onlyne_client/session/dispatch/

transport.rs

use super::*;

use super::env::missing_capability;
use super::idle::{resumable, suspend_locked};
use super::outbound::queue_outbound_locked;
use super::retire::{
    PendingClose, close_retired, keeps_idle, retire_idle_locked, stored_close_reason,
};
use super::state::{
    DispatchInner, DispatchState, FrameGuard, SessionSlot, has_attached_transport,
    rebase_generation, slot_key_named, slot_key_serving_task, slot_task,
};

/// The one sentence a session that ended a turn without a completion reads.
///
/// The wording is a contract (`docs/v2-CONTRACT.md` §3c): the plugin composes
/// none of it and keeps no copy, so this client owns the bytes and hands them
/// over verbatim through [`DispatchState::nudge_plugin`].
pub const NUDGE_TEXT: &str =
    "If this task is finished, report it with onlyne_complete; if something is missing, say what.";

/// Whether one session slot is the session an adapter mount named.
///
/// The mount carries the id the client spawned the plugin with
/// (`ONLYNE_SESSION_ID`), which is the slot's key and its stored reference.
/// Each task gets its own session, so those spellings all name one session; the
/// extra checks stay because a slot keeps the id its session was born with even
/// after its task binding changes.
pub(super) fn names_session(key: &str, slot: &SessionSlot, session_id: &str) -> bool {
    key == session_id
        || slot.session.task_id == session_id
        || slot.task_id.as_deref() == Some(session_id)
}

/// Whether a connection other than `io` is the one serving one slot.
fn attached_to_other(inner: &DispatchInner, key: &str, slot: &SessionSlot, io: &AdapterIo) -> bool {
    inner.transports.iter().any(|(session_id, (live, _))| {
        !live.same_connection(io) && names_session(key, slot, session_id)
    })
}

/// Whether `io` is the connection serving one session, as the binding rules left
/// it.
///
/// A connection this client holds read-only is never the answer: a mount that
/// finds its session already served lands in `revived` and never in
/// `transports`, which is the map this reads. Nothing else about the connection
/// is consulted — a frame carries its sender, so a connection serving one
/// session cannot answer for another by naming it.
pub(super) fn serves_session(inner: &DispatchInner, session_id: &str, io: &AdapterIo) -> bool {
    let Some(key) = slot_key_named(inner, session_id) else {
        return false;
    };
    let Some(slot) = inner.sessions.get(&key) else {
        return false;
    };
    slot_is_served_by(inner, &key, slot, io)
}

/// Whether one slot's session is served by `io`.
///
/// The body of [`serves_session`] and [`serves_task`]: the question is never
/// which name the frame carried, but whether this connection is the transport of
/// the slot that name resolves to.
fn slot_is_served_by(inner: &DispatchInner, key: &str, slot: &SessionSlot, io: &AdapterIo) -> bool {
    inner.transports.iter().any(|(served, (transport, _))| {
        transport.same_connection(io) && names_session(key, slot, served)
    }) || tools_bound_to(inner, key, io)
}

/// Whether `io` is the tools mount bound to one session.
///
/// A tools mount holds no process and is no transport, so the map the binding
/// rules fill says nothing about it: the handshake's own record is the whole
/// answer, and it is what lets a tools connection's `handoff` and `report`
/// frames travel the same authority door an agent's frames do
/// (`docs/v2-CONTRACT.md` §3b).
fn tools_bound_to(inner: &DispatchInner, key: &str, io: &AdapterIo) -> bool {
    inner
        .tools_mounts
        .iter()
        .any(|(held, bound)| held == key && bound.same_connection(io))
}

/// Whether `io` is the connection serving the session that answers for one task.
///
/// The task-spelling companion of [`serves_session`], and the authority every
/// state frame a plugin sends is measured against: a report names a task, and the
/// frame carries its sender, so a connection serving one session cannot end,
/// fault, or move another session by naming its task.
///
/// Unlike [`serves_session`] this walks every slot the task names rather than the
/// first one a map reach finds. Two slots answer to one task exactly when a
/// session came back for a task a newer session already took, and there the
/// first match is whichever one `HashMap` order happens to yield — an authority
/// check that reads that way refuses the same frame twice and applies it a third
/// time. The extra walk costs nothing on the ordinary path, where one slot
/// answers to the task, and the demoted half can never pass anyway: a read-only
/// slot is served by no transport at all.
pub(super) fn serves_task(inner: &DispatchInner, task_id: &str, io: &AdapterIo) -> bool {
    inner.sessions.iter().any(|(key, slot)| {
        names_session(key, slot, task_id) && slot_is_served_by(inner, key, slot, io)
    })
}

/// Whether `io` is any session's transport right now.
///
/// Weaker than [`serves_task`] — it says the connection serves *a* session of
/// this role, not the one a frame names — and it exists for the one window where
/// the stronger question cannot be asked: an operator's `recycle` or `cancel`
/// retires the slot before the completion it ordered arrives, and a retired slot
/// is out of `sessions` for `serves_task` to resolve. See
/// `ending_is_authorised` in `reports.rs`.
pub(super) fn is_bound_transport(inner: &DispatchInner, io: &AdapterIo) -> bool {
    inner
        .transports
        .values()
        .any(|(transport, _)| transport.same_connection(io))
}

/// Whether one connection is held read-only: it mounted a session this client
/// already serves through a different live connection.
pub(super) fn is_revived_connection(inner: &DispatchInner, io: &AdapterIo) -> bool {
    inner
        .revived
        .iter()
        .any(|(_, revived, _)| revived.same_connection(io))
}

/// Whether the connection this client holds read-only answers for one task.
///
/// The held half of the pair [`serves_task`] reads live. A demoted connection is
/// no session's transport, so the strict rule refuses its observation — and the
/// design still lets it answer for the task its own agent finished: `plugin_send`
/// routes what it sends onto the wire like any other session's, and
/// `retire_revived` retires the demoted slot with the completion that answers it.
/// A held connection is therefore the connection that may report one task ended,
/// and no other: the name it mounted with resolves to the slot the task answers
/// to.
fn held_for_task(inner: &DispatchInner, task_id: &str, io: &AdapterIo) -> bool {
    inner.revived.iter().any(|(name, revived, _)| {
        revived.same_connection(io)
            && slot_key_named(inner, name)
                .and_then(|key| inner.sessions.get(&key))
                .is_some_and(|slot| slot_task(slot) == task_id)
    })
}

/// Whether `io` may end or fault the session that answers for one task.
///
/// [`serves_task`] is the whole rule for a frame that moves state: a beat is an
/// observation of a session, and only its transport has one. An ending is a
/// different fact — the agent behind either connection ran the work, and the
/// client may have moved the binding while it was running — so the two windows
/// where the strict rule cannot see the sender are read here instead:
///
/// * the connection is held read-only for this very task ([`held_for_task`]), and
/// * this client ordered the task to its ending with `control`, which retired the
///   slot before the answer it asked for arrived. The note names the task, and a
///   transport still bound somewhere in this role names the sender; a forged
///   ending has neither, and a foreign connection answering for a task someone
///   else's command opened has the note without the binding.
pub(super) fn serves_ending(inner: &DispatchInner, task_id: &str, io: &AdapterIo) -> bool {
    serves_task(inner, task_id, io)
        || held_for_task(inner, task_id, io)
        || (inner
            .control_settles
            .iter()
            .any(|note| note.task_id == task_id)
            && is_bound_transport(inner, io))
}

/// Whether one slot is held by a connection this client keeps read-only.
///
/// A slot demoted by [`note_binding_locked`] — its task taken by a newer session
/// — is served by no transport at all: the connection that came back for it
/// waits in `revived` under the name it mounted with, which is the name that
/// names this slot. While that connection is there the slot has an owner:
/// `retire_revived` retires it with the completion that answers what the held
/// connection wrote. A demoted slot whose held connection has since gone has no
/// such owner left, so the answer is read off the held names rather than off the
/// demotion alone — and it is read by asking which slots a name names, never by
/// asking which slot a name resolves to, because two slots can answer to one
/// name and only one of them is the held connection's.
pub(super) fn held_read_only(inner: &DispatchInner, key: &str, slot: &SessionSlot) -> bool {
    slot.read_only
        && inner
            .revived
            .iter()
            .any(|(name, _, _)| names_session(key, slot, name))
}

/// Record a returning connection as read-only, once per connection.
pub(super) fn record_revived_connection(
    inner: &mut DispatchInner,
    session_id: &str,
    io: AdapterIo,
    capabilities: Vec<Capability>,
) {
    if !is_revived_connection(inner, &io) {
        inner
            .revived
            .push((session_id.to_string(), io, capabilities));
    }
}

/// Give one session back to the oldest connection that was held read-only for it.
///
/// A held connection is read-only only while another connection serves its
/// session, so the moment that connection goes is the moment the held one becomes
/// the session's only transport. Without this, a plugin that redials while the
/// client still holds the dead socket behind it is silenced for the rest of the
/// session: the assignment it came for is written to a connection nobody reads,
/// and nothing promotes it later. The demotion lifts with the promotion, so the
/// reconnected agent keeps both its session and its delivery rights.
fn promote_held_connection(inner: &mut DispatchInner, key: &str) {
    let attached = inner
        .sessions
        .get(key)
        .is_some_and(|slot| has_attached_transport(inner, key, slot));
    if attached {
        return;
    }
    let held = inner.revived.iter().position(|(name, _, _)| {
        slot_key_named(inner, name).is_some_and(|held_key| held_key == key)
    });
    let Some(index) = held else { return };
    let (name, io, capabilities) = inner.revived.remove(index);
    if let Some(slot) = inner.sessions.get_mut(key) {
        slot.read_only = false;
    }
    tracing::info!(
        session = %name,
        "a held connection takes the session its predecessor left"
    );
    attach_transport_locked(inner, &name, io, capabilities);
}

/// Settle what one mounting connection means for the session it names.
///
/// A mount that finds nothing serving its session takes it and clears the clock
/// [`DispatchState::release_connection`] started: that is the agent that came
/// back inside the reconnect grace, and the always-running agent serving task
/// after task lives in this path. A mount that finds the session already served
/// takes nothing — either another connection holds that very slot, or the task it
/// names now answers from a slot of its own, which is the case where a newer
/// session was spawned to retry the work while the old agent's process came back.
/// Such a connection is recorded read-only, and the slot it names is demoted too
/// when it owns a slot of its own.
///
/// Every binding path runs this one judgement, including the ready report, which
/// reaches an agent without writing a transport. Concurrency is what decides, not
/// the drop clock: the retry that claims an unclaimed session is served, and the
/// connection that returns to a session already served is held, whichever of the
/// two mounted first. [`DispatchState::release_connection`] promotes a held
/// connection when the live one it waited behind goes away, so a plugin that
/// redials over a socket the client has not yet seen die still gets its session.
pub(super) fn note_binding_locked(
    inner: &mut DispatchInner,
    session_id: &str,
    io: &AdapterIo,
) -> bool {
    if is_revived_connection(inner, io) {
        return false;
    }
    let Some(key) = slot_key_named(inner, session_id) else {
        return true;
    };
    let Some(slot) = inner.sessions.get(&key) else {
        return true;
    };
    let task = slot_task(slot);
    let taken = attached_to_other(inner, &key, slot, io);
    let moved_on = !taken
        && inner.sessions.iter().any(|(other, other_slot)| {
            *other != key
                && slot_task(other_slot) == task
                && attached_to_other(inner, other, other_slot, io)
        });
    let revived = taken || moved_on;
    // Whether this very connection already stands as the transport of the slot the
    // name resolves to. The answer decides what this judgement is allowed to
    // renew: a connection that already serves the session takes nothing new from
    // the mount, and the mount proves nothing the transport record did not
    // already prove. Without the distinction the stamp is a clock any caller can
    // wind, because `hand_staged` runs this judgement through `on_ready` for the
    // connection already serving a session — so every re-delivery of a task, and
    // every zombie mount that pushes one through, bought a dead agent a fresh
    // heartbeat window without one frame from the agent whose silence is what the
    // sweep reads. The stamp a returning agent does earn is written by the frames
    // it sends: the ready report of §6 stamps it when the reducer takes the row
    // (`note_beat`), and so does every accepted beat afterwards.
    let already_serving = slot_is_served_by(inner, &key, slot, io);
    // A mount that takes a session whose death clock was running is the agent
    // coming back inside the reconnect grace: the barrier had already passed, so
    // `ready` says the plugin spoke once, and the clock says the connection it
    // spoke through has since ended. That is the only shape this rebase is for.
    // A first mount has no clock running over a session that ever spoke, and a
    // settled session owes no work its reporter would answer for.
    let returning = !revived && slot.dropped_at.is_some() && slot.ready && slot.task_id.is_some();
    if let Some(slot) = inner.sessions.get_mut(&key) {
        if revived {
            // Only the spelling where this name's own slot is still served by
            // the newer connection leaves a slot of its own to silence; when it
            // is, the slot belongs to that live connection and keeps its rights.
            slot.read_only = moved_on;
        } else {
            slot.dropped_at = None;
            slot.read_only = false;
            // The mount is a frame this session's agent sent, and taking it is
            // what says the agent is here now. Without this the liveness stamp
            // would still read the moment before the drop, and the sweep's
            // silence arm would judge a returning agent that has not beaten yet
            // on the age of a frame from the process before it.
            if !already_serving {
                slot.last_beat = Some(Instant::now());
            }
        }
    }
    if revived {
        tracing::warn!(
            session = %session_id,
            task = %task,
            "a plugin mounted a session this client already serves; it is held read-only"
        );
        return false;
    }
    if returning {
        rebase_returned_reporter(inner, &key);
    }
    true
}

/// Rebase the watermark of a session whose agent came back, so its next frame
/// lands.
///
/// A plugin restarts its own sequence at its base and its generation is a
/// constant, while the watermark the row holds is the last sequence the process
/// that left reached. Left alone, every frame the returning reporter sends reads
/// at or below that watermark and is dropped as a stale duplicate until its
/// sequence climbs past it — for a session that had been running a while, the
/// whole rest of its work, reported into a tuple that never moves.
///
/// The rebase itself is [`rebase_generation`]'s: a new generation over the tuple
/// the client already holds, because the generation is the half the reporter
/// cannot be talked out of — its `generation` field is a constant it never
/// raises, so stamping the beat with the session's own generation is what lets a
/// frame from the new generation through at all, and the sequence starts again
/// under it.
///
/// The body is the stored tuple with the agent dimension put back to `Booting`
/// and the recovery line beside it dropped. That is not a guess about the agent:
/// the connection that witnessed the last agent fact is the one that ended, and
/// the plugin that just mounted has reported no turn fact yet, which is exactly
/// what `Booting` means. `recovery` goes because the reducer's own coupling
/// refuses a recovery substate beside a booting agent, and an accepted receipt
/// goes to `Pending` for the same reason — the repair `compose_observation`
/// already makes for a plugin that reports a booting process over a closed
/// drain. Everything else the client owns — the delivery drain, the reconcile
/// tuning, the counter — rides forward untouched, and so does the resource.
///
/// The generation being replaced is the one whose connection ended: this client
/// is the only authority on its own connection bookkeeping, and it carries the
/// observation content forward rather than replacing it, which is what the
/// attestation the reducer asks for is guarding against. A reporter from a
/// connection this client holds read-only never reaches the reducer at all —
/// `serves_session` refuses its frames before the gate — so lowering the
/// watermark here admits a returning reporter and nothing else.
fn rebase_returned_reporter(inner: &mut DispatchInner, key: &str) {
    let Some(slot) = inner.sessions.get(key) else {
        return;
    };
    let task_id = slot.session.task_id.clone();
    let verdict = rebase_generation(inner, &task_id, |stored| {
        let mut body = stored.clone();
        body.agent = AgentPhase::Booting;
        body.recovery = RecoveryPhase::NoRecovery;
        if body.delivery == DeliveryPhase::Accepted {
            body.delivery = DeliveryPhase::Pending;
        }
        body
    });
    match verdict {
        Ok(Some(Verdict::Applied(next))) => tracing::info!(
            task = %task_id,
            generation = next.version.generation,
            "a returning plugin's watermark was rebased onto a new generation"
        ),
        Ok(Some(verdict)) => tracing::warn!(
            task = %task_id,
            ?verdict,
            "the returning plugin's watermark was not rebased"
        ),
        Ok(None) => tracing::warn!(
            task = %task_id,
            "the returning plugin's row was not there to rebase"
        ),
        Err(error) => tracing::warn!(
            task = %task_id,
            error = %error,
            "the returning plugin's watermark was not rebased"
        ),
    }
}

/// Attach one plugin connection to the session it names, or hold it read-only.
///
/// This is where a mount that named a session becomes a transport, so the
/// read-only connection of §1 (b) never lands in `transports` and never steals
/// the assignment, delivery, or note addressed to the connection that serves the
/// session now. The one other writer of that map is the parked claim, which runs
/// the same judgement and keeps a refused agent in the park instead of recording
/// it read-only for a session it never named. Answers whether the connection took
/// the session.
fn attach_transport_locked(
    inner: &mut DispatchInner,
    session_id: &str,
    io: AdapterIo,
    capabilities: Vec<Capability>,
) -> bool {
    if !note_binding_locked(inner, session_id, &io) {
        record_revived_connection(inner, session_id, io, capabilities);
        return false;
    }
    inner
        .transports
        .insert(session_id.to_string(), (io, capabilities));
    true
}

impl DispatchState {
    /// Hold `io` for as long as one of its inbound frames is being handled.
    pub fn hold_frame(&self, io: &AdapterIo) -> FrameGuard<'_> {
        self.inner.lock().in_frame.push(io.clone());
        FrameGuard {
            state: self,
            io: io.clone(),
        }
    }

    /// Bind one adapter connection to the session it named.
    ///
    /// The name is the session id the client spawned the plugin with, which is
    /// enough on its own: a plugin that mounts before the client staged its
    /// session is remembered here and takes the payload the moment it is
    /// staged, and a plugin that mounts after finds its session waiting.
    pub fn bind_adapter(&self, session_id: &str, io: AdapterIo, capabilities: Vec<Capability>) {
        attach_transport_locked(&mut self.inner.lock(), session_id, io, capabilities);
    }

    /// Remember the delivery handle for one task.
    ///
    /// The handle goes to the session serving the task, not to a read-only slot
    /// that came back for it, so the ack this earns answers the live delivery.
    pub fn attach_msg_id(&self, task_id: &str, msg_id: &str) {
        let mut inner = self.inner.lock();
        let Some(key) = slot_key_serving_task(&inner, task_id) else {
            return;
        };
        if let Some(slot) = inner.sessions.get_mut(&key) {
            slot.msg_id = Some(msg_id.to_string());
        }
    }

    /// Take one plugin `send` frame and answer what the plugin is told.
    ///
    /// One path, whichever connection sent the frame. A handoff is a real
    /// delivery to the role it names, and this client is the only process
    /// holding a link that could carry it, so a connection that came back for a
    /// task another session now serves routes here exactly like the serving one.
    /// Nothing is held for a later merge (§3c): a session's handoffs travel as
    /// its own frames, answered where it sends them, so a settlement is only
    /// the completion's half of the account.
    pub fn plugin_send(&self, io: &AdapterIo, envelope: &Envelope) -> Result<serde_json::Value> {
        let mut inner = self.inner.lock();
        // The envelope leaves on this client's authenticated link, so the server
        // reads what it carries as this role's own message.
        send_is_authorised(&inner, envelope)?;
        let op_id = queue_outbound_locked(&mut inner, envelope)?;
        // A send the client carried is a delivery to the role it names, and it is
        // the evidence the relay guard reads at this session's next completion
        // (`guards.rs`). Only a connection that serves a session records: a
        // connection that came back for a task another session holds speaks for
        // no session of this client's.
        if let Some(key) = super::guards::session_key_of_connection(&inner, io) {
            super::guards::record_delivery(&mut inner, &key, &envelope.to);
        }
        Ok(serde_json::json!({"queued": true, "op_id": op_id}))
    }

    /// Park one plugin connection as this role's waiting agent.
    ///
    /// Only a mount that names no session parks: it is a plugin that attached
    /// before any work existed, so it takes the next session this role stages
    /// (plan §6 line 285). A role can host more than one such agent, and each
    /// that arrives joins the back of the queue: the record is a queue and not
    /// one slot, because overwriting it dropped the connection that was already
    /// waiting with no accounting of any kind — no release, no log, and no word
    /// to the plugin, which is how a quiet role lost a worker.
    ///
    /// A connection already in the queue is refreshed where it stands rather
    /// than sent to the back: a mount that names nothing twice over is the same
    /// agent re-helloing, and its place in line is that agent's due.
    pub fn park_transport(&self, io: AdapterIo, capabilities: Vec<Capability>) {
        let mut inner = self.inner.lock();
        let waiting = inner
            .parked
            .iter()
            .position(|(parked, _)| parked.same_connection(&io));
        match waiting {
            Some(index) => inner.parked[index].1 = capabilities,
            None => inner.parked.push((io, capabilities)),
        }
    }

    /// Claim the longest-waiting agent of this role for one staged session.
    ///
    /// An always-running plugin mounts naming no session, so the park holds the
    /// connection that can serve the session staged next (plan §6 line 285). The
    /// queue answers in the order it filled, so the agent that has waited longest
    /// is the one that takes the work. A claim left unbound strands the staged
    /// work: the session has a payload and this client holds no record of the
    /// socket that serves it, so a claim the binding judgement refuses returns the
    /// connection to the back of the park instead of spending it.
    ///
    /// A refused parked connection is not a mount that named a session and found
    /// it served, and the difference is what it is kept for: recording it
    /// read-only under a session it never mounted would hold it answerable to
    /// that session's task, and a role with more work to stage would lose the one
    /// waiting agent it has. It waits, and the socket's own end takes it out of
    /// the queue through `release_connection`.
    pub(super) fn claim_parked_transport(
        &self,
        session_id: &str,
    ) -> Option<(AdapterIo, Vec<Capability>)> {
        let mut inner = self.inner.lock();
        if inner.parked.is_empty() {
            return None;
        }
        // The queue is served oldest first.
        let (io, capabilities) = inner.parked.remove(0);
        // The judgement of §1 (b) runs first, and the transport is written only
        // once it answers that this connection takes the session: a refusal here
        // must leave no read-only record behind, which is why this path does not
        // run `attach_transport_locked`.
        if note_binding_locked(&mut inner, session_id, &io) {
            inner
                .transports
                .insert(session_id.to_string(), (io.clone(), capabilities.clone()));
            return Some((io, capabilities));
        }
        tracing::warn!(
            session = %session_id,
            "the longest-waiting agent took nothing from this session; it waits for the next"
        );
        inner.parked.push((io, capabilities));
        None
    }

    /// Hold `io` as this role's **standing** transport.
    ///
    /// A hosting runtime's connection is not a worker waiting for one job, so it
    /// does not join [`Self::park_transport`]'s queue: a claim takes the oldest
    /// entry and spends it, and a role whose hosting runtime serves four
    /// sessions would have nothing left for the other three. It is held here
    /// instead, where [`Self::staged_hosting_session`] finds it: the connection
    /// is asked for each session the role opens, and one connection serves as
    /// many as the role needs.
    ///
    /// A connection already standing is refreshed where it stands, the rule the
    /// park uses too: a mount that says this twice is one runtime re-helloing.
    pub fn stand_transport(&self, io: AdapterIo, capabilities: Vec<Capability>) {
        let mut inner = self.inner.lock();
        let standing = inner
            .standing
            .iter()
            .position(|(held, _)| held.same_connection(&io));
        match standing {
            Some(index) => inner.standing[index].1 = capabilities,
            None => inner.standing.push((io, capabilities)),
        }
    }

    /// Take the answer to an `open` a hosting runtime gave, and bind the
    /// connection that asked.
    ///
    /// The host named the session when it staged it, and the runtime names the
    /// conversation it opened. Both are kept: the slot stays under the host's name
    /// — that is the key every other lookup uses, and `session_tasks` binds through
    /// it — while the runtime's own name and its resume handle go beside it, so the
    /// next `open` for this family hands the handle back and the runtime resumes
    /// rather than starting a second conversation for one chain.
    ///
    /// The transport is written here because the judgement has already run: the
    /// runtime answered an `open`, and a refusal now would have to leave no
    /// binding behind.
    pub(crate) fn hosted_session_ready(
        &self,
        session_id: &str,
        opened: &onlyne_proto::OpenedArgs,
        io: AdapterIo,
        capabilities: Vec<Capability>,
    ) -> bool {
        let mut inner = self.inner.lock();
        let Some(slot) = inner.sessions.get_mut(session_id) else {
            return false;
        };
        if !opened.session_id.is_empty() {
            slot.session.backend_ref = serde_json::Value::String(opened.session_id.clone());
        }
        slot.resume_handle = opened.resume_handle.clone();
        inner
            .transports
            .insert(session_id.to_string(), (io, capabilities));
        true
    }

    /// The connection that serves one session, when its plugin is attached.
    ///
    /// A plugin names the session it was spawned for, and the slot's key is the
    /// other spelling worth trying.
    pub fn session_transport(&self, session_id: &str) -> Option<(AdapterIo, Vec<Capability>)> {
        let inner = self.inner.lock();
        if let Some(transport) = inner.transports.get(session_id) {
            return Some(transport.clone());
        }
        let key = inner
            .sessions
            .iter()
            .find(|(key, slot)| names_session(key, slot, session_id))
            .map(|(key, _)| key.clone())?;
        inner.transports.get(&key).cloned()
    }

    /// Tell the plugin serving one session to tear itself down, when that plugin
    /// implements `recycle`. A plugin without the capability is skipped: the
    /// caller's backend close stops the process either way.
    pub async fn recycle_plugin(&self, task_id: &str, reason: &str, outcome: Option<Outcome>) {
        let Some((io, capabilities)) = self.session_transport(task_id) else {
            return;
        };
        if missing_capability(&capabilities, Capability::Recycle) {
            tracing::debug!(
                task = %task_id,
                "plugin does not implement recycle; the host closes the resource"
            );
            return;
        }
        let args = RecycleArgs {
            task_id: task_id.to_string(),
            reason: reason.to_string(),
            outcome,
        };
        if let Err(error) = io.notify(AdapterMsg::Host(HostOp::Recycle(args))).await {
            tracing::warn!(error = %error, task = %task_id, "recycle frame did not reach the plugin");
        }
    }

    /// Ask the plugin serving one session for a fresh observation.
    ///
    /// The plugin answers with a heartbeat report, which is the reducer's
    /// evidence and the projection the operator reads. Answers whether a probe
    /// frame actually went out, and `false` says nothing was asked: a session no
    /// connection serves has no plugin to put the question to, and a `notify` that
    /// failed left the frame in this process. The caller must not record either
    /// as a probe that landed, because the answer a live plugin would have given
    /// never existed — the projection that says a plugin is gone is written when
    /// the session ends, never by this call.
    pub async fn probe_plugin(&self, task_id: &str) -> bool {
        let Some((io, _)) = self.session_transport(task_id) else {
            tracing::warn!(task = %task_id, "probe found no plugin connection to ask");
            return false;
        };
        let request = serde_json::json!({"task_id": task_id});
        if let Err(error) = io.notify(AdapterMsg::Host(HostOp::Probe(request))).await {
            tracing::warn!(error = %error, task = %task_id, "probe frame did not reach the plugin");
            return false;
        }
        true
    }

    /// Hand one session's own sentence to its agent through the plugin's input.
    ///
    /// The turn-end rule has one owner, this client, so the client is what tells
    /// a plugin-driven session that its turn ended without a completion
    /// (`docs/v2-CONTRACT.md` §3c). The plugin injects [`NUDGE_TEXT`] through the
    /// same channel an assignment's text takes and keeps no copy of it; the
    /// frame carries no envelope, no prose, and no attachments, and it is not a
    /// delivery: the task stays open and nothing in the plugin's turn
    /// bookkeeping is reset by it.
    ///
    /// Answers whether the frame went out. `false` is the honest answer for a
    /// session with no plugin to ask and for a plugin that declared no `inject`:
    /// a drive that cannot be nudged must not be told it was, and the caller
    /// settles the delivery at that turn end instead.
    pub async fn nudge_plugin(&self, task_id: &str) -> bool {
        let Some((io, capabilities)) = self.session_transport(task_id) else {
            tracing::debug!(
                task = %task_id,
                "nudge found no plugin connection to hand the sentence to"
            );
            return false;
        };
        if missing_capability(&capabilities, Capability::Inject) {
            tracing::debug!(
                task = %task_id,
                "plugin does not implement inject; the turn end settles the delivery"
            );
            return false;
        }
        let frame = AdapterMsg::Host(HostOp::Nudge {
            task_id: task_id.to_string(),
            text: NUDGE_TEXT.to_string(),
        });
        if let Err(error) = io.notify(frame).await {
            tracing::warn!(error = %error, task = %task_id, "nudge frame did not reach the plugin");
            return false;
        }
        true
    }

    /// Release the bindings served by one plugin connection.
    ///
    /// A graceful detach retires each idle session because the agent that owned
    /// it has left. An attached transport preserves the idle resource because
    /// the same agent is still reachable. A connection ending through
    /// another path preserves the slot and resource for an agent reconnection and
    /// starts the reconnect clock on it, which is what bounds how long a session
    /// waits for an agent that is never coming back. Every released binding
    /// retires its task progress clock. A slot carrying work remains under
    /// lifecycle ownership, and it carries that same clock: a goodbye and a
    /// silent drop both leave no heartbeat coming for the task it owes, and the
    /// window is what ends a session whose agent never returns.
    ///
    /// The session ids a goodbye retired come back, because that retirement wrote
    /// their rows: the resource close and, for a completed session, the agent's
    /// exit. The server mirrors only what this client reports, and a publish
    /// cannot run under this lock, so the caller is handed what to publish — the
    /// same answer [`DispatchState::retire_dropped_ghosts`] gives its sweep.
    pub fn release_connection(
        &self,
        session_id: Option<&str>,
        io: &AdapterIo,
        graceful_detach: bool,
    ) -> Vec<String> {
        let mut inner = self.inner.lock();
        // A read-only connection ending is not the session losing its agent: the
        // live connection still serves it, and its drop clock stays untouched.
        let revived_connection = {
            let before = inner.revived.len();
            inner
                .revived
                .retain(|(_, revived, _)| !revived.same_connection(io));
            before != inner.revived.len()
        };
        // A waiting agent whose socket ended leaves the queue and nothing else:
        // the agents behind it are other connections, still waiting for work.
        inner
            .parked
            .retain(|(parked, _)| !parked.same_connection(io));
        let released: Vec<String> = inner
            .transports
            .iter()
            .filter(|(session, (transport, _))| {
                transport.same_connection(io)
                    && session_id.is_none_or(|mounted| mounted == session.as_str())
            })
            .map(|(session, _)| session.clone())
            .collect();
        let served_tasks: Vec<String> = released
            .iter()
            .map(|session| {
                inner
                    .sessions
                    .iter()
                    .find(|(key, slot)| names_session(key, slot, session))
                    .map(|(_, slot)| {
                        slot.task_id
                            .clone()
                            .unwrap_or_else(|| slot.session.task_id.clone())
                    })
                    .unwrap_or_else(|| session.clone())
            })
            .collect();
        for task_id in served_tasks {
            inner.stall.forget(&task_id);
        }
        for session in &released {
            inner.transports.remove(session);
        }
        if !revived_connection {
            // The window of `[client] reconnect_grace_secs` starts wherever a
            // session loses the connection that would have sent its next
            // heartbeat and no other one is attached: the agent left without
            // saying so — every session that connection served — or it said
            // goodbye while its session still owed a task, which leaves no beat
            // coming either. The session keeps its slot and its resource for the
            // window, and the mount that returns inside it clears the stamp. A
            // gracefully detached session holding no task needs no window: the
            // idle retirement below takes that slot now.
            let now = Instant::now();
            for session in &released {
                if let Some((_, slot)) = inner.sessions.iter_mut().find(|(key, slot)| {
                    names_session(key, slot, session)
                        && (!graceful_detach || slot.task_id.is_some())
                }) {
                    slot.dropped_at = Some(now);
                }
            }
        }
        for session in &released {
            // Nothing serves this name any more, so the first connection that
            // mounted it read-only behind the one that just went becomes its
            // transport; a session with no such connection keeps waiting out the
            // reconnect grace, which is the sweep's to answer.
            if let Some(key) = slot_key_named(&inner, session) {
                promote_held_connection(&mut inner, &key);
            }
        }
        let mut retired: Vec<String> = Vec::new();
        let mut pending: Vec<PendingClose> = Vec::new();
        if graceful_detach {
            let idle: Vec<String> = released
                .iter()
                .filter_map(|session| {
                    inner
                        .sessions
                        .iter()
                        .find(|(key, slot)| {
                            names_session(key, slot, session) && slot.task_id.is_none()
                        })
                        .map(|(key, _)| key.clone())
                })
                .collect();
            for key in idle {
                // The id that travels is the session's own, the one whose row the
                // retirement below is about to write.
                let Some(task_id) = inner
                    .sessions
                    .get(&key)
                    .map(|slot| slot.session.task_id.clone())
                else {
                    continue;
                };
                // A scoped session whose runtime can resume is not this goodbye's
                // to end: the conversation is in the runtime's own store, and the
                // delivery that comes next starts the command again. Releasing the
                // process now is the same act the idle bound performs, and the
                // session is resumed the same way.
                if keeps_idle(&inner, &key) && resumable(&inner, &key) {
                    if suspend_locked(&mut inner, &key, &mut pending) {
                        retired.push(task_id);
                    }
                    continue;
                }
                let reason = inner
                    .sessions
                    .get(&key)
                    .and_then(|slot| stored_close_reason(&inner, &slot.session.task_id))
                    .unwrap_or(crate::backend::CloseReason::Completed);
                if retire_idle_locked(&mut inner, &key, reason, &mut pending) {
                    retired.push(task_id);
                }
            }
        }
        // The goodbye above took the idle slots; the host work their sessions owe
        // runs with the dispatch lock off it, so a plugin leaving does not make
        // every other session of this role wait out a pane close.
        drop(inner);
        close_retired(pending);
        retired
    }
}

/// Whether this client may carry one plugin `send` to the server as its own.
///
/// Both rules are the client's to enforce because both are about what this
/// process is willing to sign for: the server's acl answers for the wire, and it
/// answers for a link this client fills with whichever `from` a plugin thought to
/// write.
///
/// * A role principal besides this role is another session's voice. Every
///   legitimate sender stamps its own name: the plugin's `send` carries the role
///   its `hello` answer gave it, which is this client's role, and
///   [`DispatchState::plugin_handoff`] builds its envelope from the stored role.
///   A principal naming no role at all — a gateway or cluster sender — is left
///   alone: no agent plugin mounts with one, and the host paths that build such
///   envelopes never come through this door.
/// * `control` is the operator's plane. The server mints every command and admits
///   one from an admin link alone; a plugin that could queue one here would be
///   handing its own orders to `on_control` through the back of a message send.
///
/// A refusal is an error the sender reads, and reads correctly: unlike a state
/// report, a `send` the client will not carry has to be answered as a failure, or
/// the plugin records a handoff that never left.
///
/// [`DispatchState::plugin_handoff`]: DispatchState::plugin_handoff
fn send_is_authorised(inner: &DispatchInner, envelope: &Envelope) -> Result<()> {
    if envelope.kind == MsgKind::Control {
        tracing::warn!(
            role = %inner.role,
            to = %envelope.to,
            "a plugin send naming a control command was refused"
        );
        return Err(anyhow!("this connection may not queue a control command"));
    }
    if let Some(from) = envelope.from.role_name()
        && from != inner.role
    {
        tracing::warn!(
            role = %inner.role,
            from,
            "a plugin send written as another role was refused"
        );
        return Err(anyhow!("this role may not send as {from}"));
    }
    Ok(())
}

#[cfg(test)]
mod tests;