aion_server/config/

sections.rs

//! Typed `[section]` sub-configurations that make up [`ServerConfig`].
//!
//! Each struct/enum here maps one `[section]` (or nested value) of the server
//! TOML surface, with its `serde` attributes, `Default` impl, and the small
//! amount of per-section validation that belongs with the type
//! ([`WebSocketConfig::validate`]). They are re-exported from the `config`
//! module so every existing `crate::config::X` path resolves identically.
//!
//! [`ServerConfig`]: super::ServerConfig

use std::{net::SocketAddr, path::PathBuf, time::Duration};

use serde::{Deserialize, Deserializer, de};

use crate::error::ServerError;

use super::{
    config_error,
    defaults::{
        CLUSTER_BROADCAST_CAPACITY_REQUIRED, DEFAULT_GRPC_ADDRESS, DEFAULT_HTTP_ADDRESS,
        DEFAULT_MAX_IN_FLIGHT_ACTIVITIES, DEFAULT_MCP_AWAIT_POLL_INTERVAL_MS,
        DEFAULT_MCP_DISCOVER_TTL_MS, DEFAULT_MCP_TASK_POLL_INTERVAL_MS, DEFAULT_MCP_TASK_TTL_MS,
        DEFAULT_MCP_TOOLS_LIST_TTL_MS, DEFAULT_OBSERVABILITY_MAX_EVENT_BYTES,
        DEFAULT_OBSERVABILITY_MAX_STREAM_EVENTS, EVENT_BROADCAST_CAPACITY_REQUIRED,
        MCP_AWAIT_POLL_INTERVAL_REQUIRED, MCP_TASK_POLL_INTERVAL_REQUIRED,
        OBSERVABILITY_MAX_EVENT_BYTES_REQUIRED, OBSERVABILITY_MAX_STREAM_EVENTS_REQUIRED,
    },
};

/// Public transport listener addresses from `[server]`.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct ServerSection {
    /// HTTP/JSON and dashboard listener.
    pub listen_address: SocketAddr,
    /// gRPC API and worker-protocol listener.
    pub grpc_address: SocketAddr,
    /// Browser origins allowed to make cross-origin (CORS) requests to the
    /// public HTTP API. Empty (the default) is the SECURE default: no
    /// cross-origin request is permitted and no `CorsLayer` is installed, so a
    /// same-origin deployment behaves byte-identically to before this field
    /// existed. When set, each entry is an exact origin (scheme + host + port,
    /// e.g. `http://localhost:5173`) the browser dashboard is served from; the
    /// router then answers preflight and emits `Access-Control-Allow-Origin`
    /// for exactly those origins. There is no wildcard/allow-all default
    /// (ADR-001): cross-origin access is an explicit operator decision, and the
    /// layer never pairs `Any` with credentials.
    #[serde(default)]
    pub cors_allowed_origins: Vec<String>,
}

/// Supported event-store backend names.
///
/// Deliberately NOT `Deserialize`. `StoreConfigWire` parses the operator's
/// spelling itself, case-insensitively, so that the file door and the
/// `AION_STORE_BACKEND` door accept exactly the same set — a derived
/// `rename_all = "lowercase"` impl was a second, case-SENSITIVE parser for the
/// same value that nothing called and that disagreed with the one that runs.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum StoreBackend {
    /// In-memory store for local development.
    Memory,
    /// haematite durable store (single-node, shardable).
    Haematite,
}

/// A store-selection input that named the retired libSQL backend.
///
/// Recorded at parse/overlay time and refused by `ServerConfig::validate`, which
/// names the input it found and prescribes haematite. Recorded rather than
/// refused on the spot so all five doors converge on ONE refusal with one
/// remedy, instead of five differently-worded errors from five parsers.
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub(crate) enum RetiredStoreInput {
    /// `backend = "libsql"` in the config file's `[store]` section.
    Backend,
    /// `AION_STORE_BACKEND=libsql` in the environment.
    BackendEnvironment,
    /// `url = ...` in the config file's `[store]` section.
    Url,
    /// The `AION_STORE_URL` environment variable.
    Environment,
    /// The `--store-url` command-line flag.
    Flag,
}

/// Event-store backend configuration from `[store]`.
#[derive(Clone, Debug)]
pub struct StoreConfig {
    /// Selected backing store implementation.
    pub backend: StoreBackend,
    pub(crate) retired_input: Option<RetiredStoreInput>,
    /// Static distribution-shard assignment for this node (multi-shard
    /// active-active). When empty (the default) the node owns ALL shards — the
    /// single-node default, byte-identical to today. When set, the engine boot
    /// path scopes recovery and enumeration to exactly these shards. The
    /// single-shard memory backend ignores the assignment; it is meaningful only
    /// for a sharded backend. No election is performed: assignment is static
    /// config.
    pub owned_shards: Vec<usize>,
    /// Filesystem data directory for the haematite backend. Required when
    /// `backend = haematite`; ignored by every other backend. The directory is
    /// opened if it already holds a haematite database, otherwise created.
    pub data_dir: Option<String>,
    /// Number of haematite shards to create on a fresh database. Defaults to 64.
    ///
    /// This is an IMMUTABLE virtual-shard count: nodes own shard *ranges* and
    /// routing is `BLAKE3(key) % shard_count` with no reshard path, so a
    /// single-node deployment can later grow into a cluster WITHOUT a data
    /// migration — but only up to `shard_count` nodes, and the value is fixed
    /// at create.
    ///
    /// The default was briefly 4096 on the premise that lazy shard-actor
    /// materialization (haematite >= 0.4.0) made a high count ~free. That
    /// premise fails in practice (#187): aion-server's boot restores
    /// packages/routes/namespaces via full-prefix scans, which materialize
    /// EVERY shard, and haematite 0.4.0 then fans each commit out to every
    /// materialized shard with an unconditional fsync — ~2 fsyncs x 4096 per
    /// logical commit — blowing the 5s shard-actor timeout and bricking
    /// deploy/start/timers/outbox on a fresh server. Re-raise only after
    /// haematite makes commit O(dirty shards) and the scaffold e2es pass at
    /// the new default. Set explicitly (config or `AION_STORE_SHARD_COUNT`)
    /// to override. Ignored by every other backend, and ignored when opening
    /// an existing haematite database (the on-disk shard count wins).
    pub shard_count: usize,
    /// Optional distributed-cluster membership for the haematite backend (SS-2).
    ///
    /// Absent (the default) selects the SINGLE-NODE haematite path, byte-identical
    /// to today: no endpoint is bound, no shard is elected, the store owns
    /// everything locally. Present selects the DISTRIBUTED path: the boot path
    /// binds a replication endpoint, builds a quorum membership from `members` +
    /// `peers`, and the engine boot path elects (`acquire_shard_and_serve`) this
    /// node's `owned_shards` before recovery. Ignored by every non-haematite
    /// backend.
    pub cluster: Option<ClusterConfig>,
    /// The byte ceiling the haematite node caches may hold.
    /// **Required; no default** — the haematite boot path refuses to start
    /// without it. Ignored by every other backend (only haematite has a node
    /// cache), which is why it is required where it is USED rather than at
    /// parse time: a `memory` deployment is not asked to rule on a cache it does
    /// not have.
    ///
    /// A node is not a fixed-size thing — under byte-aware chunking a leaf
    /// reaches the ~96KB class — so a cache bounded only by `max_entries`
    /// carries a standing footprint of `entries x (whatever a node weighs)`,
    /// unbounded in bytes by construction. This is the missing bound, and
    /// haematite 0.8.2 refuses a `DatabaseConfig` that does not carry it.
    ///
    /// Deserialized through haematite's own wire shape, so there is exactly one
    /// parser for this value estate-wide:
    ///
    /// ```toml
    /// [store]
    /// node_cache_budget = { bytes = 1073741824 }  # 1 GiB
    /// # or, the pre-budget behaviour said out loud:
    /// node_cache_budget = "unlimited"
    /// ```
    ///
    /// Absent fails startup with [`STORE_NODE_CACHE_BUDGET_REQUIRED`]; a zero
    /// byte count is refused by haematite at parse time (a zero ceiling admits
    /// nothing, which is "disable the cache" and must be spelled differently).
    ///
    /// [`STORE_NODE_CACHE_BUDGET_REQUIRED`]: super::STORE_NODE_CACHE_BUDGET_REQUIRED
    pub node_cache_budget: Option<haematite::NodeCacheBudget>,
}

#[derive(Default, Deserialize)]
#[serde(default, deny_unknown_fields)]
struct StoreConfigWire {
    backend: Option<String>,
    /// Retired: recorded so `validate` can refuse it. Kept on the wire — and
    /// NOT left to `deny_unknown_fields` — so an operator carrying a 0.15
    /// config is told the key is retired and what replaces it, rather than
    /// getting an unknown-field parse error that names no remedy.
    url: Option<String>,
    owned_shards: Vec<usize>,
    data_dir: Option<String>,
    shard_count: Option<usize>,
    cluster: Option<ClusterConfig>,
    node_cache_budget: Option<haematite::NodeCacheBudget>,
}

impl<'de> Deserialize<'de> for StoreConfig {
    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
    where
        D: Deserializer<'de>,
    {
        let wire = StoreConfigWire::deserialize(deserializer)?;
        let mut config = Self::default();
        if let Some(backend) = wire.backend.as_deref() {
            match backend.to_ascii_lowercase().as_str() {
                "memory" => config.backend = StoreBackend::Memory,
                "haematite" => config.backend = StoreBackend::Haematite,
                "libsql" => config.retired_input = Some(RetiredStoreInput::Backend),
                _ => {
                    return Err(de::Error::custom(
                        "store.backend must be one of: memory, haematite",
                    ));
                }
            }
        }
        if wire.url.is_some() && config.retired_input.is_none() {
            config.retired_input = Some(RetiredStoreInput::Url);
        }
        config.owned_shards = wire.owned_shards;
        config.data_dir = wire.data_dir;
        if let Some(shard_count) = wire.shard_count {
            config.shard_count = shard_count;
        }
        config.cluster = wire.cluster;
        config.node_cache_budget = wire.node_cache_budget;
        Ok(config)
    }
}

/// Distributed-cluster membership for the haematite backend, from `[store.cluster]`.
///
/// This is the minimal, well-defaulted seam that turns the single-node haematite
/// store into a distributed one (SS-2). A "cluster of one" — `node_id` set,
/// `members` either empty or naming only `node_id`, and no `peers` — is a valid,
/// non-flaky configuration: election self-quorums (quorum denominator 1) and the
/// node boots through the production builder as the fenced owner of its shards.
#[derive(Clone, Debug, Deserialize, PartialEq, Eq)]
#[serde(deny_unknown_fields)]
pub struct ClusterConfig {
    /// This node's globally-unique distribution name (e.g. `node-0@127.0.0.1`).
    /// Used as the local endpoint name and the local membership identity.
    pub node_id: String,
    /// The replication endpoint listen address this node binds for peer
    /// quorum/election traffic (e.g. `127.0.0.1:7000`).
    pub bind_address: SocketAddr,
    /// The FULL cluster membership by node id — the quorum DENOMINATOR. Never the
    /// reachable subset. May be empty or omit peers for a cluster of one, in which
    /// case it is treated as `[node_id]` (denominator 1). `node_id` is always
    /// counted in the denominator whether or not it appears here.
    #[serde(default)]
    pub members: Vec<String>,
    /// Dialable peers (name + address) this node connects to for replication. A
    /// cluster of one leaves this empty. Peers not in `members` do not inflate the
    /// quorum denominator.
    #[serde(default)]
    pub peers: Vec<ClusterPeer>,
    /// SS-5b automatic-failover poll interval in milliseconds: how often the
    /// cluster supervisor checks each watched peer's replication liveness.
    /// Defaults to [`DEFAULT_FAILOVER_POLL_INTERVAL_MS`] when omitted.
    ///
    /// [`DEFAULT_FAILOVER_POLL_INTERVAL_MS`]: super::DEFAULT_FAILOVER_POLL_INTERVAL_MS
    #[serde(default)]
    pub failover_poll_interval_ms: Option<u64>,
    /// SS-5b debounce: the number of CONSECUTIVE polls a peer must be observed
    /// disconnected before its shards are adopted, so a transient blip does not
    /// trigger a disruptive failover. Defaults to
    /// [`DEFAULT_FAILOVER_CONFIRMATIONS`] when omitted; must be at least one.
    ///
    /// [`DEFAULT_FAILOVER_CONFIRMATIONS`]: super::DEFAULT_FAILOVER_CONFIRMATIONS
    #[serde(default)]
    pub failover_confirmations: Option<u32>,
}

/// One dialable cluster peer: its distribution name and replication address.
#[derive(Clone, Debug, Deserialize, PartialEq, Eq)]
#[serde(deny_unknown_fields)]
pub struct ClusterPeer {
    /// The peer's globally-unique distribution name (matches its `node_id`).
    pub name: String,
    /// The peer's replication endpoint address to dial.
    pub address: SocketAddr,
    /// The peer's gRPC client-API address, for request forwarding (R-2/R-3).
    /// This is DISTINCT from `address` (the replication/quorum endpoint): a
    /// forwarded client `signal`/`query`/`cancel` is dialed here, not on the
    /// replication port. Absent (the default) means the peer is not
    /// forwardable — its shards still resolve to a remote owner, but routing
    /// falls back to returning the typed `NotOwner` instead of forwarding (R-3).
    #[serde(default)]
    pub grpc_address: Option<SocketAddr>,
    /// The distribution shards this peer owns. Empty (the default) means the
    /// operator did not declare the peer's shards, so the SS-5b cluster
    /// supervisor cannot adopt them automatically when the peer dies — automatic
    /// failover for a peer requires its `owned_shards` to be declared here so the
    /// survivor knows exactly which shards to elect + resume. Declaring them does
    /// not change replication or quorum; it only tells the supervisor what to
    /// adopt on this peer's death.
    #[serde(default)]
    pub owned_shards: Vec<usize>,
}

/// Engine runtime settings from `[runtime]`.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct RuntimeSection {
    /// Number of scheduler worker threads.
    pub scheduler_threads: usize,
    /// Engine reply deadline for workflow queries, in milliseconds.
    /// REQUIRED — the server always mounts `/workflows/query`, so the query
    /// reply deadline must be an explicit operator decision; there is no
    /// default. The engine builder is equally explicit-no-default.
    pub query_timeout_ms: Option<u64>,
}

/// Graceful drain settings from `[drain]`.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct DrainConfig {
    /// Maximum drain duration in seconds.
    pub timeout_seconds: u64,
}

/// Authentication configuration applied at adapter boundaries.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct AuthConfig {
    /// Whether authentication is enabled.
    pub enabled: bool,
    /// JWKS URL used by AO-006 auth validation.
    pub jwks_url: Option<String>,
    /// JWKS refresh interval in seconds.
    pub jwks_refresh_seconds: u64,
}

/// Metrics endpoint settings from `[metrics]`.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct MetricsConfig {
    /// Whether metrics are exposed.
    pub enabled: bool,
}

/// Namespace defaults from `[namespaces]`.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct NamespacesConfig {
    /// Default namespace used for local callers and worker dispatch.
    pub default: String,
    /// Minted-on-use policy: whether referencing an unseen namespace at worker
    /// registration durably mints it ([`AutoCreate::Open`], the zero-config
    /// default) or is rejected ([`AutoCreate::Closed`]).
    pub auto_create: AutoCreate,
    /// Platform-wide default for a namespace's **cluster-wide** concurrent
    /// in-flight-activity ceiling, applied when a namespace record carries no
    /// explicit `max_in_flight_activities` override (Control-Plane Phase 2,
    /// P2-Q1). This is the GENEROUS platform default, NOT a low hard cap: it is
    /// a cluster-wide tenant contract (never "per-node × N"), so a tenant is
    /// promised ≈this many concurrent activities across the whole cluster.
    /// Defaults to [`DEFAULT_MAX_IN_FLIGHT_ACTIVITIES`].
    ///
    /// **Stored-only in this slice.** Nothing enforces it yet — the outbox
    /// dispatcher's keyed backpressure (P2-Q2) consults it in a later slice.
    ///
    /// [`DEFAULT_MAX_IN_FLIGHT_ACTIVITIES`]: super::DEFAULT_MAX_IN_FLIGHT_ACTIVITIES
    pub max_in_flight_activities: u32,
}

/// Minted-on-use namespace policy (Control-Plane Phase 1).
///
/// Governs what happens when a worker registers for a namespace that has no
/// durable registry record. Defaults to [`AutoCreate::Open`] to preserve the
/// zero-ceremony, no-pre-provision model: a namespace comes into being on first
/// reference.
#[derive(Clone, Copy, Debug, Default, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum AutoCreate {
    /// A worker registering for an unseen namespace durably mints it (an
    /// idempotent upsert through the registry). Zero-config default.
    #[default]
    Open,
    /// A worker registering for a namespace with no durable registry record is
    /// rejected; the namespace is never created at the registration hook. The
    /// escape hatch for locked-down deployments is an explicit create
    /// (`POST /namespaces`).
    Closed,
}

/// Public transport listener addresses retained for existing adapter code.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct ListenConfig {
    /// gRPC API and worker-protocol listener.
    pub grpc: SocketAddr,
    /// HTTP/JSON and dashboard listener.
    pub http: SocketAddr,
}

/// TLS certificate and private-key material.
#[derive(Clone, Debug, Deserialize)]
#[serde(deny_unknown_fields)]
pub struct TlsConfig {
    /// Certificate chain path supplied by the operator.
    pub certificate_chain_path: PathBuf,
    /// Private-key path supplied by the operator.
    pub private_key_path: PathBuf,
}

/// Static ops-console asset configuration.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct OpsConsoleConfig {
    /// Operator-selected bundle source.
    pub source: OpsConsoleAssetSource,
}

/// Static ops-console bundle source.
#[derive(Clone, Debug, Deserialize)]
pub enum OpsConsoleAssetSource {
    /// Serve the built bundle from an operator-supplied directory.
    FileSystem {
        /// Directory containing `index.html` and built asset files.
        asset_path: PathBuf,
    },
    /// Serve the compile-time embedded bundle.
    Embedded,
}

/// Namespace resolver construction mode.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct NamespaceConfig {
    /// Deployment-selected namespace mapping mode.
    pub mode: NamespaceMode,
}

/// Supported namespace mapping modes.
#[derive(Clone, Debug, Deserialize)]
pub enum NamespaceMode {
    /// All authorized namespaces share the configured engine instance.
    SharedEngine,
    /// Namespace authorization is disabled only for single-tenant deployments.
    SingleTenant {
        /// The only namespace accepted by the deployment.
        namespace: String,
    },
}

/// Remote worker heartbeat configuration.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct WorkerConfig {
    /// Window after which a silent worker is considered lost.
    #[serde(with = "duration_millis")]
    pub heartbeat_window: Duration,
    /// R1 queue-service policies and clocks (`[worker.queue_service]`).
    ///
    /// Absent, every queue is served `strict` with no clocks: structural
    /// unservability refuses, everything else waits — loudly, typed, and
    /// queryable. Writing a deadline is how an operator declares how long a
    /// dispatch may wait; writing a `durable_pending` override is how a queue
    /// declares that its runs park rather than refuse.
    #[serde(default)]
    pub queue_service: crate::worker::queue_service::QueueServiceConfig,
}

/// WebSocket stream configuration.
#[derive(Clone, Debug, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct WebSocketConfig {
    /// Per-connection outbound buffer bound.
    pub outbound_buffer_bound: usize,
    /// Capacity of the engine-global event broadcast channel that backs
    /// `/events/stream`. REQUIRED — the server always mounts the streaming
    /// endpoint, so streaming capacity must be an explicit operator decision;
    /// there is no default. Lag is filter-blind, so size this for global event
    /// volume across all namespaces, not per-subscription volume.
    pub event_broadcast_capacity: Option<usize>,
    /// Capacity of the deployment-global cluster topology/ownership broadcast
    /// channel that backs the WS3 `cluster` subscription on `/events/stream`.
    /// REQUIRED with the same non-zero startup guard as
    /// [`Self::event_broadcast_capacity`]: the cluster channel uses the same
    /// lag -> one-error-frame -> close contract, which has no defined buffer to
    /// lag against unless a capacity is configured. Cluster events are low-rate
    /// (peer/shard/worker topology deltas), so this is typically far smaller
    /// than the workflow event capacity.
    pub cluster_broadcast_capacity: Option<usize>,
}

impl WebSocketConfig {
    /// Validate the three unconditionally-mounted WebSocket seams: the
    /// per-connection buffer bound, the workflow event broadcast capacity, and
    /// the WS3 cluster broadcast capacity. The two broadcast capacities are
    /// explicit-no-default with a non-zero guard, since both back a lag ->
    /// one-error-frame -> close contract that needs a defined buffer to lag
    /// against.
    pub(super) fn validate(&self) -> Result<(), ServerError> {
        if self.outbound_buffer_bound == 0 {
            return config_error("websocket.outbound_buffer_bound must be greater than zero");
        }
        match self.event_broadcast_capacity {
            None | Some(0) => return config_error(EVENT_BROADCAST_CAPACITY_REQUIRED),
            Some(_) => {}
        }
        match self.cluster_broadcast_capacity {
            None | Some(0) => return config_error(CLUSTER_BROADCAST_CAPACITY_REQUIRED),
            Some(_) => {}
        }
        Ok(())
    }
}

/// Model Context Protocol surface settings from `[mcp]`.
///
/// The surface is DARK by default: with `enabled = false` (or the section
/// absent) the `/mcp` route is not mounted and the path is a plain 404. An MCP
/// endpoint hands a model the ability to start and cancel durable executions,
/// so commissioning it is an explicit operator decision, never a side effect of
/// upgrading.
///
/// `allowed_origins` is EMPTY by default and empty means "no browser origin".
/// A programmatic client sends no `Origin` header and is unaffected; a page in
/// a browser always sends one, so the empty list refuses every browser-borne
/// request. That is the DNS-rebinding defence the transport requires, and it is
/// a real policy rather than an unset value.
#[derive(Clone, Debug, Default, Deserialize, PartialEq, Eq)]
#[serde(default, deny_unknown_fields)]
pub struct McpConfig {
    /// Whether the `/mcp` endpoint is mounted. Defaults to false.
    pub enabled: bool,
    /// Browser origins permitted to reach the MCP endpoint. Empty permits none.
    pub allowed_origins: Vec<String>,
    /// `ttlMs` on the `server/discover` result.
    ///
    /// Defaults to [`DEFAULT_MCP_DISCOVER_TTL_MS`] when omitted and
    /// `enabled = true`, so switching the surface on does not force sizing two
    /// cache lifetimes first.
    ///
    /// [`DEFAULT_MCP_DISCOVER_TTL_MS`]: super::DEFAULT_MCP_DISCOVER_TTL_MS
    pub discover_ttl_ms: Option<u64>,
    /// `ttlMs` on the `tools/list` result. Defaults to
    /// [`DEFAULT_MCP_TOOLS_LIST_TTL_MS`].
    ///
    /// [`DEFAULT_MCP_TOOLS_LIST_TTL_MS`]: super::DEFAULT_MCP_TOOLS_LIST_TTL_MS
    pub tools_list_ttl_ms: Option<u64>,
    /// How long a created task lives before the server may fail and delete it.
    ///
    /// Zero means UNLIMITED and is written that way on the wire (`ttlMs: null`).
    /// An unlimited task is a real operator choice for a workflow that
    /// legitimately runs for months; it is not the default, because a task
    /// nobody polls holds an executor for as long as it lives. Defaults to
    /// [`DEFAULT_MCP_TASK_TTL_MS`].
    ///
    /// [`DEFAULT_MCP_TASK_TTL_MS`]: super::DEFAULT_MCP_TASK_TTL_MS
    pub task_ttl_ms: Option<u64>,
    /// The `pollIntervalMs` advertised on a created task. Defaults to
    /// [`DEFAULT_MCP_TASK_POLL_INTERVAL_MS`].
    ///
    /// [`DEFAULT_MCP_TASK_POLL_INTERVAL_MS`]: super::DEFAULT_MCP_TASK_POLL_INTERVAL_MS
    pub task_poll_interval_ms: Option<u64>,
    /// How often an awaited `start_run` re-reads the run's projected status.
    /// Defaults to [`DEFAULT_MCP_AWAIT_POLL_INTERVAL_MS`].
    ///
    /// [`DEFAULT_MCP_AWAIT_POLL_INTERVAL_MS`]: super::DEFAULT_MCP_AWAIT_POLL_INTERVAL_MS
    pub await_poll_interval_ms: Option<u64>,
}

impl McpConfig {
    /// Refuse a zero poll interval at startup.
    ///
    /// Only meaningful when the surface is enabled: a dark surface's knobs are
    /// never read, so refusing a boot over one would be refusing over a value
    /// nothing uses.
    ///
    /// # Errors
    ///
    /// [`ServerError::Config`] when an explicitly-set interval is zero. A zero
    /// poll interval is a busy loop, not a fast one, and the operator is told
    /// so rather than being handed one.
    pub(super) fn validate(&self) -> Result<(), ServerError> {
        if !self.enabled {
            return Ok(());
        }
        if self.task_poll_interval_ms == Some(0) {
            return config_error(MCP_TASK_POLL_INTERVAL_REQUIRED);
        }
        if self.await_poll_interval_ms == Some(0) {
            return config_error(MCP_AWAIT_POLL_INTERVAL_REQUIRED);
        }
        Ok(())
    }

    /// Resolve every omitted knob to its default.
    ///
    /// A zero poll interval resolves to the DEFAULT rather than to zero.
    /// [`Self::validate`] has already refused it on any loaded configuration,
    /// so this arm is reachable only from a hand-constructed config that
    /// skipped validation — and for that case a documented default is the safe
    /// direction, where a literal zero would be a busy loop.
    #[must_use]
    pub fn resolved(&self) -> ResolvedMcpConfig {
        ResolvedMcpConfig {
            enabled: self.enabled,
            allowed_origins: self.allowed_origins.clone(),
            discover_ttl_ms: self.discover_ttl_ms.unwrap_or(DEFAULT_MCP_DISCOVER_TTL_MS),
            tools_list_ttl_ms: self
                .tools_list_ttl_ms
                .unwrap_or(DEFAULT_MCP_TOOLS_LIST_TTL_MS),
            task_ttl_ms: match self.task_ttl_ms.unwrap_or(DEFAULT_MCP_TASK_TTL_MS) {
                0 => None,
                value => Some(value),
            },
            task_poll_interval_ms: self
                .task_poll_interval_ms
                .filter(|value| *value > 0)
                .unwrap_or(DEFAULT_MCP_TASK_POLL_INTERVAL_MS),
            await_poll_interval_ms: self
                .await_poll_interval_ms
                .filter(|value| *value > 0)
                .unwrap_or(DEFAULT_MCP_AWAIT_POLL_INTERVAL_MS),
        }
    }
}

impl Default for ResolvedMcpConfig {
    /// The dark surface: not mounted, no origin permitted, every lifetime at
    /// its documented default. This is what an embedder that never configures
    /// MCP gets, and it serves nothing.
    fn default() -> Self {
        McpConfig::default().resolved()
    }
}

/// The `[mcp]` section with every omitted knob resolved.
///
/// Held in [`RuntimeConfig`](super::RuntimeConfig) so the mount reads settled
/// values rather than re-deriving defaults at each composition point.
#[derive(Clone, Debug, PartialEq, Eq)]
pub struct ResolvedMcpConfig {
    /// Whether the `/mcp` endpoint is mounted.
    pub enabled: bool,
    /// Browser origins permitted to reach the endpoint.
    pub allowed_origins: Vec<String>,
    /// `ttlMs` on `server/discover`.
    pub discover_ttl_ms: u64,
    /// `ttlMs` on `tools/list`.
    pub tools_list_ttl_ms: u64,
    /// Task lifetime; `None` is unlimited.
    pub task_ttl_ms: Option<u64>,
    /// Advertised client poll interval for a task.
    pub task_poll_interval_ms: u64,
    /// Server-side poll interval for an awaited start.
    pub await_poll_interval_ms: u64,
}

/// Agent-observability transcript settings from `[observability]`: retention
/// bounds, and the transcript drain's flush policy.
///
/// The two RETENTION knobs default (they guard the durable `O` keyspace against
/// unbounded growth): `max_event_bytes` truncates one oversized transcript event
/// before it is persisted, and `max_stream_events` caps how many events one
/// `(workflow, activity, attempt)` stream retains (one marker record is
/// persisted at the cap; live streaming continues).
///
/// The two FLUSH knobs have **no default and are required** — see
/// [`Self::max_batch_events`].
#[derive(Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
#[serde(default, deny_unknown_fields)]
pub struct ObservabilityConfig {
    /// Ceiling on one persisted transcript event's serialized size, bytes.
    pub max_event_bytes: usize,
    /// Ceiling on retained events per `(workflow, activity, attempt)` stream.
    pub max_stream_events: u64,
    /// Maximum transcript events the drain commits in ONE durable append.
    /// **Required; no default** — the server refuses to start without it.
    ///
    /// Every durable commit re-persists its whole containing storage leaf as a
    /// new permanent blob, so committing one event at a time made the store cost
    /// of a run linear in event COUNT: measured 2026-08-17, a 136 KB status
    /// sweep left 73.7 MB of permanent store (540x) and a single 2,359-byte
    /// append bought a 2,303,416-byte blob. Batching divides that count by (up
    /// to) this number.
    ///
    /// There is deliberately no shipped value: the trade between store cost and
    /// how many events one refused commit can lose is the operator's, and an
    /// invented default would silently make it for them. Absent or zero fails
    /// startup with [`OBSERVABILITY_MAX_BATCH_EVENTS_REQUIRED`], the same
    /// explicit-no-default guard `websocket.cluster_broadcast_capacity` uses.
    ///
    /// [`OBSERVABILITY_MAX_BATCH_EVENTS_REQUIRED`]: super::OBSERVABILITY_MAX_BATCH_EVENTS_REQUIRED
    pub max_batch_events: Option<usize>,
    /// How long, in milliseconds, the drain may hold a PARTIAL batch open
    /// waiting for it to fill. **Required; no default** — but `0` is a valid,
    /// meaningful setting: never wait, commit whatever is already queued.
    ///
    /// This is the only knob here that touches durability timing. Transcript
    /// events already wait in an in-memory queue before they are committed, and
    /// everything queued there is lost if the process dies; this value bounds
    /// how much LONGER an event may wait, and therefore how much transcript a
    /// kill-9 can cost. It buys commits: at `0` only events that genuinely
    /// arrived together share a commit. The `O` keyspace is observability, never
    /// workflow replay authority, which is why the trade is offered at all.
    ///
    /// Absent fails startup with [`OBSERVABILITY_MAX_BATCH_HOLD_MS_REQUIRED`].
    ///
    /// [`OBSERVABILITY_MAX_BATCH_HOLD_MS_REQUIRED`]: super::OBSERVABILITY_MAX_BATCH_HOLD_MS_REQUIRED
    pub max_batch_hold_ms: Option<u64>,
}

impl ObservabilityConfig {
    /// The default retention bounds with the REQUIRED flush policy stated.
    ///
    /// For an embedder that builds its `ServerConfig` in code rather than from
    /// a file: [`Default`] deliberately leaves the flush policy unruled (so a
    /// server built from it refuses to start), and this is how such a caller
    /// states its ruling in one line. `max_batch_hold_ms` of `0` means the drain
    /// never holds a partial batch open.
    #[must_use]
    pub fn with_flush_policy(max_batch_events: usize, max_batch_hold_ms: u64) -> Self {
        Self {
            max_batch_events: Some(max_batch_events),
            max_batch_hold_ms: Some(max_batch_hold_ms),
            ..Self::default()
        }
    }

    /// Validate the retention bounds: both must be positive — a zero event
    /// ceiling truncates every event to nothing and a zero stream cap retains
    /// no transcript at all, so each is a genuine misconfiguration caught at
    /// startup with an operator-facing message (the `WebSocketConfig` pattern).
    pub(super) fn validate(&self) -> Result<(), ServerError> {
        if self.max_event_bytes == 0 {
            return config_error(OBSERVABILITY_MAX_EVENT_BYTES_REQUIRED);
        }
        if self.max_stream_events == 0 {
            return config_error(OBSERVABILITY_MAX_STREAM_EVENTS_REQUIRED);
        }
        // The FLUSH policy is deliberately not checked here. Both retention
        // bounds have documented defaults, so an omitted value is a resolved
        // value and this function can decide it at parse time; the flush policy
        // has no default at all, and a parsed `ServerConfig` is not yet a
        // running server (embedders and `aion config` inspect one without
        // booting). It is required where it is USED — at publisher
        // construction on the server boot path, `state.rs`'s
        // `required_transcript_batch_policy` — which is the same seam
        // `websocket.cluster_broadcast_capacity` is finally required at, and
        // which fails startup loudly, naming the missing key.
        Ok(())
    }
}

impl Default for ObservabilityConfig {
    /// The retention bounds fall back to their documented defaults; the flush
    /// policy does NOT — `None` here is "the operator has not ruled", which the
    /// boot path turns into a refusal to start, not into a guess.
    fn default() -> Self {
        Self {
            max_event_bytes: DEFAULT_OBSERVABILITY_MAX_EVENT_BYTES,
            max_stream_events: DEFAULT_OBSERVABILITY_MAX_STREAM_EVENTS,
            max_batch_events: None,
            max_batch_hold_ms: None,
        }
    }
}

/// Operator deploy API settings from `[deploy]`.
///
/// The deploy surface is dark by default: with `enabled = false` (or the
/// section absent) neither the `/deploy/*` HTTP routes nor the gRPC
/// `DeployService` are mounted, so a workflow server that is not a deploy
/// target exposes no deploy attack surface at all.
#[derive(Clone, Debug, Default, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct DeployConfig {
    /// Whether the deploy surface is mounted. Defaults to false.
    pub enabled: bool,
    /// Upload-size ceiling for `.aion` archives, in bytes. Defaults to
    /// [`DEFAULT_DEPLOY_MAX_ARCHIVE_BYTES`] when omitted and `enabled = true`,
    /// so turning deploy on does not force sizing a security ceiling; the
    /// operator overrides it for their packages.
    ///
    /// [`DEFAULT_DEPLOY_MAX_ARCHIVE_BYTES`]: super::DEFAULT_DEPLOY_MAX_ARCHIVE_BYTES
    pub max_archive_bytes: Option<u64>,
    /// Inflate ceiling for uploaded archive contents, in bytes: the total
    /// decompressed size of all archive entries an upload may extract to
    /// (DEFLATE bombs inflate ~1000:1 past `max_archive_bytes`). Defaults to
    /// [`DEFAULT_DEPLOY_MAX_INFLATED_BYTES`] when omitted and `enabled = true`;
    /// must be at least `max_archive_bytes`.
    ///
    /// [`DEFAULT_DEPLOY_MAX_INFLATED_BYTES`]: super::DEFAULT_DEPLOY_MAX_INFLATED_BYTES
    pub max_inflated_bytes: Option<u64>,
}

/// Local dev-server surface settings from `[dev]`.
///
/// The dev surface is dark by default, gated on `enabled`: with it false (the
/// section absent or `enabled = false`) the `/dev/*` routes are not mounted,
/// the engine installs the bare production activity dispatcher (no mocking
/// decorator), and nothing dev-specific is ever reachable. Setting `enabled =
/// true` mounts the dev endpoints and installs the per-run activity-mock
/// decorator — a development affordance, never on in production. It adds no
/// arbitrary defaults (ADR-001): the only knob is the on/off gate.
#[derive(Clone, Debug, Default, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct DevConfig {
    /// Whether the local dev-server surface is mounted. Defaults to false.
    pub enabled: bool,
}

/// Durable-outbox fan-out dispatcher settings from `[outbox]`.
///
/// The outbox dispatcher is dark by default, gated on `enabled`: with it false
/// (the section absent or `enabled = false`) the non-replayed background task
/// that claims pending outbox rows and dispatches them to connected workers is
/// never spawned, so default server behaviour is unchanged and the live
/// workflow dispatch path is the only dispatch path. Setting `enabled = true`
/// commissions the dispatcher; its operational knobs below — poll interval,
/// claim batch size, retry budget, and the backoff curve — are pure tuning, so
/// each resolves to a sane default when omitted rather than forcing the
/// operator to hand-author tuning values just to turn the feature on. An
/// explicitly set value (including a misconfigured `0`) is still validated.
///
/// Scope: this Phase-2 dispatcher dispatches claimed rows and marks each row's
/// terminal outbox state (done / retry / failed). Routing the worker completion
/// back into workflow history through the Recorder is Phase 3 and is not wired
/// here; with the flag off there is no behavioural difference at all.
#[derive(Clone, Debug, Default, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct OutboxConfig {
    /// Whether the outbox dispatcher background task is spawned. Defaults to
    /// false, leaving the dispatcher dark and server behaviour unchanged.
    pub enabled: bool,
    /// Interval between successive claim sweeps, in milliseconds. Defaults to
    /// [`DEFAULT_OUTBOX_POLL_INTERVAL_MS`] when omitted and `enabled = true`;
    /// override to size the poll cadence for fan-out volume and latency budget.
    ///
    /// [`DEFAULT_OUTBOX_POLL_INTERVAL_MS`]: super::DEFAULT_OUTBOX_POLL_INTERVAL_MS
    pub poll_interval_ms: Option<u64>,
    /// Maximum number of pending rows claimed per sweep. Defaults to
    /// [`DEFAULT_OUTBOX_BATCH_SIZE`] when omitted and `enabled = true`.
    ///
    /// [`DEFAULT_OUTBOX_BATCH_SIZE`]: super::DEFAULT_OUTBOX_BATCH_SIZE
    pub batch_size: Option<u32>,
    /// Dispatch attempts before a row is dead-lettered to `failed`. Defaults to
    /// [`DEFAULT_OUTBOX_MAX_ATTEMPTS`] when omitted and `enabled = true`. Must
    /// be at least one.
    ///
    /// [`DEFAULT_OUTBOX_MAX_ATTEMPTS`]: super::DEFAULT_OUTBOX_MAX_ATTEMPTS
    pub max_attempts: Option<u32>,
    /// Base retry backoff applied to the first retry, in milliseconds. Defaults
    /// to [`DEFAULT_OUTBOX_BACKOFF_BASE_MS`] when omitted and `enabled = true`.
    /// Successive retries multiply this by `backoff_multiplier` raised to the
    /// prior-attempt count, capped at `backoff_max_ms`.
    ///
    /// [`DEFAULT_OUTBOX_BACKOFF_BASE_MS`]: super::DEFAULT_OUTBOX_BACKOFF_BASE_MS
    pub backoff_base_ms: Option<u64>,
    /// Geometric growth factor applied to the backoff per prior attempt.
    /// Defaults to [`DEFAULT_OUTBOX_BACKOFF_MULTIPLIER`] when omitted and
    /// `enabled = true`. Must be at least one so backoff never shrinks.
    ///
    /// [`DEFAULT_OUTBOX_BACKOFF_MULTIPLIER`]: super::DEFAULT_OUTBOX_BACKOFF_MULTIPLIER
    pub backoff_multiplier: Option<u32>,
    /// Upper bound on a single retry's backoff, in milliseconds. Defaults to
    /// [`DEFAULT_OUTBOX_BACKOFF_MAX_MS`] when omitted and `enabled = true`. Must
    /// be at least `backoff_base_ms`.
    ///
    /// [`DEFAULT_OUTBOX_BACKOFF_MAX_MS`]: super::DEFAULT_OUTBOX_BACKOFF_MAX_MS
    pub backoff_max_ms: Option<u64>,
    /// Interval between live stale-claim reconciliation sweeps, in milliseconds. When both
    /// reconciliation knobs are absent the live sweep remains dark; setting either knob opts into
    /// reconciliation and requires both values to be positive.
    pub reconcile_interval_ms: Option<u64>,
    /// Age after which a durable `claimed` outbox row is considered stranded, in milliseconds. The
    /// reconciler re-arms only rows with `claimed_at` older than this threshold, preserving their
    /// attempt count.
    pub reconcile_stale_after_ms: Option<u64>,
    /// Wire transport the dispatcher uses to place a claimed row with a worker.
    /// Defaults to [`OutboxTransport::Liminal`] whenever the `liminal-transport`
    /// Cargo feature is compiled (the default ablative-stack build), so an
    /// outbox-enabled server uses the liminal cross-node transport out of the box;
    /// a slim build without that feature defaults to [`OutboxTransport::Grpc`].
    /// Selecting `liminal` in a build without the feature is a configuration error
    /// surfaced at spawn. The transport only matters when `outbox.enabled = true`.
    pub transport: OutboxTransport,
    /// Address (`host:port`) the aion-server LISTENS on for inbound liminal
    /// worker connections, used only when `transport = liminal`. REQUIRED in that
    /// mode; ignored otherwise.
    ///
    /// The aion-server HOSTS the liminal listener: a remote `LiminalActivityWorker`
    /// connects IN to this address and self-registers in-band, so the server's
    /// [`ConnectionSupervisor`](liminal_server::server::connection::ConnectionSupervisor)
    /// owns the worker's connection and can push a dispatch out on it
    /// (`push_to_connection`). This replaces the superseded 13-0 spike's
    /// client-connect address: the dispatcher no longer *connects out* to publish
    /// to a channel — it pushes to a connected worker the server already owns.
    ///
    /// The dispatch *channel* is not configured here: it is derived per-row from
    /// each row's durable `(namespace, task_queue)` via `dispatch_channel_name`
    /// (NSTQ-5), so one listener fans different worker pools out by selection.
    pub liminal_listen_address: Option<String>,
}

/// Wire transport selected for outbox dispatch.
///
/// `liminal` (the default whenever the `liminal-transport` feature is compiled —
/// which it is in the default ablative-stack build) routes the dispatch over the
/// liminal cross-node bus. `grpc` keeps the connected-worker registry path. A
/// slim build compiled WITHOUT `liminal-transport` falls back to `grpc` as the
/// default so the default is always constructible under the active feature set.
#[derive(Clone, Copy, Debug, Default, Eq, PartialEq, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum OutboxTransport {
    /// Dispatch over the in-process connected-worker gRPC registry. The default
    /// in a slim build compiled WITHOUT `liminal-transport`, which cannot
    /// construct the liminal path.
    // The default variant is feature-selected: gRPC when the liminal transport is
    // absent (it is the only constructible path), liminal otherwise.
    #[cfg_attr(not(feature = "liminal-transport"), default)]
    Grpc,
    /// Dispatch over the liminal cross-node bus (requires `liminal-transport`).
    /// The out-of-box default whenever `liminal-transport` is compiled (the
    /// default ablative-stack feature set), so an outbox-enabled server uses the
    /// ablative messaging transport without extra configuration.
    #[cfg_attr(feature = "liminal-transport", default)]
    Liminal,
}

/// Server-side authoring settings from `[authoring]`.
///
/// The AWL studio is on by default, rooted at `<AION_HOME>/authoring`, so a
/// stock `aion server` provides its document, layout, check, deploy, revision,
/// run, and scaffold surfaces without preliminary configuration. Operators can
/// override it explicitly.
///
/// Only the separate Gleam authoring loop remains dark by default: without
/// `gleam_path`, `/authoring/*` is not mounted and nothing invokes `gleam`
/// (CN7). Setting `gleam_path` commissions that loop and makes `project_root`
/// required.
#[derive(Clone, Debug, Default, Deserialize)]
#[serde(default, deny_unknown_fields)]
pub struct AuthoringConfig {
    /// Path to the external `gleam` binary the toolchain spawns. `None`
    /// (the default) leaves only the Gleam authoring loop dark; setting it gates
    /// the `/authoring/*` endpoints on. There is no default binary — the
    /// operator names it explicitly.
    pub gleam_path: Option<PathBuf>,
    /// Built Gleam workflow project root submitted source is written into and
    /// packaged from. REQUIRED when `gleam_path` is set; no default (house
    /// rule) — a Gleam project needs `gleam.toml`, the `aion_flow` dependency,
    /// `workflow.toml`, and `schemas/`, so the operator provisions and names
    /// the project root.
    pub project_root: Option<PathBuf>,
    /// Root directory exposed by the full AWL studio surface. The merged server
    /// loader defaults this to `<AION_HOME>/authoring` after config and
    /// environment overlays. `authoring.workspace_dir` or
    /// `AION_AUTHORING_WORKSPACE_DIR` therefore overrides it explicitly.
    pub workspace_dir: Option<PathBuf>,
}

impl Default for ServerSection {
    fn default() -> Self {
        Self {
            listen_address: DEFAULT_HTTP_ADDRESS,
            grpc_address: DEFAULT_GRPC_ADDRESS,
            cors_allowed_origins: Vec::new(),
        }
    }
}

impl Default for StoreConfig {
    fn default() -> Self {
        Self {
            // The ablative stack is the out-of-box durable default: an empty
            // config selects the haematite backend. The merged loader fills
            // `<AION_HOME>/data`; `memory` remains an explicit dev choice.
            backend: StoreBackend::Haematite,
            retired_input: None,
            owned_shards: Vec::new(),
            data_dir: None,
            // NOT a default: `None` is "the operator has not ruled", which the
            // haematite boot path turns into a refusal to start, not a guess.
            node_cache_budget: None,
            // 64, NOT 4096 (#187): raising the default to 4096 bricked every
            // fresh server — engine boot's scan_prefix materializes every
            // shard, then each haematite 0.4.0 commit fans out one thread +
            // fsync PER MATERIALIZED SHARD (~8k fsyncs/commit), blowing the
            // 5s shard-actor timeout on deploy/start/timers/outbox. Re-raise
            // only after haematite makes commit O(dirty shards) and the
            // scaffold e2es pass at the new default (see #187 fix plan).
            shard_count: 64,
            cluster: None,
        }
    }
}

impl Default for RuntimeSection {
    fn default() -> Self {
        Self {
            scheduler_threads: 1,
            // Deliberately absent: validation fails loudly until the operator
            // sets the workflow query reply deadline for the deployment.
            query_timeout_ms: None,
        }
    }
}

impl Default for DrainConfig {
    fn default() -> Self {
        Self {
            timeout_seconds: 30,
        }
    }
}

impl Default for AuthConfig {
    fn default() -> Self {
        Self {
            enabled: false,
            jwks_url: None,
            jwks_refresh_seconds: 300,
        }
    }
}

impl Default for MetricsConfig {
    fn default() -> Self {
        Self { enabled: true }
    }
}

impl Default for NamespacesConfig {
    fn default() -> Self {
        Self {
            default: "default".to_owned(),
            auto_create: AutoCreate::default(),
            max_in_flight_activities: DEFAULT_MAX_IN_FLIGHT_ACTIVITIES,
        }
    }
}

impl Default for ListenConfig {
    fn default() -> Self {
        Self {
            grpc: DEFAULT_GRPC_ADDRESS,
            http: DEFAULT_HTTP_ADDRESS,
        }
    }
}

impl Default for OpsConsoleConfig {
    fn default() -> Self {
        Self {
            source: OpsConsoleAssetSource::Embedded,
        }
    }
}

impl Default for NamespaceConfig {
    fn default() -> Self {
        Self {
            mode: NamespaceMode::SharedEngine,
        }
    }
}

impl Default for WorkerConfig {
    fn default() -> Self {
        Self {
            heartbeat_window: Duration::from_secs(30),
            queue_service: crate::worker::queue_service::QueueServiceConfig::default(),
        }
    }
}

impl Default for WebSocketConfig {
    fn default() -> Self {
        Self {
            outbound_buffer_bound: 32,
            // Deliberately absent: validation fails loudly until the operator
            // sizes the engine-global broadcast channel for the deployment.
            event_broadcast_capacity: None,
            // Deliberately absent for the same reason: the cluster channel has
            // no defined lag buffer until sized.
            cluster_broadcast_capacity: None,
        }
    }
}

mod duration_millis {
    use std::time::Duration;

    use serde::{Deserialize, Deserializer};

    pub(super) fn deserialize<'de, D>(deserializer: D) -> Result<Duration, D::Error>
    where
        D: Deserializer<'de>,
    {
        let millis = u64::deserialize(deserializer)?;
        Ok(Duration::from_millis(millis))
    }
}