//! HTTP client supporting custom request and response types. Pass any type into
//! a `send` function or method, and it will return a result of your desired
//! response type. `send` handles request serialization, http messaging, and
//! response deserialization.
//!
//! To keep this crate simple, it is is oriented towards a specific but very
//! common pattern. If your use case meets the following conditions, this crate
//! will work for you:
//! 1. request-response communication
//! 2. async rust functions
//! 3. communicate over http (uses reqwest under the hood)
//! 4. http body is serialized as json
//! 5. status codes outside the 200 range are considered errors
//! 6. request and response types must be serializable and deserializable using
//!    serde
//! 7. the path and HTTP method can be determined from the concrete rust type
//!    used for the request
//!
//! ## Usage
//!
//! ### Typical
//!
//! To use this library, your request and response types must implement
//! serde::Serialize and serde::Deserialize, respectively.
//!
//! To take full advantage of all library features, you can implement `Request`
//! for each of your request types, instantiate a `Client`, and then you can
//! simply invoke `Client::send` to send requests.
//!
//! ```rust
//! let client = Client::new("http://example.com");
//! let response = client.send(MyRequest::new()).await?;
//! ```
//!
//! ### Basic
//!
//! If you don't want to implement Request or create a Client, the most manual
//! and basic way to use this library is by using `send_custom`.
//!
//! ```rust
//! let my_response: MyResponse = send_custom(
//!     "http://example.com/path/to/my/request/",
//!     HttpMethod::Get,
//!     MyRequest::new()
//! )
//! .await?;
//! ```
//!
//! ### Client
//!
//! One downside of the `send_custom` (and `send`) *function* is that it
//! instantiates a client for every request, which is expensive. To improve
//! performance, you can use the `Client::send_custom` (and `Client::send`)
//! *method* instead to re-use an existing client for every request.
//!
//! ```rust
//! let client = Client::default();
//! let my_response: MyResponse = client.send_custom(
//!     "http://example.com/path/to/my/request/",
//!     HttpMethod::Get,
//!     MyRequest::new()
//! )
//! .await?;
//! ```
//!
//! ### Request
//!
//! You may also prefer not to specify metadata about the request every time you
//! send a request, since these things will likely be the same for every request
//! of this type. Describe the request metadata in the type system by
//! implementing the Request trait.
//!
//! ```rust
//! pub trait Request {
//!     type Response;
//!     fn method(&self) -> HttpMethod;
//!     fn path(&self) -> String;
//! }
//! ```
//!
//! This increases design flexibility and reduces boilerplate. See the [API
//! Client Design](#api-client-design) section below for an explanation.
//!
//! If you do not control the crate with the request and response structs, you
//! can implement any traits for them using the newtype pattern, or with a
//! reusable generic wrapper struct.
//!
//! After implementing this trait, you can use the send function and method,
//! which requires the base url to be included, instead of the full url. All
//! other information about how to send the request and response is implied by
//! the type of the input. This still creates a client on every request, so the
//! performance is not optimal if you are sending multiple requests.
//!
//! ```rust
//! let my_response = send("http://example.com", MyRequest::new()).await?;
//! // The type of my_response is determined by the trait's associated type.
//! // It does not need to be inferrable from the calling context.
//! return my_response.some_field
//! ```
//!
//! If you want to send multiple requests, or if you don't want to include the
//! base url when calling `send`, instantiate a Client:
//!
//! ```rust
//! let client = Client::new("http://example.com");
//! let my_response = client.send(MyRequest::new()).await?;
//! ```
//!
//! ### Request groups
//!
//! You can also define request groups. This defines a client type that is
//! explicit about exactly which requests it can handle. The code will not
//! compile if you try to send a request with the wrong client.
//!
//! ```rust
//! request_group!(MyApi { MyRequest1, MyRequest2 });
//! ```
//! ```rust
//! let my_client = Client::<MyApi>::new("http://example.com");
//! let my_response1 = my_client.send(MyRequest1::new()).await?; // works
//! let other_response = my_client.send(OtherRequest::new()).await?; // does not compile
//! ```
//!
//! ### send_to
//!
//! If you want to restrict the request group, but still want to include the url
//! for every call to `send`, `MyClient` has a `send_to` method that can be used
//! with the default client to specify the url at the call-site.
//! ```rust
//! let my_client = Client::<MyApi>::default();
//! let my_response2 = my_client.send_to("http://example.com", MyRequest2::new()).await?; // works
//! let other_response = my_client.send_to("http://example.com", OtherRequest::new()).await?; // does not compile
//! ```
//!
//! The send_to method can also be used to insert a string after the base_url
//! and before the Request path.
//!
//! ```rust
//! let my_client = Client::new("http://example.com");
//! let my_response = my_client.send_to("/api/v2", MyRequest::new()).await?;
//! ```
//!
//! ## Cargo Features
//!
//! Typically, the default features should be fine:
//!
//! ```toml
//! http-typed = "0.3"
//! ```
//!
//! The default features include the full Client implementation, and depend on
//! system tls libraries.
//!
//! All features:
//!
//! - **default** = ["client", "native-tls"]
//! - **client**: Includes the Client implementation described above and depends
//!   on reqwest.
//! - **native-tls**: Depend on dynamically linked system tls libraries.
//! - **rustls-tls**: Statically link all tls dependencies with webpki, no tls
//!   is required in the system.
//!
//!
//! ### No system tls? Use rustls
//!
//! To statically link the tls dependencies, use this:
//!
//! ```toml
//! http-typed = { version = "0.3", default-features = false, features = ["client", "rustls-tls"] }
//! ```
//!
//! ### No Client
//!
//! If you'd like to exclude the `Client` implementation and all of its
//! dependencies on reqwest and tls libraries, use this:
//!
//! ```toml
//! http-typed = { version = "0.3", default-features = false }
//! ```
//!
//! This allows you, as a server developer, to exclude unnecessary dependencies
//! from your server. For example, you may have an API crate with all the
//! request and response structs, which you both import in the server and also
//! make available to clients. You can feature gate the client in your API
//! library:
//!
//! ```toml
//! # api library's Cargo.toml
//!
//! [features]
//! default = ["client"]
//! client = ["http-typed/client"]
//! ```
//!
//! ...and then you can disable it in the server's Cargo.toml. Something like
//! this:
//!
//! ```toml
//! # server binary's Cargo.toml
//!
//! [dependencies]
//! my-api-client = { path = "../client", default-features = false }
//! ```
//!
//! A similar pattern can be used to give clients the option between native-tls
//! and rustls-tls.
//!
//! ## API Client Design
//! Normally, a you might implement a custom client struct to connect to an API,
//! including a custom method for every request. In doing so, you've forced all
//! dependents of the API to make a choice between two options:
//! 1. use the specific custom client struct that was already implemented,
//!    accepting any issues with it.
//! 2. implement a custom client from scratch, re-writing and maintaining all
//!    the details about each request, including what http method to use, what
//!    path to use, how to serialize/deserialized the message, etc.
//!
//! Instead, you can describe the metadata through trait definitions for
//! ultimate flexibility, without locking dependents into a client
//! implementation or needing to implement any custom clients structs.
//! Dependents of the API now have better options:
//! 1. use the Client struct provided by http-typed, accepting any issues with
//!    it. This is easier for you to support because you don't need to worry
//!    about implementation details of sending requests in general. You can just
//!    export or alias the `Client` struct.
//! 2. implement a custom client that can generically handle types implementing
//!    Request by using the data returned by their methods. This is easier for
//!    dependents because they don't need to write any request-specific code.
//!    The Request trait exposes that information without locking them into a
//!    client implementation. Only a single generic request handler is
//!    sufficient.
//!
//! ## Other use cases
//! If your use case does not meet some of the conditions 2-7 described in the
//! introduction, you'll find my other crate useful, which individually
//! generalizes each of those, allowing any of them to be individually
//! customized with minimal boilerplate. It is currently a work in progress, but
//! almost complete. This crate and that crate will be source-code-compatible,
//! meaning the other crate can be used as a drop-in replacement of this one
//! without changing any code, just with more customization available.

#[cfg(feature = "client")]
mod client;

use std::convert::Infallible;

#[cfg(feature = "client")]
pub use client::*;

pub trait Request: Sized {
    // TODO: use when stable: https://github.com/rust-lang/rust/issues/29661
    /// Specify a pre-defined approach to serialize a request body. For example:
    /// - SerdeJson
    /// - NoBody
    type Serializer: SerializeBody<Self>;

    /// Type to deserialize from the http response body
    type Response;

    /// HTTP method that the request will be sent with
    fn method(&self) -> HttpMethod;

    /// String to appended to the end of url when sending this request.
    fn path(&self) -> String;
}

pub struct SerdeJson;
pub struct NoBody;

pub trait SerializeBody<T> {
    type Error;
    fn serialize_body(request: &T) -> Result<Vec<u8>, Self::Error>;
}

impl<T> SerializeBody<T> for SerdeJson
where
    T: serde::Serialize,
{
    type Error = serde_json::error::Error;

    fn serialize_body(request: &T) -> Result<Vec<u8>, Self::Error> {
        Ok(serde_json::to_string(&request)?.into_bytes())
    }
}

impl<T> SerializeBody<T> for NoBody {
    type Error = Infallible;

    fn serialize_body(_: &T) -> Result<Vec<u8>, Self::Error> {
        Ok(vec![])
    }
}

/// Define a request group to constrain which requests can be used with a client.
/// ```ignore
/// request_group!(MyApi { MyRequest1, MyRequest2 });
/// ```
#[macro_export]
macro_rules! request_group {
    ($viz:vis $Name:ident { $($Request:ident),*$(,)? }) => {
        $viz struct $Name;
        $(impl $crate::InRequestGroup<$Name> for $Request {})*
    };
}

/// Indicates that a request is part of a request group. If you use the
/// request_group macro to define the group, it will handle the implementation
/// of this trait automatically.
pub trait InRequestGroup<Group> {}

/// The default group. All requests are in this group.
pub struct All;
impl<T> InRequestGroup<All> for T {}

#[derive(Debug, Clone, Copy)]
pub enum HttpMethod {
    Options,
    Get,
    Post,
    Put,
    Delete,
    Head,
    Trace,
    Connect,
    Patch,
}