message: classify intent (Haiku) → generate response (Sonnet) → safety check
• 4 custom tools registered in a ToolRegistry via FunctionTool: word_count, char_count, reverse_string, to_uppercase
• 5 built-in skills registered in a SkillRegistry: code_review, test_writing, refactoring, rust_debugging, api_design
• Tool-assisted pipeline: type "count", "reverse", or "uppercase" to trigger a 4th ToolCall step automatically
• Readline input with history (rustyline)
• Colored output (colored)

Slash Commands

Command	Description
/tools	Run a demo pipeline with all 4 tools and print results
/skills	List registered skills and run a code_review demo
/model <haiku|sonnet>	Switch the active model
/clear	Clear conversation history
/help	Show all commands
/quit	Exit

Setup

Requires an OpenAI-compatible LLM endpoint. Edit the constants in src/main.rs:

const BASE_URL: &str = "http://your-endpoint/v1";
const API_KEY: &str  = "your-api-key";

Then:

cargo run

What it demonstrates

This demo shows how verdict is meant to be used:

• StepAction::LlmCall with per-step ProviderSpec model routing
• StepAction::Custom with tokio::task::block_in_place for async LLM calls
• StepAction::ToolCall with ToolSet::Allow scoping
• StepAction::UseSkill with SkillMode::Pipeline
• Guard::AllOf, Guard::NonEmptyOutput, Guard::NoSecretsInOutput
• Verdict::Automated
• PipelineRunner with wired ToolRegistry and SkillRegistry

Related

• verdict — the framework library
• verdict-micro-agent — TDD micro-agent demo