fabro/docs/public/reference/sdk.mdx

---
title: "Fabro SDK"
description: "Using Fabro as a Rust library for AI agents and multi-provider LLM completions"
---

Fabro can be used as a Rust SDK with two primary entry points:

- **`fabro-agent`** — a full AI coding agent with tool use, sandboxed execution, event streaming, and context management. Use this when you want to build an agent that can read files, run commands, and interact with a codebase.
- **`fabro-llm`** — a standalone LLM client for multi-provider completions, streaming, and tool execution loops. Use this when you want direct control over LLM calls without the agent layer.

Both crates can be used independently of Fabro's workflow engine.

## Agent (`fabro-agent`)

The `fabro-agent` crate provides a session-based AI agent that runs an LLM with tool use in a sandboxed environment. The agent loop streams LLM responses, executes tool calls (`shell`, `read_file`, `write_file`, `edit_file`, `glob`, `grep`, `web_fetch`, `web_search`), feeds results back, and repeats until the model responds with text or hits a safety limit.

```toml title="Cargo.toml"
[dependencies]
fabro-auth = { git = "https://github.com/fabro-sh/fabro" }
fabro-agent = { git = "https://github.com/fabro-sh/fabro" }
fabro-llm = { git = "https://github.com/fabro-sh/fabro" }
fabro-model = { git = "https://github.com/fabro-sh/fabro" }
tokio = { version = "1", features = ["full"] }
```

### Quick start

```rust
use fabro_agent::{
    AnthropicProfile, LocalSandbox, Session, SessionOptions,
};
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_model::catalog::LlmCatalogSettings;
use fabro_model::Catalog;
use std::path::PathBuf;
use std::sync::Arc;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let source = EnvCredentialSource::new();
    let catalog = Arc::new(Catalog::from_builtin_with_overrides(&LlmCatalogSettings::default())?);
    let client = Client::from_source(&source, Arc::clone(&catalog)).await?;
    let sandbox = Arc::new(LocalSandbox::new(PathBuf::from(".")));
    let profile = Arc::new(AnthropicProfile::new("claude-sonnet-4-5"));
    let config = SessionOptions::default();

    let mut session = Session::new(client, profile, sandbox, config);
    session.initialize().await?;

    // Subscribe to events before sending input
    let mut events = session.subscribe();
    tokio::spawn(async move {
        while let Ok(event) = events.recv().await {
            if let fabro_agent::AgentEvent::TextDelta { delta } = &event.event {
                print!("{delta}");
            }
        }
    });

    session.process_input("List the files in this directory").await?;
    session.close();
    Ok(())
}
```

### Session

`Session` is the core type. It holds the LLM client, a provider profile, a sandbox, and configuration. The main loop lives inside `process_input()`.

**Constructor:**

```rust
pub fn new(
    llm_client: Client,
    provider_profile: Arc<dyn AgentProfile>,
    sandbox: Arc<dyn Sandbox>,
    config: SessionOptions,
) -> Self
```

**Lifecycle methods:**

| Method | Description |
|---|---|
| `initialize().await` | Discovers project docs, skills, and MCP servers. Call before `process_input`. |
| `process_input(input).await` | Sends user input and runs the agent loop until the model stops, the session is interrupted, or an error occurs. |
| `close()` | Ends the session and emits `SessionEnded`. |
| `interrupt()` | Cancels the current `process_input` call. |
| `cancel_token()` | Returns a `CancellationToken` for external cancellation. |

**Inspection:**

| Method | Description |
|---|---|
| `state()` | Returns `SessionState`: `Idle`, `Thinking`, `Executing`, or `Closed`. |
| `history()` | Returns the conversation as `&History` (a sequence of `Turn` values). |
| `subscribe()` | Returns a broadcast receiver for `SessionEvent` values. |

**Steering:**

| Method | Description |
|---|---|
| `steer(message)` | Injects a system-level guidance message into the next LLM call. |
| `follow_up(message)` | Queues a follow-up user message after the current turn completes. |

### SessionOptions

All fields are public. Key settings with their defaults:

| Field | Default | Description |
|---|---|---|
| `default_command_timeout_ms` | `10,000` | Default timeout for Bash tool commands. |
| `max_command_timeout_ms` | `600,000` | Maximum allowed timeout for Bash tool commands. |
| `enable_loop_detection` | `true` | Detect and break out of repetitive tool call patterns. |
| `enable_context_compaction` | `true` | Automatically summarize old turns when approaching the context window limit. |
| `compaction_threshold_percent` | `80` | Context window usage percentage that triggers compaction. |
| `max_subagent_depth` | `1` | Maximum nesting depth for sub-agents. |
| `wall_clock_timeout` | `None` | Hard timeout for `process_input`. Triggers `InterruptReason::WallClockTimeout`. |
| `tool_hooks` | `None` | Pre/post hooks around tool execution (see [Tool hooks](#tool-hooks)). |
| `mcp_servers` | `[]` | MCP server configurations to connect on startup. |
| `skill_dirs` | `None` | Directories to discover `SKILL.md` files. `None` uses convention defaults. |

### Sandbox

The `Sandbox` trait abstracts where tools execute — local filesystem, Docker container, SSH remote, or a cloud sandbox. All tool operations go through this interface.

```rust
#[async_trait]
pub trait Sandbox: Send + Sync {
    async fn read_file_bytes(&self, path: &str) -> Result<Vec<u8>, String>;
    async fn read_file_text(&self, path: &str) -> Result<String, String>;
    async fn read_file(&self, path: &str, offset: Option<usize>, limit: Option<usize>) -> Result<String, String>;
    async fn write_file(&self, path: &str, content: &str) -> Result<(), String>;
    async fn delete_file(&self, path: &str) -> Result<(), String>;
    async fn file_exists(&self, path: &str) -> Result<bool, String>;
    async fn list_directory(&self, path: &str, depth: Option<usize>) -> Result<Vec<DirEntry>, String>;
    async fn exec_command(
        &self,
        command: &str,
        timeout_ms: u64,
        working_dir: Option<&str>,
        env_vars: Option<&HashMap<String, String>>,
        cancel_token: Option<CancellationToken>,
    ) -> Result<ExecResult, String>;
    async fn grep(&self, pattern: &str, path: &str, options: &GrepOptions) -> Result<Vec<String>, String>;
    async fn walk_files(&self, base: &str, relative_start: &str, options: &WalkOptions) -> Result<Vec<SandboxFile>, String>;
    async fn glob(&self, pattern: &str, path: Option<&str>) -> Result<Vec<String>, String>;
    async fn initialize(&self) -> Result<(), String>;
    async fn cleanup(&self) -> Result<(), String>;
    fn working_directory(&self) -> &str;
    fn platform(&self) -> &str;
    fn os_version(&self) -> String;
    // ... optional methods with defaults: setup_git(), git_push_ref(), etc.
}
```

**Built-in implementations:**

| Type | Description |
|---|---|
| `LocalSandbox` | Executes directly on the local filesystem. |
| `DockerSandbox` | Runs inside a Docker container (feature-gated: `docker`). |

The `DaytonaSandbox` implementation (feature-gated: `daytona`) runs inside a Daytona cloud sandbox.

### Provider profiles

The `AgentProfile` trait encapsulates LLM-specific system prompts, tool definitions, and capability metadata. It controls how the agent presents itself to the model.

```rust
pub trait AgentProfile: Send + Sync {
    fn provider(&self) -> Provider;
    fn model(&self) -> &str;
    fn tool_registry(&self) -> &ToolRegistry;
    fn tool_registry_mut(&mut self) -> &mut ToolRegistry;
    fn build_system_prompt(&self, env: &dyn Sandbox, ...) -> String;
    fn capabilities(&self) -> ProfileCapabilities;
    fn tools(&self) -> Vec<ToolDefinition>;
    // ...
}
```

Built-in profiles: `AnthropicProfile`, `OpenAiProfile`, `GeminiProfile`.

### Events

All operations emit `AgentEvent` values through a tokio broadcast channel. Subscribe before calling `process_input()`.

```rust
let mut rx = session.subscribe();
tokio::spawn(async move {
    while let Ok(event) = rx.recv().await {
        match event.event {
            AgentEvent::TextDelta { delta } => print!("{delta}"),
            AgentEvent::ToolCallStarted { tool_name, .. } => {
                println!("[calling {tool_name}]");
            }
            AgentEvent::ToolCallCompleted { tool_name, is_error, .. } => {
                println!("[{tool_name} done, error={is_error}]");
            }
            AgentEvent::LoopDetected => println!("[loop detected]"),
            AgentEvent::CompactionCompleted { .. } => println!("[context compacted]"),
            _ => {}
        }
    }
});
```

Key `AgentEvent` variants:

| Variant | Description |
|---|---|
| `SessionStarted` / `SessionEnded` | Session lifecycle. |
| `TextDelta { delta }` | Incremental text from the model. |
| `ReasoningDelta { delta }` | Incremental reasoning/thinking text. |
| `AssistantMessage { text, model, usage, tool_call_count }` | Complete assistant turn with token usage. |
| `ToolCallStarted { tool_name, tool_call_id, arguments }` | A tool call is about to execute. |
| `ToolCallCompleted { tool_name, tool_call_id, output, is_error }` | A tool call finished. |
| `Error { error }` | An `AgentError` occurred. |
| `LoopDetected` | The agent is repeating itself. |
| `CompactionStarted` / `CompactionCompleted` | Context window compaction. |
| `SubAgentSpawned` / `SubAgentCompleted` | Sub-agent lifecycle. |
| `McpServerReady` / `McpServerFailed` | MCP server connection status. |

### Tool hooks

Implement `ToolHookCallback` to intercept tool calls for approval, logging, or transformation:

```rust
use fabro_agent::{ToolHookCallback, ToolHookDecision};
use async_trait::async_trait;

struct MyHooks;

#[async_trait]
impl ToolHookCallback for MyHooks {
    async fn pre_tool_use(
        &self,
        tool_name: &str,
        tool_input: &serde_json::Value,
    ) -> ToolHookDecision {
        if tool_name == "shell" {
            println!("Agent wants to run: {}", tool_input["command"]);
        }
        ToolHookDecision::Proceed // or Block { reason }
    }

    async fn post_tool_use(&self, tool_name: &str, _call_id: &str, _output: &str) {
        println!("{tool_name} completed");
    }

    async fn post_tool_use_failure(&self, tool_name: &str, _call_id: &str, error: &str) {
        eprintln!("{tool_name} failed: {error}");
    }
}
```

Pass hooks via `SessionOptions`:

```rust
let config = SessionOptions {
    tool_hooks: Some(Arc::new(MyHooks)),
    ..Default::default()
};
```

For simple sync approval, use `ToolApprovalAdapter` to wrap a closure:

```rust
use fabro_agent::ToolApprovalAdapter;
use std::sync::Arc;

let config = SessionOptions {
    tool_hooks: Some(Arc::new(ToolApprovalAdapter(Arc::new(|tool_name, _args| {
        if tool_name == "shell" {
            Err("shell is not allowed".into())
        } else {
            Ok(())
        }
    })))),
    ..Default::default()
};
```

### Error handling

All fallible `Session` methods return `Result<T, AgentError>`:

| Variant | Description |
|---|---|
| `Llm(SdkError)` | An error from the LLM provider (wraps `fabro_llm::error::SdkError`). |
| `SessionClosed` | `process_input` was called on a closed session. |
| `InvalidState(String)` | The session is in an unexpected state. |
| `ToolExecution(String)` | A tool execution failed. |
| `Interrupted(InterruptReason)` | The session was cancelled or timed out. |

---

## LLM client (`fabro-llm`)

The `fabro-llm` crate is a standalone Rust library for calling LLM providers. It provides a unified client that routes requests to Anthropic, OpenAI, Gemini, and other providers, with built-in streaming, tool execution, retries, and middleware.

You can use it independently of Fabro's workflow engine — add it as a dependency in any Rust project.

```toml title="Cargo.toml"
[dependencies]
fabro-auth = { git = "https://github.com/fabro-sh/fabro" }
fabro-llm = { git = "https://github.com/fabro-sh/fabro" }
fabro-model = { git = "https://github.com/fabro-sh/fabro" }
tokio = { version = "1", features = ["full"] }
serde_json = "1"
```

### Quick start

The simplest path is an environment-backed `CredentialSource`, an explicit `Arc<Catalog>`, then `Client::from_source(&source, catalog)`. That keeps credential and model resolution explicit while still auto-reading environment variables such as `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, and `GEMINI_API_KEY`.

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::generate::{generate, GenerateParams};
use fabro_model::catalog::LlmCatalogSettings;
use fabro_model::Catalog;
use std::sync::Arc;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let source = EnvCredentialSource::new();
    let catalog = Arc::new(Catalog::from_builtin_with_overrides(&LlmCatalogSettings::default())?);
    let client = Client::from_source(&source, Arc::clone(&catalog)).await?;

    let result = generate(
        GenerateParams::new("claude-sonnet-4-5", client.clone())
            .prompt("Explain ownership in Rust in two sentences.")
    ).await?;

    println!("{}", result.text());
    println!("Tokens used: {}", result.total_usage.total_tokens);
    Ok(())
}
```

### Client

`Client` is the core type that holds provider adapters and middleware. It routes each request to the appropriate provider.

#### Creating from a credential source

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_model::catalog::LlmCatalogSettings;
use fabro_model::Catalog;
use std::sync::Arc;

let source = EnvCredentialSource::new();
let catalog = Arc::new(Catalog::from_builtin_with_overrides(&LlmCatalogSettings::default())?);
let client = Client::from_source(&source, Arc::clone(&catalog)).await?;
```

For env-backed usage, `EnvCredentialSource` checks for API key environment variables and registers adapters for each provider found:

| Environment variable | Provider |
|---|---|
| `ANTHROPIC_API_KEY` | Anthropic |
| `OPENAI_API_KEY` | OpenAI |
| `GEMINI_API_KEY` or `GOOGLE_API_KEY` | Gemini |
| `MOONSHOT_API_KEY` or `KIMI_API_KEY` | Moonshot AI; `MOONSHOT_API_KEY` takes precedence |
| `ZAI_API_KEY` | ZAI |
| `MINIMAX_API_KEY` | Minimax |
| `INCEPTION_API_KEY` | Inception |
| `POOLSIDE_API_KEY` | Poolside |
| `DEEPSEEK_API_KEY` | DeepSeek |
| `OPENROUTER_API_KEY` | OpenRouter, when enabled in settings |

The first provider registered becomes the default. Provider base URLs come from the model catalog. For vault-backed usage inside Fabro, use `fabro_auth::VaultCredentialSource` instead.

The built-in Modal definition reads two proxy-token headers from the vault, so `EnvCredentialSource` does not configure it automatically. For direct SDK use, enable Modal and set its endpoint URL in the catalog:

```toml
[llm.providers.modal]
enabled = true
base_url = "https://your-endpoint.modal.run/v1"
```

Then read the two environment variables explicitly and create a typed credential after constructing `catalog` from those settings:

```rust
use fabro_auth::ApiCredential;
use fabro_llm::client::Client;
use std::collections::HashMap;

let credential = ApiCredential::with_extra_headers(
    "modal",
    HashMap::from([
        ("Modal-Key".to_string(), std::env::var("MODAL_TOKEN_ID")?),
        (
            "Modal-Secret".to_string(),
            std::env::var("MODAL_TOKEN_SECRET")?,
        ),
    ]),
);
let client = Client::from_credentials(vec![credential], catalog).await?;
```

#### Creating manually

```rust
use fabro_llm::client::Client;
use fabro_llm::providers::AnthropicAdapter;
use std::collections::HashMap;
use std::sync::Arc;

let adapter = AnthropicAdapter::new("sk-ant-...")
    .with_base_url("https://custom-proxy.example.com");

let mut providers = HashMap::new();
providers.insert("anthropic".to_string(), Arc::new(adapter) as _);

let client = Client::new(providers, Some("anthropic".to_string()), vec![]);
```

#### Low-level calls

For direct control without the tool loop, use `complete()` and `stream()` on the client:

```rust
use fabro_llm::types::{Request, Message};

let request = Request {
    model: "claude-sonnet-4-5".into(),
    messages: vec![Message::user("Hello")],
    ..Default::default()
};

let response = client.complete(&request).await?;
println!("{}", response.text());
```

### High-level generation

The `generate()` function wraps the client with automatic tool execution loops, retries, and timeouts. It is the recommended entry point for most use cases.

#### Basic completion

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::generate::{generate, GenerateParams};

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let result = generate(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .system("You are a helpful assistant.")
        .prompt("What is the capital of France?")
        .temperature(0.0)
).await?;

println!("{}", result.text());
```

#### Multi-turn conversations

Use `.messages()` instead of `.prompt()` to pass a full conversation history:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::types::Message;

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let result = generate(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .messages(vec![
            Message::user("My name is Alice."),
            Message::assistant("Hello Alice! How can I help you?"),
            Message::user("What's my name?"),
        ])
).await?;
```

<Note>
You cannot use both `.prompt()` and `.messages()` on the same request — this returns `SdkError::Configuration`.
</Note>

#### GenerateParams reference

| Method | Type | Description |
|---|---|---|
| `new(model, client)` | `(impl Into<String>, Arc<Client>)` | Required. Model ID or alias plus the client to use |
| `.prompt(text)` | `impl Into<String>` | Convenience: sends a single user message |
| `.messages(msgs)` | `Vec<Message>` | Full conversation history |
| `.system(text)` | `impl Into<String>` | System prompt |
| `.tools(tools)` | `Vec<Tool>` | Tools available to the model |
| `.tool_choice(choice)` | `ToolChoice` | How the model selects tools |
| `.max_tool_rounds(n)` | `u32` | Max tool execution rounds (default: 1) |
| `.temperature(t)` | `f64` | Sampling temperature |
| `.top_p(p)` | `f64` | Nucleus sampling |
| `.max_tokens(n)` | `i64` | Maximum output tokens |
| `.stop_sequences(seqs)` | `Vec<String>` | Stop sequences |
| `.reasoning_effort(level)` | `impl Into<String>` | e.g. `"low"`, `"medium"`, `"high"` |
| `.provider(name)` | `impl Into<String>` | Force a specific provider |
| `.max_retries(n)` | `u32` | Retry count for transient errors (default: 2) |
| `.timeout(config)` | `TimeoutConfig` | Total and per-step timeouts |
| `.abort_signal(token)` | `CancellationToken` | Cancel generation |
| `.stop_when(f)` | `Fn(&[StepResult]) -> bool` | Custom stop condition after each tool round |

#### GenerateResult

`GenerateResult` dereferences to `Response`, so you can call response methods directly:

```rust
let result = generate(params).await?;

// Response methods (via Deref)
result.text();            // concatenated text output
result.tool_calls();      // Vec<ToolCall> from the final response
result.reasoning();       // Option<String> — extended thinking content

// GenerateResult fields
result.response;          // Response — the final LLM response
result.tool_results;      // Vec<ToolResult> — from the final step
result.total_usage;       // Usage — aggregated across all steps
result.steps;             // Vec<StepResult> — one per tool round
result.output;            // Option<Value> — for structured output
```

### Tools

Tools let the model call functions during generation. There are two kinds:

- **Active tools** have an execute handler — Fabro runs them automatically and feeds results back to the model.
- **Passive tools** have no handler — Fabro returns the tool calls to you in the response.

#### Defining an active tool

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::tools::Tool;
use serde_json::json;

let weather = Tool::active(
    "get_weather",
    "Get the current weather for a city",
    json!({
        "type": "object",
        "properties": {
            "city": { "type": "string", "description": "City name" }
        },
        "required": ["city"]
    }),
    |args, _ctx| async move {
        let city = args["city"].as_str().unwrap_or("unknown");
        Ok(json!({ "temperature": "72°F", "city": city }))
    },
);
```

#### Using tools with generate

```rust
# use fabro_auth::EnvCredentialSource;
# use fabro_llm::client::Client;
# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let result = generate(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .prompt("What's the weather in San Francisco?")
        .tools(vec![weather])
        .max_tool_rounds(5)
).await?;

// Inspect the tool execution history
for (i, step) in result.steps.iter().enumerate() {
    let calls = step.response.tool_calls();
    println!("Step {i}: {} tool calls, {} results", calls.len(), step.tool_results.len());
}
```

The `generate()` function loops automatically: the model calls tools, Fabro executes them, feeds results back, and repeats until the model stops or `max_tool_rounds` is reached.

#### Tool choice

Control how the model selects tools:

```rust
use fabro_llm::types::ToolChoice;

// Let the model decide (default)
# use fabro_auth::EnvCredentialSource;
# use fabro_llm::client::Client;
# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
GenerateParams::new("opus", client.clone()).tool_choice(ToolChoice::Auto);

// Force a specific tool
GenerateParams::new("opus", client.clone()).tool_choice(ToolChoice::Named {
    tool_name: "get_weather".into()
});

// Force the model to use some tool
GenerateParams::new("opus", client.clone()).tool_choice(ToolChoice::Required);

// Prevent tool use
GenerateParams::new("opus", client.clone()).tool_choice(ToolChoice::None);
```

#### Passive tools

Passive tools let you handle execution yourself:

```rust
# use fabro_auth::EnvCredentialSource;
# use fabro_llm::client::Client;
# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let search = Tool::passive(
    "search",
    "Search the codebase",
    json!({
        "type": "object",
        "properties": {
            "query": { "type": "string" }
        },
        "required": ["query"]
    }),
);

let result = generate(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .prompt("Find all uses of the Config struct")
        .tools(vec![search])
).await?;

// Handle tool calls yourself
for call in result.tool_calls() {
    println!("Model wants to call {} with {}", call.name, call.arguments);
}
```

### Streaming

#### Text stream

For simple cases where you only need the text deltas:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::generate::{stream, GenerateParams};
use futures::StreamExt;

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let stream_result = stream(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .prompt("Write a haiku about Rust")
).await?;

let mut text_stream = stream_result.text_stream();
while let Some(chunk) = text_stream.next().await {
    print!("{}", chunk?);
}
```

#### Full event stream

For fine-grained control, consume `StreamEvent` variants directly:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::generate::{stream, GenerateParams};
use fabro_llm::types::StreamEvent;
use futures::StreamExt;

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let mut stream_result = stream(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .prompt("Explain monads")
).await?;

while let Some(event) = stream_result.next().await {
    match event? {
        StreamEvent::TextDelta { delta, .. } => print!("{delta}"),
        StreamEvent::ReasoningDelta { delta } => eprint!("[thinking] {delta}"),
        StreamEvent::ToolCallStart { tool_call } => {
            println!("\n> Calling tool: {}", tool_call.name);
        }
        StreamEvent::StepFinish { usage, .. } => {
            println!("\n[step done, {} tokens]", usage.total_tokens);
        }
        StreamEvent::Finish { response, .. } => {
            println!("\n[done: {:?}]", response.finish_reason);
        }
        _ => {}
    }
}
```

#### StreamEvent variants

| Variant | Description |
|---|---|
| `StreamStart` | Stream opened |
| `TextStart { text_id }` | Text block started |
| `TextDelta { delta, text_id }` | Incremental text chunk |
| `TextEnd { text_id }` | Text block ended |
| `ReasoningStart` | Extended thinking started |
| `ReasoningDelta { delta }` | Incremental reasoning chunk |
| `ReasoningEnd` | Extended thinking ended |
| `ToolCallStart { tool_call }` | Tool call started |
| `ToolCallDelta { tool_call }` | Incremental tool call arguments |
| `ToolCallEnd { tool_call }` | Tool call complete |
| `StepFinish { finish_reason, usage, response, tool_calls, tool_results }` | A tool round completed (more rounds may follow) |
| `Finish { finish_reason, usage, response }` | Generation complete |
| `Error { error, raw }` | Provider error |

### Structured output

Generate typed JSON objects that conform to a JSON Schema:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_llm::generate::{generate_object, GenerateParams};
use serde_json::json;

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let schema = json!({
    "type": "object",
    "properties": {
        "name": { "type": "string" },
        "age": { "type": "integer" },
        "hobbies": {
            "type": "array",
            "items": { "type": "string" }
        }
    },
    "required": ["name", "age", "hobbies"]
});

let result = generate_object(
    GenerateParams::new("claude-sonnet-4-5", client.clone())
        .prompt("Generate a profile for a fictional character"),
    schema,
).await?;

let profile = result.output.expect("structured output");
println!("Name: {}", profile["name"]);
```

### Middleware

Middleware intercepts requests and responses for logging, caching, or transformation:

```rust
use fabro_llm::middleware::{Middleware, NextFn, NextStreamFn};
use fabro_llm::provider::StreamEventStream;
use fabro_llm::types::{Request, Response};
use fabro_llm::error::SdkError;
use async_trait::async_trait;

struct LoggingMiddleware;

#[async_trait]
impl Middleware for LoggingMiddleware {
    async fn handle_complete(
        &self,
        request: Request,
        next: NextFn,
    ) -> Result<Response, SdkError> {
        println!("Request to model: {}", request.model);
        let response = next(request).await?;
        println!("Response: {} tokens", response.usage.total_tokens);
        Ok(response)
    }

    async fn handle_stream(
        &self,
        request: Request,
        next: NextStreamFn,
    ) -> Result<StreamEventStream, SdkError> {
        println!("Streaming request to model: {}", request.model);
        next(request).await
    }
}
```

Add middleware to the client:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use fabro_model::catalog::LlmCatalogSettings;
use fabro_model::Catalog;

let source = EnvCredentialSource::new();
let catalog = std::sync::Arc::new(Catalog::from_builtin_with_overrides(&LlmCatalogSettings::default())?);
let mut client = Client::from_source(&source, catalog).await?;
client.add_middleware(std::sync::Arc::new(LoggingMiddleware));
```

### Model catalog

The crate embeds a catalog of known models with metadata:

```rust
use fabro_llm::catalog;

// Look up a model by ID or alias
let info = catalog::get_model_info("opus").unwrap();
println!("{} ({})", info.display_name, info.provider);
println!("Context: {} tokens", info.limits.context_window);
println!("Tools: {}, Vision: {}", info.features.tools, info.features.vision);

// List all models for a provider
let models = catalog::list_models(Some("anthropic"));

// Get the default model for a provider
let default = catalog::default_model_for_provider("openai").unwrap();

// Find a capability-matched model on a different provider
let equivalent = catalog::closest_model("gemini", &info);
```

See [Models](/core-concepts/models) for the full catalog table.

### Error handling

All fallible operations return `Result<T, SdkError>`. The error type classifies failures to enable retry and failover decisions:

```rust
use fabro_llm::error::SdkError;

match result {
    Err(SdkError::Provider { kind, detail }) => {
        println!("Provider error ({}): {}", detail.provider, detail.message);
        if let Some(code) = detail.status_code {
            println!("HTTP {code}");
        }
    }
    Err(SdkError::RequestTimeout { message, .. }) => println!("Timeout: {message}"),
    Err(SdkError::Network { message, .. }) => println!("Network: {message}"),
    Err(SdkError::Interrupt { message }) => println!("Cancelled: {message}"),
    Err(e) => println!("Other: {e}"),
    Ok(_) => {}
}
```

#### Error classification

Every `SdkError` exposes classification methods:

| Method | Returns | Description |
|---|---|---|
| `retryable()` | `bool` | Safe to retry with the same provider (e.g. rate limit, server error) |
| `failover_eligible()` | `bool` | Safe to try a different provider |
| `retry_after()` | `Option<f64>` | Seconds to wait before retrying (from provider `Retry-After` header) |
| `status_code()` | `Option<u16>` | HTTP status code, if applicable |
| `provider_name()` | `&str` | Which provider returned the error |

#### Provider error kinds

| Kind | HTTP status | Retryable | Failover |
|---|---|---|---|
| `Authentication` | 401 | No | No |
| `AccessDenied` | 403 | No | No |
| `NotFound` | 404 | No | No |
| `InvalidRequest` | 400 | No | No |
| `RateLimit` | 429 | Yes | Yes |
| `Server` | 500, 502, 503 | Yes | Yes |
| `ContentFilter` | varies | No | No |
| `ContextLength` | varies | No | No |
| `QuotaExceeded` | varies | No | Yes |

### Retries

The `generate()` function retries automatically based on `max_retries` (default: 2). For low-level use, the `retry` function wraps any async operation:

```rust
use fabro_llm::retry::retry;
use fabro_llm::types::RetryPolicy;

let policy = RetryPolicy {
    max_retries: 3,
    base_delay: 1.0,
    max_delay: 60.0,
    backoff_multiplier: 2.0,
    jitter: true,
    on_retry: None,
};

let response = retry(&policy, || {
    let c = client.clone();
    let r = request.clone();
    async move { c.complete(&r).await }
}).await?;
```

Retry only fires when `error.retryable()` returns `true` and respects `Retry-After` headers.

### Cancellation

Pass a `CancellationToken` to interrupt long-running generation:

```rust
use fabro_auth::EnvCredentialSource;
use fabro_llm::client::Client;
use tokio_util::sync::CancellationToken;

# let source = EnvCredentialSource::new();
# let catalog = std::sync::Arc::new(fabro_model::Catalog::from_builtin_with_overrides(&fabro_model::catalog::LlmCatalogSettings::default()).unwrap());
# let client = Client::from_source(&source, catalog).await?;
let token = CancellationToken::new();
let token_clone = token.clone();

// Cancel after 30 seconds
tokio::spawn(async move {
    tokio::time::sleep(std::time::Duration::from_secs(30)).await;
    token_clone.cancel();
});

let result = generate(
    GenerateParams::new("opus", client.clone())
        .prompt("Write a novel")
        .abort_signal(token)
).await;
// Returns SdkError::Interrupt if cancelled
```

### Provider adapters

Each provider has a dedicated adapter. All adapters implement the `ProviderAdapter` trait and are interchangeable.

| Adapter | Provider | Constructor |
|---|---|---|
| `AnthropicAdapter` | Anthropic Messages API | `::new(api_key)` |
| `OpenAiAdapter` | OpenAI Responses API | `::new(api_key)` |
| `GeminiAdapter` | Google Gemini API | `::new(api_key)` |
| `OpenAiCompatibleAdapter` | Any OpenAI-compatible endpoint | `::new(api_key, base_url)` |

All adapters support `.with_base_url()` for proxies or custom endpoints. `OpenAiAdapter` also supports `.with_org_id()` and `.with_project_id()`.

#### Custom provider

Implement the `ProviderAdapter` trait to add a new provider:

```rust
use fabro_llm::provider::{ProviderAdapter, StreamEventStream};
use fabro_llm::types::{Request, Response};
use fabro_llm::error::SdkError;
use async_trait::async_trait;

struct MyProvider;

#[async_trait]
impl ProviderAdapter for MyProvider {
    fn name(&self) -> &str { "my-provider" }

    async fn complete(&self, request: &Request) -> Result<Response, SdkError> {
        // Call your provider's API
        todo!()
    }

    async fn stream(&self, request: &Request) -> Result<StreamEventStream, SdkError> {
        // Return a stream of events
        todo!()
    }
}
```

Register it on the client:

```rust
client.register_provider(Arc::new(MyProvider)).await?;
```