alien_core/resources/

sandbox.rs

//! Sandbox resource for running untrusted code in an isolated environment.
//!
//! A Sandbox is a session-oriented resource: the declaration provisions a durable parent, and
//! the application creates and destroys individual sessions through its binding at runtime.
//!
//! The capability set differs per platform and is published rather than assumed. Calling an
//! unsupported capability is a typed error naming both the platform and the capability, so a
//! portable application can branch on `SandboxCapabilities` before it calls.

use crate::error::{ErrorData, Result};
use crate::resource::{ResourceDefinition, ResourceOutputsDefinition, ResourceRef, ResourceType};
use crate::resources::ToolchainConfig;
use crate::Platform;
use alien_error::AlienError;
use bon::Builder;
use serde::{Deserialize, Serialize};
use std::any::Any;
use std::fmt::Debug;

/// Specifies where the sandbox's root filesystem comes from.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", tag = "type")]
pub enum SandboxCode {
    /// A prebuilt container image used as the sandbox root filesystem.
    #[serde(rename_all = "camelCase")]
    Image {
        /// Image reference (e.g. `ubuntu:24.04`, `ghcr.io/myorg/sandbox:latest`).
        ///
        /// Two backends narrow it in opposite directions: AWS wants an `s3://` bundle, Azure a
        /// bare catalog name such as `ubuntu`. Each refuses the other's shape while planning.
        image: String,
    },
    /// Source built into a sandbox image at deploy time.
    #[serde(rename_all = "camelCase")]
    Source {
        /// The source directory to build from
        src: String,
        /// Toolchain configuration with type-safe options
        toolchain: ToolchainConfig,
    },
}

/// Hard ceilings enforced on a sandbox session.
///
/// These are limits, not scheduling requests. Untrusted code does not respect a hint, so every
/// field is enforced by the platform and a platform that cannot enforce one is rejected at plan
/// time rather than silently ignoring it.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", deny_unknown_fields)]
pub struct SandboxLimits {
    /// CPU ceiling in cores or millicores (e.g. `"1"`, `"500m"`)
    pub cpu: String,
    /// Memory ceiling (e.g. `"2Gi"`, `"512Mi"`)
    pub memory: String,
    /// Disk ceiling (e.g. `"20Gi"`)
    pub disk: String,
    /// Maximum number of processes, which bounds fork bombs.
    ///
    /// Optional because only a container runtime has the primitive: Kubernetes sets a pid ceiling
    /// per node, not per pod, and neither AWS MicroVMs nor Azure sandboxes expose one. Declaring
    /// it on a platform that cannot apply it is refused at plan time.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub max_processes: Option<u32>,
}

/// One of the five sizes a Lambda MicroVM can be built at.
///
/// AWS has no ceiling knob: `minimumMemoryInMiB` sets a *baseline* and a running MicroVM bursts
/// vertically to four times it with no way to opt out. A declared ceiling is therefore honoured by
/// picking the tier whose **peak** stays inside it, not the tier whose baseline matches it.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct MicrovmTier {
    /// What `minimumMemoryInMiB` is set to.
    pub baseline_memory_mib: i64,
    /// The most memory the MicroVM can reach, in MiB.
    pub peak_memory_mib: i64,
    /// The most vCPU the MicroVM can reach.
    pub peak_vcpu: u32,
    /// The most disk the MicroVM can use, in MiB.
    pub max_disk_mib: i64,
}

/// The published sizes, smallest first. Baseline memory to vCPU is 2 GB per vCPU, peak is four
/// times baseline, and disk is fixed per tier rather than independently selectable.
/// Longest life AWS will run a MicroVM for, from `RunMicrovm`'s `maximumDurationInSeconds`.
const AWS_MAX_SESSION_LIFETIME_SECONDS: u32 = 28_800;

const MICROVM_TIERS: &[MicrovmTier] = &[
    MicrovmTier {
        baseline_memory_mib: 512,
        peak_memory_mib: 2048,
        peak_vcpu: 1,
        max_disk_mib: 8192,
    },
    MicrovmTier {
        baseline_memory_mib: 1024,
        peak_memory_mib: 4096,
        peak_vcpu: 2,
        max_disk_mib: 8192,
    },
    MicrovmTier {
        baseline_memory_mib: 2048,
        peak_memory_mib: 8192,
        peak_vcpu: 4,
        max_disk_mib: 8192,
    },
    MicrovmTier {
        baseline_memory_mib: 4096,
        peak_memory_mib: 16384,
        peak_vcpu: 8,
        max_disk_mib: 16384,
    },
    MicrovmTier {
        baseline_memory_mib: 8192,
        peak_memory_mib: 32768,
        peak_vcpu: 16,
        max_disk_mib: 32768,
    },
];

/// Outbound network policy for a sandbox.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", tag = "mode")]
pub enum SandboxEgress {
    /// No outbound network access.
    ///
    /// Routed traffic only. Link-local is not outbound and no backend's egress control reaches
    /// it, so this is not a boundary against instance metadata.
    Deny,
    /// Unrestricted outbound access to the public internet, and none to private ranges or the
    /// deployment's own network.
    ///
    /// Link-local carries the same exception as `Deny`. AWS and Kubernetes deliver both halves.
    /// Azure and GCP deliver the first only: one matches host patterns and the other is a single
    /// switch, so neither can name an address range to exclude.
    Allow,
    /// Outbound access only to the listed hostnames.
    ///
    /// Azure alone expresses it: its egress proxy matches on host pattern. The others filter by
    /// CIDR or carry a single switch, and both would approximate the list rather than keep it.
    #[serde(rename_all = "camelCase")]
    AllowDomains {
        /// Hostnames the sandbox may reach
        domains: Vec<String>,
    },
}

impl SandboxEgress {
    /// The single outbound switch for a backend that has no host matcher, or `None` for a mode a
    /// boolean cannot carry.
    ///
    /// `AllowDomains` needs a host list, so it maps to nothing and each caller refuses it in its
    /// own error naming the sandbox. One source for what a mode means, so a template and a session
    /// cannot disagree on it.
    pub fn internet_access_switch(&self) -> Option<bool> {
        match self {
            SandboxEgress::Allow => Some(true),
            SandboxEgress::Deny => Some(false),
            SandboxEgress::AllowDomains { .. } => None,
        }
    }
}

/// How long a session may live and when it is suspended.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", deny_unknown_fields)]
pub struct SandboxSessionPolicy {
    /// Wall-clock ceiling on a single session, after which the platform terminates it.
    ///
    /// Optional because not every backend has the primitive: Kubernetes has
    /// `activeDeadlineSeconds` and AWS `maximumDurationInSeconds`, while neither Azure nor Local
    /// expose one, so declaring a ceiling there is refused at plan time rather than accepted and
    /// never applied. AWS caps it at 8 hours.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub max_lifetime_seconds: Option<u32>,
    /// Idle period after which the session is suspended, where the platform supports it
    #[serde(skip_serializing_if = "Option::is_none")]
    pub idle_suspend_seconds: Option<u32>,
}

/// What a platform's sandbox backend can actually do.
///
/// Published so portable code can branch before calling rather than discovering a gap through
/// an error. Every field here corresponds to a capability that at least one platform lacks;
/// create, exec and terminate are the guaranteed floor and are therefore not listed.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", deny_unknown_fields)]
pub struct SandboxCapabilities {
    /// Files can be moved in and out of a session
    pub files: bool,
    /// A later call can reach a session created by an earlier one
    pub reconnect: bool,
    /// An authenticated, port-scoped capability to reach a service inside the sandbox
    pub preview: bool,
    /// Session state can be suspended and resumed
    pub suspend_resume: bool,
    /// A session's full state can be captured and used to create another
    pub snapshot: bool,
    /// Egress can be restricted to a hostname allowlist
    pub domain_egress_rules: bool,
    /// Whether a declared `deny` is actually enforced, rather than accepted and dropped
    pub egress_deny: bool,
    /// The platform enforces the declared cpu, memory and disk ceilings
    pub enforced_limits: bool,
    /// The platform can cap how many processes a session runs
    pub process_limit: bool,
    /// The platform terminates a session at a declared wall-clock deadline
    pub session_lifetime: bool,
    /// A command runs in its own PID namespace and cannot see or signal the agent's processes.
    ///
    /// Only where an agent runs as root. Creating the namespace needs `CAP_SYS_ADMIN`, and the
    /// Kubernetes sandbox pod drops every capability — which is also what denies `ptrace` by
    /// construction, so granting it there would remove a lock to add one.
    pub supervisor_pid_namespace: bool,
    /// The process supervising a command is a different identity from the command.
    ///
    /// False where a command runs as the agent's own user: it can then read the supervisor's
    /// environment and signal it. Separate from `supervisorPidNamespace`, which is about
    /// visibility rather than identity — a backend can have one without the other.
    pub supervisor_isolation: bool,
}

impl SandboxCapabilities {
    /// Returns what the given platform's sandbox backend supports.
    ///
    /// Errors for platforms with no sandbox backend, rather than returning an all-false set —
    /// "every capability is missing" and "this platform has no sandboxes" are different
    /// conditions and an application should not have to tell them apart by inspection.
    pub fn for_platform(platform: Platform) -> Result<Self> {
        match platform {
            Platform::Aws => Ok(Self {
                files: true,
                reconnect: true,
                preview: true,
                suspend_resume: true,
                snapshot: false,
                domain_egress_rules: false,
                egress_deny: true,
                enforced_limits: true,
                // Nothing in the API bounds process count.
                process_limit: false,
                // `maximumDurationInSeconds` on `RunMicrovm`, which Lambda enforces by
                // terminating the MicroVM. Capped at 8 hours by the service.
                session_lifetime: true,
                // Measured, not assumed: the agent inside a Lambda MicroVM runs as uid 0 with
                // `CapEff: 00000000a80425fb`, the standard container default set, which excludes
                // `CAP_SYS_ADMIN`. It can drop privilege (`CAP_SETUID`/`CAP_SETGID` are held) and
                // it cannot create a namespace. No backend offers this today.
                supervisor_pid_namespace: false,
                // The agent runs as uid 0 and `setuid`s the command to uid 60000, so the command
                // runs under a different identity than the process supervising it.
                supervisor_isolation: true,
            }),
            Platform::Azure => Ok(Self {
                files: true,
                reconnect: true,
                // A sandbox port carries a URL and an auth config, and the auth config offers two
                // things: anonymous, or Entra ID with an allowlist of human email addresses.
                // Neither is a credential scoped to a port for a fixed time, which is what a
                // preview capability is. Returning the anonymous URL would publish the port.
                preview: false,
                suspend_resume: true,
                // The one cloud of the five that could offer this, and the blocker is ours:
                // `snapshot()` returns an id and `CreateSessionRequest` has no field to consume
                // one, so no backend can complete the round trip. Nothing in the resource model
                // owns such an artifact either, and Microsoft states snapshots are not garbage
                // collected — an id with no owner is a bill that grows.
                snapshot: false,
                domain_egress_rules: true,
                egress_deny: true,
                enforced_limits: false,
                process_limit: false,
                // Auto-suspend and auto-delete exist; a wall-clock ceiling does not. Accepting
                // `maxLifetimeSeconds` here would be the silent no-op the capability set exists
                // to prevent, so this is a decision rather than a gap.
                session_lifetime: false,
                // No Alien process inside an Azure sandbox, so there is no supervisor to isolate.
                supervisor_pid_namespace: false,
                // No Alien process runs the command at all — the platform's own data plane does,
                // so there is no separate supervisor identity to speak of.
                supervisor_isolation: false,
            }),
            Platform::Gcp => Ok(Self::gcp_agent_platform()),
            // Preview needs a gateway that validates a session-and-port capability, and that
            // gateway does not exist yet.
            Platform::Kubernetes => Ok(Self {
                files: true,
                reconnect: true,
                preview: false,
                suspend_resume: false,
                snapshot: false,
                domain_egress_rules: false,
                egress_deny: true,
                enforced_limits: true,
                // A pid ceiling is a kubelet setting per node, not a pod field.
                process_limit: false,
                // `activeDeadlineSeconds` on the pod, which the kubelet enforces.
                session_lifetime: true,
                // The pod drops every capability, including the `CAP_SYS_ADMIN` the agent would
                // need to unshare. That is also what denies `ptrace`, so this stays false rather
                // than the pod being weakened to make it true.
                supervisor_pid_namespace: false,
                // The pod pins one uid (`run_as_user: 65534` on both pod and container) with
                // `capabilities.drop: [ALL]` and `allow_privilege_escalation: false`, so no
                // process can setuid to split the command off from a supervisor. No uid split is
                // possible, so none exists.
                supervisor_isolation: false,
            }),
            Platform::Local => Ok(Self {
                files: true,
                reconnect: true,
                preview: true,
                suspend_resume: false,
                snapshot: false,
                domain_egress_rules: false,
                egress_deny: true,
                enforced_limits: true,
                // Docker's `--pids-limit`.
                process_limit: true,
                session_lifetime: false,
                // Local has no in-sandbox agent: the manager drives Docker from outside, so
                // there is no supervisor inside the sandbox to isolate from.
                supervisor_pid_namespace: false,
                // The supervisor is the manager on the host, outside the container entirely, and
                // `docker exec` runs the command as the workload uid — a different identity by
                // construction.
                supervisor_isolation: true,
            }),
            Platform::Machines | Platform::Test => {
                Err(AlienError::new(ErrorData::SandboxPlatformUnsupported {
                    platform: platform.to_string(),
                }))
            }
        }
    }

    /// What the GCP Agent Platform sandbox backend supports; the body of the `Platform::Gcp` arm.
    pub fn gcp_agent_platform() -> Self {
        Self {
            // Agent file operations move over the session envelope.
            files: true,
            // Reaching a session across processes is safe because `generation` is derived from the
            // container boot id read through the agent's health op, so a caller detects a container
            // replaced under a stable session name rather than reconnecting to a blank one.
            reconnect: true,
            // No method mints a port-scoped ingress capability; the only ingress is `:execute`.
            preview: false,
            // `:pause` and `:resume` preserve the running container.
            suspend_resume: true,
            // A session's state can be captured and used to create another.
            snapshot: true,
            // Egress is shaped by VPC and DNS peering, which is not a hostname allowlist.
            domain_egress_rules: false,
            // A declared `deny` blocks both routed egress and DNS.
            egress_deny: true,
            // The declared ceilings are enforced, but by terminating the session on breach rather
            // than by refusing the allocation — a caller reading `true` should expect the session
            // to die, not a clean error at the point of the request.
            enforced_limits: true,
            // No ceiling on process count is observed.
            process_limit: false,
            // `ttl` maps to a session `expireTime` the platform terminates at.
            session_lifetime: true,
            // No PID-namespace isolation between the command and anything supervising it.
            supervisor_pid_namespace: false,
            // No separate supervisor identity: the command is not run under a different identity
            // than the process supervising it.
            supervisor_isolation: false,
        }
    }

    /// Returns a typed error if the named capability is absent on this platform.
    pub fn require(&self, capability: SandboxCapability, platform: Platform) -> Result<()> {
        let available = match capability {
            SandboxCapability::Files => self.files,
            SandboxCapability::Reconnect => self.reconnect,
            SandboxCapability::Preview => self.preview,
            SandboxCapability::SuspendResume => self.suspend_resume,
            SandboxCapability::Snapshot => self.snapshot,
            SandboxCapability::DomainEgressRules => self.domain_egress_rules,
            SandboxCapability::EgressDeny => self.egress_deny,
            SandboxCapability::EnforcedLimits => self.enforced_limits,
            SandboxCapability::ProcessLimit => self.process_limit,
            SandboxCapability::SessionLifetime => self.session_lifetime,
            SandboxCapability::SupervisorPidNamespace => self.supervisor_pid_namespace,
            SandboxCapability::SupervisorIsolation => self.supervisor_isolation,
        };

        if available {
            return Ok(());
        }

        Err(AlienError::new(ErrorData::SandboxCapabilityUnsupported {
            capability: capability.as_str().to_string(),
            platform: platform.to_string(),
        }))
    }
}

/// Names a single sandbox capability, so an unsupported call can report which one it needed.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase")]
pub enum SandboxCapability {
    /// Moving files in and out of a session
    Files,
    /// Reaching a session created by an earlier call
    Reconnect,
    /// An authenticated, port-scoped ingress capability
    Preview,
    /// Suspending and resuming session state
    SuspendResume,
    /// Capturing full session state
    Snapshot,
    /// Restricting egress to a hostname allowlist
    DomainEgressRules,
    /// Refusing outbound access when a sandbox declares none
    EgressDeny,
    /// Platform-enforced resource ceilings
    EnforcedLimits,
    /// A ceiling on the number of processes a session may run
    ProcessLimit,
    /// A wall-clock ceiling on a session, applied by the platform rather than by a caller
    SessionLifetime,
    /// A command runs in its own PID namespace, isolated from the agent supervising it
    SupervisorPidNamespace,
    /// A command runs under a different identity than the process supervising it
    SupervisorIsolation,
}

impl SandboxCapability {
    /// Returns the stable identifier used in errors and capability queries.
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::Files => "files",
            Self::Reconnect => "reconnect",
            Self::Preview => "preview",
            Self::SuspendResume => "suspendResume",
            Self::Snapshot => "snapshot",
            Self::DomainEgressRules => "domainEgressRules",
            Self::EgressDeny => "egressDeny",
            Self::EnforcedLimits => "enforcedLimits",
            Self::ProcessLimit => "processLimit",
            Self::SessionLifetime => "sessionLifetime",
            Self::SupervisorPidNamespace => "supervisorPidNamespace",
            Self::SupervisorIsolation => "supervisorIsolation",
        }
    }
}

/// An isolated environment for running untrusted code, created per session at runtime.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Builder)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase", deny_unknown_fields)]
#[builder(start_fn = new)]
pub struct Sandbox {
    /// Identifier for the sandbox. Must contain only alphanumeric characters, hyphens, and
    /// underscores ([A-Za-z0-9-_]). Maximum 64 characters.
    #[builder(start_fn)]
    pub id: String,
    /// Where the sandbox's root filesystem comes from
    pub code: SandboxCode,
    /// Enforced resource ceilings.
    ///
    /// Optional because not every platform can enforce them, and a declaration that names none
    /// takes the platform's own defaults. Naming them on a platform that cannot enforce them is
    /// rejected at plan time rather than silently ignored.
    #[serde(skip_serializing_if = "Option::is_none")]
    pub limits: Option<SandboxLimits>,
    /// Outbound network policy
    pub egress: SandboxEgress,
    /// Session lifetime and idle behaviour
    pub session: SandboxSessionPolicy,
    /// Ports eligible for a preview capability. A port not listed here can never be exposed,
    /// so an application cannot widen its own ingress at runtime.
    #[builder(default)]
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub preview_ports: Vec<u16>,
}

/// Whether the artifact being rendered restricts which network modes it accepts.
///
/// A sandbox is not emitted on a Kubernetes target, so nothing there routes egress through a
/// connector and the default network stays a working answer. Every site that withholds the mode,
/// explains the restriction, or renders a branch for it has to ask this one question — asking the
/// stack directly is how they came to disagree.
pub fn restricts_network_mode(stack: &crate::Stack, targets_kubernetes: bool) -> bool {
    !targets_kubernetes && stack_needs_named_subnets_at_setup(stack)
}

/// Whether any sandbox in the stack forces setup to name subnets.
///
/// A restricted sandbox routes session egress through a VPC connector, and neither generator can
/// enumerate the account default VPC's subnets, so that mode leaves the connector without any and
/// it fails at create. Callers that render an artifact want [`restricts_network_mode`] instead:
/// this one answers for the declaration, which on a Kubernetes target is not what gets emitted.
pub fn stack_needs_named_subnets_at_setup(stack: &crate::Stack) -> bool {
    stack.resources().any(|(_resource_id, resource)| {
        resource
            .config
            .downcast_ref::<Sandbox>()
            .is_some_and(|sandbox| !matches!(sandbox.egress, SandboxEgress::Allow))
    })
}

impl Sandbox {
    /// The resource type identifier for Sandbox
    pub const RESOURCE_TYPE: ResourceType = ResourceType::from_static("sandbox");

    /// Returns the sandbox's unique identifier.
    pub fn id(&self) -> &str {
        &self.id
    }

    /// The declared ceilings, or the defaults a platform applies when none were named.
    ///
    /// Backends want a concrete set: a sandbox with no declared ceilings still runs inside
    /// whatever the platform gives it, and a backend that had to branch on `None` would end up
    /// inventing its own default anyway.
    pub fn resolved_limits(&self) -> SandboxLimits {
        self.limits.clone().unwrap_or_else(default_limits)
    }

    /// Validates the declaration against what the target platform can enforce.
    ///
    /// Runs at plan time so an unenforceable limit or an unsupported egress mode fails before
    /// anything is provisioned, rather than at the first exec.
    pub fn validate_for_platform(&self, platform: Platform) -> Result<()> {
        let capabilities = SandboxCapabilities::for_platform(platform)?;

        // No backend builds a sandbox image from source: an empty image string schedules a pod
        // that can never run, the silent no-op the capability contract forbids — the failure
        // has to land here instead.
        if let SandboxCode::Source { .. } = &self.code {
            return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "code".to_string(),
                value: "source".to_string(),
                reason: "no sandbox backend builds an image from source yet; give code.image a \
                         prebuilt reference"
                    .to_string(),
            }));
        }

        // Read before the limits, because the image is declared whether or not any are.
        if platform == Platform::Azure {
            self.azure_catalog_image()?;
        }

        let Some(limits) = self.limits.as_ref() else {
            // Nothing declared, so nothing to enforce and nothing to reject.
            return self.validate_capabilities(&capabilities, platform);
        };

        validate_quantity(&self.id, "cpu", &limits.cpu)?;
        validate_quantity(&self.id, "memory", &limits.memory)?;
        validate_quantity(&self.id, "disk", &limits.disk)?;

        if let Some(max_processes) = limits.max_processes {
            if max_processes == 0 {
                return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                    resource_id: self.id.clone(),
                    field: "maxProcesses".to_string(),
                    value: "0".to_string(),
                    reason: "a sandbox that may run no processes cannot run code".to_string(),
                }));
            }
            capabilities.require(SandboxCapability::ProcessLimit, platform)?;
        }

        // Declaring limits a platform ignores is worse than not declaring them: the stack reads
        // as bounded while the sandbox is not.
        capabilities.require(SandboxCapability::EnforcedLimits, platform)?;

        if platform == Platform::Aws {
            // Refused here rather than at emit so a customer sees it while planning, and so both
            // package formats inherit the same answer.
            self.microvm_tier()?;

            // The ceiling is Lambda's, and it rejects the run rather than clamping — so a value
            // outside it would pass planning, render into the package, and fail at the first
            // session. Kubernetes takes the same field with no such bound, which is why this
            // sits under the AWS gate rather than on the type.
            if let Some(seconds) = self.session.max_lifetime_seconds {
                if !(1..=AWS_MAX_SESSION_LIFETIME_SECONDS).contains(&seconds) {
                    return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                        resource_id: self.id.clone(),
                        field: "maxLifetimeSeconds".to_string(),
                        value: seconds.to_string(),
                        reason: format!(
                            "AWS runs a MicroVM for between 1 and \
                             {AWS_MAX_SESSION_LIFETIME_SECONDS} seconds"
                        ),
                    }));
                }
            }
        }

        self.validate_capabilities(&capabilities, platform)
    }

    /// The catalog disk image Azure creates a session from.
    ///
    /// Azure names a public catalog entry rather than pulling a reference, so a registry path,
    /// tag or digest has nowhere to go. An allowlist, because the answer to "what else could be
    /// in there" is a name the data plane rejects at the first session, long after the apply.
    pub fn azure_catalog_image(&self) -> Result<&str> {
        let refused = |value: &str, reason: &str| {
            AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "code.image".to_string(),
                value: value.to_string(),
                reason: reason.to_string(),
            })
        };

        let SandboxCode::Image { image } = &self.code else {
            return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "code".to_string(),
                value: "source".to_string(),
                reason: "no sandbox backend builds an image from source yet".to_string(),
            }));
        };

        let image = image.trim();
        if image.is_empty() {
            return Err(refused(image, "a sandbox has to name an image"));
        }
        if !image
            .chars()
            .all(|c| c.is_ascii_alphanumeric() || matches!(c, '.' | '_' | '-'))
        {
            return Err(refused(
                image,
                "Azure creates a session from a public catalog disk image, so code.image must be \
                 a bare catalog name such as 'ubuntu'",
            ));
        }
        Ok(image)
    }

    /// The MicroVM size that keeps every declared ceiling, or why none does.
    ///
    /// AWS sizes are discrete and a running MicroVM bursts to four times its baseline, so the
    /// only tier that honours a ceiling is one whose peak fits inside it. A declaration no tier
    /// satisfies is refused: shipping the nearest size would give the customer a sandbox that
    /// exceeds the bound they wrote down.
    pub fn microvm_tier(&self) -> Result<MicrovmTier> {
        let Some(limits) = self.limits.as_ref() else {
            // Nothing declared: AWS's own default baseline, which is also `default_limits`.
            return Ok(MICROVM_TIERS[2]);
        };

        let memory_mib = quantity_mib(&limits.memory).ok_or_else(|| {
            AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "memory".to_string(),
                value: limits.memory.clone(),
                reason: "AWS sizes a MicroVM in whole MiB".to_string(),
            })
        })?;
        let disk_mib = quantity_mib(&limits.disk).ok_or_else(|| {
            AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "disk".to_string(),
                value: limits.disk.clone(),
                reason: "AWS sizes a MicroVM's disk in whole MiB".to_string(),
            })
        })?;
        let cpu_millicores = millicores(&limits.cpu).ok_or_else(|| {
            AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "cpu".to_string(),
                value: limits.cpu.clone(),
                reason: "expected cores or millicores".to_string(),
            })
        })?;

        // Memory and disk choose the size; cpu is then checked rather than used to choose.
        // AWS couples cpu to memory at 2 GB per vCPU, so letting a low cpu ceiling select the
        // size too would quietly hand back a machine four times smaller than the memory ceiling
        // asked for, with nothing to indicate it.
        let sized = |tier: &&MicrovmTier| {
            tier.peak_memory_mib <= memory_mib && tier.max_disk_mib <= disk_mib
        };

        let tier = MICROVM_TIERS
            .iter()
            .rev()
            .find(sized)
            .copied()
            .ok_or_else(|| {
                AlienError::new(ErrorData::SandboxLimitInvalid {
                    resource_id: self.id.clone(),
                    field: "memory".to_string(),
                    value: limits.memory.clone(),
                    reason: format!(
                        "a Lambda MicroVM bursts to four times its baseline, so the smallest \
                         ceiling AWS can hold is 2Gi memory with 8Gi disk; '{}' memory and '{}' \
                         disk fit no size",
                        limits.memory, limits.disk
                    ),
                })
            })?;

        let required_millicores = i64::from(tier.peak_vcpu) * 1000;
        if cpu_millicores < required_millicores {
            return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                resource_id: self.id.clone(),
                field: "cpu".to_string(),
                value: limits.cpu.clone(),
                reason: format!(
                    "AWS allocates one vCPU per 2GB, so a MicroVM sized to a '{}' memory ceiling \
                     reaches {} vCPU; declare cpu '{}' or lower the memory ceiling",
                    limits.memory, tier.peak_vcpu, tier.peak_vcpu
                ),
            }));
        }

        Ok(tier)
    }

    /// The capability checks that do not depend on declared limits.
    fn validate_capabilities(
        &self,
        capabilities: &SandboxCapabilities,
        platform: Platform,
    ) -> Result<()> {
        if matches!(self.egress, SandboxEgress::AllowDomains { .. }) {
            capabilities.require(SandboxCapability::DomainEgressRules, platform)?;
        }

        // `allow` asks for no restriction, so a backend that ignores it fails loudly on the first
        // blocked connection. `deny` asks for one, and a backend that ignores it puts untrusted
        // code on the internet with nothing to notice — so only this direction is gated.
        // An empty list is not a restriction anyone wrote down: it renders as a deny-all wearing
        // an allowlist's label, which reads at a glance as the opposite of what it does.
        if let SandboxEgress::AllowDomains { domains } = &self.egress {
            if domains.is_empty() {
                return Err(AlienError::new(ErrorData::SandboxLimitInvalid {
                    resource_id: self.id.clone(),
                    field: "egress.domains".to_string(),
                    value: "[]".to_string(),
                    reason: "an allowlist naming no domain denies everything; declare \
                             egress: deny if that is what was meant"
                        .to_string(),
                }));
            }
        }

        if matches!(self.egress, SandboxEgress::Deny) {
            capabilities.require(SandboxCapability::EgressDeny, platform)?;
        }

        if !self.preview_ports.is_empty() {
            capabilities.require(SandboxCapability::Preview, platform)?;
        }

        if self.session.idle_suspend_seconds.is_some() {
            capabilities.require(SandboxCapability::SuspendResume, platform)?;
        }

        if self.session.max_lifetime_seconds.is_some() {
            capabilities.require(SandboxCapability::SessionLifetime, platform)?;
        }

        Ok(())
    }
}

/// Ceilings applied when a declaration names none.
///
/// Modest on purpose: an undeclared sandbox is one whose author did not think about sizing, and
/// the safe reading of that is a small box rather than a generous one.
fn default_limits() -> SandboxLimits {
    SandboxLimits {
        cpu: "1".to_string(),
        memory: "2Gi".to_string(),
        disk: "8Gi".to_string(),
        max_processes: None,
    }
}

/// Validates a Kubernetes-style resource quantity such as `500m`, `2Gi` or `1`.
fn validate_quantity(resource_id: &str, field: &str, value: &str) -> Result<()> {
    let invalid = |reason: &str| {
        AlienError::new(ErrorData::SandboxLimitInvalid {
            resource_id: resource_id.to_string(),
            field: field.to_string(),
            value: value.to_string(),
            reason: reason.to_string(),
        })
    };

    let digits_end = value
        .find(|c: char| !c.is_ascii_digit() && c != '.')
        .unwrap_or(value.len());
    let (number, suffix) = value.split_at(digits_end);

    let parsed: f64 = number
        .parse()
        .map_err(|_| invalid("expected a number, optionally followed by a unit suffix"))?;

    if parsed <= 0.0 {
        return Err(invalid("must be greater than zero"));
    }

    const SUFFIXES: &[&str] = &["", "m", "k", "M", "G", "T", "Ki", "Mi", "Gi", "Ti"];
    if !SUFFIXES.contains(&suffix) {
        return Err(invalid(
            "unit must be one of m, k, M, G, T, Ki, Mi, Gi, Ti, or absent",
        ));
    }

    Ok(())
}

/// Splits a quantity into its number and unit suffix.
fn split_quantity(value: &str) -> Option<(f64, &str)> {
    let trimmed = value.trim();
    let digits_end = trimmed
        .find(|c: char| !c.is_ascii_digit() && c != '.')
        .unwrap_or(trimmed.len());
    let (number, suffix) = trimmed.split_at(digits_end);
    number.parse().ok().map(|number| (number, suffix))
}

/// A memory or disk quantity in whole MiB, rounded down.
///
/// Every suffix `validate_quantity` accepts is handled here. Reading only `Gi` and `Mi` and
/// falling back for the rest would turn a declared `4G` into a different size than the customer
/// asked for, which for a ceiling means a sandbox larger than its bound.
pub fn quantity_mib(value: &str) -> Option<i64> {
    let (number, suffix) = split_quantity(value)?;
    let bytes = match suffix {
        "" => number,
        "k" => number * 1e3,
        "M" => number * 1e6,
        "G" => number * 1e9,
        "T" => number * 1e12,
        "Ki" => number * 1024.0,
        "Mi" => number * 1024.0 * 1024.0,
        "Gi" => number * 1024.0 * 1024.0 * 1024.0,
        "Ti" => number * 1024.0 * 1024.0 * 1024.0 * 1024.0,
        // `m` is a millicore suffix; memory has no use for it.
        _ => return None,
    };
    Some((bytes / (1024.0 * 1024.0)) as i64)
}

/// A CPU quantity in millicores.
pub fn millicores(value: &str) -> Option<i64> {
    let (number, suffix) = split_quantity(value)?;
    match suffix {
        "" => Some((number * 1000.0) as i64),
        "m" => Some(number as i64),
        _ => None,
    }
}

/// Outputs generated by a successfully provisioned Sandbox parent.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[cfg_attr(feature = "openapi", derive(utoipa::ToSchema))]
#[serde(rename_all = "camelCase")]
pub struct SandboxOutputs {
    /// Name of the durable parent that sessions are created inside
    pub parent_name: String,
    /// Platform-specific identifier for the parent (image ARN, sandbox group id, namespace)
    #[serde(skip_serializing_if = "Option::is_none")]
    pub identifier: Option<String>,
    /// Data-plane endpoint sessions are created through, where the platform has one
    #[serde(skip_serializing_if = "Option::is_none")]
    pub endpoint: Option<String>,
}

impl ResourceOutputsDefinition for SandboxOutputs {
    fn get_resource_type(&self) -> ResourceType {
        Sandbox::RESOURCE_TYPE
    }

    fn as_any(&self) -> &dyn Any {
        self
    }

    fn box_clone(&self) -> Box<dyn ResourceOutputsDefinition> {
        Box::new(self.clone())
    }

    fn outputs_eq(&self, other: &dyn ResourceOutputsDefinition) -> bool {
        other.as_any().downcast_ref::<SandboxOutputs>() == Some(self)
    }

    fn to_json_value(&self) -> serde_json::Result<serde_json::Value> {
        serde_json::to_value(self)
    }
}

impl ResourceDefinition for Sandbox {
    fn get_resource_type(&self) -> ResourceType {
        Self::RESOURCE_TYPE
    }

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

    fn get_dependencies(&self) -> Vec<ResourceRef> {
        Vec::new()
    }

    fn validate_update(&self, new_config: &dyn ResourceDefinition) -> Result<()> {
        let new_sandbox = new_config
            .as_any()
            .downcast_ref::<Sandbox>()
            .ok_or_else(|| {
                AlienError::new(ErrorData::UnexpectedResourceType {
                    resource_id: self.id.clone(),
                    expected: Self::RESOURCE_TYPE,
                    actual: new_config.get_resource_type(),
                })
            })?;

        if self.id != new_sandbox.id {
            return Err(AlienError::new(ErrorData::InvalidResourceUpdate {
                resource_id: self.id.clone(),
                reason: "the 'id' field is immutable".to_string(),
            }));
        }

        Ok(())
    }

    fn as_any(&self) -> &dyn Any {
        self
    }

    fn as_any_mut(&mut self) -> &mut dyn Any {
        self
    }

    fn box_clone(&self) -> Box<dyn ResourceDefinition> {
        Box::new(self.clone())
    }

    fn resource_eq(&self, other: &dyn ResourceDefinition) -> bool {
        other.as_any().downcast_ref::<Sandbox>() == Some(self)
    }

    fn to_json_value(&self) -> serde_json::Result<serde_json::Value> {
        serde_json::to_value(self)
    }
}

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

    fn sandbox_with(egress: SandboxEgress, preview_ports: Vec<u16>) -> Sandbox {
        Sandbox::new("agent-sbx".to_string())
            .code(SandboxCode::Image {
                image: "ubuntu".to_string(),
            })
            .limits(SandboxLimits {
                cpu: "1".to_string(),
                memory: "2Gi".to_string(),
                disk: "20Gi".to_string(),
                max_processes: None,
            })
            .egress(egress)
            .session(SandboxSessionPolicy {
                max_lifetime_seconds: None,
                idle_suspend_seconds: None,
            })
            .preview_ports(preview_ports)
            .build()
    }

    #[test]
    fn resource_type_is_stable() {
        assert_eq!(Sandbox::RESOURCE_TYPE.as_ref(), "sandbox");
    }

    #[test]
    fn capability_sets_are_per_platform() {
        let gcp = SandboxCapabilities::for_platform(Platform::Gcp).expect("gcp is supported");
        assert!(
            gcp.reconnect,
            "generation from the container boot id makes a session reachable across processes"
        );
        assert!(!gcp.preview);
        assert!(gcp.enforced_limits);

        let azure = SandboxCapabilities::for_platform(Platform::Azure).expect("azure is supported");
        assert!(azure.files, "every backend moves files");
        assert!(gcp.files);
        // Azure is the only backend whose egress policy matches on host pattern, and the only
        // one where `deny` and a hostname list are the same object.
        assert!(azure.domain_egress_rules);
        assert!(azure.egress_deny);
        // The data plane takes no ceiling, so a declaration of one is refused rather than
        // accepted and dropped.
        assert!(!azure.enforced_limits);
        assert!(azure.suspend_resume);
        // Both stay false for reasons that are not "unbuilt": a snapshot id has nothing to
        // consume it on any backend, and an Azure port's auth is anonymous or a human allowlist,
        // neither of which is a port-scoped credential.
        assert!(!azure.snapshot);
        assert!(!azure.preview);

        let aws = SandboxCapabilities::for_platform(Platform::Aws).expect("aws is supported");
        assert!(!aws.snapshot, "AWS has no user-callable session snapshot");
        assert!(aws.suspend_resume);

        let k8s =
            SandboxCapabilities::for_platform(Platform::Kubernetes).expect("k8s is supported");
        assert!(
            !k8s.preview,
            "the session-scoped ingress gateway does not exist yet"
        );
    }

    /// Whether the process supervising a command is a separate identity from the command.
    ///
    /// Values are measured, not inferred. AWS: the agent runs as uid 0 with
    /// `CapEff: 00000000a80425fb` and `setuid`s the command to uid 60000, so the two differ.
    /// Kubernetes: the sandbox pod pins `run_as_user: 65534` on both pod and container with
    /// `capabilities.drop: [ALL]` and `allow_privilege_escalation: false`, so no uid split is
    /// possible (`kubernetes_spec.rs`). Local: `docker exec` runs as the workload uid while the
    /// manager supervises from the host. Azure and Agent Platform have no in-sandbox supervisor.
    #[test]
    fn supervisor_isolation_is_per_platform() {
        let value = |platform| {
            SandboxCapabilities::for_platform(platform)
                .expect("supported")
                .supervisor_isolation
        };

        assert!(
            value(Platform::Aws),
            "root agent setuids the command to 60000"
        );
        assert!(
            value(Platform::Local),
            "the supervisor is on the host, outside the container"
        );
        assert!(
            !value(Platform::Kubernetes),
            "a single pinned uid cannot be split"
        );
        assert!(!value(Platform::Azure), "no Alien process runs the command");
        assert!(
            !value(Platform::Gcp),
            "no separate supervisor identity runs the command"
        );
    }

    /// The point of the field: AWS and GCP report the *same* `supervisor_pid_namespace` (neither
    /// has `CAP_SYS_ADMIN`), so that axis alone reads them as equivalent. They are not — AWS
    /// separates the command's identity from the supervisor's and Agent Platform does not.
    #[test]
    fn supervisor_isolation_separates_aws_from_a_subprocess_backend() {
        let aws = SandboxCapabilities::for_platform(Platform::Aws).expect("aws is supported");
        let gcp = SandboxCapabilities::for_platform(Platform::Gcp).expect("gcp is supported");

        assert_eq!(
            aws.supervisor_pid_namespace, gcp.supervisor_pid_namespace,
            "the older axis cannot tell them apart"
        );
        assert!(
            aws.supervisor_isolation,
            "AWS setuids the command off the supervisor"
        );
        assert!(
            !gcp.supervisor_isolation,
            "the command runs under no separate supervisor identity"
        );
    }

    /// The Agent Platform row, each value against the behaviour it was measured from. `reconnect`
    /// is the tripwire: it is `true` only because `generation` is derived from the container boot
    /// id read through the agent's health op, so a caller detects a replaced container instead of
    /// reconnecting to a blank one. It is also the body of the `Platform::Gcp` arm, asserted below.
    #[test]
    fn gcp_agent_platform_row_matches_measured_backend() {
        let row = SandboxCapabilities::gcp_agent_platform();

        assert!(row.files, "agent file ops move over the session envelope");
        assert!(
            row.reconnect,
            "generation is derived from the container boot id, so a session is reachable across \
             processes"
        );
        assert!(
            !row.preview,
            "the only ingress is :execute; no port-scoped capability"
        );
        assert!(
            row.suspend_resume,
            ":pause and :resume preserve the container"
        );
        assert!(
            row.snapshot,
            "session state can be captured and restored into a new session"
        );
        assert!(
            !row.domain_egress_rules,
            "VPC and DNS peering is not a hostname allowlist"
        );
        assert!(
            row.egress_deny,
            "a declared deny blocks both egress and DNS"
        );
        assert!(
            row.enforced_limits,
            "ceilings are enforced, by terminating the session on breach"
        );
        assert!(!row.process_limit, "no process-count ceiling is observed");
        assert!(row.session_lifetime, "ttl maps to a session expireTime");
        assert!(!row.supervisor_pid_namespace, "no PID-namespace isolation");
        assert!(
            !row.supervisor_isolation,
            "the command is not run under a separate supervisor identity"
        );

        // Agent Platform is the registered GCP backend, so the arm returns exactly this row.
        let live = SandboxCapabilities::for_platform(Platform::Gcp).expect("gcp is supported");
        assert_eq!(
            live, row,
            "the Platform::Gcp arm is the Agent Platform capability row"
        );
    }

    #[test]
    fn platforms_without_a_backend_are_an_error_not_an_empty_set() {
        let error = SandboxCapabilities::for_platform(Platform::Machines)
            .expect_err("Machines has no sandbox backend");
        assert_eq!(error.code, "SANDBOX_PLATFORM_UNSUPPORTED");
    }

    #[test]
    fn unsupported_capability_names_platform_and_capability() {
        let capabilities = SandboxCapabilities::for_platform(Platform::Gcp).expect("supported");
        let error = capabilities
            .require(SandboxCapability::Preview, Platform::Gcp)
            .expect_err("GCP has no preview");

        assert_eq!(error.code, "SANDBOX_CAPABILITY_UNSUPPORTED");
        let rendered = error.to_string();
        assert!(
            rendered.contains("preview"),
            "names the capability: {rendered}"
        );
        assert!(rendered.contains("gcp"), "names the platform: {rendered}");
    }

    /// Azure matches on hostname; AWS and Kubernetes match CIDRs, and Local and GCP have a
    /// switch rather than a filter. Accepting a hostname list on those four would leave a stack
    /// reading as restricted while the sandbox reaches the whole internet.
    #[test]
    fn a_hostname_allowlist_is_refused_everywhere_it_would_be_approximated() {
        let sandbox = sandbox_with(
            SandboxEgress::AllowDomains {
                domains: vec!["example.com".to_string()],
            },
            vec![],
        );

        for platform in [
            Platform::Aws,
            Platform::Gcp,
            Platform::Kubernetes,
            Platform::Local,
        ] {
            let error = sandbox
                .validate_for_platform(platform)
                .expect_err("only Azure expresses a hostname allowlist");
            assert_eq!(
                error.code, "SANDBOX_CAPABILITY_UNSUPPORTED",
                "on {platform:?}"
            );
        }

        assert!(
            SandboxCapabilities::for_platform(Platform::Azure)
                .expect("supported")
                .domain_egress_rules,
            "Azure's egress policy matches on host pattern"
        );
    }

    /// `deny` is the declaration that carries a security promise, so a backend that cannot keep
    /// it has to refuse rather than accept it and run the code with open egress.
    #[test]
    fn a_denied_egress_is_refused_where_it_would_not_be_enforced() {
        let sandbox = sandbox_with(SandboxEgress::Deny, vec![]);

        // GCP is asserted at the capability rather than through validation: this sandbox declares
        // ceilings GCP cannot enforce, so it is refused for a reason unrelated to egress.
        assert!(
            SandboxCapabilities::for_platform(Platform::Gcp)
                .expect("supported")
                .egress_deny
        );

        for platform in [Platform::Aws, Platform::Kubernetes, Platform::Local] {
            sandbox
                .validate_for_platform(platform)
                .expect("deny is enforced here");
        }

        // Declares no ceilings, which Azure refuses for its own reason, so this isolates egress.
        let egress_only = Sandbox::new("sbx".to_string())
            .code(SandboxCode::Image {
                image: "alpine".to_string(),
            })
            .egress(SandboxEgress::Deny)
            .session(SandboxSessionPolicy {
                max_lifetime_seconds: None,
                idle_suspend_seconds: None,
            })
            .build();

        egress_only
            .validate_for_platform(Platform::Azure)
            .expect("Azure creates the sandbox under a Deny policy with full inspection");
    }

    /// Ceilings are rejected per-platform where unsupported — rejected when *declared*. With
    /// limits mandatory that would read as "GCP sandboxes cannot exist", contradicting the
    /// create, exec, files and terminate GCP does support.
    #[test]
    fn a_platform_that_cannot_enforce_limits_still_takes_a_sandbox_without_them() {
        let declared = sandbox_with(SandboxEgress::Deny, Vec::new());
        declared
            .validate_for_platform(Platform::Azure)
            .expect_err("declaring ceilings Azure cannot enforce is rejected");

        let undeclared = Sandbox::new("sbx".to_string())
            .code(SandboxCode::Image {
                image: "alpine".to_string(),
            })
            .egress(SandboxEgress::Deny)
            .session(SandboxSessionPolicy {
                max_lifetime_seconds: None,
                idle_suspend_seconds: None,
            })
            .build();

        undeclared
            .validate_for_platform(Platform::Azure)
            .expect("a sandbox naming no ceilings takes the platform's own");

        // A backend still gets a concrete set, so nothing downstream has to invent one.
        assert_eq!(undeclared.resolved_limits().cpu, "1");
    }

    #[test]
    fn preview_ports_require_the_preview_capability() {
        let sandbox = sandbox_with(SandboxEgress::Deny, vec![8080]);

        sandbox
            .validate_for_platform(Platform::Aws)
            .expect("AWS mints a port-scoped JWE");

        let error = sandbox
            .validate_for_platform(Platform::Kubernetes)
            .expect_err("Kubernetes preview is deferred");
        assert_eq!(error.code, "SANDBOX_CAPABILITY_UNSUPPORTED");
    }

    #[test]
    fn gcp_accepts_a_sandbox_declaring_enforced_limits() {
        let sandbox = sandbox_with(SandboxEgress::Allow, vec![]);
        sandbox
            .validate_for_platform(Platform::Gcp)
            .expect("Agent Platform enforces declared ceilings, by terminating on breach");
    }

    #[test]
    fn invalid_quantities_are_rejected_with_the_offending_field() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .memory = "2Gb".to_string();

        let error = sandbox
            .validate_for_platform(Platform::Aws)
            .expect_err("Gb is not a valid suffix");
        assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");
        assert!(error.to_string().contains("memory"));

        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .memory = "2Gi".to_string();
        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .cpu = "0".to_string();
        let error = sandbox
            .validate_for_platform(Platform::Aws)
            .expect_err("zero cpu is not a ceiling");
        assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");
    }

    #[test]
    fn zero_max_processes_is_rejected() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .max_processes = Some(0);

        let error = sandbox
            .validate_for_platform(Platform::Local)
            .expect_err("a sandbox must be able to run at least one process");
        assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");
        assert!(error.to_string().contains("maxProcesses"));
    }

    /// A process ceiling needs a container runtime. Kubernetes sets one per node rather than per
    /// pod, and neither MicroVMs nor Azure sandboxes expose one, so accepting the declaration
    /// anywhere else would mean carrying a bound nothing applies.
    #[test]
    fn a_process_ceiling_is_accepted_only_where_a_runtime_can_apply_it() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .max_processes = Some(256);

        sandbox
            .validate_for_platform(Platform::Local)
            .expect("Docker takes a pids limit");

        for platform in [Platform::Aws, Platform::Azure, Platform::Kubernetes] {
            let error = sandbox
                .validate_for_platform(platform)
                .expect_err("a process ceiling nothing applies must be refused");
            assert_eq!(error.code, "SANDBOX_CAPABILITY_UNSUPPORTED");
        }
    }

    /// Lambda rejects a run outside 1–28,800 rather than clamping it, so a value beyond that
    /// would pass planning, render into the package, and fail at the first session. Kubernetes
    /// takes the same field with no such bound, so the check is AWS's alone.
    #[test]
    fn a_lifetime_aws_would_reject_is_refused_while_planning() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);

        for seconds in [0, 28_801, 100_000] {
            sandbox.session.max_lifetime_seconds = Some(seconds);
            let error = sandbox
                .validate_for_platform(Platform::Aws)
                .expect_err("a lifetime outside what AWS runs is refused");
            assert_eq!(error.code, "SANDBOX_LIMIT_INVALID", "{seconds}s");

            // Kubernetes has no such ceiling, so the same declaration is fine there.
            sandbox
                .validate_for_platform(Platform::Kubernetes)
                .expect("the kubelet takes any activeDeadlineSeconds");
        }

        sandbox.session.max_lifetime_seconds = Some(28_800);
        sandbox
            .validate_for_platform(Platform::Aws)
            .expect("the ceiling itself is allowed");
    }

    /// An image reference Azure cannot honour is refused while planning, not at the first session.
    ///
    /// `code.image`'s own documentation gives a tag and a registry path as examples — exactly
    /// what Azure cannot take, so this is the shape a customer is most likely to declare.
    #[test]
    fn an_image_azure_cannot_pull_is_refused_while_planning() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        // Azure enforces no declared ceiling, so a sandbox carrying limits is refused before the
        // image is ever read.
        sandbox.limits = None;

        for image in [
            "ubuntu:24.04",
            "ghcr.io/myorg/sandbox:latest",
            "ubuntu@sha256:abc",
            "",
            "   ",
            "ubuntu latest",
            "ubuntu?x",
        ] {
            sandbox.code = SandboxCode::Image {
                image: image.to_string(),
            };
            let error = sandbox
                .validate_for_platform(Platform::Azure)
                .expect_err("an image Azure has nowhere to put is refused");
            assert_eq!(error.code, "SANDBOX_LIMIT_INVALID", "image '{image}'");

            // The same declaration is ordinary everywhere that pulls a reference.
            sandbox
                .validate_for_platform(Platform::Kubernetes)
                .expect("a registry reference is what every other backend takes");
        }

        for image in ["ubuntu", "ubuntu-22.04", "debian_slim"] {
            sandbox.code = SandboxCode::Image {
                image: image.to_string(),
            };
            sandbox
                .validate_for_platform(Platform::Azure)
                .unwrap_or_else(|error| panic!("'{image}' is a catalog name: {error}"));
        }

        // Surrounding space is trimmed rather than carried into the create body.
        sandbox.code = SandboxCode::Image {
            image: " ubuntu ".to_string(),
        };
        assert_eq!(
            sandbox
                .azure_catalog_image()
                .expect("a padded name is still a name"),
            "ubuntu"
        );
    }

    /// A deadline is accepted only where the platform itself terminates on it — the kubelet's
    /// `activeDeadlineSeconds` and Lambda's `maximumDurationInSeconds`. Everywhere else it would
    /// need a reaper that does not exist, so it is refused rather than accepted and dropped.
    #[test]
    fn a_session_deadline_is_accepted_only_where_the_platform_applies_it() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        sandbox.session.max_lifetime_seconds = Some(3600);

        sandbox
            .validate_for_platform(Platform::Kubernetes)
            .expect("the kubelet enforces activeDeadlineSeconds");
        sandbox
            .validate_for_platform(Platform::Aws)
            .expect("Lambda terminates the MicroVM at maximumDurationInSeconds");

        for platform in [Platform::Azure, Platform::Local] {
            let error = sandbox
                .validate_for_platform(platform)
                .expect_err("a deadline nothing applies must be refused");
            assert_eq!(error.code, "SANDBOX_CAPABILITY_UNSUPPORTED");
        }
    }

    /// A MicroVM bursts to four times its baseline with no way to opt out, so a ceiling is kept
    /// by choosing the size whose *peak* fits inside it. Sizing by baseline would hand back a
    /// sandbox that can reach four times what the customer declared.
    #[test]
    fn an_aws_size_is_chosen_so_its_peak_stays_inside_the_declared_ceiling() {
        let sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        let tier = sandbox
            .microvm_tier()
            .expect("2Gi/1cpu/20Gi is satisfiable");

        assert_eq!(
            tier.peak_memory_mib, 2048,
            "the peak is the declared ceiling"
        );
        assert_eq!(
            tier.baseline_memory_mib, 512,
            "which is a quarter of it as the baseline"
        );
        assert!(tier.max_disk_mib <= 20 * 1024);
    }

    /// AWS allocates one vCPU per 2GB, so a cpu ceiling below what the memory ceiling implies
    /// cannot be honoured together with it. Letting cpu choose the size instead would hand back a
    /// machine four times smaller than the memory asked for, with nothing to indicate it.
    #[test]
    fn a_cpu_ceiling_below_what_the_memory_implies_is_refused_not_quietly_downsized() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        {
            let limits = sandbox
                .limits
                .as_mut()
                .expect("the fixture declares limits");
            limits.cpu = "1".to_string();
            limits.memory = "8Gi".to_string();
        }

        let error = sandbox
            .microvm_tier()
            .expect_err("1 cpu and 8Gi cannot both be ceilings on AWS");
        assert!(
            error.to_string().contains("4 vCPU"),
            "the refusal must say what the memory ceiling implies: {error}"
        );

        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .cpu = "4".to_string();
        let tier = sandbox.microvm_tier().expect("4 cpu matches 8Gi");
        assert_eq!(tier.peak_memory_mib, 8192);
    }

    /// Below AWS's smallest peak there is no size that holds the ceiling, and rounding up to the
    /// nearest one would silently exceed it.
    #[test]
    fn an_aws_ceiling_smaller_than_any_size_is_refused_rather_than_rounded() {
        let mut sandbox = sandbox_with(SandboxEgress::Deny, vec![]);
        sandbox
            .limits
            .as_mut()
            .expect("the fixture declares limits")
            .memory = "1Gi".to_string();

        let error = sandbox
            .validate_for_platform(Platform::Aws)
            .expect_err("no MicroVM size peaks at or below 1Gi");
        assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");
        assert!(
            error.to_string().contains("2Gi"),
            "the refusal must say what the smallest holdable ceiling is: {error}"
        );
    }

    /// `Source` is a public part of the type that no backend builds: an empty image string
    /// schedules a pod that can never run, so the refusal has to happen at plan time and on
    /// every platform, not in one emitter.
    #[test]
    fn source_code_is_refused_everywhere_rather_than_producing_a_broken_manifest() {
        let sandbox = Sandbox::new("agent".to_string())
            .code(SandboxCode::Source {
                src: "./sandbox".to_string(),
                toolchain: ToolchainConfig::Docker {
                    dockerfile: None,
                    build_args: None,
                    target: None,
                },
            })
            .egress(SandboxEgress::Deny)
            .session(SandboxSessionPolicy {
                max_lifetime_seconds: None,
                idle_suspend_seconds: None,
            })
            .build();

        for platform in [
            Platform::Aws,
            Platform::Azure,
            Platform::Gcp,
            Platform::Kubernetes,
            Platform::Local,
        ] {
            let error = sandbox
                .validate_for_platform(platform)
                .expect_err("no backend builds a sandbox image from source");
            assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");
            assert!(
                error.to_string().contains("code.image"),
                "the refusal must say what to write instead: {error}"
            );
        }
    }

    /// `validate_quantity` accepts nine suffixes. Reading only `Gi` and `Mi` would size a
    /// declared `4G` as though it were `4Gi`, which for a ceiling means exceeding it.
    #[test]
    fn every_accepted_unit_converts_rather_than_falling_back() {
        assert_eq!(quantity_mib("2Gi"), Some(2048));
        assert_eq!(quantity_mib("512Mi"), Some(512));
        assert_eq!(quantity_mib("4G"), Some(3814));
        assert_eq!(quantity_mib("1Ti"), Some(1024 * 1024));
        assert_eq!(millicores("1"), Some(1000));
        assert_eq!(millicores("500m"), Some(500));
    }

    #[test]
    fn unknown_fields_are_rejected() {
        let json = r#"{
            "id": "sbx",
            "code": {"type": "image", "image": "ubuntu:24.04"},
            "limits": {"cpu": "1", "memory": "2Gi", "disk": "20Gi"},
            "egress": {"mode": "deny"},
            "session": {},
            "unexpected": true
        }"#;

        serde_json::from_str::<Sandbox>(json).expect_err("deny_unknown_fields must reject");
    }

    #[test]
    fn serialization_roundtrips() {
        let sandbox = sandbox_with(
            SandboxEgress::AllowDomains {
                domains: vec!["example.com".to_string()],
            },
            vec![8080, 9090],
        );

        let json = serde_json::to_string(&sandbox).expect("serializes");
        let restored: Sandbox = serde_json::from_str(&json).expect("deserializes");
        assert_eq!(sandbox, restored);
    }

    #[test]
    fn id_is_immutable_across_updates() {
        let original = sandbox_with(SandboxEgress::Deny, vec![]);
        let renamed = Sandbox::new("other".to_string())
            .code(SandboxCode::Image {
                image: "ubuntu".to_string(),
            })
            .limits(
                original
                    .limits
                    .clone()
                    .expect("the fixture declares limits"),
            )
            .egress(SandboxEgress::Deny)
            .session(SandboxSessionPolicy {
                max_lifetime_seconds: None,
                idle_suspend_seconds: None,
            })
            .build();

        original
            .validate_update(&original.clone())
            .expect("an unchanged config is a valid update");
        original
            .validate_update(&renamed)
            .expect_err("renaming a sandbox is not an update");
    }

    /// Azure declares an idle-suspend policy but not a wall-clock ceiling.
    ///
    /// The two travel together in `SandboxSessionPolicy` and are gated separately on purpose:
    /// Azure suspends on idle and has no maximum lifetime, so accepting one and refusing the
    /// other is the honest split rather than an inconsistency.
    #[test]
    fn azure_takes_an_idle_policy_and_still_refuses_a_lifetime_ceiling() {
        let with_policy = |session: SandboxSessionPolicy| {
            Sandbox::new("sbx".to_string())
                .code(SandboxCode::Image {
                    image: "ubuntu".to_string(),
                })
                .egress(SandboxEgress::Allow)
                .session(session)
                .build()
                .validate_for_platform(Platform::Azure)
        };

        with_policy(SandboxSessionPolicy {
            max_lifetime_seconds: None,
            idle_suspend_seconds: Some(900),
        })
        .expect("Azure suspends a session on idle");

        let error = with_policy(SandboxSessionPolicy {
            max_lifetime_seconds: Some(3600),
            idle_suspend_seconds: None,
        })
        .expect_err("Azure has no wall-clock ceiling to enforce one with");
        assert_eq!(error.code, "SANDBOX_CAPABILITY_UNSUPPORTED");
        assert!(
            error.message.contains("sessionLifetime"),
            "names the capability: {}",
            error.message
        );
    }

    /// An allowlist naming nothing is a deny-all wearing an allowlist's label.
    ///
    /// It renders as a `Deny` default with no rules — the shape the Azure provider adds a
    /// catch-all to avoid — and a reader scanning the declaration sees "allowDomains" and reads
    /// the opposite of what it does.
    #[test]
    fn an_allowlist_with_no_domains_is_refused() {
        let declared = |domains: Vec<String>| {
            Sandbox::new("sbx".to_string())
                .code(SandboxCode::Image {
                    image: "ubuntu".to_string(),
                })
                .egress(SandboxEgress::AllowDomains { domains })
                .session(SandboxSessionPolicy {
                    max_lifetime_seconds: None,
                    idle_suspend_seconds: None,
                })
                .build()
                .validate_for_platform(Platform::Azure)
        };

        let error = declared(vec![]).expect_err("an empty allowlist must be refused");
        assert_eq!(error.code, "SANDBOX_LIMIT_INVALID");

        declared(vec!["api.example.com".to_string()])
            .expect("a named domain is what an allowlist is for");
    }

    /// The two expressible modes map to the boolean; a host list maps to nothing so the caller has
    /// to refuse rather than silently pick a side.
    #[test]
    fn internet_access_switch_maps_only_the_two_expressible_modes() {
        assert_eq!(SandboxEgress::Allow.internet_access_switch(), Some(true));
        assert_eq!(SandboxEgress::Deny.internet_access_switch(), Some(false));
        assert_eq!(
            SandboxEgress::AllowDomains {
                domains: vec!["api.example.com".to_string()]
            }
            .internet_access_switch(),
            None,
            "a host list has no boolean and must not be approximated"
        );
    }
}