Public dry run dryrun-20b59dc46b72

Source commit: 20b59dc46b72103c2f8a516c692b5cc3d54fab19

Public tree identity: sha256:aaa5ac7b58b2a0d23c6b11e5f76324bf3839ca1f62653a83deb736932adcfbd9

crates/disasmer-coordinator/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "disasmer-coordinator"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[dependencies]
+base64.workspace = true
+disasmer-core = { path = "../disasmer-core" }
+disasmer-wasm-runtime = { path = "../disasmer-wasm-runtime" }
+postgres.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+sha2.workspace = true
+tempfile.workspace = true
+thiserror.workspace = true
+wasmparser.workspace = true

crates/disasmer-coordinator/src/agents.rs

@@ -0,0 +1,174 @@
+use disasmer_core::{
+    verify_agent_workflow_signature, Actor, AgentId, AgentSignedRequest, AgentWorkflowScope,
+    AuthContext, CredentialKind, Digest, ProjectId, TenantId, UserId,
+};
+
+use crate::{
+    AgentPublicKeyRecord, Coordinator, CoordinatorError, CredentialRecord, ProjectPermissionRecord,
+};
+
+impl Coordinator {
+    pub fn grant_project_debug(&mut self, tenant: TenantId, project: ProjectId, user: UserId) {
+        self.durable.project_permissions.insert(
+            (tenant.clone(), project.clone(), user.clone()),
+            ProjectPermissionRecord {
+                tenant,
+                project,
+                user,
+                can_debug: true,
+            },
+        );
+    }
+
+    pub fn register_agent_public_key(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        user: UserId,
+        agent: AgentId,
+        public_key: impl Into<String>,
+    ) -> AgentPublicKeyRecord {
+        let key = (tenant.clone(), project.clone(), agent.clone());
+        let version = self
+            .durable
+            .agent_public_keys
+            .get(&key)
+            .map_or(1, |record| record.version.saturating_add(1));
+        let public_key = public_key.into();
+        let public_key_fingerprint = Digest::sha256(&public_key);
+        let record = AgentPublicKeyRecord {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            user: user.clone(),
+            agent: agent.clone(),
+            public_key,
+            public_key_fingerprint: public_key_fingerprint.clone(),
+            version,
+            revoked: false,
+            scopes: vec!["project:read".to_owned(), "project:run".to_owned()],
+            human_account_creation_privilege: false,
+            browser_interaction_required_each_run: false,
+        };
+        self.durable.agent_public_keys.insert(key, record.clone());
+        let subject = format!("agent:{tenant}:{project}:{agent}");
+        self.durable.credentials.insert(
+            subject.clone(),
+            CredentialRecord {
+                subject,
+                tenant,
+                project: Some(project),
+                kind: CredentialKind::PublicKey,
+                public_key_fingerprint: Some(public_key_fingerprint),
+            },
+        );
+        record
+    }
+
+    pub fn list_agent_public_keys(&self, context: &AuthContext) -> Vec<AgentPublicKeyRecord> {
+        self.durable
+            .agent_public_keys
+            .values()
+            .filter(|record| record.tenant == context.tenant && record.project == context.project)
+            .filter(|record| match &context.actor {
+                Actor::User(user) => &record.user == user,
+                Actor::Agent(agent) => &record.agent == agent,
+                Actor::Node(_) | Actor::Task(_) => false,
+            })
+            .cloned()
+            .collect()
+    }
+
+    pub fn revoke_agent_public_key(
+        &mut self,
+        context: &AuthContext,
+        agent: &AgentId,
+    ) -> Result<AgentPublicKeyRecord, CoordinatorError> {
+        let key = (
+            context.tenant.clone(),
+            context.project.clone(),
+            agent.clone(),
+        );
+        let record = self
+            .durable
+            .agent_public_keys
+            .get_mut(&key)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "agent public key is not registered for this project".to_owned(),
+                )
+            })?;
+        match &context.actor {
+            Actor::User(user) if &record.user == user => {}
+            Actor::User(_) => {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent public key is outside the signed-in user scope".to_owned(),
+                ));
+            }
+            _ => {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent public-key revocation requires a user identity".to_owned(),
+                ));
+            }
+        }
+        record.revoked = true;
+        let subject = format!(
+            "agent:{}:{}:{}",
+            context.tenant, context.project, record.agent
+        );
+        self.durable.credentials.remove(&subject);
+        Ok(record.clone())
+    }
+
+    pub fn authorize_agent_project_run(
+        &self,
+        scope: AgentWorkflowScope<'_>,
+        public_key_fingerprint: Option<&Digest>,
+        payload_digest: &Digest,
+        signature: &AgentSignedRequest,
+        now_epoch_seconds: u64,
+    ) -> Result<AgentPublicKeyRecord, CoordinatorError> {
+        let record = self
+            .durable
+            .agent_public_keys
+            .get(&(
+                scope.tenant.clone(),
+                scope.project.clone(),
+                scope.agent.clone(),
+            ))
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "agent public key is not registered for this tenant/project".to_owned(),
+                )
+            })?;
+        if record.revoked {
+            return Err(CoordinatorError::Unauthorized(
+                "agent public key has been revoked".to_owned(),
+            ));
+        }
+        if let Some(public_key_fingerprint) = public_key_fingerprint {
+            if &record.public_key_fingerprint != public_key_fingerprint {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent public key fingerprint does not match the registered key".to_owned(),
+                ));
+            }
+        }
+        let max_signature_skew_seconds = 300;
+        if signature
+            .issued_at_epoch_seconds
+            .abs_diff(now_epoch_seconds)
+            > max_signature_skew_seconds
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "agent signed request is expired or outside the allowed clock skew".to_owned(),
+            ));
+        }
+        if !record.scopes.iter().any(|scope| scope == "project:run") {
+            return Err(CoordinatorError::Unauthorized(
+                "agent public key is not scoped for project runs".to_owned(),
+            ));
+        }
+        verify_agent_workflow_signature(&record.public_key, scope, payload_digest, signature)
+            .map_err(CoordinatorError::Unauthorized)?;
+        Ok(record.clone())
+    }
+}

crates/disasmer-coordinator/src/durable.rs

@@ -0,0 +1,145 @@
+use std::collections::BTreeMap;
+
+use disasmer_core::{
+    AgentId, CredentialKind, Digest, NodeId, ProjectId, SourceProviderKind, TenantId, UserId,
+};
+use serde::{Deserialize, Serialize};
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct TenantRecord {
+    pub id: TenantId,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct UserRecord {
+    pub id: UserId,
+    pub tenant: TenantId,
+    pub credential_kind: CredentialKind,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ProjectRecord {
+    pub id: ProjectId,
+    pub tenant: TenantId,
+    pub name: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct NodeIdentityRecord {
+    pub id: NodeId,
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub public_key: String,
+    pub enrollment_scope: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct CredentialRecord {
+    pub subject: String,
+    pub tenant: TenantId,
+    pub project: Option<ProjectId>,
+    pub kind: CredentialKind,
+    pub public_key_fingerprint: Option<Digest>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct CliSessionRecord {
+    pub session_digest: Digest,
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub user: UserId,
+    pub credential_kind: CredentialKind,
+    pub expires_at_epoch_seconds: Option<u64>,
+    pub revoked: bool,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct SourceProviderConfigRecord {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub provider: SourceProviderKind,
+    pub manifest_digest: Digest,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ServicePolicyRecord {
+    pub tenant: TenantId,
+    pub name: String,
+    pub digest: Digest,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct AccountPolicyState {
+    pub account_status: String,
+    pub suspended: bool,
+    pub disabled: bool,
+    pub deleted: bool,
+    pub manual_review: bool,
+    pub sanitized_reason: Option<String>,
+    pub next_actions: Vec<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ProjectPermissionRecord {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub user: UserId,
+    pub can_debug: bool,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct AgentPublicKeyRecord {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub user: UserId,
+    pub agent: AgentId,
+    pub public_key: String,
+    pub public_key_fingerprint: Digest,
+    pub version: u64,
+    pub revoked: bool,
+    pub scopes: Vec<String>,
+    pub human_account_creation_privilege: bool,
+    pub browser_interaction_required_each_run: bool,
+}
+
+#[derive(Clone, Debug, Default, Serialize, Deserialize)]
+pub struct DurableState {
+    pub tenants: BTreeMap<TenantId, TenantRecord>,
+    pub users: BTreeMap<UserId, UserRecord>,
+    pub projects: BTreeMap<ProjectId, ProjectRecord>,
+    pub node_identities: BTreeMap<NodeId, NodeIdentityRecord>,
+    pub credentials: BTreeMap<String, CredentialRecord>,
+    pub cli_sessions: BTreeMap<Digest, CliSessionRecord>,
+    pub source_provider_configs:
+        BTreeMap<(TenantId, ProjectId, String), SourceProviderConfigRecord>,
+    pub service_policy_records: BTreeMap<(TenantId, String), ServicePolicyRecord>,
+    pub project_permissions: BTreeMap<(TenantId, ProjectId, UserId), ProjectPermissionRecord>,
+    pub agent_public_keys: BTreeMap<(TenantId, ProjectId, AgentId), AgentPublicKeyRecord>,
+}
+
+pub trait DurableStore {
+    fn load(&self) -> DurableState;
+    fn save(&mut self, state: DurableState);
+}
+
+pub trait FallibleDurableStore {
+    type Error;
+
+    fn load_state(&mut self) -> Result<DurableState, Self::Error>;
+    fn save_state(&mut self, state: &DurableState) -> Result<(), Self::Error>;
+}
+
+#[derive(Clone, Debug, Default)]
+pub struct InMemoryDurableStore {
+    state: DurableState,
+}
+
+impl DurableStore for InMemoryDurableStore {
+    fn load(&self) -> DurableState {
+        self.state.clone()
+    }
+
+    fn save(&mut self, state: DurableState) {
+        self.state = state;
+    }
+}

crates/disasmer-coordinator/src/main.rs

@@ -0,0 +1,62 @@
+use std::io::Write;
+
+use disasmer_coordinator::{service::bind_listener, CoordinatorService};
+use disasmer_core::{ProjectId, TenantId, UserId};
+use serde_json::json;
+
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    let mut listen = "127.0.0.1:0".to_owned();
+    let mut allow_local_trusted = std::env::var("DISASMER_ALLOW_LOCAL_TRUSTED_LOOPBACK")
+        .ok()
+        .as_deref()
+        == Some("1");
+    let mut args = std::env::args().skip(1);
+    while let Some(arg) = args.next() {
+        if arg == "--listen" {
+            listen = args.next().ok_or("--listen requires an address")?;
+        } else if arg == "--allow-local-trusted-loopback" {
+            allow_local_trusted = true;
+        }
+    }
+
+    let (listener, addr) = bind_listener(&listen)?;
+    let database_url = std::env::var("DATABASE_URL").ok();
+    let mut service = CoordinatorService::new_with_database_url(1, database_url.as_deref())?;
+    let self_hosted_session_secret = std::env::var("DISASMER_SELF_HOSTED_SESSION_SECRET")
+        .ok()
+        .filter(|secret| !secret.trim().is_empty());
+    if let Some(session_secret) = self_hosted_session_secret.as_deref() {
+        service.issue_cli_session(
+            TenantId::new(
+                std::env::var("DISASMER_SELF_HOSTED_TENANT")
+                    .unwrap_or_else(|_| "tenant".to_owned()),
+            ),
+            ProjectId::new(
+                std::env::var("DISASMER_SELF_HOSTED_PROJECT")
+                    .unwrap_or_else(|_| "project".to_owned()),
+            ),
+            UserId::new(
+                std::env::var("DISASMER_SELF_HOSTED_USER").unwrap_or_else(|_| "user".to_owned()),
+            ),
+            session_secret,
+            None,
+        )?;
+    }
+    println!(
+        "{}",
+        json!({
+            "listen": addr.to_string(),
+            "client_authority": if allow_local_trusted { "local_trusted_loopback" } else { "strict" },
+            "self_hosted_session_bootstrapped": self_hosted_session_secret.is_some(),
+            "durable_store": service.durable_store_kind(),
+        })
+    );
+    std::io::stdout().flush()?;
+
+    if allow_local_trusted {
+        service.serve_tcp_local_trusted(listener)?;
+    } else {
+        service.serve_tcp(listener)?;
+    }
+    Ok(())
+}

crates/disasmer-coordinator/src/postgres_store.rs

@@ -0,0 +1,576 @@
+use postgres::{Client, NoTls};
+use serde::{de::DeserializeOwned, Serialize};
+use thiserror::Error;
+
+use crate::{
+    AgentPublicKeyRecord, CliSessionRecord, CredentialRecord, DurableState, FallibleDurableStore,
+    NodeIdentityRecord, ProjectPermissionRecord, ProjectRecord, ServicePolicyRecord,
+    SourceProviderConfigRecord, TenantRecord, UserRecord,
+};
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct PostgresTable {
+    pub name: &'static str,
+    pub durable_record: &'static str,
+    pub restart_surviving: bool,
+}
+
+pub const POSTGRES_DURABLE_TABLES: &[PostgresTable] = &[
+    PostgresTable {
+        name: "disasmer_tenants",
+        durable_record: "tenants",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_users",
+        durable_record: "users",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_projects",
+        durable_record: "projects",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_node_identities",
+        durable_record: "node identities",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_credentials",
+        durable_record: "credentials",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_cli_sessions",
+        durable_record: "CLI sessions",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_agent_public_keys",
+        durable_record: "agent public keys",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_source_provider_configs",
+        durable_record: "source-provider configuration",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_service_policy_records",
+        durable_record: "durable service policy records",
+        restart_surviving: true,
+    },
+    PostgresTable {
+        name: "disasmer_project_permissions",
+        durable_record: "explicit project permissions",
+        restart_surviving: true,
+    },
+];
+
+#[derive(Debug, Error)]
+pub enum PostgresStoreError {
+    #[error("postgres durable store error: {0}")]
+    Postgres(#[from] postgres::Error),
+    #[error("durable state serialization error: {0}")]
+    Serialization(#[from] serde_json::Error),
+}
+
+pub struct PostgresDurableStore {
+    client: Client,
+}
+
+impl PostgresDurableStore {
+    pub fn connect(connection_string: &str) -> Result<Self, PostgresStoreError> {
+        let mut store = Self {
+            client: Client::connect(connection_string, NoTls)?,
+        };
+        store.migrate()?;
+        Ok(store)
+    }
+
+    pub fn from_client(client: Client) -> Result<Self, PostgresStoreError> {
+        let mut store = Self { client };
+        store.migrate()?;
+        Ok(store)
+    }
+
+    pub fn schema_sql() -> &'static str {
+        POSTGRES_SCHEMA_SQL
+    }
+
+    pub fn durable_tables() -> &'static [PostgresTable] {
+        POSTGRES_DURABLE_TABLES
+    }
+
+    pub fn migrate(&mut self) -> Result<(), PostgresStoreError> {
+        self.client.batch_execute(Self::schema_sql())?;
+        Ok(())
+    }
+
+    fn query_records<T: DeserializeOwned>(
+        &mut self,
+        sql: &str,
+    ) -> Result<Vec<T>, PostgresStoreError> {
+        self.client
+            .query(sql, &[])?
+            .into_iter()
+            .map(|row| {
+                let value: serde_json::Value = row.get("record");
+                Ok(serde_json::from_value(value)?)
+            })
+            .collect()
+    }
+
+    fn record_value(record: &impl Serialize) -> Result<serde_json::Value, PostgresStoreError> {
+        Ok(serde_json::to_value(record)?)
+    }
+}
+
+impl FallibleDurableStore for PostgresDurableStore {
+    type Error = PostgresStoreError;
+
+    fn load_state(&mut self) -> Result<DurableState, Self::Error> {
+        let mut state = DurableState::default();
+
+        for record in self.query_records::<TenantRecord>(
+            "SELECT record FROM disasmer_tenants ORDER BY tenant_id",
+        )? {
+            state.tenants.insert(record.id.clone(), record);
+        }
+        for record in
+            self.query_records::<UserRecord>("SELECT record FROM disasmer_users ORDER BY user_id")?
+        {
+            state.users.insert(record.id.clone(), record);
+        }
+        for record in self.query_records::<ProjectRecord>(
+            "SELECT record FROM disasmer_projects ORDER BY project_id",
+        )? {
+            state.projects.insert(record.id.clone(), record);
+        }
+        for record in self.query_records::<NodeIdentityRecord>(
+            "SELECT record FROM disasmer_node_identities ORDER BY node_id",
+        )? {
+            state.node_identities.insert(record.id.clone(), record);
+        }
+        for record in self.query_records::<CredentialRecord>(
+            "SELECT record FROM disasmer_credentials ORDER BY subject",
+        )? {
+            state.credentials.insert(record.subject.clone(), record);
+        }
+        for record in self.query_records::<CliSessionRecord>(
+            "SELECT record FROM disasmer_cli_sessions ORDER BY session_digest",
+        )? {
+            state
+                .cli_sessions
+                .insert(record.session_digest.clone(), record);
+        }
+        for record in self.query_records::<AgentPublicKeyRecord>(
+            "SELECT record FROM disasmer_agent_public_keys ORDER BY tenant_id, project_id, agent_id",
+        )? {
+            state.agent_public_keys.insert(
+                (
+                    record.tenant.clone(),
+                    record.project.clone(),
+                    record.agent.clone(),
+                ),
+                record,
+            );
+        }
+        for record in self.query_records::<SourceProviderConfigRecord>(
+            "SELECT record FROM disasmer_source_provider_configs ORDER BY tenant_id, project_id, provider_key",
+        )? {
+            let provider_key = format!("{:?}", record.provider);
+            state.source_provider_configs.insert(
+                (record.tenant.clone(), record.project.clone(), provider_key),
+                record,
+            );
+        }
+        for record in self.query_records::<ServicePolicyRecord>(
+            "SELECT record FROM disasmer_service_policy_records ORDER BY tenant_id, name",
+        )? {
+            state
+                .service_policy_records
+                .insert((record.tenant.clone(), record.name.clone()), record);
+        }
+        for record in self.query_records::<ProjectPermissionRecord>(
+            "SELECT record FROM disasmer_project_permissions ORDER BY tenant_id, project_id, user_id",
+        )? {
+            state.project_permissions.insert(
+                (
+                    record.tenant.clone(),
+                    record.project.clone(),
+                    record.user.clone(),
+                ),
+                record,
+            );
+        }
+
+        Ok(state)
+    }
+
+    fn save_state(&mut self, state: &DurableState) -> Result<(), Self::Error> {
+        let mut tx = self.client.transaction()?;
+        tx.batch_execute(
+            "
+            DELETE FROM disasmer_project_permissions;
+            DELETE FROM disasmer_service_policy_records;
+            DELETE FROM disasmer_source_provider_configs;
+            DELETE FROM disasmer_agent_public_keys;
+            DELETE FROM disasmer_cli_sessions;
+            DELETE FROM disasmer_credentials;
+            DELETE FROM disasmer_node_identities;
+            DELETE FROM disasmer_projects;
+            DELETE FROM disasmer_users;
+            DELETE FROM disasmer_tenants;
+            ",
+        )?;
+
+        for record in state.tenants.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_tenants (tenant_id, record) VALUES ($1, $2)",
+                &[&record.id.as_str(), &value],
+            )?;
+        }
+        for record in state.users.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_users (user_id, tenant_id, record) VALUES ($1, $2, $3)",
+                &[&record.id.as_str(), &record.tenant.as_str(), &value],
+            )?;
+        }
+        for record in state.projects.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_projects (project_id, tenant_id, record) VALUES ($1, $2, $3)",
+                &[&record.id.as_str(), &record.tenant.as_str(), &value],
+            )?;
+        }
+        for record in state.node_identities.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_node_identities (node_id, tenant_id, project_id, record) VALUES ($1, $2, $3, $4)",
+                &[
+                    &record.id.as_str(),
+                    &record.tenant.as_str(),
+                    &record.project.as_str(),
+                    &value,
+                ],
+            )?;
+        }
+        for record in state.credentials.values() {
+            let value = Self::record_value(record)?;
+            let project_id = record.project.as_ref().map(|project| project.as_str());
+            tx.execute(
+                "INSERT INTO disasmer_credentials (subject, tenant_id, project_id, record) VALUES ($1, $2, $3, $4)",
+                &[&record.subject.as_str(), &record.tenant.as_str(), &project_id, &value],
+            )?;
+        }
+        for record in state.cli_sessions.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_cli_sessions (session_digest, tenant_id, project_id, user_id, record) VALUES ($1, $2, $3, $4, $5)",
+                &[
+                    &record.session_digest.as_str(),
+                    &record.tenant.as_str(),
+                    &record.project.as_str(),
+                    &record.user.as_str(),
+                    &value,
+                ],
+            )?;
+        }
+        for record in state.agent_public_keys.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_agent_public_keys (tenant_id, project_id, user_id, agent_id, record) VALUES ($1, $2, $3, $4, $5)",
+                &[
+                    &record.tenant.as_str(),
+                    &record.project.as_str(),
+                    &record.user.as_str(),
+                    &record.agent.as_str(),
+                    &value,
+                ],
+            )?;
+        }
+        for ((_, _, provider_key), record) in &state.source_provider_configs {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_source_provider_configs (tenant_id, project_id, provider_key, record) VALUES ($1, $2, $3, $4)",
+                &[
+                    &record.tenant.as_str(),
+                    &record.project.as_str(),
+                    &provider_key.as_str(),
+                    &value,
+                ],
+            )?;
+        }
+        for record in state.service_policy_records.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_service_policy_records (tenant_id, name, record) VALUES ($1, $2, $3)",
+                &[&record.tenant.as_str(), &record.name.as_str(), &value],
+            )?;
+        }
+        for record in state.project_permissions.values() {
+            let value = Self::record_value(record)?;
+            tx.execute(
+                "INSERT INTO disasmer_project_permissions (tenant_id, project_id, user_id, record) VALUES ($1, $2, $3, $4)",
+                &[
+                    &record.tenant.as_str(),
+                    &record.project.as_str(),
+                    &record.user.as_str(),
+                    &value,
+                ],
+            )?;
+        }
+
+        tx.commit()?;
+        Ok(())
+    }
+}
+
+const POSTGRES_SCHEMA_SQL: &str = r#"
+CREATE TABLE IF NOT EXISTS disasmer_tenants (
+    tenant_id TEXT PRIMARY KEY,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_users (
+    user_id TEXT PRIMARY KEY,
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_projects (
+    project_id TEXT PRIMARY KEY,
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_node_identities (
+    node_id TEXT PRIMARY KEY,
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT NOT NULL REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_credentials (
+    subject TEXT PRIMARY KEY,
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_cli_sessions (
+    session_digest TEXT PRIMARY KEY,
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT NOT NULL REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    user_id TEXT NOT NULL REFERENCES disasmer_users(user_id) ON DELETE CASCADE,
+    record JSONB NOT NULL
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_agent_public_keys (
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT NOT NULL REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    user_id TEXT NOT NULL REFERENCES disasmer_users(user_id) ON DELETE CASCADE,
+    agent_id TEXT NOT NULL,
+    record JSONB NOT NULL,
+    PRIMARY KEY (tenant_id, project_id, agent_id)
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_source_provider_configs (
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT NOT NULL REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    provider_key TEXT NOT NULL,
+    record JSONB NOT NULL,
+    PRIMARY KEY (tenant_id, project_id, provider_key)
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_service_policy_records (
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    name TEXT NOT NULL,
+    record JSONB NOT NULL,
+    PRIMARY KEY (tenant_id, name)
+);
+
+CREATE TABLE IF NOT EXISTS disasmer_project_permissions (
+    tenant_id TEXT NOT NULL REFERENCES disasmer_tenants(tenant_id) ON DELETE CASCADE,
+    project_id TEXT NOT NULL REFERENCES disasmer_projects(project_id) ON DELETE CASCADE,
+    user_id TEXT NOT NULL REFERENCES disasmer_users(user_id) ON DELETE CASCADE,
+    record JSONB NOT NULL,
+    PRIMARY KEY (tenant_id, project_id, user_id)
+);
+"#;
+
+#[cfg(test)]
+mod tests {
+    use disasmer_core::{
+        CredentialKind, Digest, NodeId, ProjectId, SourceProviderKind, TenantId, UserId,
+    };
+
+    use super::*;
+    use crate::{Coordinator, DurableStore, FallibleDurableStore, InMemoryDurableStore};
+
+    #[test]
+    fn postgres_schema_contains_only_restart_surviving_durable_tables() {
+        let names = PostgresDurableStore::durable_tables()
+            .iter()
+            .map(|table| table.name)
+            .collect::<Vec<_>>();
+
+        assert_eq!(names.len(), 10);
+        assert!(names.contains(&"disasmer_tenants"));
+        assert!(names.contains(&"disasmer_users"));
+        assert!(names.contains(&"disasmer_projects"));
+        assert!(names.contains(&"disasmer_node_identities"));
+        assert!(names.contains(&"disasmer_credentials"));
+        assert!(names.contains(&"disasmer_cli_sessions"));
+        assert!(names.contains(&"disasmer_agent_public_keys"));
+        assert!(names.contains(&"disasmer_source_provider_configs"));
+        assert!(names.contains(&"disasmer_service_policy_records"));
+        assert!(names.contains(&"disasmer_project_permissions"));
+        assert!(PostgresDurableStore::durable_tables()
+            .iter()
+            .all(|table| table.restart_surviving));
+
+        for runtime_only in [
+            "active_process",
+            "virtual_thread",
+            "scheduler_state",
+            "debug_epoch",
+            "vfs_manifest",
+            "transient_artifact_location",
+        ] {
+            assert!(
+                !PostgresDurableStore::schema_sql().contains(runtime_only),
+                "{runtime_only} must remain outside Postgres durable state"
+            );
+        }
+    }
+
+    #[test]
+    fn fallible_store_boot_uses_durable_state_and_still_drops_live_processes() {
+        #[derive(Default)]
+        struct FallibleMemoryStore {
+            inner: InMemoryDurableStore,
+        }
+
+        impl FallibleDurableStore for FallibleMemoryStore {
+            type Error = std::convert::Infallible;
+
+            fn load_state(&mut self) -> Result<DurableState, Self::Error> {
+                Ok(self.inner.load())
+            }
+
+            fn save_state(&mut self, state: &DurableState) -> Result<(), Self::Error> {
+                self.inner.save(state.clone());
+                Ok(())
+            }
+        }
+
+        let mut store = FallibleMemoryStore::default();
+        let mut first = Coordinator::try_boot(&mut store, 1).unwrap();
+        first.upsert_tenant(TenantId::from("tenant"));
+        first.upsert_user(
+            TenantId::from("tenant"),
+            UserId::from("user"),
+            CredentialKind::CliDeviceSession,
+        );
+        first.upsert_project(TenantId::from("tenant"), ProjectId::from("project"), "demo");
+        first.enroll_node(
+            TenantId::from("tenant"),
+            ProjectId::from("project"),
+            NodeId::from("node"),
+            "public-key",
+            "node:attach",
+        );
+        first.upsert_source_provider_config(
+            TenantId::from("tenant"),
+            ProjectId::from("project"),
+            SourceProviderKind::Git,
+            Digest::sha256("git-manifest"),
+        );
+        first.start_process(
+            TenantId::from("tenant"),
+            ProjectId::from("project"),
+            disasmer_core::ProcessId::from("process"),
+        );
+        first.try_persist(&mut store).unwrap();
+
+        let restarted = Coordinator::try_boot(&mut store, 2).unwrap();
+
+        assert!(restarted.project(&ProjectId::from("project")).is_some());
+        assert!(restarted.node_identity(&NodeId::from("node")).is_some());
+        assert_eq!(restarted.active_process_count(), 0);
+    }
+
+    #[test]
+    fn postgres_round_trip_runs_when_dsn_is_configured() {
+        let Ok(dsn) = std::env::var("DISASMER_TEST_POSTGRES") else {
+            return;
+        };
+
+        let mut store = PostgresDurableStore::connect(&dsn).unwrap();
+        let mut state = DurableState::default();
+        state.tenants.insert(
+            TenantId::from("tenant"),
+            TenantRecord {
+                id: TenantId::from("tenant"),
+            },
+        );
+        state.projects.insert(
+            ProjectId::from("project"),
+            ProjectRecord {
+                id: ProjectId::from("project"),
+                tenant: TenantId::from("tenant"),
+                name: "demo".to_owned(),
+            },
+        );
+        state.users.insert(
+            UserId::from("user"),
+            UserRecord {
+                id: UserId::from("user"),
+                tenant: TenantId::from("tenant"),
+                credential_kind: CredentialKind::CliDeviceSession,
+            },
+        );
+        let session_digests = [
+            Digest::sha256("postgres-round-trip-session-one"),
+            Digest::sha256("postgres-round-trip-session-two"),
+        ];
+        for session_digest in &session_digests {
+            state.cli_sessions.insert(
+                session_digest.clone(),
+                CliSessionRecord {
+                    session_digest: session_digest.clone(),
+                    tenant: TenantId::from("tenant"),
+                    project: ProjectId::from("project"),
+                    user: UserId::from("user"),
+                    credential_kind: CredentialKind::CliDeviceSession,
+                    expires_at_epoch_seconds: None,
+                    revoked: false,
+                },
+            );
+            let subject = format!("cli-session:{}", session_digest.as_str());
+            state.credentials.insert(
+                subject.clone(),
+                CredentialRecord {
+                    subject,
+                    tenant: TenantId::from("tenant"),
+                    project: Some(ProjectId::from("project")),
+                    kind: CredentialKind::CliDeviceSession,
+                    public_key_fingerprint: None,
+                },
+            );
+        }
+        store.save_state(&state).unwrap();
+        let loaded = store.load_state().unwrap();
+
+        assert!(loaded.projects.contains_key(&ProjectId::from("project")));
+        assert!(session_digests
+            .iter()
+            .all(|session_digest| loaded.cli_sessions.contains_key(session_digest)));
+        assert_eq!(loaded.credentials.len(), 2);
+    }
+}

crates/disasmer-coordinator/src/service.rs

@@ -0,0 +1,423 @@
+// Request handlers intentionally spell out their deserialized protocol fields.
+// Keeping those authority and payload values explicit at this boundary is safer
+// than passing an unvalidated wire request deeper into the service.
+#![allow(clippy::too_many_arguments)]
+
+use std::collections::{BTreeMap, BTreeSet, VecDeque};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use disasmer_core::{
+    Actor, AgentId, ArtifactRegistry, CapabilityReportError, CredentialKind, Digest, LimitError,
+    NativeQuicTransport, NodeDescriptor, NodeId, PanelState, Placement, ProcessId, ProjectId,
+    RateLimit, TenantId, TransportError, UserId,
+};
+use thiserror::Error;
+
+use crate::{Coordinator, CoordinatorError};
+
+mod admin;
+mod artifacts;
+mod authenticated;
+mod authorization;
+mod debug;
+mod debug_requests;
+mod durable_runtime;
+mod keys;
+mod logs;
+mod main_runtime;
+mod nodes;
+mod panels;
+mod process_launch;
+mod processes;
+mod protocol;
+mod quota;
+mod routing;
+mod signed_nodes;
+mod tcp;
+mod wire_protocol;
+use authorization::authorize_authenticated_user_operation;
+use durable_runtime::RuntimeDurableStore;
+use keys::{
+    artifact_id_from_path, enrollment_grant_key, EnrollmentGrantKey, PanelStopKey,
+    ProcessControlKey, TaskAssignmentKey, TaskControlKey, TaskRestartKey,
+};
+pub use protocol::{
+    ArtifactTransferAssignment, AuthenticatedCoordinatorRequest, CoordinatorRequest,
+    CoordinatorResponse, DebugAcknowledgementState, DebugAuditEvent,
+    DebugParticipantAcknowledgement, SourcePreparationDisposition, SourcePreparationStatus,
+    TaskAssignment, TaskCancellationTarget, TaskCompletionEvent, TaskExecutor,
+    TaskReplacementBundle, TaskTerminalState, VirtualProcessStatus, WorkflowActor,
+};
+pub use quota::CoordinatorQuotaConfiguration;
+pub use tcp::{bind_listener, ClientAuthorityMode};
+pub use wire_protocol::CoordinatorWireRequest;
+
+const MAX_TASK_LOG_TAIL_BYTES: usize = 256 * 1024;
+const DEBUG_CONTROL_READ_BYTES: u64 = 1024;
+const MAX_REPLAY_NONCES_PER_AUTHORITY: usize = 1_024;
+const NODE_SIGNATURE_WINDOW_SECONDS: u64 = 30;
+const MAX_NODE_REPLAY_NONCES_PER_AUTHORITY: usize = 4_096;
+const MAX_ENROLLMENT_GRANTS_PER_PROJECT: usize = 64;
+const MAX_TASK_EVENTS_PER_PROCESS: usize = 128;
+const MAX_DEBUG_AUDIT_EVENTS_PER_PROCESS: usize = 256;
+const MAX_RESTART_CHECKPOINTS_PER_PROCESS: usize = 128;
+const MAX_IN_FLIGHT_TASKS_PER_PROCESS: usize = 256;
+const MAX_NODE_REPORTED_OBJECTS_PER_KIND: usize = 1_024;
+fn bounded_ttl(requested: u64, maximum: u64) -> u64 {
+    requested.clamp(1, maximum)
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct CoordinatorAdmission {
+    pub workflow_placement_allowed: bool,
+    pub max_node_enrollment_ttl_seconds: u64,
+    pub max_artifact_download_ttl_seconds: u64,
+}
+
+impl Default for CoordinatorAdmission {
+    fn default() -> Self {
+        Self {
+            workflow_placement_allowed: true,
+            max_node_enrollment_ttl_seconds: 15 * 60,
+            max_artifact_download_ttl_seconds: 15 * 60,
+        }
+    }
+}
+
+#[derive(Debug, Error)]
+pub enum CoordinatorServiceError {
+    #[error("coordinator protocol I/O error: {0}")]
+    Io(#[from] std::io::Error),
+    #[error("coordinator protocol JSON error: {0}")]
+    Json(#[from] serde_json::Error),
+    #[error("coordinator protocol error: {0}")]
+    Protocol(String),
+    #[error("coordinator request failed: {0}")]
+    Coordinator(#[from] CoordinatorError),
+    #[error("artifact download request failed: {0}")]
+    Download(#[from] disasmer_core::DownloadError),
+    #[error("scheduler placement failed: {0}")]
+    Scheduler(#[from] disasmer_core::PlacementError),
+    #[error("transport request failed: {0}")]
+    Transport(#[from] TransportError),
+    #[error("resource limit failed: {0}")]
+    Resource(#[from] LimitError),
+    #[error("operator panel request failed: {0}")]
+    Panel(#[from] disasmer_core::PanelError),
+    #[error("invalid node capability report: {0}")]
+    CapabilityReport(#[from] CapabilityReportError),
+    #[error("invalid VFS artifact path reported by node: {0}")]
+    InvalidArtifactPath(String),
+    #[error("invalid task log tail reported by node: {0}")]
+    InvalidTaskLogTail(String),
+    #[error("durable coordinator state failed: {0}")]
+    Durable(String),
+}
+
+pub struct CoordinatorService {
+    coordinator: Coordinator,
+    store: RuntimeDurableStore,
+    node_descriptors: BTreeMap<NodeId, NodeDescriptor>,
+    enrollment_grants: BTreeMap<EnrollmentGrantKey, disasmer_core::EnrollmentGrant>,
+    task_events: VecDeque<TaskCompletionEvent>,
+    debug_audit_events: VecDeque<DebugAuditEvent>,
+    debug_epochs: BTreeMap<ProcessControlKey, u64>,
+    debug_epoch_runtime: BTreeMap<ProcessControlKey, debug::DebugEpochRuntime>,
+    debug_breakpoints: BTreeMap<ProcessControlKey, debug::DebugBreakpointPlan>,
+    debug_commands: BTreeMap<TaskControlKey, debug::DebugPendingCommand>,
+    task_assignments: BTreeMap<TaskAssignmentKey, VecDeque<TaskAssignment>>,
+    task_restart_checkpoints: BTreeMap<TaskRestartKey, processes::TaskRestartCheckpoint>,
+    task_restart_checkpoint_order: VecDeque<TaskRestartKey>,
+    main_runtime: main_runtime::CoordinatorMainRuntime,
+    pending_task_launches: VecDeque<processes::PendingTaskLaunch>,
+    task_placements: BTreeMap<TaskControlKey, Placement>,
+    active_tasks: BTreeSet<TaskControlKey>,
+    task_cancellations: BTreeSet<TaskControlKey>,
+    task_aborts: BTreeSet<TaskControlKey>,
+    process_cancellations: BTreeSet<ProcessControlKey>,
+    process_aborts: BTreeSet<ProcessControlKey>,
+    agent_replay_nonces: BTreeMap<(TenantId, ProjectId, AgentId, String), u64>,
+    node_replay_nonces: BTreeMap<(NodeId, String), u64>,
+    panel_snapshots: BTreeMap<PanelStopKey, PanelState>,
+    stopped_panels: BTreeSet<PanelStopKey>,
+    panel_event_limits: BTreeMap<(TenantId, ProjectId, ProcessId, String), RateLimit>,
+    artifact_registry: ArtifactRegistry,
+    artifact_reverse_transfers: BTreeMap<String, artifacts::ArtifactReverseTransfer>,
+    artifact_transfer_by_token: BTreeMap<Digest, String>,
+    transport: NativeQuicTransport,
+    quota: quota::CoordinatorQuota,
+    admission: CoordinatorAdmission,
+    #[cfg(test)]
+    server_time_override: Option<u64>,
+    admin_token_digest: Option<Digest>,
+    admin_replay_nonces: BTreeMap<String, u64>,
+}
+
+impl CoordinatorService {
+    pub(super) fn authorize_node_for_process_or_termination(
+        &self,
+        node: &NodeId,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) -> Result<(), CoordinatorServiceError> {
+        let process_key = keys::process_control_key(tenant, project, process);
+        if self.process_cancellations.contains(&process_key)
+            || self.process_aborts.contains(&process_key)
+        {
+            let identity = self
+                .coordinator
+                .node_identity(node)
+                .ok_or(CoordinatorError::UnknownNode)?;
+            if &identity.tenant != tenant || &identity.project != project {
+                return Err(CoordinatorError::Unauthorized(
+                    "node process-control request is outside its enrolled tenant/project scope"
+                        .to_owned(),
+                )
+                .into());
+            }
+            return Ok(());
+        }
+        self.coordinator
+            .authorize_node_for_process(node, tenant, project, process)?;
+        Ok(())
+    }
+
+    pub fn new(coordinator_epoch: u64) -> Self {
+        Self::new_with_optional_admin_token_and_admission(
+            coordinator_epoch,
+            std::env::var("DISASMER_ADMIN_TOKEN").ok(),
+            CoordinatorAdmission::default(),
+        )
+    }
+
+    pub fn new_with_admin_token(coordinator_epoch: u64, admin_token: impl Into<String>) -> Self {
+        Self::new_with_optional_admin_token_and_admission(
+            coordinator_epoch,
+            Some(admin_token.into()),
+            CoordinatorAdmission::default(),
+        )
+    }
+
+    pub fn new_with_admission(coordinator_epoch: u64, admission: CoordinatorAdmission) -> Self {
+        Self::new_with_optional_admin_token_and_admission(
+            coordinator_epoch,
+            std::env::var("DISASMER_ADMIN_TOKEN").ok(),
+            admission,
+        )
+    }
+
+    fn new_with_optional_admin_token_and_admission(
+        coordinator_epoch: u64,
+        admin_token: Option<String>,
+        admission: CoordinatorAdmission,
+    ) -> Self {
+        Self::try_new_with_optional_admin_token_admission_and_database_url(
+            coordinator_epoch,
+            admin_token,
+            admission,
+            None,
+            CoordinatorQuotaConfiguration::default(),
+        )
+        .expect("in-memory durable coordinator store initialization cannot fail")
+    }
+
+    pub fn new_with_database_url(
+        coordinator_epoch: u64,
+        database_url: Option<&str>,
+    ) -> Result<Self, CoordinatorServiceError> {
+        Self::try_new_with_optional_admin_token_admission_and_database_url(
+            coordinator_epoch,
+            std::env::var("DISASMER_ADMIN_TOKEN").ok(),
+            CoordinatorAdmission::default(),
+            database_url,
+            CoordinatorQuotaConfiguration::default(),
+        )
+    }
+
+    pub fn new_with_admin_token_and_database_url(
+        coordinator_epoch: u64,
+        admin_token: impl Into<String>,
+        database_url: Option<&str>,
+    ) -> Result<Self, CoordinatorServiceError> {
+        Self::new_with_admin_token_database_url_and_quota(
+            coordinator_epoch,
+            admin_token,
+            database_url,
+            CoordinatorQuotaConfiguration::default(),
+        )
+    }
+
+    pub fn new_with_admin_token_database_url_and_quota(
+        coordinator_epoch: u64,
+        admin_token: impl Into<String>,
+        database_url: Option<&str>,
+        quota_configuration: CoordinatorQuotaConfiguration,
+    ) -> Result<Self, CoordinatorServiceError> {
+        Self::try_new_with_optional_admin_token_admission_and_database_url(
+            coordinator_epoch,
+            Some(admin_token.into()),
+            CoordinatorAdmission::default(),
+            database_url,
+            quota_configuration,
+        )
+    }
+
+    fn try_new_with_optional_admin_token_admission_and_database_url(
+        coordinator_epoch: u64,
+        admin_token: Option<String>,
+        admission: CoordinatorAdmission,
+        database_url: Option<&str>,
+        quota_configuration: CoordinatorQuotaConfiguration,
+    ) -> Result<Self, CoordinatorServiceError> {
+        let mut store = RuntimeDurableStore::from_database_url(database_url)
+            .map_err(CoordinatorServiceError::Durable)?;
+        let coordinator = Coordinator::try_boot(&mut store, coordinator_epoch)
+            .map_err(CoordinatorServiceError::Durable)?;
+        let admin_token_digest = admin_token
+            .filter(|token| !token.trim().is_empty())
+            .map(Digest::sha256);
+        Ok(Self {
+            coordinator,
+            store,
+            node_descriptors: BTreeMap::new(),
+            enrollment_grants: BTreeMap::new(),
+            task_events: VecDeque::new(),
+            debug_audit_events: VecDeque::new(),
+            debug_epochs: BTreeMap::new(),
+            debug_epoch_runtime: BTreeMap::new(),
+            debug_breakpoints: BTreeMap::new(),
+            debug_commands: BTreeMap::new(),
+            task_assignments: BTreeMap::new(),
+            task_restart_checkpoints: BTreeMap::new(),
+            task_restart_checkpoint_order: VecDeque::new(),
+            main_runtime: main_runtime::CoordinatorMainRuntime::default(),
+            pending_task_launches: VecDeque::new(),
+            task_placements: BTreeMap::new(),
+            active_tasks: BTreeSet::new(),
+            task_cancellations: BTreeSet::new(),
+            task_aborts: BTreeSet::new(),
+            process_cancellations: BTreeSet::new(),
+            process_aborts: BTreeSet::new(),
+            agent_replay_nonces: BTreeMap::new(),
+            node_replay_nonces: BTreeMap::new(),
+            panel_snapshots: BTreeMap::new(),
+            stopped_panels: BTreeSet::new(),
+            panel_event_limits: BTreeMap::new(),
+            artifact_registry: ArtifactRegistry::default(),
+            artifact_reverse_transfers: BTreeMap::new(),
+            artifact_transfer_by_token: BTreeMap::new(),
+            transport: NativeQuicTransport,
+            quota: quota::CoordinatorQuota::new(quota_configuration),
+            admission,
+            #[cfg(test)]
+            server_time_override: None,
+            admin_token_digest,
+            admin_replay_nonces: BTreeMap::new(),
+        })
+    }
+
+    pub fn durable_store_kind(&self) -> &'static str {
+        self.store.kind()
+    }
+
+    fn persist_durable_state(&mut self) -> Result<(), CoordinatorServiceError> {
+        self.coordinator
+            .try_persist(&mut self.store)
+            .map_err(CoordinatorServiceError::Durable)
+    }
+
+    fn current_epoch_seconds(&self) -> Result<u64, CoordinatorServiceError> {
+        #[cfg(test)]
+        if let Some(now) = self.server_time_override {
+            return Ok(now);
+        }
+        SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .map(|duration| duration.as_secs())
+            .map_err(|err| CoordinatorServiceError::Protocol(format!("system clock error: {err}")))
+    }
+
+    fn handle_quota_status(
+        &self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let durable_project = self.coordinator.project(&project).ok_or_else(|| {
+            CoordinatorError::Unauthorized("quota status requires an existing project".to_owned())
+        })?;
+        if durable_project.tenant != tenant {
+            return Err(CoordinatorError::Unauthorized(
+                "quota status project is outside the tenant scope".to_owned(),
+            )
+            .into());
+        }
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        let status = self
+            .quota
+            .project_status(&tenant, &project, now_epoch_seconds);
+        Ok(CoordinatorResponse::QuotaStatus {
+            tenant,
+            project,
+            actor,
+            policy_label: status.policy_label,
+            limits: status.limits,
+            window_seconds: status.window_seconds,
+            usage: status.usage,
+            window_started_epoch_seconds: status.window_started_epoch_seconds,
+        })
+    }
+
+    #[cfg(test)]
+    fn set_server_time(&mut self, now_epoch_seconds: u64) {
+        self.server_time_override = Some(now_epoch_seconds);
+    }
+
+    pub fn issue_cli_session(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        user: UserId,
+        session_secret: &str,
+        expires_at_epoch_seconds: Option<u64>,
+    ) -> Result<crate::CliSessionRecord, CoordinatorServiceError> {
+        if let Some(existing) = self.coordinator.project(&project) {
+            if existing.tenant != tenant {
+                return Err(CoordinatorError::Unauthorized(
+                    "CLI session project belongs to a different tenant".to_owned(),
+                )
+                .into());
+            }
+        } else {
+            self.coordinator
+                .upsert_project(tenant.clone(), project.clone(), "Session project");
+        }
+        self.coordinator
+            .grant_project_debug(tenant.clone(), project.clone(), user.clone());
+        let record = self.coordinator.issue_cli_session(
+            tenant,
+            project,
+            user,
+            session_secret,
+            expires_at_epoch_seconds,
+        );
+        self.persist_durable_state()?;
+        Ok(record)
+    }
+
+    pub fn revoke_cli_session(
+        &mut self,
+        session_secret: &str,
+    ) -> Result<crate::CliSessionRecord, CoordinatorServiceError> {
+        let record = self.coordinator.revoke_cli_session(session_secret)?;
+        self.persist_durable_state()?;
+        Ok(record)
+    }
+}
+
+#[cfg(test)]
+mod tests;

crates/disasmer-coordinator/src/service/admin.rs

@@ -0,0 +1,147 @@
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use disasmer_core::{
+    admin_request_proof_from_token_digest, CredentialKind, Digest, TenantId, UserId,
+};
+
+use crate::CoordinatorError;
+
+use super::{CoordinatorResponse, CoordinatorService, CoordinatorServiceError};
+
+const ADMIN_REQUEST_MAX_CLOCK_SKEW_SECONDS: u64 = 300;
+
+impl CoordinatorService {
+    pub(super) fn handle_admin_status(
+        &mut self,
+        tenant: String,
+        actor_user: String,
+        admin_proof: Digest,
+        admin_nonce: String,
+        issued_at_epoch_seconds: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        self.verify_admin_request(
+            "admin_status",
+            &tenant,
+            &actor_user,
+            &tenant,
+            &admin_proof,
+            &admin_nonce,
+            issued_at_epoch_seconds,
+        )?;
+        let tenant = TenantId::new(tenant);
+        let actor = UserId::new(actor_user);
+        Ok(CoordinatorResponse::AdminStatus {
+            suspended: self.coordinator.tenant_suspended(&tenant),
+            tenant,
+            actor,
+            safe_default: "read_only".to_owned(),
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_suspend_tenant(
+        &mut self,
+        tenant: String,
+        actor_user: String,
+        target_tenant: String,
+        admin_proof: Digest,
+        admin_nonce: String,
+        issued_at_epoch_seconds: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        self.verify_admin_request(
+            "suspend_tenant",
+            &tenant,
+            &actor_user,
+            &target_tenant,
+            &admin_proof,
+            &admin_nonce,
+            issued_at_epoch_seconds,
+        )?;
+        let actor_tenant = TenantId::new(tenant);
+        let actor = UserId::new(actor_user);
+        let target_tenant = TenantId::new(target_tenant);
+        self.coordinator.upsert_tenant(actor_tenant.clone());
+        self.coordinator.upsert_user(
+            actor_tenant,
+            actor.clone(),
+            CredentialKind::CliDeviceSession,
+        );
+        let policy = self
+            .coordinator
+            .suspend_tenant(target_tenant.clone(), actor.clone());
+        self.persist_durable_state()?;
+        Ok(CoordinatorResponse::TenantSuspended {
+            tenant: target_tenant,
+            actor,
+            policy,
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    fn verify_admin_request(
+        &mut self,
+        operation: &str,
+        tenant: &str,
+        actor_user: &str,
+        target_tenant: &str,
+        admin_proof: &Digest,
+        admin_nonce: &str,
+        issued_at_epoch_seconds: u64,
+    ) -> Result<(), CoordinatorServiceError> {
+        let expected = self.admin_token_digest.as_ref().ok_or_else(|| {
+            CoordinatorError::Unauthorized(
+                "self-hosted admin credential is not configured".to_owned(),
+            )
+        })?;
+        if admin_nonce.trim().is_empty() || admin_nonce.len() > 256 {
+            return Err(CoordinatorError::Unauthorized(
+                "admin request nonce is missing or invalid".to_owned(),
+            )
+            .into());
+        }
+        let now = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .map(|duration| duration.as_secs())
+            .unwrap_or_default();
+        if now.abs_diff(issued_at_epoch_seconds) > ADMIN_REQUEST_MAX_CLOCK_SKEW_SECONDS {
+            return Err(CoordinatorError::Unauthorized(
+                "admin request timestamp is outside the allowed 300-second window".to_owned(),
+            )
+            .into());
+        }
+        let expected_proof = admin_request_proof_from_token_digest(
+            expected,
+            operation,
+            tenant,
+            actor_user,
+            target_tenant,
+            admin_nonce,
+            issued_at_epoch_seconds,
+        );
+        if admin_proof != &expected_proof {
+            return Err(CoordinatorError::Unauthorized(
+                "admin request proof is invalid".to_owned(),
+            )
+            .into());
+        }
+        self.admin_replay_nonces.retain(|_, issued_at| {
+            now <= issued_at.saturating_add(ADMIN_REQUEST_MAX_CLOCK_SKEW_SECONDS)
+        });
+        if self.admin_replay_nonces.contains_key(admin_nonce) {
+            return Err(CoordinatorError::Unauthorized(
+                "admin request nonce was already used".to_owned(),
+            )
+            .into());
+        }
+        if self.admin_replay_nonces.len() >= super::MAX_REPLAY_NONCES_PER_AUTHORITY {
+            return Err(CoordinatorError::Unauthorized(
+                "admin request replay window is full; retry after the bounded signature window advances"
+                    .to_owned(),
+            )
+            .into());
+        }
+        self.admin_replay_nonces
+            .insert(admin_nonce.to_owned(), issued_at_epoch_seconds);
+        Ok(())
+    }
+}

crates/disasmer-coordinator/src/service/artifacts.rs

@@ -0,0 +1,649 @@
+use std::io::{Read, Seek, SeekFrom, Write};
+
+use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine as _};
+use disasmer_core::{
+    generate_opaque_token, Actor, ArtifactId, AuthContext, DataPlaneObject, DataPlaneScope, Digest,
+    DownloadPolicy, NodeEndpoint, NodeId, ProjectId, RendezvousRequest, ResourceLimits,
+    ResourceMeter, StorageLocation, TenantId, UserId,
+};
+use sha2::{Digest as _, Sha256};
+
+use crate::CoordinatorError;
+
+use super::{
+    bounded_ttl, ArtifactTransferAssignment, CoordinatorResponse, CoordinatorService,
+    CoordinatorServiceError,
+};
+
+pub(super) const MAX_ARTIFACT_REVERSE_CHUNK_BYTES: u64 = 256 * 1024;
+const MAX_CONCURRENT_ARTIFACT_REVERSE_TRANSFERS: usize = 4;
+
+#[derive(Debug)]
+pub(super) struct ArtifactReverseTransfer {
+    transfer_id: String,
+    token_digest: Digest,
+    tenant: TenantId,
+    project: ProjectId,
+    source_node: NodeId,
+    artifact: ArtifactId,
+    expected_digest: Digest,
+    expected_size_bytes: u64,
+    expires_at_epoch_seconds: u64,
+    spool: tempfile::NamedTempFile,
+    received_bytes: u64,
+    content_hasher: Sha256,
+    delivered_offset: u64,
+    error: Option<String>,
+}
+
+impl CoordinatorService {
+    pub(super) fn handle_create_artifact_download_link(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        max_bytes: u64,
+        ttl_seconds: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let context = user_context(tenant, project, actor_user);
+        let artifact = ArtifactId::new(artifact);
+        let policy = DownloadPolicy { max_bytes };
+        let action = self
+            .artifact_registry
+            .download_action(&context, &artifact, &policy)?;
+        self.ensure_download_source_connectivity(&action.source)?;
+        let downloadable_size = self
+            .artifact_registry
+            .downloadable_size(&context, &artifact, &policy)?;
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.quota.can_charge_download(
+            &context.tenant,
+            &context.project,
+            downloadable_size,
+            now_epoch_seconds,
+        )?;
+        let token_nonce = generate_opaque_token("artifact_download")
+            .map_err(CoordinatorServiceError::Protocol)?;
+        let ttl_seconds = bounded_ttl(
+            ttl_seconds,
+            self.admission.max_artifact_download_ttl_seconds,
+        );
+        let link = self.artifact_registry.create_download_link(
+            &context,
+            &artifact,
+            &policy,
+            &token_nonce,
+            now_epoch_seconds,
+            ttl_seconds,
+        )?;
+        Ok(CoordinatorResponse::ArtifactDownloadLink { link })
+    }
+
+    pub(super) fn handle_open_artifact_download_stream(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        max_bytes: u64,
+        token_digest: Digest,
+        chunk_bytes: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let context = user_context(tenant, project, actor_user);
+        let artifact = ArtifactId::new(artifact);
+        let policy = DownloadPolicy { max_bytes };
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.artifact_registry
+            .expire_download_links(now_epoch_seconds);
+        let downloadable_size = self
+            .artifact_registry
+            .downloadable_size(&context, &artifact, &policy)?;
+        let validation_limits = ResourceLimits::unlimited();
+        let mut validation_meter = ResourceMeter::default();
+        let mut stream = self.artifact_registry.open_download_stream(
+            disasmer_core::DownloadStreamRequest {
+                context: &context,
+                artifact: &artifact,
+                policy: &policy,
+                presented_token_digest: &token_digest,
+                now_epoch_seconds,
+                limits: &validation_limits,
+            },
+            &mut validation_meter,
+        )?;
+        self.ensure_download_source_connectivity(&stream.link.source)?;
+        self.expire_artifact_reverse_transfers(now_epoch_seconds);
+
+        if let Some(transfer_id) = self.artifact_transfer_by_token.get(&token_digest).cloned() {
+            if let Some(message) = self
+                .artifact_reverse_transfers
+                .get(&transfer_id)
+                .and_then(|transfer| transfer.error.clone())
+            {
+                self.artifact_reverse_transfers.remove(&transfer_id);
+                self.artifact_transfer_by_token.remove(&token_digest);
+                return Err(CoordinatorServiceError::Protocol(format!(
+                    "retaining node could not stream artifact: {message}"
+                )));
+            }
+            let (content_offset, content, end, complete) = {
+                let transfer = self
+                    .artifact_reverse_transfers
+                    .get_mut(&transfer_id)
+                    .ok_or_else(|| {
+                        CoordinatorServiceError::Protocol(
+                            "artifact reverse transfer index is inconsistent".to_owned(),
+                        )
+                    })?;
+                if transfer.tenant != context.tenant
+                    || transfer.project != context.project
+                    || transfer.artifact != artifact
+                    || transfer.token_digest != token_digest
+                {
+                    return Err(disasmer_core::DownloadError::InvalidToken.into());
+                }
+                if transfer.received_bytes != transfer.expected_size_bytes {
+                    return Ok(CoordinatorResponse::ArtifactDownloadStream {
+                        link: stream.link,
+                        streamed_bytes: 0,
+                        charged_download_bytes: self.quota.used_download_bytes(
+                            &context.tenant,
+                            &context.project,
+                            now_epoch_seconds,
+                        ),
+                        content_bytes_available: false,
+                        content_offset: None,
+                        content_eof: false,
+                        content_base64: None,
+                        content_source: Some("retaining_node_reverse_stream_pending".to_owned()),
+                    });
+                }
+                if chunk_bytes == 0 && downloadable_size != 0 {
+                    return Err(CoordinatorServiceError::Protocol(
+                        "artifact download chunk_bytes must be greater than zero".to_owned(),
+                    ));
+                }
+                let requested = chunk_bytes
+                    .min(downloadable_size)
+                    .min(MAX_ARTIFACT_REVERSE_CHUNK_BYTES);
+                let start = transfer.delivered_offset;
+                let end = start.saturating_add(requested).min(transfer.received_bytes);
+                let length = usize::try_from(end.saturating_sub(start)).map_err(|_| {
+                    CoordinatorServiceError::Protocol(
+                        "artifact download chunk length does not fit memory bounds".to_owned(),
+                    )
+                })?;
+                let mut content = vec![0_u8; length];
+                let mut spool = transfer.spool.reopen().map_err(|error| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "open bounded artifact transfer spool: {error}"
+                    ))
+                })?;
+                spool.seek(SeekFrom::Start(start)).map_err(|error| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "seek bounded artifact transfer spool: {error}"
+                    ))
+                })?;
+                spool.read_exact(&mut content).map_err(|error| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "read bounded artifact transfer spool: {error}"
+                    ))
+                })?;
+                (start, content, end, end == transfer.received_bytes)
+            };
+            let streamed_bytes = content.len() as u64;
+            self.artifact_registry.stream_download_chunk(
+                &mut stream,
+                &validation_limits,
+                &mut validation_meter,
+                streamed_bytes,
+            )?;
+            let charged_download_bytes = self.quota.charge_download(
+                &context.tenant,
+                &context.project,
+                streamed_bytes,
+                now_epoch_seconds,
+            )?;
+            if let Some(transfer) = self.artifact_reverse_transfers.get_mut(&transfer_id) {
+                transfer.delivered_offset = end;
+            }
+            if complete {
+                self.artifact_reverse_transfers.remove(&transfer_id);
+                self.artifact_transfer_by_token.remove(&token_digest);
+            }
+            return Ok(CoordinatorResponse::ArtifactDownloadStream {
+                link: stream.link,
+                streamed_bytes,
+                charged_download_bytes,
+                content_bytes_available: true,
+                content_offset: Some(content_offset),
+                content_eof: complete,
+                content_base64: Some(BASE64_STANDARD.encode(content)),
+                content_source: Some("retaining_node_reverse_stream".to_owned()),
+            });
+        }
+
+        if self.artifact_reverse_transfers.len() >= MAX_CONCURRENT_ARTIFACT_REVERSE_TRANSFERS {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "artifact reverse transfer concurrency limit of {MAX_CONCURRENT_ARTIFACT_REVERSE_TRANSFERS} reached"
+            )));
+        }
+        let StorageLocation::RetainedNode(source_node) = &stream.link.source else {
+            return Err(disasmer_core::DownloadError::Unavailable.into());
+        };
+        let metadata = self
+            .artifact_registry
+            .metadata(&artifact)
+            .ok_or(disasmer_core::DownloadError::NotFound)?;
+        let transfer_id = generate_opaque_token("artifact_transfer")
+            .map_err(CoordinatorServiceError::Protocol)?;
+        let spool = tempfile::Builder::new()
+            .prefix("disasmer-artifact-transfer-")
+            .tempfile()
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "create bounded artifact transfer spool: {error}"
+                ))
+            })?;
+        let transfer = ArtifactReverseTransfer {
+            transfer_id: transfer_id.clone(),
+            token_digest: token_digest.clone(),
+            tenant: context.tenant.clone(),
+            project: context.project.clone(),
+            source_node: source_node.clone(),
+            artifact,
+            expected_digest: metadata.digest.clone(),
+            expected_size_bytes: metadata.size,
+            expires_at_epoch_seconds: stream.link.expires_at_epoch_seconds,
+            spool,
+            received_bytes: 0,
+            content_hasher: Sha256::new(),
+            delivered_offset: 0,
+            error: None,
+        };
+        self.artifact_reverse_transfers
+            .insert(transfer_id.clone(), transfer);
+        self.artifact_transfer_by_token
+            .insert(token_digest, transfer_id);
+        Ok(CoordinatorResponse::ArtifactDownloadStream {
+            link: stream.link,
+            streamed_bytes: 0,
+            charged_download_bytes: self.quota.used_download_bytes(
+                &context.tenant,
+                &context.project,
+                now_epoch_seconds,
+            ),
+            content_bytes_available: false,
+            content_offset: None,
+            content_eof: false,
+            content_base64: None,
+            content_source: Some("retaining_node_reverse_stream_pending".to_owned()),
+        })
+    }
+
+    pub(super) fn handle_revoke_artifact_download_link(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        token_digest: Digest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let context = user_context(tenant, project, actor_user);
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.artifact_registry
+            .expire_download_links(now_epoch_seconds);
+        let link = self.artifact_registry.revoke_download_link(
+            &context,
+            &ArtifactId::new(artifact),
+            &token_digest,
+        )?;
+        if let Some(transfer_id) = self.artifact_transfer_by_token.remove(&token_digest) {
+            self.artifact_reverse_transfers.remove(&transfer_id);
+        }
+        Ok(CoordinatorResponse::ArtifactDownloadLinkRevoked { link })
+    }
+
+    pub(super) fn handle_export_artifact_to_node(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        receiver_node: String,
+        direct_connectivity: bool,
+        failure_reason: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let context = user_context(tenant, project, actor_user);
+        let artifact = ArtifactId::new(artifact);
+        let receiver_node = NodeId::new(receiver_node);
+        let action = self.artifact_registry.download_action(
+            &context,
+            &artifact,
+            &DownloadPolicy {
+                max_bytes: u64::MAX,
+            },
+        )?;
+        let StorageLocation::RetainedNode(source_node) = action.source else {
+            return Err(disasmer_core::DownloadError::Unavailable.into());
+        };
+        let metadata = self
+            .artifact_registry
+            .metadata(&artifact)
+            .ok_or(disasmer_core::DownloadError::NotFound)?;
+        let source = self.export_endpoint(&source_node, &context.tenant, &context.project)?;
+        let destination =
+            self.export_endpoint(&receiver_node, &context.tenant, &context.project)?;
+        let plan = self.transport.plan_authenticated_direct_bulk_transfer(
+            RendezvousRequest {
+                scope: DataPlaneScope {
+                    tenant: context.tenant.clone(),
+                    project: context.project.clone(),
+                    process: metadata.process.clone(),
+                    object: DataPlaneObject::Artifact(artifact.clone()),
+                    authorization_subject: format!(
+                        "artifact-export:{}-to-{}",
+                        source_node, receiver_node
+                    ),
+                },
+                source,
+                destination,
+            },
+            direct_connectivity,
+            failure_reason,
+        )?;
+        Ok(CoordinatorResponse::ArtifactExportPlan {
+            plan,
+            source_node,
+            receiver_node,
+            artifact_size_bytes: metadata.size,
+        })
+    }
+
+    pub(super) fn handle_poll_artifact_transfer(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        self.authorize_artifact_transfer_node(&tenant, &project, &node)?;
+        let transfer = self.artifact_reverse_transfers.values().find(|transfer| {
+            transfer.tenant == tenant
+                && transfer.project == project
+                && transfer.source_node == node
+                && transfer.error.is_none()
+                && transfer.received_bytes < transfer.expected_size_bytes
+        });
+        Ok(CoordinatorResponse::ArtifactTransferAssignment {
+            transfer: transfer.map(|transfer| ArtifactTransferAssignment {
+                transfer_id: transfer.transfer_id.clone(),
+                artifact: transfer.artifact.clone(),
+                expected_digest: transfer.expected_digest.clone(),
+                expected_size_bytes: transfer.expected_size_bytes,
+                offset: transfer.received_bytes,
+                max_chunk_bytes: MAX_ARTIFACT_REVERSE_CHUNK_BYTES,
+            }),
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_upload_artifact_transfer_chunk(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        transfer_id: String,
+        artifact: String,
+        offset: u64,
+        content_base64: String,
+        chunk_digest: Digest,
+        eof: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        self.authorize_artifact_transfer_node(&tenant, &project, &node)?;
+        let transfer = self
+            .artifact_reverse_transfers
+            .get_mut(&transfer_id)
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol("unknown artifact reverse transfer".to_owned())
+            })?;
+        if transfer.tenant != tenant
+            || transfer.project != project
+            || transfer.source_node != node
+            || transfer.artifact.as_str() != artifact
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "artifact reverse transfer is outside the signed node scope".to_owned(),
+            )
+            .into());
+        }
+        if transfer.received_bytes != offset {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "artifact reverse transfer expected offset {}, received {offset}",
+                transfer.received_bytes
+            )));
+        }
+        let content = BASE64_STANDARD.decode(content_base64).map_err(|error| {
+            CoordinatorServiceError::Protocol(format!(
+                "artifact reverse transfer chunk is not valid base64: {error}"
+            ))
+        })?;
+        if content.len() as u64 > MAX_ARTIFACT_REVERSE_CHUNK_BYTES {
+            return Err(CoordinatorServiceError::Protocol(
+                "artifact reverse transfer chunk exceeds the control-plane limit".to_owned(),
+            ));
+        }
+        if Digest::sha256(&content) != chunk_digest {
+            return Err(CoordinatorServiceError::Protocol(
+                "artifact reverse transfer chunk digest mismatch".to_owned(),
+            ));
+        }
+        let next_offset = offset.saturating_add(content.len() as u64);
+        if next_offset > transfer.expected_size_bytes {
+            return Err(CoordinatorServiceError::Protocol(
+                "artifact reverse transfer exceeds retained metadata size".to_owned(),
+            ));
+        }
+        let complete = next_offset == transfer.expected_size_bytes;
+        if eof != complete {
+            return Err(CoordinatorServiceError::Protocol(
+                "artifact reverse transfer EOF does not match retained metadata size".to_owned(),
+            ));
+        }
+        transfer
+            .spool
+            .as_file_mut()
+            .seek(SeekFrom::Start(offset))
+            .and_then(|_| transfer.spool.as_file_mut().write_all(&content))
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "write bounded artifact transfer spool: {error}"
+                ))
+            })?;
+        transfer.content_hasher.update(&content);
+        transfer.received_bytes = next_offset;
+        if complete {
+            transfer.spool.as_file().sync_all().map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "sync bounded artifact transfer spool: {error}"
+                ))
+            })?;
+            let digest_hex = format!("{:x}", transfer.content_hasher.clone().finalize());
+            let digest =
+                Digest::from_sha256_hex(&digest_hex).map_err(CoordinatorServiceError::Protocol)?;
+            if digest != transfer.expected_digest {
+                transfer.spool.as_file_mut().set_len(0).map_err(|error| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "reset invalid artifact transfer spool: {error}"
+                    ))
+                })?;
+                transfer.received_bytes = 0;
+                transfer.content_hasher = Sha256::new();
+                return Err(CoordinatorServiceError::Protocol(
+                    "artifact reverse transfer content digest mismatch".to_owned(),
+                ));
+            }
+        }
+        Ok(CoordinatorResponse::ArtifactTransferChunkAccepted {
+            transfer_id,
+            next_offset,
+            complete,
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_fail_artifact_transfer(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        transfer_id: String,
+        artifact: String,
+        message: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        self.authorize_artifact_transfer_node(&tenant, &project, &node)?;
+        let transfer = self
+            .artifact_reverse_transfers
+            .get_mut(&transfer_id)
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol("unknown artifact reverse transfer".to_owned())
+            })?;
+        if transfer.tenant != tenant
+            || transfer.project != project
+            || transfer.source_node != node
+            || transfer.artifact.as_str() != artifact
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "artifact reverse transfer failure is outside the signed node scope".to_owned(),
+            )
+            .into());
+        }
+        let message = message.trim();
+        transfer.error = Some(if message.is_empty() {
+            "retained artifact bytes are unavailable".to_owned()
+        } else {
+            message.chars().take(1024).collect()
+        });
+        Ok(CoordinatorResponse::ArtifactTransferFailed { transfer_id })
+    }
+
+    fn authorize_artifact_transfer_node(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        node: &NodeId,
+    ) -> Result<(), CoordinatorServiceError> {
+        let identity = self
+            .coordinator
+            .node_identity(node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        if &identity.tenant != tenant || &identity.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "artifact reverse transfer node is outside its enrolled tenant/project scope"
+                    .to_owned(),
+            )
+            .into());
+        }
+        Ok(())
+    }
+
+    fn expire_artifact_reverse_transfers(&mut self, now_epoch_seconds: u64) {
+        let expired = self
+            .artifact_reverse_transfers
+            .iter()
+            .filter(|(_, transfer)| transfer.expires_at_epoch_seconds < now_epoch_seconds)
+            .map(|(id, transfer)| (id.clone(), transfer.token_digest.clone()))
+            .collect::<Vec<_>>();
+        for (id, token) in expired {
+            self.artifact_reverse_transfers.remove(&id);
+            self.artifact_transfer_by_token.remove(&token);
+        }
+    }
+
+    fn export_endpoint(
+        &self,
+        node: &NodeId,
+        tenant: &TenantId,
+        project: &ProjectId,
+    ) -> Result<NodeEndpoint, CoordinatorServiceError> {
+        let identity = self
+            .coordinator
+            .node_identity(node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        if &identity.tenant != tenant || &identity.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "artifact export node is outside the tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        let descriptor = self.node_descriptors.get(node).ok_or_else(|| {
+            disasmer_core::DownloadError::DirectConnectivityUnavailable(format!(
+                "node {node} has not reported export connectivity"
+            ))
+        })?;
+        if descriptor.tenant != *tenant || descriptor.project != *project {
+            return Err(CoordinatorError::Unauthorized(
+                "artifact export node descriptor is outside the tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        if !descriptor.online {
+            return Err(
+                disasmer_core::DownloadError::DirectConnectivityUnavailable(format!(
+                    "node {node} is offline for artifact export"
+                ))
+                .into(),
+            );
+        }
+        if !descriptor.direct_connectivity {
+            return Err(
+                disasmer_core::DownloadError::DirectConnectivityUnavailable(format!(
+                    "direct connectivity unavailable to node {node} for artifact export"
+                ))
+                .into(),
+            );
+        }
+        Ok(NodeEndpoint {
+            node: node.clone(),
+            advertised_addr: format!("{node}.mesh.invalid:4433"),
+            public_key_fingerprint: Digest::sha256(&identity.public_key),
+        })
+    }
+
+    fn ensure_download_source_connectivity(
+        &self,
+        source: &StorageLocation,
+    ) -> Result<(), disasmer_core::DownloadError> {
+        let StorageLocation::RetainedNode(node) = source else {
+            return Ok(());
+        };
+        let descriptor = self.node_descriptors.get(node).ok_or_else(|| {
+            disasmer_core::DownloadError::DirectConnectivityUnavailable(format!(
+                "retaining node {node} has not reported online status for artifact download"
+            ))
+        })?;
+        if !descriptor.online {
+            return Err(disasmer_core::DownloadError::DirectConnectivityUnavailable(
+                format!("retaining node {node} is offline for artifact download"),
+            ));
+        }
+        Ok(())
+    }
+}
+
+fn user_context(tenant: String, project: String, actor_user: String) -> AuthContext {
+    AuthContext {
+        tenant: TenantId::new(tenant),
+        project: ProjectId::new(project),
+        actor: Actor::User(UserId::new(actor_user)),
+    }
+}

crates/disasmer-coordinator/src/service/authenticated.rs

@@ -0,0 +1,139 @@
+use disasmer_core::{AuthContext, UserId};
+
+use super::{
+    AuthenticatedCoordinatorRequest, CoordinatorRequest, CoordinatorResponse, CoordinatorService,
+    CoordinatorServiceError,
+};
+
+impl CoordinatorService {
+    pub(super) fn handle_authenticated_agent_key_request(
+        &mut self,
+        context: &AuthContext,
+        actor: &UserId,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let request = match request {
+            AuthenticatedCoordinatorRequest::RegisterAgentPublicKey { agent, public_key } => {
+                CoordinatorRequest::RegisterAgentPublicKey {
+                    tenant: context.tenant.as_str().to_owned(),
+                    project: context.project.as_str().to_owned(),
+                    user: actor.as_str().to_owned(),
+                    agent,
+                    public_key,
+                }
+            }
+            AuthenticatedCoordinatorRequest::ListAgentPublicKeys => {
+                CoordinatorRequest::ListAgentPublicKeys {
+                    tenant: context.tenant.as_str().to_owned(),
+                    project: context.project.as_str().to_owned(),
+                    user: actor.as_str().to_owned(),
+                }
+            }
+            AuthenticatedCoordinatorRequest::RotateAgentPublicKey { agent, public_key } => {
+                CoordinatorRequest::RotateAgentPublicKey {
+                    tenant: context.tenant.as_str().to_owned(),
+                    project: context.project.as_str().to_owned(),
+                    user: actor.as_str().to_owned(),
+                    agent,
+                    public_key,
+                }
+            }
+            AuthenticatedCoordinatorRequest::RevokeAgentPublicKey { agent } => {
+                CoordinatorRequest::RevokeAgentPublicKey {
+                    tenant: context.tenant.as_str().to_owned(),
+                    project: context.project.as_str().to_owned(),
+                    user: actor.as_str().to_owned(),
+                    agent,
+                }
+            }
+            _ => unreachable!("caller filters authenticated agent key operations"),
+        };
+        self.handle_request(request)
+    }
+
+    pub(super) fn handle_authenticated_launch_task(
+        &mut self,
+        context: &AuthContext,
+        actor: &UserId,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let AuthenticatedCoordinatorRequest::LaunchTask {
+            task_spec,
+            wait_for_node,
+            artifact_path,
+            wasm_module_base64,
+        } = request
+        else {
+            unreachable!("caller filters authenticated task launches");
+        };
+        self.handle_launch_task(
+            context.tenant.as_str().to_owned(),
+            context.project.as_str().to_owned(),
+            Some(actor.as_str().to_owned()),
+            None,
+            None,
+            None,
+            None,
+            *task_spec,
+            wait_for_node,
+            artifact_path,
+            wasm_module_base64,
+        )
+    }
+
+    pub(super) fn handle_authenticated_schedule_task(
+        &mut self,
+        context: &AuthContext,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let AuthenticatedCoordinatorRequest::ScheduleTask {
+            environment,
+            environment_digest,
+            required_capabilities,
+            dependency_cache,
+            source_snapshot,
+            required_artifacts,
+            prefer_node,
+        } = request
+        else {
+            unreachable!("caller filters authenticated task scheduling");
+        };
+        self.handle_schedule_task(
+            context.tenant.as_str().to_owned(),
+            context.project.as_str().to_owned(),
+            environment,
+            environment_digest,
+            required_capabilities,
+            dependency_cache,
+            source_snapshot,
+            required_artifacts,
+            prefer_node,
+        )
+    }
+
+    pub(super) fn handle_authenticated_artifact_export(
+        &mut self,
+        context: &AuthContext,
+        actor: &UserId,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let AuthenticatedCoordinatorRequest::ExportArtifactToNode {
+            artifact,
+            receiver_node,
+            direct_connectivity,
+            failure_reason,
+        } = request
+        else {
+            unreachable!("caller filters authenticated artifact exports");
+        };
+        self.handle_export_artifact_to_node(
+            context.tenant.as_str().to_owned(),
+            context.project.as_str().to_owned(),
+            actor.as_str().to_owned(),
+            artifact,
+            receiver_node,
+            direct_connectivity,
+            failure_reason,
+        )
+    }
+}

crates/disasmer-coordinator/src/service/authorization.rs

@@ -0,0 +1,198 @@
+use disasmer_core::{Actor, AuthContext, UserId};
+
+use crate::CoordinatorError;
+
+use super::{AuthenticatedCoordinatorRequest, CoordinatorServiceError};
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(super) enum PublicUserOperation {
+    AuthStatus,
+    RevokeCliSession,
+    CreateProject,
+    SelectProject,
+    ListProjects,
+    RegisterAgentPublicKey,
+    ListAgentPublicKeys,
+    RotateAgentPublicKey,
+    RevokeAgentPublicKey,
+    CreateNodeEnrollmentGrant,
+    ListNodeDescriptors,
+    RevokeNodeCredential,
+    StartProcess,
+    ScheduleTask,
+    LaunchTask,
+    CancelProcess,
+    AbortProcess,
+    ListProcesses,
+    QuotaStatus,
+    RestartTask,
+    DebugAttach,
+    SetDebugBreakpoints,
+    InspectDebugBreakpoints,
+    CreateDebugEpoch,
+    ResumeDebugEpoch,
+    InspectDebugEpoch,
+    ListTaskEvents,
+    JoinTask,
+    CreateArtifactDownloadLink,
+    OpenArtifactDownloadStream,
+    RevokeArtifactDownloadLink,
+    ExportArtifactToNode,
+}
+
+impl PublicUserOperation {
+    pub(super) fn as_str(self) -> &'static str {
+        match self {
+            Self::AuthStatus => "auth_status",
+            Self::RevokeCliSession => "revoke_cli_session",
+            Self::CreateProject => "create_project",
+            Self::SelectProject => "select_project",
+            Self::ListProjects => "list_projects",
+            Self::RegisterAgentPublicKey => "register_agent_public_key",
+            Self::ListAgentPublicKeys => "list_agent_public_keys",
+            Self::RotateAgentPublicKey => "rotate_agent_public_key",
+            Self::RevokeAgentPublicKey => "revoke_agent_public_key",
+            Self::CreateNodeEnrollmentGrant => "create_node_enrollment_grant",
+            Self::ListNodeDescriptors => "list_node_descriptors",
+            Self::RevokeNodeCredential => "revoke_node_credential",
+            Self::StartProcess => "start_process",
+            Self::ScheduleTask => "schedule_task",
+            Self::LaunchTask => "launch_task",
+            Self::CancelProcess => "cancel_process",
+            Self::AbortProcess => "abort_process",
+            Self::ListProcesses => "list_processes",
+            Self::QuotaStatus => "quota_status",
+            Self::RestartTask => "restart_task",
+            Self::DebugAttach => "debug_attach",
+            Self::SetDebugBreakpoints => "set_debug_breakpoints",
+            Self::InspectDebugBreakpoints => "inspect_debug_breakpoints",
+            Self::CreateDebugEpoch => "create_debug_epoch",
+            Self::ResumeDebugEpoch => "resume_debug_epoch",
+            Self::InspectDebugEpoch => "inspect_debug_epoch",
+            Self::ListTaskEvents => "list_task_events",
+            Self::JoinTask => "join_task",
+            Self::CreateArtifactDownloadLink => "create_artifact_download_link",
+            Self::OpenArtifactDownloadStream => "open_artifact_download_stream",
+            Self::RevokeArtifactDownloadLink => "revoke_artifact_download_link",
+            Self::ExportArtifactToNode => "export_artifact_to_node",
+        }
+    }
+}
+
+impl From<&AuthenticatedCoordinatorRequest> for PublicUserOperation {
+    fn from(request: &AuthenticatedCoordinatorRequest) -> Self {
+        match request {
+            AuthenticatedCoordinatorRequest::AuthStatus => Self::AuthStatus,
+            AuthenticatedCoordinatorRequest::RevokeCliSession => Self::RevokeCliSession,
+            AuthenticatedCoordinatorRequest::CreateProject { .. } => Self::CreateProject,
+            AuthenticatedCoordinatorRequest::SelectProject { .. } => Self::SelectProject,
+            AuthenticatedCoordinatorRequest::ListProjects => Self::ListProjects,
+            AuthenticatedCoordinatorRequest::RegisterAgentPublicKey { .. } => {
+                Self::RegisterAgentPublicKey
+            }
+            AuthenticatedCoordinatorRequest::ListAgentPublicKeys => Self::ListAgentPublicKeys,
+            AuthenticatedCoordinatorRequest::RotateAgentPublicKey { .. } => {
+                Self::RotateAgentPublicKey
+            }
+            AuthenticatedCoordinatorRequest::RevokeAgentPublicKey { .. } => {
+                Self::RevokeAgentPublicKey
+            }
+            AuthenticatedCoordinatorRequest::CreateNodeEnrollmentGrant { .. } => {
+                Self::CreateNodeEnrollmentGrant
+            }
+            AuthenticatedCoordinatorRequest::ListNodeDescriptors => Self::ListNodeDescriptors,
+            AuthenticatedCoordinatorRequest::RevokeNodeCredential { .. } => {
+                Self::RevokeNodeCredential
+            }
+            AuthenticatedCoordinatorRequest::StartProcess { .. } => Self::StartProcess,
+            AuthenticatedCoordinatorRequest::ScheduleTask { .. } => Self::ScheduleTask,
+            AuthenticatedCoordinatorRequest::LaunchTask { .. } => Self::LaunchTask,
+            AuthenticatedCoordinatorRequest::CancelProcess { .. } => Self::CancelProcess,
+            AuthenticatedCoordinatorRequest::AbortProcess { .. } => Self::AbortProcess,
+            AuthenticatedCoordinatorRequest::ListProcesses => Self::ListProcesses,
+            AuthenticatedCoordinatorRequest::QuotaStatus => Self::QuotaStatus,
+            AuthenticatedCoordinatorRequest::RestartTask { .. } => Self::RestartTask,
+            AuthenticatedCoordinatorRequest::DebugAttach { .. } => Self::DebugAttach,
+            AuthenticatedCoordinatorRequest::SetDebugBreakpoints { .. } => {
+                Self::SetDebugBreakpoints
+            }
+            AuthenticatedCoordinatorRequest::InspectDebugBreakpoints { .. } => {
+                Self::InspectDebugBreakpoints
+            }
+            AuthenticatedCoordinatorRequest::CreateDebugEpoch { .. } => Self::CreateDebugEpoch,
+            AuthenticatedCoordinatorRequest::ResumeDebugEpoch { .. } => Self::ResumeDebugEpoch,
+            AuthenticatedCoordinatorRequest::InspectDebugEpoch { .. } => Self::InspectDebugEpoch,
+            AuthenticatedCoordinatorRequest::ListTaskEvents { .. } => Self::ListTaskEvents,
+            AuthenticatedCoordinatorRequest::JoinTask { .. } => Self::JoinTask,
+            AuthenticatedCoordinatorRequest::CreateArtifactDownloadLink { .. } => {
+                Self::CreateArtifactDownloadLink
+            }
+            AuthenticatedCoordinatorRequest::OpenArtifactDownloadStream { .. } => {
+                Self::OpenArtifactDownloadStream
+            }
+            AuthenticatedCoordinatorRequest::RevokeArtifactDownloadLink { .. } => {
+                Self::RevokeArtifactDownloadLink
+            }
+            AuthenticatedCoordinatorRequest::ExportArtifactToNode { .. } => {
+                Self::ExportArtifactToNode
+            }
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(super) struct AuthorizedPublicUser {
+    pub(super) actor: UserId,
+    pub(super) operation: PublicUserOperation,
+}
+
+pub(super) fn authorize_authenticated_user_operation(
+    context: &AuthContext,
+    request: &AuthenticatedCoordinatorRequest,
+) -> Result<AuthorizedPublicUser, CoordinatorServiceError> {
+    let operation = PublicUserOperation::from(request);
+    let actor = match &context.actor {
+        Actor::User(user) => user.clone(),
+        _ => {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "authenticated {} request requires a user CLI session",
+                operation.as_str()
+            ))
+            .into());
+        }
+    };
+    Ok(AuthorizedPublicUser { actor, operation })
+}
+
+#[cfg(test)]
+mod tests {
+    use disasmer_core::{AgentId, ProjectId, TenantId};
+
+    use super::*;
+
+    #[test]
+    fn authenticated_public_authorization_requires_user_context_and_names_operation() {
+        let request = AuthenticatedCoordinatorRequest::DebugAttach {
+            process: "vp".to_owned(),
+        };
+        let agent_context = AuthContext {
+            tenant: TenantId::from("tenant"),
+            project: ProjectId::from("project"),
+            actor: Actor::Agent(AgentId::from("agent-ci")),
+        };
+
+        let denied = authorize_authenticated_user_operation(&agent_context, &request).unwrap_err();
+        assert!(denied
+            .to_string()
+            .contains("authenticated debug_attach request requires a user CLI session"));
+
+        let user_context = AuthContext {
+            tenant: TenantId::from("tenant"),
+            project: ProjectId::from("project"),
+            actor: Actor::User(UserId::from("user")),
+        };
+        let authorized = authorize_authenticated_user_operation(&user_context, &request).unwrap();
+        assert_eq!(authorized.actor, UserId::from("user"));
+        assert_eq!(authorized.operation, PublicUserOperation::DebugAttach);
+    }
+}

crates/disasmer-coordinator/src/service/debug.rs

@@ -0,0 +1,995 @@
+use std::collections::{BTreeMap, BTreeSet};
+
+use disasmer_core::{
+    Actor, NodeId, ProcessId, ProjectId, TaskInstanceId, TenantId, UserId,
+    WasmHostDebugProbeResult, WASM_TASK_ABI_VERSION,
+};
+
+use crate::{CoordinatorError, CoordinatorServiceError};
+
+use super::keys::{process_control_key, task_control_key, task_restart_key, TaskControlKey};
+use super::{
+    CoordinatorResponse, CoordinatorService, DebugAcknowledgementState, DebugAuditEvent,
+    DebugParticipantAcknowledgement, TaskCancellationTarget, DEBUG_CONTROL_READ_BYTES,
+};
+
+mod validation;
+use validation::{runtime_all_in_state, validate_debug_snapshot, validate_probe_symbols};
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(super) struct DebugPendingCommand {
+    pub(super) epoch: u64,
+    pub(super) command: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(super) struct DebugEpochRuntime {
+    pub(super) epoch: u64,
+    pub(super) command: String,
+    pub(super) expected: BTreeSet<TaskControlKey>,
+    pub(super) acknowledgements: BTreeMap<TaskControlKey, DebugParticipantAcknowledgement>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(super) struct DebugBreakpointPlan {
+    pub(super) actor: UserId,
+    pub(super) probe_symbols: BTreeSet<String>,
+    pub(super) hit_epoch: Option<u64>,
+    pub(super) hit_task: Option<TaskInstanceId>,
+    pub(super) hit_probe_symbol: Option<String>,
+}
+
+impl CoordinatorService {
+    pub(super) fn handle_debug_attach(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        let audit_event = self.record_debug_audit_event(
+            tenant,
+            project,
+            process.clone(),
+            None,
+            actor.clone(),
+            "debug_attach",
+            authorization.allowed,
+            authorization.reason.clone(),
+        )?;
+        Ok(CoordinatorResponse::DebugAttach {
+            process,
+            actor,
+            authorization,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+
+    pub(super) fn handle_set_debug_breakpoints(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        probe_symbols: Vec<String>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let probe_symbols = validate_probe_symbols(probe_symbols)?;
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        let audit_event = self.record_debug_audit_event(
+            tenant.clone(),
+            project.clone(),
+            process.clone(),
+            None,
+            actor.clone(),
+            "set_debug_breakpoints",
+            authorization.allowed,
+            authorization.reason.clone(),
+        )?;
+        if !authorization.allowed {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "set_debug_breakpoints denied: {}",
+                authorization.reason
+            ))
+            .into());
+        }
+        self.debug_breakpoints.insert(
+            process_control_key(&tenant, &project, &process),
+            DebugBreakpointPlan {
+                actor: actor.clone(),
+                probe_symbols: probe_symbols.iter().cloned().collect(),
+                hit_epoch: None,
+                hit_task: None,
+                hit_probe_symbol: None,
+            },
+        );
+        Ok(CoordinatorResponse::DebugBreakpoints {
+            process,
+            actor,
+            probe_symbols,
+            hit_epoch: None,
+            hit_task: None,
+            hit_probe_symbol: None,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+
+    pub(super) fn handle_inspect_debug_breakpoints(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        let plan = self
+            .debug_breakpoints
+            .get(&process_control_key(&tenant, &project, &process))
+            .cloned()
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol(format!(
+                    "no debug breakpoints are configured for {process}"
+                ))
+            })?;
+        if plan.actor != actor {
+            return Err(CoordinatorError::Unauthorized(
+                "debug breakpoint inspection belongs to another authenticated user".to_owned(),
+            )
+            .into());
+        }
+        let audit_event = self.record_debug_audit_event(
+            tenant,
+            project,
+            process.clone(),
+            plan.hit_task.clone(),
+            actor.clone(),
+            "inspect_debug_breakpoints",
+            authorization.allowed,
+            authorization.reason.clone(),
+        )?;
+        if !authorization.allowed {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "inspect_debug_breakpoints denied: {}",
+                authorization.reason
+            ))
+            .into());
+        }
+        Ok(CoordinatorResponse::DebugBreakpoints {
+            process,
+            actor,
+            probe_symbols: plan.probe_symbols.into_iter().collect(),
+            hit_epoch: plan.hit_epoch,
+            hit_task: plan.hit_task,
+            hit_probe_symbol: plan.hit_probe_symbol,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+
+    pub(super) fn handle_create_debug_epoch(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        stopped_task: String,
+        reason: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        self.handle_debug_epoch_control(
+            tenant,
+            project,
+            actor_user,
+            process,
+            Some(stopped_task),
+            None,
+            "create_debug_epoch",
+            "freeze",
+            true,
+            reason,
+        )
+    }
+
+    pub(super) fn handle_resume_debug_epoch(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        epoch: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        self.handle_debug_epoch_control(
+            tenant,
+            project,
+            actor_user,
+            process,
+            None,
+            Some(epoch),
+            "resume_debug_epoch",
+            "resume",
+            false,
+            "continue requested by debugger",
+        )
+    }
+
+    pub(super) fn handle_poll_debug_command(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.coordinator
+            .authorize_node_for_process(&node, &tenant, &project, &process)?;
+        let pending = self
+            .debug_commands
+            .remove(&task_control_key(&tenant, &project, &process, &node, &task));
+        Ok(CoordinatorResponse::DebugCommand {
+            process,
+            task,
+            epoch: pending.as_ref().map(|command| command.epoch),
+            command: pending.map(|command| command.command),
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_report_debug_state(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        epoch: u64,
+        state: DebugAcknowledgementState,
+        stack_frames: Vec<String>,
+        local_values: Vec<(String, String)>,
+        task_args: Vec<(String, String)>,
+        handles: Vec<(String, String)>,
+        command_status: Option<String>,
+        recent_output: Vec<String>,
+        message: Option<String>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        validate_debug_snapshot(
+            &stack_frames,
+            &local_values,
+            &task_args,
+            &handles,
+            command_status.as_deref(),
+            &recent_output,
+            message.as_deref(),
+        )?;
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.coordinator
+            .authorize_node_for_process(&node, &tenant, &project, &process)?;
+        let participant_key = task_control_key(&tenant, &project, &process, &node, &task);
+        let process_key = process_control_key(&tenant, &project, &process);
+        let runtime = self
+            .debug_epoch_runtime
+            .get_mut(&process_key)
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol(format!(
+                    "cannot acknowledge debug epoch {epoch} for {process}: no active debug epoch"
+                ))
+            })?;
+        if runtime.epoch != epoch {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "cannot acknowledge debug epoch {epoch} for {process}: current debug epoch is {}",
+                runtime.epoch
+            )));
+        }
+        if !runtime.expected.contains(&participant_key) {
+            return Err(CoordinatorError::Unauthorized(
+                "debug acknowledgement is not from an expected active task participant".to_owned(),
+            )
+            .into());
+        }
+        let task_definition = self
+            .task_restart_checkpoints
+            .get(&task_restart_key(&tenant, &project, &process, &task))
+            .map(|checkpoint| checkpoint.assignment.task_spec.task_definition.clone())
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "debug acknowledgement does not name a coordinator-issued task instance"
+                        .to_owned(),
+                )
+            })?;
+        let valid_state = matches!(
+            (runtime.command.as_str(), &state),
+            ("freeze", DebugAcknowledgementState::Frozen)
+                | ("resume", DebugAcknowledgementState::Running)
+                | (_, DebugAcknowledgementState::Failed)
+        );
+        if !valid_state {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "debug epoch {epoch} command `{}` cannot be acknowledged as {state:?}",
+                runtime.command
+            )));
+        }
+        runtime.acknowledgements.insert(
+            participant_key,
+            DebugParticipantAcknowledgement {
+                node: node.clone(),
+                task_definition,
+                task: task.clone(),
+                epoch,
+                state: state.clone(),
+                stack_frames,
+                local_values,
+                task_args,
+                handles,
+                command_status,
+                recent_output,
+                message,
+            },
+        );
+        Ok(CoordinatorResponse::DebugStateRecorded {
+            process,
+            node,
+            task,
+            epoch,
+            state,
+        })
+    }
+
+    pub(super) fn handle_report_debug_probe_hit(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        probe_symbol: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let probe_symbol = validate_probe_symbols(vec![probe_symbol])?
+            .into_iter()
+            .next()
+            .expect("one validated probe symbol remains one symbol");
+        let tenant_id = TenantId::new(tenant.clone());
+        let project_id = ProjectId::new(project.clone());
+        let process_id = ProcessId::new(process.clone());
+        let node_id = NodeId::new(node.clone());
+        let task_id = TaskInstanceId::new(task.clone());
+        self.coordinator.authorize_node_for_process(
+            &node_id,
+            &tenant_id,
+            &project_id,
+            &process_id,
+        )?;
+        let participant_key =
+            task_control_key(&tenant_id, &project_id, &process_id, &node_id, &task_id);
+        if !self.active_tasks.contains(&participant_key) {
+            return Err(CoordinatorError::Unauthorized(
+                "debug probe hit is not from an active task participant".to_owned(),
+            )
+            .into());
+        }
+        let process_key = process_control_key(&tenant_id, &project_id, &process_id);
+        let Some(plan) = self.debug_breakpoints.get(&process_key).cloned() else {
+            return Ok(CoordinatorResponse::DebugProbeHit {
+                process: process_id,
+                node: node_id,
+                task: task_id,
+                probe_symbol,
+                breakpoint_matched: false,
+                debug_epoch: None,
+            });
+        };
+        if !plan.probe_symbols.contains(&probe_symbol) {
+            return Ok(CoordinatorResponse::DebugProbeHit {
+                process: process_id,
+                node: node_id,
+                task: task_id,
+                probe_symbol,
+                breakpoint_matched: false,
+                debug_epoch: None,
+            });
+        }
+        let response = self.handle_debug_epoch_control(
+            tenant,
+            project,
+            plan.actor.as_str().to_owned(),
+            process,
+            Some(task.clone()),
+            None,
+            "wasm_debug_probe_hit",
+            "freeze",
+            true,
+            format!("executing Wasm reached probe `{probe_symbol}`"),
+        )?;
+        let CoordinatorResponse::DebugEpoch { epoch, .. } = response else {
+            unreachable!("debug epoch control always returns DebugEpoch")
+        };
+        if let Some(plan) = self.debug_breakpoints.get_mut(&process_key) {
+            plan.hit_epoch = Some(epoch);
+            plan.hit_task = Some(task_id.clone());
+            plan.hit_probe_symbol = Some(probe_symbol.clone());
+        }
+        Ok(CoordinatorResponse::DebugProbeHit {
+            process: process_id,
+            node: node_id,
+            task: task_id,
+            probe_symbol,
+            breakpoint_matched: true,
+            debug_epoch: Some(epoch),
+        })
+    }
+
+    pub(super) fn handle_coordinator_main_debug_probe(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        process: ProcessId,
+        task: TaskInstanceId,
+        probe_symbol: String,
+    ) -> Result<WasmHostDebugProbeResult, CoordinatorServiceError> {
+        let probe_symbol = validate_probe_symbols(vec![probe_symbol])?
+            .into_iter()
+            .next()
+            .expect("one validated probe symbol remains one symbol");
+        let process_key = process_control_key(&tenant, &project, &process);
+        let main_matches = self
+            .main_runtime
+            .controls
+            .get(&process_key)
+            .is_some_and(|control| control.task_instance == task);
+        if !main_matches {
+            return Err(CoordinatorError::Unauthorized(
+                "debug probe does not belong to the active coordinator main participant".to_owned(),
+            )
+            .into());
+        }
+        let Some(plan) = self.debug_breakpoints.get(&process_key).cloned() else {
+            return Ok(WasmHostDebugProbeResult {
+                abi_version: WASM_TASK_ABI_VERSION,
+                breakpoint_matched: false,
+                debug_epoch: None,
+            });
+        };
+        if !plan.probe_symbols.contains(&probe_symbol) {
+            return Ok(WasmHostDebugProbeResult {
+                abi_version: WASM_TASK_ABI_VERSION,
+                breakpoint_matched: false,
+                debug_epoch: None,
+            });
+        }
+        if let Some(control) = self.main_runtime.controls.get_mut(&process_key) {
+            control.stopped_probe_symbol = Some(probe_symbol.clone());
+        }
+        let response = self.handle_debug_epoch_control(
+            tenant.as_str().to_owned(),
+            project.as_str().to_owned(),
+            plan.actor.as_str().to_owned(),
+            process.as_str().to_owned(),
+            Some(task.as_str().to_owned()),
+            None,
+            "wasm_debug_probe_hit",
+            "freeze",
+            true,
+            format!("coordinator main reached probe `{probe_symbol}`"),
+        )?;
+        let CoordinatorResponse::DebugEpoch { epoch, .. } = response else {
+            unreachable!("debug epoch control always returns DebugEpoch")
+        };
+        if let Some(plan) = self.debug_breakpoints.get_mut(&process_key) {
+            plan.hit_epoch = Some(epoch);
+            plan.hit_task = Some(task);
+            plan.hit_probe_symbol = Some(probe_symbol);
+        }
+        Ok(WasmHostDebugProbeResult {
+            abi_version: WASM_TASK_ABI_VERSION,
+            breakpoint_matched: true,
+            debug_epoch: Some(epoch),
+        })
+    }
+
+    pub(super) fn handle_inspect_debug_epoch(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        epoch: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        let process_key = process_control_key(&tenant, &project, &process);
+        self.sync_coordinator_main_debug_acknowledgement(&process_key, epoch);
+        let runtime = self
+            .debug_epoch_runtime
+            .get(&process_key)
+            .filter(|runtime| runtime.epoch == epoch)
+            .cloned()
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol(format!(
+                    "debug epoch {epoch} is not active for {process}"
+                ))
+            })?;
+        let audit_event = self.record_debug_audit_event(
+            tenant,
+            project,
+            process.clone(),
+            None,
+            actor.clone(),
+            "inspect_debug_epoch",
+            authorization.allowed,
+            authorization.reason.clone(),
+        )?;
+        if !authorization.allowed {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "inspect_debug_epoch denied: {}",
+                authorization.reason
+            ))
+            .into());
+        }
+        let acknowledgements = runtime
+            .acknowledgements
+            .values()
+            .cloned()
+            .collect::<Vec<_>>();
+        let all_acknowledged = !runtime.expected.is_empty()
+            && runtime
+                .expected
+                .iter()
+                .all(|key| runtime.acknowledgements.contains_key(key));
+        let fully_frozen = runtime.command == "freeze"
+            && all_acknowledged
+            && acknowledgements
+                .iter()
+                .all(|ack| ack.state == DebugAcknowledgementState::Frozen);
+        let fully_resumed = runtime.command == "resume"
+            && all_acknowledged
+            && acknowledgements
+                .iter()
+                .all(|ack| ack.state == DebugAcknowledgementState::Running);
+        let failed = acknowledgements
+            .iter()
+            .any(|ack| ack.state == DebugAcknowledgementState::Failed);
+        let failure_messages = acknowledgements
+            .iter()
+            .filter(|ack| ack.state == DebugAcknowledgementState::Failed)
+            .map(|ack| {
+                ack.message
+                    .clone()
+                    .unwrap_or_else(|| format!("task {} failed debug control", ack.task))
+            })
+            .collect();
+        let expected_tasks = runtime
+            .expected
+            .into_iter()
+            .map(|(_, _, process, node, task)| TaskCancellationTarget {
+                process,
+                node,
+                task,
+            })
+            .collect();
+        Ok(CoordinatorResponse::DebugEpochStatus {
+            process,
+            actor,
+            epoch,
+            command: runtime.command,
+            expected_tasks,
+            acknowledgements,
+            fully_frozen,
+            fully_resumed,
+            failed,
+            failure_messages,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+
+    pub(super) fn clear_debug_state_for_process(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) {
+        self.debug_epochs
+            .remove(&process_control_key(tenant, project, process));
+        self.debug_epoch_runtime
+            .remove(&process_control_key(tenant, project, process));
+        self.debug_breakpoints
+            .remove(&process_control_key(tenant, project, process));
+        self.debug_commands
+            .retain(|(task_tenant, task_project, task_process, _, _), _| {
+                task_tenant != tenant || task_project != project || task_process != process
+            });
+    }
+
+    fn handle_debug_epoch_control(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        stopped_task: Option<String>,
+        expected_epoch: Option<u64>,
+        operation: &'static str,
+        command: &'static str,
+        all_stop_requested: bool,
+        reason: impl Into<String>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        let reason = reason.into();
+        let audit_event = self.record_debug_audit_event(
+            tenant.clone(),
+            project.clone(),
+            process.clone(),
+            stopped_task.as_ref().map(TaskInstanceId::new),
+            actor.clone(),
+            operation,
+            authorization.allowed,
+            if authorization.allowed {
+                reason
+            } else {
+                authorization.reason.clone()
+            },
+        )?;
+        if !authorization.allowed {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "{operation} denied: {}",
+                authorization.reason
+            ))
+            .into());
+        }
+
+        let process_key = process_control_key(&tenant, &project, &process);
+        let epoch = match expected_epoch {
+            Some(expected) => {
+                let current = self
+                    .debug_epochs
+                    .get(&process_key)
+                    .copied()
+                    .ok_or_else(|| {
+                        CoordinatorServiceError::Protocol(format!(
+                        "cannot resume debug epoch {expected} for {process}: no active debug epoch"
+                    ))
+                    })?;
+                if current != expected {
+                    return Err(CoordinatorServiceError::Protocol(format!(
+                        "cannot resume debug epoch {expected} for {process}: current debug epoch is {current}"
+                    )));
+                }
+                let runtime = self.debug_epoch_runtime.get(&process_key).ok_or_else(|| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "cannot resume debug epoch {expected} for {process}: participant state is unavailable"
+                    ))
+                })?;
+                if runtime.command != "freeze"
+                    || !runtime_all_in_state(runtime, DebugAcknowledgementState::Frozen)
+                {
+                    return Err(CoordinatorServiceError::Protocol(format!(
+                        "cannot resume debug epoch {expected} for {process}: all expected participants have not acknowledged frozen state"
+                    )));
+                }
+                current
+            }
+            None => {
+                if let Some(runtime) = self.debug_epoch_runtime.get(&process_key) {
+                    let previous_complete = runtime.command == "resume"
+                        && runtime_all_in_state(runtime, DebugAcknowledgementState::Running);
+                    if !previous_complete {
+                        return Err(CoordinatorServiceError::Protocol(format!(
+                            "cannot create a new debug epoch for {process}: epoch {} has not fully resumed",
+                            runtime.epoch
+                        )));
+                    }
+                }
+                let next = self.debug_epochs.get(&process_key).copied().unwrap_or(0) + 1;
+                self.debug_epochs.insert(process_key.clone(), next);
+                next
+            }
+        };
+        let mut affected_tasks = self
+            .enqueue_debug_command_for_active_tasks(&tenant, &project, &process, epoch, command);
+        let mut expected = self
+            .active_tasks
+            .iter()
+            .filter(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant == &tenant && task_project == &project && task_process == &process
+            })
+            .cloned()
+            .collect::<BTreeSet<_>>();
+        if let Some(control) = self
+            .main_runtime
+            .controls
+            .get(&process_key)
+            .filter(|control| matches!(control.state.as_str(), "running" | "stopping"))
+        {
+            let key = task_control_key(
+                &tenant,
+                &project,
+                &process,
+                &NodeId::from("coordinator-main"),
+                &control.task_instance,
+            );
+            expected.insert(key);
+            affected_tasks.push(TaskCancellationTarget {
+                process: process.clone(),
+                node: NodeId::from("coordinator-main"),
+                task: control.task_instance.clone(),
+            });
+        }
+        self.debug_epoch_runtime.insert(
+            process_key.clone(),
+            DebugEpochRuntime {
+                epoch,
+                command: command.to_owned(),
+                expected,
+                acknowledgements: BTreeMap::new(),
+            },
+        );
+        if let Some(control) = self
+            .main_runtime
+            .controls
+            .get(&process_key)
+            .filter(|control| matches!(control.state.as_str(), "running" | "stopping"))
+        {
+            if command == "freeze" {
+                control.debug.request_freeze(epoch);
+            } else {
+                control.debug.request_resume(epoch);
+            }
+        }
+        self.sync_coordinator_main_debug_acknowledgement(&process_key, epoch);
+        Ok(CoordinatorResponse::DebugEpoch {
+            process,
+            actor,
+            epoch,
+            command: command.to_owned(),
+            affected_tasks,
+            all_stop_requested,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+
+    fn sync_coordinator_main_debug_acknowledgement(
+        &mut self,
+        process_key: &super::keys::ProcessControlKey,
+        epoch: u64,
+    ) {
+        let Some(control) = self.main_runtime.controls.get(process_key) else {
+            return;
+        };
+        let Some(runtime) = self.debug_epoch_runtime.get_mut(process_key) else {
+            return;
+        };
+        if runtime.epoch != epoch {
+            return;
+        }
+        let state = if runtime.command == "freeze" && control.debug.frozen_epoch() == Some(epoch) {
+            Some(DebugAcknowledgementState::Frozen)
+        } else if runtime.command == "resume"
+            && control.debug.resume_requested(epoch)
+            && control.debug.frozen_epoch() != Some(epoch)
+        {
+            Some(DebugAcknowledgementState::Running)
+        } else {
+            None
+        };
+        let Some(state) = state else {
+            return;
+        };
+        let handles = control
+            .handles
+            .lock()
+            .map(|handles| {
+                let mut snapshot = handles
+                    .iter()
+                    .map(|(handle_id, spec)| {
+                        (
+                            format!("task_handle_{handle_id}"),
+                            format!(
+                                "definition={} instance={} state=active",
+                                spec.task_definition, spec.task_instance
+                            ),
+                        )
+                    })
+                    .collect::<Vec<_>>();
+                snapshot.sort_by(|left, right| left.0.cmp(&right.0));
+                snapshot
+            })
+            .unwrap_or_else(|_| {
+                vec![(
+                    "handle-registry-diagnostic".to_owned(),
+                    "coordinator main handle registry was unavailable".to_owned(),
+                )]
+            });
+        let node = NodeId::from("coordinator-main");
+        let key = task_control_key(
+            &process_key.0,
+            &process_key.1,
+            &process_key.2,
+            &node,
+            &control.task_instance,
+        );
+        if !runtime.expected.contains(&key) {
+            return;
+        }
+        runtime.acknowledgements.insert(
+            key,
+            DebugParticipantAcknowledgement {
+                node,
+                task_definition: control.task_definition.clone(),
+                task: control.task_instance.clone(),
+                epoch,
+                state,
+                stack_frames: vec![control
+                    .stopped_probe_symbol
+                    .as_deref()
+                    .and_then(|symbol| symbol.strip_prefix("disasmer.probe."))
+                    .map(|function| format!("{function}::wasm"))
+                    .unwrap_or_else(|| {
+                        format!("coordinator_main::wasm ({})", control.task_definition)
+                    })],
+                local_values: Vec::new(),
+                task_args: Vec::new(),
+                handles,
+                command_status: None,
+                recent_output: vec![
+                    "coordinator-hosted capless main acknowledged all-stop control".to_owned(),
+                ],
+                message: None,
+            },
+        );
+    }
+
+    fn enqueue_debug_command_for_active_tasks(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+        epoch: u64,
+        command: &str,
+    ) -> Vec<TaskCancellationTarget> {
+        let task_keys = self
+            .active_tasks
+            .iter()
+            .filter(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant == tenant && task_project == project && task_process == process
+            })
+            .cloned()
+            .collect::<Vec<_>>();
+        for key in &task_keys {
+            self.debug_commands.insert(
+                key.clone(),
+                DebugPendingCommand {
+                    epoch,
+                    command: command.to_owned(),
+                },
+            );
+        }
+        task_keys
+            .into_iter()
+            .map(|(_, _, process, node, task)| TaskCancellationTarget {
+                process,
+                task,
+                node,
+            })
+            .collect()
+    }
+
+    pub(super) fn record_debug_audit_event(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        process: ProcessId,
+        task: Option<TaskInstanceId>,
+        actor: UserId,
+        operation: &'static str,
+        allowed: bool,
+        reason: impl Into<String>,
+    ) -> Result<DebugAuditEvent, CoordinatorServiceError> {
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        let charged_debug_read_bytes = if allowed {
+            self.quota.charge_debug_read(
+                &tenant,
+                &project,
+                DEBUG_CONTROL_READ_BYTES,
+                now_epoch_seconds,
+            )?;
+            DEBUG_CONTROL_READ_BYTES
+        } else {
+            0
+        };
+        let used_debug_read_bytes =
+            self.quota
+                .used_debug_read_bytes(&tenant, &project, now_epoch_seconds);
+        let event = DebugAuditEvent {
+            tenant,
+            project,
+            process,
+            task,
+            actor,
+            operation: operation.to_owned(),
+            allowed,
+            reason: reason.into(),
+            charged_debug_read_bytes,
+            used_debug_read_bytes,
+        };
+        while self
+            .debug_audit_events
+            .iter()
+            .filter(|retained| {
+                retained.tenant == event.tenant
+                    && retained.project == event.project
+                    && retained.process == event.process
+            })
+            .count()
+            >= super::MAX_DEBUG_AUDIT_EVENTS_PER_PROCESS
+        {
+            let Some(index) = self.debug_audit_events.iter().position(|retained| {
+                retained.tenant == event.tenant
+                    && retained.project == event.project
+                    && retained.process == event.process
+            }) else {
+                break;
+            };
+            self.debug_audit_events.remove(index);
+        }
+        self.debug_audit_events.push_back(event.clone());
+        Ok(event)
+    }
+}

crates/disasmer-coordinator/src/service/debug/validation.rs

@@ -0,0 +1,92 @@
+use std::collections::BTreeSet;
+
+use crate::CoordinatorServiceError;
+
+use super::{DebugAcknowledgementState, DebugEpochRuntime};
+
+pub(super) fn runtime_all_in_state(
+    runtime: &DebugEpochRuntime,
+    state: DebugAcknowledgementState,
+) -> bool {
+    !runtime.expected.is_empty()
+        && runtime.expected.iter().all(|key| {
+            runtime
+                .acknowledgements
+                .get(key)
+                .is_some_and(|ack| ack.state == state)
+        })
+}
+
+pub(super) fn validate_probe_symbols(
+    probe_symbols: Vec<String>,
+) -> Result<Vec<String>, CoordinatorServiceError> {
+    if probe_symbols.len() > 128 {
+        return Err(CoordinatorServiceError::Protocol(
+            "debug breakpoint request exceeds 128 probe symbols".to_owned(),
+        ));
+    }
+    let mut unique = BTreeSet::new();
+    for symbol in probe_symbols {
+        if symbol.trim().is_empty()
+            || symbol.len() > 256
+            || !symbol
+                .chars()
+                .all(|character| character.is_ascii_alphanumeric() || "._:-/".contains(character))
+        {
+            return Err(CoordinatorServiceError::Protocol(
+                "debug breakpoint probe symbol is invalid".to_owned(),
+            ));
+        }
+        unique.insert(symbol);
+    }
+    Ok(unique.into_iter().collect())
+}
+
+pub(super) fn validate_debug_snapshot(
+    stack_frames: &[String],
+    local_values: &[(String, String)],
+    task_args: &[(String, String)],
+    handles: &[(String, String)],
+    command_status: Option<&str>,
+    recent_output: &[String],
+    message: Option<&str>,
+) -> Result<(), CoordinatorServiceError> {
+    const MAX_ITEMS: usize = 128;
+    const MAX_TEXT_BYTES: usize = 16 * 1024;
+    if [
+        stack_frames.len(),
+        local_values.len(),
+        task_args.len(),
+        handles.len(),
+        recent_output.len(),
+    ]
+    .into_iter()
+    .any(|count| count > MAX_ITEMS)
+    {
+        return Err(CoordinatorServiceError::Protocol(
+            "debug participant snapshot exceeds the 128-item field limit".to_owned(),
+        ));
+    }
+    let text_bytes = stack_frames.iter().map(String::len).sum::<usize>()
+        + local_values
+            .iter()
+            .map(|(name, value)| name.len() + value.len())
+            .sum::<usize>()
+        + task_args
+            .iter()
+            .map(|(name, value)| name.len() + value.len())
+            .sum::<usize>()
+        + handles
+            .iter()
+            .map(|(name, value)| name.len() + value.len())
+            .sum::<usize>()
+        + recent_output.iter().map(String::len).sum::<usize>()
+        + command_status.map(str::len).unwrap_or(0)
+        + message.map(str::len).unwrap_or(0);
+    if text_bytes > MAX_TEXT_BYTES {
+        return Err(CoordinatorServiceError::Protocol(format!(
+            "debug participant snapshot is {text_bytes} bytes; maximum is {MAX_TEXT_BYTES}"
+        )));
+    }
+    Ok(())
+}

crates/disasmer-coordinator/src/service/debug_requests.rs

@@ -0,0 +1,470 @@
+use std::collections::BTreeSet;
+
+use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine as _};
+use disasmer_core::{
+    Actor, Capability, CredentialKind, Digest, EnvironmentResource, ProcessId, ProjectId,
+    RestartDecision, RestartPolicy, RestartRequest, TaskDispatch, TaskInstanceId, TenantId, UserId,
+    WasmExportAbi,
+};
+
+use crate::{CoordinatorError, CoordinatorServiceError};
+
+use super::keys::task_restart_key;
+use super::{
+    AuthenticatedCoordinatorRequest, CoordinatorRequest, CoordinatorResponse, CoordinatorService,
+    TaskReplacementBundle, WorkflowActor,
+};
+
+impl CoordinatorService {
+    pub(super) fn handle_debug_request(
+        &mut self,
+        request: CoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        match request {
+            CoordinatorRequest::DebugAttach {
+                tenant,
+                project,
+                actor_user,
+                process,
+            } => self.handle_debug_attach(tenant, project, actor_user, process),
+            CoordinatorRequest::SetDebugBreakpoints {
+                tenant,
+                project,
+                actor_user,
+                process,
+                probe_symbols,
+            } => self.handle_set_debug_breakpoints(
+                tenant,
+                project,
+                actor_user,
+                process,
+                probe_symbols,
+            ),
+            CoordinatorRequest::InspectDebugBreakpoints {
+                tenant,
+                project,
+                actor_user,
+                process,
+            } => self.handle_inspect_debug_breakpoints(tenant, project, actor_user, process),
+            CoordinatorRequest::CreateDebugEpoch {
+                tenant,
+                project,
+                actor_user,
+                process,
+                stopped_task,
+                reason,
+            } => self.handle_create_debug_epoch(
+                tenant,
+                project,
+                actor_user,
+                process,
+                stopped_task,
+                reason,
+            ),
+            CoordinatorRequest::ResumeDebugEpoch {
+                tenant,
+                project,
+                actor_user,
+                process,
+                epoch,
+            } => self.handle_resume_debug_epoch(tenant, project, actor_user, process, epoch),
+            CoordinatorRequest::InspectDebugEpoch {
+                tenant,
+                project,
+                actor_user,
+                process,
+                epoch,
+            } => self.handle_inspect_debug_epoch(tenant, project, actor_user, process, epoch),
+            _ => unreachable!("handle_debug_request only accepts debug coordinator requests"),
+        }
+    }
+
+    pub(super) fn handle_authenticated_debug_request(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        actor: &UserId,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        match request {
+            AuthenticatedCoordinatorRequest::DebugAttach { process } => self.handle_debug_attach(
+                tenant.as_str().to_owned(),
+                project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+            ),
+            AuthenticatedCoordinatorRequest::SetDebugBreakpoints {
+                process,
+                probe_symbols,
+            } => self.handle_set_debug_breakpoints(
+                tenant.as_str().to_owned(),
+                project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+                probe_symbols,
+            ),
+            AuthenticatedCoordinatorRequest::InspectDebugBreakpoints { process } => self
+                .handle_inspect_debug_breakpoints(
+                    tenant.as_str().to_owned(),
+                    project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    process,
+                ),
+            AuthenticatedCoordinatorRequest::CreateDebugEpoch {
+                process,
+                stopped_task,
+                reason,
+            } => self.handle_create_debug_epoch(
+                tenant.as_str().to_owned(),
+                project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+                stopped_task,
+                reason,
+            ),
+            AuthenticatedCoordinatorRequest::ResumeDebugEpoch { process, epoch } => self
+                .handle_resume_debug_epoch(
+                    tenant.as_str().to_owned(),
+                    project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    process,
+                    epoch,
+                ),
+            AuthenticatedCoordinatorRequest::InspectDebugEpoch { process, epoch } => self
+                .handle_inspect_debug_epoch(
+                    tenant.as_str().to_owned(),
+                    project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    process,
+                    epoch,
+                ),
+            _ => unreachable!(
+                "handle_authenticated_debug_request only accepts debug coordinator requests"
+            ),
+        }
+    }
+
+    pub(super) fn handle_restart_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        task: String,
+        replacement_bundle: Option<TaskReplacementBundle>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let task = TaskInstanceId::new(task);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        let authorization = self.coordinator.authorize_debug_attach(&context, &process);
+        if !authorization.allowed {
+            let _ = self.record_debug_audit_event(
+                tenant,
+                project,
+                process,
+                Some(task),
+                actor,
+                "restart_task",
+                false,
+                authorization.reason.clone(),
+            )?;
+            return Err(CoordinatorError::Unauthorized(format!(
+                "task restart denied: {}",
+                authorization.reason
+            ))
+            .into());
+        }
+        let active_key = self
+            .active_tasks
+            .iter()
+            .find(|(task_tenant, task_project, task_process, _, task_id)| {
+                task_tenant == &tenant
+                    && task_project == &project
+                    && task_process == &process
+                    && task_id == &task
+            })
+            .cloned();
+        let process_key = super::keys::process_control_key(&tenant, &project, &process);
+        let active_main = self
+            .main_runtime
+            .controls
+            .get(&process_key)
+            .is_some_and(|control| {
+                control.task_instance == task
+                    && matches!(control.state.as_str(), "running" | "stopping")
+            });
+        let active_task = active_key.is_some() || active_main;
+        let completed_event_observed = self.task_events.iter().any(|event| {
+            event.tenant == tenant
+                && event.project == project
+                && event.process == process
+                && event.task == task
+        });
+        let checkpoint_key = task_restart_key(&tenant, &project, &process, &task);
+        let checkpoint = self.task_restart_checkpoints.get(&checkpoint_key).cloned();
+        let mut accepted = false;
+        let mut restarted_task_instance = None;
+        let mut clean_boundary_available = false;
+        let mut requires_whole_process_restart = true;
+        let message = if active_main {
+            "selected coordinator main is still active; restart the whole virtual process to rerun its capless entry boundary".to_owned()
+        } else if active_task {
+            "selected task is still active; wait for its terminal event or abort the whole virtual process before restarting from its clean entry boundary".to_owned()
+        } else if !completed_event_observed {
+            "selected task is not known in the active process; restart the whole virtual process or inspect task list".to_owned()
+        } else if let Some(checkpoint) = checkpoint {
+            let vfs_available =
+                checkpoint
+                    .checkpoint
+                    .vfs_manifest
+                    .objects
+                    .iter()
+                    .all(|(path, object)| {
+                        let artifact = super::keys::artifact_id_from_path(path);
+                        self.artifact_registry
+                            .metadata(&artifact)
+                            .is_some_and(|metadata| {
+                                metadata.tenant == tenant
+                                    && metadata.project == project
+                                    && metadata.digest == object.digest
+                                    && metadata.size == object.size
+                                    && !metadata.retaining_nodes.is_empty()
+                            })
+                    });
+            if !vfs_available {
+                "selected task checkpoint references VFS artifacts that are no longer retained; restart the whole virtual process".to_owned()
+            } else {
+                let replacement = replacement_bundle
+                    .as_ref()
+                    .map(|bundle| {
+                        validate_task_replacement(
+                            bundle,
+                            &checkpoint.assignment.task_spec.task_definition,
+                            checkpoint.assignment.task_spec.environment_id.as_deref(),
+                        )
+                    })
+                    .transpose()?;
+                let request = RestartRequest {
+                    task: task.clone(),
+                    entrypoint: replacement.as_ref().map_or_else(
+                        || checkpoint.checkpoint.boundary.task_entrypoint.clone(),
+                        |replacement| replacement.export.clone(),
+                    ),
+                    serialized_args: checkpoint.checkpoint.boundary.serialized_args.clone(),
+                    environment_digest: replacement.as_ref().map_or_else(
+                        || checkpoint.checkpoint.boundary.environment_digest.clone(),
+                        |replacement| replacement.environment_digest.clone(),
+                    ),
+                    task_abi: replacement.as_ref().map_or_else(
+                        || checkpoint.checkpoint.boundary.task_abi.clone(),
+                        |replacement| replacement.restart_compatibility.clone(),
+                    ),
+                    source_edited: replacement.is_some(),
+                };
+                match RestartPolicy.decide(&checkpoint.checkpoint, &request) {
+                    RestartDecision::RestartTask { from_vfs_epoch, .. } => {
+                        clean_boundary_available = true;
+                        let mut assignment = checkpoint.assignment;
+                        let new_task_instance = disasmer_core::TaskInstanceId::new(
+                            disasmer_core::generate_opaque_token("ti")
+                                .map_err(CoordinatorServiceError::Protocol)?,
+                        );
+                        assignment.task = new_task_instance.clone();
+                        assignment.task_spec.task_instance = new_task_instance.clone();
+                        assignment.artifact_path =
+                            format!("/vfs/artifacts/{}-result.json", new_task_instance.as_str());
+                        if let Some(replacement) = replacement {
+                            assignment.task_spec.dispatch = TaskDispatch::CoordinatorNodeWasm {
+                                export: Some(replacement.export),
+                                abi: WasmExportAbi::TaskV1,
+                            };
+                            assignment.task_spec.environment = replacement
+                                .environment
+                                .map(|environment| environment.requirements);
+                            assignment.task_spec.environment_digest =
+                                Some(replacement.environment_digest);
+                            assignment.task_spec.required_capabilities =
+                                replacement.required_capabilities;
+                            assignment.task_spec.bundle_digest =
+                                Some(replacement.bundle_digest.clone());
+                            assignment.wasm_module_base64 = replacement.wasm_module_base64;
+                        }
+                        let launch = self.handle_launch_task_with_actor(
+                            tenant.clone(),
+                            project.clone(),
+                            WorkflowActor {
+                                kind: "user".to_owned(),
+                                user: Some(actor.clone()),
+                                agent: None,
+                                credential_kind: CredentialKind::CliDeviceSession,
+                                public_key_fingerprint: None,
+                                authenticated_without_browser: false,
+                                scopes: vec!["process:restart-task".to_owned()],
+                            },
+                            assignment.task_spec,
+                            true,
+                            assignment.artifact_path,
+                            assignment.wasm_module_base64,
+                        )?;
+                        accepted = matches!(
+                            launch,
+                            CoordinatorResponse::TaskLaunched { .. }
+                                | CoordinatorResponse::TaskQueued { .. }
+                        );
+                        if accepted {
+                            restarted_task_instance = Some(new_task_instance.clone());
+                        }
+                        requires_whole_process_restart = !accepted;
+                        format!(
+                            "selected task restarted as new instance {new_task_instance} from clean VFS entry boundary epoch {from_vfs_epoch}; unflushed task changes were discarded"
+                        )
+                    }
+                    RestartDecision::RestartWholeVirtualProcess { message } => message,
+                }
+            }
+        } else {
+            "selected task has terminal metadata but no captured clean VFS entry boundary; restart the whole virtual process".to_owned()
+        };
+        let audit_event = self.record_debug_audit_event(
+            tenant,
+            project,
+            process.clone(),
+            Some(task.clone()),
+            actor.clone(),
+            "restart_task",
+            true,
+            &message,
+        )?;
+        Ok(CoordinatorResponse::TaskRestart {
+            process,
+            task,
+            restarted_task_instance,
+            actor,
+            accepted,
+            clean_boundary_available,
+            active_task,
+            completed_event_observed,
+            requires_whole_process_restart,
+            message,
+            charged_debug_read_bytes: audit_event.charged_debug_read_bytes,
+            used_debug_read_bytes: audit_event.used_debug_read_bytes,
+            audit_event,
+        })
+    }
+}
+
+struct ValidatedTaskReplacement {
+    bundle_digest: Digest,
+    wasm_module_base64: String,
+    export: String,
+    restart_compatibility: Digest,
+    environment: Option<EnvironmentResource>,
+    environment_digest: Digest,
+    required_capabilities: BTreeSet<Capability>,
+}
+
+fn validate_task_replacement(
+    replacement: &TaskReplacementBundle,
+    task_definition: &disasmer_core::TaskDefinitionId,
+    environment_id: Option<&str>,
+) -> Result<ValidatedTaskReplacement, CoordinatorServiceError> {
+    let module = BASE64_STANDARD
+        .decode(&replacement.wasm_module_base64)
+        .map_err(|error| {
+            CoordinatorServiceError::Protocol(format!(
+                "replacement task bundle is not valid base64: {error}"
+            ))
+        })?;
+    let actual_digest = Digest::sha256(&module);
+    if actual_digest != replacement.bundle_digest {
+        return Err(CoordinatorServiceError::Protocol(format!(
+            "replacement task bundle digest mismatch: expected {}, actual {actual_digest}",
+            replacement.bundle_digest
+        )));
+    }
+    let descriptors = super::main_runtime::task_descriptors(&module)?;
+    let descriptor = descriptors.get(task_definition.as_str()).ok_or_else(|| {
+        CoordinatorServiceError::Protocol(format!(
+            "replacement bundle has no task definition `{task_definition}`"
+        ))
+    })?;
+    if descriptor
+        .get("abi_version")
+        .and_then(serde_json::Value::as_u64)
+        != Some(disasmer_core::WASM_TASK_ABI_VERSION as u64)
+    {
+        return Err(CoordinatorServiceError::Protocol(format!(
+            "replacement task `{task_definition}` uses an unsupported task ABI"
+        )));
+    }
+    let export = descriptor
+        .get("export")
+        .and_then(serde_json::Value::as_str)
+        .filter(|export| !export.is_empty())
+        .ok_or_else(|| {
+            CoordinatorServiceError::Protocol(format!(
+                "replacement task `{task_definition}` omitted its export"
+            ))
+        })?
+        .to_owned();
+    let restart_compatibility = serde_json::from_value(
+        descriptor
+            .get("restart_compatibility_hash")
+            .cloned()
+            .ok_or_else(|| {
+                CoordinatorServiceError::Protocol(format!(
+                    "replacement task `{task_definition}` omitted restart compatibility metadata"
+                ))
+            })?,
+    )?;
+    let mut required_capabilities = descriptor
+        .get("required_capabilities")
+        .and_then(serde_json::Value::as_array)
+        .into_iter()
+        .flatten()
+        .map(|capability| {
+            super::main_runtime::capability_from_descriptor(capability.as_str().ok_or_else(
+                || {
+                    CoordinatorServiceError::Protocol(
+                        "replacement task capability is not a string".to_owned(),
+                    )
+                },
+            )?)
+            .map_err(CoordinatorServiceError::Protocol)
+        })
+        .collect::<Result<BTreeSet<_>, _>>()?;
+    let environment = environment_id
+        .map(|environment_id| {
+            super::main_runtime::bundle_environments(&module)?
+                .remove(environment_id)
+                .ok_or_else(|| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "replacement bundle has no environment `{environment_id}`"
+                    ))
+                })
+        })
+        .transpose()?;
+    if let Some(environment) = &environment {
+        required_capabilities.extend(environment.requirements.capabilities.iter().cloned());
+    }
+    let environment_digest = environment.as_ref().map_or_else(
+        || Digest::sha256("disasmer.environment.unconstrained.v1"),
+        |environment| environment.digest.clone(),
+    );
+    Ok(ValidatedTaskReplacement {
+        bundle_digest: replacement.bundle_digest.clone(),
+        wasm_module_base64: replacement.wasm_module_base64.clone(),
+        export,
+        restart_compatibility,
+        environment,
+        environment_digest,
+        required_capabilities,
+    })
+}

crates/disasmer-coordinator/src/service/durable_runtime.rs

@@ -0,0 +1,47 @@
+use crate::{
+    DurableState, DurableStore, FallibleDurableStore, InMemoryDurableStore, PostgresDurableStore,
+};
+
+pub(super) enum RuntimeDurableStore {
+    InMemory(InMemoryDurableStore),
+    Postgres(PostgresDurableStore),
+}
+
+impl RuntimeDurableStore {
+    pub(super) fn from_database_url(database_url: Option<&str>) -> Result<Self, String> {
+        match database_url.map(str::trim).filter(|url| !url.is_empty()) {
+            Some(url) => PostgresDurableStore::connect(url)
+                .map(Self::Postgres)
+                .map_err(|error| error.to_string()),
+            None => Ok(Self::InMemory(InMemoryDurableStore::default())),
+        }
+    }
+
+    pub(super) fn kind(&self) -> &'static str {
+        match self {
+            Self::InMemory(_) => "in_memory",
+            Self::Postgres(_) => "postgres",
+        }
+    }
+}
+
+impl FallibleDurableStore for RuntimeDurableStore {
+    type Error = String;
+
+    fn load_state(&mut self) -> Result<DurableState, Self::Error> {
+        match self {
+            Self::InMemory(store) => Ok(store.load()),
+            Self::Postgres(store) => store.load_state().map_err(|error| error.to_string()),
+        }
+    }
+
+    fn save_state(&mut self, state: &DurableState) -> Result<(), Self::Error> {
+        match self {
+            Self::InMemory(store) => {
+                store.save(state.clone());
+                Ok(())
+            }
+            Self::Postgres(store) => store.save_state(state).map_err(|error| error.to_string()),
+        }
+    }
+}

crates/disasmer-coordinator/src/service/keys.rs

@@ -0,0 +1,71 @@
+use disasmer_core::{ArtifactId, NodeId, ProcessId, ProjectId, TaskInstanceId, TenantId, VfsPath};
+
+pub(super) type TaskControlKey = (TenantId, ProjectId, ProcessId, NodeId, TaskInstanceId);
+pub(super) type TaskRestartKey = (TenantId, ProjectId, ProcessId, TaskInstanceId);
+pub(super) type TaskAssignmentKey = (TenantId, ProjectId, NodeId);
+pub(super) type PanelStopKey = (TenantId, ProjectId, ProcessId);
+pub(super) type EnrollmentGrantKey = (TenantId, ProjectId, String);
+pub(super) type ProcessControlKey = (TenantId, ProjectId, ProcessId);
+
+pub(super) fn task_control_key(
+    tenant: &TenantId,
+    project: &ProjectId,
+    process: &ProcessId,
+    node: &NodeId,
+    task: &TaskInstanceId,
+) -> TaskControlKey {
+    (
+        tenant.clone(),
+        project.clone(),
+        process.clone(),
+        node.clone(),
+        task.clone(),
+    )
+}
+
+pub(super) fn task_restart_key(
+    tenant: &TenantId,
+    project: &ProjectId,
+    process: &ProcessId,
+    task: &TaskInstanceId,
+) -> TaskRestartKey {
+    (
+        tenant.clone(),
+        project.clone(),
+        process.clone(),
+        task.clone(),
+    )
+}
+
+pub(super) fn process_control_key(
+    tenant: &TenantId,
+    project: &ProjectId,
+    process: &ProcessId,
+) -> ProcessControlKey {
+    (tenant.clone(), project.clone(), process.clone())
+}
+
+pub(super) fn panel_stop_key(
+    tenant: &TenantId,
+    project: &ProjectId,
+    process: &ProcessId,
+) -> PanelStopKey {
+    (tenant.clone(), project.clone(), process.clone())
+}
+
+pub(super) fn enrollment_grant_key(
+    tenant: &TenantId,
+    project: &ProjectId,
+    grant: &str,
+) -> EnrollmentGrantKey {
+    (tenant.clone(), project.clone(), grant.to_owned())
+}
+
+pub(super) fn artifact_id_from_path(path: &VfsPath) -> ArtifactId {
+    let value = path
+        .as_str()
+        .strip_prefix("/vfs/artifacts/")
+        .unwrap_or(path.as_str())
+        .replace('/', ":");
+    ArtifactId::new(value)
+}

crates/disasmer-coordinator/src/service/logs.rs

@@ -0,0 +1,535 @@
+use disasmer_core::{
+    ArtifactFlush, ArtifactId, Digest, NodeId, ProcessId, ProjectId, TaskBoundaryValue,
+    TaskInstanceId, TaskJoinResult, TaskJoinState, TenantId, UserId, VfsPath,
+};
+
+use crate::CoordinatorError;
+
+use super::keys::{process_control_key, task_control_key};
+use super::{
+    artifact_id_from_path, CoordinatorResponse, CoordinatorService, CoordinatorServiceError,
+    TaskCompletionEvent, TaskTerminalState, MAX_TASK_LOG_TAIL_BYTES,
+};
+
+impl CoordinatorService {
+    pub(super) fn handle_report_task_log(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        stdout_bytes: u64,
+        stderr_bytes: u64,
+        stdout_tail: String,
+        stderr_tail: String,
+        stdout_truncated: bool,
+        stderr_truncated: bool,
+        backpressured: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        validate_task_log_tail("stdout_tail", &stdout_tail)?;
+        validate_task_log_tail("stderr_tail", &stderr_tail)?;
+        let reported_bytes = checked_reported_log_bytes(stdout_bytes, stderr_bytes)?;
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.quota
+            .can_charge_log_bytes(&tenant, &project, reported_bytes, now_epoch_seconds)?;
+        self.quota
+            .charge_log_bytes(&tenant, &project, reported_bytes, now_epoch_seconds)?;
+        Ok(CoordinatorResponse::TaskLogRecorded {
+            process,
+            task,
+            stdout_bytes,
+            stderr_bytes,
+            stdout_tail: if stdout_truncated {
+                format!("{stdout_tail}\n... truncated")
+            } else {
+                stdout_tail
+            },
+            stderr_tail: if stderr_truncated {
+                format!("{stderr_tail}\n... truncated")
+            } else {
+                stderr_tail
+            },
+            backpressured,
+        })
+    }
+
+    pub(super) fn handle_report_vfs_metadata(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        artifact_path: Option<String>,
+        artifact_digest: Option<Digest>,
+        artifact_size_bytes: Option<u64>,
+        large_bytes_uploaded: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let artifact_path = artifact_path
+            .map(VfsPath::new)
+            .transpose()
+            .map_err(|err| CoordinatorServiceError::InvalidArtifactPath(err.to_string()))?;
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        if let (Some(path), Some(digest)) = (&artifact_path, artifact_digest) {
+            self.flush_artifact_metadata(ArtifactFlush {
+                id: artifact_id_from_path(path),
+                tenant,
+                project,
+                process: process.clone(),
+                producer_task: task.clone(),
+                retaining_node: node,
+                digest,
+                size: artifact_size_bytes.unwrap_or_default(),
+            })?;
+        }
+        Ok(CoordinatorResponse::VfsMetadataRecorded {
+            process,
+            task,
+            artifact_path,
+            large_bytes_uploaded,
+        })
+    }
+
+    pub(super) fn handle_task_completed(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        terminal_state: Option<TaskTerminalState>,
+        status_code: Option<i32>,
+        stdout_bytes: u64,
+        stderr_bytes: u64,
+        stdout_tail: String,
+        stderr_tail: String,
+        stdout_truncated: bool,
+        stderr_truncated: bool,
+        artifact_path: Option<String>,
+        artifact_digest: Option<Digest>,
+        artifact_size_bytes: Option<u64>,
+        result: Option<TaskBoundaryValue>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        validate_task_log_tail("stdout_tail", &stdout_tail)?;
+        validate_task_log_tail("stderr_tail", &stderr_tail)?;
+        let artifact_path = artifact_path
+            .map(VfsPath::new)
+            .transpose()
+            .map_err(|err| CoordinatorServiceError::InvalidArtifactPath(err.to_string()))?;
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        let checkpoint = self
+            .task_restart_checkpoints
+            .get(&super::keys::task_restart_key(
+                &tenant, &project, &process, &task,
+            ))
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "signed node task completion does not name a coordinator-issued task instance"
+                        .to_owned(),
+                )
+            })?;
+        if checkpoint.assignment.node != node {
+            return Err(CoordinatorError::Unauthorized(
+                "signed node task completion came from a node other than the assigned node"
+                    .to_owned(),
+            )
+            .into());
+        }
+        let mut event = TaskCompletionEvent {
+            tenant,
+            project,
+            process,
+            node,
+            executor: super::TaskExecutor::Node,
+            task_definition: checkpoint.assignment.task_spec.task_definition.clone(),
+            task,
+            placement: None,
+            terminal_state: terminal_state
+                .unwrap_or_else(|| TaskTerminalState::from_status_code(status_code)),
+            status_code,
+            stdout_bytes,
+            stderr_bytes,
+            stdout_tail,
+            stderr_tail,
+            stdout_truncated,
+            stderr_truncated,
+            artifact_path,
+            artifact_digest: artifact_digest.clone(),
+            artifact_size_bytes,
+            result,
+        };
+        let reported_bytes = checked_reported_log_bytes(event.stdout_bytes, event.stderr_bytes)?;
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.quota.can_charge_log_bytes(
+            &event.tenant,
+            &event.project,
+            reported_bytes,
+            now_epoch_seconds,
+        )?;
+        self.quota.charge_log_bytes(
+            &event.tenant,
+            &event.project,
+            reported_bytes,
+            now_epoch_seconds,
+        )?;
+        let task_key = task_control_key(
+            &event.tenant,
+            &event.project,
+            &event.process,
+            &event.node,
+            &event.task,
+        );
+        let process_key = process_control_key(&event.tenant, &event.project, &event.process);
+        let process_was_aborted = self.process_aborts.contains(&process_key);
+        event.placement = self.task_placements.remove(&task_key);
+        if let (Some(path), Some(digest)) = (&event.artifact_path, artifact_digest) {
+            self.flush_artifact_metadata(ArtifactFlush {
+                id: artifact_id_from_path(path),
+                tenant: event.tenant.clone(),
+                project: event.project.clone(),
+                process: event.process.clone(),
+                producer_task: event.task.clone(),
+                retaining_node: event.node.clone(),
+                digest,
+                size: artifact_size_bytes.unwrap_or(stdout_bytes),
+            })?;
+        }
+        self.task_cancellations.remove(&task_key);
+        self.task_aborts.remove(&task_key);
+        self.debug_commands.remove(&task_key);
+        self.active_tasks.remove(&task_key);
+        if !self
+            .active_tasks
+            .iter()
+            .any(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant == &event.tenant
+                    && task_project == &event.project
+                    && task_process == &event.process
+            })
+        {
+            self.process_aborts.remove(&process_key);
+        }
+        if process_was_aborted {
+            let checkpoint_key = super::keys::task_restart_key(
+                &event.tenant,
+                &event.project,
+                &event.process,
+                &event.task,
+            );
+            self.task_restart_checkpoints.remove(&checkpoint_key);
+            self.task_restart_checkpoint_order
+                .retain(|retained| retained != &checkpoint_key);
+        }
+        self.record_task_completion_event(event.clone());
+        self.notify_coordinator_main_waiters(&event);
+        Ok(CoordinatorResponse::TaskRecorded {
+            process: event.process,
+            task: event.task,
+            events_recorded: self.task_events.len(),
+        })
+    }
+
+    pub(super) fn handle_list_task_events(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: Option<String>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let _actor = UserId::new(actor_user);
+        let process = process.map(ProcessId::new);
+        if let Some(process) = &process {
+            self.authorize_task_event_process_scope(&tenant, &project, process)?;
+        }
+        let events = self
+            .task_events
+            .iter()
+            .filter(|event| {
+                event.tenant == tenant
+                    && event.project == project
+                    && process
+                        .as_ref()
+                        .is_none_or(|process| event.process == *process)
+            })
+            .cloned()
+            .collect();
+        Ok(CoordinatorResponse::TaskEvents { events })
+    }
+
+    fn authorize_task_event_process_scope(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) -> Result<(), CoordinatorServiceError> {
+        let active_in_scope = self
+            .coordinator
+            .active_process(tenant, project, process)
+            .is_some();
+        let historical_in_scope = self.task_events.iter().any(|event| {
+            event.tenant == *tenant && event.project == *project && event.process == *process
+        });
+        let process_exists_outside_scope = self
+            .coordinator
+            .active_process_exists_outside_scope(tenant, project, process)
+            || self.task_events.iter().any(|event| {
+                event.process == *process && (event.tenant != *tenant || event.project != *project)
+            });
+        if !active_in_scope && !historical_in_scope && process_exists_outside_scope {
+            return Err(CoordinatorError::Unauthorized(
+                "task event access is outside the virtual process tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        Ok(())
+    }
+
+    pub(super) fn handle_join_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        task: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let _actor = UserId::new(actor_user);
+        let process = ProcessId::new(process);
+        let task = TaskInstanceId::new(task);
+        Ok(CoordinatorResponse::TaskJoined {
+            join: self.task_join_result(tenant, project, process, task),
+        })
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_join_child_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        parent_task: String,
+        task: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let parent_task = TaskInstanceId::new(parent_task);
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        if !self.active_tasks.contains(&super::keys::task_control_key(
+            &tenant,
+            &project,
+            &process,
+            &node,
+            &parent_task,
+        )) {
+            return Err(CoordinatorError::Unauthorized(
+                "child task join requires a currently active parent task on the signed node"
+                    .to_owned(),
+            )
+            .into());
+        }
+        Ok(CoordinatorResponse::TaskJoined {
+            join: self.task_join_result(tenant, project, process, TaskInstanceId::new(task)),
+        })
+    }
+
+    pub(super) fn task_join_result(
+        &self,
+        tenant: TenantId,
+        project: ProjectId,
+        process: ProcessId,
+        task: TaskInstanceId,
+    ) -> TaskJoinResult {
+        let event = self.task_events.iter().rev().find(|event| {
+            event.tenant == tenant
+                && event.project == project
+                && event.process == process
+                && event.task == task
+        });
+
+        if let Some(event) = event {
+            TaskJoinResult::from_remote_completion(
+                event.process.clone(),
+                event.task.clone(),
+                event.node.clone(),
+                join_state_for_terminal(&event.terminal_state),
+                event.result.clone(),
+                event.status_code,
+                join_message_for_event(event),
+            )
+        } else {
+            let known = self.task_is_known_or_active(&tenant, &project, &process, &task);
+            TaskJoinResult::pending(
+                process,
+                task,
+                if known {
+                    "waiting for signed node task_completed event before join returns"
+                } else {
+                    "no signed node completion event has been observed for this task"
+                },
+            )
+        }
+    }
+
+    pub(super) fn record_task_completion_event(&mut self, mut event: TaskCompletionEvent) {
+        event.stdout_tail = bounded_log_tail(event.stdout_tail, &mut event.stdout_truncated);
+        event.stderr_tail = bounded_log_tail(event.stderr_tail, &mut event.stderr_truncated);
+        while self
+            .task_events
+            .iter()
+            .filter(|retained| {
+                retained.tenant == event.tenant
+                    && retained.project == event.project
+                    && retained.process == event.process
+            })
+            .count()
+            >= super::MAX_TASK_EVENTS_PER_PROCESS
+        {
+            let Some(index) = self.task_events.iter().position(|retained| {
+                retained.tenant == event.tenant
+                    && retained.project == event.project
+                    && retained.process == event.process
+            }) else {
+                break;
+            };
+            self.task_events.remove(index);
+        }
+        self.task_events.push_back(event);
+    }
+
+    fn flush_artifact_metadata(
+        &mut self,
+        flush: ArtifactFlush,
+    ) -> Result<(), CoordinatorServiceError> {
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.artifact_registry
+            .expire_download_links(now_epoch_seconds);
+        let pinned = self
+            .task_restart_checkpoints
+            .values()
+            .flat_map(|checkpoint| checkpoint.assignment.task_spec.required_artifacts.iter())
+            .chain(
+                self.pending_task_launches
+                    .iter()
+                    .flat_map(|pending| pending.task_spec.required_artifacts.iter()),
+            )
+            .cloned()
+            .collect::<std::collections::BTreeSet<ArtifactId>>();
+        self.artifact_registry
+            .flush_metadata_bounded(flush, &pinned)
+            .map(|_| ())
+            .map_err(CoordinatorServiceError::Protocol)
+    }
+
+    fn task_is_known_or_active(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+        task: &TaskInstanceId,
+    ) -> bool {
+        self.active_tasks
+            .iter()
+            .any(|(task_tenant, task_project, task_process, _, task_id)| {
+                task_tenant == tenant
+                    && task_project == project
+                    && task_process == process
+                    && task_id == task
+            })
+            || self.pending_task_launches.iter().any(|pending| {
+                &pending.tenant == tenant
+                    && &pending.project == project
+                    && &pending.process == process
+                    && &pending.task == task
+            })
+            || self.task_assignments.values().any(|assignments| {
+                assignments.iter().any(|assignment| {
+                    &assignment.tenant == tenant
+                        && &assignment.project == project
+                        && &assignment.process == process
+                        && &assignment.task == task
+                })
+            })
+    }
+}
+
+fn checked_reported_log_bytes(
+    stdout_bytes: u64,
+    stderr_bytes: u64,
+) -> Result<u64, CoordinatorServiceError> {
+    stdout_bytes.checked_add(stderr_bytes).ok_or_else(|| {
+        CoordinatorServiceError::Protocol(
+            "reported task log byte counts exceed the supported range".to_owned(),
+        )
+    })
+}
+
+fn validate_task_log_tail(kind: &str, value: &str) -> Result<(), CoordinatorServiceError> {
+    if value.len() > MAX_TASK_LOG_TAIL_BYTES {
+        return Err(CoordinatorServiceError::InvalidTaskLogTail(format!(
+            "{kind} is {} bytes; max is {MAX_TASK_LOG_TAIL_BYTES}",
+            value.len()
+        )));
+    }
+    Ok(())
+}
+
+fn bounded_log_tail(mut value: String, truncated: &mut bool) -> String {
+    if value.len() <= MAX_TASK_LOG_TAIL_BYTES {
+        return value;
+    }
+    let mut boundary = MAX_TASK_LOG_TAIL_BYTES;
+    while !value.is_char_boundary(boundary) {
+        boundary -= 1;
+    }
+    value.truncate(boundary);
+    *truncated = true;
+    value
+}
+
+fn join_state_for_terminal(terminal: &TaskTerminalState) -> TaskJoinState {
+    match terminal {
+        TaskTerminalState::Completed => TaskJoinState::Completed,
+        TaskTerminalState::Failed => TaskJoinState::Failed,
+        TaskTerminalState::Cancelled => TaskJoinState::Cancelled,
+    }
+}
+
+fn join_message_for_event(event: &TaskCompletionEvent) -> String {
+    match event.terminal_state {
+        TaskTerminalState::Completed => {
+            "joined result from signed node task_completed event".to_owned()
+        }
+        TaskTerminalState::Failed => {
+            let stderr = event.stderr_tail.trim();
+            if stderr.is_empty() {
+                "remote task failed".to_owned()
+            } else {
+                format!("remote task failed: {stderr}")
+            }
+        }
+        TaskTerminalState::Cancelled => "remote task was cancelled".to_owned(),
+    }
+}

crates/disasmer-coordinator/src/service/main_runtime.rs

@@ -0,0 +1,999 @@
+use std::collections::{BTreeMap, BTreeSet, HashMap};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::mpsc::{self, Receiver, Sender, SyncSender, TryRecvError};
+use std::sync::{Arc, Mutex};
+
+use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine as _};
+use disasmer_core::{
+    Capability, CredentialKind, Digest, EnvironmentResource, NodeId, ProcessId, ProjectId,
+    TaskBoundaryValue, TaskDefinitionId, TaskDispatch, TaskInstanceId, TaskJoinState, TaskSpec,
+    TenantId, WasmExportAbi, WasmHostCommandRequest, WasmHostCommandResult,
+    WasmHostDebugProbeRequest, WasmHostDebugProbeResult, WasmHostSourceSnapshotRequest,
+    WasmHostSourceSnapshotResult, WasmHostTaskControlRequest, WasmHostTaskControlResult,
+    WasmHostTaskHandle, WasmHostTaskJoinRequest, WasmHostTaskJoinResult, WasmHostTaskStartRequest,
+    WasmHostVfsRequest, WasmHostVfsResult, WasmTaskInvocation, WasmTaskOutcome, WasmTaskResult,
+};
+use disasmer_wasm_runtime::{WasmDebugControl, WasmTaskHost, WasmtimeTaskRuntime};
+use wasmparser::{Parser, Payload};
+
+use crate::{CoordinatorError, CoordinatorServiceError};
+
+use super::keys::{process_control_key, task_restart_key, ProcessControlKey, TaskRestartKey};
+use super::{
+    CoordinatorResponse, CoordinatorService, TaskCompletionEvent, TaskExecutor, TaskTerminalState,
+    WorkflowActor,
+};
+
+#[derive(Clone)]
+struct MainScope {
+    tenant: TenantId,
+    project: ProjectId,
+    process: ProcessId,
+    task_definition: TaskDefinitionId,
+    task_instance: TaskInstanceId,
+    epoch: u64,
+    launch_id: u64,
+}
+
+enum MainCommand {
+    StartTask {
+        scope: MainScope,
+        handle_id: u64,
+        task_spec: TaskSpec,
+        wasm_module_base64: String,
+        response: SyncSender<Result<WasmHostTaskHandle, String>>,
+    },
+    JoinTask {
+        scope: MainScope,
+        task_instance: TaskInstanceId,
+        response: SyncSender<Result<WasmHostTaskJoinResult, String>>,
+    },
+    DebugProbe {
+        scope: MainScope,
+        request: WasmHostDebugProbeRequest,
+        response: SyncSender<Result<WasmHostDebugProbeResult, String>>,
+    },
+    Finished {
+        scope: MainScope,
+        result: Result<WasmTaskResult, String>,
+    },
+}
+
+pub(super) struct CoordinatorMainControl {
+    pub(super) task_definition: TaskDefinitionId,
+    pub(super) task_instance: TaskInstanceId,
+    pub(super) abort: Arc<AtomicBool>,
+    pub(super) debug: Arc<WasmDebugControl>,
+    pub(super) state: String,
+    pub(super) stopped_probe_symbol: Option<String>,
+    pub(super) handles: Arc<Mutex<HashMap<u64, TaskSpec>>>,
+    pub(super) launch_id: u64,
+}
+
+pub(super) struct CoordinatorMainRuntime {
+    sender: Sender<MainCommand>,
+    receiver: Receiver<MainCommand>,
+    pub(super) controls: BTreeMap<ProcessControlKey, CoordinatorMainControl>,
+    join_waiters: BTreeMap<TaskRestartKey, Vec<SyncSender<Result<WasmHostTaskJoinResult, String>>>>,
+    next_launch_id: u64,
+}
+
+impl Default for CoordinatorMainRuntime {
+    fn default() -> Self {
+        let (sender, receiver) = mpsc::channel();
+        Self {
+            sender,
+            receiver,
+            controls: BTreeMap::new(),
+            join_waiters: BTreeMap::new(),
+            next_launch_id: 1,
+        }
+    }
+}
+
+impl CoordinatorMainRuntime {
+    pub(super) fn is_waiting_for_task(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) -> bool {
+        self.join_waiters
+            .keys()
+            .any(|(waiter_tenant, waiter_project, waiter_process, _)| {
+                waiter_tenant == tenant && waiter_project == project && waiter_process == process
+            })
+    }
+
+    fn drain_commands(&self) -> Vec<MainCommand> {
+        let mut commands = Vec::new();
+        loop {
+            match self.receiver.try_recv() {
+                Ok(command) => commands.push(command),
+                Err(TryRecvError::Empty | TryRecvError::Disconnected) => break,
+            }
+        }
+        commands
+    }
+
+    fn launch(
+        &mut self,
+        mut scope: MainScope,
+        export: String,
+        module: Vec<u8>,
+        wasm_module_base64: String,
+        bundle_digest: Digest,
+        task_descriptors: HashMap<String, serde_json::Value>,
+        environments: BTreeMap<String, EnvironmentResource>,
+    ) -> Result<(), CoordinatorServiceError> {
+        let process_key = process_control_key(&scope.tenant, &scope.project, &scope.process);
+        if self.controls.contains_key(&process_key) {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "virtual process {} already has a coordinator main instance",
+                scope.process
+            )));
+        }
+        scope.launch_id = self.next_launch_id;
+        self.next_launch_id = self.next_launch_id.saturating_add(1);
+        let abort = Arc::new(AtomicBool::new(false));
+        let debug = Arc::new(WasmDebugControl::default());
+        let handles = Arc::new(Mutex::new(HashMap::new()));
+        self.controls.insert(
+            process_key,
+            CoordinatorMainControl {
+                task_definition: scope.task_definition.clone(),
+                task_instance: scope.task_instance.clone(),
+                abort: Arc::clone(&abort),
+                debug: Arc::clone(&debug),
+                state: "running".to_owned(),
+                stopped_probe_symbol: None,
+                handles: Arc::clone(&handles),
+                launch_id: scope.launch_id,
+            },
+        );
+        let sender = self.sender.clone();
+        let invocation = WasmTaskInvocation::new(
+            scope.task_definition.clone(),
+            scope.task_instance.clone(),
+            Vec::new(),
+        );
+        std::thread::Builder::new()
+            .name(format!("disasmer-main-{}", scope.process))
+            .spawn(move || {
+                let host = CoordinatorMainHost {
+                    scope: scope.clone(),
+                    sender: sender.clone(),
+                    abort,
+                    debug,
+                    task_descriptors,
+                    environments,
+                    bundle_digest: bundle_digest.clone(),
+                    wasm_module_base64,
+                    next_handle_id: 1,
+                    handles,
+                };
+                let result = WasmtimeTaskRuntime::new()
+                    .and_then(|runtime| {
+                        runtime.run_task_export_verified_with_task_host(
+                            &module,
+                            &bundle_digest,
+                            &export,
+                            &invocation,
+                            Box::new(host),
+                        )
+                    })
+                    .map_err(|error| error.to_string());
+                let _ = sender.send(MainCommand::Finished { scope, result });
+            })
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "start coordinator main runtime thread: {error}"
+                ))
+            })?;
+        Ok(())
+    }
+
+    pub(super) fn interrupt_process(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+        reason: &str,
+    ) {
+        let process_key = process_control_key(tenant, project, process);
+        if let Some(control) = self.controls.get_mut(&process_key) {
+            control.abort.store(true, Ordering::Release);
+            control.state = "stopping".to_owned();
+        }
+        let waiter_keys = self
+            .join_waiters
+            .keys()
+            .filter(|(waiter_tenant, waiter_project, waiter_process, _)| {
+                waiter_tenant == tenant && waiter_project == project && waiter_process == process
+            })
+            .cloned()
+            .collect::<Vec<_>>();
+        for key in waiter_keys {
+            if let Some(waiters) = self.join_waiters.remove(&key) {
+                for waiter in waiters {
+                    let _ = waiter.send(Err(reason.to_owned()));
+                }
+            }
+        }
+    }
+
+    fn is_current_scope(&self, scope: &MainScope) -> bool {
+        self.controls
+            .get(&process_control_key(
+                &scope.tenant,
+                &scope.project,
+                &scope.process,
+            ))
+            .is_some_and(|control| control.launch_id == scope.launch_id)
+    }
+}
+
+struct CoordinatorMainHost {
+    scope: MainScope,
+    sender: Sender<MainCommand>,
+    abort: Arc<AtomicBool>,
+    debug: Arc<WasmDebugControl>,
+    task_descriptors: HashMap<String, serde_json::Value>,
+    environments: BTreeMap<String, EnvironmentResource>,
+    bundle_digest: Digest,
+    wasm_module_base64: String,
+    next_handle_id: u64,
+    handles: Arc<Mutex<HashMap<u64, TaskSpec>>>,
+}
+
+impl WasmTaskHost for CoordinatorMainHost {
+    fn abort_signal(&self) -> Option<Arc<AtomicBool>> {
+        Some(Arc::clone(&self.abort))
+    }
+
+    fn debug_control(&self) -> Option<Arc<WasmDebugControl>> {
+        Some(Arc::clone(&self.debug))
+    }
+
+    fn start_task(
+        &mut self,
+        request: WasmHostTaskStartRequest,
+    ) -> Result<WasmHostTaskHandle, String> {
+        request.validate()?;
+        if self.abort.load(Ordering::Acquire) {
+            return Err("coordinator main is stopping".to_owned());
+        }
+        let descriptor = self
+            .task_descriptors
+            .get(request.task_definition.as_str())
+            .ok_or_else(|| {
+                format!(
+                    "bundle has no task descriptor named `{}`",
+                    request.task_definition
+                )
+            })?;
+        let export = descriptor
+            .get("export")
+            .and_then(serde_json::Value::as_str)
+            .filter(|export| !export.trim().is_empty())
+            .ok_or_else(|| {
+                format!(
+                    "task `{}` descriptor omitted its Wasm export",
+                    request.task_definition
+                )
+            })?;
+        let mut required_capabilities = descriptor
+            .get("required_capabilities")
+            .and_then(serde_json::Value::as_array)
+            .into_iter()
+            .flatten()
+            .map(|value| {
+                capability_from_descriptor(
+                    value
+                        .as_str()
+                        .ok_or("task capability descriptor is not a string")?,
+                )
+            })
+            .collect::<Result<BTreeSet<_>, _>>()?;
+        let selected_environment = request
+            .environment_id
+            .as_deref()
+            .map(|environment| {
+                self.environments.get(environment).cloned().ok_or_else(|| {
+                    format!("bundle environment manifest has no environment `{environment}`")
+                })
+            })
+            .transpose()?;
+        let environment = selected_environment
+            .as_ref()
+            .map(|environment| environment.requirements.clone());
+        let environment_digest = selected_environment
+            .as_ref()
+            .map(|environment| environment.digest.clone());
+        if let Some(environment) = &environment {
+            required_capabilities.extend(environment.capabilities.iter().cloned());
+        }
+        let required_artifacts = request
+            .args
+            .iter()
+            .flat_map(TaskBoundaryValue::required_artifacts)
+            .collect::<BTreeSet<_>>()
+            .into_iter()
+            .collect::<Vec<_>>();
+        let source_snapshots = request
+            .args
+            .iter()
+            .flat_map(TaskBoundaryValue::source_snapshots)
+            .collect::<BTreeSet<_>>();
+        if source_snapshots.len() > 1 {
+            return Err(
+                "one task invocation cannot require multiple distinct source snapshots".to_owned(),
+            );
+        }
+        if self
+            .handles
+            .lock()
+            .map_err(|_| "coordinator main handle registry is unavailable".to_owned())?
+            .len()
+            >= super::MAX_IN_FLIGHT_TASKS_PER_PROCESS
+        {
+            return Err(format!(
+                "coordinator main task-handle limit of {} reached",
+                super::MAX_IN_FLIGHT_TASKS_PER_PROCESS
+            ));
+        }
+        let handle_id = self.next_handle_id;
+        let task_instance =
+            TaskInstanceId::new(format!("{}:child:{handle_id}", self.scope.task_instance));
+        let task_spec = TaskSpec {
+            tenant: self.scope.tenant.clone(),
+            project: self.scope.project.clone(),
+            process: self.scope.process.clone(),
+            task_definition: request.task_definition,
+            task_instance,
+            dispatch: TaskDispatch::CoordinatorNodeWasm {
+                export: Some(export.to_owned()),
+                abi: WasmExportAbi::TaskV1,
+            },
+            environment_id: request.environment_id,
+            environment,
+            environment_digest,
+            required_capabilities,
+            dependency_cache: None,
+            source_snapshot: source_snapshots.into_iter().next(),
+            required_artifacts,
+            args: request.args,
+            vfs_epoch: self.scope.epoch,
+            bundle_digest: Some(self.bundle_digest.clone()),
+        };
+        let (response, receiver) = mpsc::sync_channel(1);
+        self.sender
+            .send(MainCommand::StartTask {
+                scope: self.scope.clone(),
+                handle_id,
+                task_spec: task_spec.clone(),
+                wasm_module_base64: self.wasm_module_base64.clone(),
+                response,
+            })
+            .map_err(|_| "coordinator main command channel closed".to_owned())?;
+        let handle = receiver
+            .recv()
+            .map_err(|_| "coordinator main task-start response channel closed".to_owned())??;
+        self.handles
+            .lock()
+            .map_err(|_| "coordinator main handle registry is unavailable".to_owned())?
+            .insert(handle_id, task_spec);
+        self.next_handle_id = self.next_handle_id.saturating_add(1);
+        Ok(handle)
+    }
+
+    fn join_task(
+        &mut self,
+        request: WasmHostTaskJoinRequest,
+    ) -> Result<WasmHostTaskJoinResult, String> {
+        let task_spec = self
+            .handles
+            .lock()
+            .map_err(|_| "coordinator main handle registry is unavailable".to_owned())?
+            .get(&request.handle_id)
+            .cloned()
+            .ok_or_else(|| format!("unknown Wasm task handle {}", request.handle_id))?;
+        let (response, receiver) = mpsc::sync_channel(1);
+        self.sender
+            .send(MainCommand::JoinTask {
+                scope: self.scope.clone(),
+                task_instance: task_spec.task_instance,
+                response,
+            })
+            .map_err(|_| "coordinator main command channel closed".to_owned())?;
+        let joined = receiver
+            .recv()
+            .map_err(|_| "coordinator main task-join response channel closed".to_owned())?;
+        self.handles
+            .lock()
+            .map_err(|_| "coordinator main handle registry is unavailable".to_owned())?
+            .remove(&request.handle_id);
+        joined
+    }
+
+    fn run_command(
+        &mut self,
+        _request: WasmHostCommandRequest,
+    ) -> Result<WasmHostCommandResult, String> {
+        Err("coordinator main is capless and cannot run native commands".to_owned())
+    }
+
+    fn poll_task_control(
+        &mut self,
+        request: WasmHostTaskControlRequest,
+    ) -> Result<WasmHostTaskControlResult, String> {
+        request.validate()?;
+        Ok(WasmHostTaskControlResult {
+            abi_version: disasmer_core::WASM_TASK_ABI_VERSION,
+            cancellation_requested: self.abort.load(Ordering::Acquire),
+        })
+    }
+
+    fn debug_probe(
+        &mut self,
+        request: WasmHostDebugProbeRequest,
+    ) -> Result<WasmHostDebugProbeResult, String> {
+        request.validate()?;
+        let (response, receiver) = mpsc::sync_channel(1);
+        self.sender
+            .send(MainCommand::DebugProbe {
+                scope: self.scope.clone(),
+                request,
+                response,
+            })
+            .map_err(|_| "coordinator main command channel closed".to_owned())?;
+        receiver
+            .recv()
+            .map_err(|_| "coordinator main debug-probe response channel closed".to_owned())?
+    }
+
+    fn vfs_operation(&mut self, _request: WasmHostVfsRequest) -> Result<WasmHostVfsResult, String> {
+        Err("coordinator main is capless and cannot access task VFS files".to_owned())
+    }
+
+    fn snapshot_source(
+        &mut self,
+        _request: WasmHostSourceSnapshotRequest,
+    ) -> Result<WasmHostSourceSnapshotResult, String> {
+        Err("coordinator main is capless and cannot access source checkouts".to_owned())
+    }
+}
+
+impl CoordinatorService {
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_launch_coordinator_main(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: Option<String>,
+        actor_agent: Option<String>,
+        agent_public_key_fingerprint: Option<Digest>,
+        agent_signature: Option<disasmer_core::AgentSignedRequest>,
+        request_payload_digest: Option<&Digest>,
+        task_spec: TaskSpec,
+        wasm_module_base64: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = task_spec.process.clone();
+        let task_instance = task_spec.task_instance.clone();
+        let actor = self.workflow_actor(
+            &tenant,
+            &project,
+            actor_user,
+            actor_agent,
+            agent_public_key_fingerprint,
+            agent_signature,
+            request_payload_digest,
+            "launch_task",
+            &process,
+            Some(&task_instance),
+        )?;
+        let active = self
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "coordinator main launch requires an active virtual process".to_owned(),
+                )
+            })?;
+        debug_assert_eq!(active.tenant, tenant);
+        debug_assert_eq!(active.project, project);
+        if task_spec.tenant != tenant || task_spec.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "coordinator main TaskSpec is outside the authenticated scope".to_owned(),
+            )
+            .into());
+        }
+        let export = match &task_spec.dispatch {
+            TaskDispatch::CoordinatorNodeWasm {
+                export: Some(export),
+                abi: WasmExportAbi::EntrypointV1,
+            } => export.clone(),
+            _ => {
+                return Err(CoordinatorServiceError::Protocol(
+                    "coordinator main requires an explicit EntrypointV1 Wasm export".to_owned(),
+                ))
+            }
+        };
+        if task_spec.environment_id.is_some()
+            || task_spec.environment.is_some()
+            || task_spec.environment_digest.is_some()
+            || !task_spec.required_capabilities.is_empty()
+            || task_spec.dependency_cache.is_some()
+            || task_spec.source_snapshot.is_some()
+            || !task_spec.required_artifacts.is_empty()
+            || !task_spec.args.is_empty()
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "coordinator main must be capless and may not receive environment, source, artifact, cache, or argument authority"
+                    .to_owned(),
+            )
+            .into());
+        }
+        let bundle_digest = task_spec.bundle_digest.clone().ok_or_else(|| {
+            CoordinatorServiceError::Protocol(
+                "coordinator main TaskSpec omitted bundle digest".to_owned(),
+            )
+        })?;
+        let module = BASE64_STANDARD
+            .decode(&wasm_module_base64)
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "coordinator main module is not valid base64: {error}"
+                ))
+            })?;
+        let actual_digest = Digest::sha256(&module);
+        if actual_digest != bundle_digest {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "coordinator main module digest mismatch: expected {bundle_digest}, actual {actual_digest}"
+            ))
+            .into());
+        }
+        WasmTaskInvocation::new(
+            task_spec.task_definition.clone(),
+            task_instance.clone(),
+            Vec::new(),
+        )
+        .validate()
+        .map_err(CoordinatorServiceError::Protocol)?;
+        let descriptors = task_descriptors(&module)?;
+        let environments = bundle_environments(&module)?;
+        let scope = MainScope {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            process: process.clone(),
+            task_definition: task_spec.task_definition.clone(),
+            task_instance: task_instance.clone(),
+            epoch: task_spec.vfs_epoch,
+            launch_id: 0,
+        };
+        self.main_runtime.launch(
+            scope,
+            export,
+            module,
+            wasm_module_base64,
+            bundle_digest,
+            descriptors,
+            environments,
+        )?;
+        Ok(CoordinatorResponse::MainLaunched {
+            process,
+            task_definition: task_spec.task_definition,
+            task_instance,
+            actor,
+            state: "running".to_owned(),
+        })
+    }
+
+    pub(super) fn pump_main_runtime_commands(&mut self) {
+        let commands = self.main_runtime.drain_commands();
+        for command in commands {
+            match command {
+                MainCommand::StartTask {
+                    scope,
+                    handle_id,
+                    task_spec,
+                    wasm_module_base64,
+                    response,
+                } => {
+                    if !self.main_runtime.is_current_scope(&scope) {
+                        let _ = response.send(Err(
+                            "coordinator main process incarnation was replaced".to_owned(),
+                        ));
+                        continue;
+                    }
+                    let actor = WorkflowActor {
+                        kind: "task".to_owned(),
+                        user: None,
+                        agent: None,
+                        credential_kind: CredentialKind::TaskCredential,
+                        public_key_fingerprint: None,
+                        authenticated_without_browser: true,
+                        scopes: vec!["process:spawn-child".to_owned()],
+                    };
+                    let result = self
+                        .handle_launch_task_with_actor(
+                            scope.tenant,
+                            scope.project,
+                            actor,
+                            task_spec.clone(),
+                            true,
+                            format!("/vfs/artifacts/{}-result.json", task_spec.task_instance),
+                            wasm_module_base64,
+                        )
+                        .and_then(|launch| match launch {
+                            CoordinatorResponse::TaskLaunched { .. }
+                            | CoordinatorResponse::TaskQueued { .. } => Ok(WasmHostTaskHandle {
+                                abi_version: disasmer_core::WASM_TASK_ABI_VERSION,
+                                handle_id,
+                                task_spec,
+                            }),
+                            other => Err(CoordinatorServiceError::Protocol(format!(
+                                "unexpected coordinator-main child launch response: {other:?}"
+                            ))),
+                        })
+                        .map_err(|error| error.to_string());
+                    let _ = response.send(result);
+                }
+                MainCommand::JoinTask {
+                    scope,
+                    task_instance,
+                    response,
+                } => {
+                    if !self.main_runtime.is_current_scope(&scope) {
+                        let _ = response.send(Err(
+                            "coordinator main process incarnation was replaced".to_owned(),
+                        ));
+                        continue;
+                    }
+                    let join = self.task_join_result(
+                        scope.tenant.clone(),
+                        scope.project.clone(),
+                        scope.process.clone(),
+                        task_instance.clone(),
+                    );
+                    match join.state {
+                        TaskJoinState::Completed => {
+                            let result = join
+                                .result
+                                .ok_or_else(|| {
+                                    "completed child task omitted its boundary result".to_owned()
+                                })
+                                .map(|result| WasmHostTaskJoinResult {
+                                    abi_version: disasmer_core::WASM_TASK_ABI_VERSION,
+                                    task_instance,
+                                    result,
+                                });
+                            let _ = response.send(result);
+                        }
+                        TaskJoinState::Failed | TaskJoinState::Cancelled => {
+                            let _ = response.send(Err(join.message));
+                        }
+                        TaskJoinState::Pending => {
+                            self.main_runtime
+                                .join_waiters
+                                .entry(task_restart_key(
+                                    &scope.tenant,
+                                    &scope.project,
+                                    &scope.process,
+                                    &task_instance,
+                                ))
+                                .or_default()
+                                .push(response);
+                        }
+                    }
+                }
+                MainCommand::DebugProbe {
+                    scope,
+                    request,
+                    response,
+                } => {
+                    if !self.main_runtime.is_current_scope(&scope) {
+                        let _ = response.send(Err(
+                            "coordinator main process incarnation was replaced".to_owned(),
+                        ));
+                        continue;
+                    }
+                    let result = self
+                        .handle_coordinator_main_debug_probe(
+                            scope.tenant,
+                            scope.project,
+                            scope.process,
+                            scope.task_instance,
+                            request.symbol,
+                        )
+                        .map_err(|error| error.to_string());
+                    let _ = response.send(result);
+                }
+                MainCommand::Finished { scope, result } => {
+                    if !self.main_runtime.is_current_scope(&scope) {
+                        continue;
+                    }
+                    self.record_coordinator_main_completion(scope, result);
+                }
+            }
+        }
+    }
+
+    pub(super) fn notify_coordinator_main_waiters(&mut self, event: &TaskCompletionEvent) {
+        let key = task_restart_key(&event.tenant, &event.project, &event.process, &event.task);
+        let Some(waiters) = self.main_runtime.join_waiters.remove(&key) else {
+            return;
+        };
+        let result = match event.terminal_state {
+            TaskTerminalState::Completed => event
+                .result
+                .clone()
+                .ok_or_else(|| "completed child task omitted its boundary result".to_owned())
+                .map(|result| WasmHostTaskJoinResult {
+                    abi_version: disasmer_core::WASM_TASK_ABI_VERSION,
+                    task_instance: event.task.clone(),
+                    result,
+                }),
+            TaskTerminalState::Failed | TaskTerminalState::Cancelled => {
+                Err(event.stderr_tail.clone())
+            }
+        };
+        for waiter in waiters {
+            let _ = waiter.send(result.clone());
+        }
+    }
+
+    fn record_coordinator_main_completion(
+        &mut self,
+        scope: MainScope,
+        result: Result<WasmTaskResult, String>,
+    ) {
+        let (terminal_state, boundary, error) = match result {
+            Ok(result) if result.outcome == WasmTaskOutcome::Completed => {
+                (TaskTerminalState::Completed, result.result, String::new())
+            }
+            Ok(result) => (
+                TaskTerminalState::Failed,
+                None,
+                result
+                    .error
+                    .unwrap_or_else(|| "coordinator main failed without an error".to_owned()),
+            ),
+            Err(error) => (TaskTerminalState::Failed, None, error),
+        };
+        let main_state = match terminal_state {
+            TaskTerminalState::Completed => "completed",
+            TaskTerminalState::Failed => "failed",
+            TaskTerminalState::Cancelled => "cancelled",
+        };
+        let event = TaskCompletionEvent {
+            tenant: scope.tenant.clone(),
+            project: scope.project.clone(),
+            process: scope.process.clone(),
+            node: NodeId::from("coordinator-main"),
+            executor: TaskExecutor::CoordinatorMain,
+            task_definition: scope.task_definition,
+            task: scope.task_instance,
+            placement: None,
+            terminal_state,
+            status_code: if error.is_empty() { Some(0) } else { None },
+            stdout_bytes: 0,
+            stderr_bytes: error.len() as u64,
+            stdout_tail: String::new(),
+            stderr_tail: error,
+            stdout_truncated: false,
+            stderr_truncated: false,
+            artifact_path: None,
+            artifact_digest: None,
+            artifact_size_bytes: None,
+            result: boundary,
+        };
+        self.record_task_completion_event(event);
+        if let Some(control) = self.main_runtime.controls.get_mut(&process_control_key(
+            &scope.tenant,
+            &scope.project,
+            &scope.process,
+        )) {
+            control.state = main_state.to_owned();
+            control.stopped_probe_symbol = None;
+            if let Ok(mut handles) = control.handles.lock() {
+                handles.clear();
+            }
+        }
+        // Completion keeps the virtual process and its final debug handshake
+        // inspectable. In particular, the DAP client may still be waiting for
+        // every participant's resume acknowledgement when the main Wasm task
+        // exits. A new process incarnation or an explicit abort clears this
+        // ephemeral state.
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn completed_main_keeps_process_and_debug_handshake_visible_until_explicit_abort() {
+        let mut service = CoordinatorService::new(7);
+        let tenant = TenantId::from("tenant");
+        let project = ProjectId::from("project");
+        let process = ProcessId::from("vp-current");
+        let main_task = TaskInstanceId::from("ti:vp-current:main");
+        let scope = MainScope {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            process: process.clone(),
+            task_definition: TaskDefinitionId::from("build"),
+            task_instance: main_task.clone(),
+            epoch: 7,
+            launch_id: 1,
+        };
+        service
+            .coordinator
+            .start_process(tenant.clone(), project.clone(), process.clone());
+        let process_key = process_control_key(&tenant, &project, &process);
+        service.main_runtime.controls.insert(
+            process_key.clone(),
+            CoordinatorMainControl {
+                task_definition: scope.task_definition.clone(),
+                task_instance: main_task.clone(),
+                abort: Arc::new(AtomicBool::new(false)),
+                debug: Arc::new(WasmDebugControl::default()),
+                state: "running".to_owned(),
+                stopped_probe_symbol: None,
+                handles: Arc::new(Mutex::new(HashMap::new())),
+                launch_id: 1,
+            },
+        );
+        service.debug_epochs.insert(process_key.clone(), 2);
+        service.debug_epoch_runtime.insert(
+            process_key.clone(),
+            super::super::debug::DebugEpochRuntime {
+                epoch: 2,
+                command: "resume".to_owned(),
+                expected: BTreeSet::new(),
+                acknowledgements: BTreeMap::new(),
+            },
+        );
+
+        service.record_coordinator_main_completion(
+            scope,
+            Ok(WasmTaskResult::completed(
+                main_task,
+                TaskBoundaryValue::SmallJson(serde_json::Value::Null),
+            )),
+        );
+
+        assert!(service
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .is_some());
+        let CoordinatorResponse::ProcessStatuses { processes, .. } = service
+            .handle_list_processes(
+                tenant.as_str().to_owned(),
+                project.as_str().to_owned(),
+                "user".to_owned(),
+            )
+            .unwrap()
+        else {
+            panic!("expected process statuses");
+        };
+        assert_eq!(processes.len(), 1);
+        assert_eq!(processes[0].state, "completed");
+        assert_eq!(processes[0].main_state.as_deref(), Some("completed"));
+        assert_eq!(service.debug_epochs.get(&process_key), Some(&2));
+        assert_eq!(
+            service
+                .debug_epoch_runtime
+                .get(&process_key)
+                .map(|runtime| runtime.command.as_str()),
+            Some("resume")
+        );
+
+        service
+            .handle_abort_process(
+                tenant.as_str().to_owned(),
+                project.as_str().to_owned(),
+                "user".to_owned(),
+                process.as_str().to_owned(),
+            )
+            .unwrap();
+        assert!(service
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .is_none());
+        assert!(!service.main_runtime.controls.contains_key(&process_key));
+        assert!(!service.debug_epochs.contains_key(&process_key));
+        assert!(!service.debug_epoch_runtime.contains_key(&process_key));
+    }
+}
+
+pub(super) fn task_descriptors(
+    module: &[u8],
+) -> Result<HashMap<String, serde_json::Value>, CoordinatorServiceError> {
+    let mut descriptors = HashMap::new();
+    for payload in Parser::new(0).parse_all(module) {
+        let Payload::CustomSection(section) = payload.map_err(|error| {
+            CoordinatorServiceError::Protocol(format!("parse coordinator main bundle: {error}"))
+        })?
+        else {
+            continue;
+        };
+        if section.name() != "disasmer.tasks" {
+            continue;
+        }
+        for record in section
+            .data()
+            .split(|byte| *byte == b'\n' || *byte == 0)
+            .filter(|record| !record.is_empty())
+        {
+            let descriptor: serde_json::Value = serde_json::from_slice(record)?;
+            let name = descriptor
+                .get("name")
+                .and_then(serde_json::Value::as_str)
+                .ok_or_else(|| {
+                    CoordinatorServiceError::Protocol("task descriptor omitted its name".to_owned())
+                })?
+                .to_owned();
+            descriptors.insert(name, descriptor);
+        }
+    }
+    Ok(descriptors)
+}
+
+pub(super) fn bundle_environments(
+    module: &[u8],
+) -> Result<BTreeMap<String, EnvironmentResource>, CoordinatorServiceError> {
+    for payload in Parser::new(0).parse_all(module) {
+        let Payload::CustomSection(section) = payload.map_err(|error| {
+            CoordinatorServiceError::Protocol(format!(
+                "parse coordinator main environment manifest: {error}"
+            ))
+        })?
+        else {
+            continue;
+        };
+        if section.name() != "disasmer.environments" {
+            continue;
+        }
+        let environments: Vec<EnvironmentResource> = serde_json::from_slice(section.data())
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "bundle environment manifest is invalid: {error}"
+                ))
+            })?;
+        let mut by_name = BTreeMap::new();
+        for environment in environments {
+            if by_name
+                .insert(environment.name.clone(), environment)
+                .is_some()
+            {
+                return Err(CoordinatorServiceError::Protocol(
+                    "bundle environment manifest contains duplicate names".to_owned(),
+                ));
+            }
+        }
+        return Ok(by_name);
+    }
+    Ok(BTreeMap::new())
+}
+
+pub(super) fn capability_from_descriptor(capability: &str) -> Result<Capability, String> {
+    match capability.to_ascii_lowercase().as_str() {
+        "command" => Ok(Capability::Command),
+        "rootless_podman" => Ok(Capability::RootlessPodman),
+        "source_git" => Ok(Capability::SourceGit),
+        "source_filesystem" => Ok(Capability::SourceFilesystem),
+        "network" => Ok(Capability::Network),
+        "host_filesystem" => Ok(Capability::HostFilesystem),
+        "secrets" => Ok(Capability::Secrets),
+        "inbound_ports" => Ok(Capability::InboundPorts),
+        "arbitrary_syscalls" => Ok(Capability::ArbitrarySyscalls),
+        "vfs_artifacts" => Ok(Capability::VfsArtifacts),
+        "windows_command_dev" => Ok(Capability::WindowsCommandDev),
+        "quic_direct" => Ok(Capability::QuicDirect),
+        other => Err(format!("unknown task capability `{other}`")),
+    }
+}

crates/disasmer-coordinator/src/service/nodes.rs

@@ -0,0 +1,407 @@
+use std::collections::BTreeSet;
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use disasmer_core::{
+    generate_opaque_token, verify_node_request_signature, Actor, ArtifactId, CredentialKind,
+    Digest, NodeCapabilities, NodeDescriptor, NodeId, NodeSignedRequest, ProjectId,
+    SourceProviderKind, TenantId, UserId,
+};
+
+use crate::CoordinatorError;
+
+use super::{
+    bounded_ttl, enrollment_grant_key, CoordinatorResponse, CoordinatorService,
+    CoordinatorServiceError,
+};
+
+impl CoordinatorService {
+    pub(super) fn handle_attach_node(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        public_key: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        self.coordinator.ensure_tenant_active(&tenant)?;
+        self.coordinator.upsert_tenant(tenant.clone());
+        self.coordinator.upsert_user(
+            tenant.clone(),
+            UserId::from("local-user"),
+            CredentialKind::CliDeviceSession,
+        );
+        self.coordinator
+            .upsert_project(tenant.clone(), project.clone(), "local");
+        self.coordinator.upsert_source_provider_config(
+            tenant.clone(),
+            project.clone(),
+            SourceProviderKind::Filesystem,
+            Digest::sha256("local-filesystem"),
+        );
+        self.coordinator.enroll_node(
+            tenant.clone(),
+            project.clone(),
+            node.clone(),
+            public_key,
+            "node:attach",
+        );
+        self.persist_durable_state()?;
+        Ok(CoordinatorResponse::NodeAttached {
+            node,
+            tenant,
+            project,
+        })
+    }
+
+    pub(super) fn handle_create_node_enrollment_grant(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        ttl_seconds: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        self.coordinator.ensure_tenant_active(&tenant)?;
+        self.coordinator.upsert_tenant(tenant.clone());
+        self.coordinator
+            .upsert_user(tenant.clone(), actor, CredentialKind::CliDeviceSession);
+        self.coordinator
+            .upsert_project(tenant.clone(), project.clone(), "local");
+        self.coordinator.upsert_source_provider_config(
+            tenant.clone(),
+            project.clone(),
+            SourceProviderKind::Filesystem,
+            Digest::sha256("local-filesystem"),
+        );
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.enrollment_grants.retain(|_, grant| {
+            !grant.consumed && grant.expires_at_epoch_seconds >= now_epoch_seconds
+        });
+        if self
+            .enrollment_grants
+            .values()
+            .filter(|grant| grant.tenant == tenant && grant.project == project)
+            .count()
+            >= super::MAX_ENROLLMENT_GRANTS_PER_PROJECT
+        {
+            return Err(CoordinatorServiceError::Protocol(
+                "node enrollment grant limit reached for this project; consume a grant or wait for one to expire"
+                    .to_owned(),
+            ));
+        }
+        let grant =
+            generate_opaque_token("node_grant").map_err(CoordinatorServiceError::Protocol)?;
+        let ttl_seconds = bounded_ttl(ttl_seconds, self.admission.max_node_enrollment_ttl_seconds);
+        let scope = "node:attach".to_owned();
+        let expires_at_epoch_seconds = now_epoch_seconds.saturating_add(ttl_seconds);
+        let enrollment = self.coordinator.create_node_enrollment_grant(
+            tenant.clone(),
+            project.clone(),
+            grant.clone(),
+            scope.clone(),
+            expires_at_epoch_seconds,
+        );
+        self.enrollment_grants
+            .insert(enrollment_grant_key(&tenant, &project, &grant), enrollment);
+        self.persist_durable_state()?;
+        Ok(CoordinatorResponse::NodeEnrollmentGrantCreated {
+            tenant,
+            project,
+            grant,
+            scope,
+            expires_at_epoch_seconds,
+        })
+    }
+
+    pub(super) fn handle_exchange_node_enrollment_grant(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        public_key: String,
+        enrollment_grant: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.enrollment_grants.retain(|_, grant| {
+            !grant.consumed && grant.expires_at_epoch_seconds >= now_epoch_seconds
+        });
+        self.coordinator.ensure_tenant_active(&tenant)?;
+        let grant_key = enrollment_grant_key(&tenant, &project, &enrollment_grant);
+        let grant =
+            self.enrollment_grants
+                .get_mut(&grant_key)
+                .ok_or(CoordinatorError::Enrollment(
+                    disasmer_core::EnrollmentError::Expired,
+                ))?;
+        let credential = self.coordinator.exchange_node_enrollment_grant(
+            grant,
+            node.clone(),
+            &public_key,
+            "node:attach",
+            now_epoch_seconds,
+        )?;
+        self.enrollment_grants.remove(&grant_key);
+        self.persist_durable_state()?;
+        Ok(CoordinatorResponse::NodeEnrollmentExchanged {
+            node,
+            tenant,
+            project,
+            credential,
+        })
+    }
+
+    pub(super) fn handle_node_heartbeat(
+        &mut self,
+        node: String,
+        node_signature: Option<NodeSignedRequest>,
+        payload_digest: &Digest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let node = NodeId::new(node);
+        self.authenticate_node_request(&node, node_signature, "node_heartbeat", payload_digest)?;
+        Ok(CoordinatorResponse::NodeHeartbeat {
+            node,
+            epoch: self.coordinator.coordinator_epoch(),
+        })
+    }
+
+    pub(super) fn handle_report_node_capabilities(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        capabilities: NodeCapabilities,
+        cached_environment_digests: Vec<Digest>,
+        dependency_cache_digests: Vec<Digest>,
+        source_snapshots: Vec<Digest>,
+        artifact_locations: Vec<String>,
+        direct_connectivity: bool,
+        online: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        let identity = self
+            .coordinator
+            .node_identity(&node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        if identity.tenant != tenant || identity.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "node capability report is outside the enrolled tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        capabilities.validate_public_report()?;
+        for (kind, count) in [
+            ("cached environments", cached_environment_digests.len()),
+            ("dependency caches", dependency_cache_digests.len()),
+            ("source snapshots", source_snapshots.len()),
+            ("artifact locations", artifact_locations.len()),
+        ] {
+            if count > super::MAX_NODE_REPORTED_OBJECTS_PER_KIND {
+                return Err(CoordinatorServiceError::Protocol(format!(
+                    "node capability report contains {count} {kind}; limit is {}",
+                    super::MAX_NODE_REPORTED_OBJECTS_PER_KIND
+                )));
+            }
+        }
+        if cached_environment_digests
+            .iter()
+            .chain(&dependency_cache_digests)
+            .chain(&source_snapshots)
+            .any(|digest| !digest.is_valid_sha256())
+        {
+            return Err(CoordinatorServiceError::Protocol(
+                "node capability report contains an invalid digest".to_owned(),
+            ));
+        }
+        if artifact_locations.iter().any(|artifact| {
+            artifact.trim().is_empty()
+                || artifact.len() > 256
+                || artifact
+                    .chars()
+                    .any(|character| matches!(character, '/' | '\\' | '\0'))
+        }) {
+            return Err(CoordinatorServiceError::Protocol(
+                "node capability report contains an invalid artifact id".to_owned(),
+            ));
+        }
+        let artifact_locations = artifact_locations
+            .into_iter()
+            .map(ArtifactId::new)
+            .collect::<BTreeSet<_>>();
+        self.artifact_registry
+            .reconcile_node_retention(&node, &artifact_locations);
+
+        self.node_descriptors.insert(
+            node.clone(),
+            NodeDescriptor {
+                id: node.clone(),
+                tenant,
+                project,
+                capabilities,
+                cached_environments: cached_environment_digests.into_iter().collect(),
+                dependency_caches: dependency_cache_digests.into_iter().collect(),
+                source_snapshots: source_snapshots.into_iter().collect(),
+                artifact_locations,
+                direct_connectivity,
+                online,
+            },
+        );
+        Ok(CoordinatorResponse::NodeCapabilitiesRecorded {
+            node,
+            node_descriptors: self.node_descriptors.len(),
+        })
+    }
+
+    pub(super) fn handle_list_node_descriptors(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let descriptors = self
+            .node_descriptors
+            .values()
+            .filter(|descriptor| descriptor.tenant == tenant && descriptor.project == project)
+            .cloned()
+            .collect();
+        Ok(CoordinatorResponse::NodeDescriptors { descriptors, actor })
+    }
+
+    pub(super) fn handle_revoke_node_credential(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        node: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let node = NodeId::new(node);
+        let context = disasmer_core::AuthContext {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            actor: Actor::User(actor.clone()),
+        };
+        self.coordinator.revoke_node_credential(&context, &node)?;
+        let descriptor_removed = self.node_descriptors.remove(&node).is_some();
+        self.artifact_registry.garbage_collect_node(&node);
+        let queued_assignments_removed = self
+            .task_assignments
+            .remove(&(tenant.clone(), project.clone(), node.clone()))
+            .map_or(0, |assignments| assignments.len());
+        self.active_tasks
+            .retain(|(task_tenant, task_project, _, task_node, _)| {
+                task_tenant != &tenant || task_project != &project || task_node != &node
+            });
+        self.task_cancellations
+            .retain(|(task_tenant, task_project, _, task_node, _)| {
+                task_tenant != &tenant || task_project != &project || task_node != &node
+            });
+        self.task_aborts
+            .retain(|(task_tenant, task_project, _, task_node, _)| {
+                task_tenant != &tenant || task_project != &project || task_node != &node
+            });
+        self.task_placements
+            .retain(|(task_tenant, task_project, _, task_node, _), _| {
+                task_tenant != &tenant || task_project != &project || task_node != &node
+            });
+        self.persist_durable_state()?;
+        Ok(CoordinatorResponse::NodeCredentialRevoked {
+            node,
+            tenant,
+            project,
+            actor,
+            descriptor_removed,
+            queued_assignments_removed,
+        })
+    }
+
+    pub(super) fn authenticate_node_request(
+        &mut self,
+        node: &NodeId,
+        node_signature: Option<NodeSignedRequest>,
+        request_kind: &str,
+        payload_digest: &Digest,
+    ) -> Result<(), CoordinatorServiceError> {
+        let identity = self
+            .coordinator
+            .node_identity(node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        let signature = node_signature.ok_or_else(|| {
+            CoordinatorError::Unauthorized(
+                "node request requires a signed proof of enrolled private-key possession"
+                    .to_owned(),
+            )
+        })?;
+        if signature.nonce.trim().is_empty() || signature.nonce.len() > 256 {
+            return Err(CoordinatorError::Unauthorized(
+                "node signed request nonce is missing or invalid".to_owned(),
+            )
+            .into());
+        }
+        let now_epoch_seconds = unix_timestamp_seconds();
+        if signature
+            .issued_at_epoch_seconds
+            .abs_diff(now_epoch_seconds)
+            > super::NODE_SIGNATURE_WINDOW_SECONDS
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "node signed request is expired or outside the allowed clock skew".to_owned(),
+            )
+            .into());
+        }
+        let replay_key = (node.clone(), signature.nonce.clone());
+        self.node_replay_nonces.retain(|_, accepted_at| {
+            now_epoch_seconds <= accepted_at.saturating_add(super::NODE_SIGNATURE_WINDOW_SECONDS)
+        });
+        if self.node_replay_nonces.contains_key(&replay_key) {
+            return Err(CoordinatorError::Unauthorized(
+                "node signed request nonce has already been used".to_owned(),
+            )
+            .into());
+        }
+        verify_node_request_signature(
+            &identity.public_key,
+            node,
+            request_kind,
+            payload_digest,
+            &signature,
+        )
+        .map_err(CoordinatorError::Unauthorized)?;
+        if self
+            .node_replay_nonces
+            .keys()
+            .filter(|(retained_node, _)| retained_node == node)
+            .count()
+            >= super::MAX_NODE_REPLAY_NONCES_PER_AUTHORITY
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "node signed request replay window is full; retry after the bounded signature window advances"
+                    .to_owned(),
+            )
+            .into());
+        }
+        self.node_replay_nonces
+            .insert(replay_key, now_epoch_seconds);
+        Ok(())
+    }
+}
+
+fn unix_timestamp_seconds() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|duration| duration.as_secs())
+        .unwrap_or(0)
+}

crates/disasmer-coordinator/src/service/panels.rs

@@ -0,0 +1,307 @@
+use disasmer_core::{
+    Actor, DownloadPolicy, PanelEvent, PanelEventKind, PanelState, PanelWidget, PanelWidgetKind,
+    ProcessId, ProjectId, RateLimit, TenantId, UserId,
+};
+
+use crate::CoordinatorError;
+
+use super::artifact_id_from_path;
+use super::keys::panel_stop_key;
+use super::{CoordinatorResponse, CoordinatorService, CoordinatorServiceError};
+
+impl CoordinatorService {
+    pub(super) fn clear_operator_panel_state(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) {
+        let stop_key = panel_stop_key(tenant, project, process);
+        self.panel_snapshots.remove(&stop_key);
+        self.stopped_panels.remove(&stop_key);
+        self.panel_event_limits
+            .retain(|(event_tenant, event_project, event_process, _), _| {
+                event_tenant != tenant || event_project != project || event_process != process
+            });
+    }
+
+    pub(super) fn handle_render_operator_panel(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        max_download_bytes: u64,
+        stopped: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let stop_key = panel_stop_key(&tenant, &project, &process);
+
+        let panel = if stopped {
+            self.ensure_operator_panel_scope(&tenant, &project, &process)?;
+            self.stopped_panels.insert(stop_key);
+            let stop_key = panel_stop_key(&tenant, &project, &process);
+            let panel = match self.panel_snapshots.get(&stop_key).cloned() {
+                Some(panel) => stopped_panel_snapshot(panel),
+                None => self.render_operator_panel(
+                    tenant.clone(),
+                    project.clone(),
+                    process.clone(),
+                    UserId::new(actor_user),
+                    max_download_bytes,
+                    true,
+                )?,
+            };
+            self.panel_snapshots.insert(stop_key, panel.clone());
+            panel
+        } else {
+            self.stopped_panels.remove(&stop_key);
+            let panel = self.render_operator_panel(
+                tenant.clone(),
+                project.clone(),
+                process.clone(),
+                UserId::new(actor_user),
+                max_download_bytes,
+                false,
+            )?;
+            self.panel_snapshots.insert(stop_key, panel.clone());
+            panel
+        };
+        Ok(CoordinatorResponse::OperatorPanel { panel })
+    }
+
+    pub(super) fn handle_submit_panel_event(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        widget_id: String,
+        kind: PanelEventKind,
+        max_events: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let stop_key = panel_stop_key(&tenant, &project, &process);
+        let stopped = self.stopped_panels.contains(&stop_key);
+        let panel = if stopped {
+            match self.panel_snapshots.get(&stop_key).cloned() {
+                Some(panel) => stopped_panel_snapshot(panel),
+                None => {
+                    let panel = self.render_operator_panel(
+                        tenant.clone(),
+                        project.clone(),
+                        process.clone(),
+                        UserId::from("panel-user"),
+                        self.quota.download_limit(),
+                        true,
+                    )?;
+                    self.panel_snapshots.insert(stop_key.clone(), panel.clone());
+                    panel
+                }
+            }
+        } else {
+            let panel = self.render_operator_panel(
+                tenant.clone(),
+                project.clone(),
+                process.clone(),
+                UserId::from("panel-user"),
+                self.quota.download_limit(),
+                false,
+            )?;
+            self.panel_snapshots.insert(stop_key, panel.clone());
+            panel
+        };
+        let event = PanelEvent {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            process: process.clone(),
+            widget_id: widget_id.clone(),
+            kind,
+        };
+        let limit_key = (tenant, project, process, widget_id);
+        let mut limit = self
+            .panel_event_limits
+            .get(&limit_key)
+            .cloned()
+            .unwrap_or(RateLimit {
+                max_events,
+                used_events: 0,
+            });
+        limit.max_events = max_events;
+        panel.accept_event(&event, &mut limit)?;
+        self.panel_event_limits.insert(limit_key, limit.clone());
+        Ok(CoordinatorResponse::PanelEventAccepted {
+            used_events: limit.used_events,
+            max_events: limit.max_events,
+        })
+    }
+
+    fn render_operator_panel(
+        &self,
+        tenant: TenantId,
+        project: ProjectId,
+        process: ProcessId,
+        actor_user: UserId,
+        max_download_bytes: u64,
+        stopped: bool,
+    ) -> Result<PanelState, CoordinatorServiceError> {
+        self.ensure_operator_panel_scope(&tenant, &project, &process)?;
+
+        let events = self
+            .task_events
+            .iter()
+            .filter(|event| {
+                event.tenant == tenant && event.project == project && event.process == process
+            })
+            .collect::<Vec<_>>();
+        let completed = events
+            .iter()
+            .filter(|event| event.status_code == Some(0))
+            .count() as u64;
+        let total = events.len().max(1) as u64;
+        let stdout_bytes = events.iter().map(|event| event.stdout_bytes).sum::<u64>();
+        let stderr_bytes = events.iter().map(|event| event.stderr_bytes).sum::<u64>();
+        let last_task = events.last().map(|event| event.task.clone());
+
+        let mut panel = PanelState::new(tenant.clone(), project.clone(), process.clone());
+        if stopped {
+            panel.freeze_program_ui_events();
+        }
+        panel.add_widget(PanelWidget {
+            id: "process-status".to_owned(),
+            label: "Process Status".to_owned(),
+            kind: PanelWidgetKind::Text {
+                value: if stopped {
+                    "stopped".to_owned()
+                } else {
+                    "running".to_owned()
+                },
+            },
+        })?;
+        panel.add_widget(PanelWidget {
+            id: "task-progress".to_owned(),
+            label: "Tasks".to_owned(),
+            kind: PanelWidgetKind::Progress {
+                current: completed,
+                total,
+            },
+        })?;
+        panel.add_widget(PanelWidget {
+            id: "task-summary".to_owned(),
+            label: "Task Summary".to_owned(),
+            kind: PanelWidgetKind::Text {
+                value: if events.is_empty() {
+                    "no task events recorded".to_owned()
+                } else {
+                    events
+                        .iter()
+                        .map(|event| {
+                            format!(
+                                "{} [{}]:{:?}:{}",
+                                event.task_definition, event.task, event.status_code, event.node
+                            )
+                        })
+                        .collect::<Vec<_>>()
+                        .join(", ")
+                },
+            },
+        })?;
+        panel.add_widget(PanelWidget {
+            id: "recent-logs".to_owned(),
+            label: "Recent Logs".to_owned(),
+            kind: PanelWidgetKind::Text {
+                value: format!("stdout={stdout_bytes} stderr={stderr_bytes}"),
+            },
+        })?;
+        panel.add_widget(PanelWidget {
+            id: "debug-process".to_owned(),
+            label: "Debug Process".to_owned(),
+            kind: PanelWidgetKind::Button {
+                action: "debug-process".to_owned(),
+            },
+        })?;
+        panel.add_widget(PanelWidget {
+            id: "cancel-process".to_owned(),
+            label: "Cancel Process".to_owned(),
+            kind: PanelWidgetKind::Button {
+                action: "cancel-process".to_owned(),
+            },
+        })?;
+        if last_task.is_some() {
+            panel.add_widget(PanelWidget {
+                id: "restart-selected-task".to_owned(),
+                label: "Restart Selected Task".to_owned(),
+                kind: PanelWidgetKind::Button {
+                    action: "restart-task".to_owned(),
+                },
+            })?;
+        }
+
+        let mut actions = vec![
+            disasmer_core::ControlPlaneAction::DebugProcess,
+            disasmer_core::ControlPlaneAction::CancelProcess,
+        ];
+        if let Some(task) = last_task.clone() {
+            actions.push(disasmer_core::ControlPlaneAction::RestartTask(task));
+        }
+        panel.set_control_plane_actions(actions);
+
+        if let Some(path) = events
+            .iter()
+            .rev()
+            .find_map(|event| event.artifact_path.as_ref())
+        {
+            let artifact = artifact_id_from_path(path);
+            let context = disasmer_core::AuthContext {
+                tenant,
+                project,
+                actor: Actor::User(actor_user),
+            };
+            panel.add_download_widget_from_action(
+                "download-artifact",
+                "Download Artifact",
+                self.artifact_registry.download_action(
+                    &context,
+                    &artifact,
+                    &DownloadPolicy {
+                        max_bytes: max_download_bytes,
+                    },
+                ),
+            )?;
+        }
+
+        Ok(panel)
+    }
+
+    fn ensure_operator_panel_scope(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+    ) -> Result<(), CoordinatorServiceError> {
+        let active = self
+            .coordinator
+            .active_process(tenant, project, process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "operator panel requires an active virtual process".to_owned(),
+                )
+            })?;
+        debug_assert_eq!(active.tenant, *tenant);
+        debug_assert_eq!(active.project, *project);
+        Ok(())
+    }
+}
+
+fn stopped_panel_snapshot(mut panel: PanelState) -> PanelState {
+    panel.freeze_program_ui_events();
+    if let Some(status) = panel.widgets.get_mut("process-status") {
+        status.kind = PanelWidgetKind::Text {
+            value: "stopped".to_owned(),
+        };
+    }
+    panel
+}

crates/disasmer-coordinator/src/service/process_launch.rs

@@ -0,0 +1,555 @@
+use std::collections::BTreeMap;
+
+use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine as _};
+use disasmer_core::{
+    AgentSignedRequest, ArtifactId, CheckpointBoundary, CredentialKind, DefaultScheduler, Digest,
+    NodeId, PlacementRequest, ProcessId, ProjectId, Scheduler, TaskCheckpoint, TaskDispatch,
+    TaskInstanceId, TaskSpec, TenantId, VfsManifest, VfsObject, VfsPath, WasmTaskInvocation,
+};
+
+use crate::CoordinatorError;
+
+use super::keys::{process_control_key, task_control_key, task_restart_key};
+use super::{
+    CoordinatorResponse, CoordinatorService, CoordinatorServiceError, TaskAssignment, WorkflowActor,
+};
+
+use super::processes::*;
+
+impl CoordinatorService {
+    pub(super) fn capture_task_restart_checkpoint(
+        &mut self,
+        assignment: &TaskAssignment,
+    ) -> Result<(), CoordinatorServiceError> {
+        let task_spec = &assignment.task_spec;
+        let environment_digest = task_spec.environment_digest.clone().unwrap_or_else(|| {
+            task_spec.environment.as_ref().map_or_else(
+                || Digest::sha256("disasmer.environment.unconstrained.v1"),
+                |environment| {
+                    Digest::sha256(
+                        serde_json::to_vec(environment)
+                            .expect("serializable environment requirements"),
+                    )
+                },
+            )
+        });
+        let task_entrypoint = match &task_spec.dispatch {
+            disasmer_core::TaskDispatch::CoordinatorNodeWasm { export, .. } => export
+                .clone()
+                .or_else(|| assignment_task_descriptor(assignment)?.get("export")?.as_str().map(str::to_owned))
+                .ok_or_else(|| {
+                    CoordinatorServiceError::Protocol(format!(
+                        "cannot capture restart checkpoint for task `{}`: bundle descriptor omitted its Wasm export",
+                        task_spec.task_definition
+                    ))
+                })?,
+        };
+        let mut objects = BTreeMap::new();
+        let mut missing_required_artifact = false;
+        for artifact in &task_spec.required_artifacts {
+            let Some(metadata) = self.artifact_registry.metadata(artifact) else {
+                missing_required_artifact = true;
+                continue;
+            };
+            if metadata.tenant != assignment.tenant
+                || metadata.project != assignment.project
+                || metadata.retaining_nodes.is_empty()
+            {
+                missing_required_artifact = true;
+                continue;
+            }
+            let path = VfsPath::new(format!("/vfs/artifacts/{artifact}"))
+                .map_err(|error| CoordinatorServiceError::InvalidArtifactPath(error.to_string()))?;
+            objects.insert(
+                path.clone(),
+                VfsObject {
+                    path,
+                    digest: metadata.digest.clone(),
+                    size: metadata.size,
+                    producer: metadata.producer_task.clone(),
+                    node: metadata.producer_node.clone(),
+                },
+            );
+        }
+        let checkpoint = TaskCheckpoint {
+            task: assignment.task.clone(),
+            boundary: CheckpointBoundary {
+                task_entrypoint,
+                serialized_args: Digest::sha256(serde_json::to_vec(&task_spec.args)?),
+                environment_digest,
+                vfs_epoch: task_spec.vfs_epoch,
+                task_abi: assignment_task_compatibility(assignment)
+                    .unwrap_or(Digest::sha256(serde_json::to_vec(&task_spec.dispatch)?)),
+            },
+            vfs_manifest: VfsManifest {
+                epoch: task_spec.vfs_epoch,
+                producer: assignment.task.clone(),
+                node: assignment.node.clone(),
+                objects,
+                large_bytes_uploaded: false,
+            },
+            depends_on_live_stack: false,
+            depends_on_live_socket: false,
+            depends_on_ephemeral_artifact_durability: missing_required_artifact,
+        };
+        let key = task_restart_key(
+            &assignment.tenant,
+            &assignment.project,
+            &assignment.process,
+            &assignment.task,
+        );
+        self.task_restart_checkpoint_order
+            .retain(|retained| retained != &key);
+        self.task_restart_checkpoints.insert(
+            key.clone(),
+            TaskRestartCheckpoint {
+                checkpoint,
+                assignment: assignment.clone(),
+            },
+        );
+        self.task_restart_checkpoint_order.push_back(key);
+        while self
+            .task_restart_checkpoint_order
+            .iter()
+            .filter(|(tenant, project, process, _)| {
+                tenant == &assignment.tenant
+                    && project == &assignment.project
+                    && process == &assignment.process
+            })
+            .count()
+            > super::MAX_RESTART_CHECKPOINTS_PER_PROCESS
+        {
+            let Some(index) = self.task_restart_checkpoint_order.iter().position(
+                |(tenant, project, process, _)| {
+                    tenant == &assignment.tenant
+                        && project == &assignment.project
+                        && process == &assignment.process
+                },
+            ) else {
+                break;
+            };
+            if let Some(expired) = self.task_restart_checkpoint_order.remove(index) {
+                self.task_restart_checkpoints.remove(&expired);
+            }
+        }
+        Ok(())
+    }
+
+    pub(super) fn handle_schedule_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        environment: Option<disasmer_core::EnvironmentRequirements>,
+        environment_digest: Option<Digest>,
+        required_capabilities: Vec<disasmer_core::Capability>,
+        dependency_cache: Option<Digest>,
+        source_snapshot: Option<Digest>,
+        required_artifacts: Vec<String>,
+        prefer_node: Option<String>,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        let request = PlacementRequest {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            environment,
+            environment_digest,
+            required_capabilities: required_capabilities.into_iter().collect(),
+            dependency_cache,
+            source_snapshot,
+            required_artifacts: required_artifacts
+                .into_iter()
+                .map(ArtifactId::new)
+                .collect(),
+            quota_available: self
+                .quota
+                .can_charge_workflow_spawn(&tenant, &project, now_epoch_seconds)
+                .is_ok(),
+            policy_allowed: self.admission.workflow_placement_allowed,
+            prefer_node: prefer_node.map(NodeId::new),
+        };
+        let nodes = self.node_descriptors.values().cloned().collect::<Vec<_>>();
+        let placement = DefaultScheduler.place(&nodes, &request)?;
+        Ok(CoordinatorResponse::TaskPlacement { placement })
+    }
+
+    pub(super) fn handle_launch_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: Option<String>,
+        actor_agent: Option<String>,
+        agent_public_key_fingerprint: Option<Digest>,
+        agent_signature: Option<AgentSignedRequest>,
+        request_payload_digest: Option<&Digest>,
+        task_spec: TaskSpec,
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        if matches!(
+            &task_spec.dispatch,
+            TaskDispatch::CoordinatorNodeWasm {
+                abi: disasmer_core::WasmExportAbi::EntrypointV1,
+                ..
+            }
+        ) {
+            return self.handle_launch_coordinator_main(
+                tenant,
+                project,
+                actor_user,
+                actor_agent,
+                agent_public_key_fingerprint,
+                agent_signature,
+                request_payload_digest,
+                task_spec,
+                wasm_module_base64,
+            );
+        }
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = task_spec.process.clone();
+        let task = task_spec.task_instance.clone();
+        let actor = self.workflow_actor(
+            &tenant,
+            &project,
+            actor_user,
+            actor_agent,
+            agent_public_key_fingerprint,
+            agent_signature,
+            request_payload_digest,
+            "launch_task",
+            &process,
+            Some(&task),
+        )?;
+        self.handle_launch_task_with_actor(
+            tenant,
+            project,
+            actor,
+            task_spec,
+            wait_for_node,
+            artifact_path,
+            wasm_module_base64,
+        )
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(super) fn handle_launch_child_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        parent_task: String,
+        task_spec: TaskSpec,
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let parent_task = TaskInstanceId::new(parent_task);
+        if task_spec.process != process {
+            return Err(CoordinatorError::Unauthorized(
+                "child task must remain in its parent virtual process".to_owned(),
+            )
+            .into());
+        }
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        let parent_key = task_control_key(&tenant, &project, &process, &node, &parent_task);
+        if !self.active_tasks.contains(&parent_key) {
+            return Err(CoordinatorError::Unauthorized(
+                "child task launch requires a currently active parent task on the signed node"
+                    .to_owned(),
+            )
+            .into());
+        }
+        let actor = WorkflowActor {
+            kind: "task".to_owned(),
+            user: None,
+            agent: None,
+            credential_kind: CredentialKind::TaskCredential,
+            public_key_fingerprint: None,
+            authenticated_without_browser: true,
+            scopes: vec!["process:spawn-child".to_owned()],
+        };
+        self.handle_launch_task_with_actor(
+            tenant,
+            project,
+            actor,
+            task_spec,
+            wait_for_node,
+            artifact_path,
+            wasm_module_base64,
+        )
+    }
+
+    pub(super) fn handle_launch_task_with_actor(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        actor: WorkflowActor,
+        task_spec: TaskSpec,
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        if task_spec.tenant != tenant || task_spec.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "task specification is outside the authenticated tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        if !task_spec.product_mode_uses_remote_dispatch() {
+            return Err(CoordinatorError::Unauthorized(
+                "task specification must use the Wasm coordinator/node dispatch ABI".to_owned(),
+            )
+            .into());
+        }
+        if task_spec
+            .environment_id
+            .as_deref()
+            .is_some_and(|environment| environment.trim().is_empty() || environment.len() > 128)
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "task specification environment id is invalid".to_owned(),
+            )
+            .into());
+        }
+        let process = task_spec.process.clone();
+        let task = task_spec.task_instance.clone();
+        let active = self
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "task launch requires an active coordinator-side virtual process".to_owned(),
+                )
+            })?;
+        debug_assert_eq!(active.tenant, tenant);
+        debug_assert_eq!(active.project, project);
+        if self
+            .process_cancellations
+            .contains(&process_control_key(&tenant, &project, &process))
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "task launch is blocked because the virtual process is cancelling".to_owned(),
+            )
+            .into());
+        }
+        if self.task_instance_exists(&tenant, &project, &process, &task) {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "task instance {task} already exists in virtual process {process}; every spawn must use a fresh task-instance id"
+            )));
+        }
+        let in_flight = self
+            .active_tasks
+            .iter()
+            .filter(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant == &tenant && task_project == &project && task_process == &process
+            })
+            .count()
+            + self
+                .pending_task_launches
+                .iter()
+                .filter(|pending| {
+                    pending.tenant == tenant
+                        && pending.project == project
+                        && pending.process == process
+                })
+                .count();
+        if in_flight >= super::MAX_IN_FLIGHT_TASKS_PER_PROCESS {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "virtual process task limit of {} reached; join or cancel existing work before spawning more",
+                super::MAX_IN_FLIGHT_TASKS_PER_PROCESS
+            )));
+        }
+        if task_spec.vfs_epoch != active.coordinator_epoch {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "task specification VFS epoch {} does not match active process epoch {}",
+                task_spec.vfs_epoch, active.coordinator_epoch
+            ))
+            .into());
+        }
+        let bundle_digest = task_spec.bundle_digest.as_ref().ok_or_else(|| {
+            CoordinatorError::Unauthorized(
+                "Wasm task specification is missing its bundle digest".to_owned(),
+            )
+        })?;
+        if !bundle_digest.is_valid_sha256() {
+            return Err(CoordinatorError::Unauthorized(
+                "Wasm task specification has an invalid bundle digest".to_owned(),
+            )
+            .into());
+        }
+        let module = BASE64_STANDARD
+            .decode(&wasm_module_base64)
+            .map_err(|error| {
+                CoordinatorServiceError::Protocol(format!(
+                    "Wasm task module is not valid base64: {error}"
+                ))
+            })?;
+        let actual_digest = Digest::sha256(&module);
+        if &actual_digest != bundle_digest {
+            return Err(CoordinatorError::Unauthorized(format!(
+                "Wasm task module digest does not match bundle digest: expected {bundle_digest}, actual {actual_digest}"
+            ))
+            .into());
+        }
+        WasmTaskInvocation::new(
+            task_spec.task_definition.clone(),
+            task.clone(),
+            task_spec.args.clone(),
+        )
+        .validate()
+        .map_err(CoordinatorServiceError::Protocol)?;
+        for artifact in &task_spec.required_artifacts {
+            let metadata = self.artifact_registry.metadata(artifact).ok_or_else(|| {
+                CoordinatorError::Unauthorized(format!(
+                    "required artifact {artifact} is unavailable or has expired"
+                ))
+            })?;
+            if metadata.tenant != tenant || metadata.project != project {
+                return Err(CoordinatorError::Unauthorized(format!(
+                    "required artifact {artifact} is outside the task tenant/project scope"
+                ))
+                .into());
+            }
+            if metadata.retaining_nodes.is_empty() {
+                return Err(CoordinatorError::Unauthorized(format!(
+                    "required artifact {artifact} has no retaining node"
+                ))
+                .into());
+            }
+        }
+        VfsPath::new(&artifact_path)
+            .map_err(|error| CoordinatorServiceError::InvalidArtifactPath(error.to_string()))?;
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.quota
+            .can_charge_workflow_spawn(&tenant, &project, now_epoch_seconds)?;
+        let request = PlacementRequest {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            environment: task_spec.environment.clone(),
+            environment_digest: task_spec.environment_digest.clone(),
+            required_capabilities: task_spec.required_capabilities.clone(),
+            dependency_cache: task_spec.dependency_cache.clone(),
+            source_snapshot: task_spec.source_snapshot.clone(),
+            required_artifacts: task_spec.required_artifacts.iter().cloned().collect(),
+            quota_available: self
+                .quota
+                .can_charge_workflow_spawn(&tenant, &project, now_epoch_seconds)
+                .is_ok(),
+            policy_allowed: self.admission.workflow_placement_allowed,
+            prefer_node: None,
+        };
+        let nodes = self.node_descriptors.values().cloned().collect::<Vec<_>>();
+        let placement = match DefaultScheduler.place(&nodes, &request) {
+            Ok(placement) => placement,
+            Err(err) if wait_for_node => {
+                let reason = if err.message.is_empty() {
+                    "waiting for any capable node".to_owned()
+                } else {
+                    err.message
+                };
+                let charged_spawns =
+                    self.quota
+                        .charge_workflow_spawn(&tenant, &project, now_epoch_seconds)?;
+                self.pending_task_launches.push_back(PendingTaskLaunch {
+                    tenant: tenant.clone(),
+                    project: project.clone(),
+                    process: process.clone(),
+                    task: task.clone(),
+                    request,
+                    epoch: active.coordinator_epoch,
+                    artifact_path,
+                    task_spec,
+                    wasm_module_base64,
+                });
+                return Ok(CoordinatorResponse::TaskQueued {
+                    process,
+                    task,
+                    actor,
+                    reason,
+                    charged_spawns,
+                    queued_tasks: self.pending_task_launches.len(),
+                });
+            }
+            Err(err) => return Err(err.into()),
+        };
+        let charged_spawns =
+            self.quota
+                .charge_workflow_spawn(&tenant, &project, now_epoch_seconds)?;
+        let assignment = TaskAssignment {
+            tenant: tenant.clone(),
+            project: project.clone(),
+            process: process.clone(),
+            task: task.clone(),
+            node: placement.node.clone(),
+            epoch: active.coordinator_epoch,
+            artifact_path,
+            task_spec,
+            wasm_module_base64,
+        };
+        self.capture_task_restart_checkpoint(&assignment)?;
+        let task_key = task_control_key(&tenant, &project, &process, &placement.node, &task);
+        self.task_placements
+            .insert(task_key.clone(), placement.clone());
+        self.active_tasks.insert(task_key);
+        self.task_assignments
+            .entry((tenant, project, placement.node.clone()))
+            .or_default()
+            .push_back(assignment.clone());
+        Ok(CoordinatorResponse::TaskLaunched {
+            process,
+            task,
+            actor,
+            placement,
+            assignment: Box::new(assignment),
+            charged_spawns,
+        })
+    }
+
+    fn task_instance_exists(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        process: &ProcessId,
+        task_instance: &disasmer_core::TaskInstanceId,
+    ) -> bool {
+        self.active_tasks.iter().any(
+            |(task_tenant, task_project, task_process, _, existing_instance)| {
+                task_tenant == tenant
+                    && task_project == project
+                    && task_process == process
+                    && existing_instance == task_instance
+            },
+        ) || self.pending_task_launches.iter().any(|pending| {
+            &pending.tenant == tenant
+                && &pending.project == project
+                && &pending.process == process
+                && &pending.task == task_instance
+        }) || self.task_events.iter().any(|event| {
+            &event.tenant == tenant
+                && &event.project == project
+                && &event.process == process
+                && &event.task == task_instance
+        })
+    }
+}
+
+fn assignment_task_compatibility(assignment: &TaskAssignment) -> Option<Digest> {
+    let descriptor = assignment_task_descriptor(assignment)?;
+    serde_json::from_value(descriptor.get("restart_compatibility_hash")?.clone()).ok()
+}
+
+fn assignment_task_descriptor(assignment: &TaskAssignment) -> Option<serde_json::Value> {
+    let module = BASE64_STANDARD
+        .decode(&assignment.wasm_module_base64)
+        .ok()?;
+    let mut descriptors = super::main_runtime::task_descriptors(&module).ok()?;
+    descriptors.remove(assignment.task_spec.task_definition.as_str())
+}

crates/disasmer-coordinator/src/service/processes.rs

@@ -0,0 +1,788 @@
+use std::collections::{BTreeSet, VecDeque};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use disasmer_core::{
+    AgentId, AgentSignedRequest, CredentialKind, DefaultScheduler, Digest, NodeId,
+    PlacementRequest, ProcessId, ProjectId, RendezvousRequest, Scheduler, SourcePreparation,
+    TaskCheckpoint, TaskInstanceId, TaskSpec, TenantId, UserId,
+};
+
+use crate::CoordinatorError;
+
+use super::keys::{process_control_key, task_control_key};
+use super::{
+    CoordinatorResponse, CoordinatorService, CoordinatorServiceError, SourcePreparationDisposition,
+    SourcePreparationStatus, TaskAssignment, TaskCancellationTarget, VirtualProcessStatus,
+    WorkflowActor,
+};
+
+#[derive(Clone, Debug)]
+pub(super) struct PendingTaskLaunch {
+    pub(super) tenant: TenantId,
+    pub(super) project: ProjectId,
+    pub(super) process: ProcessId,
+    pub(super) task: TaskInstanceId,
+    pub(super) request: PlacementRequest,
+    pub(super) epoch: u64,
+    pub(super) artifact_path: String,
+    pub(super) task_spec: TaskSpec,
+    pub(super) wasm_module_base64: String,
+}
+
+#[derive(Clone, Debug)]
+pub(super) struct TaskRestartCheckpoint {
+    pub(super) checkpoint: TaskCheckpoint,
+    pub(super) assignment: TaskAssignment,
+}
+
+impl CoordinatorService {
+    pub(super) fn handle_poll_task_assignment(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        let identity = self
+            .coordinator
+            .node_identity(&node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        if identity.tenant != tenant || identity.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "task assignment poll is outside the enrolled tenant/project scope".to_owned(),
+            )
+            .into());
+        }
+        let assignment_key = (tenant.clone(), project.clone(), node.clone());
+        let assignment = self
+            .task_assignments
+            .get_mut(&assignment_key)
+            .and_then(VecDeque::pop_front);
+        if assignment.is_some() {
+            return Ok(CoordinatorResponse::TaskAssignment {
+                assignment: assignment.map(Box::new),
+            });
+        }
+        let assignment = self.assign_pending_task_to_node(&tenant, &project, &node)?;
+        Ok(CoordinatorResponse::TaskAssignment {
+            assignment: assignment.map(Box::new),
+        })
+    }
+
+    fn assign_pending_task_to_node(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        node: &NodeId,
+    ) -> Result<Option<TaskAssignment>, CoordinatorServiceError> {
+        let Some(descriptor) = self.node_descriptors.get(node).cloned() else {
+            return Ok(None);
+        };
+        let mut remaining = VecDeque::new();
+        let mut selected = None;
+
+        while let Some(pending) = self.pending_task_launches.pop_front() {
+            if selected.is_some() {
+                remaining.push_back(pending);
+                continue;
+            }
+            if &pending.tenant != tenant || &pending.project != project {
+                remaining.push_back(pending);
+                continue;
+            }
+            if self.process_cancellations.contains(&process_control_key(
+                &pending.tenant,
+                &pending.project,
+                &pending.process,
+            )) {
+                continue;
+            }
+            let Some(active) = self.coordinator.active_process(
+                &pending.tenant,
+                &pending.project,
+                &pending.process,
+            ) else {
+                continue;
+            };
+            if active.tenant != pending.tenant
+                || active.project != pending.project
+                || active.coordinator_epoch != pending.epoch
+            {
+                continue;
+            }
+            let Ok(placement) =
+                DefaultScheduler.place(std::slice::from_ref(&descriptor), &pending.request)
+            else {
+                remaining.push_back(pending);
+                continue;
+            };
+            let assignment = TaskAssignment {
+                tenant: pending.tenant.clone(),
+                project: pending.project.clone(),
+                process: pending.process.clone(),
+                task: pending.task.clone(),
+                node: placement.node.clone(),
+                epoch: pending.epoch,
+                artifact_path: pending.artifact_path,
+                task_spec: pending.task_spec,
+                wasm_module_base64: pending.wasm_module_base64,
+            };
+            self.capture_task_restart_checkpoint(&assignment)?;
+            let task_key = task_control_key(
+                &pending.tenant,
+                &pending.project,
+                &pending.process,
+                &placement.node,
+                &pending.task,
+            );
+            self.task_placements.insert(task_key.clone(), placement);
+            self.active_tasks.insert(task_key);
+            selected = Some(assignment);
+        }
+
+        self.pending_task_launches = remaining;
+        Ok(selected)
+    }
+
+    pub(super) fn handle_request_rendezvous(
+        &mut self,
+        scope: disasmer_core::DataPlaneScope,
+        source: disasmer_core::NodeEndpoint,
+        destination: disasmer_core::NodeEndpoint,
+        direct_connectivity: bool,
+        failure_reason: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        let charged_rendezvous_attempts = self.quota.charge_rendezvous_attempt(
+            &scope.tenant,
+            &scope.project,
+            now_epoch_seconds,
+        )?;
+        let plan = self.transport.plan_authenticated_direct_bulk_transfer(
+            RendezvousRequest {
+                scope,
+                source,
+                destination,
+            },
+            direct_connectivity,
+            failure_reason,
+        )?;
+        Ok(CoordinatorResponse::RendezvousPlan {
+            plan,
+            charged_rendezvous_attempts,
+        })
+    }
+
+    pub(super) fn handle_request_source_preparation(
+        &mut self,
+        tenant: String,
+        project: String,
+        provider: disasmer_core::SourceProviderKind,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let preparation = SourcePreparation::node_task(tenant.clone(), project.clone(), provider);
+        let request = PlacementRequest {
+            tenant,
+            project,
+            environment: None,
+            environment_digest: None,
+            required_capabilities: preparation.required_capabilities.clone(),
+            dependency_cache: None,
+            source_snapshot: None,
+            required_artifacts: Default::default(),
+            quota_available: true,
+            policy_allowed: true,
+            prefer_node: None,
+        };
+        let nodes = self.node_descriptors.values().cloned().collect::<Vec<_>>();
+        let disposition = match DefaultScheduler.place(&nodes, &request) {
+            Ok(placement) => SourcePreparationDisposition::Assigned {
+                node: placement.node,
+            },
+            Err(err) => SourcePreparationDisposition::Pending {
+                reason: if err.message.is_empty() {
+                    "waiting for any capable node to prepare source".to_owned()
+                } else {
+                    err.message
+                },
+            },
+        };
+        Ok(CoordinatorResponse::SourcePreparation {
+            status: SourcePreparationStatus {
+                preparation,
+                disposition,
+            },
+        })
+    }
+
+    pub(super) fn handle_complete_source_preparation(
+        &mut self,
+        tenant: String,
+        project: String,
+        node: String,
+        provider: disasmer_core::SourceProviderKind,
+        source_snapshot: Digest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let node = NodeId::new(node);
+        let identity = self
+            .coordinator
+            .node_identity(&node)
+            .ok_or(CoordinatorError::UnknownNode)?;
+        if identity.tenant != tenant || identity.project != project {
+            return Err(CoordinatorError::Unauthorized(
+                "source preparation completion is outside the enrolled tenant/project scope"
+                    .to_owned(),
+            )
+            .into());
+        }
+        let descriptor = self.node_descriptors.get_mut(&node).ok_or_else(|| {
+            CoordinatorError::Unauthorized(
+                "source preparation completion requires a node capability report".to_owned(),
+            )
+        })?;
+        if !descriptor.source_snapshots.contains(&source_snapshot)
+            && descriptor.source_snapshots.len() >= super::MAX_NODE_REPORTED_OBJECTS_PER_KIND
+        {
+            return Err(CoordinatorServiceError::Protocol(format!(
+                "node source snapshot retention limit of {} reached; refresh the node capability report",
+                super::MAX_NODE_REPORTED_OBJECTS_PER_KIND
+            )));
+        }
+        descriptor.source_snapshots.insert(source_snapshot.clone());
+        Ok(CoordinatorResponse::SourcePreparationCompleted {
+            node,
+            provider,
+            source_snapshot,
+        })
+    }
+
+    pub(super) fn handle_start_process(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: Option<String>,
+        actor_agent: Option<String>,
+        agent_public_key_fingerprint: Option<Digest>,
+        agent_signature: Option<AgentSignedRequest>,
+        request_payload_digest: Option<&Digest>,
+        process: String,
+        restart: bool,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        self.coordinator.ensure_tenant_active(&tenant)?;
+        let actor = self.workflow_actor(
+            &tenant,
+            &project,
+            actor_user,
+            actor_agent,
+            agent_public_key_fingerprint,
+            agent_signature,
+            request_payload_digest,
+            "start_process",
+            &process,
+            None,
+        )?;
+        let replacing_existing = if let Some(active) = self
+            .coordinator
+            .active_process_for_project(&tenant, &project)
+        {
+            if active.id != process || !restart {
+                return Err(CoordinatorError::Unauthorized(format!(
+                    "project already has active virtual process {}; attach to or restart it, request cooperative cancellation, abort it, or use another Coordinator Project",
+                    active.id
+                ))
+                .into());
+            }
+            true
+        } else {
+            false
+        };
+        if replacing_existing {
+            self.main_runtime.interrupt_process(
+                &tenant,
+                &project,
+                &process,
+                "virtual process incarnation replaced",
+            );
+            self.main_runtime
+                .controls
+                .remove(&process_control_key(&tenant, &project, &process));
+        }
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        let charged_spawns =
+            self.quota
+                .charge_workflow_spawn(&tenant, &project, now_epoch_seconds)?;
+        self.process_cancellations
+            .remove(&process_control_key(&tenant, &project, &process));
+        self.process_aborts
+            .remove(&process_control_key(&tenant, &project, &process));
+        self.clear_debug_state_for_process(&tenant, &project, &process);
+        self.clear_operator_panel_state(&tenant, &project, &process);
+        self.task_cancellations
+            .retain(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant != &tenant || task_project != &project || task_process != &process
+            });
+        self.task_aborts
+            .retain(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant != &tenant || task_project != &project || task_process != &process
+            });
+        self.active_tasks
+            .retain(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant != &tenant || task_project != &project || task_process != &process
+            });
+        self.task_placements
+            .retain(|(task_tenant, task_project, task_process, _, _), _| {
+                task_tenant != &tenant || task_project != &project || task_process != &process
+            });
+        self.task_assignments.retain(|_, assignments| {
+            assignments.retain(|assignment| {
+                assignment.tenant != tenant
+                    || assignment.project != project
+                    || assignment.process != process
+            });
+            !assignments.is_empty()
+        });
+        self.pending_task_launches.retain(|pending| {
+            pending.tenant != tenant || pending.project != project || pending.process != process
+        });
+        self.task_restart_checkpoints.retain(
+            |(checkpoint_tenant, checkpoint_project, checkpoint_process, _), _| {
+                checkpoint_tenant != &tenant
+                    || checkpoint_project != &project
+                    || checkpoint_process != &process
+            },
+        );
+        self.task_restart_checkpoint_order.retain(
+            |(checkpoint_tenant, checkpoint_project, checkpoint_process, _)| {
+                checkpoint_tenant != &tenant
+                    || checkpoint_project != &project
+                    || checkpoint_process != &process
+            },
+        );
+        self.task_events.retain(|event| {
+            event.tenant != tenant || event.project != project || event.process != process
+        });
+        self.debug_audit_events.retain(|event| {
+            event.tenant != tenant || event.project != project || event.process != process
+        });
+        self.coordinator
+            .start_process(tenant, project, process.clone());
+        Ok(CoordinatorResponse::ProcessStarted {
+            process,
+            epoch: self.coordinator.coordinator_epoch(),
+            actor,
+            charged_spawns,
+        })
+    }
+
+    pub(super) fn handle_reconnect_node(
+        &mut self,
+        node: String,
+        process: String,
+        epoch: u64,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let node = NodeId::new(node);
+        let process = ProcessId::new(process);
+        self.coordinator
+            .reconnect_node(&node, Some((&process, epoch)))?;
+        Ok(CoordinatorResponse::NodeReconnected { node, process })
+    }
+
+    pub(super) fn handle_cancel_task(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.coordinator
+            .authorize_node_for_process(&node, &tenant, &project, &process)?;
+        let active = self
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "task cancellation requires an active virtual process".to_owned(),
+                )
+            })?;
+        if !active.connected_nodes.contains(&node) {
+            return Err(CoordinatorError::Unauthorized(
+                "task cancellation target node is not connected to the virtual process".to_owned(),
+            )
+            .into());
+        }
+        self.task_cancellations
+            .insert(task_control_key(&tenant, &project, &process, &node, &task));
+        Ok(CoordinatorResponse::TaskCancellationRequested {
+            process,
+            task,
+            node,
+        })
+    }
+
+    pub(super) fn handle_cancel_process(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let _actor_user = actor_user;
+        let active = self
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "process cancellation requires an active virtual process".to_owned(),
+                )
+            })?;
+        debug_assert_eq!(active.tenant, tenant);
+        debug_assert_eq!(active.project, project);
+        self.process_cancellations
+            .insert(process_control_key(&tenant, &project, &process));
+        self.main_runtime.interrupt_process(
+            &tenant,
+            &project,
+            &process,
+            "virtual process cancellation requested",
+        );
+        self.clear_debug_state_for_process(&tenant, &project, &process);
+        self.pending_task_launches.retain(|pending| {
+            pending.tenant != tenant || pending.project != project || pending.process != process
+        });
+        let mut cancelled_tasks = Vec::new();
+        let mut affected_nodes = BTreeSet::new();
+        for (task_tenant, task_project, task_process, node, task) in self.active_tasks.iter() {
+            if task_tenant == &tenant && task_project == &project && task_process == &process {
+                self.task_cancellations
+                    .insert(task_control_key(&tenant, &project, &process, node, task));
+                affected_nodes.insert(node.clone());
+                cancelled_tasks.push(TaskCancellationTarget {
+                    process: process.clone(),
+                    task: task.clone(),
+                    node: node.clone(),
+                });
+            }
+        }
+        Ok(CoordinatorResponse::ProcessCancellationRequested {
+            process,
+            cancelled_tasks,
+            affected_nodes: affected_nodes.into_iter().collect(),
+        })
+    }
+
+    pub(super) fn handle_abort_process(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let _actor_user = actor_user;
+        let active = self
+            .coordinator
+            .active_process(&tenant, &project, &process)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "process abort requires an active virtual process".to_owned(),
+                )
+            })?;
+        debug_assert_eq!(active.tenant, tenant);
+        debug_assert_eq!(active.project, project);
+
+        let process_key = process_control_key(&tenant, &project, &process);
+        self.process_cancellations.remove(&process_key);
+        self.task_cancellations
+            .retain(|(task_tenant, task_project, task_process, _, _)| {
+                task_tenant != &tenant || task_project != &project || task_process != &process
+            });
+        self.process_aborts.insert(process_key);
+        self.main_runtime
+            .interrupt_process(&tenant, &project, &process, "virtual process aborted");
+        self.main_runtime
+            .controls
+            .remove(&process_control_key(&tenant, &project, &process));
+        self.clear_debug_state_for_process(&tenant, &project, &process);
+        self.clear_operator_panel_state(&tenant, &project, &process);
+        self.pending_task_launches.retain(|pending| {
+            pending.tenant != tenant || pending.project != project || pending.process != process
+        });
+        self.task_assignments.retain(|_, assignments| {
+            assignments.retain(|assignment| {
+                assignment.tenant != tenant
+                    || assignment.project != project
+                    || assignment.process != process
+            });
+            !assignments.is_empty()
+        });
+        let mut aborted_tasks = Vec::new();
+        let mut affected_nodes = BTreeSet::new();
+        for (task_tenant, task_project, task_process, node, task) in self.active_tasks.iter() {
+            if task_tenant == &tenant && task_project == &project && task_process == &process {
+                self.task_aborts
+                    .insert(task_control_key(&tenant, &project, &process, node, task));
+                affected_nodes.insert(node.clone());
+                aborted_tasks.push(TaskCancellationTarget {
+                    process: process.clone(),
+                    task: task.clone(),
+                    node: node.clone(),
+                });
+            }
+        }
+
+        self.coordinator
+            .abort_process(&tenant, &project, &process)?;
+        let active_restart_tasks = aborted_tasks
+            .iter()
+            .map(|target| target.task.clone())
+            .collect::<BTreeSet<_>>();
+        self.task_restart_checkpoints.retain(
+            |(checkpoint_tenant, checkpoint_project, checkpoint_process, checkpoint_task), _| {
+                checkpoint_tenant != &tenant
+                    || checkpoint_project != &project
+                    || checkpoint_process != &process
+                    || active_restart_tasks.contains(checkpoint_task)
+            },
+        );
+        self.task_restart_checkpoint_order.retain(
+            |(checkpoint_tenant, checkpoint_project, checkpoint_process, checkpoint_task)| {
+                checkpoint_tenant != &tenant
+                    || checkpoint_project != &project
+                    || checkpoint_process != &process
+                    || active_restart_tasks.contains(checkpoint_task)
+            },
+        );
+        if aborted_tasks.is_empty() {
+            self.process_aborts
+                .remove(&process_control_key(&tenant, &project, &process));
+        }
+        Ok(CoordinatorResponse::ProcessAborted {
+            process,
+            aborted_tasks,
+            affected_nodes: affected_nodes.into_iter().collect(),
+        })
+    }
+
+    pub(super) fn handle_list_processes(
+        &mut self,
+        tenant: String,
+        project: String,
+        actor_user: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let actor = UserId::new(actor_user);
+        let processes = self
+            .coordinator
+            .active_processes_for_project(&tenant, &project)
+            .into_iter()
+            .map(|active| {
+                let process_key = process_control_key(&active.tenant, &active.project, &active.id);
+                let main = self.main_runtime.controls.get(&process_key);
+                let state = if self.process_cancellations.contains(&process_key) {
+                    "cancelling"
+                } else {
+                    main.map_or("running", |main| main.state.as_str())
+                };
+                let main_wait_state = main.and_then(|main| {
+                    if main.state != "running" {
+                        return None;
+                    }
+                    if self.pending_task_launches.iter().any(|pending| {
+                        pending.tenant == active.tenant
+                            && pending.project == active.project
+                            && pending.process == active.id
+                    }) {
+                        Some("waiting_for_node".to_owned())
+                    } else if self.main_runtime.is_waiting_for_task(
+                        &active.tenant,
+                        &active.project,
+                        &active.id,
+                    ) {
+                        Some("waiting_for_task".to_owned())
+                    } else {
+                        Some("executing".to_owned())
+                    }
+                });
+                VirtualProcessStatus {
+                    process: active.id,
+                    state: state.to_owned(),
+                    main_task_definition: main.map(|main| main.task_definition.clone()),
+                    main_task_instance: main.map(|main| main.task_instance.clone()),
+                    main_state: main.map(|main| main.state.clone()),
+                    main_wait_state,
+                    main_debug_epoch: main.and_then(|main| main.debug.requested_epoch()),
+                    connected_nodes: active.connected_nodes.into_iter().collect(),
+                    coordinator_epoch: active.coordinator_epoch,
+                }
+            })
+            .collect();
+        Ok(CoordinatorResponse::ProcessStatuses { processes, actor })
+    }
+
+    pub(super) fn handle_poll_task_control(
+        &mut self,
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let tenant = TenantId::new(tenant);
+        let project = ProjectId::new(project);
+        let process = ProcessId::new(process);
+        let node = NodeId::new(node);
+        let task = TaskInstanceId::new(task);
+        self.authorize_node_for_process_or_termination(&node, &tenant, &project, &process)?;
+        let cancel_requested = self
+            .task_cancellations
+            .contains(&task_control_key(&tenant, &project, &process, &node, &task))
+            || self
+                .process_cancellations
+                .contains(&process_control_key(&tenant, &project, &process));
+        let abort_requested = self
+            .task_aborts
+            .contains(&task_control_key(&tenant, &project, &process, &node, &task))
+            || self
+                .process_aborts
+                .contains(&process_control_key(&tenant, &project, &process));
+        Ok(CoordinatorResponse::TaskControl {
+            process,
+            task,
+            cancel_requested,
+            abort_requested,
+        })
+    }
+
+    pub(super) fn workflow_actor(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        actor_user: Option<String>,
+        actor_agent: Option<String>,
+        agent_public_key_fingerprint: Option<Digest>,
+        agent_signature: Option<AgentSignedRequest>,
+        request_payload_digest: Option<&Digest>,
+        request_kind: &str,
+        process: &ProcessId,
+        task: Option<&TaskInstanceId>,
+    ) -> Result<WorkflowActor, CoordinatorServiceError> {
+        if let Some(agent) = actor_agent {
+            let agent = AgentId::new(agent);
+            let signature = agent_signature.ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "agent workflow dispatch requires a signed request proving private-key possession"
+                        .to_owned(),
+                )
+            })?;
+            let request_payload_digest = request_payload_digest.ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "agent workflow dispatch requires a canonical signed request payload"
+                        .to_owned(),
+                )
+            })?;
+            if signature.nonce.trim().is_empty() || signature.nonce.len() > 256 {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent signed request nonce is missing or invalid".to_owned(),
+                )
+                .into());
+            }
+            let now_epoch_seconds = unix_timestamp_seconds();
+            let replay_key = (
+                tenant.clone(),
+                project.clone(),
+                agent.clone(),
+                signature.nonce.clone(),
+            );
+            const AGENT_SIGNATURE_WINDOW_SECONDS: u64 = 300;
+            self.agent_replay_nonces.retain(|_, issued_at| {
+                now_epoch_seconds <= issued_at.saturating_add(AGENT_SIGNATURE_WINDOW_SECONDS)
+            });
+            if self.agent_replay_nonces.contains_key(&replay_key) {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent signed request nonce has already been used".to_owned(),
+                )
+                .into());
+            }
+            let record = self.coordinator.authorize_agent_project_run(
+                disasmer_core::AgentWorkflowScope {
+                    tenant,
+                    project,
+                    agent: &agent,
+                    request_kind,
+                    process,
+                    task,
+                },
+                agent_public_key_fingerprint.as_ref(),
+                request_payload_digest,
+                &signature,
+                now_epoch_seconds,
+            )?;
+            if self
+                .agent_replay_nonces
+                .keys()
+                .filter(|(retained_tenant, retained_project, retained_agent, _)| {
+                    retained_tenant == tenant
+                        && retained_project == project
+                        && retained_agent == &agent
+                })
+                .count()
+                >= super::MAX_REPLAY_NONCES_PER_AUTHORITY
+            {
+                return Err(CoordinatorError::Unauthorized(
+                    "agent signed request replay window is full; retry after the bounded signature window advances"
+                        .to_owned(),
+                )
+                .into());
+            }
+            self.agent_replay_nonces
+                .insert(replay_key, signature.issued_at_epoch_seconds);
+            return Ok(WorkflowActor {
+                kind: "agent".to_owned(),
+                user: Some(record.user),
+                agent: Some(agent),
+                credential_kind: CredentialKind::PublicKey,
+                public_key_fingerprint: Some(record.public_key_fingerprint),
+                authenticated_without_browser: true,
+                scopes: record.scopes,
+            });
+        }
+
+        let actor = UserId::new(actor_user.unwrap_or_else(|| "user".to_owned()));
+        Ok(WorkflowActor {
+            kind: "user".to_owned(),
+            user: Some(actor),
+            agent: None,
+            credential_kind: CredentialKind::BrowserSession,
+            public_key_fingerprint: None,
+            authenticated_without_browser: false,
+            scopes: vec!["project:read".to_owned(), "project:run".to_owned()],
+        })
+    }
+}
+
+fn unix_timestamp_seconds() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|duration| duration.as_secs())
+        .unwrap_or(0)
+}

crates/disasmer-coordinator/src/service/protocol.rs

@@ -0,0 +1,644 @@
+use std::collections::BTreeMap;
+
+use disasmer_core::{
+    AgentId, AgentSignedRequest, ArtifactId, Authorization, Capability, CredentialKind,
+    DataPlaneScope, Digest, DirectBulkTransferPlan, DownloadLink, EnvironmentRequirements,
+    LimitKind, NodeCapabilities, NodeDescriptor, NodeEndpoint, NodeId, NodeSignedRequest,
+    PanelEventKind, PanelState, Placement, ProcessId, ProjectId, ResourceLimits, SourcePreparation,
+    SourceProviderKind, TaskBoundaryValue, TaskInstanceId, TaskJoinResult, TaskSpec, TenantId,
+    UserId, VfsPath,
+};
+use serde::{Deserialize, Serialize};
+
+mod responses;
+pub use responses::*;
+
+use crate::{AgentPublicKeyRecord, ProjectRecord};
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct TaskReplacementBundle {
+    pub bundle_digest: Digest,
+    pub wasm_module_base64: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case", deny_unknown_fields)]
+pub enum CoordinatorRequest {
+    Ping,
+    Authenticated {
+        session_secret: String,
+        request: AuthenticatedCoordinatorRequest,
+    },
+    AuthStatus {
+        tenant: String,
+        project: String,
+        actor_user: String,
+    },
+    AdminStatus {
+        tenant: String,
+        actor_user: String,
+        admin_proof: Digest,
+        admin_nonce: String,
+        issued_at_epoch_seconds: u64,
+    },
+    SuspendTenant {
+        tenant: String,
+        actor_user: String,
+        target_tenant: String,
+        admin_proof: Digest,
+        admin_nonce: String,
+        issued_at_epoch_seconds: u64,
+    },
+    CreateProject {
+        tenant: String,
+        actor_user: String,
+        project: String,
+        name: String,
+    },
+    SelectProject {
+        tenant: String,
+        actor_user: String,
+        project: String,
+    },
+    ListProjects {
+        tenant: String,
+        actor_user: String,
+    },
+    RegisterAgentPublicKey {
+        tenant: String,
+        project: String,
+        user: String,
+        agent: String,
+        public_key: String,
+    },
+    ListAgentPublicKeys {
+        tenant: String,
+        project: String,
+        user: String,
+    },
+    RotateAgentPublicKey {
+        tenant: String,
+        project: String,
+        user: String,
+        agent: String,
+        public_key: String,
+    },
+    RevokeAgentPublicKey {
+        tenant: String,
+        project: String,
+        user: String,
+        agent: String,
+    },
+    AttachNode {
+        tenant: String,
+        project: String,
+        node: String,
+        public_key: String,
+    },
+    CreateNodeEnrollmentGrant {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        #[serde(default = "default_node_enrollment_ttl_seconds")]
+        ttl_seconds: u64,
+    },
+    ExchangeNodeEnrollmentGrant {
+        tenant: String,
+        project: String,
+        node: String,
+        public_key: String,
+        enrollment_grant: String,
+    },
+    NodeHeartbeat {
+        node: String,
+        #[serde(default)]
+        node_signature: Option<NodeSignedRequest>,
+    },
+    SignedNode {
+        node: String,
+        node_signature: NodeSignedRequest,
+        request: Box<CoordinatorRequest>,
+    },
+    ReportNodeCapabilities {
+        tenant: String,
+        project: String,
+        node: String,
+        capabilities: NodeCapabilities,
+        cached_environment_digests: Vec<Digest>,
+        #[serde(default)]
+        dependency_cache_digests: Vec<Digest>,
+        source_snapshots: Vec<Digest>,
+        artifact_locations: Vec<String>,
+        direct_connectivity: bool,
+        online: bool,
+    },
+    ListNodeDescriptors {
+        tenant: String,
+        project: String,
+        actor_user: String,
+    },
+    RevokeNodeCredential {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        node: String,
+    },
+    ScheduleTask {
+        tenant: String,
+        project: String,
+        environment: Option<EnvironmentRequirements>,
+        environment_digest: Option<Digest>,
+        required_capabilities: Vec<Capability>,
+        #[serde(default)]
+        dependency_cache: Option<Digest>,
+        source_snapshot: Option<Digest>,
+        required_artifacts: Vec<String>,
+        prefer_node: Option<String>,
+    },
+    LaunchTask {
+        tenant: String,
+        project: String,
+        #[serde(default)]
+        actor_user: Option<String>,
+        #[serde(default)]
+        actor_agent: Option<String>,
+        #[serde(default)]
+        agent_public_key_fingerprint: Option<Digest>,
+        #[serde(default)]
+        agent_signature: Option<AgentSignedRequest>,
+        task_spec: TaskSpec,
+        #[serde(default)]
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    },
+    LaunchChildTask {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        parent_task: String,
+        task_spec: TaskSpec,
+        #[serde(default)]
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    },
+    JoinChildTask {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        parent_task: String,
+        task: String,
+    },
+    PollTaskAssignment {
+        tenant: String,
+        project: String,
+        node: String,
+    },
+    PollArtifactTransfer {
+        tenant: String,
+        project: String,
+        node: String,
+    },
+    UploadArtifactTransferChunk {
+        tenant: String,
+        project: String,
+        node: String,
+        transfer_id: String,
+        artifact: String,
+        offset: u64,
+        content_base64: String,
+        chunk_digest: Digest,
+        eof: bool,
+    },
+    FailArtifactTransfer {
+        tenant: String,
+        project: String,
+        node: String,
+        transfer_id: String,
+        artifact: String,
+        message: String,
+    },
+    RequestRendezvous {
+        scope: DataPlaneScope,
+        source: NodeEndpoint,
+        destination: NodeEndpoint,
+        direct_connectivity: bool,
+        failure_reason: String,
+    },
+    RequestSourcePreparation {
+        tenant: String,
+        project: String,
+        provider: SourceProviderKind,
+    },
+    CompleteSourcePreparation {
+        tenant: String,
+        project: String,
+        node: String,
+        provider: SourceProviderKind,
+        source_snapshot: Digest,
+    },
+    StartProcess {
+        tenant: String,
+        project: String,
+        #[serde(default)]
+        actor_user: Option<String>,
+        #[serde(default)]
+        actor_agent: Option<String>,
+        #[serde(default)]
+        agent_public_key_fingerprint: Option<Digest>,
+        #[serde(default)]
+        agent_signature: Option<AgentSignedRequest>,
+        process: String,
+        #[serde(default)]
+        restart: bool,
+    },
+    ReconnectNode {
+        node: String,
+        process: String,
+        epoch: u64,
+    },
+    CancelTask {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    },
+    CancelProcess {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    },
+    AbortProcess {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    },
+    ListProcesses {
+        tenant: String,
+        project: String,
+        actor_user: String,
+    },
+    QuotaStatus {
+        tenant: String,
+        project: String,
+        actor_user: String,
+    },
+    PollTaskControl {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    },
+    RestartTask {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        task: String,
+        #[serde(default)]
+        replacement_bundle: Option<TaskReplacementBundle>,
+    },
+    DebugAttach {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    },
+    SetDebugBreakpoints {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        probe_symbols: Vec<String>,
+    },
+    InspectDebugBreakpoints {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+    },
+    CreateDebugEpoch {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        stopped_task: String,
+        reason: String,
+    },
+    ResumeDebugEpoch {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        epoch: u64,
+    },
+    InspectDebugEpoch {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        epoch: u64,
+    },
+    PollDebugCommand {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+    },
+    ReportDebugState {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        epoch: u64,
+        state: DebugAcknowledgementState,
+        #[serde(default)]
+        stack_frames: Vec<String>,
+        #[serde(default)]
+        local_values: Vec<(String, String)>,
+        #[serde(default)]
+        task_args: Vec<(String, String)>,
+        #[serde(default)]
+        handles: Vec<(String, String)>,
+        #[serde(default)]
+        command_status: Option<String>,
+        #[serde(default)]
+        recent_output: Vec<String>,
+        #[serde(default)]
+        message: Option<String>,
+    },
+    ReportDebugProbeHit {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        probe_symbol: String,
+    },
+    ReportTaskLog {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        stdout_bytes: u64,
+        stderr_bytes: u64,
+        #[serde(default)]
+        stdout_tail: String,
+        #[serde(default)]
+        stderr_tail: String,
+        stdout_truncated: bool,
+        stderr_truncated: bool,
+        backpressured: bool,
+    },
+    ReportVfsMetadata {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        artifact_path: Option<String>,
+        artifact_digest: Option<Digest>,
+        artifact_size_bytes: Option<u64>,
+        large_bytes_uploaded: bool,
+    },
+    TaskCompleted {
+        tenant: String,
+        project: String,
+        process: String,
+        node: String,
+        task: String,
+        #[serde(default)]
+        terminal_state: Option<TaskTerminalState>,
+        status_code: Option<i32>,
+        stdout_bytes: u64,
+        stderr_bytes: u64,
+        #[serde(default)]
+        stdout_tail: String,
+        #[serde(default)]
+        stderr_tail: String,
+        #[serde(default)]
+        stdout_truncated: bool,
+        #[serde(default)]
+        stderr_truncated: bool,
+        artifact_path: Option<String>,
+        artifact_digest: Option<Digest>,
+        artifact_size_bytes: Option<u64>,
+        #[serde(default)]
+        result: Option<TaskBoundaryValue>,
+    },
+    ListTaskEvents {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        #[serde(default)]
+        process: Option<String>,
+    },
+    JoinTask {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        process: String,
+        task: String,
+    },
+    RenderOperatorPanel {
+        tenant: String,
+        project: String,
+        process: String,
+        actor_user: String,
+        max_download_bytes: u64,
+        stopped: bool,
+    },
+    SubmitPanelEvent {
+        tenant: String,
+        project: String,
+        process: String,
+        widget_id: String,
+        kind: PanelEventKind,
+        max_events: u64,
+    },
+    CreateArtifactDownloadLink {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        max_bytes: u64,
+        #[serde(default = "default_download_ttl_seconds")]
+        ttl_seconds: u64,
+    },
+    OpenArtifactDownloadStream {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        max_bytes: u64,
+        token_digest: Digest,
+        chunk_bytes: u64,
+    },
+    RevokeArtifactDownloadLink {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        token_digest: Digest,
+    },
+    ExportArtifactToNode {
+        tenant: String,
+        project: String,
+        actor_user: String,
+        artifact: String,
+        receiver_node: String,
+        direct_connectivity: bool,
+        failure_reason: String,
+    },
+}
+
+impl CoordinatorRequest {
+    pub fn operation(&self) -> Result<String, String> {
+        serde_json::to_value(self)
+            .map_err(|err| format!("failed to encode coordinator request operation: {err}"))
+            .map(|value| disasmer_core::coordinator_payload_operation(&value))
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case", deny_unknown_fields)]
+pub enum AuthenticatedCoordinatorRequest {
+    AuthStatus,
+    RevokeCliSession,
+    CreateProject {
+        project: String,
+        name: String,
+    },
+    SelectProject {
+        project: String,
+    },
+    ListProjects,
+    RegisterAgentPublicKey {
+        agent: String,
+        public_key: String,
+    },
+    ListAgentPublicKeys,
+    RotateAgentPublicKey {
+        agent: String,
+        public_key: String,
+    },
+    RevokeAgentPublicKey {
+        agent: String,
+    },
+    CreateNodeEnrollmentGrant {
+        #[serde(default = "default_node_enrollment_ttl_seconds")]
+        ttl_seconds: u64,
+    },
+    ListNodeDescriptors,
+    RevokeNodeCredential {
+        node: String,
+    },
+    StartProcess {
+        process: String,
+        #[serde(default)]
+        restart: bool,
+    },
+    ScheduleTask {
+        environment: Option<EnvironmentRequirements>,
+        environment_digest: Option<Digest>,
+        required_capabilities: Vec<Capability>,
+        #[serde(default)]
+        dependency_cache: Option<Digest>,
+        source_snapshot: Option<Digest>,
+        required_artifacts: Vec<String>,
+        prefer_node: Option<String>,
+    },
+    LaunchTask {
+        task_spec: Box<TaskSpec>,
+        #[serde(default)]
+        wait_for_node: bool,
+        artifact_path: String,
+        wasm_module_base64: String,
+    },
+    CancelProcess {
+        process: String,
+    },
+    AbortProcess {
+        process: String,
+    },
+    ListProcesses,
+    QuotaStatus,
+    RestartTask {
+        process: String,
+        task: String,
+        #[serde(default)]
+        replacement_bundle: Option<TaskReplacementBundle>,
+    },
+    DebugAttach {
+        process: String,
+    },
+    SetDebugBreakpoints {
+        process: String,
+        probe_symbols: Vec<String>,
+    },
+    InspectDebugBreakpoints {
+        process: String,
+    },
+    CreateDebugEpoch {
+        process: String,
+        stopped_task: String,
+        reason: String,
+    },
+    ResumeDebugEpoch {
+        process: String,
+        epoch: u64,
+    },
+    InspectDebugEpoch {
+        process: String,
+        epoch: u64,
+    },
+    ListTaskEvents {
+        #[serde(default)]
+        process: Option<String>,
+    },
+    JoinTask {
+        process: String,
+        task: String,
+    },
+    CreateArtifactDownloadLink {
+        artifact: String,
+        max_bytes: u64,
+        #[serde(default = "default_download_ttl_seconds")]
+        ttl_seconds: u64,
+    },
+    OpenArtifactDownloadStream {
+        artifact: String,
+        max_bytes: u64,
+        token_digest: Digest,
+        chunk_bytes: u64,
+    },
+    RevokeArtifactDownloadLink {
+        artifact: String,
+        token_digest: Digest,
+    },
+    ExportArtifactToNode {
+        artifact: String,
+        receiver_node: String,
+        direct_connectivity: bool,
+        failure_reason: String,
+    },
+}
+
+fn default_download_ttl_seconds() -> u64 {
+    900
+}
+
+fn default_node_enrollment_ttl_seconds() -> u64 {
+    900
+}

crates/disasmer-coordinator/src/service/protocol/responses.rs

@@ -0,0 +1,492 @@
+use super::*;
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TaskTerminalState {
+    Completed,
+    Failed,
+    Cancelled,
+}
+
+impl TaskTerminalState {
+    pub(crate) fn from_status_code(status_code: Option<i32>) -> Self {
+        match status_code {
+            Some(0) => Self::Completed,
+            _ => Self::Failed,
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TaskExecutor {
+    CoordinatorMain,
+    Node,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct TaskCompletionEvent {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub process: ProcessId,
+    pub node: NodeId,
+    pub executor: TaskExecutor,
+    pub task_definition: disasmer_core::TaskDefinitionId,
+    pub task: TaskInstanceId,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub placement: Option<Placement>,
+    pub terminal_state: TaskTerminalState,
+    pub status_code: Option<i32>,
+    pub stdout_bytes: u64,
+    pub stderr_bytes: u64,
+    pub stdout_tail: String,
+    pub stderr_tail: String,
+    pub stdout_truncated: bool,
+    pub stderr_truncated: bool,
+    pub artifact_path: Option<VfsPath>,
+    pub artifact_digest: Option<Digest>,
+    pub artifact_size_bytes: Option<u64>,
+    pub result: Option<TaskBoundaryValue>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DebugAuditEvent {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub process: ProcessId,
+    pub task: Option<TaskInstanceId>,
+    pub actor: UserId,
+    pub operation: String,
+    pub allowed: bool,
+    pub reason: String,
+    pub charged_debug_read_bytes: u64,
+    pub used_debug_read_bytes: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct WorkflowActor {
+    pub kind: String,
+    pub user: Option<UserId>,
+    pub agent: Option<AgentId>,
+    pub credential_kind: CredentialKind,
+    pub public_key_fingerprint: Option<Digest>,
+    pub authenticated_without_browser: bool,
+    pub scopes: Vec<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct TaskAssignment {
+    pub tenant: TenantId,
+    pub project: ProjectId,
+    pub process: ProcessId,
+    pub task: TaskInstanceId,
+    pub node: NodeId,
+    pub epoch: u64,
+    pub artifact_path: String,
+    pub task_spec: TaskSpec,
+    pub wasm_module_base64: String,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ArtifactTransferAssignment {
+    pub transfer_id: String,
+    pub artifact: ArtifactId,
+    pub expected_digest: Digest,
+    pub expected_size_bytes: u64,
+    pub offset: u64,
+    pub max_chunk_bytes: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct TaskCancellationTarget {
+    pub process: ProcessId,
+    pub task: TaskInstanceId,
+    pub node: NodeId,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum DebugAcknowledgementState {
+    Frozen,
+    Running,
+    Failed,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct DebugParticipantAcknowledgement {
+    pub node: NodeId,
+    pub task_definition: disasmer_core::TaskDefinitionId,
+    pub task: TaskInstanceId,
+    pub epoch: u64,
+    pub state: DebugAcknowledgementState,
+    pub stack_frames: Vec<String>,
+    pub local_values: Vec<(String, String)>,
+    pub task_args: Vec<(String, String)>,
+    pub handles: Vec<(String, String)>,
+    pub command_status: Option<String>,
+    pub recent_output: Vec<String>,
+    pub message: Option<String>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct VirtualProcessStatus {
+    pub process: ProcessId,
+    pub state: String,
+    pub main_task_definition: Option<disasmer_core::TaskDefinitionId>,
+    pub main_task_instance: Option<TaskInstanceId>,
+    pub main_state: Option<String>,
+    pub main_wait_state: Option<String>,
+    pub main_debug_epoch: Option<u64>,
+    pub connected_nodes: Vec<NodeId>,
+    pub coordinator_epoch: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub enum SourcePreparationDisposition {
+    Pending { reason: String },
+    Assigned { node: NodeId },
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub struct SourcePreparationStatus {
+    pub preparation: SourcePreparation,
+    pub disposition: SourcePreparationDisposition,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "type", rename_all = "snake_case")]
+pub enum CoordinatorResponse {
+    Pong {
+        epoch: u64,
+    },
+    AuthStatus {
+        tenant: TenantId,
+        project: ProjectId,
+        actor: UserId,
+        authenticated: bool,
+        account_status: String,
+        suspended: bool,
+        disabled: bool,
+        deleted: bool,
+        manual_review: bool,
+        sanitized_reason: Option<String>,
+        next_actions: Vec<String>,
+        private_moderation_details_exposed: bool,
+        signup_failure_details_exposed: bool,
+    },
+    AdminStatus {
+        tenant: TenantId,
+        actor: UserId,
+        suspended: bool,
+        safe_default: String,
+    },
+    TenantSuspended {
+        tenant: TenantId,
+        actor: UserId,
+        policy: crate::ServicePolicyRecord,
+    },
+    ProjectCreated {
+        project: ProjectRecord,
+        actor: UserId,
+    },
+    ProjectSelected {
+        project: ProjectRecord,
+        actor: UserId,
+    },
+    Projects {
+        projects: Vec<ProjectRecord>,
+        actor: UserId,
+    },
+    CliSessionRevoked {
+        tenant: TenantId,
+        project: ProjectId,
+        actor: UserId,
+    },
+    AgentPublicKey {
+        record: AgentPublicKeyRecord,
+        actor: UserId,
+    },
+    AgentPublicKeys {
+        records: Vec<AgentPublicKeyRecord>,
+        actor: UserId,
+    },
+    NodeAttached {
+        node: NodeId,
+        tenant: TenantId,
+        project: ProjectId,
+    },
+    NodeEnrollmentGrantCreated {
+        tenant: TenantId,
+        project: ProjectId,
+        grant: String,
+        scope: String,
+        expires_at_epoch_seconds: u64,
+    },
+    NodeEnrollmentExchanged {
+        node: NodeId,
+        tenant: TenantId,
+        project: ProjectId,
+        credential: disasmer_core::NodeCredential,
+    },
+    NodeHeartbeat {
+        node: NodeId,
+        epoch: u64,
+    },
+    NodeCapabilitiesRecorded {
+        node: NodeId,
+        node_descriptors: usize,
+    },
+    NodeDescriptors {
+        descriptors: Vec<NodeDescriptor>,
+        actor: UserId,
+    },
+    NodeCredentialRevoked {
+        node: NodeId,
+        tenant: TenantId,
+        project: ProjectId,
+        actor: UserId,
+        descriptor_removed: bool,
+        queued_assignments_removed: usize,
+    },
+    TaskPlacement {
+        placement: Placement,
+    },
+    TaskLaunched {
+        process: ProcessId,
+        task: TaskInstanceId,
+        actor: WorkflowActor,
+        placement: Placement,
+        assignment: Box<TaskAssignment>,
+        charged_spawns: u64,
+    },
+    MainLaunched {
+        process: ProcessId,
+        task_definition: disasmer_core::TaskDefinitionId,
+        task_instance: TaskInstanceId,
+        actor: WorkflowActor,
+        state: String,
+    },
+    TaskQueued {
+        process: ProcessId,
+        task: TaskInstanceId,
+        actor: WorkflowActor,
+        reason: String,
+        charged_spawns: u64,
+        queued_tasks: usize,
+    },
+    TaskAssignment {
+        assignment: Option<Box<TaskAssignment>>,
+    },
+    ArtifactTransferAssignment {
+        transfer: Option<ArtifactTransferAssignment>,
+    },
+    ArtifactTransferChunkAccepted {
+        transfer_id: String,
+        next_offset: u64,
+        complete: bool,
+    },
+    ArtifactTransferFailed {
+        transfer_id: String,
+    },
+    RendezvousPlan {
+        plan: DirectBulkTransferPlan,
+        charged_rendezvous_attempts: u64,
+    },
+    SourcePreparation {
+        status: SourcePreparationStatus,
+    },
+    SourcePreparationCompleted {
+        node: NodeId,
+        provider: SourceProviderKind,
+        source_snapshot: Digest,
+    },
+    ProcessStarted {
+        process: ProcessId,
+        epoch: u64,
+        actor: WorkflowActor,
+        charged_spawns: u64,
+    },
+    NodeReconnected {
+        node: NodeId,
+        process: ProcessId,
+    },
+    TaskCancellationRequested {
+        process: ProcessId,
+        task: TaskInstanceId,
+        node: NodeId,
+    },
+    ProcessCancellationRequested {
+        process: ProcessId,
+        cancelled_tasks: Vec<TaskCancellationTarget>,
+        affected_nodes: Vec<NodeId>,
+    },
+    ProcessAborted {
+        process: ProcessId,
+        aborted_tasks: Vec<TaskCancellationTarget>,
+        affected_nodes: Vec<NodeId>,
+    },
+    ProcessStatuses {
+        processes: Vec<VirtualProcessStatus>,
+        actor: UserId,
+    },
+    QuotaStatus {
+        tenant: TenantId,
+        project: ProjectId,
+        actor: UserId,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        policy_label: Option<String>,
+        limits: ResourceLimits,
+        window_seconds: BTreeMap<LimitKind, u64>,
+        usage: BTreeMap<LimitKind, u64>,
+        window_started_epoch_seconds: BTreeMap<LimitKind, u64>,
+    },
+    TaskControl {
+        process: ProcessId,
+        task: TaskInstanceId,
+        cancel_requested: bool,
+        abort_requested: bool,
+    },
+    TaskRestart {
+        process: ProcessId,
+        task: TaskInstanceId,
+        restarted_task_instance: Option<disasmer_core::TaskInstanceId>,
+        actor: UserId,
+        accepted: bool,
+        clean_boundary_available: bool,
+        active_task: bool,
+        completed_event_observed: bool,
+        requires_whole_process_restart: bool,
+        message: String,
+        audit_event: DebugAuditEvent,
+        charged_debug_read_bytes: u64,
+        used_debug_read_bytes: u64,
+    },
+    DebugCommand {
+        process: ProcessId,
+        task: TaskInstanceId,
+        epoch: Option<u64>,
+        command: Option<String>,
+    },
+    DebugStateRecorded {
+        process: ProcessId,
+        node: NodeId,
+        task: TaskInstanceId,
+        epoch: u64,
+        state: DebugAcknowledgementState,
+    },
+    DebugAttach {
+        process: ProcessId,
+        actor: UserId,
+        authorization: Authorization,
+        audit_event: DebugAuditEvent,
+        charged_debug_read_bytes: u64,
+        used_debug_read_bytes: u64,
+    },
+    DebugBreakpoints {
+        process: ProcessId,
+        actor: UserId,
+        probe_symbols: Vec<String>,
+        hit_epoch: Option<u64>,
+        hit_task: Option<TaskInstanceId>,
+        hit_probe_symbol: Option<String>,
+        audit_event: DebugAuditEvent,
+        charged_debug_read_bytes: u64,
+        used_debug_read_bytes: u64,
+    },
+    DebugProbeHit {
+        process: ProcessId,
+        node: NodeId,
+        task: TaskInstanceId,
+        probe_symbol: String,
+        breakpoint_matched: bool,
+        debug_epoch: Option<u64>,
+    },
+    DebugEpoch {
+        process: ProcessId,
+        actor: UserId,
+        epoch: u64,
+        command: String,
+        affected_tasks: Vec<TaskCancellationTarget>,
+        all_stop_requested: bool,
+        audit_event: DebugAuditEvent,
+        charged_debug_read_bytes: u64,
+        used_debug_read_bytes: u64,
+    },
+    DebugEpochStatus {
+        process: ProcessId,
+        actor: UserId,
+        epoch: u64,
+        command: String,
+        expected_tasks: Vec<TaskCancellationTarget>,
+        acknowledgements: Vec<DebugParticipantAcknowledgement>,
+        fully_frozen: bool,
+        fully_resumed: bool,
+        failed: bool,
+        failure_messages: Vec<String>,
+        audit_event: DebugAuditEvent,
+        charged_debug_read_bytes: u64,
+        used_debug_read_bytes: u64,
+    },
+    TaskLogRecorded {
+        process: ProcessId,
+        task: TaskInstanceId,
+        stdout_bytes: u64,
+        stderr_bytes: u64,
+        stdout_tail: String,
+        stderr_tail: String,
+        backpressured: bool,
+    },
+    VfsMetadataRecorded {
+        process: ProcessId,
+        task: TaskInstanceId,
+        artifact_path: Option<VfsPath>,
+        large_bytes_uploaded: bool,
+    },
+    TaskRecorded {
+        process: ProcessId,
+        task: TaskInstanceId,
+        events_recorded: usize,
+    },
+    TaskEvents {
+        events: Vec<TaskCompletionEvent>,
+    },
+    TaskJoined {
+        join: TaskJoinResult,
+    },
+    OperatorPanel {
+        panel: PanelState,
+    },
+    PanelEventAccepted {
+        used_events: u64,
+        max_events: u64,
+    },
+    ArtifactDownloadLink {
+        link: DownloadLink,
+    },
+    ArtifactDownloadLinkRevoked {
+        link: DownloadLink,
+    },
+    ArtifactDownloadStream {
+        link: DownloadLink,
+        streamed_bytes: u64,
+        charged_download_bytes: u64,
+        content_bytes_available: bool,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        content_offset: Option<u64>,
+        #[serde(default)]
+        content_eof: bool,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        content_base64: Option<String>,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        content_source: Option<String>,
+    },
+    ArtifactExportPlan {
+        plan: DirectBulkTransferPlan,
+        source_node: NodeId,
+        receiver_node: NodeId,
+        artifact_size_bytes: u64,
+    },
+    Error {
+        message: String,
+    },
+}

crates/disasmer-coordinator/src/service/quota.rs

@@ -0,0 +1,421 @@
+use std::collections::BTreeMap;
+
+use disasmer_core::{LimitError, LimitKind, ProjectId, ResourceLimits, ResourceMeter, TenantId};
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct CoordinatorQuotaConfiguration {
+    pub limits: ResourceLimits,
+    pub window_seconds: BTreeMap<LimitKind, u64>,
+    pub policy_label: Option<String>,
+}
+
+pub(super) struct CoordinatorQuotaStatus {
+    pub(super) policy_label: Option<String>,
+    pub(super) limits: ResourceLimits,
+    pub(super) window_seconds: BTreeMap<LimitKind, u64>,
+    pub(super) usage: BTreeMap<LimitKind, u64>,
+    pub(super) window_started_epoch_seconds: BTreeMap<LimitKind, u64>,
+}
+
+impl CoordinatorQuotaConfiguration {
+    pub fn new(
+        limits: ResourceLimits,
+        window_seconds: impl IntoIterator<Item = (LimitKind, u64)>,
+    ) -> Result<Self, String> {
+        let window_seconds = window_seconds.into_iter().collect::<BTreeMap<_, _>>();
+        if window_seconds.values().any(|seconds| *seconds == 0) {
+            return Err("quota windows must be at least one second".to_owned());
+        }
+        Ok(Self {
+            limits,
+            window_seconds,
+            policy_label: None,
+        })
+    }
+
+    pub fn with_policy_label(mut self, label: impl Into<String>) -> Self {
+        let label = label.into();
+        self.policy_label = (!label.trim().is_empty()).then_some(label);
+        self
+    }
+
+    pub fn unlimited() -> Self {
+        Self {
+            limits: ResourceLimits::unlimited(),
+            window_seconds: LimitKind::ALL
+                .into_iter()
+                .map(|kind| (kind, u64::MAX))
+                .collect(),
+            policy_label: None,
+        }
+    }
+
+    pub fn window_seconds(&self, kind: LimitKind) -> u64 {
+        self.window_seconds
+            .get(&kind)
+            .copied()
+            .unwrap_or(u64::MAX)
+            .max(1)
+    }
+}
+
+impl Default for CoordinatorQuotaConfiguration {
+    fn default() -> Self {
+        Self::unlimited()
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
+struct ProjectQuotaScope {
+    tenant: TenantId,
+    project: ProjectId,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord)]
+struct MeterKey {
+    scope: ProjectQuotaScope,
+    kind: LimitKind,
+    window: u64,
+}
+
+#[derive(Clone, Debug)]
+pub(super) struct CoordinatorQuota {
+    configuration: CoordinatorQuotaConfiguration,
+    meters: BTreeMap<MeterKey, ResourceMeter>,
+}
+
+impl Default for CoordinatorQuota {
+    fn default() -> Self {
+        Self::new(CoordinatorQuotaConfiguration::default())
+    }
+}
+
+impl CoordinatorQuota {
+    pub(super) fn new(configuration: CoordinatorQuotaConfiguration) -> Self {
+        Self {
+            configuration,
+            meters: BTreeMap::new(),
+        }
+    }
+
+    fn key(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        now_epoch_seconds: u64,
+    ) -> MeterKey {
+        MeterKey {
+            scope: ProjectQuotaScope {
+                tenant: tenant.clone(),
+                project: project.clone(),
+            },
+            kind,
+            window: now_epoch_seconds / self.configuration.window_seconds(kind),
+        }
+    }
+
+    fn meter(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        now_epoch_seconds: u64,
+    ) -> Option<&ResourceMeter> {
+        self.meters
+            .get(&self.key(tenant, project, kind, now_epoch_seconds))
+    }
+
+    fn meter_mut(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        now_epoch_seconds: u64,
+    ) -> &mut ResourceMeter {
+        let key = self.key(tenant, project, kind, now_epoch_seconds);
+        self.meters.retain(|existing, _| {
+            existing.scope != key.scope || existing.kind != kind || existing.window == key.window
+        });
+        self.meters.entry(key).or_default()
+    }
+
+    fn can_charge(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        amount: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<(), LimitError> {
+        self.meter(tenant, project, kind, now_epoch_seconds)
+            .cloned()
+            .unwrap_or_default()
+            .can_charge(&self.configuration.limits, kind, amount)
+    }
+
+    fn charge(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        amount: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        let limits = self.configuration.limits.clone();
+        let meter = self.meter_mut(tenant, project, kind, now_epoch_seconds);
+        meter.charge(&limits, kind, amount)?;
+        Ok(meter.used(&kind))
+    }
+
+    fn used(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        kind: LimitKind,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.meter(tenant, project, kind, now_epoch_seconds)
+            .map_or(0, |meter| meter.used(&kind))
+    }
+
+    pub(super) fn can_charge_workflow_spawn(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> Result<(), LimitError> {
+        self.can_charge(tenant, project, LimitKind::Spawn, 1, now_epoch_seconds)
+    }
+
+    pub(super) fn charge_api_call(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(tenant, project, LimitKind::ApiCall, 1, now_epoch_seconds)
+    }
+
+    pub(super) fn can_charge_log_bytes(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        bytes: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<(), LimitError> {
+        self.can_charge(
+            tenant,
+            project,
+            LimitKind::LogBytes,
+            bytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn charge_log_bytes(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        bytes: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(
+            tenant,
+            project,
+            LimitKind::LogBytes,
+            bytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn charge_workflow_spawn(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(tenant, project, LimitKind::Spawn, 1, now_epoch_seconds)
+    }
+
+    #[cfg(test)]
+    pub(super) fn used_workflow_spawns(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.used(tenant, project, LimitKind::Spawn, now_epoch_seconds)
+    }
+
+    #[cfg(test)]
+    pub(super) fn used_api_calls(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.used(tenant, project, LimitKind::ApiCall, now_epoch_seconds)
+    }
+
+    #[cfg(test)]
+    pub(super) fn used_log_bytes(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.used(tenant, project, LimitKind::LogBytes, now_epoch_seconds)
+    }
+
+    #[cfg(test)]
+    pub(super) fn active_meter_count(&self) -> usize {
+        self.meters.len()
+    }
+
+    pub(super) fn charge_rendezvous_attempt(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(
+            tenant,
+            project,
+            LimitKind::RendezvousAttempt,
+            1,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn can_charge_download(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        bytes: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<(), LimitError> {
+        self.can_charge(
+            tenant,
+            project,
+            LimitKind::ArtifactDownloadBytes,
+            bytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn charge_download(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        bytes: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(
+            tenant,
+            project,
+            LimitKind::ArtifactDownloadBytes,
+            bytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn download_limit(&self) -> u64 {
+        self.configuration
+            .limits
+            .limit(&LimitKind::ArtifactDownloadBytes)
+    }
+
+    pub(super) fn used_download_bytes(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.used(
+            tenant,
+            project,
+            LimitKind::ArtifactDownloadBytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn charge_debug_read(
+        &mut self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        bytes: u64,
+        now_epoch_seconds: u64,
+    ) -> Result<u64, LimitError> {
+        self.charge(
+            tenant,
+            project,
+            LimitKind::DebugReadBytes,
+            bytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn used_debug_read_bytes(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> u64 {
+        self.used(
+            tenant,
+            project,
+            LimitKind::DebugReadBytes,
+            now_epoch_seconds,
+        )
+    }
+
+    pub(super) fn project_status(
+        &self,
+        tenant: &TenantId,
+        project: &ProjectId,
+        now_epoch_seconds: u64,
+    ) -> CoordinatorQuotaStatus {
+        let mut usage = BTreeMap::new();
+        let mut window_starts = BTreeMap::new();
+        for kind in LimitKind::ALL {
+            let seconds = self.configuration.window_seconds(kind);
+            usage.insert(kind, self.used(tenant, project, kind, now_epoch_seconds));
+            window_starts.insert(kind, (now_epoch_seconds / seconds).saturating_mul(seconds));
+        }
+        CoordinatorQuotaStatus {
+            policy_label: self.configuration.policy_label.clone(),
+            limits: self.configuration.limits.clone(),
+            window_seconds: self.configuration.window_seconds.clone(),
+            usage,
+            window_started_epoch_seconds: window_starts,
+        }
+    }
+
+    #[cfg(test)]
+    pub(super) fn set_workflow_limits(&mut self, limits: ResourceLimits) {
+        self.configuration
+            .limits
+            .limits
+            .insert(LimitKind::Spawn, limits.limit(&LimitKind::Spawn));
+        self.meters.clear();
+    }
+
+    #[cfg(test)]
+    pub(super) fn set_download_limits(&mut self, limits: ResourceLimits) {
+        self.configuration.limits.limits.insert(
+            LimitKind::ArtifactDownloadBytes,
+            limits.limit(&LimitKind::ArtifactDownloadBytes),
+        );
+        self.meters.clear();
+    }
+
+    #[cfg(test)]
+    pub(super) fn set_rendezvous_limits(&mut self, limits: ResourceLimits) {
+        self.configuration.limits.limits.insert(
+            LimitKind::RendezvousAttempt,
+            limits.limit(&LimitKind::RendezvousAttempt),
+        );
+        self.meters.clear();
+    }
+}

crates/disasmer-coordinator/src/service/routing.rs

@@ -0,0 +1,778 @@
+use super::*;
+
+impl CoordinatorService {
+    pub fn handle_request(
+        &mut self,
+        request: CoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        self.pump_main_runtime_commands();
+        let request_payload = serde_json::to_value(&request).map_err(|error| {
+            CoordinatorServiceError::Protocol(format!(
+                "failed to canonicalize coordinator request for authentication: {error}"
+            ))
+        })?;
+        let request_payload_digest = disasmer_core::signed_request_payload_digest(&request_payload);
+        match request {
+            CoordinatorRequest::Ping => Ok(CoordinatorResponse::Pong {
+                epoch: self.coordinator.coordinator_epoch(),
+            }),
+            CoordinatorRequest::Authenticated {
+                session_secret,
+                request,
+            } => self.handle_authenticated_request(session_secret, request),
+            CoordinatorRequest::AuthStatus {
+                tenant,
+                project,
+                actor_user,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let project = ProjectId::new(project);
+                let actor = UserId::new(actor_user);
+                let account_state = self.coordinator.account_policy_state(&tenant);
+                Ok(CoordinatorResponse::AuthStatus {
+                    tenant,
+                    project,
+                    actor,
+                    authenticated: true,
+                    account_status: account_state.account_status,
+                    suspended: account_state.suspended,
+                    disabled: account_state.disabled,
+                    deleted: account_state.deleted,
+                    manual_review: account_state.manual_review,
+                    sanitized_reason: account_state.sanitized_reason,
+                    next_actions: account_state.next_actions,
+                    private_moderation_details_exposed: false,
+                    signup_failure_details_exposed: false,
+                })
+            }
+            CoordinatorRequest::AdminStatus {
+                tenant,
+                actor_user,
+                admin_proof,
+                admin_nonce,
+                issued_at_epoch_seconds,
+            } => self.handle_admin_status(
+                tenant,
+                actor_user,
+                admin_proof,
+                admin_nonce,
+                issued_at_epoch_seconds,
+            ),
+            CoordinatorRequest::SuspendTenant {
+                tenant,
+                actor_user,
+                target_tenant,
+                admin_proof,
+                admin_nonce,
+                issued_at_epoch_seconds,
+            } => self.handle_suspend_tenant(
+                tenant,
+                actor_user,
+                target_tenant,
+                admin_proof,
+                admin_nonce,
+                issued_at_epoch_seconds,
+            ),
+            CoordinatorRequest::CreateProject {
+                tenant,
+                actor_user,
+                project,
+                name,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let actor = UserId::new(actor_user);
+                let project = ProjectId::new(project);
+                self.coordinator.ensure_tenant_active(&tenant)?;
+                if let Some(existing) = self.coordinator.project(&project) {
+                    if existing.tenant != tenant {
+                        return Err(CoordinatorError::Unauthorized(
+                            "project id is outside the signed-in tenant scope".to_owned(),
+                        )
+                        .into());
+                    }
+                }
+                self.coordinator.upsert_tenant(tenant.clone());
+                self.coordinator.upsert_user(
+                    tenant.clone(),
+                    actor.clone(),
+                    CredentialKind::BrowserSession,
+                );
+                self.coordinator
+                    .upsert_project(tenant.clone(), project.clone(), name);
+                self.coordinator.grant_project_debug(
+                    tenant.clone(),
+                    project.clone(),
+                    actor.clone(),
+                );
+                self.persist_durable_state()?;
+                let project = self
+                    .coordinator
+                    .project(&project)
+                    .expect("project was just created")
+                    .clone();
+                Ok(CoordinatorResponse::ProjectCreated { project, actor })
+            }
+            CoordinatorRequest::SelectProject {
+                tenant,
+                actor_user,
+                project,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let actor = UserId::new(actor_user);
+                let project_id = ProjectId::new(project);
+                let project = self
+                    .coordinator
+                    .project(&project_id)
+                    .ok_or_else(|| {
+                        CoordinatorError::Unauthorized(
+                            "project is not visible to the signed-in user".to_owned(),
+                        )
+                    })?
+                    .clone();
+                if project.tenant != tenant {
+                    return Err(CoordinatorError::Unauthorized(
+                        "project is outside the signed-in tenant scope".to_owned(),
+                    )
+                    .into());
+                }
+                self.coordinator
+                    .upsert_user(tenant, actor.clone(), CredentialKind::BrowserSession);
+                self.persist_durable_state()?;
+                Ok(CoordinatorResponse::ProjectSelected { project, actor })
+            }
+            CoordinatorRequest::ListProjects { tenant, actor_user } => {
+                let tenant = TenantId::new(tenant);
+                let actor = UserId::new(actor_user);
+                let context = disasmer_core::AuthContext {
+                    tenant: tenant.clone(),
+                    project: ProjectId::from("__project_listing__"),
+                    actor: Actor::User(actor.clone()),
+                };
+                self.coordinator
+                    .upsert_user(tenant, actor.clone(), CredentialKind::BrowserSession);
+                self.persist_durable_state()?;
+                Ok(CoordinatorResponse::Projects {
+                    projects: self.coordinator.list_projects(&context),
+                    actor,
+                })
+            }
+            CoordinatorRequest::RegisterAgentPublicKey {
+                tenant,
+                project,
+                user,
+                agent,
+                public_key,
+            }
+            | CoordinatorRequest::RotateAgentPublicKey {
+                tenant,
+                project,
+                user,
+                agent,
+                public_key,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let project = ProjectId::new(project);
+                let actor = UserId::new(user);
+                let agent = AgentId::new(agent);
+                self.coordinator.ensure_tenant_active(&tenant)?;
+                if let Some(existing) = self.coordinator.project(&project) {
+                    if existing.tenant != tenant {
+                        return Err(CoordinatorError::Unauthorized(
+                            "project id is outside the signed-in tenant scope".to_owned(),
+                        )
+                        .into());
+                    }
+                }
+                self.coordinator.upsert_tenant(tenant.clone());
+                self.coordinator.upsert_user(
+                    tenant.clone(),
+                    actor.clone(),
+                    CredentialKind::CliDeviceSession,
+                );
+                self.coordinator
+                    .upsert_project(tenant.clone(), project.clone(), "local");
+                let record = self.coordinator.register_agent_public_key(
+                    tenant,
+                    project,
+                    actor.clone(),
+                    agent,
+                    public_key,
+                );
+                self.persist_durable_state()?;
+                Ok(CoordinatorResponse::AgentPublicKey { record, actor })
+            }
+            CoordinatorRequest::ListAgentPublicKeys {
+                tenant,
+                project,
+                user,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let project = ProjectId::new(project);
+                let actor = UserId::new(user);
+                let context = disasmer_core::AuthContext {
+                    tenant,
+                    project,
+                    actor: Actor::User(actor.clone()),
+                };
+                Ok(CoordinatorResponse::AgentPublicKeys {
+                    records: self.coordinator.list_agent_public_keys(&context),
+                    actor,
+                })
+            }
+            CoordinatorRequest::RevokeAgentPublicKey {
+                tenant,
+                project,
+                user,
+                agent,
+            } => {
+                let tenant = TenantId::new(tenant);
+                let project = ProjectId::new(project);
+                let actor = UserId::new(user);
+                let agent = AgentId::new(agent);
+                let context = disasmer_core::AuthContext {
+                    tenant,
+                    project,
+                    actor: Actor::User(actor.clone()),
+                };
+                let record = self.coordinator.revoke_agent_public_key(&context, &agent)?;
+                self.persist_durable_state()?;
+                Ok(CoordinatorResponse::AgentPublicKey { record, actor })
+            }
+            CoordinatorRequest::AttachNode {
+                tenant,
+                project,
+                node,
+                public_key,
+            } => self.handle_attach_node(tenant, project, node, public_key),
+            CoordinatorRequest::CreateNodeEnrollmentGrant {
+                tenant,
+                project,
+                actor_user,
+                ttl_seconds,
+            } => self.handle_create_node_enrollment_grant(tenant, project, actor_user, ttl_seconds),
+            CoordinatorRequest::ExchangeNodeEnrollmentGrant {
+                tenant,
+                project,
+                node,
+                public_key,
+                enrollment_grant,
+            } => self.handle_exchange_node_enrollment_grant(
+                tenant,
+                project,
+                node,
+                public_key,
+                enrollment_grant,
+            ),
+            CoordinatorRequest::NodeHeartbeat {
+                node,
+                node_signature,
+            } => self.handle_node_heartbeat(node, node_signature, &request_payload_digest),
+            CoordinatorRequest::SignedNode {
+                node,
+                node_signature,
+                request,
+            } => self.handle_signed_node_request(node, node_signature, *request),
+            CoordinatorRequest::ReportNodeCapabilities { .. } => {
+                self.reject_unsigned_node_request()
+            }
+            CoordinatorRequest::ListNodeDescriptors {
+                tenant,
+                project,
+                actor_user,
+            } => self.handle_list_node_descriptors(tenant, project, actor_user),
+            CoordinatorRequest::RevokeNodeCredential {
+                tenant,
+                project,
+                actor_user,
+                node,
+            } => self.handle_revoke_node_credential(tenant, project, actor_user, node),
+            CoordinatorRequest::ScheduleTask {
+                tenant,
+                project,
+                environment,
+                environment_digest,
+                required_capabilities,
+                dependency_cache,
+                source_snapshot,
+                required_artifacts,
+                prefer_node,
+            } => self.handle_schedule_task(
+                tenant,
+                project,
+                environment,
+                environment_digest,
+                required_capabilities,
+                dependency_cache,
+                source_snapshot,
+                required_artifacts,
+                prefer_node,
+            ),
+            CoordinatorRequest::LaunchTask {
+                tenant,
+                project,
+                actor_user,
+                actor_agent,
+                agent_public_key_fingerprint,
+                agent_signature,
+                task_spec,
+                wait_for_node,
+                artifact_path,
+                wasm_module_base64,
+            } => self.handle_launch_task(
+                tenant,
+                project,
+                actor_user,
+                actor_agent,
+                agent_public_key_fingerprint,
+                agent_signature,
+                Some(&request_payload_digest),
+                task_spec,
+                wait_for_node,
+                artifact_path,
+                wasm_module_base64,
+            ),
+            CoordinatorRequest::LaunchChildTask { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::JoinChildTask { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::PollTaskAssignment { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::RequestRendezvous {
+                scope,
+                source,
+                destination,
+                direct_connectivity,
+                failure_reason,
+            } => self.handle_request_rendezvous(
+                scope,
+                source,
+                destination,
+                direct_connectivity,
+                failure_reason,
+            ),
+            CoordinatorRequest::RequestSourcePreparation {
+                tenant,
+                project,
+                provider,
+            } => self.handle_request_source_preparation(tenant, project, provider),
+            CoordinatorRequest::CompleteSourcePreparation { .. } => {
+                self.reject_unsigned_node_request()
+            }
+            CoordinatorRequest::StartProcess {
+                tenant,
+                project,
+                actor_user,
+                actor_agent,
+                agent_public_key_fingerprint,
+                agent_signature,
+                process,
+                restart,
+            } => self.handle_start_process(
+                tenant,
+                project,
+                actor_user,
+                actor_agent,
+                agent_public_key_fingerprint,
+                agent_signature,
+                Some(&request_payload_digest),
+                process,
+                restart,
+            ),
+            CoordinatorRequest::ReconnectNode { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::CancelTask {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+            } => self.handle_cancel_task(tenant, project, process, node, task),
+            CoordinatorRequest::CancelProcess {
+                tenant,
+                project,
+                actor_user,
+                process,
+            } => self.handle_cancel_process(tenant, project, actor_user, process),
+            CoordinatorRequest::AbortProcess {
+                tenant,
+                project,
+                actor_user,
+                process,
+            } => self.handle_abort_process(tenant, project, actor_user, process),
+            CoordinatorRequest::ListProcesses {
+                tenant,
+                project,
+                actor_user,
+            } => self.handle_list_processes(tenant, project, actor_user),
+            CoordinatorRequest::QuotaStatus {
+                tenant,
+                project,
+                actor_user,
+            } => self.handle_quota_status(tenant, project, actor_user),
+            CoordinatorRequest::PollTaskControl { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::PollArtifactTransfer { .. }
+            | CoordinatorRequest::UploadArtifactTransferChunk { .. }
+            | CoordinatorRequest::FailArtifactTransfer { .. } => {
+                self.reject_unsigned_node_request()
+            }
+            CoordinatorRequest::RestartTask {
+                tenant,
+                project,
+                actor_user,
+                process,
+                task,
+                replacement_bundle,
+            } => self.handle_restart_task(
+                tenant,
+                project,
+                actor_user,
+                process,
+                task,
+                replacement_bundle,
+            ),
+            request @ (CoordinatorRequest::DebugAttach { .. }
+            | CoordinatorRequest::SetDebugBreakpoints { .. }
+            | CoordinatorRequest::InspectDebugBreakpoints { .. }
+            | CoordinatorRequest::CreateDebugEpoch { .. }
+            | CoordinatorRequest::ResumeDebugEpoch { .. }
+            | CoordinatorRequest::InspectDebugEpoch { .. }) => self.handle_debug_request(request),
+            CoordinatorRequest::PollDebugCommand { .. }
+            | CoordinatorRequest::ReportDebugState { .. }
+            | CoordinatorRequest::ReportDebugProbeHit { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::ReportTaskLog { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::ReportVfsMetadata { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::TaskCompleted { .. } => self.reject_unsigned_node_request(),
+            CoordinatorRequest::ListTaskEvents {
+                tenant,
+                project,
+                actor_user,
+                process,
+            } => self.handle_list_task_events(tenant, project, actor_user, process),
+            CoordinatorRequest::JoinTask {
+                tenant,
+                project,
+                actor_user,
+                process,
+                task,
+            } => self.handle_join_task(tenant, project, actor_user, process, task),
+            CoordinatorRequest::RenderOperatorPanel {
+                tenant,
+                project,
+                process,
+                actor_user,
+                max_download_bytes,
+                stopped,
+            } => self.handle_render_operator_panel(
+                tenant,
+                project,
+                actor_user,
+                process,
+                max_download_bytes,
+                stopped,
+            ),
+            CoordinatorRequest::SubmitPanelEvent {
+                tenant,
+                project,
+                process,
+                widget_id,
+                kind,
+                max_events,
+            } => self
+                .handle_submit_panel_event(tenant, project, process, widget_id, kind, max_events),
+            CoordinatorRequest::CreateArtifactDownloadLink {
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                max_bytes,
+                ttl_seconds,
+            } => self.handle_create_artifact_download_link(
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                max_bytes,
+                ttl_seconds,
+            ),
+            CoordinatorRequest::OpenArtifactDownloadStream {
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                max_bytes,
+                token_digest,
+                chunk_bytes,
+            } => self.handle_open_artifact_download_stream(
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                max_bytes,
+                token_digest,
+                chunk_bytes,
+            ),
+            CoordinatorRequest::RevokeArtifactDownloadLink {
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                token_digest,
+            } => self.handle_revoke_artifact_download_link(
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                token_digest,
+            ),
+            CoordinatorRequest::ExportArtifactToNode {
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                receiver_node,
+                direct_connectivity,
+                failure_reason,
+            } => self.handle_export_artifact_to_node(
+                tenant,
+                project,
+                actor_user,
+                artifact,
+                receiver_node,
+                direct_connectivity,
+                failure_reason,
+            ),
+        }
+    }
+
+    fn handle_authenticated_request(
+        &mut self,
+        session_secret: String,
+        request: AuthenticatedCoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let context = self.coordinator.authenticate_cli_session(&session_secret)?;
+        let authorized = authorize_authenticated_user_operation(&context, &request)?;
+        let now_epoch_seconds = self.current_epoch_seconds()?;
+        self.quota
+            .charge_api_call(&context.tenant, &context.project, now_epoch_seconds)?;
+        let _authorized_operation = authorized.operation;
+        let actor = authorized.actor;
+        match request {
+            AuthenticatedCoordinatorRequest::AuthStatus => {
+                let account_state = self.coordinator.account_policy_state(&context.tenant);
+                Ok(CoordinatorResponse::AuthStatus {
+                    tenant: context.tenant,
+                    project: context.project,
+                    actor,
+                    authenticated: true,
+                    account_status: account_state.account_status,
+                    suspended: account_state.suspended,
+                    disabled: account_state.disabled,
+                    deleted: account_state.deleted,
+                    manual_review: account_state.manual_review,
+                    sanitized_reason: account_state.sanitized_reason,
+                    next_actions: account_state.next_actions,
+                    private_moderation_details_exposed: false,
+                    signup_failure_details_exposed: false,
+                })
+            }
+            AuthenticatedCoordinatorRequest::RevokeCliSession => {
+                self.coordinator.revoke_cli_session(&session_secret)?;
+                self.persist_durable_state()?;
+                Ok(CoordinatorResponse::CliSessionRevoked {
+                    tenant: context.tenant,
+                    project: context.project,
+                    actor,
+                })
+            }
+            AuthenticatedCoordinatorRequest::CreateProject { project, name } => {
+                let project = ProjectId::new(project);
+                self.coordinator.ensure_tenant_active(&context.tenant)?;
+                if let Some(existing) = self.coordinator.project(&project) {
+                    if existing.tenant != context.tenant {
+                        return Err(CoordinatorError::Unauthorized(
+                            "project id is outside the authenticated tenant scope".to_owned(),
+                        )
+                        .into());
+                    }
+                }
+                self.coordinator
+                    .upsert_project(context.tenant.clone(), project.clone(), name);
+                self.coordinator.grant_project_debug(
+                    context.tenant.clone(),
+                    project.clone(),
+                    actor.clone(),
+                );
+                self.persist_durable_state()?;
+                let project = self
+                    .coordinator
+                    .project(&project)
+                    .expect("project was just created")
+                    .clone();
+                Ok(CoordinatorResponse::ProjectCreated { project, actor })
+            }
+            AuthenticatedCoordinatorRequest::SelectProject { project } => {
+                let project_id = ProjectId::new(project);
+                let project = self
+                    .coordinator
+                    .project(&project_id)
+                    .ok_or_else(|| {
+                        CoordinatorError::Unauthorized(
+                            "project is not visible to the authenticated user".to_owned(),
+                        )
+                    })?
+                    .clone();
+                if project.tenant != context.tenant {
+                    return Err(CoordinatorError::Unauthorized(
+                        "project is outside the authenticated tenant scope".to_owned(),
+                    )
+                    .into());
+                }
+                Ok(CoordinatorResponse::ProjectSelected { project, actor })
+            }
+            AuthenticatedCoordinatorRequest::ListProjects => Ok(CoordinatorResponse::Projects {
+                projects: self.coordinator.list_projects(&context),
+                actor,
+            }),
+            request @ (AuthenticatedCoordinatorRequest::RegisterAgentPublicKey { .. }
+            | AuthenticatedCoordinatorRequest::ListAgentPublicKeys
+            | AuthenticatedCoordinatorRequest::RotateAgentPublicKey { .. }
+            | AuthenticatedCoordinatorRequest::RevokeAgentPublicKey { .. }) => {
+                self.handle_authenticated_agent_key_request(&context, &actor, request)
+            }
+            AuthenticatedCoordinatorRequest::CreateNodeEnrollmentGrant { ttl_seconds } => self
+                .handle_create_node_enrollment_grant(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    ttl_seconds,
+                ),
+            AuthenticatedCoordinatorRequest::ListNodeDescriptors => self
+                .handle_list_node_descriptors(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                ),
+            AuthenticatedCoordinatorRequest::RevokeNodeCredential { node } => self
+                .handle_revoke_node_credential(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    node,
+                ),
+            AuthenticatedCoordinatorRequest::StartProcess { process, restart } => self
+                .handle_start_process(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    Some(actor.as_str().to_owned()),
+                    None,
+                    None,
+                    None,
+                    None,
+                    process,
+                    restart,
+                ),
+            request @ AuthenticatedCoordinatorRequest::ScheduleTask { .. } => {
+                self.handle_authenticated_schedule_task(&context, request)
+            }
+            request @ AuthenticatedCoordinatorRequest::LaunchTask { .. } => {
+                self.handle_authenticated_launch_task(&context, &actor, request)
+            }
+            AuthenticatedCoordinatorRequest::CancelProcess { process } => self
+                .handle_cancel_process(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    process,
+                ),
+            AuthenticatedCoordinatorRequest::AbortProcess { process } => self.handle_abort_process(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+            ),
+            AuthenticatedCoordinatorRequest::ListProcesses => self.handle_list_processes(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+            ),
+            AuthenticatedCoordinatorRequest::QuotaStatus => self.handle_quota_status(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+            ),
+            AuthenticatedCoordinatorRequest::RestartTask {
+                process,
+                task,
+                replacement_bundle,
+            } => self.handle_restart_task(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+                task,
+                replacement_bundle,
+            ),
+            request @ (AuthenticatedCoordinatorRequest::DebugAttach { .. }
+            | AuthenticatedCoordinatorRequest::SetDebugBreakpoints { .. }
+            | AuthenticatedCoordinatorRequest::InspectDebugBreakpoints { .. }
+            | AuthenticatedCoordinatorRequest::CreateDebugEpoch { .. }
+            | AuthenticatedCoordinatorRequest::ResumeDebugEpoch { .. }
+            | AuthenticatedCoordinatorRequest::InspectDebugEpoch { .. }) => self
+                .handle_authenticated_debug_request(
+                    &context.tenant,
+                    &context.project,
+                    &actor,
+                    request,
+                ),
+            AuthenticatedCoordinatorRequest::ListTaskEvents { process } => self
+                .handle_list_task_events(
+                    context.tenant.as_str().to_owned(),
+                    context.project.as_str().to_owned(),
+                    actor.as_str().to_owned(),
+                    process,
+                ),
+            AuthenticatedCoordinatorRequest::JoinTask { process, task } => self.handle_join_task(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                process,
+                task,
+            ),
+            AuthenticatedCoordinatorRequest::CreateArtifactDownloadLink {
+                artifact,
+                max_bytes,
+                ttl_seconds,
+            } => self.handle_create_artifact_download_link(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                artifact,
+                max_bytes,
+                ttl_seconds,
+            ),
+            AuthenticatedCoordinatorRequest::OpenArtifactDownloadStream {
+                artifact,
+                max_bytes,
+                token_digest,
+                chunk_bytes,
+            } => self.handle_open_artifact_download_stream(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                artifact,
+                max_bytes,
+                token_digest,
+                chunk_bytes,
+            ),
+            AuthenticatedCoordinatorRequest::RevokeArtifactDownloadLink {
+                artifact,
+                token_digest,
+            } => self.handle_revoke_artifact_download_link(
+                context.tenant.as_str().to_owned(),
+                context.project.as_str().to_owned(),
+                actor.as_str().to_owned(),
+                artifact,
+                token_digest,
+            ),
+            request @ AuthenticatedCoordinatorRequest::ExportArtifactToNode { .. } => {
+                self.handle_authenticated_artifact_export(&context, &actor, request)
+            }
+        }
+    }
+}

crates/disasmer-coordinator/src/service/signed_nodes.rs

@@ -0,0 +1,365 @@
+use disasmer_core::{NodeId, NodeSignedRequest};
+
+use crate::CoordinatorError;
+
+use super::{CoordinatorRequest, CoordinatorResponse, CoordinatorService, CoordinatorServiceError};
+
+impl CoordinatorService {
+    pub(super) fn handle_signed_node_request(
+        &mut self,
+        signed_node: String,
+        node_signature: NodeSignedRequest,
+        request: CoordinatorRequest,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        let request_kind = signed_node_request_kind(&request)?;
+        let request_node = signed_node_request_node(&request)?;
+        let request_payload = serde_json::to_value(&request).map_err(|error| {
+            CoordinatorServiceError::Protocol(format!(
+                "failed to canonicalize signed node request: {error}"
+            ))
+        })?;
+        let payload_digest = disasmer_core::signed_request_payload_digest(&request_payload);
+        let signed_node = NodeId::new(signed_node);
+        if request_node != signed_node {
+            return Err(CoordinatorError::Unauthorized(
+                "signed node request node does not match the wrapped request node".to_owned(),
+            )
+            .into());
+        }
+        self.authenticate_node_request(
+            &signed_node,
+            Some(node_signature),
+            request_kind,
+            &payload_digest,
+        )?;
+        match request {
+            CoordinatorRequest::ReportNodeCapabilities {
+                tenant,
+                project,
+                node,
+                capabilities,
+                cached_environment_digests,
+                dependency_cache_digests,
+                source_snapshots,
+                artifact_locations,
+                direct_connectivity,
+                online,
+            } => self.handle_report_node_capabilities(
+                tenant,
+                project,
+                node,
+                capabilities,
+                cached_environment_digests,
+                dependency_cache_digests,
+                source_snapshots,
+                artifact_locations,
+                direct_connectivity,
+                online,
+            ),
+            CoordinatorRequest::PollTaskAssignment {
+                tenant,
+                project,
+                node,
+            } => self.handle_poll_task_assignment(tenant, project, node),
+            CoordinatorRequest::PollArtifactTransfer {
+                tenant,
+                project,
+                node,
+            } => self.handle_poll_artifact_transfer(tenant, project, node),
+            CoordinatorRequest::UploadArtifactTransferChunk {
+                tenant,
+                project,
+                node,
+                transfer_id,
+                artifact,
+                offset,
+                content_base64,
+                chunk_digest,
+                eof,
+            } => self.handle_upload_artifact_transfer_chunk(
+                tenant,
+                project,
+                node,
+                transfer_id,
+                artifact,
+                offset,
+                content_base64,
+                chunk_digest,
+                eof,
+            ),
+            CoordinatorRequest::FailArtifactTransfer {
+                tenant,
+                project,
+                node,
+                transfer_id,
+                artifact,
+                message,
+            } => self.handle_fail_artifact_transfer(
+                tenant,
+                project,
+                node,
+                transfer_id,
+                artifact,
+                message,
+            ),
+            CoordinatorRequest::LaunchChildTask {
+                tenant,
+                project,
+                process,
+                node,
+                parent_task,
+                task_spec,
+                wait_for_node,
+                artifact_path,
+                wasm_module_base64,
+            } => self.handle_launch_child_task(
+                tenant,
+                project,
+                process,
+                node,
+                parent_task,
+                task_spec,
+                wait_for_node,
+                artifact_path,
+                wasm_module_base64,
+            ),
+            CoordinatorRequest::JoinChildTask {
+                tenant,
+                project,
+                process,
+                node,
+                parent_task,
+                task,
+            } => self.handle_join_child_task(tenant, project, process, node, parent_task, task),
+            CoordinatorRequest::CompleteSourcePreparation {
+                tenant,
+                project,
+                node,
+                provider,
+                source_snapshot,
+            } => self.handle_complete_source_preparation(
+                tenant,
+                project,
+                node,
+                provider,
+                source_snapshot,
+            ),
+            CoordinatorRequest::ReconnectNode {
+                node,
+                process,
+                epoch,
+            } => self.handle_reconnect_node(node, process, epoch),
+            CoordinatorRequest::PollTaskControl {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+            } => self.handle_poll_task_control(tenant, project, process, node, task),
+            CoordinatorRequest::PollDebugCommand {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+            } => self.handle_poll_debug_command(tenant, project, process, node, task),
+            CoordinatorRequest::ReportDebugState {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                epoch,
+                state,
+                stack_frames,
+                local_values,
+                task_args,
+                handles,
+                command_status,
+                recent_output,
+                message,
+            } => self.handle_report_debug_state(
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                epoch,
+                state,
+                stack_frames,
+                local_values,
+                task_args,
+                handles,
+                command_status,
+                recent_output,
+                message,
+            ),
+            CoordinatorRequest::ReportDebugProbeHit {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                probe_symbol,
+            } => self.handle_report_debug_probe_hit(
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                probe_symbol,
+            ),
+            CoordinatorRequest::ReportTaskLog {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                stdout_bytes,
+                stderr_bytes,
+                stdout_tail,
+                stderr_tail,
+                stdout_truncated,
+                stderr_truncated,
+                backpressured,
+            } => self.handle_report_task_log(
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                stdout_bytes,
+                stderr_bytes,
+                stdout_tail,
+                stderr_tail,
+                stdout_truncated,
+                stderr_truncated,
+                backpressured,
+            ),
+            CoordinatorRequest::ReportVfsMetadata {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                artifact_path,
+                artifact_digest,
+                artifact_size_bytes,
+                large_bytes_uploaded,
+            } => self.handle_report_vfs_metadata(
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                artifact_path,
+                artifact_digest,
+                artifact_size_bytes,
+                large_bytes_uploaded,
+            ),
+            CoordinatorRequest::TaskCompleted {
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                terminal_state,
+                status_code,
+                stdout_bytes,
+                stderr_bytes,
+                stdout_tail,
+                stderr_tail,
+                stdout_truncated,
+                stderr_truncated,
+                artifact_path,
+                artifact_digest,
+                artifact_size_bytes,
+                result,
+            } => self.handle_task_completed(
+                tenant,
+                project,
+                process,
+                node,
+                task,
+                terminal_state,
+                status_code,
+                stdout_bytes,
+                stderr_bytes,
+                stdout_tail,
+                stderr_tail,
+                stdout_truncated,
+                stderr_truncated,
+                artifact_path,
+                artifact_digest,
+                artifact_size_bytes,
+                result,
+            ),
+            _ => self.reject_unsigned_node_request(),
+        }
+    }
+
+    pub(super) fn reject_unsigned_node_request(
+        &self,
+    ) -> Result<CoordinatorResponse, CoordinatorServiceError> {
+        Err(CoordinatorError::Unauthorized(
+            "node-originated request requires signed_node envelope proof".to_owned(),
+        )
+        .into())
+    }
+}
+
+fn signed_node_request_kind(
+    request: &CoordinatorRequest,
+) -> Result<&'static str, CoordinatorServiceError> {
+    match request {
+        CoordinatorRequest::ReportNodeCapabilities { .. } => Ok("report_node_capabilities"),
+        CoordinatorRequest::PollTaskAssignment { .. } => Ok("poll_task_assignment"),
+        CoordinatorRequest::PollArtifactTransfer { .. } => Ok("poll_artifact_transfer"),
+        CoordinatorRequest::UploadArtifactTransferChunk { .. } => {
+            Ok("upload_artifact_transfer_chunk")
+        }
+        CoordinatorRequest::FailArtifactTransfer { .. } => Ok("fail_artifact_transfer"),
+        CoordinatorRequest::LaunchChildTask { .. } => Ok("launch_child_task"),
+        CoordinatorRequest::JoinChildTask { .. } => Ok("join_child_task"),
+        CoordinatorRequest::CompleteSourcePreparation { .. } => Ok("complete_source_preparation"),
+        CoordinatorRequest::ReconnectNode { .. } => Ok("reconnect_node"),
+        CoordinatorRequest::PollTaskControl { .. } => Ok("poll_task_control"),
+        CoordinatorRequest::PollDebugCommand { .. } => Ok("poll_debug_command"),
+        CoordinatorRequest::ReportDebugState { .. } => Ok("report_debug_state"),
+        CoordinatorRequest::ReportDebugProbeHit { .. } => Ok("report_debug_probe_hit"),
+        CoordinatorRequest::ReportTaskLog { .. } => Ok("report_task_log"),
+        CoordinatorRequest::ReportVfsMetadata { .. } => Ok("report_vfs_metadata"),
+        CoordinatorRequest::TaskCompleted { .. } => Ok("task_completed"),
+        _ => Err(CoordinatorError::Unauthorized(
+            "signed_node envelope only accepts node-originated coordinator requests".to_owned(),
+        )
+        .into()),
+    }
+}
+
+fn signed_node_request_node(
+    request: &CoordinatorRequest,
+) -> Result<NodeId, CoordinatorServiceError> {
+    match request {
+        CoordinatorRequest::ReportNodeCapabilities { node, .. }
+        | CoordinatorRequest::PollTaskAssignment { node, .. }
+        | CoordinatorRequest::PollArtifactTransfer { node, .. }
+        | CoordinatorRequest::UploadArtifactTransferChunk { node, .. }
+        | CoordinatorRequest::FailArtifactTransfer { node, .. }
+        | CoordinatorRequest::LaunchChildTask { node, .. }
+        | CoordinatorRequest::JoinChildTask { node, .. }
+        | CoordinatorRequest::CompleteSourcePreparation { node, .. }
+        | CoordinatorRequest::ReconnectNode { node, .. }
+        | CoordinatorRequest::PollTaskControl { node, .. }
+        | CoordinatorRequest::PollDebugCommand { node, .. }
+        | CoordinatorRequest::ReportDebugState { node, .. }
+        | CoordinatorRequest::ReportDebugProbeHit { node, .. }
+        | CoordinatorRequest::ReportTaskLog { node, .. }
+        | CoordinatorRequest::ReportVfsMetadata { node, .. }
+        | CoordinatorRequest::TaskCompleted { node, .. } => Ok(NodeId::new(node.clone())),
+        _ => Err(CoordinatorError::Unauthorized(
+            "signed_node envelope only accepts node-originated coordinator requests".to_owned(),
+        )
+        .into()),
+    }
+}

crates/disasmer-coordinator/src/service/tcp.rs

@@ -0,0 +1,200 @@
+use std::io::{BufRead, BufReader, Write};
+use std::net::{SocketAddr, TcpListener, TcpStream};
+use std::sync::{Arc, Mutex};
+
+use super::{CoordinatorRequest, CoordinatorResponse, CoordinatorService, CoordinatorServiceError};
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ClientAuthorityMode {
+    Strict,
+    LocalTrustedLoopback,
+}
+
+impl CoordinatorService {
+    pub fn serve_tcp(self, listener: TcpListener) -> Result<(), CoordinatorServiceError> {
+        if !listener.local_addr()?.ip().is_loopback() {
+            return Err(CoordinatorServiceError::Protocol(
+                "the native coordinator transport is plaintext and restricted to loopback; expose a remote coordinator only through a secure transport"
+                    .to_owned(),
+            ));
+        }
+        self.serve_tcp_with_authority(listener, ClientAuthorityMode::Strict)
+    }
+
+    pub fn serve_tcp_local_trusted(
+        self,
+        listener: TcpListener,
+    ) -> Result<(), CoordinatorServiceError> {
+        if !listener.local_addr()?.ip().is_loopback() {
+            return Err(CoordinatorServiceError::Protocol(
+                "local trusted request mode is restricted to a loopback listener".to_owned(),
+            ));
+        }
+        self.serve_tcp_with_authority(listener, ClientAuthorityMode::LocalTrustedLoopback)
+    }
+
+    fn serve_tcp_with_authority(
+        self,
+        listener: TcpListener,
+        authority_mode: ClientAuthorityMode,
+    ) -> Result<(), CoordinatorServiceError> {
+        let shared = Arc::new(Mutex::new(self));
+        for stream in listener.incoming() {
+            let stream = stream?;
+            let service = Arc::clone(&shared);
+            std::thread::spawn(move || {
+                if let Err(err) = handle_shared_stream(service, stream, authority_mode) {
+                    eprintln!("coordinator stream failed: {err}");
+                }
+            });
+        }
+        Ok(())
+    }
+
+    pub fn handle_stream(&mut self, stream: TcpStream) -> Result<(), CoordinatorServiceError> {
+        self.handle_stream_with_authority(stream, ClientAuthorityMode::Strict)
+    }
+
+    #[cfg(test)]
+    pub(super) fn handle_stream_local_trusted(
+        &mut self,
+        stream: TcpStream,
+    ) -> Result<(), CoordinatorServiceError> {
+        self.handle_stream_with_authority(stream, ClientAuthorityMode::LocalTrustedLoopback)
+    }
+
+    fn handle_stream_with_authority(
+        &mut self,
+        stream: TcpStream,
+        authority_mode: ClientAuthorityMode,
+    ) -> Result<(), CoordinatorServiceError> {
+        let mut reader = BufReader::new(stream.try_clone()?);
+        let mut writer = stream;
+        loop {
+            let mut line = String::new();
+            if reader.read_line(&mut line)? == 0 {
+                return Ok(());
+            }
+            if line.trim().is_empty() {
+                continue;
+            }
+            let response = match decode_wire_request(&line) {
+                Ok(request) => match authorize_client_request(&request, authority_mode)
+                    .and_then(|()| self.handle_request(request))
+                {
+                    Ok(response) => response,
+                    Err(err) => CoordinatorResponse::Error {
+                        message: err.to_string(),
+                    },
+                },
+                Err(err) => CoordinatorResponse::Error {
+                    message: err.to_string(),
+                },
+            };
+            serde_json::to_writer(&mut writer, &response)?;
+            writer.write_all(b"\n")?;
+            writer.flush()?;
+        }
+    }
+}
+
+fn handle_shared_stream(
+    service: Arc<Mutex<CoordinatorService>>,
+    stream: TcpStream,
+    authority_mode: ClientAuthorityMode,
+) -> Result<(), CoordinatorServiceError> {
+    let mut reader = BufReader::new(stream.try_clone()?);
+    let mut writer = stream;
+    loop {
+        let mut line = String::new();
+        if reader.read_line(&mut line)? == 0 {
+            return Ok(());
+        }
+        if line.trim().is_empty() {
+            continue;
+        }
+        let response = match decode_wire_request(&line) {
+            Ok(request) => match authorize_client_request(&request, authority_mode) {
+                Ok(()) => match service.lock() {
+                    Ok(mut service) => match service.handle_request(request) {
+                        Ok(response) => response,
+                        Err(err) => CoordinatorResponse::Error {
+                            message: err.to_string(),
+                        },
+                    },
+                    Err(_) => CoordinatorResponse::Error {
+                        message: "coordinator service lock poisoned".to_owned(),
+                    },
+                },
+                Err(err) => CoordinatorResponse::Error {
+                    message: err.to_string(),
+                },
+            },
+            Err(err) => CoordinatorResponse::Error {
+                message: err.to_string(),
+            },
+        };
+        serde_json::to_writer(&mut writer, &response)?;
+        writer.write_all(b"\n")?;
+        writer.flush()?;
+    }
+}
+
+pub fn bind_listener(addr: &str) -> Result<(TcpListener, SocketAddr), CoordinatorServiceError> {
+    let listener = TcpListener::bind(addr)?;
+    let addr = listener.local_addr()?;
+    Ok((listener, addr))
+}
+
+fn decode_wire_request(line: &str) -> Result<CoordinatorRequest, CoordinatorServiceError> {
+    serde_json::from_str::<super::CoordinatorWireRequest>(line)?
+        .into_request()
+        .map_err(CoordinatorServiceError::Protocol)
+}
+
+fn authorize_client_request(
+    request: &CoordinatorRequest,
+    authority_mode: ClientAuthorityMode,
+) -> Result<(), CoordinatorServiceError> {
+    if authority_mode == ClientAuthorityMode::LocalTrustedLoopback {
+        return Ok(());
+    }
+    match request {
+        CoordinatorRequest::Ping
+        | CoordinatorRequest::Authenticated { .. }
+        | CoordinatorRequest::ExchangeNodeEnrollmentGrant { .. }
+        | CoordinatorRequest::SignedNode { .. }
+        | CoordinatorRequest::NodeHeartbeat {
+            node_signature: Some(_),
+            ..
+        }
+        | CoordinatorRequest::StartProcess {
+            actor_agent: Some(_),
+            agent_signature: Some(_),
+            ..
+        }
+        | CoordinatorRequest::LaunchTask {
+            actor_agent: Some(_),
+            agent_signature: Some(_),
+            ..
+        }
+        | CoordinatorRequest::AdminStatus { .. }
+        | CoordinatorRequest::SuspendTenant { .. } => Ok(()),
+        _ => Err(CoordinatorServiceError::Protocol(
+            "strict Core Client authority requires an authenticated CLI session, signed Agent, signed Node, enrollment grant exchange, or admin credential; request-body identity fields are not authority"
+                .to_owned(),
+        )),
+    }
+}
+
+#[cfg(test)]
+mod transport_boundary_tests {
+    use super::*;
+
+    #[test]
+    fn native_plaintext_service_refuses_non_loopback_listener() {
+        let (listener, _) = bind_listener("0.0.0.0:0").unwrap();
+        let error = CoordinatorService::new(1).serve_tcp(listener).unwrap_err();
+        assert!(error.to_string().contains("restricted to loopback"));
+    }
+}

crates/disasmer-coordinator/src/service/wire_protocol.rs

@@ -0,0 +1,59 @@
+use disasmer_core::{COORDINATOR_PROTOCOL_VERSION, COORDINATOR_WIRE_REQUEST_TYPE};
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+
+use super::CoordinatorRequest;
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum CoordinatorWireRequest {
+    Envelope(CoordinatorRequestEnvelope),
+}
+
+impl CoordinatorWireRequest {
+    pub fn into_request(self) -> Result<CoordinatorRequest, String> {
+        match self {
+            Self::Envelope(envelope) => envelope.into_request(),
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
+pub struct CoordinatorRequestEnvelope {
+    #[serde(rename = "type")]
+    pub envelope_type: String,
+    pub protocol_version: u64,
+    pub request_id: String,
+    pub operation: String,
+    #[serde(default)]
+    pub authentication: Option<Value>,
+    pub payload: CoordinatorRequest,
+}
+
+impl CoordinatorRequestEnvelope {
+    pub fn into_request(self) -> Result<CoordinatorRequest, String> {
+        if self.envelope_type != COORDINATOR_WIRE_REQUEST_TYPE {
+            return Err(format!(
+                "unsupported coordinator wire request type {}; expected {}",
+                self.envelope_type, COORDINATOR_WIRE_REQUEST_TYPE
+            ));
+        }
+        if self.protocol_version != COORDINATOR_PROTOCOL_VERSION {
+            return Err(format!(
+                "unsupported coordinator protocol version {}; expected {}",
+                self.protocol_version, COORDINATOR_PROTOCOL_VERSION
+            ));
+        }
+        if self.request_id.trim().is_empty() {
+            return Err("coordinator wire request_id must be non-empty".to_owned());
+        }
+        let payload_operation = self.payload.operation()?;
+        if self.operation != payload_operation {
+            return Err(format!(
+                "coordinator wire operation {} does not match payload operation {}",
+                self.operation, payload_operation
+            ));
+        }
+        Ok(self.payload)
+    }
+}

crates/disasmer-coordinator/src/sessions.rs

@@ -0,0 +1,167 @@
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use disasmer_core::{Actor, AuthContext, CredentialKind, Digest, ProjectId, TenantId, UserId};
+
+use crate::{CliSessionRecord, Coordinator, CoordinatorError, CredentialRecord};
+
+impl Coordinator {
+    pub fn issue_cli_session(
+        &mut self,
+        tenant: TenantId,
+        project: ProjectId,
+        user: UserId,
+        session_secret: &str,
+        expires_at_epoch_seconds: Option<u64>,
+    ) -> CliSessionRecord {
+        self.upsert_tenant(tenant.clone());
+        self.upsert_user(
+            tenant.clone(),
+            user.clone(),
+            CredentialKind::CliDeviceSession,
+        );
+        let session_digest = Digest::sha256(session_secret);
+        let record = CliSessionRecord {
+            session_digest: session_digest.clone(),
+            tenant: tenant.clone(),
+            project: project.clone(),
+            user: user.clone(),
+            credential_kind: CredentialKind::CliDeviceSession,
+            expires_at_epoch_seconds,
+            revoked: false,
+        };
+        self.durable
+            .cli_sessions
+            .insert(session_digest.clone(), record.clone());
+        let credential_subject = format!("cli-session:{}", session_digest.as_str());
+        self.durable.credentials.insert(
+            credential_subject.clone(),
+            CredentialRecord {
+                subject: credential_subject,
+                tenant,
+                project: Some(project),
+                kind: CredentialKind::CliDeviceSession,
+                public_key_fingerprint: None,
+            },
+        );
+        record
+    }
+
+    pub fn authenticate_cli_session(
+        &self,
+        session_secret: &str,
+    ) -> Result<AuthContext, CoordinatorError> {
+        self.authenticate_cli_session_at(session_secret, unix_timestamp_seconds())
+    }
+
+    pub fn authenticate_cli_session_at(
+        &self,
+        session_secret: &str,
+        now_epoch_seconds: u64,
+    ) -> Result<AuthContext, CoordinatorError> {
+        if session_secret.trim().is_empty() {
+            return Err(CoordinatorError::Unauthorized(
+                "CLI session credential is missing".to_owned(),
+            ));
+        }
+        let session_digest = Digest::sha256(session_secret);
+        let record = self
+            .durable
+            .cli_sessions
+            .get(&session_digest)
+            .ok_or_else(|| {
+                CoordinatorError::Unauthorized(
+                    "CLI session credential is not recognized".to_owned(),
+                )
+            })?;
+        if record.revoked {
+            return Err(CoordinatorError::Unauthorized(
+                "CLI session credential has been revoked".to_owned(),
+            ));
+        }
+        if record
+            .expires_at_epoch_seconds
+            .is_some_and(|expires_at| expires_at <= now_epoch_seconds)
+        {
+            return Err(CoordinatorError::Unauthorized(
+                "CLI session credential has expired; run disasmer login --browser again".to_owned(),
+            ));
+        }
+        if record.credential_kind != CredentialKind::CliDeviceSession {
+            return Err(CoordinatorError::Unauthorized(
+                "credential is not a CLI session".to_owned(),
+            ));
+        }
+        self.ensure_tenant_active(&record.tenant)?;
+        Ok(AuthContext {
+            tenant: record.tenant.clone(),
+            project: record.project.clone(),
+            actor: Actor::User(record.user.clone()),
+        })
+    }
+
+    pub fn revoke_cli_session(
+        &mut self,
+        session_secret: &str,
+    ) -> Result<CliSessionRecord, CoordinatorError> {
+        self.authenticate_cli_session(session_secret)?;
+        let session_digest = Digest::sha256(session_secret);
+        let record = self
+            .durable
+            .cli_sessions
+            .get_mut(&session_digest)
+            .expect("authenticated session must exist");
+        record.revoked = true;
+        Ok(record.clone())
+    }
+}
+
+fn unix_timestamp_seconds() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|duration| duration.as_secs())
+        .unwrap_or(0)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::collections::BTreeSet;
+
+    use crate::InMemoryDurableStore;
+
+    use super::*;
+
+    #[test]
+    fn repeated_logins_for_one_user_create_distinct_session_credentials() {
+        let store = InMemoryDurableStore::default();
+        let mut coordinator = Coordinator::boot(&store, 1);
+        let tenant = TenantId::from("tenant");
+        let user = UserId::from("user");
+
+        coordinator.issue_cli_session(
+            tenant.clone(),
+            ProjectId::from("project-one"),
+            user.clone(),
+            "session-one",
+            None,
+        );
+        coordinator.issue_cli_session(
+            tenant,
+            ProjectId::from("project-two"),
+            user,
+            "session-two",
+            None,
+        );
+
+        let subjects = coordinator
+            .durable
+            .credentials
+            .values()
+            .map(|credential| credential.subject.clone())
+            .collect::<BTreeSet<_>>();
+        assert_eq!(coordinator.durable.cli_sessions.len(), 2);
+        assert_eq!(subjects.len(), 2);
+        assert!(subjects
+            .iter()
+            .all(|subject| subject.starts_with("cli-session:sha256:")));
+    }
+}