Development Guidelines

This document provides comprehensive guidelines for developers contributing to ADK-Rust. Following these standards ensures code quality, consistency, and maintainability across the project.

Table of Contents

• Getting Started
• Project Structure
• Code Style
• Error Handling
• Async Patterns
• Testing
• Documentation
• Pull Request Process
• Common Tasks

Getting Started

Prerequisites

• Rust: 1.95.0 or higher (edition 2024, check with rustc --version)
• Cargo: Latest stable
• Git: For version control
• sccache (recommended): Compilation cache that cuts rebuild times by ~70%

Setting Up Your Environment

# Clone the repository
git clone https://github.com/zavora-ai/adk-rust.git
cd adk-rust

# Option A: Nix/devenv (reproducible — identical on Linux, macOS, CI)
devenv shell

# Option B: Setup script (installs sccache, cmake, etc.)
./scripts/setup-dev.sh

# Option C: Manual
cargo build

# Install cargo-nextest (parallel test runner, ~10x faster)
curl -LsSf https://get.nexte.st/latest/mac | tar zxf - -C ${CARGO_HOME:-~/.cargo}/bin

# Run all tests
cargo nextest run --workspace

# Check for lints
cargo clippy --all-targets --all-features

# Format code
cargo fmt --all

Environment Variables

For running examples and tests that require API keys:

# Gemini (default provider)
export GOOGLE_API_KEY="your-api-key"

# OpenAI (optional)
export OPENAI_API_KEY="your-api-key"

# Anthropic (optional)
export ANTHROPIC_API_KEY="your-api-key"

Project Structure

ADK-Rust is organized as a Cargo workspace with multiple crates:

adk-rust/
├── adk-core/       # Foundational traits and types (Agent, Tool, Llm, Event)
├── adk-telemetry/  # OpenTelemetry integration
├── adk-model/      # LLM providers (Gemini, OpenAI, Anthropic)
├── adk-tool/       # Tool system (FunctionTool, MCP, AgentTool)
├── adk-session/    # Session management (in-memory, SQLite)
├── adk-artifact/   # Binary artifact storage
├── adk-memory/     # Long-term memory with search
├── adk-agent/      # Agent implementations (LlmAgent, workflow agents)
├── adk-runner/     # Execution runtime
├── adk-server/     # REST API and A2A protocol
├── adk-cli/        # Command-line launcher
├── adk-realtime/   # Voice/audio streaming agents
├── adk-graph/      # LangGraph-style workflows
├── adk-browser/    # Browser automation tools
├── adk-eval/       # Agent evaluation framework
├── adk-rust/       # Umbrella crate (re-exports all)
└── examples/       # Working examples

Crate Dependencies

Crates must be published in dependency order:

1. adk-core (no internal deps)
2. adk-telemetry
3. adk-model
4. adk-tool
5. adk-session
6. adk-artifact
7. adk-memory
8. adk-agent
9. adk-runner
10. adk-server
11. adk-cli
12. adk-realtime
13. adk-graph
14. adk-browser
15. adk-eval
16. adk-rust (umbrella)

Code Style

General Principles

1. Clarity over cleverness: Write code that is easy to read and understand
2. Explicit over implicit: Prefer explicit types and error handling
3. Small functions: Keep functions focused and under 50 lines when possible
4. Meaningful names: Use descriptive variable and function names

Formatting

Use rustfmt with default settings:

cargo fmt --all

The CI pipeline enforces formatting. Always run cargo fmt before committing.

Naming Conventions

Imports

Organize imports in this order:

// 1. Standard library
use std::collections::HashMap;
use std::sync::Arc;

// 2. External crates
use async_trait::async_trait;
use serde::{Deserialize, Serialize};
use tokio::sync::RwLock;

// 3. Internal crates (adk-*)
use adk_core::{Agent, Event, Result};

// 4. Local modules
use crate::config::Config;
use super::utils;

Clippy

All code must pass clippy with no warnings:

cargo clippy --all-targets --all-features

Address clippy warnings rather than suppressing them. If suppression is necessary, document why:

#[allow(clippy::too_many_arguments)]
// Builder pattern requires many parameters; refactoring would hurt usability
fn complex_builder(...) { }

Error Handling

Structured Error Envelope

AdkError is a structured error type with component (where), category (what kind), code (machine key), message (human text), retry hint, and optional details:

use adk_core::{AdkError, ErrorComponent, ErrorCategory, Result};

// Return Result<T> (aliased to Result<T, AdkError>)
pub async fn my_function() -> Result<String> {
    let data = fetch_data().await?;

    if data.is_empty() {
        return Err(AdkError::new(
            ErrorComponent::Tool,
            ErrorCategory::NotFound,
            "tool.data.not_found",
            "No data found for the given query",
        ));
    }

    Ok(data)
}

Choosing Component and Category

ErrorComponent identifies where the failure happened (the origin subsystem, not the trait boundary it surfaces through):

ErrorCategory classifies what went wrong:

Convenience Constructors

For common patterns:

// Structured (preferred for new code)
AdkError::not_found(ErrorComponent::Session, "session.not_found", "Session xyz not found")
AdkError::rate_limited(ErrorComponent::Model, "model.openai.rate_limited", "Too many requests")
AdkError::unauthorized(ErrorComponent::Auth, "auth.token_expired", "Bearer token expired")
AdkError::timeout(ErrorComponent::Tool, "tool.execution_timeout", "Tool timed out after 30s")

// Backward-compatible (for migration — produces .legacy codes)
AdkError::tool("No data found")
AdkError::model("Provider returned 500")
AdkError::session("Session not found")

Builder API

Attach structured metadata for richer error context:

let err = AdkError::new(
    ErrorComponent::Model,
    ErrorCategory::RateLimited,
    "model.openai.rate_limited",
    "OpenAI rate limit exceeded",
)
.with_provider("openai")
.with_upstream_status(429)
.with_request_id("req-abc123")
.with_retry(RetryHint {
    should_retry: true,
    retry_after_ms: Some(5000),
    max_attempts: Some(3),
});

Retry Hints

Retryable categories (RateLimited, Unavailable, Timeout) automatically set should_retry: true. Check retryability with err.is_retryable(), which reads retry.should_retry as the single source of truth:

if err.is_retryable() {
    if let Some(delay) = err.retry.retry_after() {
        tokio::time::sleep(delay).await;
    }
    // retry the operation
}

Category Checks

err.is_retryable()    // retry.should_retry (RateLimited, Unavailable, Timeout by default)
err.is_not_found()    // category == NotFound
err.is_unauthorized() // category == Unauthorized
err.is_rate_limited() // category == RateLimited
err.is_timeout()      // category == Timeout

Component Checks (backward compat)

err.is_model()   // component == Model
err.is_tool()    // component == Tool
err.is_session() // component == Session
err.is_config()  // code == "config.legacy" (temporary bridge)

HTTP Status and Problem JSON

AdkError maps directly to HTTP responses:

let status = err.http_status_code(); // u16 based on category
let body = err.to_problem_json();    // structured JSON error body
// body: { "error": { "code", "message", "component", "category", "requestId", "retryAfter", ... } }

Crate-Local Errors with From Impls

Crates with domain-specific errors implement From<CrateLocalError> for AdkError:

// In your crate
#[derive(Debug, thiserror::Error)]
pub enum MyToolError {
    #[error("connection failed: {0}")]
    ConnectionFailed(String),
    #[error("timeout after {0}ms")]
    Timeout(u64),
}

impl From<MyToolError> for AdkError {
    fn from(err: MyToolError) -> Self {
        let (category, code) = match &err {
            MyToolError::ConnectionFailed(_) => (ErrorCategory::Unavailable, "mytool.connection"),
            MyToolError::Timeout(_) => (ErrorCategory::Timeout, "mytool.timeout"),
        };
        AdkError::new(ErrorComponent::Tool, category, code, err.to_string())
            .with_source(err)
    }
}

No Blanket From Impls

std::io::Error and serde_json::Error cross too many subsystem boundaries for a single blanket conversion. Use explicit map_err with the correct component:

// Good: explicit component and category
let data = std::fs::read_to_string(path)
    .map_err(|e| AdkError::new(
        ErrorComponent::Session,
        ErrorCategory::Internal,
        "session.io_read",
        format!("failed to read session file: {e}"),
    ).with_source(e))?;

// Good: for quick migration
let data = serde_json::from_str(&raw)
    .map_err(|e| AdkError::session(format!("JSON parse failed: {e}")))?;

Error Messages

Write clear, actionable error messages:

// Good: specific and actionable
AdkError::new(
    ErrorComponent::Model,
    ErrorCategory::InvalidInput,
    "model.missing_api_key",
    "API key not found. Set GOOGLE_API_KEY environment variable.",
)

// Bad: vague
AdkError::model("Invalid config")

No Panics in Library Code

Library crates must never panic on recoverable errors. Avoid unwrap(), expect(), and panic!() in src/ code (test code is exempt).

RwLock / Mutex: Use graceful degradation instead of unwrap(). A poisoned lock means another thread panicked — crashing the current thread too makes things worse.

// Bad: panics if lock is poisoned
let state = self.state.write().unwrap();

// Good: log and return a safe default
let Ok(state) = self.state.write() else {
    tracing::error!("state lock poisoned — returning default");
    return Default::default();
};

// Good: recover through the poison (data may be stale but won't crash)
let state = self.state.read().unwrap_or_else(|e| e.into_inner());

Constructors: Return Result when initialization can fail (e.g., connecting to external services).

// Bad: panics if Docker is not running
pub fn new(config: Config) -> Self {
    let client = connect().expect("connection failed");
    Self { client }
}

// Good: caller decides how to handle the failure
pub fn new(config: Config) -> Result<Self, MyError> {
    let client = connect().map_err(|e| MyError::Init(e.to_string()))?;
    Ok(Self { client })
}

Builder methods: Use if let Some instead of expect() for Arc::get_mut():

// Bad: panics if Arc is shared
pub fn add_callback(mut self, cb: Callback) -> Self {
    Arc::get_mut(&mut self.callbacks).expect("not shared").push(cb);
    self
}

// Good: silent no-op (builder pattern guarantees single ownership)
pub fn add_callback(mut self, cb: Callback) -> Self {
    if let Some(callbacks) = Arc::get_mut(&mut self.callbacks) {
        callbacks.push(cb);
    }
    self
}

Async Patterns

Use Tokio

All async code uses the Tokio runtime:

use tokio::sync::{Mutex, RwLock};

// Prefer RwLock for read-heavy data
let state: Arc<RwLock<State>> = Arc::new(RwLock::new(State::default()));

// Use Mutex for write-heavy or simple cases
let counter: Arc<Mutex<u32>> = Arc::new(Mutex::new(0));

Tokio Feature Convention

Library crates (adk-*) must declare only the minimal tokio features they actually use:

# Library crates — minimal features
tokio = { workspace = true, features = ["rt", "sync", "time"] }

# Binary crates only (adk-cli, examples) — full is acceptable
tokio = { workspace = true, features = ["full"] }

Never use features = ["full"] in a library crate. This forces all downstream consumers to compile every tokio subsystem whether they need it or not.

Async Traits

Use async_trait for async trait methods:

use async_trait::async_trait;

#[async_trait]
pub trait MyTrait: Send + Sync {
    async fn do_work(&self) -> Result<()>;
}

Streaming

Use EventStream for streaming responses:

use adk_core::EventStream;
use async_stream::stream;
use futures::Stream;

fn create_stream() -> EventStream {
    let s = stream! {
        yield Ok(Event::new("inv-1"));
        yield Ok(Event::new("inv-2"));
    };
    Box::pin(s)
}

Thread Safety

All public types must be Send + Sync:

// Good: Thread-safe
pub struct MyAgent {
    name: String,
    tools: Vec<Arc<dyn Tool>>,  // Arc for shared ownership
}

// Verify with compile-time checks
fn assert_send_sync<T: Send + Sync>() {}
fn _check() {
    assert_send_sync::<MyAgent>();
}

Testing

Test Runner

ADK-Rust uses cargo-nextest for test execution. Nextest runs each test binary in a separate process with parallel scheduling, giving ~10x speedup over cargo test on this workspace.

# Install (one-time)
curl -LsSf https://get.nexte.st/latest/mac | tar zxf - -C ${CARGO_HOME:-~/.cargo}/bin

# Or via devenv (included automatically)
devenv shell

Configuration lives in .config/nextest.toml with two profiles:

• default — local development (fail-fast, no retries)
• ci — CI runs (retries for flaky tests, slow-test warnings)

Test Organization

crate/
├── src/
│   ├── lib.rs          # Unit tests at bottom of file
│   └── module.rs       # Module-specific tests
└── tests/
    └── integration.rs  # Integration tests

Unit Tests

Place unit tests in the same file as the code:

pub fn add(a: i32, b: i32) -> i32 {
    a + b
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_add() {
        assert_eq!(add(2, 3), 5);
    }

    #[tokio::test]
    async fn test_async_function() {
        let result = async_function().await;
        assert!(result.is_ok());
    }
}

Integration Tests

Place in tests/ directory:

// tests/integration_test.rs
use adk_core::*;

#[tokio::test]
async fn test_full_workflow() {
    // Setup
    let service = InMemorySessionService::new();

    // Execute
    let session = service.create(request).await.unwrap();

    // Assert
    assert_eq!(session.id(), "test-session");
}

Mock Testing

Use MockLlm for testing without API calls:

use adk_model::MockLlm;

#[tokio::test]
async fn test_agent_with_mock() {
    let mock = MockLlm::new(vec![
        "First response".to_string(),
        "Second response".to_string(),
    ]);

    let agent = LlmAgentBuilder::new("test")
        .model(Arc::new(mock))
        .build()
        .unwrap();

    // Test agent behavior
}

Test Commands

# Run all tests (nextest — parallel, fast)
cargo nextest run --workspace

# Run specific crate tests
cargo nextest run -p adk-core

# Run with CI profile (retries flaky tests)
cargo nextest run --workspace --profile ci

# Run doctests (nextest doesn't run these — use cargo test)
cargo test --workspace --doc

# Run ignored tests (require API keys)
cargo nextest run --workspace -- --run-ignored

# Run with output (nextest shows output for failing tests by default)
cargo nextest run --workspace --no-capture

# Devenv shortcuts
devenv shell ws-test          # nextest, default profile
devenv shell ws-test-ci       # nextest, CI profile
devenv shell ws-test-slow     # cargo test fallback (includes doctests)

Documentation

Doc Comments

Use /// for public items:

/// Creates a new LLM agent with the specified configuration.
///
/// # Arguments
///
/// * `name` - A unique identifier for this agent
/// * `model` - The LLM provider to use for reasoning
///
/// # Examples
///
/// ```rust
/// use adk_agent::LlmAgentBuilder;
///
/// let agent = LlmAgentBuilder::new("assistant")
///     .model(Arc::new(model))
///     .build()?;
/// ```
///
/// # Errors
///
/// Returns an error with component `Agent` if the model is not set.
pub fn new(name: impl Into<String>) -> Self {
    // ...
}

Module Documentation

Add module-level docs at the top of lib.rs:

//! # adk-core
//!
//! Core types and traits for ADK-Rust.
//!
//! ## Overview
//!
//! This crate provides the foundational types...

README Files

Each crate should have a README.md with:

1. Brief description
2. Installation instructions
3. Quick example
4. Link to full documentation

Documentation Tests

Ensure doc examples compile:

cargo test --doc --all

Pull Request Process

Before Submitting

1. Run the full test suite:

   cargo nextest run --workspace

2. Run clippy:

   cargo clippy --all-targets --all-features

3. Format code:

   cargo fmt --all

4. Update documentation if adding/changing public API

5. Add tests for new functionality

PR Guidelines

• Title: Clear, concise description of the change
• Description: Explain what and why (not how)
• Size: Keep PRs focused; split large changes
• Tests: Include tests for new functionality
• Breaking changes: Clearly document in description

Commit Messages

Follow conventional commits:

feat: add OpenAI streaming support
fix: correct tool parameter validation
docs: update quickstart guide
refactor: simplify session state management
test: add integration tests for A2A protocol

Project Scaffolding

Use cargo adk new with the Composable Template System to scaffold new projects. The built-in registry provides 12 templates, 9 add-ons, and 5 enterprise patterns:

# Basic agent (default)
cargo adk new my-agent

# Agent with tools and Docker support
cargo adk new my-agent --template tools --addon docker

# A2A protocol agent with CI and telemetry
cargo adk new my-agent --template a2a --addon ci --addon telemetry

# Graph workflow with enterprise observability
cargo adk new my-agent --template graph --addon telemetry --addon docker --addon ci

The --addon flag is composable — combine any base template with any number of addons. See the Composable Templates documentation for the full list of templates, addons, and enterprise patterns.

Common Tasks

Adding a New Tool

1. Create the tool:

use adk_core::{Tool, ToolContext, Result};
use async_trait::async_trait;
use serde_json::Value;

pub struct MyTool {
    // fields
}

#[async_trait]
impl Tool for MyTool {
    fn name(&self) -> &str {
        "my_tool"
    }

    fn description(&self) -> &str {
        "Does something useful"
    }

    fn parameters_schema(&self) -> Option<Value> {
        Some(serde_json::json!({
            "type": "object",
            "properties": {
                "input": { "type": "string" }
            },
            "required": ["input"]
        }))
    }

    async fn execute(&self, ctx: Arc<dyn ToolContext>, args: Value) -> Result<Value> {
        let input = args["input"].as_str().unwrap_or_default();
        Ok(serde_json::json!({ "result": input }))
    }
}

1. Add to agent:

let agent = LlmAgentBuilder::new("agent")
    .model(model)
    .tool(Arc::new(MyTool::new()))
    .build()?;

Adding a New Model Provider

1. Create module in adk-model/src/:

// adk-model/src/mymodel/mod.rs
mod client;
pub use client::MyModelClient;

1. Implement the Llm trait:

use adk_core::{Llm, LlmRequest, LlmResponse, LlmResponseStream, Result};

pub struct MyModelClient {
    api_key: String,
}

#[async_trait]
impl Llm for MyModelClient {
    fn name(&self) -> &str {
        "my-model"
    }

    async fn generate_content(
        &self,
        request: LlmRequest,
        stream: bool,
    ) -> Result<LlmResponseStream> {
        // Implementation
    }
}

1. Add feature flag in adk-model/Cargo.toml:

[features]
mymodel = ["dep:mymodel-sdk"]

1. Export conditionally:

#[cfg(feature = "mymodel")]
pub mod mymodel;
#[cfg(feature = "mymodel")]
pub use mymodel::MyModelClient;

Adding a New Agent Type

1. Create module in adk-agent/src/:

// adk-agent/src/my_agent.rs
use adk_core::{Agent, EventStream, InvocationContext, Result};
use async_trait::async_trait;

pub struct MyAgent {
    name: String,
}

#[async_trait]
impl Agent for MyAgent {
    fn name(&self) -> &str {
        &self.name
    }

    fn description(&self) -> &str {
        "My custom agent"
    }

    async fn run(&self, ctx: Arc<dyn InvocationContext>) -> Result<EventStream> {
        // Implementation
    }
}

1. Export in adk-agent/src/lib.rs:

mod my_agent;
pub use my_agent::MyAgent;

Debugging Tips

1. Enable tracing:

   adk_telemetry::init_telemetry();

2. Inspect events:

   while let Some(event) = stream.next().await {
       eprintln!("Event: {:?}", event);
   }

3. Use RUST_LOG:

   RUST_LOG=debug cargo run --example myexample

---

Previous: ← Access Control

Questions? Open an issue on GitHub.