stakpak_server/

sandbox.rs

//! Sandboxed MCP server management.
//!
//! When sandbox mode is enabled for a session, a `stakpak mcp start` server
//! is spawned inside a warden container. The host-side proxy connects to it
//! via HTTPS/mTLS, and tool calls from the agent loop are routed through the
//! containerized server — executing `run_command`, file I/O, etc. inside the
//! sandbox.
//!
//! ## Sandbox Modes
//!
//! - **Persistent** (default): A single sandbox container is spawned at process startup and
//!   reused across all sessions. Near-zero per-session overhead, slightly less
//!   isolation (sessions share the same container filesystem).
//!
//! - **Ephemeral**: A new sandbox container is spawned for each session
//!   and destroyed when the session ends. Maximum isolation, ~5-10s startup overhead.
//!
//! ## mTLS key exchange
//!
//! Each side generates its own identity independently:
//!
//! 1. Host generates a client identity (CA + leaf cert + key, all in memory)
//! 2. Host passes the client **CA cert** (public only) to the container via env var
//! 3. Container generates a server identity (CA + leaf cert + key, all in memory)
//! 4. Container outputs the server **CA cert** (public only) to stdout
//! 5. Host parses the server CA cert and builds a client TLS config
//!
//! Private keys never leave their respective processes.

use serde::{Deserialize, Serialize};
use stakpak_mcp_client::McpClient;
use stakpak_mcp_proxy::client::{ClientPoolConfig, ServerConfig};
use stakpak_mcp_proxy::server::start_proxy_server;
use stakpak_shared::cert_utils::{CertificateChain, MtlsIdentity};
use std::collections::HashMap;
use std::path::Path;
use std::sync::Arc;
use tokio::io::AsyncBufReadExt;
use tokio::net::TcpListener;
use tokio::process::Child;
use tokio::sync::{broadcast, watch};

/// Environment variable used to pass the client CA cert PEM to the container.
const TRUSTED_CLIENT_CA_ENV: &str = "STAKPAK_MCP_CLIENT_CA";

// ── Sandbox mode ────────────────────────────────────────────────────────────

/// Controls how sandbox containers are managed across sessions.
#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "lowercase")]
pub enum SandboxMode {
    /// Spawn a new sandbox container for each session and destroy it when the
    /// session ends. Maximum isolation, ~5-10s startup overhead per session.
    Ephemeral,
    /// Spawn a single sandbox container at process startup and reuse it for all
    /// sessions. Near-zero per-session overhead, shared container filesystem.
    #[default]
    Persistent,
}

impl std::fmt::Display for SandboxMode {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            SandboxMode::Ephemeral => write!(f, "ephemeral"),
            SandboxMode::Persistent => write!(f, "persistent"),
        }
    }
}

// ── Sandbox config ──────────────────────────────────────────────────────────

/// Configuration for spawning sandboxed MCP servers.
#[derive(Clone, Debug)]
pub struct SandboxConfig {
    /// Path to the warden binary.
    pub warden_path: String,
    /// Container image for the sandbox (e.g., `ghcr.io/stakpak/agent:v1.2.3`).
    pub image: String,
    /// Volume mounts for the container (e.g., `["./:/agent:ro"]`).
    pub volumes: Vec<String>,
    /// How sandbox containers are managed across sessions.
    pub mode: SandboxMode,
}

// ── Sandbox health ──────────────────────────────────────────────────────────

/// Health status of a persistent sandbox, updated by the background monitor.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct SandboxHealth {
    /// Whether the last health check succeeded.
    pub healthy: bool,
    /// Number of consecutive successful health checks.
    pub consecutive_ok: u64,
    /// Number of consecutive failed health checks.
    pub consecutive_failures: u64,
    /// ISO 8601 timestamp of the last successful health check.
    pub last_ok: Option<String>,
    /// Error message from the last failed health check, if any.
    pub last_error: Option<String>,
    /// Total number of respawn attempts since the sandbox was started.
    pub total_respawn_attempts: u64,
}

impl Default for SandboxHealth {
    fn default() -> Self {
        Self {
            healthy: true,
            consecutive_ok: 0,
            consecutive_failures: 0,
            last_ok: None,
            last_error: None,
            total_respawn_attempts: 0,
        }
    }
}

/// Interval between health checks for persistent sandboxes.
const HEALTH_CHECK_INTERVAL: std::time::Duration = std::time::Duration::from_secs(30);

/// Number of consecutive failures before attempting a respawn.
const RESPAWN_THRESHOLD: u64 = 3;

/// Maximum number of total respawn attempts before giving up and shutting down.
const MAX_RESPAWN_ATTEMPTS: u64 = 5;

// ── Persistent sandbox ──────────────────────────────────────────────────────

/// A long-lived sandbox that persists across sessions.
///
/// Spawned once at process startup (when `SandboxMode::Persistent` is configured)
/// and shared by all sessions. Includes a background health monitor that
/// periodically pings the sandbox and attempts to respawn it on failure.
pub struct PersistentSandbox {
    inner: Arc<tokio::sync::RwLock<SandboxedMcpServer>>,
    config: SandboxConfig,
    health_rx: watch::Receiver<SandboxHealth>,
    /// Handle to the background health monitor task.
    monitor_handle: tokio::task::JoinHandle<()>,
}

impl PersistentSandbox {
    /// Spawn a persistent sandbox with a background health monitor.
    pub async fn spawn(config: &SandboxConfig) -> Result<Self, String> {
        tracing::info!(image = %config.image, "Spawning persistent sandbox container");
        let inner = SandboxedMcpServer::spawn(config).await?;
        tracing::info!("Persistent sandbox ready");

        let initial_health = SandboxHealth {
            healthy: true,
            consecutive_ok: 1,
            consecutive_failures: 0,
            last_ok: Some(chrono::Utc::now().to_rfc3339()),
            last_error: None,
            total_respawn_attempts: 0,
        };
        let (health_tx, health_rx) = watch::channel(initial_health);

        let inner = Arc::new(tokio::sync::RwLock::new(inner));
        let monitor_inner = inner.clone();
        let monitor_config = config.clone();

        let monitor_handle = tokio::spawn(async move {
            health_monitor_loop(monitor_inner, monitor_config, health_tx).await;
        });

        Ok(Self {
            inner,
            config: config.clone(),
            health_rx,
            monitor_handle,
        })
    }

    /// Get the MCP client for routing tool calls through this sandbox.
    pub async fn client(&self) -> Arc<McpClient> {
        self.inner.read().await.client.clone()
    }

    /// Get the tools available from this sandbox.
    pub async fn tools(&self) -> Vec<stakai::Tool> {
        self.inner.read().await.tools.clone()
    }

    /// Get the current health status (non-blocking snapshot).
    pub fn health(&self) -> SandboxHealth {
        self.health_rx.borrow().clone()
    }

    /// Get the sandbox mode from the config.
    pub fn mode(&self) -> &SandboxMode {
        &self.config.mode
    }

    /// Shut down the persistent sandbox and its health monitor.
    pub async fn shutdown(self) {
        tracing::info!("Shutting down persistent sandbox");
        self.monitor_handle.abort();
        // Try to take exclusive ownership of the inner sandbox for clean shutdown.
        // If other sessions still hold references, the RwLock + Arc will be dropped
        // when all references are gone; the container process will be cleaned up by
        // the OS when the host process exits.
        if let Ok(inner) = Arc::try_unwrap(self.inner) {
            let sandbox = inner.into_inner();
            sandbox.shutdown().await;
        } else {
            tracing::warn!(
                "Other references to persistent sandbox still exist; container will be cleaned up on process exit"
            );
        }
    }

    /// Force-kill the sandbox container and abort the health monitor.
    ///
    /// Unlike `shutdown(self)`, this works through a shared reference so it can
    /// be called from the graceful-shutdown handler where only `Arc<Self>` is
    /// available. The container process is killed via the write lock.
    pub async fn kill(&self) {
        tracing::warn!(
            "Killing persistent sandbox container — in-flight sessions using this sandbox will fail"
        );
        self.monitor_handle.abort();
        self.inner.write().await.teardown().await;
        tracing::info!("Persistent sandbox container killed");
    }
}

/// Background loop that periodically health-checks the persistent sandbox
/// and attempts to respawn it after consecutive failures.
async fn health_monitor_loop(
    inner: Arc<tokio::sync::RwLock<SandboxedMcpServer>>,
    config: SandboxConfig,
    health_tx: watch::Sender<SandboxHealth>,
) {
    let mut health = SandboxHealth::default();

    loop {
        tokio::time::sleep(HEALTH_CHECK_INTERVAL).await;

        let check_result = {
            let sandbox = inner.read().await;
            // Use list_tools as a health probe — it exercises the full
            // mTLS → proxy → container → MCP server path.
            tokio::time::timeout(
                std::time::Duration::from_secs(10),
                stakpak_mcp_client::get_tools(&sandbox.client),
            )
            .await
        };

        match check_result {
            Ok(Ok(_tools)) => {
                health.healthy = true;
                health.consecutive_ok = health.consecutive_ok.saturating_add(1);
                health.consecutive_failures = 0;
                health.last_ok = Some(chrono::Utc::now().to_rfc3339());
                health.last_error = None;
                tracing::debug!(
                    consecutive_ok = health.consecutive_ok,
                    "Persistent sandbox health check passed"
                );
            }
            Ok(Err(e)) => {
                let err_msg = format!("MCP error: {e}");
                health.healthy = false;
                health.consecutive_ok = 0;
                health.consecutive_failures = health.consecutive_failures.saturating_add(1);
                health.last_error = Some(err_msg.clone());
                tracing::warn!(
                    consecutive_failures = health.consecutive_failures,
                    error = %err_msg,
                    "Persistent sandbox health check failed"
                );
            }
            Err(_timeout) => {
                health.healthy = false;
                health.consecutive_ok = 0;
                health.consecutive_failures = health.consecutive_failures.saturating_add(1);
                health.last_error = Some("Health check timed out (10s)".to_string());
                tracing::warn!(
                    consecutive_failures = health.consecutive_failures,
                    "Persistent sandbox health check timed out"
                );
            }
        }

        // Attempt respawn after RESPAWN_THRESHOLD consecutive failures
        if health.consecutive_failures >= RESPAWN_THRESHOLD {
            health.total_respawn_attempts = health.total_respawn_attempts.saturating_add(1);

            if health.total_respawn_attempts > MAX_RESPAWN_ATTEMPTS {
                tracing::error!(
                    total_attempts = health.total_respawn_attempts,
                    "Persistent sandbox exceeded maximum respawn attempts ({}) — giving up. \
                     The server cannot operate without a healthy sandbox. Shutting down.",
                    MAX_RESPAWN_ATTEMPTS
                );
                health.last_error = Some(format!(
                    "Exceeded max respawn attempts ({}); sandbox permanently failed",
                    MAX_RESPAWN_ATTEMPTS
                ));
                let _ = health_tx.send(health);
                // Exit the monitor loop — the sandbox is unrecoverable.
                // The server health endpoint will report unhealthy, and operators
                // should investigate and restart the autopilot.
                return;
            }

            tracing::error!(
                failures = health.consecutive_failures,
                attempt = health.total_respawn_attempts,
                max_attempts = MAX_RESPAWN_ATTEMPTS,
                "Persistent sandbox unhealthy — attempting respawn"
            );

            // Take write lock to replace the sandbox
            let mut sandbox = inner.write().await;

            // Shut down the old one (best-effort)
            sandbox.teardown().await;

            match SandboxedMcpServer::spawn(&config).await {
                Ok(new_sandbox) => {
                    *sandbox = new_sandbox;
                    health.healthy = true;
                    health.consecutive_ok = 1;
                    health.consecutive_failures = 0;
                    health.last_ok = Some(chrono::Utc::now().to_rfc3339());
                    health.last_error = None;
                    tracing::info!("Persistent sandbox respawned successfully");
                }
                Err(e) => {
                    health.last_error = Some(format!("Respawn failed: {e}"));
                    tracing::error!(error = %e, "Failed to respawn persistent sandbox");
                    // Don't reset consecutive_failures — next iteration will try again
                }
            }
        }

        // Publish updated health (ignore error if all receivers dropped)
        let _ = health_tx.send(health.clone());
    }
}

/// A running sandboxed MCP server with its associated proxy and client.
///
/// Drop this struct to shut down the sandbox.
pub struct SandboxedMcpServer {
    /// MCP client connected via the per-session proxy.
    pub client: Arc<McpClient>,
    /// Tools available from the sandboxed server.
    pub tools: Vec<stakai::Tool>,
    /// Channel to shut down the per-session proxy.
    proxy_shutdown_tx: broadcast::Sender<()>,
    /// The warden container child process.
    container_process: Child,
}

impl SandboxedMcpServer {
    /// Spawn a sandboxed MCP server inside a warden container and connect to it.
    ///
    /// 1. Generates a client mTLS identity (private key stays in host memory)
    /// 2. Passes the client CA cert (public) to the container via env var
    /// 3. Spawns `warden wrap <image> -- stakpak mcp start`
    /// 4. Parses the server CA cert (public) from the container's stdout
    /// 5. Builds a client TLS config trusting the server CA, using the client key
    /// 6. Starts a per-session MCP proxy pointing to the container
    /// 7. Connects a client to the proxy
    pub async fn spawn(config: &SandboxConfig) -> Result<Self, String> {
        // 1. Generate client identity — private key stays in host memory
        let client_identity = MtlsIdentity::generate_client()
            .map_err(|e| format!("Failed to generate client identity: {e}"))?;

        let client_ca_pem = client_identity
            .ca_cert_pem()
            .map_err(|e| format!("Failed to get client CA PEM: {e}"))?;

        // 2. Find a free port for the container's MCP server to expose
        let container_host_port = find_free_port()
            .await
            .map_err(|e| format!("Failed to find free port for sandbox: {e}"))?;

        // 3. Spawn warden container, passing client CA cert (public) via env var
        let mut container_process =
            spawn_warden_container(config, container_host_port, &client_ca_pem)
                .await
                .map_err(|e| format!("Failed to spawn sandbox container: {e}"))?;

        // 4. Parse the server CA cert (public) from the container's stdout
        let server_ca_pem = parse_server_ca_from_stdout(&mut container_process).await?;
        tracing::info!(
            "Parsed server CA from container stdout ({} bytes)",
            server_ca_pem.len()
        );

        // 5. Build client TLS config — trusts server CA, authenticates with our key
        let container_client_config = client_identity
            .create_client_config(&server_ca_pem)
            .map_err(|e| format!("Failed to create client TLS config: {e}"))?;

        // 6. Wait for the MCP server inside the container to be ready
        let server_url = format!("https://127.0.0.1:{container_host_port}/mcp");
        tracing::info!(url = %server_url, "Waiting for sandbox MCP server to be ready");
        wait_for_server_ready(&server_url, &container_client_config).await?;
        tracing::info!("Sandbox MCP server is ready");

        // 7. Start a per-session proxy connecting to the sandboxed server
        let (proxy_shutdown_tx, proxy_shutdown_rx) = broadcast::channel::<()>(1);

        let proxy_binding = find_available_binding("sandbox proxy").await?;
        let proxy_url = format!("https://{}/mcp", proxy_binding.address);

        let proxy_cert_chain = Arc::new(
            CertificateChain::generate()
                .map_err(|e| format!("Failed to generate proxy certificates: {e}"))?,
        );

        let pool_config = build_sandbox_proxy_config(server_url, Arc::new(container_client_config));

        let proxy_chain_for_server = proxy_cert_chain.clone();
        let proxy_listener = proxy_binding.listener;
        tokio::spawn(async move {
            if let Err(e) = start_proxy_server(
                pool_config,
                proxy_listener,
                proxy_chain_for_server,
                true,  // redact_secrets
                false, // privacy_mode
                Some(proxy_shutdown_rx),
            )
            .await
            {
                tracing::error!("Sandbox proxy error: {e}");
            }
        });

        // Small delay for proxy to start
        tokio::time::sleep(tokio::time::Duration::from_millis(100)).await;

        // 8. Connect client to proxy
        let client = connect_to_proxy(&proxy_url, proxy_cert_chain).await?;

        // 9. Get tools
        let mcp_tools = stakpak_mcp_client::get_tools(&client)
            .await
            .map_err(|e| format!("Failed to get sandbox tools: {e}"))?;

        let tools = mcp_tools
            .into_iter()
            .map(|tool| stakai::Tool {
                tool_type: "function".to_string(),
                function: stakai::ToolFunction {
                    name: tool.name.as_ref().to_string(),
                    description: tool
                        .description
                        .as_ref()
                        .map(std::string::ToString::to_string)
                        .unwrap_or_default(),
                    parameters: serde_json::Value::Object((*tool.input_schema).clone()),
                },
                provider_options: None,
            })
            .collect();

        Ok(Self {
            client,
            tools,
            proxy_shutdown_tx,
            container_process,
        })
    }

    /// Shut down the sandbox: stop the proxy and kill the container.
    pub async fn shutdown(mut self) {
        self.teardown().await;
    }

    /// Stop the proxy and gracefully shut down the warden container process.
    ///
    /// Sends SIGINT (not SIGTERM) because warden listens for `ctrl_c` (SIGINT)
    /// to trigger its cleanup — stopping the user container and sidecar.
    /// Waits up to 10s for warden to finish, then force-kills.
    pub async fn teardown(&mut self) {
        let _ = self.proxy_shutdown_tx.send(());

        // Send SIGINT so warden's `signal::ctrl_c()` handler fires and runs
        // its container cleanup (stop_user_container + cleanup sidecar).
        #[cfg(unix)]
        if let Some(pid) = self.container_process.id() {
            let _ = tokio::process::Command::new("kill")
                .args(["-INT", &pid.to_string()])
                .output()
                .await;
        }

        // Give warden up to 10s to stop containers and exit cleanly.
        match tokio::time::timeout(
            std::time::Duration::from_secs(10),
            self.container_process.wait(),
        )
        .await
        {
            Ok(Ok(status)) => {
                tracing::debug!(exit_status = ?status, "Warden process exited gracefully");
            }
            _ => {
                tracing::warn!("Warden process did not exit in 10s — force killing");
                let _ = self.container_process.kill().await;
                let _ = self.container_process.wait().await;
            }
        }
    }
}

async fn spawn_warden_container(
    config: &SandboxConfig,
    host_port: u16,
    client_ca_pem: &str,
) -> Result<Child, String> {
    use stakpak_shared::container::{expand_volume_path, is_named_volume};

    let mut cmd = tokio::process::Command::new(&config.warden_path);
    cmd.arg("wrap");
    cmd.arg(&config.image);

    // Mount configured volumes
    for vol in &config.volumes {
        let expanded = expand_volume_path(vol);
        let host_path = expanded.split(':').next().unwrap_or(&expanded);
        // Named volumes (e.g. "stakpak-aqua-cache:/container/path") don't have a
        // host filesystem path — mount them unconditionally. Bind mounts are only
        // added when the host path actually exists.
        if is_named_volume(host_path) || Path::new(host_path).exists() {
            cmd.args(["--volume", &expanded]);
        }
    }

    // Port forwarding for the MCP server — publish on the sidecar so the
    // host can reach the container's MCP server port directly.
    cmd.args(["-p", &format!("127.0.0.1:{host_port}:8080")]);

    // Prevent warden re-entry
    cmd.args(["--env", "STAKPAK_SKIP_WARDEN=1"]);

    // Tell the MCP server to bind to a fixed port inside the container
    // so it matches the published port on the sidecar.
    cmd.args(["--env", "STAKPAK_MCP_PORT=8080"]);

    // Pass the client CA cert (public only) so the server can trust the client.
    cmd.args(["--env", &format!("{TRUSTED_CLIENT_CA_ENV}={client_ca_pem}")]);

    // Pass through API credentials if set
    if let Ok(api_key) = std::env::var("STAKPAK_API_KEY") {
        cmd.args(["--env", &format!("STAKPAK_API_KEY={api_key}")]);
    }
    if let Ok(profile) = std::env::var("STAKPAK_PROFILE") {
        cmd.args(["--env", &format!("STAKPAK_PROFILE={profile}")]);
    }
    if let Ok(endpoint) = std::env::var("STAKPAK_API_ENDPOINT") {
        cmd.args(["--env", &format!("STAKPAK_API_ENDPOINT={endpoint}")]);
    }

    // The MCP server detects STAKPAK_MCP_CLIENT_CA and generates its own
    // server identity, outputting the server CA cert to stdout.
    cmd.args(["--", "stakpak", "mcp", "start"]);

    cmd.stdout(std::process::Stdio::piped());
    cmd.stderr(std::process::Stdio::piped());
    cmd.stdin(std::process::Stdio::null());

    let child = cmd
        .spawn()
        .map_err(|e| format!("Failed to spawn warden process: {e}"))?;

    Ok(child)
}

/// Parse the server CA certificate PEM from the container's stdout.
///
/// The MCP server outputs the server CA cert between structured delimiters:
/// ```text
/// ---BEGIN STAKPAK SERVER CA---
/// -----BEGIN CERTIFICATE-----
/// ...
/// -----END CERTIFICATE-----
/// ---END STAKPAK SERVER CA---
/// ```
async fn parse_server_ca_from_stdout(process: &mut Child) -> Result<String, String> {
    let stdout = process
        .stdout
        .take()
        .ok_or_else(|| "Container stdout not captured".to_string())?;

    let mut reader = tokio::io::BufReader::new(stdout);
    let mut server_ca_pem = String::new();
    let mut in_server_ca = false;
    let mut line = String::new();

    let timeout_duration = tokio::time::Duration::from_secs(60);
    let deadline = tokio::time::Instant::now() + timeout_duration;

    tracing::debug!("Starting to read container stdout for server CA...");

    loop {
        line.clear();
        let bytes_read = tokio::time::timeout_at(deadline, reader.read_line(&mut line))
            .await
            .map_err(|_| {
                "Timed out waiting for container to output server CA certificate".to_string()
            })?
            .map_err(|e| format!("Failed to read container stdout: {e}"))?;

        if bytes_read == 0 {
            tracing::error!("Container stdout EOF before server CA was found");
            return Err("Container exited before outputting server CA certificate".to_string());
        }

        let trimmed = line.trim();
        tracing::debug!(line = %trimmed, bytes = bytes_read, "Read line from container stdout");

        if trimmed == "---BEGIN STAKPAK SERVER CA---" {
            in_server_ca = true;
            continue;
        }

        if trimmed == "---END STAKPAK SERVER CA---" {
            tracing::debug!("Found end of server CA block");
            break;
        }

        if in_server_ca {
            server_ca_pem.push_str(trimmed);
            server_ca_pem.push('\n');
        }
    }

    let server_ca_pem = server_ca_pem.trim().to_string();

    if server_ca_pem.is_empty() {
        return Err("Failed to parse server CA certificate from container output".to_string());
    }

    Ok(server_ca_pem)
}

async fn wait_for_server_ready(
    url: &str,
    client_config: &rustls::ClientConfig,
) -> Result<(), String> {
    let http_client = reqwest::Client::builder()
        .use_preconfigured_tls(client_config.clone())
        .build()
        .map_err(|e| format!("Failed to build readiness check client: {e}"))?;

    let mut last_error = String::new();
    for attempt in 0..30 {
        tokio::time::sleep(tokio::time::Duration::from_millis(if attempt < 5 {
            500
        } else {
            1000
        }))
        .await;

        match http_client.get(url).send().await {
            Ok(_) => {
                tracing::info!(attempt, "Sandbox MCP server ready");
                return Ok(());
            }
            Err(e) => {
                last_error = format!("{e:?}");
                tracing::debug!(attempt, error = %last_error, "Readiness check failed");
            }
        }
    }

    Err(format!(
        "Sandbox MCP server failed to become ready after 30 attempts: {last_error}"
    ))
}

struct ProxyBinding {
    address: String,
    listener: TcpListener,
}

async fn find_available_binding(purpose: &str) -> Result<ProxyBinding, String> {
    let listener = TcpListener::bind("127.0.0.1:0")
        .await
        .map_err(|e| format!("Failed to bind port for {purpose}: {e}"))?;
    let addr = listener
        .local_addr()
        .map_err(|e| format!("Failed to get address for {purpose}: {e}"))?;
    Ok(ProxyBinding {
        address: addr.to_string(),
        listener,
    })
}

// TODO: TOCTOU race — between dropping the listener and Docker binding the port,
// another process could claim it. Consider retrying with a different port on bind
// failure, or passing the listener fd directly if Docker supports it.
async fn find_free_port() -> Result<u16, String> {
    let listener = TcpListener::bind("127.0.0.1:0")
        .await
        .map_err(|e| format!("Failed to bind ephemeral port: {e}"))?;
    let port = listener
        .local_addr()
        .map_err(|e| format!("Failed to get ephemeral port: {e}"))?
        .port();
    // Drop the listener to free the port for Docker to use
    drop(listener);
    Ok(port)
}

fn build_sandbox_proxy_config(
    sandbox_server_url: String,
    client_tls_config: Arc<rustls::ClientConfig>,
) -> ClientPoolConfig {
    let mut servers: HashMap<String, ServerConfig> = HashMap::new();

    // Register the sandboxed MCP server under the same name ("stakpak") so
    // tool names like `stakpak__run_command` route correctly through the proxy.
    servers.insert(
        "stakpak".to_string(),
        ServerConfig::Http {
            url: sandbox_server_url,
            headers: None,
            certificate_chain: Arc::new(None),
            client_tls_config: Some(client_tls_config),
        },
    );

    // Keep the external paks server accessible
    servers.insert(
        "paks".to_string(),
        ServerConfig::Http {
            url: "https://apiv2.stakpak.dev/v1/paks/mcp".to_string(),
            headers: None,
            certificate_chain: Arc::new(None),
            client_tls_config: None,
        },
    );

    ClientPoolConfig::with_servers(servers)
}

async fn connect_to_proxy(
    proxy_url: &str,
    cert_chain: Arc<CertificateChain>,
) -> Result<Arc<McpClient>, String> {
    const MAX_RETRIES: u32 = 5;
    let mut retry_delay = tokio::time::Duration::from_millis(50);
    let mut last_error = None;

    for attempt in 1..=MAX_RETRIES {
        match stakpak_mcp_client::connect_https(proxy_url, Some(cert_chain.clone()), None).await {
            Ok(client) => return Ok(Arc::new(client)),
            Err(e) => {
                last_error = Some(e);
                if attempt < MAX_RETRIES {
                    tokio::time::sleep(retry_delay).await;
                    retry_delay *= 2;
                }
            }
        }
    }

    Err(format!(
        "Failed to connect to sandbox proxy after {MAX_RETRIES} retries: {}",
        last_error.map(|e| e.to_string()).unwrap_or_default()
    ))
}

#[cfg(test)]
mod tests {
    #[test]
    fn parse_server_ca_from_structured_output() {
        let output = "\
🔐 mTLS enabled - independent identity (sandbox mode)
---BEGIN STAKPAK SERVER CA---
-----BEGIN CERTIFICATE-----
MIIB0zCCAXmgAwIBAgIUFAKE=
-----END CERTIFICATE-----
---END STAKPAK SERVER CA---
MCP server started at https://0.0.0.0:8080/mcp
";

        let expected_ca = "\
-----BEGIN CERTIFICATE-----
MIIB0zCCAXmgAwIBAgIUFAKE=
-----END CERTIFICATE-----";

        // Parse the same way parse_server_ca_from_stdout does
        let mut server_ca_pem = String::new();
        let mut in_server_ca = false;

        for line in output.lines() {
            let trimmed = line.trim();
            if trimmed == "---BEGIN STAKPAK SERVER CA---" {
                in_server_ca = true;
                continue;
            }
            if trimmed == "---END STAKPAK SERVER CA---" {
                break;
            }
            if in_server_ca {
                server_ca_pem.push_str(trimmed);
                server_ca_pem.push('\n');
            }
        }

        assert_eq!(server_ca_pem.trim(), expected_ca);
    }

    #[test]
    fn mtls_identity_cross_trust() {
        use stakpak_shared::cert_utils::MtlsIdentity;

        // Ensure a crypto provider is installed for rustls
        let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();

        // Simulate the sandbox mTLS exchange
        let client_identity = MtlsIdentity::generate_client().expect("generate client identity");
        let server_identity = MtlsIdentity::generate_server().expect("generate server identity");

        let client_ca_pem = client_identity.ca_cert_pem().expect("client CA PEM");
        let server_ca_pem = server_identity.ca_cert_pem().expect("server CA PEM");

        // Server trusts client CA, client trusts server CA
        let _server_config = server_identity
            .create_server_config(&client_ca_pem)
            .expect("server config with client CA trust");
        let _client_config = client_identity
            .create_client_config(&server_ca_pem)
            .expect("client config with server CA trust");

        // Only public CA certs were exchanged — private keys stayed in their
        // respective MtlsIdentity structs.
        assert!(client_ca_pem.contains("BEGIN CERTIFICATE"));
        assert!(server_ca_pem.contains("BEGIN CERTIFICATE"));
        assert!(!client_ca_pem.contains("PRIVATE KEY"));
        assert!(!server_ca_pem.contains("PRIVATE KEY"));
    }

    // ── Named volume detection in expand_volume_path / mount filter ────

    #[test]
    fn expand_volume_path_leaves_named_volumes_unchanged() {
        use stakpak_shared::container::expand_volume_path;
        let named = "stakpak-aqua-cache:/home/agent/.local/share/aquaproj-aqua";
        assert_eq!(expand_volume_path(named), named);
    }

    /// Named volumes (no `/` or `.` prefix in the host part) must pass the
    /// mount filter even though they don't exist on the host filesystem.
    #[test]
    fn named_volume_is_detected_correctly() {
        use stakpak_shared::container::is_named_volume;
        let cases = vec![
            ("stakpak-aqua-cache", true),
            ("my-volume", true),
            ("./relative/path", false),
            ("/absolute/path", false),
            ("relative/with/slash", false),
            (".", false),
        ];
        for (host_part, expected) in cases {
            assert_eq!(
                is_named_volume(host_part),
                expected,
                "host_part={host_part:?} expected named={expected}"
            );
        }
    }

    #[test]
    fn sandbox_mode_default_is_persistent() {
        assert_eq!(
            super::SandboxMode::default(),
            super::SandboxMode::Persistent
        );
    }

    #[test]
    fn sandbox_mode_serde_roundtrip() {
        #[derive(serde::Serialize, serde::Deserialize)]
        struct Wrapper {
            #[serde(default)]
            mode: super::SandboxMode,
        }

        // Explicit persistent
        let json = serde_json::json!({"mode": "persistent"});
        let w: Wrapper = serde_json::from_value(json).expect("deserialize persistent");
        assert_eq!(w.mode, super::SandboxMode::Persistent);

        // Explicit ephemeral
        let json = serde_json::json!({"mode": "ephemeral"});
        let w: Wrapper = serde_json::from_value(json).expect("deserialize ephemeral");
        assert_eq!(w.mode, super::SandboxMode::Ephemeral);

        // Missing field → default (persistent)
        let json = serde_json::json!({});
        let w: Wrapper = serde_json::from_value(json).expect("deserialize default");
        assert_eq!(w.mode, super::SandboxMode::Persistent);

        // Display
        assert_eq!(super::SandboxMode::Persistent.to_string(), "persistent");
        assert_eq!(super::SandboxMode::Ephemeral.to_string(), "ephemeral");
    }

    #[test]
    fn sandbox_health_default_is_healthy() {
        let h = super::SandboxHealth::default();
        assert!(h.healthy);
        assert_eq!(h.consecutive_ok, 0);
        assert_eq!(h.consecutive_failures, 0);
        assert!(h.last_ok.is_none());
        assert!(h.last_error.is_none());
        assert_eq!(h.total_respawn_attempts, 0);
    }
}