zed/crates/copilot_chat/src/copilot_chat.rs

pub mod responses;

use std::path::PathBuf;
use std::sync::Arc;
use std::sync::OnceLock;

use anyhow::Context as _;
use anyhow::{Result, anyhow};
use collections::HashSet;
use fs::Fs;
use futures::{AsyncBufReadExt, AsyncReadExt, StreamExt, io::BufReader, stream::BoxStream};
use gpui::TaskExt;
use gpui::WeakEntity;
use gpui::{App, AsyncApp, Global, prelude::*};
use http_client::HttpRequestExt;
use http_client::{AsyncBody, HttpClient, Method, Request as HttpRequest};
use paths::home_dir;
use serde::{Deserialize, Serialize};

use settings::watch_config_dir;

pub const COPILOT_OAUTH_ENV_VAR: &str = "GH_COPILOT_TOKEN";
const DEFAULT_COPILOT_API_ENDPOINT: &str = "https://api.githubcopilot.com";

#[derive(Default, Clone, Debug, PartialEq)]
pub struct CopilotChatConfiguration {
    pub enterprise_uri: Option<String>,
}

impl CopilotChatConfiguration {
    pub fn oauth_domain(&self) -> String {
        if let Some(enterprise_uri) = &self.enterprise_uri {
            Self::parse_domain(enterprise_uri)
        } else {
            "github.com".to_string()
        }
    }

    pub fn graphql_url(&self) -> String {
        if let Some(enterprise_uri) = &self.enterprise_uri {
            let domain = Self::parse_domain(enterprise_uri);
            format!("https://{}/api/graphql", domain)
        } else {
            "https://api.github.com/graphql".to_string()
        }
    }

    pub fn chat_completions_url(&self, api_endpoint: &str) -> String {
        format!("{}/chat/completions", api_endpoint)
    }

    pub fn responses_url(&self, api_endpoint: &str) -> String {
        format!("{}/responses", api_endpoint)
    }

    pub fn messages_url(&self, api_endpoint: &str) -> String {
        format!("{}/v1/messages", api_endpoint)
    }

    pub fn models_url(&self, api_endpoint: &str) -> String {
        format!("{}/models", api_endpoint)
    }

    fn parse_domain(enterprise_uri: &str) -> String {
        let uri = enterprise_uri.trim_end_matches('/');

        if let Some(domain) = uri.strip_prefix("https://") {
            domain.split('/').next().unwrap_or(domain).to_string()
        } else if let Some(domain) = uri.strip_prefix("http://") {
            domain.split('/').next().unwrap_or(domain).to_string()
        } else {
            uri.split('/').next().unwrap_or(uri).to_string()
        }
    }
}

#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
#[serde(rename_all = "lowercase")]
pub enum Role {
    User,
    Assistant,
    System,
}

#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
pub enum ChatLocation {
    #[default]
    Panel,
    Editor,
    EditingSession,
    Terminal,
    Agent,
    Other,
}

impl ChatLocation {
    pub fn to_intent_string(self) -> &'static str {
        match self {
            ChatLocation::Panel => "conversation-panel",
            ChatLocation::Editor => "conversation-inline",
            ChatLocation::EditingSession => "conversation-edits",
            ChatLocation::Terminal => "conversation-terminal",
            ChatLocation::Agent => "conversation-agent",
            ChatLocation::Other => "conversation-other",
        }
    }
}

#[derive(Deserialize, Serialize, Debug, Clone, PartialEq)]
pub enum ModelSupportedEndpoint {
    #[serde(rename = "/chat/completions")]
    ChatCompletions,
    #[serde(rename = "/responses")]
    Responses,
    #[serde(rename = "/v1/messages")]
    Messages,
    /// Unknown endpoint that we don't explicitly support yet
    #[serde(other)]
    Unknown,
}

#[derive(Deserialize)]
struct ModelSchema {
    #[serde(deserialize_with = "deserialize_models_skip_errors")]
    data: Vec<Model>,
}

fn deserialize_models_skip_errors<'de, D>(deserializer: D) -> Result<Vec<Model>, D::Error>
where
    D: serde::Deserializer<'de>,
{
    let raw_values = Vec::<serde_json::Value>::deserialize(deserializer)?;
    let models = raw_values
        .into_iter()
        .filter_map(|value| match serde_json::from_value::<Model>(value) {
            Ok(model) => Some(model),
            Err(err) => {
                log::warn!("GitHub Copilot Chat model failed to deserialize: {:?}", err);
                None
            }
        })
        .collect();

    Ok(models)
}

#[derive(Clone, Serialize, Deserialize, Debug, PartialEq)]
pub struct Model {
    billing: ModelBilling,
    capabilities: ModelCapabilities,
    id: String,
    name: String,
    policy: Option<ModelPolicy>,
    vendor: ModelVendor,
    is_chat_default: bool,
    // The model with this value true is selected by VSCode copilot if a premium request limit is
    // reached. Zed does not currently implement this behaviour
    is_chat_fallback: bool,
    model_picker_enabled: bool,
    #[serde(default)]
    supported_endpoints: Vec<ModelSupportedEndpoint>,
}

#[derive(Clone, Serialize, Deserialize, Debug, PartialEq)]
struct ModelBilling {
    is_premium: bool,
    multiplier: f64,
    // List of plans a model is restricted to
    // Field is not present if a model is available for all plans
    #[serde(default)]
    restricted_to: Option<Vec<String>>,
}

#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
struct ModelCapabilities {
    family: String,
    #[serde(default)]
    limits: ModelLimits,
    supports: ModelSupportedFeatures,
    #[serde(rename = "type")]
    model_type: String,
    #[serde(default)]
    tokenizer: Option<String>,
}

#[derive(Default, Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
struct ModelLimits {
    #[serde(default)]
    max_context_window_tokens: usize,
    #[serde(default)]
    max_output_tokens: usize,
    #[serde(default)]
    max_prompt_tokens: u64,
}

#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
struct ModelPolicy {
    state: String,
}

#[derive(Clone, Serialize, Deserialize, Debug, Eq, PartialEq)]
struct ModelSupportedFeatures {
    #[serde(default)]
    streaming: bool,
    #[serde(default)]
    tool_calls: bool,
    #[serde(default)]
    parallel_tool_calls: bool,
    #[serde(default)]
    vision: bool,
    #[serde(default)]
    thinking: bool,
    #[serde(default)]
    adaptive_thinking: bool,
    #[serde(default)]
    max_thinking_budget: Option<u32>,
    #[serde(default)]
    min_thinking_budget: Option<u32>,
    #[serde(default)]
    reasoning_effort: Vec<String>,
}

#[derive(Clone, Copy, Serialize, Deserialize, Debug, Eq, PartialEq)]
pub enum ModelVendor {
    // Azure OpenAI should have no functional difference from OpenAI in Copilot Chat
    #[serde(alias = "Azure OpenAI")]
    OpenAI,
    Google,
    Anthropic,
    #[serde(rename = "xAI")]
    XAI,
    /// Unknown vendor that we don't explicitly support yet
    #[serde(other)]
    Unknown,
}

#[derive(Serialize, Deserialize, Debug, Eq, PartialEq, Clone)]
#[serde(tag = "type")]
pub enum ChatMessagePart {
    #[serde(rename = "text")]
    Text { text: String },
    #[serde(rename = "image_url")]
    Image { image_url: ImageUrl },
}

#[derive(Serialize, Deserialize, Debug, Eq, PartialEq, Clone)]
pub struct ImageUrl {
    pub url: String,
}

impl Model {
    pub fn uses_streaming(&self) -> bool {
        self.capabilities.supports.streaming
    }

    pub fn id(&self) -> &str {
        self.id.as_str()
    }

    pub fn display_name(&self) -> &str {
        self.name.as_str()
    }

    pub fn max_token_count(&self) -> u64 {
        self.capabilities.limits.max_context_window_tokens as u64
    }

    pub fn max_output_tokens(&self) -> usize {
        self.capabilities.limits.max_output_tokens
    }

    pub fn supports_tools(&self) -> bool {
        self.capabilities.supports.tool_calls
    }

    pub fn vendor(&self) -> ModelVendor {
        self.vendor
    }

    pub fn supports_vision(&self) -> bool {
        self.capabilities.supports.vision
    }

    pub fn supports_parallel_tool_calls(&self) -> bool {
        self.capabilities.supports.parallel_tool_calls
    }

    pub fn tokenizer(&self) -> Option<&str> {
        self.capabilities.tokenizer.as_deref()
    }

    pub fn supports_response(&self) -> bool {
        self.supported_endpoints
            .contains(&ModelSupportedEndpoint::Responses)
    }

    pub fn supports_messages(&self) -> bool {
        self.supported_endpoints
            .contains(&ModelSupportedEndpoint::Messages)
    }

    pub fn supports_thinking(&self) -> bool {
        self.capabilities.supports.thinking
    }

    pub fn supports_adaptive_thinking(&self) -> bool {
        self.capabilities.supports.adaptive_thinking
    }

    pub fn can_think(&self) -> bool {
        self.supports_thinking()
            || self.supports_adaptive_thinking()
            || self.max_thinking_budget().is_some()
            || !self.reasoning_effort_levels().is_empty()
    }

    pub fn max_thinking_budget(&self) -> Option<u32> {
        self.capabilities.supports.max_thinking_budget
    }

    pub fn min_thinking_budget(&self) -> Option<u32> {
        self.capabilities.supports.min_thinking_budget
    }

    pub fn reasoning_effort_levels(&self) -> &[String] {
        &self.capabilities.supports.reasoning_effort
    }

    pub fn family(&self) -> &str {
        &self.capabilities.family
    }

    pub fn multiplier(&self) -> f64 {
        self.billing.multiplier
    }
}

#[derive(Serialize, Deserialize)]
pub struct Request {
    pub n: usize,
    pub stream: bool,
    pub temperature: f32,
    pub model: String,
    pub messages: Vec<ChatMessage>,
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub tools: Vec<Tool>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub tool_choice: Option<ToolChoice>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub thinking_budget: Option<u32>,
}

#[derive(Serialize, Deserialize)]
pub struct Function {
    pub name: String,
    pub description: String,
    pub parameters: serde_json::Value,
}

#[derive(Serialize, Deserialize)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum Tool {
    Function { function: Function },
}

#[derive(Serialize, Deserialize, Debug)]
#[serde(rename_all = "lowercase")]
pub enum ToolChoice {
    Auto,
    Required,
    None,
}

#[derive(Serialize, Deserialize, Debug)]
#[serde(tag = "role", rename_all = "lowercase")]
pub enum ChatMessage {
    Assistant {
        content: ChatMessageContent,
        #[serde(default, skip_serializing_if = "Vec::is_empty")]
        tool_calls: Vec<ToolCall>,
        #[serde(default, skip_serializing_if = "Option::is_none")]
        reasoning_opaque: Option<String>,
        #[serde(default, skip_serializing_if = "Option::is_none")]
        reasoning_text: Option<String>,
    },
    User {
        content: ChatMessageContent,
    },
    System {
        content: String,
    },
    Tool {
        content: ChatMessageContent,
        tool_call_id: String,
    },
}

#[derive(Debug, Serialize, Deserialize)]
#[serde(untagged)]
pub enum ChatMessageContent {
    Plain(String),
    Multipart(Vec<ChatMessagePart>),
}

impl ChatMessageContent {
    pub fn empty() -> Self {
        ChatMessageContent::Multipart(vec![])
    }
}

impl From<Vec<ChatMessagePart>> for ChatMessageContent {
    fn from(mut parts: Vec<ChatMessagePart>) -> Self {
        if let [ChatMessagePart::Text { text }] = parts.as_mut_slice() {
            ChatMessageContent::Plain(std::mem::take(text))
        } else {
            ChatMessageContent::Multipart(parts)
        }
    }
}

impl From<String> for ChatMessageContent {
    fn from(text: String) -> Self {
        ChatMessageContent::Plain(text)
    }
}

#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct ToolCall {
    pub id: String,
    #[serde(flatten)]
    pub content: ToolCallContent,
}

#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
#[serde(tag = "type", rename_all = "lowercase")]
pub enum ToolCallContent {
    Function { function: FunctionContent },
}

#[derive(Serialize, Deserialize, Debug, Eq, PartialEq)]
pub struct FunctionContent {
    pub name: String,
    pub arguments: String,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub thought_signature: Option<String>,
}

#[derive(Deserialize, Debug)]
#[serde(tag = "type", rename_all = "snake_case")]
pub struct ResponseEvent {
    pub choices: Vec<ResponseChoice>,
    pub id: String,
    pub usage: Option<Usage>,
}

#[derive(Deserialize, Debug)]
pub struct Usage {
    pub completion_tokens: u64,
    pub prompt_tokens: u64,
    pub total_tokens: u64,
}

#[derive(Debug, Deserialize)]
pub struct ResponseChoice {
    pub index: Option<usize>,
    pub finish_reason: Option<String>,
    pub delta: Option<ResponseDelta>,
    pub message: Option<ResponseDelta>,
}

#[derive(Debug, Deserialize)]
pub struct ResponseDelta {
    pub content: Option<String>,
    pub role: Option<Role>,
    #[serde(default)]
    pub tool_calls: Vec<ToolCallChunk>,
    pub reasoning_opaque: Option<String>,
    pub reasoning_text: Option<String>,
}
#[derive(Deserialize, Debug, Eq, PartialEq)]
pub struct ToolCallChunk {
    pub index: Option<usize>,
    pub id: Option<String>,
    pub function: Option<FunctionChunk>,
}

#[derive(Deserialize, Debug, Eq, PartialEq)]
pub struct FunctionChunk {
    pub name: Option<String>,
    pub arguments: Option<String>,
    pub thought_signature: Option<String>,
}

struct GlobalCopilotChat(gpui::Entity<CopilotChat>);

impl Global for GlobalCopilotChat {}

pub struct CopilotChat {
    oauth_token: Option<String>,
    api_endpoint: Option<String>,
    configuration: CopilotChatConfiguration,
    models: Option<Vec<Model>>,
    client: Arc<dyn HttpClient>,
}

pub fn init(
    fs: Arc<dyn Fs>,
    client: Arc<dyn HttpClient>,
    configuration: CopilotChatConfiguration,
    cx: &mut App,
) {
    let copilot_chat = cx.new(|cx| CopilotChat::new(fs, client, configuration, cx));
    cx.set_global(GlobalCopilotChat(copilot_chat));
}

pub fn copilot_chat_config_dir() -> &'static PathBuf {
    static COPILOT_CHAT_CONFIG_DIR: OnceLock<PathBuf> = OnceLock::new();

    COPILOT_CHAT_CONFIG_DIR.get_or_init(|| {
        let config_dir = if cfg!(target_os = "windows") {
            dirs::data_local_dir().expect("failed to determine LocalAppData directory")
        } else {
            std::env::var("XDG_CONFIG_HOME")
                .map(PathBuf::from)
                .unwrap_or_else(|_| home_dir().join(".config"))
        };

        config_dir.join("github-copilot")
    })
}

fn copilot_chat_config_paths() -> [PathBuf; 2] {
    let base_dir = copilot_chat_config_dir();
    [base_dir.join("hosts.json"), base_dir.join("apps.json")]
}

impl CopilotChat {
    pub fn global(cx: &App) -> Option<gpui::Entity<Self>> {
        cx.try_global::<GlobalCopilotChat>()
            .map(|model| model.0.clone())
    }

    fn new(
        fs: Arc<dyn Fs>,
        client: Arc<dyn HttpClient>,
        configuration: CopilotChatConfiguration,
        cx: &mut Context<Self>,
    ) -> Self {
        let config_paths: HashSet<PathBuf> = copilot_chat_config_paths().into_iter().collect();
        let dir_path = copilot_chat_config_dir();

        cx.spawn(async move |this, cx| {
            let mut parent_watch_rx = watch_config_dir(
                cx.background_executor(),
                fs.clone(),
                dir_path.clone(),
                config_paths,
            );
            while let Some(contents) = parent_watch_rx.next().await {
                let oauth_domain =
                    this.read_with(cx, |this, _| this.configuration.oauth_domain())?;
                let oauth_token = extract_oauth_token(contents, &oauth_domain);

                this.update(cx, |this, cx| {
                    this.oauth_token = oauth_token.clone();
                    cx.notify();
                })?;

                if oauth_token.is_some() {
                    Self::update_models(&this, cx).await?;
                }
            }
            anyhow::Ok(())
        })
        .detach_and_log_err(cx);

        let this = Self {
            oauth_token: std::env::var(COPILOT_OAUTH_ENV_VAR).ok(),
            api_endpoint: None,
            models: None,
            configuration,
            client,
        };

        if this.oauth_token.is_some() {
            cx.spawn(async move |this, cx| Self::update_models(&this, cx).await)
                .detach_and_log_err(cx);
        }

        this
    }

    async fn update_models(this: &WeakEntity<Self>, cx: &mut AsyncApp) -> Result<()> {
        let (oauth_token, client, configuration) = this.read_with(cx, |this, _| {
            (
                this.oauth_token.clone(),
                this.client.clone(),
                this.configuration.clone(),
            )
        })?;

        let oauth_token = oauth_token
            .ok_or_else(|| anyhow!("OAuth token is missing while updating Copilot Chat models"))?;

        let api_endpoint =
            Self::resolve_api_endpoint(&this, &oauth_token, &configuration, &client, cx).await?;

        let models_url = configuration.models_url(&api_endpoint);
        let models = get_models(models_url.into(), oauth_token, client.clone()).await?;

        this.update(cx, |this, cx| {
            this.models = Some(models);
            cx.notify();
        })?;
        anyhow::Ok(())
    }

    pub fn is_authenticated(&self) -> bool {
        self.oauth_token.is_some()
    }

    pub fn models(&self) -> Option<&[Model]> {
        self.models.as_deref()
    }

    pub async fn stream_completion(
        request: Request,
        location: ChatLocation,
        is_user_initiated: bool,
        mut cx: AsyncApp,
    ) -> Result<BoxStream<'static, Result<ResponseEvent>>> {
        let (client, oauth_token, api_endpoint, configuration) =
            Self::get_auth_details(&mut cx).await?;

        let api_url = configuration.chat_completions_url(&api_endpoint);
        stream_completion(
            client.clone(),
            oauth_token,
            api_url.into(),
            request,
            is_user_initiated,
            location,
        )
        .await
    }

    pub async fn stream_response(
        request: responses::Request,
        location: ChatLocation,
        is_user_initiated: bool,
        mut cx: AsyncApp,
    ) -> Result<BoxStream<'static, Result<responses::StreamEvent>>> {
        let (client, oauth_token, api_endpoint, configuration) =
            Self::get_auth_details(&mut cx).await?;

        let api_url = configuration.responses_url(&api_endpoint);
        responses::stream_response(
            client.clone(),
            oauth_token,
            api_url,
            request,
            is_user_initiated,
            location,
        )
        .await
    }

    pub async fn stream_messages(
        body: String,
        location: ChatLocation,
        is_user_initiated: bool,
        anthropic_beta: Option<String>,
        mut cx: AsyncApp,
    ) -> Result<BoxStream<'static, Result<anthropic::Event, anthropic::AnthropicError>>> {
        let (client, oauth_token, api_endpoint, configuration) =
            Self::get_auth_details(&mut cx).await?;

        let api_url = configuration.messages_url(&api_endpoint);
        stream_messages(
            client.clone(),
            oauth_token,
            api_url,
            body,
            is_user_initiated,
            location,
            anthropic_beta,
        )
        .await
    }

    async fn get_auth_details(
        cx: &mut AsyncApp,
    ) -> Result<(
        Arc<dyn HttpClient>,
        String,
        String,
        CopilotChatConfiguration,
    )> {
        let this = cx
            .update(|cx| Self::global(cx))
            .context("Copilot chat is not enabled")?;

        let (oauth_token, api_endpoint, client, configuration) = this.read_with(cx, |this, _| {
            (
                this.oauth_token.clone(),
                this.api_endpoint.clone(),
                this.client.clone(),
                this.configuration.clone(),
            )
        });

        let oauth_token = oauth_token.context("No OAuth token available")?;

        let api_endpoint = match api_endpoint {
            Some(endpoint) => endpoint,
            None => {
                let weak = this.downgrade();
                Self::resolve_api_endpoint(&weak, &oauth_token, &configuration, &client, cx).await?
            }
        };

        Ok((client, oauth_token, api_endpoint, configuration))
    }

    async fn resolve_api_endpoint(
        this: &WeakEntity<Self>,
        oauth_token: &str,
        configuration: &CopilotChatConfiguration,
        client: &Arc<dyn HttpClient>,
        cx: &mut AsyncApp,
    ) -> Result<String> {
        let api_endpoint = match discover_api_endpoint(oauth_token, configuration, client).await {
            Ok(endpoint) => endpoint,
            Err(error) => {
                log::warn!(
                    "Failed to discover Copilot API endpoint via GraphQL, \
                         falling back to {DEFAULT_COPILOT_API_ENDPOINT}: {error:#}"
                );
                DEFAULT_COPILOT_API_ENDPOINT.to_string()
            }
        };

        this.update(cx, |this, cx| {
            this.api_endpoint = Some(api_endpoint.clone());
            cx.notify();
        })?;

        Ok(api_endpoint)
    }

    pub fn set_configuration(
        &mut self,
        configuration: CopilotChatConfiguration,
        cx: &mut Context<Self>,
    ) {
        let same_configuration = self.configuration == configuration;
        self.configuration = configuration;
        if !same_configuration {
            self.api_endpoint = None;
            cx.spawn(async move |this, cx| {
                Self::update_models(&this, cx).await?;
                Ok::<_, anyhow::Error>(())
            })
            .detach();
        }
    }
}

async fn get_models(
    models_url: Arc<str>,
    oauth_token: String,
    client: Arc<dyn HttpClient>,
) -> Result<Vec<Model>> {
    let all_models = request_models(models_url, oauth_token, client).await?;

    let mut models: Vec<Model> = all_models
        .into_iter()
        .filter(|model| {
            model.model_picker_enabled
                && model.capabilities.model_type.as_str() == "chat"
                && model
                    .policy
                    .as_ref()
                    .is_none_or(|policy| policy.state == "enabled")
        })
        .collect();

    if let Some(default_model_position) = models.iter().position(|model| model.is_chat_default) {
        let default_model = models.remove(default_model_position);
        models.insert(0, default_model);
    }

    Ok(models)
}

#[derive(Deserialize)]
struct GraphQLResponse {
    data: Option<GraphQLData>,
}

#[derive(Deserialize)]
struct GraphQLData {
    viewer: GraphQLViewer,
}

#[derive(Deserialize)]
struct GraphQLViewer {
    #[serde(rename = "copilotEndpoints")]
    copilot_endpoints: GraphQLCopilotEndpoints,
}

#[derive(Deserialize)]
struct GraphQLCopilotEndpoints {
    api: String,
}

pub(crate) async fn discover_api_endpoint(
    oauth_token: &str,
    configuration: &CopilotChatConfiguration,
    client: &Arc<dyn HttpClient>,
) -> Result<String> {
    let graphql_url = configuration.graphql_url();
    let query = serde_json::json!({
        "query": "query { viewer { copilotEndpoints { api } } }"
    });

    let request = HttpRequest::builder()
        .method(Method::POST)
        .uri(graphql_url.as_str())
        .header("Authorization", format!("Bearer {}", oauth_token))
        .header("Content-Type", "application/json")
        .body(AsyncBody::from(serde_json::to_string(&query)?))?;

    let mut response = client.send(request).await?;

    anyhow::ensure!(
        response.status().is_success(),
        "GraphQL endpoint discovery failed: {}",
        response.status()
    );

    let mut body = Vec::new();
    response.body_mut().read_to_end(&mut body).await?;
    let body_str = std::str::from_utf8(&body)?;

    let parsed: GraphQLResponse = serde_json::from_str(body_str)
        .context("Failed to parse GraphQL response for Copilot endpoint discovery")?;

    let data = parsed
        .data
        .context("GraphQL response contained no data field")?;

    Ok(data.viewer.copilot_endpoints.api)
}

pub(crate) fn copilot_request_headers(
    builder: http_client::Builder,
    oauth_token: &str,
    is_user_initiated: Option<bool>,
    location: Option<ChatLocation>,
) -> http_client::Builder {
    builder
        .header("Authorization", format!("Bearer {}", oauth_token))
        .header("Content-Type", "application/json")
        .header(
            "Editor-Version",
            format!(
                "Zed/{}",
                option_env!("CARGO_PKG_VERSION").unwrap_or("unknown")
            ),
        )
        .header("X-GitHub-Api-Version", "2025-10-01")
        .when_some(is_user_initiated, |builder, is_user_initiated| {
            builder.header(
                "X-Initiator",
                if is_user_initiated { "user" } else { "agent" },
            )
        })
        .when_some(location, |builder, loc| {
            let interaction_type = loc.to_intent_string();
            builder
                .header("X-Interaction-Type", interaction_type)
                .header("OpenAI-Intent", interaction_type)
        })
}

async fn request_models(
    models_url: Arc<str>,
    oauth_token: String,
    client: Arc<dyn HttpClient>,
) -> Result<Vec<Model>> {
    let request_builder = copilot_request_headers(
        HttpRequest::builder()
            .method(Method::GET)
            .uri(models_url.as_ref()),
        &oauth_token,
        None,
        None,
    );

    let request = request_builder.body(AsyncBody::empty())?;

    let mut response = client.send(request).await?;

    anyhow::ensure!(
        response.status().is_success(),
        "Failed to request models: {}",
        response.status()
    );
    let mut body = Vec::new();
    response.body_mut().read_to_end(&mut body).await?;

    let body_str = std::str::from_utf8(&body)?;

    let models = serde_json::from_str::<ModelSchema>(body_str)?.data;

    Ok(models)
}

fn extract_oauth_token(contents: String, domain: &str) -> Option<String> {
    serde_json::from_str::<serde_json::Value>(&contents)
        .map(|v| {
            v.as_object().and_then(|obj| {
                obj.iter().find_map(|(key, value)| {
                    if key.starts_with(domain) {
                        value["oauth_token"].as_str().map(|v| v.to_string())
                    } else {
                        None
                    }
                })
            })
        })
        .ok()
        .flatten()
}

async fn stream_completion(
    client: Arc<dyn HttpClient>,
    oauth_token: String,
    completion_url: Arc<str>,
    request: Request,
    is_user_initiated: bool,
    location: ChatLocation,
) -> Result<BoxStream<'static, Result<ResponseEvent>>> {
    let is_vision_request = request.messages.iter().any(|message| match message {
        ChatMessage::User { content }
        | ChatMessage::Assistant { content, .. }
        | ChatMessage::Tool { content, .. } => {
            matches!(content, ChatMessageContent::Multipart(parts) if parts.iter().any(|part| matches!(part, ChatMessagePart::Image { .. })))
        }
        _ => false,
    });

    let request_builder = copilot_request_headers(
        HttpRequest::builder()
            .method(Method::POST)
            .uri(completion_url.as_ref()),
        &oauth_token,
        Some(is_user_initiated),
        Some(location),
    )
    .when(is_vision_request, |builder| {
        builder.header("Copilot-Vision-Request", is_vision_request.to_string())
    });

    let is_streaming = request.stream;

    let json = serde_json::to_string(&request)?;
    let request = request_builder.body(AsyncBody::from(json))?;
    let mut response = client.send(request).await?;

    if !response.status().is_success() {
        let mut body = Vec::new();
        response.body_mut().read_to_end(&mut body).await?;
        let body_str = std::str::from_utf8(&body)?;
        anyhow::bail!(
            "Failed to connect to API: {} {}",
            response.status(),
            body_str
        );
    }

    if is_streaming {
        let reader = BufReader::new(response.into_body());
        Ok(reader
            .lines()
            .filter_map(|line| async move {
                match line {
                    Ok(line) => {
                        let line = line.strip_prefix("data: ")?;
                        if line.starts_with("[DONE]") {
                            return None;
                        }

                        match serde_json::from_str::<ResponseEvent>(line) {
                            Ok(response) => {
                                if response.choices.is_empty() {
                                    None
                                } else {
                                    Some(Ok(response))
                                }
                            }
                            Err(error) => Some(Err(anyhow!(error))),
                        }
                    }
                    Err(error) => Some(Err(anyhow!(error))),
                }
            })
            .boxed())
    } else {
        let mut body = Vec::new();
        response.body_mut().read_to_end(&mut body).await?;
        let body_str = std::str::from_utf8(&body)?;
        let response: ResponseEvent = serde_json::from_str(body_str)?;

        Ok(futures::stream::once(async move { Ok(response) }).boxed())
    }
}

async fn stream_messages(
    client: Arc<dyn HttpClient>,
    oauth_token: String,
    api_url: String,
    body: String,
    is_user_initiated: bool,
    location: ChatLocation,
    anthropic_beta: Option<String>,
) -> Result<BoxStream<'static, Result<anthropic::Event, anthropic::AnthropicError>>> {
    let mut request_builder = copilot_request_headers(
        HttpRequest::builder().method(Method::POST).uri(&api_url),
        &oauth_token,
        Some(is_user_initiated),
        Some(location),
    );

    if let Some(beta) = &anthropic_beta {
        request_builder = request_builder.header("anthropic-beta", beta.as_str());
    }

    let request = request_builder.body(AsyncBody::from(body))?;
    let mut response = client.send(request).await?;

    if !response.status().is_success() {
        let mut body = String::new();
        response.body_mut().read_to_string(&mut body).await?;
        anyhow::bail!("Failed to connect to API: {} {}", response.status(), body);
    }

    let reader = BufReader::new(response.into_body());
    Ok(reader
        .lines()
        .filter_map(|line| async move {
            match line {
                Ok(line) => {
                    let line = line
                        .strip_prefix("data: ")
                        .or_else(|| line.strip_prefix("data:"))?;
                    if line.starts_with("[DONE]") || line.is_empty() {
                        return None;
                    }
                    match serde_json::from_str(line) {
                        Ok(event) => Some(Ok(event)),
                        Err(error) => {
                            log::error!(
                                "Failed to parse Copilot messages stream event: `{}`\nResponse: `{}`",
                                error,
                                line,
                            );
                            Some(Err(anthropic::AnthropicError::DeserializeResponse(error)))
                        }
                    }
                }
                Err(error) => Some(Err(anthropic::AnthropicError::ReadResponse(error))),
            }
        })
        .boxed())
}

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

    #[test]
    fn test_resilient_model_schema_deserialize() {
        let json = r#"{
              "data": [
                {
                  "billing": {
                    "is_premium": false,
                    "multiplier": 0
                  },
                  "capabilities": {
                    "family": "gpt-4",
                    "limits": {
                      "max_context_window_tokens": 32768,
                      "max_output_tokens": 4096,
                      "max_prompt_tokens": 32768
                    },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "tokenizer": "cl100k_base",
                    "type": "chat"
                  },
                  "id": "gpt-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": false,
                  "name": "GPT 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Azure OpenAI",
                  "version": "gpt-4-0613"
                },
                {
                    "some-unknown-field": 123
                },
                {
                  "billing": {
                    "is_premium": true,
                    "multiplier": 1,
                    "restricted_to": [
                      "pro",
                      "pro_plus",
                      "business",
                      "enterprise"
                    ]
                  },
                  "capabilities": {
                    "family": "claude-3.7-sonnet",
                    "limits": {
                      "max_context_window_tokens": 200000,
                      "max_output_tokens": 16384,
                      "max_prompt_tokens": 90000,
                      "vision": {
                        "max_prompt_image_size": 3145728,
                        "max_prompt_images": 1,
                        "supported_media_types": ["image/jpeg", "image/png", "image/webp"]
                      }
                    },
                    "object": "model_capabilities",
                    "supports": {
                      "parallel_tool_calls": true,
                      "streaming": true,
                      "tool_calls": true,
                      "vision": true
                    },
                    "tokenizer": "o200k_base",
                    "type": "chat"
                  },
                  "id": "claude-3.7-sonnet",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude 3.7 Sonnet",
                  "object": "model",
                  "policy": {
                    "state": "enabled",
                    "terms": "Enable access to the latest Claude 3.7 Sonnet model from Anthropic. [Learn more about how GitHub Copilot serves Claude 3.7 Sonnet](https://docs.github.com/copilot/using-github-copilot/using-claude-sonnet-in-github-copilot)."
                  },
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-3.7-sonnet"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        assert_eq!(schema.data.len(), 2);
        assert_eq!(schema.data[0].id, "gpt-4");
        assert_eq!(schema.data[1].id, "claude-3.7-sonnet");
    }

    #[test]
    fn test_unknown_vendor_resilience() {
        let json = r#"{
              "data": [
                {
                  "billing": {
                    "is_premium": false,
                    "multiplier": 1
                  },
                  "capabilities": {
                    "family": "future-model",
                    "limits": {
                      "max_context_window_tokens": 128000,
                      "max_output_tokens": 8192,
                      "max_prompt_tokens": 120000
                    },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "future-model-v1",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Future Model v1",
                  "object": "model",
                  "preview": false,
                  "vendor": "SomeNewVendor",
                  "version": "v1.0"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        assert_eq!(schema.data.len(), 1);
        assert_eq!(schema.data[0].id, "future-model-v1");
        assert_eq!(schema.data[0].vendor, ModelVendor::Unknown);
    }

    #[test]
    fn test_max_token_count_returns_context_window_not_prompt_tokens() {
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4"
                },
                {
                  "billing": { "is_premium": false, "multiplier": 1 },
                  "capabilities": {
                    "family": "gpt-4o",
                    "limits": { "max_context_window_tokens": 128000, "max_output_tokens": 16384, "max_prompt_tokens": 110000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "gpt-4o",
                  "is_chat_default": true,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "GPT-4o",
                  "object": "model",
                  "preview": false,
                  "vendor": "Azure OpenAI",
                  "version": "gpt-4o"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        // max_token_count() should return context window (200000), not prompt tokens (90000)
        assert_eq!(schema.data[0].max_token_count(), 200000);

        // GPT-4o should return 128000 (context window), not 110000 (prompt tokens)
        assert_eq!(schema.data[1].max_token_count(), 128000);
    }

    #[test]
    fn test_models_with_pending_policy_deserialize() {
        // This test verifies that models with policy states other than "enabled"
        // (such as "pending" or "requires_consent") are properly deserialized.
        // Note: These models will be filtered out by get_models() and won't appear
        // in the model picker until the user enables them on GitHub.
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "policy": {
                    "state": "pending",
                    "terms": "Enable access to Claude models from Anthropic."
                  },
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-opus-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-opus-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Opus 4",
                  "object": "model",
                  "policy": {
                    "state": "requires_consent",
                    "terms": "Enable access to Claude models from Anthropic."
                  },
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-opus-4"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        // Both models should deserialize successfully (filtering happens in get_models)
        assert_eq!(schema.data.len(), 2);
        assert_eq!(schema.data[0].id, "claude-sonnet-4");
        assert_eq!(schema.data[1].id, "claude-opus-4");
    }

    #[test]
    fn test_multiple_anthropic_models_preserved() {
        // This test verifies that multiple Claude models from Anthropic
        // are all preserved and not incorrectly deduplicated.
        // This was the root cause of issue #47540.
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-opus-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-opus-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Opus 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-opus-4"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4.5",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4.5",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4.5",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4.5"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        // All three Anthropic models should be preserved
        assert_eq!(schema.data.len(), 3);
        assert_eq!(schema.data[0].id, "claude-sonnet-4");
        assert_eq!(schema.data[1].id, "claude-opus-4");
        assert_eq!(schema.data[2].id, "claude-sonnet-4.5");
    }

    #[test]
    fn test_models_with_same_family_both_preserved() {
        // Test that models sharing the same family (e.g., thinking variants)
        // are both preserved in the model list.
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4-thinking",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4 (Thinking)",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4-thinking"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        // Both models should be preserved even though they share the same family
        assert_eq!(schema.data.len(), 2);
        assert_eq!(schema.data[0].id, "claude-sonnet-4");
        assert_eq!(schema.data[1].id, "claude-sonnet-4-thinking");
    }

    #[test]
    fn test_mixed_vendor_models_all_preserved() {
        // Test that models from different vendors are all preserved.
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": false, "multiplier": 1 },
                  "capabilities": {
                    "family": "gpt-4o",
                    "limits": { "max_context_window_tokens": 128000, "max_output_tokens": 16384, "max_prompt_tokens": 110000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "gpt-4o",
                  "is_chat_default": true,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "GPT-4o",
                  "object": "model",
                  "preview": false,
                  "vendor": "Azure OpenAI",
                  "version": "gpt-4o"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4"
                },
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "gemini-2.0-flash",
                    "limits": { "max_context_window_tokens": 1000000, "max_output_tokens": 8192, "max_prompt_tokens": 900000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "gemini-2.0-flash",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Gemini 2.0 Flash",
                  "object": "model",
                  "preview": false,
                  "vendor": "Google",
                  "version": "gemini-2.0-flash"
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        // All three models from different vendors should be preserved
        assert_eq!(schema.data.len(), 3);
        assert_eq!(schema.data[0].id, "gpt-4o");
        assert_eq!(schema.data[1].id, "claude-sonnet-4");
        assert_eq!(schema.data[2].id, "gemini-2.0-flash");
    }

    #[test]
    fn test_model_with_messages_endpoint_deserializes() {
        // Anthropic Claude models use /v1/messages endpoint.
        // This test verifies such models deserialize correctly (issue #47540 root cause).
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "claude-sonnet-4",
                    "limits": { "max_context_window_tokens": 200000, "max_output_tokens": 16384, "max_prompt_tokens": 90000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "claude-sonnet-4",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Claude Sonnet 4",
                  "object": "model",
                  "preview": false,
                  "vendor": "Anthropic",
                  "version": "claude-sonnet-4",
                  "supported_endpoints": ["/v1/messages"]
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        assert_eq!(schema.data.len(), 1);
        assert_eq!(schema.data[0].id, "claude-sonnet-4");
        assert_eq!(
            schema.data[0].supported_endpoints,
            vec![ModelSupportedEndpoint::Messages]
        );
    }

    #[test]
    fn test_model_with_unknown_endpoint_deserializes() {
        // Future-proofing: unknown endpoints should deserialize to Unknown variant
        // instead of causing the entire model to fail deserialization.
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": false, "multiplier": 1 },
                  "capabilities": {
                    "family": "future-model",
                    "limits": { "max_context_window_tokens": 128000, "max_output_tokens": 8192, "max_prompt_tokens": 120000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "future-model-v2",
                  "is_chat_default": false,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "Future Model v2",
                  "object": "model",
                  "preview": false,
                  "vendor": "OpenAI",
                  "version": "v2.0",
                  "supported_endpoints": ["/v2/completions", "/chat/completions"]
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        assert_eq!(schema.data.len(), 1);
        assert_eq!(schema.data[0].id, "future-model-v2");
        assert_eq!(
            schema.data[0].supported_endpoints,
            vec![
                ModelSupportedEndpoint::Unknown,
                ModelSupportedEndpoint::ChatCompletions
            ]
        );
    }

    #[test]
    fn test_model_with_multiple_endpoints() {
        // Test model with multiple supported endpoints (common for newer models).
        let json = r#"{
              "data": [
                {
                  "billing": { "is_premium": true, "multiplier": 1 },
                  "capabilities": {
                    "family": "gpt-4o",
                    "limits": { "max_context_window_tokens": 128000, "max_output_tokens": 16384, "max_prompt_tokens": 110000 },
                    "object": "model_capabilities",
                    "supports": { "streaming": true, "tool_calls": true },
                    "type": "chat"
                  },
                  "id": "gpt-4o",
                  "is_chat_default": true,
                  "is_chat_fallback": false,
                  "model_picker_enabled": true,
                  "name": "GPT-4o",
                  "object": "model",
                  "preview": false,
                  "vendor": "OpenAI",
                  "version": "gpt-4o",
                  "supported_endpoints": ["/chat/completions", "/responses"]
                }
              ],
              "object": "list"
            }"#;

        let schema: ModelSchema = serde_json::from_str(json).unwrap();

        assert_eq!(schema.data.len(), 1);
        assert_eq!(schema.data[0].id, "gpt-4o");
        assert_eq!(
            schema.data[0].supported_endpoints,
            vec![
                ModelSupportedEndpoint::ChatCompletions,
                ModelSupportedEndpoint::Responses
            ]
        );
    }

    #[test]
    fn test_supports_response_method() {
        // Test the supports_response() method which determines endpoint routing.
        let model_with_responses_only = Model {
            billing: ModelBilling {
                is_premium: false,
                multiplier: 1.0,
                restricted_to: None,
            },
            capabilities: ModelCapabilities {
                family: "test".to_string(),
                limits: ModelLimits::default(),
                supports: ModelSupportedFeatures {
                    streaming: true,
                    tool_calls: true,
                    parallel_tool_calls: false,
                    vision: false,
                    thinking: false,
                    adaptive_thinking: false,
                    max_thinking_budget: None,
                    min_thinking_budget: None,
                    reasoning_effort: vec![],
                },
                model_type: "chat".to_string(),
                tokenizer: None,
            },
            id: "test-model".to_string(),
            name: "Test Model".to_string(),
            policy: None,
            vendor: ModelVendor::OpenAI,
            is_chat_default: false,
            is_chat_fallback: false,
            model_picker_enabled: true,
            supported_endpoints: vec![ModelSupportedEndpoint::Responses],
        };

        let model_with_chat_completions = Model {
            supported_endpoints: vec![ModelSupportedEndpoint::ChatCompletions],
            ..model_with_responses_only.clone()
        };

        let model_with_both = Model {
            supported_endpoints: vec![
                ModelSupportedEndpoint::ChatCompletions,
                ModelSupportedEndpoint::Responses,
            ],
            ..model_with_responses_only.clone()
        };

        let model_with_messages = Model {
            supported_endpoints: vec![ModelSupportedEndpoint::Messages],
            ..model_with_responses_only.clone()
        };

        // Only /responses endpoint -> supports_response = true
        assert!(model_with_responses_only.supports_response());

        // Only /chat/completions endpoint -> supports_response = false
        assert!(!model_with_chat_completions.supports_response());

        // Both endpoints (has /chat/completions) -> supports_response = false
        assert!(model_with_both.supports_response());

        // Only /v1/messages endpoint -> supports_response = false (doesn't have /responses)
        assert!(!model_with_messages.supports_response());
    }

    #[test]
    fn test_tool_choice_required_serializes_as_required() {
        // Regression test: ToolChoice::Required must serialize as "required" (not "any")
        // for OpenAI-compatible APIs. Reverting the rename would break this.
        assert_eq!(
            serde_json::to_string(&ToolChoice::Required).unwrap(),
            "\"required\""
        );
        assert_eq!(
            serde_json::to_string(&ToolChoice::Auto).unwrap(),
            "\"auto\""
        );
        assert_eq!(
            serde_json::to_string(&ToolChoice::None).unwrap(),
            "\"none\""
        );
    }
}