feat(grid-agent): supervise grid sessions (#121)

config/grid-agent.example.json

@@ -25,5 +25,12 @@
   "storage_path": "data/grid-agent",
   "behavior": {
     "heartbeat_seconds": 30
+  },
+  "reconnect": {
+    "initial_delay_milliseconds": 1000,
+    "maximum_delay_seconds": 60,
+    "stable_reset_seconds": 120,
+    "jitter_basis_points": 2000,
+    "offline_work_capacity": 128
   }
 }

crates/metacrate-grid-agent/Cargo.toml

@@ -18,6 +18,9 @@ sha2 = "0.11"
 tokio = { version = "1.53.1", features = ["macros", "rt", "sync", "time"] }
 url = "2.5.8"
 
+[dev-dependencies]
+tokio = { version = "1.53.1", features = ["test-util"] }
+
 [target.'cfg(any(unix, windows))'.dependencies]
 tokio = { version = "1.53.1", features = ["io-util", "net", "rt-multi-thread", "signal"] }
 

crates/metacrate-grid-agent/README.md

@@ -6,9 +6,10 @@ OpenSim grid agent. It contains a reusable library and the
 offline: it publishes a deterministic ready event, accepts control commands,
 and shuts down both owned tasks without contacting a grid or LLM. The library
 also provides the bounded exact-endpoint LLM transport and tool loop for live
-adapters. The `live-grid` feature exposes the side-effect-free owner for the existing
-`libremetaverse::GridClient`; later live adapters must extend that manager graph
-instead of adding a protocol client.
+adapters. The `live-grid` feature exposes the owner for the existing
+`libremetaverse::GridClient`. Its supervised session adapter reuses that
+client's native `NetworkManager` for login, event-queue readiness, disconnect
+notifications, and logout instead of adding a protocol client.
 
 The LLM connection identity has exactly two resolved fields:
 `llm.endpoint_url` and `llm.api_key`. The endpoint is used exactly as supplied;
@@ -44,3 +45,6 @@ compatibility envelope, retry rules, and tool-loop safety contract. The central
 origin matrix, approval binding, budgets, prompt-data boundary, and opaque
 backend authorization are specified in
 [`../../docs/grid-agent-policy.md`](../../docs/grid-agent-policy.md).
+The reconnect state machine, generation fencing, offline-work contract, and
+shutdown deadline are specified in
+[`../../docs/grid-agent-session.md`](../../docs/grid-agent-session.md).

crates/metacrate-grid-agent/src/backend.rs

@@ -7,6 +7,10 @@ use std::error::Error;
 use std::fmt;
 use std::future::Future;
 use std::pin::Pin;
+#[cfg(feature = "live-grid")]
+use std::sync::Arc;
+#[cfg(feature = "live-grid")]
+use std::sync::atomic::{AtomicBool, Ordering};
 use tokio::sync::mpsc;
 
 pub type BackendFuture<'a, T> = Pin<Box<dyn Future<Output = T> + Send + 'a>>;
@@ -82,11 +86,45 @@ impl OfflineGridBackend {
 /// Live-feature composition owner that guarantees production adapters reuse
 /// the existing `libremetaverse` manager/client graph.
 #[cfg(feature = "live-grid")]
-#[derive(Debug)]
 pub struct LibremetaverseClientOwner {
     client: libremetaverse::GridClient,
 }
 
+#[cfg(feature = "live-grid")]
+impl fmt::Debug for LibremetaverseClientOwner {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter
+            .debug_struct("LibremetaverseClientOwner")
+            .field("client", &"[NATIVE CLIENT REDACTED]")
+            .finish()
+    }
+}
+
+/// Feature-gated session adapter that reuses the existing native
+/// `NetworkManager` login, event, disconnect, and logout lifecycle.
+#[cfg(feature = "live-grid")]
+#[derive(Clone)]
+pub struct LibremetaverseSessionBackend {
+    network: libremetaverse::NetworkManager,
+    login_url: String,
+    first_name: String,
+    last_name: String,
+    password: crate::config::SecretString,
+}
+
+#[cfg(feature = "live-grid")]
+impl fmt::Debug for LibremetaverseSessionBackend {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
+        formatter
+            .debug_struct("LibremetaverseSessionBackend")
+            .field("transport_connected", &self.network.native_connected())
+            .field("login_url", &"[REDACTED ENDPOINT]")
+            .field("avatar", &"[REDACTED IDENTITY]")
+            .field("password", &self.password)
+            .finish_non_exhaustive()
+    }
+}
+
 #[cfg(feature = "live-grid")]
 impl LibremetaverseClientOwner {
     /// Builds the shared client composition root without starting login or I/O.
@@ -107,6 +145,213 @@ impl LibremetaverseClientOwner {
     pub const fn client(&self) -> &libremetaverse::GridClient {
         &self.client
     }
+
+    /// Creates the supervised live-session adapter without logging in.
+    pub fn session_backend(
+        &self,
+        connection: crate::config::GridConnection,
+    ) -> Result<LibremetaverseSessionBackend, BackendError> {
+        let avatar = connection.avatar_name.trim();
+        let (first_name, last_name) = avatar
+            .split_once(char::is_whitespace)
+            .map_or((avatar, "Resident"), |(first, last)| (first, last.trim()));
+        if first_name.is_empty() || last_name.is_empty() {
+            return Err(BackendError::Configuration {
+                component: "grid avatar identity",
+            });
+        }
+        Ok(LibremetaverseSessionBackend {
+            network: self.client.network(),
+            login_url: connection.login_url.expose_url().to_owned(),
+            first_name: first_name.to_owned(),
+            last_name: last_name.to_owned(),
+            password: connection.password,
+        })
+    }
+}
+
+#[cfg(feature = "live-grid")]
+struct LibremetaverseSession {
+    generation: u64,
+    network: libremetaverse::NetworkManager,
+    signals: mpsc::Receiver<crate::session::SessionSignal>,
+    subscriptions: Vec<libremetaverse_types::compat::Subscription>,
+}
+
+#[cfg(feature = "live-grid")]
+impl crate::session::GridSession for LibremetaverseSession {
+    fn generation(&self) -> u64 {
+        self.generation
+    }
+
+    fn next_signal(
+        &mut self,
+        cancellation: CancellationToken,
+    ) -> crate::session::SessionFuture<'_, crate::session::SessionSignal> {
+        Box::pin(async move {
+            tokio::select! {
+                signal = self.signals.recv() => signal.unwrap_or_else(|| {
+                    crate::session::SessionSignal::Disconnected(
+                        crate::session::SessionFailure::new(
+                            crate::session::SessionFailureKind::TransientTransport,
+                        ),
+                    )
+                }),
+                () = cancellation.cancelled() => crate::session::SessionSignal::Disconnected(
+                    crate::session::SessionFailure::new(
+                        crate::session::SessionFailureKind::TransientTransport,
+                    ),
+                ),
+            }
+        })
+    }
+
+    fn logout(
+        mut self: Box<Self>,
+        cancellation: CancellationToken,
+    ) -> crate::session::SessionFuture<'static, Result<(), crate::session::SessionFailure>> {
+        Box::pin(async move {
+            self.subscriptions.clear();
+            self.network
+                .native_logout_async(Some(cancellation))
+                .await
+                .map_err(|_| {
+                    crate::session::SessionFailure::new(
+                        crate::session::SessionFailureKind::TransientTransport,
+                    )
+                })
+        })
+    }
+}
+
+#[cfg(feature = "live-grid")]
+impl crate::session::GridSessionBackend for LibremetaverseSessionBackend {
+    fn login(
+        &self,
+        generation: u64,
+        cancellation: CancellationToken,
+    ) -> crate::session::SessionFuture<
+        '_,
+        Result<Box<dyn crate::session::GridSession>, crate::session::SessionFailure>,
+    > {
+        Box::pin(async move {
+            let mut params = self
+                .network
+                .native_default_login_params(
+                    self.first_name.clone(),
+                    self.last_name.clone(),
+                    self.password.expose_secret().to_owned(),
+                    "MetaCrate".to_owned(),
+                    env!("CARGO_PKG_VERSION").to_owned(),
+                )
+                .map_err(|_| {
+                    crate::session::SessionFailure::new(
+                        crate::session::SessionFailureKind::InvalidConfiguration,
+                    )
+                })?;
+            params.uri.clone_from(&self.login_url);
+            params.start = "last".to_owned();
+            let logged_in = self
+                .network
+                .native_login(params, Some(cancellation))
+                .await
+                .map_err(|_| {
+                    crate::session::SessionFailure::new(
+                        crate::session::SessionFailureKind::TransientTransport,
+                    )
+                })?;
+            if !logged_in {
+                return Err(classify_native_login_failure(
+                    &self.network.native_login_error_key(),
+                ));
+            }
+
+            // Native login has already installed the current simulator and
+            // parsed inventory state. The existing GridClient managers retain
+            // world/inventory and movement-worker ownership; this generation
+            // adds only one disconnect and one readiness subscription.
+            let (sender, signals) = mpsc::channel(8);
+            let disconnected_sender = sender.clone();
+            let disconnected = self
+                .network
+                .native_subscribe_disconnected(Arc::new(move |event| {
+                    let kind = match event.reason() {
+                        libremetaverse::NetworkManagerDisconnectType::NetworkTimeout => {
+                            crate::session::SessionFailureKind::TransientTransport
+                        }
+                        libremetaverse::NetworkManagerDisconnectType::ServerInitiated => {
+                            crate::session::SessionFailureKind::Kicked
+                        }
+                        libremetaverse::NetworkManagerDisconnectType::SimShutdown => {
+                            crate::session::SessionFailureKind::Maintenance
+                        }
+                        libremetaverse::NetworkManagerDisconnectType::ClientInitiated => {
+                            crate::session::SessionFailureKind::TransientTransport
+                        }
+                    };
+                    let _ =
+                        disconnected_sender.try_send(crate::session::SessionSignal::Disconnected(
+                            crate::session::SessionFailure::new(kind),
+                        ));
+                }));
+            let ready_sender = sender.clone();
+            let ready_once = Arc::new(AtomicBool::new(false));
+            let callback_ready = Arc::clone(&ready_once);
+            let ready = self
+                .network
+                .native_subscribe_event_queue_running(Arc::new(move |_| {
+                    if !callback_ready.swap(true, Ordering::AcqRel) {
+                        let _ = ready_sender.try_send(crate::session::SessionSignal::Ready);
+                    }
+                }));
+            if !self.network.native_connected() {
+                let _ = sender.try_send(crate::session::SessionSignal::Disconnected(
+                    crate::session::SessionFailure::new(
+                        crate::session::SessionFailureKind::TransientTransport,
+                    ),
+                ));
+            } else if self
+                .network
+                .native_current_sim()
+                .and_then(|simulator| simulator.native_is_event_queue_running(None).ok())
+                .unwrap_or(false)
+            {
+                if !ready_once.swap(true, Ordering::AcqRel) {
+                    let _ = sender.try_send(crate::session::SessionSignal::Ready);
+                }
+            }
+            let session: Box<dyn crate::session::GridSession> = Box::new(LibremetaverseSession {
+                generation,
+                network: self.network.clone(),
+                signals,
+                subscriptions: vec![disconnected, ready],
+            });
+            Ok(session)
+        })
+    }
+
+    fn flush_audit(
+        &self,
+        _cancellation: CancellationToken,
+    ) -> crate::session::SessionFuture<'_, Result<(), crate::session::SessionFailure>> {
+        Box::pin(async { Ok(()) })
+    }
+}
+
+#[cfg(feature = "live-grid")]
+fn classify_native_login_failure(error_key: &str) -> crate::session::SessionFailure {
+    let normalized = error_key.trim().to_ascii_lowercase();
+    let kind = if matches!(
+        normalized.as_str(),
+        "key" | "password" | "credential" | "account" | "username" | "user"
+    ) {
+        crate::session::SessionFailureKind::InvalidCredentials
+    } else if normalized == "canceled" {
+        crate::session::SessionFailureKind::TransientTransport
+    } else {
+        crate::session::SessionFailureKind::ServerFailure
+    };
+    crate::session::SessionFailure::new(kind)
 }
 
 impl GridBackend for OfflineGridBackend {

crates/metacrate-grid-agent/src/config.rs

@@ -213,6 +213,7 @@ pub struct AgentConfig {
     pub limits: Limits,
     pub storage_path: PathBuf,
     pub behavior: BehaviorSettings,
+    pub reconnect: crate::session::ReconnectPolicy,
 }
 
 impl AgentConfig {
@@ -239,6 +240,9 @@ impl AgentConfig {
     /// Returns the first invalid field without performing I/O.
     pub fn validate(&self) -> Result<(), ConfigError> {
         validate_limits(&self.limits)?;
+        self.reconnect
+            .validate()
+            .map_err(|_| ConfigError::InvalidReconnect)?;
         if self.mode != OperatingMode::OfflineFake && self.grid.is_none() {
             return Err(ConfigError::Missing {
                 field: "grid",
@@ -411,6 +415,7 @@ pub enum ConfigError {
         minimum: usize,
         maximum: usize,
     },
+    InvalidReconnect,
 }
 
 impl fmt::Display for ConfigError {
@@ -460,6 +465,7 @@ impl fmt::Display for ConfigError {
                 formatter,
                 "unsafe {field}={value}; expected {minimum}..={maximum}"
             ),
+            Self::InvalidReconnect => formatter.write_str("invalid reconnect policy bounds"),
         }
     }
 }
@@ -479,6 +485,7 @@ struct FileConfig {
     limits: RawLimits,
     storage_path: Option<PathBuf>,
     behavior: RawBehavior,
+    reconnect: RawReconnect,
 }
 
 #[derive(Clone, Default, Deserialize)]
@@ -534,6 +541,16 @@ struct RawBehavior {
     heartbeat_seconds: Option<u64>,
 }
 
+#[derive(Clone, Default, Deserialize)]
+#[serde(default, deny_unknown_fields)]
+struct RawReconnect {
+    initial_delay_milliseconds: Option<u64>,
+    maximum_delay_seconds: Option<u64>,
+    stable_reset_seconds: Option<u64>,
+    jitter_basis_points: Option<u16>,
+    offline_work_capacity: Option<usize>,
+}
+
 fn read_config(path: &Path) -> Result<FileConfig, ConfigError> {
     let bytes = read_bounded_regular_file("configuration file", path, MAX_CONFIG_BYTES)?;
     serde_json::from_slice(&bytes).map_err(|error| ConfigError::InvalidSchema {
@@ -666,31 +683,51 @@ fn resolve<E: Environment>(
     };
     validate_limits(&limits)?;
 
+    let timeouts = Timeouts {
+        startup: checked_duration(
+            "timeouts.startup_seconds",
+            raw.timeouts.startup.unwrap_or(30),
+            1,
+            300,
+        )?,
+        shutdown: checked_duration(
+            "timeouts.shutdown_seconds",
+            raw.timeouts.shutdown.unwrap_or(10),
+            1,
+            60,
+        )?,
+        request: checked_duration(
+            "timeouts.request_seconds",
+            raw.timeouts.request.unwrap_or(60),
+            1,
+            300,
+        )?,
+    };
+    let reconnect_defaults = crate::session::ReconnectPolicy::default();
+    let reconnect = crate::session::ReconnectPolicy {
+        initial_delay: Duration::from_millis(
+            raw.reconnect.initial_delay_milliseconds.unwrap_or(1_000),
+        ),
+        maximum_delay: Duration::from_secs(raw.reconnect.maximum_delay_seconds.unwrap_or(60)),
+        stable_reset_after: Duration::from_secs(raw.reconnect.stable_reset_seconds.unwrap_or(120)),
+        shutdown_deadline: timeouts.shutdown,
+        jitter_basis_points: raw
+            .reconnect
+            .jitter_basis_points
+            .unwrap_or(reconnect_defaults.jitter_basis_points),
+        instance_seed: reconnect_defaults.instance_seed,
+        offline_work_capacity: raw
+            .reconnect
+            .offline_work_capacity
+            .unwrap_or(reconnect_defaults.offline_work_capacity),
+    };
+
     let config = AgentConfig {
         mode,
         llm,
         grid,
         authorized_avatar_uuids,
-        timeouts: Timeouts {
-            startup: checked_duration(
-                "timeouts.startup_seconds",
-                raw.timeouts.startup.unwrap_or(30),
-                1,
-                300,
-            )?,
-            shutdown: checked_duration(
-                "timeouts.shutdown_seconds",
-                raw.timeouts.shutdown.unwrap_or(10),
-                1,
-                60,
-            )?,
-            request: checked_duration(
-                "timeouts.request_seconds",
-                raw.timeouts.request.unwrap_or(60),
-                1,
-                300,
-            )?,
-        },
+        timeouts,
         limits,
         storage_path: raw
             .storage_path
@@ -703,6 +740,7 @@ fn resolve<E: Environment>(
                 300,
             )?,
         },
+        reconnect,
     };
     config.validate()?;
     Ok(config)
@@ -1076,6 +1114,19 @@ mod tests {
             Err(ConfigError::UnsafeLimit { .. })
         ));
         let _ = fs::remove_file(path);
+
+        let reconnect = temporary_file(
+            "unsafe-reconnect.json",
+            r#"{"reconnect":{"initial_delay_milliseconds":0}}"#,
+        );
+        assert!(matches!(
+            ConfigLoader::new()
+                .with_file(&reconnect)
+                .with_environment(offline_environment())
+                .load(),
+            Err(ConfigError::InvalidReconnect)
+        ));
+        let _ = fs::remove_file(reconnect);
     }
 
     #[test]

crates/metacrate-grid-agent/src/lib.rs

@@ -9,18 +9,21 @@ pub mod config;
 pub mod llm;
 pub mod policy;
 pub mod service;
+pub mod session;
 pub mod tool_loop;
 pub mod types;
 
 #[cfg(test)]
 mod policy_tests;
+#[cfg(test)]
+mod session_tests;
 
-#[cfg(feature = "live-grid")]
-pub use backend::LibremetaverseClientOwner;
 pub use backend::{
     AuthorizedToolBackend, BackendError, BackendFuture, GridBackend, OfflineGridBackend,
     WorldMutator,
 };
+#[cfg(feature = "live-grid")]
+pub use backend::{LibremetaverseClientOwner, LibremetaverseSessionBackend};
 pub use config::{
     AgentConfig, BehaviorSettings, ConfigError, ConfigLoader, EndpointUrl, Environment,
     GridConnection, Limits, LlmConnection, MapEnvironment, OperatingMode, SecretString,
@@ -39,6 +42,12 @@ pub use policy::{
     ResourceCost, ResourceEstimator, Risk, SchedulerGrantId, UntrustedData, UntrustedSource,
 };
 pub use service::{AgentService, ServiceError, ServiceHandle, ServiceState};
+pub use session::{
+    GridSession, GridSessionBackend, ReconnectPolicy, SessionControl, SessionFailure,
+    SessionFailureKind, SessionFuture, SessionObservation, SessionSignal, SessionState,
+    SessionStatus, SessionSupervisor, SessionSupervisorError, SessionSupervisorHandle, SessionWork,
+    WorkDisposition, WorkKind,
+};
 pub use tool_loop::{
     HistorySummarizer, SessionGeneration, ToolExecution, ToolExecutor, ToolFuture, ToolLoop,
     ToolLoopError, ToolLoopLimits, ToolLoopOutcome,

crates/metacrate-grid-agent/src/main.rs

@@ -4,6 +4,8 @@ use metacrate_grid_agent::{
 use std::error::Error;
 use std::fmt;
 use std::path::PathBuf;
+#[cfg(feature = "live-grid")]
+use std::sync::Arc;
 
 const MAX_ARGUMENTS: usize = 8;
 
@@ -84,9 +86,11 @@ async fn main() -> Result<(), Box<dyn Error>> {
         return Ok(());
     }
     if config.mode != OperatingMode::OfflineFake {
+        #[cfg(feature = "live-grid")]
+        return run_live(config, options.run_once).await;
+        #[cfg(not(feature = "live-grid"))]
         return Err(CliError(
-            "this architecture issue starts only the offline backend; live login is owned by a later milestone issue"
-                .into(),
+            "live grid mode requires rebuilding with --features live-grid".into(),
         )
         .into());
     }
@@ -133,3 +137,101 @@ async fn main() -> Result<(), Box<dyn Error>> {
     println!("grid agent stopped cleanly");
     Ok(())
 }
+
+#[cfg(feature = "live-grid")]
+async fn run_live(
+    config: metacrate_grid_agent::AgentConfig,
+    run_once: bool,
+) -> Result<(), Box<dyn Error>> {
+    use metacrate_grid_agent::{
+        GridSessionBackend, LibremetaverseClientOwner, SessionObservation, SessionState,
+        SessionSupervisor,
+    };
+
+    let connection = config.grid.clone().ok_or_else(|| {
+        CliError("validated live configuration did not contain a grid connection".into())
+    })?;
+    let owner = LibremetaverseClientOwner::new()?;
+    let backend = owner.session_backend(connection)?;
+    let erased: Arc<dyn GridSessionBackend> = Arc::new(backend);
+    let mut handle = SessionSupervisor::new(
+        erased,
+        config.reconnect,
+        config.limits.control_queue,
+        config.limits.observable_queue,
+    )?
+    .start();
+
+    if run_once {
+        let readiness = tokio::time::timeout(config.timeouts.startup, async {
+            loop {
+                match handle.next_observation().await {
+                    Some(SessionObservation::Transition { status, .. }) if status.agent_ready => {
+                        return Ok::<(), CliError>(());
+                    }
+                    Some(SessionObservation::Transition {
+                        status:
+                            metacrate_grid_agent::SessionStatus {
+                                state: SessionState::AuthenticationBlocked,
+                                ..
+                            },
+                        ..
+                    }) => {
+                        return Err(CliError(
+                            "grid authentication/configuration requires operator action".into(),
+                        ));
+                    }
+                    Some(_) => {}
+                    None => {
+                        return Err(CliError(
+                            "session supervisor stopped before readiness".into(),
+                        ));
+                    }
+                }
+            }
+        })
+        .await;
+        let readiness = match readiness {
+            Ok(result) => result,
+            Err(_) => Err(CliError("timed out waiting for full grid readiness".into())),
+        };
+        if let Err(error) = readiness {
+            handle.shutdown().await?;
+            return Err(error.into());
+        }
+        handle.shutdown().await?;
+        println!("grid agent completed one supervised login/logout cycle");
+        return Ok(());
+    }
+
+    println!("grid agent session supervisor started; press Ctrl-C to stop");
+    let mut signal_error = None;
+    loop {
+        tokio::select! {
+            signal = tokio::signal::ctrl_c() => {
+                if let Err(error) = signal {
+                    signal_error = Some(error);
+                }
+                break;
+            }
+            event = handle.next_observation() => {
+                let Some(event) = event else { break; };
+                if let SessionObservation::Transition { status, reason, retry_in } = event {
+                    println!(
+                        "grid session state={} generation={} transport_connected={} agent_ready={} reason={reason:?} retry_in={retry_in:?}",
+                        status.state.as_str(),
+                        status.generation,
+                        status.transport_connected,
+                        status.agent_ready,
+                    );
+                }
+            }
+        }
+    }
+    handle.shutdown().await?;
+    if let Some(error) = signal_error {
+        return Err(error.into());
+    }
+    println!("grid agent stopped cleanly");
+    Ok(())
+}

crates/metacrate-grid-agent/src/session_tests.rs

@@ -0,0 +1,526 @@
+use crate::session::*;
+use libremetaverse_types::compat::CancellationToken;
+use std::collections::VecDeque;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+#[derive(Clone)]
+enum LoginPlan {
+    Failure {
+        after: Duration,
+        failure: SessionFailure,
+    },
+    Success {
+        after: Duration,
+        signals: VecDeque<(Duration, SessionSignal)>,
+    },
+}
+
+#[derive(Default)]
+struct FakeStats {
+    logins: AtomicUsize,
+    logouts: AtomicUsize,
+    flushes: AtomicUsize,
+    active_sessions: AtomicUsize,
+    active_callbacks: AtomicUsize,
+    active_workers: AtomicUsize,
+    active_sockets: AtomicUsize,
+    max_sessions: AtomicUsize,
+    max_callbacks: AtomicUsize,
+    max_workers: AtomicUsize,
+    max_sockets: AtomicUsize,
+}
+
+impl FakeStats {
+    fn acquire(&self) {
+        update_max(&self.active_sessions, &self.max_sessions, 1);
+        update_max(&self.active_callbacks, &self.max_callbacks, 4);
+        update_max(&self.active_workers, &self.max_workers, 2);
+        update_max(&self.active_sockets, &self.max_sockets, 1);
+    }
+
+    fn release(&self) {
+        self.active_sessions.fetch_sub(1, Ordering::AcqRel);
+        self.active_callbacks.fetch_sub(4, Ordering::AcqRel);
+        self.active_workers.fetch_sub(2, Ordering::AcqRel);
+        self.active_sockets.fetch_sub(1, Ordering::AcqRel);
+    }
+}
+
+fn update_max(active: &AtomicUsize, maximum: &AtomicUsize, amount: usize) {
+    let next = active.fetch_add(amount, Ordering::AcqRel) + amount;
+    maximum.fetch_max(next, Ordering::AcqRel);
+}
+
+struct FakeBackend {
+    plans: Mutex<VecDeque<LoginPlan>>,
+    stats: Arc<FakeStats>,
+}
+
+impl FakeBackend {
+    fn new(plans: impl IntoIterator<Item = LoginPlan>) -> (Arc<Self>, Arc<FakeStats>) {
+        let stats = Arc::new(FakeStats::default());
+        (
+            Arc::new(Self {
+                plans: Mutex::new(plans.into_iter().collect()),
+                stats: Arc::clone(&stats),
+            }),
+            stats,
+        )
+    }
+}
+
+impl GridSessionBackend for FakeBackend {
+    fn login(
+        &self,
+        generation: u64,
+        cancellation: CancellationToken,
+    ) -> SessionFuture<'_, Result<Box<dyn GridSession>, SessionFailure>> {
+        self.stats.logins.fetch_add(1, Ordering::AcqRel);
+        let plan = self
+            .plans
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .pop_front()
+            .unwrap_or(LoginPlan::Success {
+                after: Duration::ZERO,
+                signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+            });
+        let stats = Arc::clone(&self.stats);
+        Box::pin(async move {
+            let after = match &plan {
+                LoginPlan::Failure { after, .. } | LoginPlan::Success { after, .. } => *after,
+            };
+            tokio::select! {
+                () = tokio::time::sleep(after) => {}
+                () = cancellation.cancelled() => {
+                    return Err(SessionFailure::new(SessionFailureKind::TransientTransport));
+                }
+            }
+            match plan {
+                LoginPlan::Failure { failure, .. } => Err(failure),
+                LoginPlan::Success { signals, .. } => {
+                    stats.acquire();
+                    let session: Box<dyn GridSession> = Box::new(FakeSession {
+                        generation,
+                        signals,
+                        stats,
+                        released: false,
+                    });
+                    Ok(session)
+                }
+            }
+        })
+    }
+
+    fn flush_audit(
+        &self,
+        _cancellation: CancellationToken,
+    ) -> SessionFuture<'_, Result<(), SessionFailure>> {
+        self.stats.flushes.fetch_add(1, Ordering::AcqRel);
+        Box::pin(async { Ok(()) })
+    }
+}
+
+struct FakeSession {
+    generation: u64,
+    signals: VecDeque<(Duration, SessionSignal)>,
+    stats: Arc<FakeStats>,
+    released: bool,
+}
+
+impl FakeSession {
+    fn release(&mut self) {
+        if !self.released {
+            self.released = true;
+            self.stats.release();
+        }
+    }
+}
+
+impl GridSession for FakeSession {
+    fn generation(&self) -> u64 {
+        self.generation
+    }
+
+    fn next_signal(&mut self, cancellation: CancellationToken) -> SessionFuture<'_, SessionSignal> {
+        let next = self.signals.pop_front();
+        Box::pin(async move {
+            if let Some((after, signal)) = next {
+                tokio::select! {
+                    () = tokio::time::sleep(after) => signal,
+                    () = cancellation.cancelled() => SessionSignal::Disconnected(
+                        SessionFailure::new(SessionFailureKind::TransientTransport)
+                    ),
+                }
+            } else {
+                cancellation.cancelled().await;
+                SessionSignal::Disconnected(SessionFailure::new(
+                    SessionFailureKind::TransientTransport,
+                ))
+            }
+        })
+    }
+
+    fn logout(
+        mut self: Box<Self>,
+        _cancellation: CancellationToken,
+    ) -> SessionFuture<'static, Result<(), SessionFailure>> {
+        self.stats.logouts.fetch_add(1, Ordering::AcqRel);
+        self.release();
+        Box::pin(async { Ok(()) })
+    }
+}
+
+impl Drop for FakeSession {
+    fn drop(&mut self) {
+        self.release();
+    }
+}
+
+fn ready_then(after: Duration, failure: SessionFailureKind) -> LoginPlan {
+    LoginPlan::Success {
+        after: Duration::ZERO,
+        signals: VecDeque::from([
+            (Duration::ZERO, SessionSignal::Ready),
+            (
+                after,
+                SessionSignal::Disconnected(SessionFailure::new(failure)),
+            ),
+        ]),
+    }
+}
+
+fn test_policy() -> ReconnectPolicy {
+    ReconnectPolicy {
+        initial_delay: Duration::from_secs(1),
+        maximum_delay: Duration::from_secs(8),
+        stable_reset_after: Duration::from_secs(10),
+        shutdown_deadline: Duration::from_secs(2),
+        jitter_basis_points: 0,
+        instance_seed: 7,
+        offline_work_capacity: 2,
+    }
+}
+
+#[test]
+fn retry_hints_are_minimums_and_all_delays_remain_capped() {
+    let mut policy = test_policy();
+    policy.jitter_basis_points = 5_000;
+    for seed in 0..128 {
+        policy.instance_seed = seed;
+        let hinted = Duration::from_secs(6);
+        let delay = policy.retry_delay(1, Some(hinted));
+        assert!(delay >= hinted);
+        assert!(delay <= policy.maximum_delay);
+        assert!(policy.retry_delay(64, None) <= policy.maximum_delay);
+    }
+}
+
+fn start(plans: impl IntoIterator<Item = LoginPlan>) -> (SessionSupervisorHandle, Arc<FakeStats>) {
+    let (backend, stats) = FakeBackend::new(plans);
+    let erased: Arc<dyn GridSessionBackend> = backend;
+    let handle = SessionSupervisor::new(erased, test_policy(), 32, 256)
+        .expect("supervisor")
+        .start();
+    (handle, stats)
+}
+
+async fn settle() {
+    for _ in 0..20 {
+        tokio::task::yield_now().await;
+    }
+}
+
+async fn wait_state(handle: &SessionSupervisorHandle, expected: SessionState) {
+    for _ in 0..100 {
+        if handle.state() == expected {
+            return;
+        }
+        tokio::task::yield_now().await;
+    }
+    panic!("expected {expected:?}, got {:?}", handle.status());
+}
+
+#[tokio::test(start_paused = true)]
+async fn paused_time_classifies_failures_backoff_flapping_and_stable_reset() {
+    let plans = [
+        LoginPlan::Failure {
+            after: Duration::ZERO,
+            failure: SessionFailure::new(SessionFailureKind::TransientTransport),
+        },
+        LoginPlan::Failure {
+            after: Duration::ZERO,
+            failure: SessionFailure::with_retry_after(
+                SessionFailureKind::Maintenance,
+                Duration::from_secs(5),
+            ),
+        },
+        ready_then(Duration::from_secs(1), SessionFailureKind::Kicked),
+        ready_then(
+            Duration::from_secs(11),
+            SessionFailureKind::SimulatorDisconnected,
+        ),
+        LoginPlan::Success {
+            after: Duration::ZERO,
+            signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+        },
+    ];
+    let (mut handle, stats) = start(plans);
+    wait_state(&handle, SessionState::Backoff).await;
+    assert_eq!(handle.status().consecutive_failures, 1);
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Backoff).await;
+    assert_eq!(handle.status().consecutive_failures, 2);
+    tokio::time::advance(Duration::from_secs(5)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    let first_online = handle.status().generation;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Backoff).await;
+    assert!(!handle.accepts_result(first_online));
+    tokio::time::advance(Duration::from_secs(4)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    tokio::time::advance(Duration::from_secs(11)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Backoff).await;
+    assert_eq!(handle.status().consecutive_failures, 1);
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    handle.shutdown().await.expect("shutdown");
+    assert_eq!(stats.flushes.load(Ordering::Acquire), 1);
+}
+
+#[tokio::test(start_paused = true)]
+async fn invalid_credentials_and_configuration_block_until_operator_reconnect() {
+    for failure in [
+        SessionFailureKind::InvalidCredentials,
+        SessionFailureKind::InvalidConfiguration,
+    ] {
+        let plans = [
+            LoginPlan::Failure {
+                after: Duration::ZERO,
+                failure: SessionFailure::new(failure),
+            },
+            LoginPlan::Success {
+                after: Duration::ZERO,
+                signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+            },
+        ];
+        let (mut handle, stats) = start(plans);
+        wait_state(&handle, SessionState::AuthenticationBlocked).await;
+        tokio::time::advance(Duration::from_hours(1)).await;
+        settle().await;
+        assert_eq!(stats.logins.load(Ordering::Acquire), 1);
+        handle
+            .control(SessionControl::ForceReconnect)
+            .await
+            .expect("force reconnect");
+        settle().await;
+        wait_state(&handle, SessionState::Online).await;
+        handle.shutdown().await.expect("shutdown");
+    }
+}
+
+#[tokio::test(start_paused = true)]
+async fn manual_pause_logout_reconnect_and_offline_work_are_predictable() {
+    let (mut handle, stats) = start([LoginPlan::Success {
+        after: Duration::ZERO,
+        signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+    }]);
+    wait_state(&handle, SessionState::Online).await;
+    let generation = handle.status().generation;
+    let token = handle
+        .generation_token(generation)
+        .expect("generation token");
+    handle.control(SessionControl::Pause).await.expect("pause");
+    settle().await;
+    wait_state(&handle, SessionState::Paused).await;
+    assert!(token.is_cancellation_requested());
+    assert!(!handle.accepts_result(generation));
+    assert_eq!(
+        handle
+            .submit(SessionWork::new("mutate", WorkKind::IdempotentMutation).expect("work"))
+            .await
+            .expect("decision"),
+        WorkDisposition::RejectedOfflineMutation
+    );
+    assert_eq!(
+        handle
+            .submit(SessionWork::new("read", WorkKind::ReadOnly).expect("work"))
+            .await
+            .expect("decision"),
+        WorkDisposition::Queued
+    );
+    handle
+        .control(SessionControl::Resume)
+        .await
+        .expect("resume");
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    assert_ne!(handle.status().generation, generation);
+    assert_eq!(
+        handle
+            .submit(SessionWork::new("read", WorkKind::ReadOnly).expect("work"))
+            .await
+            .expect("dedupe decision"),
+        WorkDisposition::RejectedDuplicate
+    );
+    let before_force = handle.status().generation;
+    handle
+        .control(SessionControl::ForceReconnect)
+        .await
+        .expect("active reconnect");
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    assert_ne!(handle.status().generation, before_force);
+    handle
+        .control(SessionControl::Logout)
+        .await
+        .expect("logout");
+    settle().await;
+    wait_state(&handle, SessionState::Stopped).await;
+    handle
+        .control(SessionControl::ForceReconnect)
+        .await
+        .expect("reconnect");
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    handle.shutdown().await.expect("shutdown");
+    assert_eq!(stats.active_sessions.load(Ordering::Acquire), 0);
+}
+
+#[tokio::test(start_paused = true)]
+async fn transport_connection_is_distinct_from_full_agent_readiness() {
+    let (mut handle, _) = start([LoginPlan::Success {
+        after: Duration::ZERO,
+        signals: VecDeque::from([(Duration::from_secs(10), SessionSignal::Ready)]),
+    }]);
+    wait_state(&handle, SessionState::Degraded).await;
+    let degraded = handle.status();
+    assert!(degraded.transport_connected);
+    assert!(!degraded.agent_ready);
+    assert!(!handle.accepts_result(degraded.generation));
+    tokio::time::advance(Duration::from_secs(10)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    assert!(handle.status().agent_ready);
+    handle.shutdown().await.expect("shutdown");
+}
+
+#[tokio::test(start_paused = true)]
+async fn shutdown_covers_degraded_authentication_blocked_stopped_and_online_states() {
+    let cases = [
+        (
+            LoginPlan::Success {
+                after: Duration::ZERO,
+                signals: VecDeque::from([(Duration::from_hours(1), SessionSignal::Ready)]),
+            },
+            SessionState::Degraded,
+        ),
+        (
+            LoginPlan::Failure {
+                after: Duration::ZERO,
+                failure: SessionFailure::new(SessionFailureKind::InvalidCredentials),
+            },
+            SessionState::AuthenticationBlocked,
+        ),
+        (
+            LoginPlan::Success {
+                after: Duration::ZERO,
+                signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+            },
+            SessionState::Online,
+        ),
+    ];
+    for (plan, expected) in cases {
+        let (mut handle, stats) = start([plan]);
+        wait_state(&handle, expected).await;
+        if expected == SessionState::Online {
+            handle
+                .control(SessionControl::Logout)
+                .await
+                .expect("logout");
+            settle().await;
+            wait_state(&handle, SessionState::Stopped).await;
+        }
+        handle.shutdown().await.expect("shutdown");
+        assert_eq!(stats.active_sessions.load(Ordering::Acquire), 0);
+        assert_eq!(stats.flushes.load(Ordering::Acquire), 1);
+    }
+}
+
+#[tokio::test(start_paused = true)]
+async fn fake_grid_reconnects_keep_exactly_one_resource_set_and_no_stale_execution() {
+    let plans = [
+        ready_then(Duration::from_secs(1), SessionFailureKind::Maintenance),
+        ready_then(Duration::from_secs(1), SessionFailureKind::Kicked),
+        LoginPlan::Success {
+            after: Duration::ZERO,
+            signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+        },
+    ];
+    let (mut handle, stats) = start(plans);
+    wait_state(&handle, SessionState::Online).await;
+    let stale = handle.status().generation;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    tokio::time::advance(Duration::from_secs(2)).await;
+    settle().await;
+    wait_state(&handle, SessionState::Online).await;
+    assert!(!handle.accepts_result(stale));
+    assert!(handle.accepts_result(handle.status().generation));
+    assert_eq!(stats.max_sessions.load(Ordering::Acquire), 1);
+    assert_eq!(stats.max_callbacks.load(Ordering::Acquire), 4);
+    assert_eq!(stats.max_workers.load(Ordering::Acquire), 2);
+    assert_eq!(stats.max_sockets.load(Ordering::Acquire), 1);
+    handle.shutdown().await.expect("shutdown");
+    assert_eq!(stats.active_sessions.load(Ordering::Acquire), 0);
+    assert_eq!(stats.active_callbacks.load(Ordering::Acquire), 0);
+    assert_eq!(stats.active_workers.load(Ordering::Acquire), 0);
+    assert_eq!(stats.active_sockets.load(Ordering::Acquire), 0);
+    assert_eq!(stats.logouts.load(Ordering::Acquire), 3);
+    assert_eq!(stats.flushes.load(Ordering::Acquire), 1);
+}
+
+#[tokio::test(start_paused = true)]
+async fn shutdown_is_cancellation_driven_during_connect_backoff_paused_and_online() {
+    let scenarios = [
+        LoginPlan::Success {
+            after: Duration::from_hours(1),
+            signals: VecDeque::new(),
+        },
+        LoginPlan::Failure {
+            after: Duration::ZERO,
+            failure: SessionFailure::new(SessionFailureKind::TransientTransport),
+        },
+        LoginPlan::Success {
+            after: Duration::ZERO,
+            signals: VecDeque::from([(Duration::ZERO, SessionSignal::Ready)]),
+        },
+    ];
+    for (index, plan) in scenarios.into_iter().enumerate() {
+        let (mut handle, stats) = start([plan]);
+        settle().await;
+        if index == 2 {
+            wait_state(&handle, SessionState::Online).await;
+            handle.control(SessionControl::Pause).await.expect("pause");
+            settle().await;
+            wait_state(&handle, SessionState::Paused).await;
+        }
+        handle.shutdown().await.expect("bounded shutdown");
+        assert_eq!(handle.state(), SessionState::Stopped);
+        assert_eq!(stats.active_sessions.load(Ordering::Acquire), 0);
+        assert_eq!(stats.flushes.load(Ordering::Acquire), 1);
+    }
+}

crates/metacrate-grid-agent/tests/dependency_policy.rs

@@ -46,7 +46,7 @@ fn runtime_source_has_no_subprocess_or_native_abi_escape_hatch() {
     let mut files = Vec::with_capacity(8);
     collect_rust_files(&source, &mut files);
     assert!(
-        files.len() <= 10,
+        files.len() <= 12,
         "source-file count needs a reviewed bound update"
     );
     for path in files {
@@ -86,6 +86,25 @@ fn world_backend_requires_the_opaque_policy_authorization() {
     assert!(!backend.contains("decision: PolicyDecision"));
 }
 
+#[test]
+fn live_session_adapter_reuses_only_the_native_network_manager_lifecycle() {
+    let root = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
+    let backend = fs::read_to_string(root.join("src/backend.rs")).expect("backend source");
+    for required in [
+        ".native_default_login_params(",
+        ".native_login(",
+        ".native_subscribe_disconnected(",
+        ".native_subscribe_event_queue_running(",
+        ".native_logout_async(",
+    ] {
+        assert!(
+            backend.contains(required),
+            "missing native lifecycle {required}"
+        );
+    }
+    assert!(!backend.contains("reqwest"));
+}
+
 fn collect_rust_files(directory: &Path, output: &mut Vec<PathBuf>) {
     for entry in fs::read_dir(directory).expect("read source directory") {
         let path = entry.expect("source entry").path();

crates/metacrate-grid-agent/tests/session_supervisor.rs

@@ -0,0 +1,227 @@
+use libremetaverse_types::compat::CancellationToken;
+use metacrate_grid_agent::{
+    GridSession, GridSessionBackend, ReconnectPolicy, SessionFailure, SessionFailureKind,
+    SessionFuture, SessionSignal, SessionState, SessionSupervisor, SessionSupervisorHandle,
+    SessionWork, WorkDisposition, WorkKind,
+};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::time::Duration;
+
+#[derive(Default)]
+struct ResourceStats {
+    attempts: AtomicUsize,
+    logouts: AtomicUsize,
+    flushes: AtomicUsize,
+    sessions: AtomicUsize,
+    callbacks: AtomicUsize,
+    workers: AtomicUsize,
+    sockets: AtomicUsize,
+    max_sessions: AtomicUsize,
+    max_callbacks: AtomicUsize,
+    max_workers: AtomicUsize,
+    max_sockets: AtomicUsize,
+}
+
+impl ResourceStats {
+    fn acquire(&self) {
+        acquire(&self.sessions, &self.max_sessions, 1);
+        acquire(&self.callbacks, &self.max_callbacks, 3);
+        acquire(&self.workers, &self.max_workers, 2);
+        acquire(&self.sockets, &self.max_sockets, 1);
+    }
+
+    fn release(&self) {
+        self.sessions.fetch_sub(1, Ordering::AcqRel);
+        self.callbacks.fetch_sub(3, Ordering::AcqRel);
+        self.workers.fetch_sub(2, Ordering::AcqRel);
+        self.sockets.fetch_sub(1, Ordering::AcqRel);
+    }
+}
+
+fn acquire(active: &AtomicUsize, maximum: &AtomicUsize, amount: usize) {
+    let next = active.fetch_add(amount, Ordering::AcqRel) + amount;
+    maximum.fetch_max(next, Ordering::AcqRel);
+}
+
+struct ReconnectingFakeGrid {
+    stats: Arc<ResourceStats>,
+}
+
+impl GridSessionBackend for ReconnectingFakeGrid {
+    fn login(
+        &self,
+        generation: u64,
+        _cancellation: CancellationToken,
+    ) -> SessionFuture<'_, Result<Box<dyn GridSession>, SessionFailure>> {
+        let attempt = self.stats.attempts.fetch_add(1, Ordering::AcqRel) + 1;
+        self.stats.acquire();
+        let stats = Arc::clone(&self.stats);
+        Box::pin(async move {
+            let session: Box<dyn GridSession> = Box::new(FakeConnectedSession {
+                generation,
+                attempt,
+                step: 0,
+                stats,
+                released: false,
+            });
+            Ok(session)
+        })
+    }
+
+    fn flush_audit(
+        &self,
+        _cancellation: CancellationToken,
+    ) -> SessionFuture<'_, Result<(), SessionFailure>> {
+        self.stats.flushes.fetch_add(1, Ordering::AcqRel);
+        Box::pin(async { Ok(()) })
+    }
+}
+
+struct FakeConnectedSession {
+    generation: u64,
+    attempt: usize,
+    step: u8,
+    stats: Arc<ResourceStats>,
+    released: bool,
+}
+
+impl FakeConnectedSession {
+    fn release(&mut self) {
+        if !self.released {
+            self.released = true;
+            self.stats.release();
+        }
+    }
+}
+
+impl GridSession for FakeConnectedSession {
+    fn generation(&self) -> u64 {
+        self.generation
+    }
+
+    fn next_signal(&mut self, cancellation: CancellationToken) -> SessionFuture<'_, SessionSignal> {
+        let attempt = self.attempt;
+        let step = self.step;
+        self.step = self.step.saturating_add(1);
+        Box::pin(async move {
+            if step == 0 {
+                return SessionSignal::Ready;
+            }
+            if attempt <= 2 {
+                tokio::select! {
+                    () = tokio::time::sleep(Duration::from_secs(1)) => {
+                        let kind = if attempt == 1 {
+                            SessionFailureKind::Maintenance
+                        } else {
+                            SessionFailureKind::Kicked
+                        };
+                        SessionSignal::Disconnected(SessionFailure::new(kind))
+                    }
+                    () = cancellation.cancelled() => SessionSignal::Disconnected(
+                        SessionFailure::new(SessionFailureKind::TransientTransport),
+                    ),
+                }
+            } else {
+                cancellation.cancelled().await;
+                SessionSignal::Disconnected(SessionFailure::new(
+                    SessionFailureKind::TransientTransport,
+                ))
+            }
+        })
+    }
+
+    fn logout(
+        mut self: Box<Self>,
+        _cancellation: CancellationToken,
+    ) -> SessionFuture<'static, Result<(), SessionFailure>> {
+        self.stats.logouts.fetch_add(1, Ordering::AcqRel);
+        self.release();
+        Box::pin(async { Ok(()) })
+    }
+}
+
+impl Drop for FakeConnectedSession {
+    fn drop(&mut self) {
+        self.release();
+    }
+}
+
+async fn settle() {
+    for _ in 0..30 {
+        tokio::task::yield_now().await;
+    }
+}
+
+async fn wait_online(handle: &SessionSupervisorHandle) {
+    for _ in 0..100 {
+        if handle.state() == SessionState::Online {
+            return;
+        }
+        tokio::task::yield_now().await;
+    }
+    panic!("fake grid did not become ready: {:?}", handle.status());
+}
+
+#[tokio::test(start_paused = true)]
+async fn repeated_reconnects_have_one_resource_set_no_replay_and_no_shutdown_leaks() {
+    let stats = Arc::new(ResourceStats::default());
+    let backend: Arc<dyn GridSessionBackend> = Arc::new(ReconnectingFakeGrid {
+        stats: Arc::clone(&stats),
+    });
+    let policy = ReconnectPolicy {
+        initial_delay: Duration::from_secs(1),
+        maximum_delay: Duration::from_secs(4),
+        stable_reset_after: Duration::from_secs(30),
+        shutdown_deadline: Duration::from_secs(3),
+        jitter_basis_points: 0,
+        instance_seed: 1,
+        offline_work_capacity: 8,
+    };
+    let mut handle = SessionSupervisor::new(backend, policy, 32, 256)
+        .expect("supervisor")
+        .start();
+
+    wait_online(&handle).await;
+    let stale_generation = handle.status().generation;
+    assert_eq!(
+        handle
+            .submit(SessionWork::new("response-1", WorkKind::ReadOnly).expect("work"))
+            .await
+            .expect("accepted"),
+        WorkDisposition::Accepted {
+            generation: stale_generation
+        }
+    );
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    wait_online(&handle).await;
+    tokio::time::advance(Duration::from_secs(1)).await;
+    settle().await;
+    tokio::time::advance(Duration::from_secs(2)).await;
+    settle().await;
+    wait_online(&handle).await;
+
+    assert!(!handle.accepts_result(stale_generation));
+    assert_eq!(
+        handle
+            .submit(SessionWork::new("response-1", WorkKind::ReadOnly).expect("work"))
+            .await
+            .expect("dedupe"),
+        WorkDisposition::RejectedDuplicate
+    );
+    assert_eq!(stats.max_sessions.load(Ordering::Acquire), 1);
+    assert_eq!(stats.max_callbacks.load(Ordering::Acquire), 3);
+    assert_eq!(stats.max_workers.load(Ordering::Acquire), 2);
+    assert_eq!(stats.max_sockets.load(Ordering::Acquire), 1);
+
+    handle.shutdown().await.expect("shutdown");
+    assert_eq!(stats.sessions.load(Ordering::Acquire), 0);
+    assert_eq!(stats.callbacks.load(Ordering::Acquire), 0);
+    assert_eq!(stats.workers.load(Ordering::Acquire), 0);
+    assert_eq!(stats.sockets.load(Ordering::Acquire), 0);
+    assert_eq!(stats.logouts.load(Ordering::Acquire), 3);
+    assert_eq!(stats.flushes.load(Ordering::Acquire), 1);
+}

docs/grid-agent-architecture.md

@@ -10,18 +10,21 @@ platform-specific core path.
 
 ## Ownership and bounds
 
-`AgentService::start` is the sole task-creation point. It validates the complete
-configuration before allocating channels, calling a backend, or permitting
-network access. `ServiceHandle` then exclusively owns cancellation, both join
-handles, the control sender, and observable receiver.
+`AgentService::start` owns offline task creation, while
+`SessionSupervisor::start` owns the live lifecycle task. Configuration is
+validated before either allocates channels, calls a backend, or permits network
+access. Their handles exclusively own cancellation, join handles, controls, and
+observable receivers.
 
 | Resource | Owner | Hard/configured bound | Backpressure/termination |
 | --- | --- | --- | --- |
 | Grid-event queue | coordinator receives; backend sends | 8,192 / `grid_event_queue` | async send or cancellation |
 | Control queue | coordinator receives; handle sends | 256 / `control_queue` | async send; closed after stop |
 | Observable queue | handle receives; coordinator/backend send | 8,192 / `observable_queue` | async send or cancellation |
-| Backend task | `ServiceHandle.tasks[0]` | exactly one | shared cancellation token, joined first |
+| Backend task | `ServiceHandle.tasks[0]` | exactly one in offline mode | shared cancellation token, joined first |
 | Coordinator task | `ServiceHandle.tasks[1]` | exactly one | shared cancellation token, joined second |
+| Live session supervisor | `SessionSupervisorHandle` | one owner; one active generation/session | generation cancellation, exact-once logout, bounded join |
+| Offline session work | live supervisor | 1,024 hard / `reconnect.offline_work_capacity` | read-only queue; mutations rejected while not ready |
 | Body / message | typed boundary owners | 8 MiB / 64 KiB hard ceilings, with lower configured limits | rejected before enqueue |
 | Conversation / tool calls | request owner | 256 messages / 64 calls, with lower configured limits | rejected before request |
 | LLM request slots | shared `LlmClient` semaphore | 256 hard / configured concurrent requests | async acquire or cancellation |
@@ -36,7 +39,7 @@ type. Dynamic configuration can lower these absolute ceilings but cannot raise
 them. Backend error text is not forwarded; observations publish a fixed bounded
 diagnostic.
 
-## State machine and shutdown
+## State machines and shutdown
 
 The explicit service states are `starting -> running <-> paused -> stopping ->
 stopped`, with `failed` reserved for task failure. The offline backend publishes
@@ -54,6 +57,13 @@ feature-enabled live adapter drops its
 `LibremetaverseClientOwner` last, invoking the existing client ownership
 shutdown.
 
+Live modes use `stopped`, `connecting`, `degraded` (transport connected but not
+fully ready), `online`, `backoff`, `authentication-blocked`, `paused`, and
+`shutting-down`. Every attempt rotates a generation cancellation token.
+Disconnect, pause, logout, force reconnect, or shutdown invalidates it before
+cleanup, fencing old events and late LLM/tool results. See
+[`grid-agent-session.md`](grid-agent-session.md).
+
 ## Trust boundaries
 
 - JSON configuration and environment text are untrusted. Unknown fields,
@@ -83,7 +93,7 @@ shutdown.
 metacrate-grid-agent binary (portable Ctrl-C + config path)
                     |
                     v
-AgentService -> bounded Tokio channels/tasks -> injected GridBackend
+offline AgentService -> bounded Tokio channels/tasks -> injected GridBackend
      |                                           |
      v                                           v
 typed config/events/policy boundaries    live-grid feature boundary
@@ -104,3 +114,5 @@ The precise LLM compatibility and cancellation contract is documented in
 [`grid-agent-llm.md`](grid-agent-llm.md).
 The origin/capability matrix and opaque mutation boundary are documented in
 [`grid-agent-policy.md`](grid-agent-policy.md).
+The live lifecycle and generation contract is documented in
+[`grid-agent-session.md`](grid-agent-session.md).

docs/grid-agent-session.md

@@ -0,0 +1,75 @@
+# Grid-session supervision
+
+Live integrated and split modes run through `SessionSupervisor`. Its injected
+`GridSessionBackend` makes the lifecycle deterministic under a fake grid while
+the `live-grid` adapter reuses `libremetaverse::NetworkManager` for native
+login, event-queue readiness, disconnect callbacks, and logout.
+
+The native login does not yield a session until the login response has populated
+the existing inventory skeleton and a current simulator/UDP circuit exists.
+The adapter then waits for that simulator's event queue before publishing full
+readiness. Existing `GridClient` inventory/world managers and its single
+movement-update owner remain the composition owners; the adapter never creates
+duplicates. Only its generation-scoped readiness/disconnect subscriptions are
+recreated, and their RAII guards are dropped before logout.
+
+## States and reasons
+
+| State | Transport connected | Agent ready | Exit condition |
+| --- | --- | --- | --- |
+| `stopped` | no | no | startup, resume, or force reconnect |
+| `connecting` | no | no | login result, operator control, or cancellation |
+| `degraded` | yes | no | readiness, disconnect, or operator control |
+| `online` | yes | yes | readiness loss, disconnect, or operator control |
+| `backoff` | no | no | cancellation-driven timer or operator control |
+| `authentication-blocked` | no | no | resume/force reconnect or shutdown |
+| `paused` | no | no | resume/force reconnect, logout, or shutdown |
+| `shutting-down` | no | no | session logout, audit flush, and joined owner task |
+
+Every transition carries a stable `SessionReason`; raw login responses, server
+messages, credentials, capability URLs, and session tokens cannot enter the
+failure or observation types. `SessionStatus` reports transport connectivity
+and complete readiness separately.
+
+Invalid credentials and invalid local login configuration stop automatic
+retry. Transport failures, maintenance, kicks, simulator disconnects, and
+server failures retry with exponential backoff. `reconnect.maximum_delay_seconds`
+is a hard cap. A server retry hint is a minimum up to that cap. Per-instance
+jitter is bounded by `jitter_basis_points`, preventing synchronized reconnects,
+and a connection that remains up for `stable_reset_seconds` resets the failure
+streak.
+
+Backoff defaults are one second initially, 60 seconds maximum, 20 percent
+jitter, and a 120-second stable reset window. All waits use Tokio timers inside
+`select!` with cancellation/control; there are no blocking sleeps.
+
+## Generation and work safety
+
+Each attempt receives a monotonically changing generation and cancellation
+token. Session-scoped subscriptions, readiness ownership, and workers belong to
+the returned `GridSession` and are closed once by its consuming `logout` method.
+Late inference/tool results are accepted only when `accepts_result` sees their
+exact generation in fully ready state.
+
+While not ready, bounded read-only work may queue. Mutation work—including
+nominally idempotent mutation—is rejected, because the supervisor cannot prove
+that a previous policy authorization remains valid. Work IDs are deduplicated
+across reconnects with a bounded recent-ID set, so reconnect never silently
+duplicates an outbound response. Queued reads are released once into the new
+generation.
+
+## Shutdown
+
+Shutdown first rejects new work and invalidates the generation token, which
+cancels inference, tools, and scheduled consumers sharing it. It then consumes
+the active session for one logout attempt, waits within
+`timeouts.shutdown_seconds`, calls the backend's bounded audit-flush hook, and
+joins the supervisor owner task. A late task is aborted and awaited by the
+handle, so no task or socket is detached.
+
+Focused deterministic verification:
+
+```sh
+cargo test --locked -p metacrate-grid-agent --lib session_tests
+cargo clippy --locked -p metacrate-grid-agent --all-targets -- -D warnings
+```