diff --git a/crates/attractor/README.md b/crates/attractor/README.md
index b96a20a68..e8538eced 100644
--- a/crates/attractor/README.md
+++ b/crates/attractor/README.md
@@ -1,3 +1,172 @@
 # attractor
 
-A DOT-based pipeline runner for multi-stage AI workflows.
+A DOT-based pipeline runner for multi-stage AI workflows. Define workflows as Graphviz `digraph` files and execute them with pluggable handlers, conditional routing, human-in-the-loop gates, parallel branching, retry policies, and checkpoint-based recovery.
+
+## Key Concepts
+
+- **Graph** -- A directed graph parsed from DOT syntax containing nodes, edges, and attributes. The graph carries a `goal` describing the pipeline's purpose.
+- **Node** -- A workflow step. Graphviz shapes map to handler types (e.g., `Mdiamond` = start, `Msquare` = exit, `box` = codergen/LLM, `diamond` = conditional, `hexagon` = human gate, `component` = parallel).
+- **Edge** -- A connection between nodes with optional `condition`, `label`, `weight`, and `fidelity` attributes that control routing.
+- **Handler** -- An async trait implementation that executes a node and returns an `Outcome`. Built-in handlers include `StartHandler`, `ExitHandler`, `CodergenHandler`, `ConditionalHandler`, `WaitHumanHandler`, `ParallelHandler`, `FanInHandler`, `ToolHandler`, and `ManagerLoopHandler`.
+- **Outcome** -- The result of executing a handler, carrying a `StageStatus` (Success, Fail, PartialSuccess, Retry, Skipped), optional routing hints (`preferred_label`, `suggested_next_ids`), and context updates.
+- **Context** -- A thread-safe key-value store shared across pipeline stages, supporting snapshots and isolated cloning for parallel branches.
+- **Interviewer** -- A trait for human-in-the-loop interactions. Implementations include `AutoApproveInterviewer`, `QueueInterviewer`, `CallbackInterviewer`, `ConsoleInterviewer`, and `RecordingInterviewer`.
+- **Checkpoint** -- A serializable snapshot of execution state (completed nodes, context values, logs) for crash recovery and resume.
+
+## Pipeline Definition
+
+Pipelines are defined using Graphviz DOT syntax:
+
+```dot
+digraph MyPipeline {
+    graph [goal="Implement and validate a feature"]
+    rankdir=LR
+    node [shape=box, timeout="900s"]
+
+    start     [shape=Mdiamond, label="Start"]
+    exit      [shape=Msquare, label="Exit"]
+    plan      [label="Plan", prompt="Plan the implementation"]
+    implement [label="Implement", prompt="Implement the plan"]
+    validate  [label="Validate", prompt="Run tests"]
+    gate      [shape=diamond, label="Tests passing?"]
+
+    start -> plan -> implement -> validate -> gate
+    gate -> exit      [label="Yes", condition="outcome=success"]
+    gate -> implement [label="No", condition="outcome!=success"]
+}
+```
+
+## Usage
+
+### Parsing and Validating a Pipeline
+
+```rust
+use attractor::pipeline::prepare_pipeline;
+
+let dot_source = r#"digraph Simple {
+    graph [goal="Run tests"]
+    start [shape=Mdiamond]
+    exit  [shape=Msquare]
+    work  [shape=box, prompt="Run the test suite"]
+    start -> work -> exit
+}"#;
+
+let graph = prepare_pipeline(dot_source)
+    .expect("pipeline should parse and validate");
+assert_eq!(graph.name, "Simple");
+assert_eq!(graph.goal(), "Run tests");
+```
+
+`prepare_pipeline` parses the DOT source, applies built-in transforms (variable expansion, stylesheet application, preamble injection), and validates the graph against 14 built-in lint rules.
+
+### Running a Pipeline
+
+```rust
+use attractor::engine::{PipelineEngine, RunConfig};
+use attractor::event::EventEmitter;
+use attractor::handler::HandlerRegistry;
+use attractor::handler::start::StartHandler;
+use attractor::handler::exit::ExitHandler;
+use attractor::handler::codergen::CodergenHandler;
+use attractor::pipeline::prepare_pipeline;
+
+let graph = prepare_pipeline(dot_source).unwrap();
+
+let mut registry = HandlerRegistry::new(Box::new(CodergenHandler::new(None)));
+registry.register("start", Box::new(StartHandler));
+registry.register("exit", Box::new(ExitHandler));
+registry.register("codergen", Box::new(CodergenHandler::new(None)));
+
+let engine = PipelineEngine::new(registry, EventEmitter::new());
+let config = RunConfig {
+    logs_root: "/tmp/pipeline-run".into(),
+};
+
+// engine.run(&graph, &config).await
+```
+
+### Custom Handlers
+
+Implement the `Handler` trait to add custom node behavior:
+
+```rust
+use attractor::handler::Handler;
+use attractor::context::Context;
+use attractor::graph::{Graph, Node};
+use attractor::outcome::Outcome;
+use attractor::error::AttractorError;
+use async_trait::async_trait;
+use std::path::Path;
+
+struct MyHandler;
+
+#[async_trait]
+impl Handler for MyHandler {
+    async fn execute(
+        &self,
+        node: &Node,
+        context: &Context,
+        graph: &Graph,
+        logs_root: &Path,
+    ) -> Result<Outcome, AttractorError> {
+        // Custom logic here
+        Ok(Outcome::success())
+    }
+}
+```
+
+### Model Stylesheets
+
+CSS-like stylesheets control LLM model assignment with specificity-based cascading:
+
+```dot
+digraph Styled {
+    graph [
+        goal="Build feature",
+        model_stylesheet="
+            * { llm_model: claude-sonnet-4-5; llm_provider: anthropic; }
+            .code { llm_model: claude-opus-4-6; }
+            #critical_review { llm_model: gpt-5.2; llm_provider: openai; }
+        "
+    ]
+    // ...
+}
+```
+
+Selectors by specificity: `*` (universal, 0) < `shape` (1) < `.class` (2) < `#id` (3). Explicit node attributes are never overridden.
+
+### Condition Expressions
+
+Edge conditions use a simple expression syntax for routing:
+
+```
+outcome=success
+outcome!=fail
+outcome=success && context.tests_passed=true
+my_flag
+```
+
+Clauses support `=`, `!=`, and bare key truthiness checks, joined with `&&`.
+
+### Human-in-the-Loop Gates
+
+Nodes with `shape=hexagon` or `type="wait.human"` pause execution for human input. Outgoing edge labels become selectable options, with accelerator key parsing for patterns like `[A] Approve` and `F) Fix`.
+
+### Parallel Execution
+
+Nodes with `shape=component` fan out to branches concurrently. Configurable join policies: `wait_all` (default), `first_success`, `k_of_n(N)`, `quorum(0.5)`. Error policies: `continue`, `fail_fast`, `ignore`.
+
+### Checkpoints and Resume
+
+The engine saves a checkpoint after each node. Resume from a checkpoint with `engine.run_from_checkpoint(&graph, &config, &checkpoint)`.
+
+## Architecture
+
+```
+parser (DOT -> AST -> Graph)
+  -> transform (variable expansion, stylesheet, preamble)
+    -> validation (14 lint rules)
+      -> engine (execution loop with retry, edge selection, goal gates)
+        -> handler (pluggable node executors)
+          -> interviewer (human-in-the-loop I/O)
+```
diff --git a/crates/coding-agent-loop/README.md b/crates/coding-agent-loop/README.md
new file mode 100644
index 000000000..ab1f2a084
--- /dev/null
+++ b/crates/coding-agent-loop/README.md
@@ -0,0 +1,237 @@
+# coding-agent-loop
+
+A programmable agentic loop for building coding agents. This crate provides the core session management, tool execution, and LLM interaction loop used to power interactive coding assistants.
+
+## Architecture
+
+The crate is organized around a central `Session` that drives an agentic loop:
+
+1. **User input** is appended to a conversation `History`
+2. The session builds a `Request` with system prompt, history, and tools
+3. An LLM generates a response (text and/or tool calls) via `unified-llm`
+4. Tool calls are executed through a `ToolRegistry` against an `ExecutionEnvironment`
+5. Results are recorded and the loop continues until the LLM responds with text only (natural completion), a turn limit is reached, or the session is aborted
+
+```
+User Input
+    |
+    v
+[Session::process_input]
+    |
+    v
++-------------------+
+| Build Request     |  <-- system prompt + history + tools
++-------------------+
+    |
+    v
++-------------------+
+| LLM Call          |  <-- via unified-llm Client
++-------------------+
+    |
+    v
++-------------------+     +-------------------+
+| Tool Calls?  -----+-yes-| Execute Tools     |
++-------------------+     | (parallel or seq)  |
+    | no                  +-------------------+
+    v                         |
+  [Done]                      +---> loop back to Build Request
+```
+
+### Key Components
+
+- **`Session`** -- Manages the full agentic loop: LLM calls, tool execution, steering, follow-ups, abort handling, and event emission.
+- **`ProviderProfile`** (trait) -- Defines how to build system prompts, which tools to register, and what capabilities a provider supports. Ships with `AnthropicProfile`, `OpenAiProfile`, and `GeminiProfile`.
+- **`ExecutionEnvironment`** (trait) -- Abstracts filesystem, shell, grep, and glob operations. `LocalExecutionEnvironment` provides a real implementation; the trait enables sandboxing and testing.
+- **`ToolRegistry`** -- Maps tool names to definitions and async executor functions. Tools are registered per-profile.
+- **`History`** -- Ordered list of `Turn` variants (`User`, `Assistant`, `ToolResults`, `System`, `Steering`) that converts to LLM messages.
+- **`EventEmitter`** -- Broadcasts `SessionEvent`s (tool calls, text, errors, warnings) over a `tokio::sync::broadcast` channel for UI or logging.
+- **`SubAgentManager`** -- Spawns child `Session`s on background tasks for delegated work, with depth limits.
+- **`SessionConfig`** -- Tunable parameters: max turns, tool round limits, command timeouts, loop detection, output truncation limits, and user instructions.
+
+## Key Types and Traits
+
+### `Session`
+
+The main entry point. Created with an LLM client, a provider profile, an execution environment, and a config.
+
+### `ProviderProfile`
+
+```rust
+pub trait ProviderProfile: Send + Sync {
+    fn id(&self) -> String;
+    fn model(&self) -> String;
+    fn tool_registry(&self) -> &ToolRegistry;
+    fn build_system_prompt(
+        &self,
+        env: &dyn ExecutionEnvironment,
+        env_context: &EnvContext,
+        project_docs: &[String],
+        user_instructions: Option<&str>,
+    ) -> String;
+    fn capabilities(&self) -> ProfileCapabilities;
+    fn knowledge_cutoff(&self) -> &str;
+    // ... default methods for tools(), provider_options(), supports_*()
+}
+```
+
+Built-in profiles:
+- **`AnthropicProfile`** -- 200K context, extended thinking beta headers, tools: `read_file`, `write_file`, `edit_file`, `shell`, `grep`, `glob`
+- **`OpenAiProfile`** -- 128K context, reasoning effort support, tools: `read_file`, `write_file`, `shell`, `grep`, `glob`, `apply_patch` (v4a format)
+- **`GeminiProfile`** -- 1M context, safety settings, tools: all Anthropic tools plus `read_many_files`, `list_dir`, `web_search`, `web_fetch`
+
+### `ExecutionEnvironment`
+
+```rust
+pub trait ExecutionEnvironment: Send + Sync {
+    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 exec_command(&self, command: &str, timeout_ms: u64, ...) -> Result<ExecResult, String>;
+    async fn grep(&self, pattern: &str, path: &str, options: &GrepOptions) -> Result<Vec<String>, String>;
+    async fn glob(&self, pattern: &str, path: Option<&str>) -> Result<Vec<String>, String>;
+    // ... plus delete_file, file_exists, list_directory, initialize, cleanup, platform info
+}
+```
+
+`LocalExecutionEnvironment` is the real implementation with env-var filtering (strips secrets), process group management, and ripgrep/grep fallback.
+
+### `SessionConfig`
+
+```rust
+pub struct SessionConfig {
+    pub max_turns: usize,                    // 0 = unlimited
+    pub max_tool_rounds_per_input: usize,    // default: 200
+    pub default_command_timeout_ms: u64,     // default: 10s
+    pub max_command_timeout_ms: u64,         // default: 600s
+    pub enable_loop_detection: bool,         // default: true
+    pub loop_detection_window: usize,        // default: 10
+    pub max_subagent_depth: usize,           // default: 1
+    pub user_instructions: Option<String>,
+    pub reasoning_effort: Option<String>,
+    // ... plus tool_output_limits, tool_line_limits, git_root
+}
+```
+
+## Usage
+
+```rust
+use coding_agent_loop::{
+    AnthropicProfile, LocalExecutionEnvironment, Session, SessionConfig,
+};
+use std::path::PathBuf;
+use std::sync::Arc;
+use unified_llm::client::Client;
+
+// 1. Create an LLM client (via unified-llm)
+let client: Client = /* configure unified-llm client */;
+
+// 2. Choose a provider profile
+let profile = Arc::new(AnthropicProfile::new("claude-sonnet-4-20250514"));
+
+// 3. Create an execution environment
+let env = Arc::new(LocalExecutionEnvironment::new(
+    PathBuf::from("/path/to/project"),
+));
+
+// 4. Configure the session
+let config = SessionConfig {
+    max_tool_rounds_per_input: 50,
+    enable_loop_detection: true,
+    user_instructions: Some("Always write tests first".into()),
+    ..SessionConfig::default()
+};
+
+// 5. Create and initialize the session
+let mut session = Session::new(client, profile, env, config);
+session.initialize().await;
+
+// 6. Subscribe to events (for UI rendering)
+let mut rx = session.subscribe();
+tokio::spawn(async move {
+    while let Ok(event) = rx.recv().await {
+        // Handle SessionEvent: tool calls, text, errors, etc.
+    }
+});
+
+// 7. Process user input
+session.process_input("Fix the failing test in src/lib.rs").await?;
+```
+
+### Steering and Follow-ups
+
+Inject guidance mid-conversation or queue follow-up messages:
+
+```rust
+// Inject a steering message before the next LLM call
+session.steer("Focus on the root cause, not symptoms".into());
+
+// Queue a follow-up that runs after the current input completes
+session.follow_up("Now run the test suite to verify".into());
+```
+
+### Abort
+
+Cancel a running session from another thread:
+
+```rust
+let abort_flag = session.abort_flag_handle();
+// From another task:
+abort_flag.store(true, std::sync::atomic::Ordering::SeqCst);
+```
+
+### Custom Tools
+
+Register additional tools via the profile's `ToolRegistry`:
+
+```rust
+use coding_agent_loop::tool_registry::{RegisteredTool, ToolExecutor};
+use unified_llm::types::ToolDefinition;
+use std::sync::Arc;
+
+let custom_tool = RegisteredTool {
+    definition: ToolDefinition {
+        name: "my_tool".into(),
+        description: "Does something useful".into(),
+        parameters: serde_json::json!({
+            "type": "object",
+            "properties": {
+                "input": {"type": "string"}
+            },
+            "required": ["input"]
+        }),
+    },
+    executor: Arc::new(|args, env| {
+        Box::pin(async move {
+            let input = args["input"].as_str().unwrap_or("");
+            Ok(format!("Processed: {input}"))
+        })
+    }),
+};
+
+// Register on a mutable profile before creating the session
+profile.tool_registry_mut().register(custom_tool);
+```
+
+### Subagents
+
+Spawn child sessions for delegated tasks:
+
+```rust
+use coding_agent_loop::subagent::SubAgentManager;
+
+let mut profile = AnthropicProfile::new("claude-sonnet-4-20250514");
+let manager = Arc::new(tokio::sync::Mutex::new(SubAgentManager::new(3)));
+let factory = Arc::new(|| { /* create a new Session */ });
+
+// Registers spawn_agent, send_input, wait, close_agent tools
+profile.register_subagent_tools(manager, factory, 0);
+```
+
+## Safety Features
+
+- **Loop detection** -- Detects repeating tool call patterns (period 1, 2, or 3) and injects a steering warning
+- **Context window monitoring** -- Emits `ContextWindowWarning` events when estimated usage exceeds 80%
+- **Tool argument validation** -- Validates arguments against JSON Schema before execution
+- **Tool output truncation** -- Per-tool character and line limits with head/tail or tail-only truncation modes
+- **Environment variable filtering** -- `LocalExecutionEnvironment` strips secrets (`*_API_KEY`, `*_SECRET`, `*_TOKEN`, `*_PASSWORD`, `*_CREDENTIAL`) from subprocess environments
+- **Command timeouts** -- Configurable per-command with process group cleanup (SIGTERM then SIGKILL)
+- **Project doc discovery** -- Automatically discovers `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, or `.codex/instructions.md` based on provider, with a 32KB budget
diff --git a/crates/unified-llm-cli/README.md b/crates/unified-llm-cli/README.md
new file mode 100644
index 000000000..2f0c98f0f
--- /dev/null
+++ b/crates/unified-llm-cli/README.md
@@ -0,0 +1,95 @@
+# unified-llm-cli
+
+A command-line interface for interacting with LLM providers through the [unified-llm](../unified-llm/) library. Supports Anthropic, OpenAI, Google Gemini, and other providers with a single tool.
+
+## Installation
+
+```sh
+cargo install --path crates/unified-llm-cli
+```
+
+This installs the `ullm` binary.
+
+## Configuration
+
+API keys are read from environment variables. You can set them directly or place them in a `.env` file:
+
+```sh
+export ANTHROPIC_API_KEY="sk-..."
+export OPENAI_API_KEY="sk-..."
+export GEMINI_API_KEY="..."
+```
+
+## Commands
+
+### `ullm prompt`
+
+Send a prompt to an LLM and print the response.
+
+```sh
+ullm prompt "What is the capital of France?"
+```
+
+**Flags:**
+
+| Flag | Description |
+|------|-------------|
+| `-m, --model <MODEL>` | Model to use (defaults to first catalog model) |
+| `-s, --system <TEXT>` | System prompt |
+| `-o, --option <KEY=VALUE>` | Generation options: `temperature`, `max_tokens`, `top_p`, or provider-specific keys |
+| `-u, --usage` | Show token usage on stderr |
+| `--no-stream` | Disable streaming (wait for full response) |
+
+**Stdin support:** Pipe text into `ullm prompt` to use it as input. If both stdin and an argument are provided, they are concatenated.
+
+```sh
+echo "Summarize this" | ullm prompt
+cat article.txt | ullm prompt "Give me the key points"
+```
+
+**Examples:**
+
+```sh
+# Use a specific model
+ullm prompt -m claude-opus-4-6 "Explain quicksort"
+
+# Set a system prompt
+ullm prompt -s "You are a helpful translator" "Translate to French: hello"
+
+# Adjust generation parameters
+ullm prompt -o temperature=0.2 -o max_tokens=500 "Write a haiku"
+
+# Show token usage
+ullm prompt -u --no-stream "Hello"
+```
+
+### `ullm models`
+
+List available models from all providers. Running `ullm models` with no subcommand defaults to `ullm models list`.
+
+```sh
+ullm models
+```
+
+#### `ullm models list`
+
+```sh
+ullm models list
+```
+
+**Flags:**
+
+| Flag | Description |
+|------|-------------|
+| `-p, --provider <NAME>` | Filter models by provider (e.g., `anthropic`, `openai`, `gemini`) |
+| `-q, --query <TEXT>` | Search models by ID, display name, or alias (case-insensitive) |
+
+**Examples:**
+
+```sh
+# List only Anthropic models
+ullm models list --provider anthropic
+
+# Search for models matching "opus"
+ullm models list --query opus
+```
diff --git a/crates/unified-llm/README.md b/crates/unified-llm/README.md
new file mode 100644
index 000000000..92f0ed792
--- /dev/null
+++ b/crates/unified-llm/README.md
@@ -0,0 +1,240 @@
+# unified-llm
+
+A unified async Rust client library for multiple LLM providers. Write your LLM integration code once and switch between Anthropic, OpenAI, and Google Gemini without changing your application logic.
+
+## Key concepts
+
+- **Client** -- Routes requests to registered provider adapters. Can be created explicitly or auto-configured from environment variables.
+- **ProviderAdapter** -- The trait every provider implements (`complete` and `stream`). Built-in adapters: `AnthropicAdapter`, `OpenAiAdapter`, `GeminiAdapter`, `OpenAiCompatibleAdapter`.
+- **Middleware** -- Intercepts requests/responses for logging, caching, or transformation. Supports both blocking and streaming paths.
+- **generate()** -- High-level function that wraps `Client.complete()` with automatic tool execution loops, retries, timeouts, and cancellation.
+- **Tool** -- Active tools (with an execute handler) run automatically in the tool loop. Passive tools (no handler) surface tool calls back to the caller.
+- **Model catalog** -- Built-in metadata for common models. Advisory only; unknown model strings pass through.
+
+## Providers
+
+| Provider | Adapter | API | Env var |
+|----------|---------|-----|---------|
+| Anthropic | `AnthropicAdapter` | Messages API | `ANTHROPIC_API_KEY` |
+| OpenAI | `OpenAiAdapter` | Responses API | `OPENAI_API_KEY` |
+| Google Gemini | `GeminiAdapter` | generateContent | `GEMINI_API_KEY` or `GOOGLE_API_KEY` |
+| OpenAI-compatible | `OpenAiCompatibleAdapter` | Chat Completions | (custom) |
+
+All adapters support streaming, tool calling, structured output (`response_format`), and provider-specific options via `provider_options`.
+
+## Usage
+
+### Auto-configure from environment
+
+```rust
+use unified_llm::client::Client;
+use unified_llm::types::{Message, Request};
+
+let client = Client::from_env().await?;
+
+let request = Request {
+    model: "claude-sonnet-4-5".to_string(),
+    messages: vec![Message::user("What is the capital of France?")],
+    provider: None,
+    tools: None,
+    tool_choice: None,
+    response_format: None,
+    temperature: Some(0.0),
+    top_p: None,
+    max_tokens: Some(100),
+    stop_sequences: None,
+    reasoning_effort: None,
+    metadata: None,
+    provider_options: None,
+};
+
+let response = client.complete(&request).await?;
+println!("{}", response.text());
+```
+
+### High-level generate()
+
+```rust
+use unified_llm::generate::{generate, GenerateParams};
+
+let result = generate(
+    GenerateParams::new("claude-sonnet-4-5")
+        .prompt("Explain monads in one sentence")
+        .system("You are a concise programming tutor.")
+        .max_tokens(200)
+).await?;
+
+println!("{}", result.text());
+```
+
+### Tool calling
+
+```rust
+use unified_llm::generate::{generate, GenerateParams};
+use unified_llm::tools::Tool;
+use std::sync::Arc;
+
+let weather_tool = Tool::active(
+    "get_weather",
+    "Get the current weather for a city",
+    serde_json::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(serde_json::json!({"temp": "72F", "city": city}))
+    },
+);
+
+let result = generate(
+    GenerateParams::new("claude-sonnet-4-5")
+        .prompt("What's the weather in San Francisco?")
+        .tools(vec![weather_tool])
+        .max_tool_rounds(3)
+).await?;
+```
+
+### Streaming
+
+```rust
+use unified_llm::client::Client;
+use unified_llm::types::{Message, Request, StreamEvent};
+use futures::StreamExt;
+
+let client = Client::from_env().await?;
+let request = Request {
+    model: "claude-sonnet-4-5".to_string(),
+    messages: vec![Message::user("Tell me a joke")],
+    // ...other fields set to None/defaults
+    # provider: None, tools: None, tool_choice: None,
+    # response_format: None, temperature: None, top_p: None,
+    # max_tokens: None, stop_sequences: None, reasoning_effort: None,
+    # metadata: None, provider_options: None,
+};
+
+let mut stream = client.stream(&request).await?;
+while let Some(event) = stream.next().await {
+    match event? {
+        StreamEvent::TextDelta { delta, .. } => print!("{delta}"),
+        StreamEvent::Finish { response, .. } => {
+            println!("\nTokens used: {}", response.usage.total_tokens);
+        }
+        _ => {}
+    }
+}
+```
+
+### Middleware
+
+```rust
+use unified_llm::middleware::{Middleware, NextFn, NextStreamFn};
+use unified_llm::types::{Request, Response};
+use unified_llm::provider::StreamEventStream;
+use unified_llm::error::SdkError;
+
+struct LoggingMiddleware;
+
+#[async_trait::async_trait]
+impl Middleware for LoggingMiddleware {
+    async fn handle_complete(
+        &self,
+        request: Request,
+        next: NextFn,
+    ) -> Result<Response, SdkError> {
+        eprintln!("Request to model: {}", request.model);
+        let response = next(request).await?;
+        eprintln!("Response tokens: {}", response.usage.total_tokens);
+        Ok(response)
+    }
+
+    async fn handle_stream(
+        &self,
+        request: Request,
+        next: NextStreamFn,
+    ) -> Result<StreamEventStream, SdkError> {
+        next(request).await
+    }
+}
+```
+
+### OpenAI-compatible providers
+
+```rust
+use unified_llm::providers::OpenAiCompatibleAdapter;
+use std::sync::Arc;
+
+let adapter = OpenAiCompatibleAdapter::new("your-api-key", "https://api.groq.com/openai/v1")
+    .with_name("groq");
+```
+
+### Model catalog
+
+```rust
+use unified_llm::catalog::{get_model_info, list_models, get_latest_model};
+
+let info = get_model_info("claude-opus-4-6");
+let anthropic_models = list_models(Some("anthropic"));
+let best_reasoner = get_latest_model("anthropic", Some("reasoning"));
+```
+
+## Key types
+
+| Type | Description |
+|------|-------------|
+| `Request` | Unified request with model, messages, tools, temperature, etc. |
+| `Response` | Unified response with message, finish reason, usage, rate limit info |
+| `Message` | A message with role, content parts, and optional tool call ID |
+| `ContentPart` | Text, Image, Audio, Document, ToolCall, ToolResult, Thinking |
+| `StreamEvent` | Events for streaming: TextDelta, ToolCallStart/Delta/End, Finish, etc. |
+| `SdkError` | Typed errors with retryability, status codes, and provider error kinds |
+| `GenerateParams` | Builder for the high-level `generate()` function |
+| `GenerateResult` | Result containing response, tool results, total usage, and step history |
+| `ToolDefinition` | Tool name, description, and JSON Schema parameters |
+| `ToolChoice` | Auto, None, Required, or Named tool selection |
+| `Usage` | Token counts including input, output, reasoning, and cache tokens |
+| `RetryPolicy` | Configurable retry with exponential backoff, jitter, and max delay |
+| `ModelInfo` | Metadata about a model (context window, capabilities, costs) |
+
+## Error handling
+
+`SdkError` provides structured error variants with built-in retryability classification:
+
+- **Retryable**: `RateLimit`, `Server`, `Network`, `Stream`, `RequestTimeout`
+- **Non-retryable**: `Authentication`, `AccessDenied`, `InvalidRequest`, `ContextLength`, `Configuration`
+
+The `retry()` function and `generate()` respect `Retry-After` headers and use exponential backoff with jitter.
+
+## Provider-specific options
+
+Pass provider-specific parameters via `provider_options` without losing portability:
+
+```rust
+use unified_llm::types::Request;
+
+let request = Request {
+    provider_options: Some(serde_json::json!({
+        "anthropic": {
+            "thinking": {"type": "enabled", "budget_tokens": 10000},
+            "auto_cache": true
+        },
+        "openai": {
+            "store": true,
+            "previous_response_id": "resp_abc123"
+        },
+        "gemini": {
+            "safetySettings": [
+                {"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"}
+            ]
+        }
+    })),
+    // ...other fields
+    # model: String::new(), messages: vec![], provider: None, tools: None,
+    # tool_choice: None, response_format: None, temperature: None,
+    # top_p: None, max_tokens: None, stop_sequences: None,
+    # reasoning_effort: None, metadata: None,
+};
+```