初始化提交

crates/openfang-kernel/Cargo.toml

@@ -0,0 +1,42 @@
+[package]
+name = "openfang-kernel"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+description = "Core kernel for the OpenFang Agent OS"
+
+[dependencies]
+openfang-types = { path = "../openfang-types" }
+openfang-memory = { path = "../openfang-memory" }
+openfang-runtime = { path = "../openfang-runtime" }
+openfang-skills = { path = "../openfang-skills" }
+openfang-hands = { path = "../openfang-hands" }
+openfang-extensions = { path = "../openfang-extensions" }
+openfang-wire = { path = "../openfang-wire" }
+openfang-channels = { path = "../openfang-channels" }
+tokio = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+toml = { workspace = true }
+dashmap = { workspace = true }
+crossbeam = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+chrono = { workspace = true }
+uuid = { workspace = true }
+thiserror = { workspace = true }
+async-trait = { workspace = true }
+dirs = { workspace = true }
+futures = { workspace = true }
+subtle = { workspace = true }
+rand = { workspace = true }
+hex = { workspace = true }
+reqwest = { workspace = true }
+cron = "0.15"
+
+[target.'cfg(unix)'.dependencies]
+libc = "0.2"
+
+[dev-dependencies]
+tokio-test = { workspace = true }
+tempfile = { workspace = true }

crates/openfang-kernel/src/approval.rs

@@ -0,0 +1,405 @@
+//! Execution approval manager — gates dangerous operations behind human approval.
+
+use chrono::Utc;
+use dashmap::DashMap;
+use openfang_types::approval::{
+    ApprovalDecision, ApprovalPolicy, ApprovalRequest, ApprovalResponse, RiskLevel,
+};
+use tracing::{debug, info, warn};
+use uuid::Uuid;
+
+/// Max pending requests per agent.
+const MAX_PENDING_PER_AGENT: usize = 5;
+
+/// Manages approval requests with oneshot channels for blocking resolution.
+pub struct ApprovalManager {
+    pending: DashMap<Uuid, PendingRequest>,
+    policy: std::sync::RwLock<ApprovalPolicy>,
+}
+
+struct PendingRequest {
+    request: ApprovalRequest,
+    sender: tokio::sync::oneshot::Sender<ApprovalDecision>,
+}
+
+impl ApprovalManager {
+    pub fn new(policy: ApprovalPolicy) -> Self {
+        Self {
+            pending: DashMap::new(),
+            policy: std::sync::RwLock::new(policy),
+        }
+    }
+
+    /// Check if a tool requires approval based on current policy.
+    pub fn requires_approval(&self, tool_name: &str) -> bool {
+        let policy = self.policy.read().unwrap_or_else(|e| e.into_inner());
+        policy.require_approval.iter().any(|t| t == tool_name)
+    }
+
+    /// Submit an approval request. Returns a future that resolves when approved/denied/timed out.
+    pub async fn request_approval(&self, req: ApprovalRequest) -> ApprovalDecision {
+        // Check per-agent pending limit
+        let agent_pending = self
+            .pending
+            .iter()
+            .filter(|r| r.value().request.agent_id == req.agent_id)
+            .count();
+        if agent_pending >= MAX_PENDING_PER_AGENT {
+            warn!(agent_id = %req.agent_id, "Approval request rejected: too many pending");
+            return ApprovalDecision::Denied;
+        }
+
+        let timeout = std::time::Duration::from_secs(req.timeout_secs);
+        let id = req.id;
+
+        let (tx, rx) = tokio::sync::oneshot::channel();
+        self.pending.insert(
+            id,
+            PendingRequest {
+                request: req,
+                sender: tx,
+            },
+        );
+
+        info!(request_id = %id, "Approval request submitted, waiting for resolution");
+
+        match tokio::time::timeout(timeout, rx).await {
+            Ok(Ok(decision)) => {
+                debug!(request_id = %id, ?decision, "Approval resolved");
+                decision
+            }
+            _ => {
+                self.pending.remove(&id);
+                warn!(request_id = %id, "Approval request timed out");
+                ApprovalDecision::TimedOut
+            }
+        }
+    }
+
+    /// Resolve a pending request (called by API/UI).
+    pub fn resolve(
+        &self,
+        request_id: Uuid,
+        decision: ApprovalDecision,
+        decided_by: Option<String>,
+    ) -> Result<ApprovalResponse, String> {
+        match self.pending.remove(&request_id) {
+            Some((_, pending)) => {
+                let response = ApprovalResponse {
+                    request_id,
+                    decision,
+                    decided_at: Utc::now(),
+                    decided_by,
+                };
+                // Send decision to waiting agent (ignore error if receiver dropped)
+                let _ = pending.sender.send(decision);
+                info!(request_id = %request_id, ?decision, "Approval request resolved");
+                Ok(response)
+            }
+            None => Err(format!("No pending approval request with id {request_id}")),
+        }
+    }
+
+    /// List all pending requests (for API/dashboard display).
+    pub fn list_pending(&self) -> Vec<ApprovalRequest> {
+        self.pending
+            .iter()
+            .map(|r| r.value().request.clone())
+            .collect()
+    }
+
+    /// Number of pending requests.
+    pub fn pending_count(&self) -> usize {
+        self.pending.len()
+    }
+
+    /// Update the approval policy (for hot-reload).
+    pub fn update_policy(&self, policy: ApprovalPolicy) {
+        *self.policy.write().unwrap_or_else(|e| e.into_inner()) = policy;
+    }
+
+    /// Get a copy of the current policy.
+    pub fn policy(&self) -> ApprovalPolicy {
+        self.policy
+            .read()
+            .unwrap_or_else(|e| e.into_inner())
+            .clone()
+    }
+
+    /// Classify the risk level of a tool invocation.
+    pub fn classify_risk(tool_name: &str) -> RiskLevel {
+        match tool_name {
+            "shell_exec" => RiskLevel::Critical,
+            "file_write" | "file_delete" => RiskLevel::High,
+            "web_fetch" | "browser_navigate" => RiskLevel::Medium,
+            _ => RiskLevel::Low,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openfang_types::approval::ApprovalPolicy;
+    use std::sync::Arc;
+
+    fn default_manager() -> ApprovalManager {
+        ApprovalManager::new(ApprovalPolicy::default())
+    }
+
+    fn make_request(agent_id: &str, tool_name: &str, timeout_secs: u64) -> ApprovalRequest {
+        ApprovalRequest {
+            id: Uuid::new_v4(),
+            agent_id: agent_id.to_string(),
+            tool_name: tool_name.to_string(),
+            description: "test operation".to_string(),
+            action_summary: "test action".to_string(),
+            risk_level: RiskLevel::High,
+            requested_at: Utc::now(),
+            timeout_secs,
+        }
+    }
+
+    // -----------------------------------------------------------------------
+    // requires_approval
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_requires_approval_default() {
+        let mgr = default_manager();
+        assert!(mgr.requires_approval("shell_exec"));
+        assert!(!mgr.requires_approval("file_read"));
+    }
+
+    #[test]
+    fn test_requires_approval_custom_policy() {
+        let policy = ApprovalPolicy {
+            require_approval: vec!["file_write".to_string(), "file_delete".to_string()],
+            timeout_secs: 30,
+            auto_approve_autonomous: false,
+            auto_approve: false,
+        };
+        let mgr = ApprovalManager::new(policy);
+        assert!(mgr.requires_approval("file_write"));
+        assert!(mgr.requires_approval("file_delete"));
+        assert!(!mgr.requires_approval("shell_exec"));
+        assert!(!mgr.requires_approval("file_read"));
+    }
+
+    // -----------------------------------------------------------------------
+    // classify_risk
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_classify_risk() {
+        assert_eq!(
+            ApprovalManager::classify_risk("shell_exec"),
+            RiskLevel::Critical
+        );
+        assert_eq!(
+            ApprovalManager::classify_risk("file_write"),
+            RiskLevel::High
+        );
+        assert_eq!(
+            ApprovalManager::classify_risk("file_delete"),
+            RiskLevel::High
+        );
+        assert_eq!(
+            ApprovalManager::classify_risk("web_fetch"),
+            RiskLevel::Medium
+        );
+        assert_eq!(
+            ApprovalManager::classify_risk("browser_navigate"),
+            RiskLevel::Medium
+        );
+        assert_eq!(ApprovalManager::classify_risk("file_read"), RiskLevel::Low);
+        assert_eq!(
+            ApprovalManager::classify_risk("unknown_tool"),
+            RiskLevel::Low
+        );
+    }
+
+    // -----------------------------------------------------------------------
+    // resolve nonexistent
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_resolve_nonexistent() {
+        let mgr = default_manager();
+        let result = mgr.resolve(Uuid::new_v4(), ApprovalDecision::Approved, None);
+        assert!(result.is_err());
+        assert!(result.unwrap_err().contains("No pending approval request"));
+    }
+
+    // -----------------------------------------------------------------------
+    // list_pending empty
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_list_pending_empty() {
+        let mgr = default_manager();
+        assert!(mgr.list_pending().is_empty());
+    }
+
+    // -----------------------------------------------------------------------
+    // update_policy
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_update_policy() {
+        let mgr = default_manager();
+        assert!(mgr.requires_approval("shell_exec"));
+        assert!(!mgr.requires_approval("file_write"));
+
+        let new_policy = ApprovalPolicy {
+            require_approval: vec!["file_write".to_string()],
+            timeout_secs: 120,
+            auto_approve_autonomous: true,
+            auto_approve: false,
+        };
+        mgr.update_policy(new_policy);
+
+        assert!(!mgr.requires_approval("shell_exec"));
+        assert!(mgr.requires_approval("file_write"));
+
+        let policy = mgr.policy();
+        assert_eq!(policy.timeout_secs, 120);
+        assert!(policy.auto_approve_autonomous);
+    }
+
+    // -----------------------------------------------------------------------
+    // pending_count
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_pending_count() {
+        let mgr = default_manager();
+        assert_eq!(mgr.pending_count(), 0);
+    }
+
+    // -----------------------------------------------------------------------
+    // request_approval — timeout
+    // -----------------------------------------------------------------------
+
+    #[tokio::test]
+    async fn test_request_approval_timeout() {
+        let mgr = Arc::new(default_manager());
+        let req = make_request("agent-1", "shell_exec", 10);
+        let decision = mgr.request_approval(req).await;
+        assert_eq!(decision, ApprovalDecision::TimedOut);
+        // After timeout, pending map should be cleaned up
+        assert_eq!(mgr.pending_count(), 0);
+    }
+
+    // -----------------------------------------------------------------------
+    // request_approval — approve
+    // -----------------------------------------------------------------------
+
+    #[tokio::test]
+    async fn test_request_approval_approve() {
+        let mgr = Arc::new(default_manager());
+        let req = make_request("agent-1", "shell_exec", 60);
+        let request_id = req.id;
+
+        let mgr2 = Arc::clone(&mgr);
+        tokio::spawn(async move {
+            // Small delay to let the request register
+            tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+            let result = mgr2.resolve(
+                request_id,
+                ApprovalDecision::Approved,
+                Some("admin".to_string()),
+            );
+            assert!(result.is_ok());
+            let resp = result.unwrap();
+            assert_eq!(resp.decision, ApprovalDecision::Approved);
+            assert_eq!(resp.decided_by, Some("admin".to_string()));
+        });
+
+        let decision = mgr.request_approval(req).await;
+        assert_eq!(decision, ApprovalDecision::Approved);
+    }
+
+    // -----------------------------------------------------------------------
+    // request_approval — deny
+    // -----------------------------------------------------------------------
+
+    #[tokio::test]
+    async fn test_request_approval_deny() {
+        let mgr = Arc::new(default_manager());
+        let req = make_request("agent-1", "shell_exec", 60);
+        let request_id = req.id;
+
+        let mgr2 = Arc::clone(&mgr);
+        tokio::spawn(async move {
+            tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+            let result = mgr2.resolve(request_id, ApprovalDecision::Denied, None);
+            assert!(result.is_ok());
+        });
+
+        let decision = mgr.request_approval(req).await;
+        assert_eq!(decision, ApprovalDecision::Denied);
+    }
+
+    // -----------------------------------------------------------------------
+    // max pending per agent
+    // -----------------------------------------------------------------------
+
+    #[tokio::test]
+    async fn test_max_pending_per_agent() {
+        let mgr = Arc::new(default_manager());
+
+        // Fill up 5 pending requests for agent-1 (they will all be waiting)
+        let mut ids = Vec::new();
+        for _ in 0..MAX_PENDING_PER_AGENT {
+            let req = make_request("agent-1", "shell_exec", 300);
+            ids.push(req.id);
+            let mgr_clone = Arc::clone(&mgr);
+            tokio::spawn(async move {
+                mgr_clone.request_approval(req).await;
+            });
+        }
+
+        // Give spawned tasks time to register
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        assert_eq!(mgr.pending_count(), MAX_PENDING_PER_AGENT);
+
+        // 6th request for the same agent should be immediately denied
+        let req6 = make_request("agent-1", "shell_exec", 300);
+        let decision = mgr.request_approval(req6).await;
+        assert_eq!(decision, ApprovalDecision::Denied);
+
+        // A different agent should still be able to submit
+        let req_other = make_request("agent-2", "shell_exec", 300);
+        let other_id = req_other.id;
+        let mgr2 = Arc::clone(&mgr);
+        tokio::spawn(async move {
+            mgr2.request_approval(req_other).await;
+        });
+        tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+        assert_eq!(mgr.pending_count(), MAX_PENDING_PER_AGENT + 1);
+
+        // Cleanup: resolve all pending to avoid hanging tasks
+        for id in &ids {
+            let _ = mgr.resolve(*id, ApprovalDecision::Denied, None);
+        }
+        let _ = mgr.resolve(other_id, ApprovalDecision::Denied, None);
+    }
+
+    // -----------------------------------------------------------------------
+    // policy defaults
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_policy_defaults() {
+        let mgr = default_manager();
+        let policy = mgr.policy();
+        assert_eq!(policy.require_approval, vec!["shell_exec".to_string()]);
+        assert_eq!(policy.timeout_secs, 60);
+        assert!(!policy.auto_approve_autonomous);
+    }
+}

crates/openfang-kernel/src/auth.rs

@@ -0,0 +1,316 @@
+//! RBAC authentication and authorization for multi-user access control.
+//!
+//! The AuthManager maps platform user identities (Telegram ID, Discord ID, etc.)
+//! to OpenFang users with roles, then enforces permission checks on actions.
+
+use dashmap::DashMap;
+use openfang_types::agent::UserId;
+use openfang_types::config::UserConfig;
+use openfang_types::error::{OpenFangError, OpenFangResult};
+use std::fmt;
+use tracing::info;
+
+/// User roles with hierarchical permissions.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+pub enum UserRole {
+    /// Read-only access — can view agent output but cannot interact.
+    Viewer = 0,
+    /// Standard user — can chat with agents.
+    User = 1,
+    /// Admin — can spawn/kill agents, install skills, view usage.
+    Admin = 2,
+    /// Owner — full access including user management and config changes.
+    Owner = 3,
+}
+
+impl fmt::Display for UserRole {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self {
+            UserRole::Viewer => write!(f, "viewer"),
+            UserRole::User => write!(f, "user"),
+            UserRole::Admin => write!(f, "admin"),
+            UserRole::Owner => write!(f, "owner"),
+        }
+    }
+}
+
+impl UserRole {
+    /// Parse a role from a string.
+    pub fn from_str_role(s: &str) -> Self {
+        match s.to_lowercase().as_str() {
+            "owner" => UserRole::Owner,
+            "admin" => UserRole::Admin,
+            "viewer" => UserRole::Viewer,
+            _ => UserRole::User,
+        }
+    }
+}
+
+/// Actions that can be authorized.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum Action {
+    /// Chat with an agent.
+    ChatWithAgent,
+    /// Spawn a new agent.
+    SpawnAgent,
+    /// Kill a running agent.
+    KillAgent,
+    /// Install a skill.
+    InstallSkill,
+    /// View kernel configuration.
+    ViewConfig,
+    /// Modify kernel configuration.
+    ModifyConfig,
+    /// View usage/billing data.
+    ViewUsage,
+    /// Manage users (create, delete, change roles).
+    ManageUsers,
+}
+
+impl Action {
+    /// Minimum role required for this action.
+    fn required_role(&self) -> UserRole {
+        match self {
+            Action::ChatWithAgent => UserRole::User,
+            Action::ViewConfig => UserRole::User,
+            Action::ViewUsage => UserRole::Admin,
+            Action::SpawnAgent => UserRole::Admin,
+            Action::KillAgent => UserRole::Admin,
+            Action::InstallSkill => UserRole::Admin,
+            Action::ModifyConfig => UserRole::Owner,
+            Action::ManageUsers => UserRole::Owner,
+        }
+    }
+}
+
+/// A resolved user identity.
+#[derive(Debug, Clone)]
+pub struct UserIdentity {
+    /// OpenFang user ID.
+    pub id: UserId,
+    /// Display name.
+    pub name: String,
+    /// Role.
+    pub role: UserRole,
+}
+
+/// RBAC authentication and authorization manager.
+pub struct AuthManager {
+    /// Known users by their OpenFang user ID.
+    users: DashMap<UserId, UserIdentity>,
+    /// Channel binding index: "channel_type:platform_id" → UserId.
+    channel_index: DashMap<String, UserId>,
+}
+
+impl AuthManager {
+    /// Create a new AuthManager from kernel user configuration.
+    pub fn new(user_configs: &[UserConfig]) -> Self {
+        let manager = Self {
+            users: DashMap::new(),
+            channel_index: DashMap::new(),
+        };
+
+        for config in user_configs {
+            let user_id = UserId::new();
+            let role = UserRole::from_str_role(&config.role);
+            let identity = UserIdentity {
+                id: user_id,
+                name: config.name.clone(),
+                role,
+            };
+
+            manager.users.insert(user_id, identity);
+
+            // Index channel bindings
+            for (channel_type, platform_id) in &config.channel_bindings {
+                let key = format!("{channel_type}:{platform_id}");
+                manager.channel_index.insert(key, user_id);
+            }
+
+            info!(
+                user = %config.name,
+                role = %role,
+                bindings = config.channel_bindings.len(),
+                "Registered user"
+            );
+        }
+
+        manager
+    }
+
+    /// Identify a user from a channel identity.
+    ///
+    /// Returns the OpenFang UserId if a matching channel binding exists,
+    /// or None for unrecognized users.
+    pub fn identify(&self, channel_type: &str, platform_id: &str) -> Option<UserId> {
+        let key = format!("{channel_type}:{platform_id}");
+        self.channel_index.get(&key).map(|r| *r.value())
+    }
+
+    /// Get a user's identity by their UserId.
+    pub fn get_user(&self, user_id: UserId) -> Option<UserIdentity> {
+        self.users.get(&user_id).map(|r| r.value().clone())
+    }
+
+    /// Authorize a user for an action.
+    ///
+    /// Returns Ok(()) if the user has sufficient permissions, or AuthDenied error.
+    pub fn authorize(&self, user_id: UserId, action: &Action) -> OpenFangResult<()> {
+        let identity = self
+            .users
+            .get(&user_id)
+            .ok_or_else(|| OpenFangError::AuthDenied("Unknown user".to_string()))?;
+
+        let required = action.required_role();
+        if identity.role >= required {
+            Ok(())
+        } else {
+            Err(OpenFangError::AuthDenied(format!(
+                "User '{}' (role: {}) lacks permission for {:?} (requires: {})",
+                identity.name, identity.role, action, required
+            )))
+        }
+    }
+
+    /// Check if RBAC is configured (any users registered).
+    pub fn is_enabled(&self) -> bool {
+        !self.users.is_empty()
+    }
+
+    /// Get the count of registered users.
+    pub fn user_count(&self) -> usize {
+        self.users.len()
+    }
+
+    /// List all registered users.
+    pub fn list_users(&self) -> Vec<UserIdentity> {
+        self.users.iter().map(|r| r.value().clone()).collect()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::collections::HashMap;
+
+    fn test_configs() -> Vec<UserConfig> {
+        vec![
+            UserConfig {
+                name: "Alice".to_string(),
+                role: "owner".to_string(),
+                channel_bindings: {
+                    let mut m = HashMap::new();
+                    m.insert("telegram".to_string(), "123456".to_string());
+                    m.insert("discord".to_string(), "987654".to_string());
+                    m
+                },
+                api_key_hash: None,
+            },
+            UserConfig {
+                name: "Guest".to_string(),
+                role: "user".to_string(),
+                channel_bindings: {
+                    let mut m = HashMap::new();
+                    m.insert("telegram".to_string(), "999999".to_string());
+                    m
+                },
+                api_key_hash: None,
+            },
+            UserConfig {
+                name: "ReadOnly".to_string(),
+                role: "viewer".to_string(),
+                channel_bindings: HashMap::new(),
+                api_key_hash: None,
+            },
+        ]
+    }
+
+    #[test]
+    fn test_user_registration() {
+        let manager = AuthManager::new(&test_configs());
+        assert!(manager.is_enabled());
+        assert_eq!(manager.user_count(), 3);
+    }
+
+    #[test]
+    fn test_identify_from_channel() {
+        let manager = AuthManager::new(&test_configs());
+
+        // Alice on Telegram
+        let owner_tg = manager.identify("telegram", "123456");
+        assert!(owner_tg.is_some());
+
+        // Alice on Discord
+        let owner_dc = manager.identify("discord", "987654");
+        assert!(owner_dc.is_some());
+
+        // Same user across channels
+        assert_eq!(owner_tg.unwrap(), owner_dc.unwrap());
+
+        // Unknown user
+        assert!(manager.identify("telegram", "unknown").is_none());
+    }
+
+    #[test]
+    fn test_owner_can_do_everything() {
+        let manager = AuthManager::new(&test_configs());
+        let owner_id = manager.identify("telegram", "123456").unwrap();
+
+        assert!(manager.authorize(owner_id, &Action::ChatWithAgent).is_ok());
+        assert!(manager.authorize(owner_id, &Action::SpawnAgent).is_ok());
+        assert!(manager.authorize(owner_id, &Action::KillAgent).is_ok());
+        assert!(manager.authorize(owner_id, &Action::ManageUsers).is_ok());
+        assert!(manager.authorize(owner_id, &Action::ModifyConfig).is_ok());
+    }
+
+    #[test]
+    fn test_user_limited_access() {
+        let manager = AuthManager::new(&test_configs());
+        let guest_id = manager.identify("telegram", "999999").unwrap();
+
+        // User can chat and view config
+        assert!(manager.authorize(guest_id, &Action::ChatWithAgent).is_ok());
+        assert!(manager.authorize(guest_id, &Action::ViewConfig).is_ok());
+
+        // User cannot spawn/kill/manage
+        assert!(manager.authorize(guest_id, &Action::SpawnAgent).is_err());
+        assert!(manager.authorize(guest_id, &Action::KillAgent).is_err());
+        assert!(manager.authorize(guest_id, &Action::ManageUsers).is_err());
+    }
+
+    #[test]
+    fn test_viewer_read_only() {
+        let manager = AuthManager::new(&test_configs());
+        let users = manager.list_users();
+        let viewer = users.iter().find(|u| u.name == "ReadOnly").unwrap();
+
+        // Viewer cannot even chat
+        assert!(manager
+            .authorize(viewer.id, &Action::ChatWithAgent)
+            .is_err());
+    }
+
+    #[test]
+    fn test_unknown_user_denied() {
+        let manager = AuthManager::new(&test_configs());
+        let fake_id = UserId::new();
+        assert!(manager.authorize(fake_id, &Action::ChatWithAgent).is_err());
+    }
+
+    #[test]
+    fn test_no_users_means_disabled() {
+        let manager = AuthManager::new(&[]);
+        assert!(!manager.is_enabled());
+        assert_eq!(manager.user_count(), 0);
+    }
+
+    #[test]
+    fn test_role_parsing() {
+        assert_eq!(UserRole::from_str_role("owner"), UserRole::Owner);
+        assert_eq!(UserRole::from_str_role("admin"), UserRole::Admin);
+        assert_eq!(UserRole::from_str_role("viewer"), UserRole::Viewer);
+        assert_eq!(UserRole::from_str_role("user"), UserRole::User);
+        assert_eq!(UserRole::from_str_role("OWNER"), UserRole::Owner);
+        assert_eq!(UserRole::from_str_role("unknown"), UserRole::User);
+    }
+}

crates/openfang-kernel/src/auto_reply.rs

@@ -0,0 +1,211 @@
+//! Auto-reply background engine — trigger-driven background replies with concurrency control.
+
+use openfang_types::agent::AgentId;
+use openfang_types::config::AutoReplyConfig;
+use std::sync::Arc;
+use tokio::sync::Semaphore;
+use tracing::{debug, info, warn};
+
+/// Where to deliver the auto-reply result.
+#[derive(Debug, Clone)]
+pub struct AutoReplyChannel {
+    /// Channel type string (e.g., "telegram", "discord").
+    pub channel_type: String,
+    /// Peer/user ID to send the reply to.
+    pub peer_id: String,
+    /// Optional thread ID for threaded replies.
+    pub thread_id: Option<String>,
+}
+
+/// Auto-reply engine with concurrency limits and suppression patterns.
+pub struct AutoReplyEngine {
+    config: AutoReplyConfig,
+    semaphore: Arc<Semaphore>,
+}
+
+impl AutoReplyEngine {
+    /// Create a new auto-reply engine from configuration.
+    pub fn new(config: AutoReplyConfig) -> Self {
+        let permits = config.max_concurrent.max(1);
+        Self {
+            semaphore: Arc::new(Semaphore::new(permits)),
+            config,
+        }
+    }
+
+    /// Check if a message should trigger auto-reply.
+    /// Returns `None` if suppressed or disabled, `Some(agent_id)` if should auto-reply.
+    pub fn should_reply(
+        &self,
+        message: &str,
+        _channel_type: &str,
+        agent_id: AgentId,
+    ) -> Option<AgentId> {
+        if !self.config.enabled {
+            return None;
+        }
+
+        // Check suppression patterns
+        let lower = message.to_lowercase();
+        for pattern in &self.config.suppress_patterns {
+            if lower.contains(&pattern.to_lowercase()) {
+                debug!(pattern = %pattern, "Auto-reply suppressed by pattern");
+                return None;
+            }
+        }
+
+        Some(agent_id)
+    }
+
+    /// Execute an auto-reply in the background.
+    /// Returns a JoinHandle for the spawned task.
+    ///
+    /// The `send_fn` is called with the agent response to deliver it back to the channel.
+    pub async fn execute_reply<F>(
+        &self,
+        kernel_handle: Arc<dyn openfang_runtime::kernel_handle::KernelHandle>,
+        agent_id: AgentId,
+        message: String,
+        reply_channel: AutoReplyChannel,
+        send_fn: F,
+    ) -> Result<tokio::task::JoinHandle<()>, String>
+    where
+        F: Fn(String, AutoReplyChannel) -> futures::future::BoxFuture<'static, ()>
+            + Send
+            + Sync
+            + 'static,
+    {
+        // Try to acquire a semaphore permit
+        let permit = match self.semaphore.clone().try_acquire_owned() {
+            Ok(p) => p,
+            Err(_) => {
+                return Err(format!(
+                    "Auto-reply concurrency limit reached ({} max)",
+                    self.config.max_concurrent
+                ));
+            }
+        };
+
+        let timeout_secs = self.config.timeout_secs;
+
+        let handle = tokio::spawn(async move {
+            let _permit = permit; // Hold permit until task completes
+
+            info!(
+                agent = %agent_id,
+                channel = %reply_channel.channel_type,
+                peer = %reply_channel.peer_id,
+                "Starting auto-reply"
+            );
+
+            let result = tokio::time::timeout(
+                std::time::Duration::from_secs(timeout_secs),
+                kernel_handle.send_to_agent(&agent_id.to_string(), &message),
+            )
+            .await;
+
+            match result {
+                Ok(Ok(response)) => {
+                    send_fn(response, reply_channel).await;
+                }
+                Ok(Err(e)) => {
+                    warn!(agent = %agent_id, error = %e, "Auto-reply agent error");
+                }
+                Err(_) => {
+                    warn!(agent = %agent_id, timeout = timeout_secs, "Auto-reply timed out");
+                }
+            }
+        });
+
+        Ok(handle)
+    }
+
+    /// Check if auto-reply is enabled.
+    pub fn is_enabled(&self) -> bool {
+        self.config.enabled
+    }
+
+    /// Get the current configuration (read-only).
+    pub fn config(&self) -> &AutoReplyConfig {
+        &self.config
+    }
+
+    /// Get available permits (for monitoring).
+    pub fn available_permits(&self) -> usize {
+        self.semaphore.available_permits()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn test_config(enabled: bool) -> AutoReplyConfig {
+        AutoReplyConfig {
+            enabled,
+            max_concurrent: 3,
+            timeout_secs: 120,
+            suppress_patterns: vec!["/stop".to_string(), "/pause".to_string()],
+        }
+    }
+
+    #[test]
+    fn test_disabled_engine() {
+        let engine = AutoReplyEngine::new(test_config(false));
+        let agent_id = AgentId::new();
+        assert!(engine.should_reply("hello", "telegram", agent_id).is_none());
+    }
+
+    #[test]
+    fn test_enabled_engine_allows() {
+        let engine = AutoReplyEngine::new(test_config(true));
+        let agent_id = AgentId::new();
+        let result = engine.should_reply("hello there", "telegram", agent_id);
+        assert_eq!(result, Some(agent_id));
+    }
+
+    #[test]
+    fn test_suppression_patterns() {
+        let engine = AutoReplyEngine::new(test_config(true));
+        let agent_id = AgentId::new();
+
+        // Should be suppressed
+        assert!(engine.should_reply("/stop", "telegram", agent_id).is_none());
+        assert!(engine
+            .should_reply("please /pause this", "telegram", agent_id)
+            .is_none());
+
+        // Not suppressed
+        assert!(engine.should_reply("hello", "telegram", agent_id).is_some());
+    }
+
+    #[test]
+    fn test_concurrency_limit() {
+        let config = AutoReplyConfig {
+            enabled: true,
+            max_concurrent: 2,
+            timeout_secs: 120,
+            suppress_patterns: Vec::new(),
+        };
+        let engine = AutoReplyEngine::new(config);
+        assert_eq!(engine.available_permits(), 2);
+    }
+
+    #[test]
+    fn test_is_enabled() {
+        let on = AutoReplyEngine::new(test_config(true));
+        assert!(on.is_enabled());
+
+        let off = AutoReplyEngine::new(test_config(false));
+        assert!(!off.is_enabled());
+    }
+
+    #[test]
+    fn test_default_config() {
+        let config = AutoReplyConfig::default();
+        assert!(!config.enabled);
+        assert_eq!(config.max_concurrent, 3);
+        assert_eq!(config.timeout_secs, 120);
+        assert!(config.suppress_patterns.contains(&"/stop".to_string()));
+    }
+}

crates/openfang-kernel/src/background.rs

@@ -0,0 +1,457 @@
+//! Background agent executor — runs agents autonomously on schedules, timers, and conditions.
+//!
+//! Supports three autonomous modes:
+//! - **Continuous**: Agent self-prompts on a fixed interval.
+//! - **Periodic**: Agent wakes on a simplified cron schedule (e.g. "every 5m").
+//! - **Proactive**: Agent wakes when matching events fire (via the trigger engine).
+
+use crate::triggers::TriggerPattern;
+use dashmap::DashMap;
+use openfang_types::agent::{AgentId, ScheduleMode};
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use tokio::sync::watch;
+use tokio::task::JoinHandle;
+use tracing::{debug, info, warn};
+
+/// Maximum number of concurrent background LLM calls across all agents.
+const MAX_CONCURRENT_BG_LLM: usize = 5;
+
+/// Manages background task loops for autonomous agents.
+pub struct BackgroundExecutor {
+    /// Running background task handles, keyed by agent ID.
+    tasks: DashMap<AgentId, JoinHandle<()>>,
+    /// Shutdown signal receiver (from Supervisor).
+    shutdown_rx: watch::Receiver<bool>,
+    /// SECURITY: Global semaphore to limit concurrent background LLM calls.
+    llm_semaphore: Arc<tokio::sync::Semaphore>,
+}
+
+impl BackgroundExecutor {
+    /// Create a new executor bound to the supervisor's shutdown signal.
+    pub fn new(shutdown_rx: watch::Receiver<bool>) -> Self {
+        Self {
+            tasks: DashMap::new(),
+            shutdown_rx,
+            llm_semaphore: Arc::new(tokio::sync::Semaphore::new(MAX_CONCURRENT_BG_LLM)),
+        }
+    }
+
+    /// Start a background loop for an agent based on its schedule mode.
+    ///
+    /// For `Continuous` and `Periodic` modes, spawns a tokio task that
+    /// periodically sends a self-prompt message to the agent.
+    /// For `Proactive` mode, registers triggers — no dedicated task needed.
+    ///
+    /// `send_message` is a closure that sends a message to the given agent
+    /// and returns a result. It captures an `Arc<OpenFangKernel>` from the caller.
+    pub fn start_agent<F>(
+        &self,
+        agent_id: AgentId,
+        agent_name: &str,
+        schedule: &ScheduleMode,
+        send_message: F,
+    ) where
+        F: Fn(AgentId, String) -> tokio::task::JoinHandle<()> + Send + Sync + 'static,
+    {
+        match schedule {
+            ScheduleMode::Reactive => {} // nothing to do
+            ScheduleMode::Continuous {
+                check_interval_secs,
+            } => {
+                let interval = std::time::Duration::from_secs(*check_interval_secs);
+                let name = agent_name.to_string();
+                let mut shutdown = self.shutdown_rx.clone();
+                let busy = Arc::new(AtomicBool::new(false));
+                let semaphore = self.llm_semaphore.clone();
+
+                info!(
+                    agent = %name, id = %agent_id,
+                    interval_secs = check_interval_secs,
+                    "Starting continuous background loop"
+                );
+
+                let handle = tokio::spawn(async move {
+                    loop {
+                        tokio::select! {
+                            _ = tokio::time::sleep(interval) => {}
+                            _ = shutdown.changed() => {
+                                info!(agent = %name, "Continuous loop: shutdown signal received");
+                                break;
+                            }
+                        }
+
+                        // Skip if previous tick is still running
+                        if busy
+                            .compare_exchange(false, true, Ordering::SeqCst, Ordering::SeqCst)
+                            .is_err()
+                        {
+                            debug!(agent = %name, "Continuous loop: skipping tick (busy)");
+                            continue;
+                        }
+
+                        // SECURITY: Acquire global LLM concurrency permit
+                        let permit = match semaphore.clone().acquire_owned().await {
+                            Ok(p) => p,
+                            Err(_) => {
+                                busy.store(false, Ordering::SeqCst);
+                                break; // Semaphore closed
+                            }
+                        };
+
+                        let prompt = format!(
+                            "[AUTONOMOUS TICK] You are running in continuous mode. \
+                             Check your goals, review shared memory for pending tasks, \
+                             and take any necessary actions. Agent: {name}"
+                        );
+                        debug!(agent = %name, "Continuous loop: sending self-prompt");
+                        let busy_clone = busy.clone();
+                        let jh = (send_message)(agent_id, prompt);
+                        // Spawn a watcher that clears the busy flag and drops permit when done
+                        tokio::spawn(async move {
+                            let _ = jh.await;
+                            drop(permit);
+                            busy_clone.store(false, Ordering::SeqCst);
+                        });
+                    }
+                });
+
+                self.tasks.insert(agent_id, handle);
+            }
+            ScheduleMode::Periodic { cron } => {
+                let interval_secs = parse_cron_to_secs(cron);
+                let interval = std::time::Duration::from_secs(interval_secs);
+                let name = agent_name.to_string();
+                let cron_owned = cron.clone();
+                let mut shutdown = self.shutdown_rx.clone();
+                let busy = Arc::new(AtomicBool::new(false));
+                let semaphore = self.llm_semaphore.clone();
+
+                info!(
+                    agent = %name, id = %agent_id,
+                    cron = %cron, interval_secs = interval_secs,
+                    "Starting periodic background loop"
+                );
+
+                let handle = tokio::spawn(async move {
+                    loop {
+                        tokio::select! {
+                            _ = tokio::time::sleep(interval) => {}
+                            _ = shutdown.changed() => {
+                                info!(agent = %name, "Periodic loop: shutdown signal received");
+                                break;
+                            }
+                        }
+
+                        if busy
+                            .compare_exchange(false, true, Ordering::SeqCst, Ordering::SeqCst)
+                            .is_err()
+                        {
+                            debug!(agent = %name, "Periodic loop: skipping tick (busy)");
+                            continue;
+                        }
+
+                        // SECURITY: Acquire global LLM concurrency permit
+                        let permit = match semaphore.clone().acquire_owned().await {
+                            Ok(p) => p,
+                            Err(_) => {
+                                busy.store(false, Ordering::SeqCst);
+                                break; // Semaphore closed
+                            }
+                        };
+
+                        let prompt = format!(
+                            "[SCHEDULED TICK] You are running on a periodic schedule ({cron_owned}). \
+                             Perform your routine duties. Agent: {name}"
+                        );
+                        debug!(agent = %name, "Periodic loop: sending scheduled prompt");
+                        let busy_clone = busy.clone();
+                        let jh = (send_message)(agent_id, prompt);
+                        tokio::spawn(async move {
+                            let _ = jh.await;
+                            drop(permit);
+                            busy_clone.store(false, Ordering::SeqCst);
+                        });
+                    }
+                });
+
+                self.tasks.insert(agent_id, handle);
+            }
+            ScheduleMode::Proactive { .. } => {
+                // Proactive agents rely on triggers, not a dedicated loop.
+                // Triggers are registered by the kernel during spawn_agent / start_background_agents.
+                debug!(agent = %agent_name, "Proactive agent — triggers handle activation");
+            }
+        }
+    }
+
+    /// Stop the background loop for an agent, if one is running.
+    pub fn stop_agent(&self, agent_id: AgentId) {
+        if let Some((_, handle)) = self.tasks.remove(&agent_id) {
+            handle.abort();
+            info!(id = %agent_id, "Background loop stopped");
+        }
+    }
+
+    /// Number of actively running background loops.
+    pub fn active_count(&self) -> usize {
+        self.tasks.len()
+    }
+}
+
+/// Parse a proactive condition string into a `TriggerPattern`.
+///
+/// Supported formats:
+/// - `"event:agent_spawned"` → `TriggerPattern::AgentSpawned { name_pattern: "*" }`
+/// - `"event:agent_terminated"` → `TriggerPattern::AgentTerminated`
+/// - `"event:lifecycle"` → `TriggerPattern::Lifecycle`
+/// - `"event:system"` → `TriggerPattern::System`
+/// - `"memory:some_key"` → `TriggerPattern::MemoryKeyPattern { key_pattern: "some_key" }`
+/// - `"all"` → `TriggerPattern::All`
+pub fn parse_condition(condition: &str) -> Option<TriggerPattern> {
+    let condition = condition.trim();
+
+    if condition.eq_ignore_ascii_case("all") {
+        return Some(TriggerPattern::All);
+    }
+
+    if let Some(event_kind) = condition.strip_prefix("event:") {
+        let kind = event_kind.trim().to_lowercase();
+        return match kind.as_str() {
+            "agent_spawned" => Some(TriggerPattern::AgentSpawned {
+                name_pattern: "*".to_string(),
+            }),
+            "agent_terminated" => Some(TriggerPattern::AgentTerminated),
+            "lifecycle" => Some(TriggerPattern::Lifecycle),
+            "system" => Some(TriggerPattern::System),
+            "memory_update" => Some(TriggerPattern::MemoryUpdate),
+            other => {
+                warn!(condition = %condition, "Unknown event condition: {other}");
+                None
+            }
+        };
+    }
+
+    if let Some(key) = condition.strip_prefix("memory:") {
+        return Some(TriggerPattern::MemoryKeyPattern {
+            key_pattern: key.trim().to_string(),
+        });
+    }
+
+    warn!(condition = %condition, "Unrecognized proactive condition format");
+    None
+}
+
+/// Parse a simplified cron expression into seconds.
+///
+/// Supported formats:
+/// - `"every 30s"` → 30
+/// - `"every 5m"` → 300
+/// - `"every 1h"` → 3600
+/// - `"every 2d"` → 172800
+///
+/// Falls back to 300 seconds (5 minutes) for unparseable expressions.
+pub fn parse_cron_to_secs(cron: &str) -> u64 {
+    let cron = cron.trim().to_lowercase();
+
+    // Try "every <N><unit>" format
+    if let Some(rest) = cron.strip_prefix("every ") {
+        let rest = rest.trim();
+        if let Some(num_str) = rest.strip_suffix('s') {
+            if let Ok(n) = num_str.trim().parse::<u64>() {
+                return n;
+            }
+        }
+        if let Some(num_str) = rest.strip_suffix('m') {
+            if let Ok(n) = num_str.trim().parse::<u64>() {
+                return n * 60;
+            }
+        }
+        if let Some(num_str) = rest.strip_suffix('h') {
+            if let Ok(n) = num_str.trim().parse::<u64>() {
+                return n * 3600;
+            }
+        }
+        if let Some(num_str) = rest.strip_suffix('d') {
+            if let Ok(n) = num_str.trim().parse::<u64>() {
+                return n * 86400;
+            }
+        }
+    }
+
+    warn!(cron = %cron, "Unparseable cron expression, defaulting to 300s");
+    300
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_parse_cron_seconds() {
+        assert_eq!(parse_cron_to_secs("every 30s"), 30);
+        assert_eq!(parse_cron_to_secs("every 1s"), 1);
+    }
+
+    #[test]
+    fn test_parse_cron_minutes() {
+        assert_eq!(parse_cron_to_secs("every 5m"), 300);
+        assert_eq!(parse_cron_to_secs("every 1m"), 60);
+    }
+
+    #[test]
+    fn test_parse_cron_hours() {
+        assert_eq!(parse_cron_to_secs("every 1h"), 3600);
+        assert_eq!(parse_cron_to_secs("every 2h"), 7200);
+    }
+
+    #[test]
+    fn test_parse_cron_days() {
+        assert_eq!(parse_cron_to_secs("every 1d"), 86400);
+    }
+
+    #[test]
+    fn test_parse_cron_fallback() {
+        // Unparseable → 300
+        assert_eq!(parse_cron_to_secs("*/5 * * * *"), 300);
+        assert_eq!(parse_cron_to_secs("gibberish"), 300);
+    }
+
+    #[test]
+    fn test_parse_condition_events() {
+        assert!(matches!(
+            parse_condition("event:agent_spawned"),
+            Some(TriggerPattern::AgentSpawned { .. })
+        ));
+        assert!(matches!(
+            parse_condition("event:agent_terminated"),
+            Some(TriggerPattern::AgentTerminated)
+        ));
+        assert!(matches!(
+            parse_condition("event:lifecycle"),
+            Some(TriggerPattern::Lifecycle)
+        ));
+        assert!(matches!(
+            parse_condition("event:system"),
+            Some(TriggerPattern::System)
+        ));
+        assert!(matches!(
+            parse_condition("event:memory_update"),
+            Some(TriggerPattern::MemoryUpdate)
+        ));
+    }
+
+    #[test]
+    fn test_parse_condition_memory() {
+        match parse_condition("memory:agent.*.status") {
+            Some(TriggerPattern::MemoryKeyPattern { key_pattern }) => {
+                assert_eq!(key_pattern, "agent.*.status");
+            }
+            other => panic!("Expected MemoryKeyPattern, got {:?}", other),
+        }
+    }
+
+    #[test]
+    fn test_parse_condition_all() {
+        assert!(matches!(parse_condition("all"), Some(TriggerPattern::All)));
+    }
+
+    #[test]
+    fn test_parse_condition_unknown() {
+        assert!(parse_condition("event:unknown_thing").is_none());
+        assert!(parse_condition("badprefix:foo").is_none());
+    }
+
+    #[tokio::test]
+    async fn test_continuous_shutdown() {
+        let (shutdown_tx, shutdown_rx) = watch::channel(false);
+        let executor = BackgroundExecutor::new(shutdown_rx);
+        let agent_id = AgentId::new();
+
+        let tick_count = Arc::new(std::sync::atomic::AtomicU64::new(0));
+        let tick_clone = tick_count.clone();
+
+        let schedule = ScheduleMode::Continuous {
+            check_interval_secs: 1, // 1 second for fast test
+        };
+
+        executor.start_agent(agent_id, "test-agent", &schedule, move |_id, _msg| {
+            let tc = tick_clone.clone();
+            tokio::spawn(async move {
+                tc.fetch_add(1, Ordering::SeqCst);
+            })
+        });
+
+        assert_eq!(executor.active_count(), 1);
+
+        // Wait for at least 1 tick
+        tokio::time::sleep(std::time::Duration::from_millis(1500)).await;
+        assert!(tick_count.load(Ordering::SeqCst) >= 1);
+
+        // Shutdown
+        let _ = shutdown_tx.send(true);
+        tokio::time::sleep(std::time::Duration::from_millis(200)).await;
+
+        // The loop should have exited (handle finished)
+        // Active count still shows the entry until stop_agent is called
+        executor.stop_agent(agent_id);
+        assert_eq!(executor.active_count(), 0);
+    }
+
+    #[tokio::test]
+    async fn test_skip_if_busy() {
+        let (_shutdown_tx, shutdown_rx) = watch::channel(false);
+        let executor = BackgroundExecutor::new(shutdown_rx);
+        let agent_id = AgentId::new();
+
+        let tick_count = Arc::new(std::sync::atomic::AtomicU64::new(0));
+        let tick_clone = tick_count.clone();
+
+        let schedule = ScheduleMode::Continuous {
+            check_interval_secs: 1,
+        };
+
+        // Each tick takes 3 seconds — should cause subsequent ticks to be skipped
+        executor.start_agent(agent_id, "slow-agent", &schedule, move |_id, _msg| {
+            let tc = tick_clone.clone();
+            tokio::spawn(async move {
+                tc.fetch_add(1, Ordering::SeqCst);
+                tokio::time::sleep(std::time::Duration::from_secs(3)).await;
+            })
+        });
+
+        // Wait 2.5 seconds: 1 tick should fire at t=1s, second at t=2s should be skipped (busy)
+        tokio::time::sleep(std::time::Duration::from_millis(2500)).await;
+        let ticks = tick_count.load(Ordering::SeqCst);
+        // Should be exactly 1 because the first tick is still "busy" when the second arrives
+        assert_eq!(ticks, 1, "Expected 1 tick (skip-if-busy), got {ticks}");
+
+        executor.stop_agent(agent_id);
+    }
+
+    #[test]
+    fn test_executor_active_count() {
+        let (_tx, rx) = watch::channel(false);
+        let executor = BackgroundExecutor::new(rx);
+        assert_eq!(executor.active_count(), 0);
+
+        // Reactive mode → no background task
+        let id = AgentId::new();
+        executor.start_agent(id, "reactive", &ScheduleMode::Reactive, |_id, _msg| {
+            tokio::spawn(async {})
+        });
+        assert_eq!(executor.active_count(), 0);
+
+        // Proactive mode → no dedicated task
+        let id2 = AgentId::new();
+        executor.start_agent(
+            id2,
+            "proactive",
+            &ScheduleMode::Proactive {
+                conditions: vec!["event:agent_spawned".to_string()],
+            },
+            |_id, _msg| tokio::spawn(async {}),
+        );
+        assert_eq!(executor.active_count(), 0);
+    }
+}

crates/openfang-kernel/src/capabilities.rs

@@ -0,0 +1,95 @@
+//! Capability manager — enforces capability-based security.
+
+use dashmap::DashMap;
+use openfang_types::agent::AgentId;
+use openfang_types::capability::{capability_matches, Capability, CapabilityCheck};
+use tracing::debug;
+
+/// Manages capability grants for all agents.
+pub struct CapabilityManager {
+    /// Granted capabilities per agent.
+    grants: DashMap<AgentId, Vec<Capability>>,
+}
+
+impl CapabilityManager {
+    /// Create a new capability manager.
+    pub fn new() -> Self {
+        Self {
+            grants: DashMap::new(),
+        }
+    }
+
+    /// Grant capabilities to an agent.
+    pub fn grant(&self, agent_id: AgentId, capabilities: Vec<Capability>) {
+        self.grants.insert(agent_id, capabilities);
+    }
+
+    /// Check whether an agent has a specific capability.
+    pub fn check(&self, agent_id: AgentId, required: &Capability) -> CapabilityCheck {
+        let grants = match self.grants.get(&agent_id) {
+            Some(g) => g,
+            None => {
+                return CapabilityCheck::Denied(format!(
+                    "No capabilities registered for agent {agent_id}"
+                ))
+            }
+        };
+
+        for granted in grants.value() {
+            if capability_matches(granted, required) {
+                debug!(agent = %agent_id, ?required, "Capability granted");
+                return CapabilityCheck::Granted;
+            }
+        }
+
+        CapabilityCheck::Denied(format!(
+            "Agent {agent_id} does not have capability: {required:?}"
+        ))
+    }
+
+    /// List all capabilities for an agent.
+    pub fn list(&self, agent_id: AgentId) -> Vec<Capability> {
+        self.grants
+            .get(&agent_id)
+            .map(|g| g.value().clone())
+            .unwrap_or_default()
+    }
+
+    /// Remove all capabilities for an agent.
+    pub fn revoke_all(&self, agent_id: AgentId) {
+        self.grants.remove(&agent_id);
+    }
+}
+
+impl Default for CapabilityManager {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_grant_and_check() {
+        let mgr = CapabilityManager::new();
+        let id = AgentId::new();
+        mgr.grant(id, vec![Capability::ToolInvoke("file_read".to_string())]);
+        assert!(mgr
+            .check(id, &Capability::ToolInvoke("file_read".to_string()))
+            .is_granted());
+        assert!(!mgr
+            .check(id, &Capability::ToolInvoke("shell_exec".to_string()))
+            .is_granted());
+    }
+
+    #[test]
+    fn test_no_grants() {
+        let mgr = CapabilityManager::new();
+        let id = AgentId::new();
+        assert!(!mgr
+            .check(id, &Capability::ToolInvoke("anything".to_string()))
+            .is_granted());
+    }
+}

crates/openfang-kernel/src/config.rs

@@ -0,0 +1,434 @@
+//! Configuration loading from `~/.openfang/config.toml` with defaults.
+//!
+//! Supports config includes: the `include` field specifies additional TOML files
+//! to load and deep-merge before the root config (root overrides includes).
+
+use openfang_types::config::KernelConfig;
+use std::collections::HashSet;
+use std::path::{Path, PathBuf};
+use tracing::info;
+
+/// Maximum include nesting depth.
+const MAX_INCLUDE_DEPTH: u32 = 10;
+
+/// Load kernel configuration from a TOML file, with defaults.
+///
+/// If the config contains an `include` field, included files are loaded
+/// and deep-merged first, then the root config overrides them.
+pub fn load_config(path: Option<&Path>) -> KernelConfig {
+    let config_path = path
+        .map(|p| p.to_path_buf())
+        .unwrap_or_else(default_config_path);
+
+    if config_path.exists() {
+        match std::fs::read_to_string(&config_path) {
+            Ok(contents) => match toml::from_str::<toml::Value>(&contents) {
+                Ok(mut root_value) => {
+                    // Process includes before deserializing
+                    let config_dir = config_path
+                        .parent()
+                        .unwrap_or_else(|| Path::new("."))
+                        .to_path_buf();
+                    let mut visited = HashSet::new();
+                    if let Ok(canonical) = std::fs::canonicalize(&config_path) {
+                        visited.insert(canonical);
+                    } else {
+                        visited.insert(config_path.clone());
+                    }
+
+                    if let Err(e) =
+                        resolve_config_includes(&mut root_value, &config_dir, &mut visited, 0)
+                    {
+                        tracing::warn!(
+                            error = %e,
+                            "Config include resolution failed, using root config only"
+                        );
+                    }
+
+                    // Remove the `include` field before deserializing to avoid confusion
+                    if let toml::Value::Table(ref mut tbl) = root_value {
+                        tbl.remove("include");
+                    }
+
+                    match root_value.try_into::<KernelConfig>() {
+                        Ok(config) => {
+                            info!(path = %config_path.display(), "Loaded configuration");
+                            return config;
+                        }
+                        Err(e) => {
+                            tracing::warn!(
+                                error = %e,
+                                path = %config_path.display(),
+                                "Failed to deserialize merged config, using defaults"
+                            );
+                        }
+                    }
+                }
+                Err(e) => {
+                    tracing::warn!(
+                        error = %e,
+                        path = %config_path.display(),
+                        "Failed to parse config, using defaults"
+                    );
+                }
+            },
+            Err(e) => {
+                tracing::warn!(
+                    error = %e,
+                    path = %config_path.display(),
+                    "Failed to read config file, using defaults"
+                );
+            }
+        }
+    } else {
+        info!(
+            path = %config_path.display(),
+            "Config file not found, using defaults"
+        );
+    }
+
+    KernelConfig::default()
+}
+
+/// Resolve config includes by deep-merging included files into the root value.
+///
+/// Included files are loaded first and the root config overrides them.
+/// Security: rejects absolute paths, `..` components, and circular references.
+fn resolve_config_includes(
+    root_value: &mut toml::Value,
+    config_dir: &Path,
+    visited: &mut HashSet<PathBuf>,
+    depth: u32,
+) -> Result<(), String> {
+    if depth > MAX_INCLUDE_DEPTH {
+        return Err(format!(
+            "Config include depth exceeded maximum of {MAX_INCLUDE_DEPTH}"
+        ));
+    }
+
+    // Extract include list from the current value
+    let includes = match root_value {
+        toml::Value::Table(tbl) => {
+            if let Some(toml::Value::Array(arr)) = tbl.get("include") {
+                arr.iter()
+                    .filter_map(|v| v.as_str().map(String::from))
+                    .collect::<Vec<_>>()
+            } else {
+                return Ok(());
+            }
+        }
+        _ => return Ok(()),
+    };
+
+    if includes.is_empty() {
+        return Ok(());
+    }
+
+    // Merge each include (earlier includes are overridden by later ones,
+    // and the root config overrides everything).
+    let mut merged_base = toml::Value::Table(toml::map::Map::new());
+
+    for include_path_str in &includes {
+        // SECURITY: reject absolute paths
+        let include_path = Path::new(include_path_str);
+        if include_path.is_absolute() {
+            return Err(format!(
+                "Config include rejects absolute path: {include_path_str}"
+            ));
+        }
+        // SECURITY: reject `..` components
+        for component in include_path.components() {
+            if let std::path::Component::ParentDir = component {
+                return Err(format!(
+                    "Config include rejects path traversal: {include_path_str}"
+                ));
+            }
+        }
+
+        let resolved = config_dir.join(include_path);
+        // SECURITY: verify resolved path stays within config dir
+        let canonical = std::fs::canonicalize(&resolved).map_err(|e| {
+            format!(
+                "Config include '{}' cannot be resolved: {e}",
+                include_path_str
+            )
+        })?;
+        let canonical_dir = std::fs::canonicalize(config_dir)
+            .map_err(|e| format!("Config dir cannot be canonicalized: {e}"))?;
+        if !canonical.starts_with(&canonical_dir) {
+            return Err(format!(
+                "Config include '{}' escapes config directory",
+                include_path_str
+            ));
+        }
+
+        // SECURITY: circular detection
+        if !visited.insert(canonical.clone()) {
+            return Err(format!(
+                "Circular config include detected: {include_path_str}"
+            ));
+        }
+
+        info!(include = %include_path_str, "Loading config include");
+
+        let contents = std::fs::read_to_string(&canonical)
+            .map_err(|e| format!("Failed to read config include '{}': {e}", include_path_str))?;
+        let mut include_value: toml::Value = toml::from_str(&contents)
+            .map_err(|e| format!("Failed to parse config include '{}': {e}", include_path_str))?;
+
+        // Recursively resolve includes in the included file
+        let include_dir = canonical.parent().unwrap_or(config_dir).to_path_buf();
+        resolve_config_includes(&mut include_value, &include_dir, visited, depth + 1)?;
+
+        // Remove include field from the included file
+        if let toml::Value::Table(ref mut tbl) = include_value {
+            tbl.remove("include");
+        }
+
+        // Deep merge: include overrides the base built so far
+        deep_merge_toml(&mut merged_base, &include_value);
+    }
+
+    // Now deep merge: root overrides the merged includes
+    // Save root's current values (minus include), then merge root on top
+    let root_without_include = {
+        let mut v = root_value.clone();
+        if let toml::Value::Table(ref mut tbl) = v {
+            tbl.remove("include");
+        }
+        v
+    };
+    deep_merge_toml(&mut merged_base, &root_without_include);
+    *root_value = merged_base;
+
+    Ok(())
+}
+
+/// Deep-merge two TOML values. `overlay` values override `base` values.
+/// For tables, recursively merge. For everything else, overlay wins.
+pub fn deep_merge_toml(base: &mut toml::Value, overlay: &toml::Value) {
+    match (base, overlay) {
+        (toml::Value::Table(base_tbl), toml::Value::Table(overlay_tbl)) => {
+            for (key, overlay_val) in overlay_tbl {
+                if let Some(base_val) = base_tbl.get_mut(key) {
+                    deep_merge_toml(base_val, overlay_val);
+                } else {
+                    base_tbl.insert(key.clone(), overlay_val.clone());
+                }
+            }
+        }
+        (base, overlay) => {
+            *base = overlay.clone();
+        }
+    }
+}
+
+/// Get the default config file path.
+pub fn default_config_path() -> PathBuf {
+    dirs::home_dir()
+        .unwrap_or_else(std::env::temp_dir)
+        .join(".openfang")
+        .join("config.toml")
+}
+
+/// Get the default OpenFang home directory.
+pub fn openfang_home() -> PathBuf {
+    dirs::home_dir()
+        .unwrap_or_else(std::env::temp_dir)
+        .join(".openfang")
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::Write;
+
+    #[test]
+    fn test_load_config_defaults() {
+        let config = load_config(None);
+        assert_eq!(config.log_level, "info");
+    }
+
+    #[test]
+    fn test_load_config_missing_file() {
+        let config = load_config(Some(Path::new("/nonexistent/config.toml")));
+        assert_eq!(config.log_level, "info");
+    }
+
+    #[test]
+    fn test_deep_merge_simple() {
+        let mut base: toml::Value = toml::from_str(
+            r#"
+            log_level = "debug"
+            api_listen = "0.0.0.0:4200"
+        "#,
+        )
+        .unwrap();
+        let overlay: toml::Value = toml::from_str(
+            r#"
+            log_level = "info"
+            network_enabled = true
+        "#,
+        )
+        .unwrap();
+        deep_merge_toml(&mut base, &overlay);
+        assert_eq!(base["log_level"].as_str(), Some("info"));
+        assert_eq!(base["api_listen"].as_str(), Some("0.0.0.0:4200"));
+        assert_eq!(base["network_enabled"].as_bool(), Some(true));
+    }
+
+    #[test]
+    fn test_deep_merge_nested_tables() {
+        let mut base: toml::Value = toml::from_str(
+            r#"
+            [memory]
+            decay_rate = 0.1
+            consolidation_threshold = 10000
+        "#,
+        )
+        .unwrap();
+        let overlay: toml::Value = toml::from_str(
+            r#"
+            [memory]
+            decay_rate = 0.5
+        "#,
+        )
+        .unwrap();
+        deep_merge_toml(&mut base, &overlay);
+        let mem = base["memory"].as_table().unwrap();
+        assert_eq!(mem["decay_rate"].as_float(), Some(0.5));
+        assert_eq!(mem["consolidation_threshold"].as_integer(), Some(10000));
+    }
+
+    #[test]
+    fn test_basic_include() {
+        let dir = tempfile::tempdir().unwrap();
+        let base_path = dir.path().join("base.toml");
+        let root_path = dir.path().join("config.toml");
+
+        // Base config
+        let mut f = std::fs::File::create(&base_path).unwrap();
+        writeln!(f, "log_level = \"debug\"").unwrap();
+        writeln!(f, "api_listen = \"0.0.0.0:9999\"").unwrap();
+        drop(f);
+
+        // Root config (includes base, overrides log_level)
+        let mut f = std::fs::File::create(&root_path).unwrap();
+        writeln!(f, "include = [\"base.toml\"]").unwrap();
+        writeln!(f, "log_level = \"warn\"").unwrap();
+        drop(f);
+
+        let config = load_config(Some(&root_path));
+        assert_eq!(config.log_level, "warn"); // root overrides
+        assert_eq!(config.api_listen, "0.0.0.0:9999"); // from base
+    }
+
+    #[test]
+    fn test_nested_include() {
+        let dir = tempfile::tempdir().unwrap();
+        let grandchild = dir.path().join("grandchild.toml");
+        let child = dir.path().join("child.toml");
+        let root = dir.path().join("config.toml");
+
+        let mut f = std::fs::File::create(&grandchild).unwrap();
+        writeln!(f, "log_level = \"trace\"").unwrap();
+        drop(f);
+
+        let mut f = std::fs::File::create(&child).unwrap();
+        writeln!(f, "include = [\"grandchild.toml\"]").unwrap();
+        writeln!(f, "log_level = \"debug\"").unwrap();
+        drop(f);
+
+        let mut f = std::fs::File::create(&root).unwrap();
+        writeln!(f, "include = [\"child.toml\"]").unwrap();
+        writeln!(f, "log_level = \"info\"").unwrap();
+        drop(f);
+
+        let config = load_config(Some(&root));
+        assert_eq!(config.log_level, "info"); // root wins
+    }
+
+    #[test]
+    fn test_circular_include_detected() {
+        let dir = tempfile::tempdir().unwrap();
+        let a_path = dir.path().join("a.toml");
+        let b_path = dir.path().join("b.toml");
+
+        let mut f = std::fs::File::create(&a_path).unwrap();
+        writeln!(f, "include = [\"b.toml\"]").unwrap();
+        writeln!(f, "log_level = \"info\"").unwrap();
+        drop(f);
+
+        let mut f = std::fs::File::create(&b_path).unwrap();
+        writeln!(f, "include = [\"a.toml\"]").unwrap();
+        drop(f);
+
+        // Should not panic — circular detection triggers, falls back gracefully
+        let config = load_config(Some(&a_path));
+        // Falls back to defaults due to the circular error
+        assert!(!config.log_level.is_empty());
+    }
+
+    #[test]
+    fn test_path_traversal_blocked() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path().join("config.toml");
+
+        let mut f = std::fs::File::create(&root).unwrap();
+        writeln!(f, "include = [\"../etc/passwd\"]").unwrap();
+        drop(f);
+
+        // Should not panic — path traversal triggers error, falls back
+        let config = load_config(Some(&root));
+        assert_eq!(config.log_level, "info"); // defaults
+    }
+
+    #[test]
+    fn test_max_depth_exceeded() {
+        let dir = tempfile::tempdir().unwrap();
+
+        // Create a chain of 12 files (exceeds MAX_INCLUDE_DEPTH=10)
+        for i in (0..12).rev() {
+            let name = format!("level{i}.toml");
+            let path = dir.path().join(&name);
+            let mut f = std::fs::File::create(&path).unwrap();
+            if i < 11 {
+                let next = format!("level{}.toml", i + 1);
+                writeln!(f, "include = [\"{next}\"]").unwrap();
+            }
+            writeln!(f, "log_level = \"level{i}\"").unwrap();
+            drop(f);
+        }
+
+        let root = dir.path().join("level0.toml");
+        let config = load_config(Some(&root));
+        // Falls back due to depth limit — but should not panic
+        assert!(!config.log_level.is_empty());
+    }
+
+    #[test]
+    fn test_absolute_path_rejected() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path().join("config.toml");
+
+        let mut f = std::fs::File::create(&root).unwrap();
+        writeln!(f, "include = [\"/etc/shadow\"]").unwrap();
+        drop(f);
+
+        let config = load_config(Some(&root));
+        assert_eq!(config.log_level, "info"); // defaults
+    }
+
+    #[test]
+    fn test_no_includes_works() {
+        let dir = tempfile::tempdir().unwrap();
+        let root = dir.path().join("config.toml");
+
+        let mut f = std::fs::File::create(&root).unwrap();
+        writeln!(f, "log_level = \"trace\"").unwrap();
+        drop(f);
+
+        let config = load_config(Some(&root));
+        assert_eq!(config.log_level, "trace");
+    }
+}

crates/openfang-kernel/src/config_reload.rs

@@ -0,0 +1,674 @@
+//! Config hot-reload — diffs two `KernelConfig` instances and produces a `ReloadPlan`.
+//!
+//! **Hot-reload safe**: channels, skills, usage footer, web config, browser,
+//! approval policy, cron settings, webhook triggers, extensions.
+//!
+//! **No-op** (informational only): log_level, language, mode.
+//!
+//! **Restart required**: api_listen, api_key, network, memory, default_model.
+
+use openfang_types::config::{KernelConfig, ReloadMode};
+use tracing::{info, warn};
+
+// ---------------------------------------------------------------------------
+// HotAction — what can be changed at runtime without restart
+// ---------------------------------------------------------------------------
+
+/// An individual action that can be applied at runtime (hot-reload).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum HotAction {
+    /// Channel configuration changed — reload channel bridges.
+    ReloadChannels,
+    /// Skill configuration changed — reload skill registry.
+    ReloadSkills,
+    /// Usage footer mode changed.
+    UpdateUsageFooter,
+    /// Web config changed — rebuild web tools context.
+    ReloadWebConfig,
+    /// Browser config changed.
+    ReloadBrowserConfig,
+    /// Approval policy changed.
+    UpdateApprovalPolicy,
+    /// Cron max jobs changed.
+    UpdateCronConfig,
+    /// Webhook trigger config changed.
+    UpdateWebhookConfig,
+    /// Extension config changed.
+    ReloadExtensions,
+    /// MCP server list changed — reconnect MCP clients.
+    ReloadMcpServers,
+    /// A2A config changed.
+    ReloadA2aConfig,
+    /// Fallback provider chain changed.
+    ReloadFallbackProviders,
+    /// Provider base URL overrides changed.
+    ReloadProviderUrls,
+}
+
+// ---------------------------------------------------------------------------
+// ReloadPlan — the output of diffing two configs
+// ---------------------------------------------------------------------------
+
+/// A categorized plan for applying config changes.
+///
+/// After building a plan via [`build_reload_plan`], callers inspect
+/// `restart_required` to decide whether a full restart is needed or
+/// the `hot_actions` can be applied in-place.
+#[derive(Debug, Clone)]
+pub struct ReloadPlan {
+    /// Whether a full restart is needed.
+    pub restart_required: bool,
+    /// Human-readable reasons why restart is required.
+    pub restart_reasons: Vec<String>,
+    /// Actions that can be hot-reloaded without restart.
+    pub hot_actions: Vec<HotAction>,
+    /// Fields that changed but are no-ops (informational only).
+    pub noop_changes: Vec<String>,
+}
+
+impl ReloadPlan {
+    /// Whether any changes were detected at all.
+    pub fn has_changes(&self) -> bool {
+        self.restart_required || !self.hot_actions.is_empty() || !self.noop_changes.is_empty()
+    }
+
+    /// Whether the plan can be applied without restart.
+    pub fn is_hot_reloadable(&self) -> bool {
+        !self.restart_required
+    }
+
+    /// Log a human-readable summary of the plan.
+    pub fn log_summary(&self) {
+        if !self.has_changes() {
+            info!("config reload: no changes detected");
+            return;
+        }
+        if self.restart_required {
+            warn!(
+                "config reload: restart required — {}",
+                self.restart_reasons.join("; ")
+            );
+        }
+        for action in &self.hot_actions {
+            info!("config reload: hot-reload action queued — {action:?}");
+        }
+        for noop in &self.noop_changes {
+            info!("config reload: no-op change — {noop}");
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// build_reload_plan
+// ---------------------------------------------------------------------------
+
+/// Compare JSON-serialized forms of a field. Returns `true` when the
+/// serialized representations differ (or if one side fails to serialize).
+fn field_changed<T: serde::Serialize>(old: &T, new: &T) -> bool {
+    let old_json = serde_json::to_string(old).ok();
+    let new_json = serde_json::to_string(new).ok();
+    old_json != new_json
+}
+
+/// Diff two configurations and produce a reload plan.
+///
+/// The plan categorizes every detected change into one of three buckets:
+///
+/// 1. **restart_required** — the change touches something that cannot be
+///    patched at runtime (e.g. the listen address or database path).
+/// 2. **hot_actions** — the change can be applied without restarting.
+/// 3. **noop_changes** — the change is informational; no action needed.
+pub fn build_reload_plan(old: &KernelConfig, new: &KernelConfig) -> ReloadPlan {
+    let mut plan = ReloadPlan {
+        restart_required: false,
+        restart_reasons: Vec::new(),
+        hot_actions: Vec::new(),
+        noop_changes: Vec::new(),
+    };
+
+    // ----- Restart-required fields -----
+
+    if old.api_listen != new.api_listen {
+        plan.restart_required = true;
+        plan.restart_reasons.push(format!(
+            "api_listen changed: {} -> {}",
+            old.api_listen, new.api_listen
+        ));
+    }
+
+    if old.api_key != new.api_key {
+        plan.restart_required = true;
+        plan.restart_reasons.push("api_key changed".to_string());
+    }
+
+    if old.network_enabled != new.network_enabled {
+        plan.restart_required = true;
+        plan.restart_reasons
+            .push("network_enabled changed".to_string());
+    }
+
+    // Network config (shared_secret, listen_addresses, etc.)
+    if field_changed(&old.network, &new.network) {
+        plan.restart_required = true;
+        plan.restart_reasons
+            .push("network config changed".to_string());
+    }
+
+    // Memory config (requires restarting SQLite connections)
+    if field_changed(&old.memory, &new.memory) {
+        plan.restart_required = true;
+        plan.restart_reasons
+            .push("memory config changed".to_string());
+    }
+
+    // Default model (driver needs recreation)
+    if field_changed(&old.default_model, &new.default_model) {
+        plan.restart_required = true;
+        plan.restart_reasons
+            .push("default_model changed".to_string());
+    }
+
+    // Home/data directory changes
+    if old.home_dir != new.home_dir {
+        plan.restart_required = true;
+        plan.restart_reasons.push(format!(
+            "home_dir changed: {:?} -> {:?}",
+            old.home_dir, new.home_dir
+        ));
+    }
+    if old.data_dir != new.data_dir {
+        plan.restart_required = true;
+        plan.restart_reasons.push(format!(
+            "data_dir changed: {:?} -> {:?}",
+            old.data_dir, new.data_dir
+        ));
+    }
+
+    // Vault config (encryption key derivation)
+    if field_changed(&old.vault, &new.vault) {
+        plan.restart_required = true;
+        plan.restart_reasons
+            .push("vault config changed".to_string());
+    }
+
+    // ----- Hot-reloadable fields -----
+
+    if field_changed(&old.channels, &new.channels) {
+        plan.hot_actions.push(HotAction::ReloadChannels);
+    }
+
+    if old.usage_footer != new.usage_footer {
+        plan.hot_actions.push(HotAction::UpdateUsageFooter);
+    }
+
+    if field_changed(&old.web, &new.web) {
+        plan.hot_actions.push(HotAction::ReloadWebConfig);
+    }
+
+    if field_changed(&old.browser, &new.browser) {
+        plan.hot_actions.push(HotAction::ReloadBrowserConfig);
+    }
+
+    if field_changed(&old.approval, &new.approval) {
+        plan.hot_actions.push(HotAction::UpdateApprovalPolicy);
+    }
+
+    if old.max_cron_jobs != new.max_cron_jobs {
+        plan.hot_actions.push(HotAction::UpdateCronConfig);
+    }
+
+    if field_changed(&old.webhook_triggers, &new.webhook_triggers) {
+        plan.hot_actions.push(HotAction::UpdateWebhookConfig);
+    }
+
+    if field_changed(&old.extensions, &new.extensions) {
+        plan.hot_actions.push(HotAction::ReloadExtensions);
+    }
+
+    if field_changed(&old.mcp_servers, &new.mcp_servers) {
+        plan.hot_actions.push(HotAction::ReloadMcpServers);
+    }
+
+    if field_changed(&old.a2a, &new.a2a) {
+        plan.hot_actions.push(HotAction::ReloadA2aConfig);
+    }
+
+    if field_changed(&old.fallback_providers, &new.fallback_providers) {
+        plan.hot_actions.push(HotAction::ReloadFallbackProviders);
+    }
+
+    if field_changed(&old.provider_urls, &new.provider_urls) {
+        plan.hot_actions.push(HotAction::ReloadProviderUrls);
+    }
+
+    // ----- No-op fields -----
+
+    if old.log_level != new.log_level {
+        plan.noop_changes
+            .push(format!("log_level: {} -> {}", old.log_level, new.log_level));
+    }
+
+    if old.language != new.language {
+        plan.noop_changes
+            .push(format!("language: {} -> {}", old.language, new.language));
+    }
+
+    if old.mode != new.mode {
+        plan.noop_changes
+            .push(format!("mode: {:?} -> {:?}", old.mode, new.mode));
+    }
+
+    plan
+}
+
+// ---------------------------------------------------------------------------
+// validate_config_for_reload
+// ---------------------------------------------------------------------------
+
+/// Validate a new config before applying it.
+///
+/// Returns `Ok(())` if the config passes basic sanity checks, or `Err` with
+/// a list of human-readable error messages.
+pub fn validate_config_for_reload(config: &KernelConfig) -> Result<(), Vec<String>> {
+    let mut errors = Vec::new();
+
+    if config.api_listen.is_empty() {
+        errors.push("api_listen cannot be empty".to_string());
+    }
+
+    if config.max_cron_jobs > 10_000 {
+        errors.push("max_cron_jobs exceeds reasonable limit (10000)".to_string());
+    }
+
+    // Validate approval policy
+    if let Err(e) = config.approval.validate() {
+        errors.push(format!("approval policy: {e}"));
+    }
+
+    // Network config: if network is enabled, shared_secret must be set
+    if config.network_enabled && config.network.shared_secret.is_empty() {
+        errors.push("network_enabled is true but network.shared_secret is empty".to_string());
+    }
+
+    if errors.is_empty() {
+        Ok(())
+    } else {
+        Err(errors)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// should_reload — convenience helper for the reload mode
+// ---------------------------------------------------------------------------
+
+/// Given the configured [`ReloadMode`] and a [`ReloadPlan`], decide whether
+/// the caller should apply hot actions.
+///
+/// Returns `true` if hot-reload actions should be applied.
+pub fn should_apply_hot(mode: ReloadMode, plan: &ReloadPlan) -> bool {
+    match mode {
+        ReloadMode::Off => false,
+        ReloadMode::Restart => false, // caller must do a full restart
+        ReloadMode::Hot => !plan.hot_actions.is_empty(),
+        ReloadMode::Hybrid => !plan.hot_actions.is_empty(),
+    }
+}
+
+// ===========================================================================
+// Tests
+// ===========================================================================
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openfang_types::config::KernelConfig;
+
+    /// Helper: create a default config for diffing.
+    fn default_cfg() -> KernelConfig {
+        KernelConfig::default()
+    }
+
+    // -----------------------------------------------------------------------
+    // Plan detection tests
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_no_changes_detected() {
+        let a = default_cfg();
+        let b = default_cfg();
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.has_changes());
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.is_empty());
+        assert!(plan.noop_changes.is_empty());
+    }
+
+    #[test]
+    fn test_api_listen_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.api_listen = "0.0.0.0:8080".to_string();
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan
+            .restart_reasons
+            .iter()
+            .any(|r| r.contains("api_listen")));
+    }
+
+    #[test]
+    fn test_api_key_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.api_key = "super-secret-key".to_string();
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan.restart_reasons.iter().any(|r| r.contains("api_key")));
+    }
+
+    #[test]
+    fn test_network_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.network_enabled = true;
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan
+            .restart_reasons
+            .iter()
+            .any(|r| r.contains("network_enabled")));
+    }
+
+    #[test]
+    fn test_network_config_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.network.shared_secret = "new-secret".to_string();
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan
+            .restart_reasons
+            .iter()
+            .any(|r| r.contains("network config")));
+    }
+
+    #[test]
+    fn test_memory_config_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.memory.consolidation_threshold = 99_999;
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan
+            .restart_reasons
+            .iter()
+            .any(|r| r.contains("memory config")));
+    }
+
+    #[test]
+    fn test_default_model_requires_restart() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.default_model.model = "gpt-4".to_string();
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan
+            .restart_reasons
+            .iter()
+            .any(|r| r.contains("default_model")));
+    }
+
+    // -----------------------------------------------------------------------
+    // Hot-reload tests
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_channels_hot_reload() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        // Change the channels config by adding a Telegram config
+        b.channels.telegram = Some(openfang_types::config::TelegramConfig {
+            bot_token_env: "TG_TOKEN".to_string(),
+            ..Default::default()
+        });
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.contains(&HotAction::ReloadChannels));
+    }
+
+    #[test]
+    fn test_usage_footer_hot_reload() {
+        use openfang_types::config::UsageFooterMode;
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.usage_footer = UsageFooterMode::Off;
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.contains(&HotAction::UpdateUsageFooter));
+    }
+
+    #[test]
+    fn test_max_cron_jobs_hot_reload() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.max_cron_jobs = 1000;
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.contains(&HotAction::UpdateCronConfig));
+    }
+
+    #[test]
+    fn test_extensions_hot_reload() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.extensions.reconnect_max_attempts = 20;
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.contains(&HotAction::ReloadExtensions));
+    }
+
+    #[test]
+    fn test_provider_urls_hot_reload() {
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.provider_urls
+            .insert("ollama".to_string(), "http://10.0.0.5:11434/v1".to_string());
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.contains(&HotAction::ReloadProviderUrls));
+    }
+
+    // -----------------------------------------------------------------------
+    // Mixed changes
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_mixed_changes() {
+        use openfang_types::config::UsageFooterMode;
+        let a = default_cfg();
+        let mut b = default_cfg();
+        // Restart-required
+        b.api_listen = "0.0.0.0:9999".to_string();
+        // Hot-reloadable
+        b.usage_footer = UsageFooterMode::Tokens;
+        b.max_cron_jobs = 100;
+        // No-op
+        b.log_level = "debug".to_string();
+
+        let plan = build_reload_plan(&a, &b);
+        assert!(plan.restart_required);
+        assert!(plan.has_changes());
+        // Hot actions are still collected even if restart is required,
+        // so the caller knows what will need re-initialization after restart.
+        assert!(plan.hot_actions.contains(&HotAction::UpdateUsageFooter));
+        assert!(plan.hot_actions.contains(&HotAction::UpdateCronConfig));
+        assert!(plan.noop_changes.iter().any(|c| c.contains("log_level")));
+    }
+
+    // -----------------------------------------------------------------------
+    // No-op changes
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_noop_changes() {
+        use openfang_types::config::KernelMode;
+        let a = default_cfg();
+        let mut b = default_cfg();
+        b.log_level = "debug".to_string();
+        b.language = "de".to_string();
+        b.mode = KernelMode::Dev;
+
+        let plan = build_reload_plan(&a, &b);
+        assert!(!plan.restart_required);
+        assert!(plan.hot_actions.is_empty());
+        assert_eq!(plan.noop_changes.len(), 3);
+        assert!(plan.noop_changes.iter().any(|c| c.contains("log_level")));
+        assert!(plan.noop_changes.iter().any(|c| c.contains("language")));
+        assert!(plan.noop_changes.iter().any(|c| c.contains("mode")));
+    }
+
+    // -----------------------------------------------------------------------
+    // has_changes / is_hot_reloadable helpers
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_has_changes() {
+        // No changes
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![],
+            noop_changes: vec![],
+        };
+        assert!(!plan.has_changes());
+
+        // Only noop
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![],
+            noop_changes: vec!["log_level: info -> debug".to_string()],
+        };
+        assert!(plan.has_changes());
+
+        // Only hot
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![HotAction::UpdateCronConfig],
+            noop_changes: vec![],
+        };
+        assert!(plan.has_changes());
+
+        // Only restart
+        let plan = ReloadPlan {
+            restart_required: true,
+            restart_reasons: vec!["api_listen changed".to_string()],
+            hot_actions: vec![],
+            noop_changes: vec![],
+        };
+        assert!(plan.has_changes());
+    }
+
+    #[test]
+    fn test_is_hot_reloadable() {
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![HotAction::ReloadChannels],
+            noop_changes: vec![],
+        };
+        assert!(plan.is_hot_reloadable());
+
+        let plan = ReloadPlan {
+            restart_required: true,
+            restart_reasons: vec!["api_listen changed".to_string()],
+            hot_actions: vec![HotAction::ReloadChannels],
+            noop_changes: vec![],
+        };
+        assert!(!plan.is_hot_reloadable());
+    }
+
+    // -----------------------------------------------------------------------
+    // Validation tests
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_validate_config_for_reload_valid() {
+        let config = default_cfg();
+        assert!(validate_config_for_reload(&config).is_ok());
+    }
+
+    #[test]
+    fn test_validate_config_for_reload_invalid() {
+        // Empty api_listen
+        let mut config = default_cfg();
+        config.api_listen = String::new();
+        let err = validate_config_for_reload(&config).unwrap_err();
+        assert!(err.iter().any(|e| e.contains("api_listen")));
+
+        // Excessive max_cron_jobs
+        let mut config = default_cfg();
+        config.max_cron_jobs = 100_000;
+        let err = validate_config_for_reload(&config).unwrap_err();
+        assert!(err.iter().any(|e| e.contains("max_cron_jobs")));
+    }
+
+    #[test]
+    fn test_validate_network_enabled_no_secret() {
+        let mut config = default_cfg();
+        config.network_enabled = true;
+        config.network.shared_secret = String::new();
+        let err = validate_config_for_reload(&config).unwrap_err();
+        assert!(err.iter().any(|e| e.contains("shared_secret")));
+    }
+
+    // -----------------------------------------------------------------------
+    // should_apply_hot
+    // -----------------------------------------------------------------------
+
+    #[test]
+    fn test_should_apply_hot_off() {
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![HotAction::ReloadChannels],
+            noop_changes: vec![],
+        };
+        assert!(!should_apply_hot(ReloadMode::Off, &plan));
+    }
+
+    #[test]
+    fn test_should_apply_hot_restart_mode() {
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![HotAction::ReloadChannels],
+            noop_changes: vec![],
+        };
+        assert!(!should_apply_hot(ReloadMode::Restart, &plan));
+    }
+
+    #[test]
+    fn test_should_apply_hot_hybrid() {
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![HotAction::ReloadChannels],
+            noop_changes: vec![],
+        };
+        assert!(should_apply_hot(ReloadMode::Hybrid, &plan));
+        assert!(should_apply_hot(ReloadMode::Hot, &plan));
+    }
+
+    #[test]
+    fn test_should_apply_hot_empty() {
+        let plan = ReloadPlan {
+            restart_required: false,
+            restart_reasons: vec![],
+            hot_actions: vec![],
+            noop_changes: vec![],
+        };
+        assert!(!should_apply_hot(ReloadMode::Hybrid, &plan));
+    }
+}

crates/openfang-kernel/src/cron.rs

@@ -0,0 +1,790 @@
+//! Cron job scheduler engine for the OpenFang kernel.
+//!
+//! Manages scheduled jobs (recurring and one-shot) across all agents.
+//! This is separate from `scheduler.rs` which handles agent resource tracking.
+//!
+//! The scheduler stores jobs in a `DashMap` for concurrent access, persists
+//! them to a JSON file on disk, and exposes methods for the kernel tick loop
+//! to query due jobs and record outcomes.
+
+use chrono::{Duration, Utc};
+use dashmap::DashMap;
+use openfang_types::agent::AgentId;
+use openfang_types::error::{OpenFangError, OpenFangResult};
+use openfang_types::scheduler::{CronJob, CronJobId, CronSchedule};
+use serde::{Deserialize, Serialize};
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicUsize, Ordering};
+use tracing::{debug, info, warn};
+
+/// Maximum consecutive errors before a job is auto-disabled.
+const MAX_CONSECUTIVE_ERRORS: u32 = 5;
+
+// ---------------------------------------------------------------------------
+// JobMeta — extra runtime state not stored in CronJob itself
+// ---------------------------------------------------------------------------
+
+/// Runtime metadata for a cron job that extends the base `CronJob` type.
+///
+/// The `CronJob` struct in `openfang-types` is intentionally lean (no
+/// `one_shot`, `last_status`, or error tracking). The scheduler tracks
+/// these operational details separately.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct JobMeta {
+    /// The underlying job definition.
+    pub job: CronJob,
+    /// Whether this job should be removed after a single successful execution.
+    pub one_shot: bool,
+    /// Human-readable status of the last execution (e.g. `"ok"` or `"error: ..."`).
+    pub last_status: Option<String>,
+    /// Number of consecutive failed executions.
+    pub consecutive_errors: u32,
+}
+
+impl JobMeta {
+    /// Wrap a `CronJob` with default metadata.
+    pub fn new(job: CronJob, one_shot: bool) -> Self {
+        Self {
+            job,
+            one_shot,
+            last_status: None,
+            consecutive_errors: 0,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// CronScheduler
+// ---------------------------------------------------------------------------
+
+/// Cron job scheduler — manages scheduled jobs for all agents.
+///
+/// Thread-safe via `DashMap`. The kernel should call [`due_jobs`] on a
+/// regular interval (e.g. every 10-30 seconds) to discover jobs that need
+/// to fire, then call [`record_success`] or [`record_failure`] after
+/// execution completes.
+pub struct CronScheduler {
+    /// All tracked jobs, keyed by their unique ID.
+    jobs: DashMap<CronJobId, JobMeta>,
+    /// Path to the persistence file (`<home>/cron_jobs.json`).
+    persist_path: PathBuf,
+    /// Global cap on total jobs across all agents (atomic for hot-reload).
+    max_total_jobs: AtomicUsize,
+}
+
+impl CronScheduler {
+    /// Create a new scheduler.
+    ///
+    /// `home_dir` is the OpenFang data directory; jobs are persisted to
+    /// `<home_dir>/cron_jobs.json`. `max_total_jobs` caps the total number
+    /// of jobs across all agents.
+    pub fn new(home_dir: &Path, max_total_jobs: usize) -> Self {
+        Self {
+            jobs: DashMap::new(),
+            persist_path: home_dir.join("cron_jobs.json"),
+            max_total_jobs: AtomicUsize::new(max_total_jobs),
+        }
+    }
+
+    /// Update the max total jobs limit (for hot-reload).
+    pub fn set_max_total_jobs(&self, new_max: usize) {
+        self.max_total_jobs.store(new_max, Ordering::Relaxed);
+    }
+
+    // -- Persistence --------------------------------------------------------
+
+    /// Load persisted jobs from disk.
+    ///
+    /// Returns the number of jobs loaded. If the persistence file does not
+    /// exist, returns `Ok(0)` without error.
+    pub fn load(&self) -> OpenFangResult<usize> {
+        if !self.persist_path.exists() {
+            return Ok(0);
+        }
+        let data = std::fs::read_to_string(&self.persist_path)
+            .map_err(|e| OpenFangError::Internal(format!("Failed to read cron jobs: {e}")))?;
+        let metas: Vec<JobMeta> = serde_json::from_str(&data)
+            .map_err(|e| OpenFangError::Internal(format!("Failed to parse cron jobs: {e}")))?;
+        let count = metas.len();
+        for meta in metas {
+            self.jobs.insert(meta.job.id, meta);
+        }
+        info!(count, "Loaded cron jobs from disk");
+        Ok(count)
+    }
+
+    /// Persist all jobs to disk via atomic write (write to `.tmp`, then rename).
+    pub fn persist(&self) -> OpenFangResult<()> {
+        let metas: Vec<JobMeta> = self.jobs.iter().map(|r| r.value().clone()).collect();
+        let data = serde_json::to_string_pretty(&metas)
+            .map_err(|e| OpenFangError::Internal(format!("Failed to serialize cron jobs: {e}")))?;
+        let tmp_path = self.persist_path.with_extension("json.tmp");
+        std::fs::write(&tmp_path, data.as_bytes()).map_err(|e| {
+            OpenFangError::Internal(format!("Failed to write cron jobs temp file: {e}"))
+        })?;
+        std::fs::rename(&tmp_path, &self.persist_path).map_err(|e| {
+            OpenFangError::Internal(format!("Failed to rename cron jobs file: {e}"))
+        })?;
+        debug!(count = metas.len(), "Persisted cron jobs");
+        Ok(())
+    }
+
+    // -- CRUD ---------------------------------------------------------------
+
+    /// Add a new job. Validates fields, computes the initial `next_run`,
+    /// and inserts it into the scheduler.
+    ///
+    /// `one_shot` controls whether the job is removed after a single
+    /// successful execution.
+    pub fn add_job(&self, mut job: CronJob, one_shot: bool) -> OpenFangResult<CronJobId> {
+        // Global limit
+        let max_jobs = self.max_total_jobs.load(Ordering::Relaxed);
+        if self.jobs.len() >= max_jobs {
+            return Err(OpenFangError::Internal(format!(
+                "Global cron job limit reached ({})",
+                max_jobs
+            )));
+        }
+
+        // Per-agent count
+        let agent_count = self
+            .jobs
+            .iter()
+            .filter(|r| r.value().job.agent_id == job.agent_id)
+            .count();
+
+        // CronJob.validate returns Result<(), String>
+        job.validate(agent_count)
+            .map_err(OpenFangError::InvalidInput)?;
+
+        // Compute initial next_run
+        job.next_run = Some(compute_next_run(&job.schedule));
+
+        let id = job.id;
+        self.jobs.insert(id, JobMeta::new(job, one_shot));
+        Ok(id)
+    }
+
+    /// Remove a job by ID. Returns the removed `CronJob`.
+    pub fn remove_job(&self, id: CronJobId) -> OpenFangResult<CronJob> {
+        self.jobs
+            .remove(&id)
+            .map(|(_, meta)| meta.job)
+            .ok_or_else(|| OpenFangError::Internal(format!("Cron job {id} not found")))
+    }
+
+    /// Enable or disable a job. Re-enabling resets errors and recomputes
+    /// `next_run`.
+    pub fn set_enabled(&self, id: CronJobId, enabled: bool) -> OpenFangResult<()> {
+        match self.jobs.get_mut(&id) {
+            Some(mut meta) => {
+                meta.job.enabled = enabled;
+                if enabled {
+                    meta.consecutive_errors = 0;
+                    meta.job.next_run = Some(compute_next_run(&meta.job.schedule));
+                }
+                Ok(())
+            }
+            None => Err(OpenFangError::Internal(format!("Cron job {id} not found"))),
+        }
+    }
+
+    // -- Queries ------------------------------------------------------------
+
+    /// Get a single job by ID.
+    pub fn get_job(&self, id: CronJobId) -> Option<CronJob> {
+        self.jobs.get(&id).map(|r| r.value().job.clone())
+    }
+
+    /// Get the full metadata for a job (includes `one_shot`, `last_status`,
+    /// `consecutive_errors`).
+    pub fn get_meta(&self, id: CronJobId) -> Option<JobMeta> {
+        self.jobs.get(&id).map(|r| r.value().clone())
+    }
+
+    /// List all jobs for a specific agent.
+    pub fn list_jobs(&self, agent_id: AgentId) -> Vec<CronJob> {
+        self.jobs
+            .iter()
+            .filter(|r| r.value().job.agent_id == agent_id)
+            .map(|r| r.value().job.clone())
+            .collect()
+    }
+
+    /// List all jobs across all agents.
+    pub fn list_all_jobs(&self) -> Vec<CronJob> {
+        self.jobs.iter().map(|r| r.value().job.clone()).collect()
+    }
+
+    /// Total number of tracked jobs.
+    pub fn total_jobs(&self) -> usize {
+        self.jobs.len()
+    }
+
+    /// Return jobs whose `next_run` is at or before `now` and are enabled.
+    ///
+    /// **Important**: This also pre-advances each due job's `next_run` to the
+    /// next scheduled time. This prevents the same job from being returned as
+    /// "due" on subsequent tick iterations while it's still executing.
+    pub fn due_jobs(&self) -> Vec<CronJob> {
+        let now = Utc::now();
+        let mut due = Vec::new();
+        for mut entry in self.jobs.iter_mut() {
+            let meta = entry.value_mut();
+            if meta.job.enabled && meta.job.next_run.map(|t| t <= now).unwrap_or(false) {
+                due.push(meta.job.clone());
+                // Pre-advance next_run so the job won't fire again on the next
+                // tick while it's still executing. Use `now` as the base so the
+                // next fire time is computed strictly after the current moment.
+                meta.job.next_run = Some(compute_next_run_after(&meta.job.schedule, now));
+            }
+        }
+        due
+    }
+
+    // -- Outcome recording --------------------------------------------------
+
+    /// Record a successful execution for a job.
+    ///
+    /// Updates `last_run`, resets errors, and either removes the job (if
+    /// one-shot) or advances `next_run`.
+    pub fn record_success(&self, id: CronJobId) {
+        // We need to check one_shot first, then potentially remove.
+        let should_remove = {
+            if let Some(mut meta) = self.jobs.get_mut(&id) {
+                meta.job.last_run = Some(Utc::now());
+                meta.last_status = Some("ok".to_string());
+                meta.consecutive_errors = 0;
+                // one_shot jobs get removed; recurring jobs keep the next_run
+                // already pre-advanced by due_jobs() — no recompute needed.
+                meta.one_shot
+            } else {
+                return;
+            }
+        };
+        if should_remove {
+            self.jobs.remove(&id);
+        }
+    }
+
+    /// Record a failed execution for a job.
+    ///
+    /// Increments the consecutive error counter. If it reaches
+    /// [`MAX_CONSECUTIVE_ERRORS`], the job is automatically disabled.
+    pub fn record_failure(&self, id: CronJobId, error_msg: &str) {
+        if let Some(mut meta) = self.jobs.get_mut(&id) {
+            meta.job.last_run = Some(Utc::now());
+            meta.last_status = Some(format!(
+                "error: {}",
+                openfang_types::truncate_str(error_msg, 256)
+            ));
+            meta.consecutive_errors += 1;
+            if meta.consecutive_errors >= MAX_CONSECUTIVE_ERRORS {
+                warn!(
+                    job_id = %id,
+                    errors = meta.consecutive_errors,
+                    "Auto-disabling cron job after repeated failures"
+                );
+                meta.job.enabled = false;
+            } else {
+                meta.job.next_run =
+                    Some(compute_next_run_after(&meta.job.schedule, Utc::now()));
+            }
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// compute_next_run
+// ---------------------------------------------------------------------------
+
+/// Compute the next fire time for a schedule, based on `now`.
+///
+/// - `At { at }` — returns `at` directly.
+/// - `Every { every_secs }` — returns `now + every_secs`.
+/// - `Cron { expr, tz }` — parses the cron expression and computes the next
+///   matching time. Supports standard 5-field (`min hour dom month dow`) and
+///   6-field (`sec min hour dom month dow`) formats by converting to the
+///   7-field format required by the `cron` crate.
+pub fn compute_next_run(schedule: &CronSchedule) -> chrono::DateTime<Utc> {
+    compute_next_run_after(schedule, Utc::now())
+}
+
+/// Compute the next fire time for a schedule, strictly after `after`.
+///
+/// Uses `after + 1 second` as the base time so the `cron` crate's
+/// inclusive `.after()` always returns a strictly future time. Without
+/// this offset, calling `compute_next_run` right after a job fires can
+/// return the same minute (or even the same second), causing the
+/// scheduler to re-fire immediately.
+pub fn compute_next_run_after(
+    schedule: &CronSchedule,
+    after: chrono::DateTime<Utc>,
+) -> chrono::DateTime<Utc> {
+    match schedule {
+        CronSchedule::At { at } => *at,
+        CronSchedule::Every { every_secs } => after + Duration::seconds(*every_secs as i64),
+        CronSchedule::Cron { expr, tz: _ } => {
+            // Convert standard 5/6-field cron to 7-field for the `cron` crate.
+            // Standard 5-field: min hour dom month dow
+            // 6-field:          sec min hour dom month dow
+            // cron crate:       sec min hour dom month dow year
+            let trimmed = expr.trim();
+            let fields: Vec<&str> = trimmed.split_whitespace().collect();
+            let seven_field = match fields.len() {
+                5 => format!("0 {trimmed} *"),
+                6 => format!("{trimmed} *"),
+                _ => expr.clone(),
+            };
+
+            // Add 1 second so `.after()` (inclusive) skips the current second.
+            let base = after + Duration::seconds(1);
+
+            match seven_field.parse::<cron::Schedule>() {
+                Ok(sched) => sched
+                    .after(&base)
+                    .next()
+                    .unwrap_or_else(|| after + Duration::hours(1)),
+                Err(e) => {
+                    warn!("Failed to parse cron expression '{}': {}", expr, e);
+                    after + Duration::hours(1)
+                }
+            }
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use chrono::Duration;
+    use openfang_types::scheduler::{CronAction, CronDelivery};
+
+    /// Build a minimal valid `CronJob` with an `Every` schedule.
+    fn make_job(agent_id: AgentId) -> CronJob {
+        CronJob {
+            id: CronJobId::new(),
+            agent_id,
+            name: "test-job".into(),
+            enabled: true,
+            schedule: CronSchedule::Every { every_secs: 3600 },
+            action: CronAction::SystemEvent {
+                text: "ping".into(),
+            },
+            delivery: CronDelivery::None,
+            created_at: Utc::now(),
+            last_run: None,
+            next_run: None,
+        }
+    }
+
+    /// Create a scheduler backed by a temp directory.
+    fn make_scheduler(max_total: usize) -> (CronScheduler, tempfile::TempDir) {
+        let tmp = tempfile::tempdir().unwrap();
+        let sched = CronScheduler::new(tmp.path(), max_total);
+        (sched, tmp)
+    }
+
+    // -- test_add_job_and_list ----------------------------------------------
+
+    #[test]
+    fn test_add_job_and_list() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+
+        let id = sched.add_job(job, false).unwrap();
+
+        // Should appear in agent list
+        let jobs = sched.list_jobs(agent);
+        assert_eq!(jobs.len(), 1);
+        assert_eq!(jobs[0].id, id);
+        assert_eq!(jobs[0].name, "test-job");
+
+        // Should appear in global list
+        let all = sched.list_all_jobs();
+        assert_eq!(all.len(), 1);
+
+        // get_job should return it
+        let fetched = sched.get_job(id).unwrap();
+        assert_eq!(fetched.agent_id, agent);
+
+        // next_run should have been computed
+        assert!(fetched.next_run.is_some());
+        assert_eq!(sched.total_jobs(), 1);
+    }
+
+    // -- test_remove_job ----------------------------------------------------
+
+    #[test]
+    fn test_remove_job() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        let removed = sched.remove_job(id).unwrap();
+        assert_eq!(removed.name, "test-job");
+        assert_eq!(sched.total_jobs(), 0);
+
+        // Removing again should fail
+        assert!(sched.remove_job(id).is_err());
+    }
+
+    // -- test_add_job_global_limit ------------------------------------------
+
+    #[test]
+    fn test_add_job_global_limit() {
+        let (sched, _tmp) = make_scheduler(2);
+        let agent = AgentId::new();
+
+        let j1 = make_job(agent);
+        let j2 = make_job(agent);
+        let j3 = make_job(agent);
+
+        sched.add_job(j1, false).unwrap();
+        sched.add_job(j2, false).unwrap();
+
+        // Third should hit global limit
+        let err = sched.add_job(j3, false).unwrap_err();
+        let msg = err.to_string();
+        assert!(
+            msg.contains("limit"),
+            "Expected global limit error, got: {msg}"
+        );
+    }
+
+    // -- test_add_job_per_agent_limit ---------------------------------------
+
+    #[test]
+    fn test_add_job_per_agent_limit() {
+        // MAX_JOBS_PER_AGENT = 50 in openfang-types
+        let (sched, _tmp) = make_scheduler(1000);
+        let agent = AgentId::new();
+
+        for i in 0..50 {
+            let mut job = make_job(agent);
+            job.name = format!("job-{i}");
+            sched.add_job(job, false).unwrap();
+        }
+
+        // 51st should be rejected by validate()
+        let mut overflow = make_job(agent);
+        overflow.name = "overflow".into();
+        let err = sched.add_job(overflow, false).unwrap_err();
+        let msg = err.to_string();
+        assert!(
+            msg.contains("50"),
+            "Expected per-agent limit error, got: {msg}"
+        );
+    }
+
+    // -- test_record_success_removes_one_shot --------------------------------
+
+    #[test]
+    fn test_record_success_removes_one_shot() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+        let id = sched.add_job(job, true).unwrap(); // one_shot = true
+
+        assert_eq!(sched.total_jobs(), 1);
+
+        sched.record_success(id);
+
+        // One-shot job should have been removed
+        assert_eq!(sched.total_jobs(), 0);
+        assert!(sched.get_job(id).is_none());
+    }
+
+    #[test]
+    fn test_record_success_keeps_recurring() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+        let id = sched.add_job(job, false).unwrap(); // one_shot = false
+
+        sched.record_success(id);
+
+        // Recurring job should still be there
+        assert_eq!(sched.total_jobs(), 1);
+        let meta = sched.get_meta(id).unwrap();
+        assert_eq!(meta.last_status.as_deref(), Some("ok"));
+        assert_eq!(meta.consecutive_errors, 0);
+        assert!(meta.job.last_run.is_some());
+    }
+
+    // -- test_record_failure_auto_disable -----------------------------------
+
+    #[test]
+    fn test_record_failure_auto_disable() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        // Fail MAX_CONSECUTIVE_ERRORS - 1 times: should still be enabled
+        for i in 0..(MAX_CONSECUTIVE_ERRORS - 1) {
+            sched.record_failure(id, &format!("error {i}"));
+            let meta = sched.get_meta(id).unwrap();
+            assert!(
+                meta.job.enabled,
+                "Job should still be enabled after {} failures",
+                i + 1
+            );
+            assert_eq!(meta.consecutive_errors, i + 1);
+        }
+
+        // One more failure should auto-disable
+        sched.record_failure(id, "final error");
+        let meta = sched.get_meta(id).unwrap();
+        assert!(
+            !meta.job.enabled,
+            "Job should be auto-disabled after {MAX_CONSECUTIVE_ERRORS} failures"
+        );
+        assert_eq!(meta.consecutive_errors, MAX_CONSECUTIVE_ERRORS);
+        assert!(
+            meta.last_status.as_ref().unwrap().starts_with("error:"),
+            "last_status should record the error"
+        );
+    }
+
+    // -- test_due_jobs_only_enabled -----------------------------------------
+
+    #[test]
+    fn test_due_jobs_only_enabled() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+
+        // Job 1: enabled, next_run in the past
+        let mut j1 = make_job(agent);
+        j1.name = "enabled-due".into();
+        let id1 = sched.add_job(j1, false).unwrap();
+
+        // Job 2: disabled
+        let mut j2 = make_job(agent);
+        j2.name = "disabled-job".into();
+        let id2 = sched.add_job(j2, false).unwrap();
+        sched.set_enabled(id2, false).unwrap();
+
+        // Force job 1's next_run to the past
+        if let Some(mut meta) = sched.jobs.get_mut(&id1) {
+            meta.job.next_run = Some(Utc::now() - Duration::seconds(10));
+        }
+
+        // Force job 2's next_run to the past too (but it's disabled)
+        if let Some(mut meta) = sched.jobs.get_mut(&id2) {
+            meta.job.next_run = Some(Utc::now() - Duration::seconds(10));
+        }
+
+        let due = sched.due_jobs();
+        assert_eq!(due.len(), 1);
+        assert_eq!(due[0].name, "enabled-due");
+    }
+
+    #[test]
+    fn test_due_jobs_future_not_included() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+
+        let job = make_job(agent);
+        sched.add_job(job, false).unwrap();
+
+        // The job was just added with next_run = now + 3600s, so it should
+        // not be due yet.
+        let due = sched.due_jobs();
+        assert!(due.is_empty());
+    }
+
+    // -- test_set_enabled ---------------------------------------------------
+
+    #[test]
+    fn test_set_enabled() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+
+        let job = make_job(agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        // Disable
+        sched.set_enabled(id, false).unwrap();
+        let meta = sched.get_meta(id).unwrap();
+        assert!(!meta.job.enabled);
+
+        // Re-enable resets error count
+        sched.record_failure(id, "ignored because disabled");
+        // Actually the job is disabled so record_failure still updates it.
+        // Let's first re-enable to test reset.
+        sched.set_enabled(id, true).unwrap();
+        let meta = sched.get_meta(id).unwrap();
+        assert!(meta.job.enabled);
+        assert_eq!(meta.consecutive_errors, 0);
+        assert!(meta.job.next_run.is_some());
+
+        // Non-existent ID should fail
+        let fake_id = CronJobId::new();
+        assert!(sched.set_enabled(fake_id, true).is_err());
+    }
+
+    // -- test_persist_and_load ----------------------------------------------
+
+    #[test]
+    fn test_persist_and_load() {
+        let tmp = tempfile::tempdir().unwrap();
+        let agent = AgentId::new();
+
+        // Create scheduler, add jobs, persist
+        {
+            let sched = CronScheduler::new(tmp.path(), 100);
+            let mut j1 = make_job(agent);
+            j1.name = "persist-a".into();
+            let mut j2 = make_job(agent);
+            j2.name = "persist-b".into();
+
+            sched.add_job(j1, false).unwrap();
+            sched.add_job(j2, true).unwrap(); // one_shot
+
+            sched.persist().unwrap();
+        }
+
+        // Create a new scheduler and load from disk
+        {
+            let sched = CronScheduler::new(tmp.path(), 100);
+            let count = sched.load().unwrap();
+            assert_eq!(count, 2);
+            assert_eq!(sched.total_jobs(), 2);
+
+            let jobs = sched.list_jobs(agent);
+            assert_eq!(jobs.len(), 2);
+
+            let names: Vec<&str> = jobs.iter().map(|j| j.name.as_str()).collect();
+            assert!(names.contains(&"persist-a"));
+            assert!(names.contains(&"persist-b"));
+
+            // Verify one_shot flag was preserved
+            let b_id = jobs.iter().find(|j| j.name == "persist-b").unwrap().id;
+            let meta = sched.get_meta(b_id).unwrap();
+            assert!(meta.one_shot);
+        }
+    }
+
+    #[test]
+    fn test_load_no_file_returns_zero() {
+        let tmp = tempfile::tempdir().unwrap();
+        let sched = CronScheduler::new(tmp.path(), 100);
+        assert_eq!(sched.load().unwrap(), 0);
+    }
+
+    // -- compute_next_run ---------------------------------------------------
+
+    #[test]
+    fn test_compute_next_run_at() {
+        let target = Utc::now() + Duration::hours(2);
+        let schedule = CronSchedule::At { at: target };
+        let next = compute_next_run(&schedule);
+        assert_eq!(next, target);
+    }
+
+    #[test]
+    fn test_compute_next_run_every() {
+        let before = Utc::now();
+        let schedule = CronSchedule::Every { every_secs: 300 };
+        let next = compute_next_run(&schedule);
+        let after = Utc::now();
+
+        // Should be roughly now + 300s
+        assert!(next >= before + Duration::seconds(300));
+        assert!(next <= after + Duration::seconds(300));
+    }
+
+    #[test]
+    fn test_compute_next_run_cron_daily() {
+        let now = Utc::now();
+        let schedule = CronSchedule::Cron {
+            expr: "0 9 * * *".into(),
+            tz: None,
+        };
+        let next = compute_next_run(&schedule);
+
+        // Should be within the next 24 hours (next 09:00 UTC)
+        assert!(next > now);
+        assert!(next <= now + Duration::hours(24));
+        assert_eq!(next.format("%M").to_string(), "00");
+        assert_eq!(next.format("%H").to_string(), "09");
+    }
+
+    #[test]
+    fn test_compute_next_run_cron_with_dow() {
+        let now = Utc::now();
+        let schedule = CronSchedule::Cron {
+            expr: "30 14 * * 1-5".into(),
+            tz: None,
+        };
+        let next = compute_next_run(&schedule);
+
+        // Should be within the next 7 days and at 14:30
+        assert!(next > now);
+        assert!(next <= now + Duration::days(7));
+        assert_eq!(next.format("%H:%M").to_string(), "14:30");
+    }
+
+    #[test]
+    fn test_compute_next_run_cron_invalid_expr() {
+        let now = Utc::now();
+        let schedule = CronSchedule::Cron {
+            expr: "not a cron".into(),
+            tz: None,
+        };
+        let next = compute_next_run(&schedule);
+        // Invalid expression falls back to 1 hour from now
+        assert!(next > now + Duration::minutes(59));
+        assert!(next <= now + Duration::minutes(61));
+    }
+
+    // -- error message truncation in record_failure -------------------------
+
+    #[test]
+    fn test_compute_next_run_after_skips_current_second() {
+        // A "every 4 hours" cron: next_run should be >= 4 hours from now,
+        // not in the same minute (the bug from #55).
+        let schedule = CronSchedule::Cron {
+            expr: "0 */4 * * *".into(),
+            tz: None,
+        };
+        let now = Utc::now();
+        let next = compute_next_run_after(&schedule, now);
+        // Must be strictly after `now` and at least ~1 hour away
+        // (the closest 4-hourly boundary is at least minutes away).
+        assert!(next > now, "next_run should be strictly after now");
+        let diff = next - now;
+        assert!(
+            diff.num_minutes() >= 1,
+            "Expected next_run at least 1 min away, got {} seconds",
+            diff.num_seconds()
+        );
+    }
+
+    #[test]
+    fn test_record_failure_truncates_long_error() {
+        let (sched, _tmp) = make_scheduler(100);
+        let agent = AgentId::new();
+        let job = make_job(agent);
+        let id = sched.add_job(job, false).unwrap();
+
+        let long_error = "x".repeat(1000);
+        sched.record_failure(id, &long_error);
+
+        let meta = sched.get_meta(id).unwrap();
+        let status = meta.last_status.unwrap();
+        // "error: " is 7 chars + 256 chars of truncated message = 263 max
+        assert!(
+            status.len() <= 263,
+            "Status should be truncated, got {} chars",
+            status.len()
+        );
+    }
+}

crates/openfang-kernel/src/error.rs

@@ -0,0 +1,19 @@
+//! Kernel-specific error types.
+
+use openfang_types::error::OpenFangError;
+use thiserror::Error;
+
+/// Kernel error type wrapping OpenFangError with kernel-specific context.
+#[derive(Error, Debug)]
+pub enum KernelError {
+    /// A wrapped OpenFangError.
+    #[error(transparent)]
+    OpenFang(#[from] OpenFangError),
+
+    /// The kernel failed to boot.
+    #[error("Boot failed: {0}")]
+    BootFailed(String),
+}
+
+/// Alias for kernel results.
+pub type KernelResult<T> = Result<T, KernelError>;

crates/openfang-kernel/src/event_bus.rs

@@ -0,0 +1,149 @@
+//! Event bus — pub/sub with pattern matching and history ring buffer.
+
+use dashmap::DashMap;
+use openfang_types::agent::AgentId;
+use openfang_types::event::{Event, EventTarget};
+use std::collections::VecDeque;
+use std::sync::Arc;
+use tokio::sync::{broadcast, RwLock};
+use tracing::debug;
+
+/// Maximum events retained in the history ring buffer.
+const HISTORY_SIZE: usize = 1000;
+
+/// The central event bus for inter-agent and system communication.
+pub struct EventBus {
+    /// Broadcast channel for all events.
+    sender: broadcast::Sender<Event>,
+    /// Per-agent event channels.
+    agent_channels: DashMap<AgentId, broadcast::Sender<Event>>,
+    /// Event history ring buffer.
+    history: Arc<RwLock<VecDeque<Event>>>,
+}
+
+impl EventBus {
+    /// Create a new event bus.
+    pub fn new() -> Self {
+        let (sender, _) = broadcast::channel(1024);
+        Self {
+            sender,
+            agent_channels: DashMap::new(),
+            history: Arc::new(RwLock::new(VecDeque::with_capacity(HISTORY_SIZE))),
+        }
+    }
+
+    /// Publish an event to the bus.
+    pub async fn publish(&self, event: Event) {
+        debug!(
+            event_id = %event.id,
+            source = %event.source,
+            "Publishing event"
+        );
+
+        // Store in history
+        {
+            let mut history = self.history.write().await;
+            if history.len() >= HISTORY_SIZE {
+                history.pop_front();
+            }
+            history.push_back(event.clone());
+        }
+
+        // Route to target
+        match &event.target {
+            EventTarget::Agent(agent_id) => {
+                if let Some(sender) = self.agent_channels.get(agent_id) {
+                    let _ = sender.send(event.clone());
+                }
+            }
+            EventTarget::Broadcast => {
+                let _ = self.sender.send(event.clone());
+                for entry in self.agent_channels.iter() {
+                    let _ = entry.value().send(event.clone());
+                }
+            }
+            EventTarget::Pattern(_pattern) => {
+                // Phase 1: broadcast to all for pattern matching
+                let _ = self.sender.send(event.clone());
+            }
+            EventTarget::System => {
+                let _ = self.sender.send(event.clone());
+            }
+        }
+    }
+
+    /// Subscribe to events for a specific agent.
+    pub fn subscribe_agent(&self, agent_id: AgentId) -> broadcast::Receiver<Event> {
+        let entry = self.agent_channels.entry(agent_id).or_insert_with(|| {
+            let (tx, _) = broadcast::channel(256);
+            tx
+        });
+        entry.subscribe()
+    }
+
+    /// Subscribe to all broadcast/system events.
+    pub fn subscribe_all(&self) -> broadcast::Receiver<Event> {
+        self.sender.subscribe()
+    }
+
+    /// Get recent event history.
+    pub async fn history(&self, limit: usize) -> Vec<Event> {
+        let history = self.history.read().await;
+        history.iter().rev().take(limit).cloned().collect()
+    }
+
+    /// Remove an agent's channel when it's terminated.
+    pub fn unsubscribe_agent(&self, agent_id: AgentId) {
+        self.agent_channels.remove(&agent_id);
+    }
+}
+
+impl Default for EventBus {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openfang_types::event::{EventPayload, SystemEvent};
+
+    #[tokio::test]
+    async fn test_publish_and_history() {
+        let bus = EventBus::new();
+        let agent_id = AgentId::new();
+        let event = Event::new(
+            agent_id,
+            EventTarget::System,
+            EventPayload::System(SystemEvent::KernelStarted),
+        );
+        bus.publish(event).await;
+        let history = bus.history(10).await;
+        assert_eq!(history.len(), 1);
+    }
+
+    #[tokio::test]
+    async fn test_agent_subscribe() {
+        let bus = EventBus::new();
+        let agent_id = AgentId::new();
+        let mut rx = bus.subscribe_agent(agent_id);
+
+        let event = Event::new(
+            AgentId::new(),
+            EventTarget::Agent(agent_id),
+            EventPayload::System(SystemEvent::HealthCheck {
+                status: "ok".to_string(),
+            }),
+        );
+        bus.publish(event).await;
+
+        let received = rx.recv().await.unwrap();
+        match received.payload {
+            EventPayload::System(SystemEvent::HealthCheck { status }) => {
+                assert_eq!(status, "ok");
+            }
+            _ => panic!("Wrong payload"),
+        }
+    }
+}

crates/openfang-kernel/src/heartbeat.rs

@@ -0,0 +1,245 @@
+//! Heartbeat monitor — detects unresponsive agents for 24/7 autonomous operation.
+//!
+//! The heartbeat monitor runs as a background tokio task, periodically checking
+//! each running agent's `last_active` timestamp. If an agent hasn't been active
+//! for longer than 2x its heartbeat interval, a `HealthCheckFailed` event is
+//! published to the event bus.
+
+use crate::registry::AgentRegistry;
+use chrono::Utc;
+use openfang_types::agent::{AgentId, AgentState};
+use tracing::{debug, warn};
+
+/// Default heartbeat check interval (seconds).
+const DEFAULT_CHECK_INTERVAL_SECS: u64 = 30;
+
+/// Multiplier: agent is considered unresponsive if inactive for this many
+/// multiples of its heartbeat interval.
+const UNRESPONSIVE_MULTIPLIER: u64 = 2;
+
+/// Result of a heartbeat check.
+#[derive(Debug, Clone)]
+pub struct HeartbeatStatus {
+    /// Agent ID.
+    pub agent_id: AgentId,
+    /// Agent name.
+    pub name: String,
+    /// Seconds since last activity.
+    pub inactive_secs: i64,
+    /// Whether the agent is considered unresponsive.
+    pub unresponsive: bool,
+}
+
+/// Heartbeat monitor configuration.
+#[derive(Debug, Clone)]
+pub struct HeartbeatConfig {
+    /// How often to run the heartbeat check (seconds).
+    pub check_interval_secs: u64,
+    /// Default threshold for unresponsiveness (seconds).
+    /// Overridden per-agent by AutonomousConfig.heartbeat_interval_secs.
+    pub default_timeout_secs: u64,
+}
+
+impl Default for HeartbeatConfig {
+    fn default() -> Self {
+        Self {
+            check_interval_secs: DEFAULT_CHECK_INTERVAL_SECS,
+            default_timeout_secs: DEFAULT_CHECK_INTERVAL_SECS * UNRESPONSIVE_MULTIPLIER,
+        }
+    }
+}
+
+/// Check all running agents and return their heartbeat status.
+///
+/// This is a pure function — it doesn't start a background task.
+/// The caller (kernel) can run this periodically or in a background task.
+pub fn check_agents(registry: &AgentRegistry, config: &HeartbeatConfig) -> Vec<HeartbeatStatus> {
+    let now = Utc::now();
+    let mut statuses = Vec::new();
+
+    for entry_ref in registry.list() {
+        // Only check running agents
+        if entry_ref.state != AgentState::Running {
+            continue;
+        }
+
+        let inactive_secs = (now - entry_ref.last_active).num_seconds();
+
+        // Determine timeout: use agent's autonomous config if set, else default
+        let timeout_secs = entry_ref
+            .manifest
+            .autonomous
+            .as_ref()
+            .map(|a| a.heartbeat_interval_secs * UNRESPONSIVE_MULTIPLIER)
+            .unwrap_or(config.default_timeout_secs) as i64;
+
+        let unresponsive = inactive_secs > timeout_secs;
+
+        if unresponsive {
+            warn!(
+                agent = %entry_ref.name,
+                inactive_secs,
+                timeout_secs,
+                "Agent is unresponsive"
+            );
+        } else {
+            debug!(
+                agent = %entry_ref.name,
+                inactive_secs,
+                "Agent heartbeat OK"
+            );
+        }
+
+        statuses.push(HeartbeatStatus {
+            agent_id: entry_ref.id,
+            name: entry_ref.name.clone(),
+            inactive_secs,
+            unresponsive,
+        });
+    }
+
+    statuses
+}
+
+/// Check if an agent is currently within its quiet hours.
+///
+/// Quiet hours format: "HH:MM-HH:MM" (24-hour format, UTC).
+/// Returns true if the current time falls within the quiet period.
+pub fn is_quiet_hours(quiet_hours: &str) -> bool {
+    let parts: Vec<&str> = quiet_hours.split('-').collect();
+    if parts.len() != 2 {
+        return false;
+    }
+
+    let now = Utc::now();
+    let current_minutes = now.format("%H").to_string().parse::<u32>().unwrap_or(0) * 60
+        + now.format("%M").to_string().parse::<u32>().unwrap_or(0);
+
+    let parse_time = |s: &str| -> Option<u32> {
+        let hm: Vec<&str> = s.trim().split(':').collect();
+        if hm.len() != 2 {
+            return None;
+        }
+        let h = hm[0].parse::<u32>().ok()?;
+        let m = hm[1].parse::<u32>().ok()?;
+        if h > 23 || m > 59 {
+            return None;
+        }
+        Some(h * 60 + m)
+    };
+
+    let start = match parse_time(parts[0]) {
+        Some(v) => v,
+        None => return false,
+    };
+    let end = match parse_time(parts[1]) {
+        Some(v) => v,
+        None => return false,
+    };
+
+    if start <= end {
+        // Same-day range: e.g., 22:00-06:00 would be cross-midnight
+        // This is start <= current < end
+        current_minutes >= start && current_minutes < end
+    } else {
+        // Cross-midnight: e.g., 22:00-06:00
+        current_minutes >= start || current_minutes < end
+    }
+}
+
+/// Aggregate heartbeat summary.
+#[derive(Debug, Clone, Default)]
+pub struct HeartbeatSummary {
+    /// Total agents checked.
+    pub total_checked: usize,
+    /// Number of responsive agents.
+    pub responsive: usize,
+    /// Number of unresponsive agents.
+    pub unresponsive: usize,
+    /// Details of unresponsive agents.
+    pub unresponsive_agents: Vec<HeartbeatStatus>,
+}
+
+/// Produce a summary from heartbeat statuses.
+pub fn summarize(statuses: &[HeartbeatStatus]) -> HeartbeatSummary {
+    let unresponsive_agents: Vec<HeartbeatStatus> = statuses
+        .iter()
+        .filter(|s| s.unresponsive)
+        .cloned()
+        .collect();
+
+    HeartbeatSummary {
+        total_checked: statuses.len(),
+        responsive: statuses.len() - unresponsive_agents.len(),
+        unresponsive: unresponsive_agents.len(),
+        unresponsive_agents,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_quiet_hours_parsing() {
+        // We can't easily test time-dependent logic, but we can test format parsing
+        assert!(!is_quiet_hours("invalid"));
+        assert!(!is_quiet_hours(""));
+        assert!(!is_quiet_hours("25:00-06:00")); // Invalid hours handled gracefully
+    }
+
+    #[test]
+    fn test_quiet_hours_format_valid() {
+        // The function returns true/false based on current time
+        // We just verify it doesn't panic on valid input
+        let _ = is_quiet_hours("22:00-06:00");
+        let _ = is_quiet_hours("00:00-23:59");
+        let _ = is_quiet_hours("09:00-17:00");
+    }
+
+    #[test]
+    fn test_heartbeat_config_default() {
+        let config = HeartbeatConfig::default();
+        assert_eq!(config.check_interval_secs, 30);
+        assert_eq!(config.default_timeout_secs, 60);
+    }
+
+    #[test]
+    fn test_summarize_empty() {
+        let summary = summarize(&[]);
+        assert_eq!(summary.total_checked, 0);
+        assert_eq!(summary.responsive, 0);
+        assert_eq!(summary.unresponsive, 0);
+    }
+
+    #[test]
+    fn test_summarize_mixed() {
+        let statuses = vec![
+            HeartbeatStatus {
+                agent_id: AgentId::new(),
+                name: "agent-1".to_string(),
+                inactive_secs: 10,
+                unresponsive: false,
+            },
+            HeartbeatStatus {
+                agent_id: AgentId::new(),
+                name: "agent-2".to_string(),
+                inactive_secs: 120,
+                unresponsive: true,
+            },
+            HeartbeatStatus {
+                agent_id: AgentId::new(),
+                name: "agent-3".to_string(),
+                inactive_secs: 5,
+                unresponsive: false,
+            },
+        ];
+
+        let summary = summarize(&statuses);
+        assert_eq!(summary.total_checked, 3);
+        assert_eq!(summary.responsive, 2);
+        assert_eq!(summary.unresponsive, 1);
+        assert_eq!(summary.unresponsive_agents.len(), 1);
+        assert_eq!(summary.unresponsive_agents[0].name, "agent-2");
+    }
+}

crates/openfang-kernel/src/lib.rs

@@ -0,0 +1,29 @@
+//! Core kernel for the OpenFang Agent Operating System.
+//!
+//! The kernel manages agent lifecycles, memory, permissions, scheduling,
+//! and inter-agent communication.
+
+pub mod approval;
+pub mod auth;
+pub mod auto_reply;
+pub mod background;
+pub mod capabilities;
+pub mod config;
+pub mod config_reload;
+pub mod cron;
+pub mod error;
+pub mod event_bus;
+pub mod heartbeat;
+pub mod kernel;
+pub mod metering;
+pub mod pairing;
+pub mod registry;
+pub mod scheduler;
+pub mod supervisor;
+pub mod triggers;
+pub mod whatsapp_gateway;
+pub mod wizard;
+pub mod workflow;
+
+pub use kernel::DeliveryTracker;
+pub use kernel::OpenFangKernel;

crates/openfang-kernel/src/metering.rs

@@ -0,0 +1,753 @@
+//! Metering engine — tracks LLM cost and enforces spending quotas.
+
+use openfang_memory::usage::{ModelUsage, UsageRecord, UsageStore, UsageSummary};
+use openfang_types::agent::{AgentId, ResourceQuota};
+use openfang_types::error::{OpenFangError, OpenFangResult};
+use std::sync::Arc;
+
+/// The metering engine tracks usage cost and enforces quota limits.
+pub struct MeteringEngine {
+    /// Persistent usage store (SQLite-backed).
+    store: Arc<UsageStore>,
+}
+
+impl MeteringEngine {
+    /// Create a new metering engine with the given usage store.
+    pub fn new(store: Arc<UsageStore>) -> Self {
+        Self { store }
+    }
+
+    /// Record a usage event (persists to SQLite).
+    pub fn record(&self, record: &UsageRecord) -> OpenFangResult<()> {
+        self.store.record(record)
+    }
+
+    /// Check if an agent is within its spending quotas (hourly, daily, monthly).
+    /// Returns Ok(()) if under all quotas, or QuotaExceeded error if over any.
+    pub fn check_quota(&self, agent_id: AgentId, quota: &ResourceQuota) -> OpenFangResult<()> {
+        // Hourly check
+        if quota.max_cost_per_hour_usd > 0.0 {
+            let hourly_cost = self.store.query_hourly(agent_id)?;
+            if hourly_cost >= quota.max_cost_per_hour_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Agent {} exceeded hourly cost quota: ${:.4} / ${:.4}",
+                    agent_id, hourly_cost, quota.max_cost_per_hour_usd
+                )));
+            }
+        }
+
+        // Daily check
+        if quota.max_cost_per_day_usd > 0.0 {
+            let daily_cost = self.store.query_daily(agent_id)?;
+            if daily_cost >= quota.max_cost_per_day_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Agent {} exceeded daily cost quota: ${:.4} / ${:.4}",
+                    agent_id, daily_cost, quota.max_cost_per_day_usd
+                )));
+            }
+        }
+
+        // Monthly check
+        if quota.max_cost_per_month_usd > 0.0 {
+            let monthly_cost = self.store.query_monthly(agent_id)?;
+            if monthly_cost >= quota.max_cost_per_month_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Agent {} exceeded monthly cost quota: ${:.4} / ${:.4}",
+                    agent_id, monthly_cost, quota.max_cost_per_month_usd
+                )));
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Check global budget limits (across all agents).
+    pub fn check_global_budget(
+        &self,
+        budget: &openfang_types::config::BudgetConfig,
+    ) -> OpenFangResult<()> {
+        if budget.max_hourly_usd > 0.0 {
+            let cost = self.store.query_global_hourly()?;
+            if cost >= budget.max_hourly_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Global hourly budget exceeded: ${:.4} / ${:.4}",
+                    cost, budget.max_hourly_usd
+                )));
+            }
+        }
+
+        if budget.max_daily_usd > 0.0 {
+            let cost = self.store.query_today_cost()?;
+            if cost >= budget.max_daily_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Global daily budget exceeded: ${:.4} / ${:.4}",
+                    cost, budget.max_daily_usd
+                )));
+            }
+        }
+
+        if budget.max_monthly_usd > 0.0 {
+            let cost = self.store.query_global_monthly()?;
+            if cost >= budget.max_monthly_usd {
+                return Err(OpenFangError::QuotaExceeded(format!(
+                    "Global monthly budget exceeded: ${:.4} / ${:.4}",
+                    cost, budget.max_monthly_usd
+                )));
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Get budget status — current spend vs limits for all time windows.
+    pub fn budget_status(&self, budget: &openfang_types::config::BudgetConfig) -> BudgetStatus {
+        let hourly = self.store.query_global_hourly().unwrap_or(0.0);
+        let daily = self.store.query_today_cost().unwrap_or(0.0);
+        let monthly = self.store.query_global_monthly().unwrap_or(0.0);
+
+        BudgetStatus {
+            hourly_spend: hourly,
+            hourly_limit: budget.max_hourly_usd,
+            hourly_pct: if budget.max_hourly_usd > 0.0 {
+                hourly / budget.max_hourly_usd
+            } else {
+                0.0
+            },
+            daily_spend: daily,
+            daily_limit: budget.max_daily_usd,
+            daily_pct: if budget.max_daily_usd > 0.0 {
+                daily / budget.max_daily_usd
+            } else {
+                0.0
+            },
+            monthly_spend: monthly,
+            monthly_limit: budget.max_monthly_usd,
+            monthly_pct: if budget.max_monthly_usd > 0.0 {
+                monthly / budget.max_monthly_usd
+            } else {
+                0.0
+            },
+            alert_threshold: budget.alert_threshold,
+        }
+    }
+
+    /// Get a usage summary, optionally filtered by agent.
+    pub fn get_summary(&self, agent_id: Option<AgentId>) -> OpenFangResult<UsageSummary> {
+        self.store.query_summary(agent_id)
+    }
+
+    /// Get usage grouped by model.
+    pub fn get_by_model(&self) -> OpenFangResult<Vec<ModelUsage>> {
+        self.store.query_by_model()
+    }
+
+    /// Estimate the cost of an LLM call based on model and token counts.
+    ///
+    /// Pricing table (approximate, per million tokens):
+    ///
+    /// | Model Family          | Input $/M | Output $/M |
+    /// |-----------------------|-----------|------------|
+    /// | claude-haiku          |     0.25  |      1.25  |
+    /// | claude-sonnet-4-6     |     3.00  |     15.00  |
+    /// | claude-opus-4-6       |     5.00  |     25.00  |
+    /// | claude-opus (legacy)  |    15.00  |     75.00  |
+    /// | gpt-5.2(-pro)         |     1.75  |     14.00  |
+    /// | gpt-5(.1)             |     1.25  |     10.00  |
+    /// | gpt-5-mini            |     0.25  |      2.00  |
+    /// | gpt-5-nano            |     0.05  |      0.40  |
+    /// | gpt-4o                |     2.50  |     10.00  |
+    /// | gpt-4o-mini           |     0.15  |      0.60  |
+    /// | gpt-4.1               |     2.00  |      8.00  |
+    /// | gpt-4.1-mini          |     0.40  |      1.60  |
+    /// | gpt-4.1-nano          |     0.10  |      0.40  |
+    /// | o3-mini               |     1.10  |      4.40  |
+    /// | gemini-3.1            |     2.50  |     15.00  |
+    /// | gemini-3              |     0.50  |      3.00  |
+    /// | gemini-2.5-flash-lite |     0.04  |      0.15  |
+    /// | gemini-2.5-pro        |     1.25  |     10.00  |
+    /// | gemini-2.5-flash      |     0.15  |      0.60  |
+    /// | gemini-2.0-flash      |     0.10  |      0.40  |
+    /// | deepseek-chat/v3      |     0.27  |      1.10  |
+    /// | deepseek-reasoner/r1  |     0.55  |      2.19  |
+    /// | llama-4-maverick      |     0.50  |      0.77  |
+    /// | llama-4-scout         |     0.11  |      0.34  |
+    /// | llama/mixtral (groq)  |     0.05  |      0.10  |
+    /// | grok-4.1              |     0.20  |      0.50  |
+    /// | grok-4                |     3.00  |     15.00  |
+    /// | grok-3                |     3.00  |     15.00  |
+    /// | qwen                  |     0.20  |      0.60  |
+    /// | mistral-large         |     2.00  |      6.00  |
+    /// | mistral-small         |     0.10  |      0.30  |
+    /// | command-r-plus        |     2.50  |     10.00  |
+    /// | Default (unknown)     |     1.00  |      3.00  |
+    pub fn estimate_cost(model: &str, input_tokens: u64, output_tokens: u64) -> f64 {
+        let model_lower = model.to_lowercase();
+        let (input_per_m, output_per_m) = estimate_cost_rates(&model_lower);
+
+        let input_cost = (input_tokens as f64 / 1_000_000.0) * input_per_m;
+        let output_cost = (output_tokens as f64 / 1_000_000.0) * output_per_m;
+        input_cost + output_cost
+    }
+
+    /// Estimate cost using the model catalog as the pricing source.
+    ///
+    /// Falls back to the default rate ($1/$3 per million) if the model is not
+    /// found in the catalog.
+    pub fn estimate_cost_with_catalog(
+        catalog: &openfang_runtime::model_catalog::ModelCatalog,
+        model: &str,
+        input_tokens: u64,
+        output_tokens: u64,
+    ) -> f64 {
+        let (input_per_m, output_per_m) = catalog.pricing(model).unwrap_or((1.0, 3.0));
+        let input_cost = (input_tokens as f64 / 1_000_000.0) * input_per_m;
+        let output_cost = (output_tokens as f64 / 1_000_000.0) * output_per_m;
+        input_cost + output_cost
+    }
+
+    /// Clean up old usage records.
+    pub fn cleanup(&self, days: u32) -> OpenFangResult<usize> {
+        self.store.cleanup_old(days)
+    }
+}
+
+/// Budget status snapshot — current spend vs limits for all time windows.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct BudgetStatus {
+    pub hourly_spend: f64,
+    pub hourly_limit: f64,
+    pub hourly_pct: f64,
+    pub daily_spend: f64,
+    pub daily_limit: f64,
+    pub daily_pct: f64,
+    pub monthly_spend: f64,
+    pub monthly_limit: f64,
+    pub monthly_pct: f64,
+    pub alert_threshold: f64,
+}
+
+/// Returns (input_per_million, output_per_million) pricing for a model.
+///
+/// Order matters: more specific patterns must come before generic ones
+/// (e.g. "gpt-4o-mini" before "gpt-4o", "gpt-4.1-mini" before "gpt-4.1").
+fn estimate_cost_rates(model: &str) -> (f64, f64) {
+    // ── Anthropic ──────────────────────────────────────────────
+    if model.contains("haiku") {
+        return (0.25, 1.25);
+    }
+    if model.contains("opus-4-6") || model.contains("claude-opus-4-6") {
+        return (5.0, 25.0);
+    }
+    if model.contains("opus") {
+        return (15.0, 75.0);
+    }
+    if model.contains("sonnet-4-6") || model.contains("claude-sonnet-4-6") {
+        return (3.0, 15.0);
+    }
+    if model.contains("sonnet") {
+        return (3.0, 15.0);
+    }
+
+    // ── OpenAI ─────────────────────────────────────────────────
+    if model.contains("gpt-5.2-pro") {
+        return (1.75, 14.0);
+    }
+    if model.contains("gpt-5.2") {
+        return (1.75, 14.0);
+    }
+    if model.contains("gpt-5.1") {
+        return (1.25, 10.0);
+    }
+    if model.contains("gpt-5-nano") {
+        return (0.05, 0.40);
+    }
+    if model.contains("gpt-5-mini") {
+        return (0.25, 2.0);
+    }
+    if model.contains("gpt-5") {
+        return (1.25, 10.0);
+    }
+    if model.contains("gpt-4o-mini") {
+        return (0.15, 0.60);
+    }
+    if model.contains("gpt-4o") {
+        return (2.50, 10.0);
+    }
+    if model.contains("gpt-4.1-nano") {
+        return (0.10, 0.40);
+    }
+    if model.contains("gpt-4.1-mini") {
+        return (0.40, 1.60);
+    }
+    if model.contains("gpt-4.1") {
+        return (2.00, 8.00);
+    }
+    if model.contains("o4-mini") {
+        return (1.10, 4.40);
+    }
+    if model.contains("o3-mini") {
+        return (1.10, 4.40);
+    }
+    if model.contains("o3") {
+        return (2.00, 8.00);
+    }
+    // Generic gpt-4 fallback
+    if model.contains("gpt-4") {
+        return (2.50, 10.0);
+    }
+
+    // ── Google Gemini ──────────────────────────────────────────
+    if model.contains("gemini-3.1") {
+        return (2.50, 15.0);
+    }
+    if model.contains("gemini-3") {
+        return (0.50, 3.0);
+    }
+    if model.contains("gemini-2.5-flash-lite") {
+        return (0.04, 0.15);
+    }
+    if model.contains("gemini-2.5-pro") {
+        return (1.25, 10.0);
+    }
+    if model.contains("gemini-2.5-flash") {
+        return (0.15, 0.60);
+    }
+    if model.contains("gemini-2.0-flash") || model.contains("gemini-flash") {
+        return (0.10, 0.40);
+    }
+    // Generic gemini fallback
+    if model.contains("gemini") {
+        return (0.15, 0.60);
+    }
+
+    // ── DeepSeek ───────────────────────────────────────────────
+    if model.contains("deepseek-reasoner") || model.contains("deepseek-r1") {
+        return (0.55, 2.19);
+    }
+    if model.contains("deepseek") {
+        return (0.27, 1.10);
+    }
+
+    // ── Cerebras (ultra-fast, cheap) ── must come before llama ─
+    if model.contains("cerebras") {
+        return (0.06, 0.06);
+    }
+
+    // ── SambaNova ── must come before llama ──────────────────────
+    if model.contains("sambanova") {
+        return (0.06, 0.06);
+    }
+
+    // ── Replicate ── must come before llama ─────────────────────
+    if model.contains("replicate") {
+        return (0.40, 0.40);
+    }
+
+    // ── Open-source (Groq, Together, etc.) ─────────────────────
+    if model.contains("llama-4-maverick") {
+        return (0.50, 0.77);
+    }
+    if model.contains("llama-4-scout") {
+        return (0.11, 0.34);
+    }
+    if model.contains("llama") || model.contains("mixtral") {
+        return (0.05, 0.10);
+    }
+    // ── Qwen (Alibaba) ──────────────────────────────────────────
+    if model.contains("qwen-max") {
+        return (4.00, 12.00);
+    }
+    if model.contains("qwen-vl") {
+        return (1.50, 4.50);
+    }
+    if model.contains("qwen-plus") {
+        return (0.80, 2.00);
+    }
+    if model.contains("qwen-turbo") {
+        return (0.30, 0.60);
+    }
+    if model.contains("qwen") {
+        return (0.20, 0.60);
+    }
+
+    // ── MiniMax ──────────────────────────────────────────────────
+    if model.contains("minimax") {
+        return (1.00, 3.00);
+    }
+
+    // ── Zhipu / GLM ─────────────────────────────────────────────
+    if model.contains("glm-4-flash") {
+        return (0.10, 0.10);
+    }
+    if model.contains("glm") {
+        return (1.50, 5.00);
+    }
+    if model.contains("codegeex") {
+        return (0.10, 0.10);
+    }
+
+    // ── Moonshot / Kimi ─────────────────────────────────────────
+    if model.contains("moonshot") || model.contains("kimi") {
+        return (0.80, 0.80);
+    }
+
+    // ── Baidu ERNIE ─────────────────────────────────────────────
+    if model.contains("ernie") {
+        return (2.00, 6.00);
+    }
+
+    // ── AWS Bedrock ─────────────────────────────────────────────
+    if model.contains("nova-pro") {
+        return (0.80, 3.20);
+    }
+    if model.contains("nova-lite") {
+        return (0.06, 0.24);
+    }
+
+    // ── Mistral ────────────────────────────────────────────────
+    if model.contains("mistral-large") {
+        return (2.00, 6.00);
+    }
+    if model.contains("mistral-small") || model.contains("mistral") {
+        return (0.10, 0.30);
+    }
+
+    // ── Cohere ─────────────────────────────────────────────────
+    if model.contains("command-r-plus") {
+        return (2.50, 10.0);
+    }
+    if model.contains("command-r") {
+        return (0.15, 0.60);
+    }
+
+    // ── Perplexity ──────────────────────────────────────────────
+    if model.contains("sonar-pro") {
+        return (3.0, 15.0);
+    }
+    if model.contains("sonar") {
+        return (1.0, 5.0);
+    }
+
+    // ── xAI / Grok ──────────────────────────────────────────────
+    if model.contains("grok-4.1") {
+        return (0.20, 0.50);
+    }
+    if model.contains("grok-4") {
+        return (3.0, 15.0);
+    }
+    if model.contains("grok-3-mini") || model.contains("grok-2-mini") || model.contains("grok-mini")
+    {
+        return (0.30, 0.50);
+    }
+    if model.contains("grok-3") {
+        return (3.0, 15.0);
+    }
+    if model.contains("grok") {
+        return (2.0, 10.0);
+    }
+
+    // ── AI21 / Jamba ────────────────────────────────────────────
+    if model.contains("jamba") {
+        return (2.0, 8.0);
+    }
+
+    // ── Default (conservative) ─────────────────────────────────
+    (1.0, 3.0)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openfang_memory::MemorySubstrate;
+
+    fn setup() -> MeteringEngine {
+        let substrate = MemorySubstrate::open_in_memory(0.1).unwrap();
+        let store = Arc::new(UsageStore::new(substrate.usage_conn()));
+        MeteringEngine::new(store)
+    }
+
+    #[test]
+    fn test_record_and_check_quota_under() {
+        let engine = setup();
+        let agent_id = AgentId::new();
+        let quota = ResourceQuota {
+            max_cost_per_hour_usd: 1.0,
+            ..Default::default()
+        };
+
+        engine
+            .record(&UsageRecord {
+                agent_id,
+                model: "claude-haiku".to_string(),
+                input_tokens: 100,
+                output_tokens: 50,
+                cost_usd: 0.001,
+                tool_calls: 0,
+            })
+            .unwrap();
+
+        assert!(engine.check_quota(agent_id, &quota).is_ok());
+    }
+
+    #[test]
+    fn test_check_quota_exceeded() {
+        let engine = setup();
+        let agent_id = AgentId::new();
+        let quota = ResourceQuota {
+            max_cost_per_hour_usd: 0.01,
+            ..Default::default()
+        };
+
+        engine
+            .record(&UsageRecord {
+                agent_id,
+                model: "claude-sonnet".to_string(),
+                input_tokens: 10000,
+                output_tokens: 5000,
+                cost_usd: 0.05,
+                tool_calls: 0,
+            })
+            .unwrap();
+
+        let result = engine.check_quota(agent_id, &quota);
+        assert!(result.is_err());
+        let err = result.unwrap_err().to_string();
+        assert!(err.contains("exceeded hourly cost quota"));
+    }
+
+    #[test]
+    fn test_check_quota_zero_limit_skipped() {
+        let engine = setup();
+        let agent_id = AgentId::new();
+        let quota = ResourceQuota {
+            max_cost_per_hour_usd: 0.0,
+            ..Default::default()
+        };
+
+        // Even with high usage, a zero limit means no enforcement
+        engine
+            .record(&UsageRecord {
+                agent_id,
+                model: "claude-opus".to_string(),
+                input_tokens: 100000,
+                output_tokens: 50000,
+                cost_usd: 100.0,
+                tool_calls: 0,
+            })
+            .unwrap();
+
+        assert!(engine.check_quota(agent_id, &quota).is_ok());
+    }
+
+    #[test]
+    fn test_estimate_cost_haiku() {
+        let cost = MeteringEngine::estimate_cost("claude-haiku-4-5-20251001", 1_000_000, 1_000_000);
+        assert!((cost - 1.50).abs() < 0.01); // $0.25 + $1.25
+    }
+
+    #[test]
+    fn test_estimate_cost_sonnet() {
+        let cost = MeteringEngine::estimate_cost("claude-sonnet-4-20250514", 1_000_000, 1_000_000);
+        assert!((cost - 18.0).abs() < 0.01); // $3.00 + $15.00
+    }
+
+    #[test]
+    fn test_estimate_cost_opus() {
+        let cost = MeteringEngine::estimate_cost("claude-opus-4-20250514", 1_000_000, 1_000_000);
+        assert!((cost - 90.0).abs() < 0.01); // $15.00 + $75.00
+    }
+
+    #[test]
+    fn test_estimate_cost_gpt4o() {
+        let cost = MeteringEngine::estimate_cost("gpt-4o-2024-11-20", 1_000_000, 1_000_000);
+        assert!((cost - 12.50).abs() < 0.01); // $2.50 + $10.00
+    }
+
+    #[test]
+    fn test_estimate_cost_gpt4o_mini() {
+        let cost = MeteringEngine::estimate_cost("gpt-4o-mini", 1_000_000, 1_000_000);
+        assert!((cost - 0.75).abs() < 0.01); // $0.15 + $0.60
+    }
+
+    #[test]
+    fn test_estimate_cost_gpt41() {
+        let cost = MeteringEngine::estimate_cost("gpt-4.1", 1_000_000, 1_000_000);
+        assert!((cost - 10.0).abs() < 0.01); // $2.00 + $8.00
+    }
+
+    #[test]
+    fn test_estimate_cost_gpt41_mini() {
+        let cost = MeteringEngine::estimate_cost("gpt-4.1-mini", 1_000_000, 1_000_000);
+        assert!((cost - 2.0).abs() < 0.01); // $0.40 + $1.60
+    }
+
+    #[test]
+    fn test_estimate_cost_gpt41_nano() {
+        let cost = MeteringEngine::estimate_cost("gpt-4.1-nano", 1_000_000, 1_000_000);
+        assert!((cost - 0.50).abs() < 0.01); // $0.10 + $0.40
+    }
+
+    #[test]
+    fn test_estimate_cost_o3_mini() {
+        let cost = MeteringEngine::estimate_cost("o3-mini", 1_000_000, 1_000_000);
+        assert!((cost - 5.50).abs() < 0.01); // $1.10 + $4.40
+    }
+
+    #[test]
+    fn test_estimate_cost_gemini_20_flash() {
+        let cost = MeteringEngine::estimate_cost("gemini-2.0-flash", 1_000_000, 1_000_000);
+        assert!((cost - 0.50).abs() < 0.01); // $0.10 + $0.40
+    }
+
+    #[test]
+    fn test_estimate_cost_gemini_25_pro() {
+        let cost = MeteringEngine::estimate_cost("gemini-2.5-pro", 1_000_000, 1_000_000);
+        assert!((cost - 11.25).abs() < 0.01); // $1.25 + $10.00
+    }
+
+    #[test]
+    fn test_estimate_cost_gemini_25_flash() {
+        let cost = MeteringEngine::estimate_cost("gemini-2.5-flash", 1_000_000, 1_000_000);
+        assert!((cost - 0.75).abs() < 0.01); // $0.15 + $0.60
+    }
+
+    #[test]
+    fn test_estimate_cost_deepseek_chat() {
+        let cost = MeteringEngine::estimate_cost("deepseek-chat", 1_000_000, 1_000_000);
+        assert!((cost - 1.37).abs() < 0.01); // $0.27 + $1.10
+    }
+
+    #[test]
+    fn test_estimate_cost_deepseek_reasoner() {
+        let cost = MeteringEngine::estimate_cost("deepseek-reasoner", 1_000_000, 1_000_000);
+        assert!((cost - 2.74).abs() < 0.01); // $0.55 + $2.19
+    }
+
+    #[test]
+    fn test_estimate_cost_llama() {
+        let cost = MeteringEngine::estimate_cost("llama-3.3-70b-versatile", 1_000_000, 1_000_000);
+        assert!((cost - 0.15).abs() < 0.01); // $0.05 + $0.10
+    }
+
+    #[test]
+    fn test_estimate_cost_mixtral() {
+        let cost = MeteringEngine::estimate_cost("mixtral-8x7b", 1_000_000, 1_000_000);
+        assert!((cost - 0.15).abs() < 0.01); // $0.05 + $0.10
+    }
+
+    #[test]
+    fn test_estimate_cost_qwen() {
+        let cost = MeteringEngine::estimate_cost("qwen-2.5-72b", 1_000_000, 1_000_000);
+        assert!((cost - 0.80).abs() < 0.01); // $0.20 + $0.60
+    }
+
+    #[test]
+    fn test_estimate_cost_mistral_large() {
+        let cost = MeteringEngine::estimate_cost("mistral-large-latest", 1_000_000, 1_000_000);
+        assert!((cost - 8.0).abs() < 0.01); // $2.00 + $6.00
+    }
+
+    #[test]
+    fn test_estimate_cost_mistral_small() {
+        let cost = MeteringEngine::estimate_cost("mistral-small-latest", 1_000_000, 1_000_000);
+        assert!((cost - 0.40).abs() < 0.01); // $0.10 + $0.30
+    }
+
+    #[test]
+    fn test_estimate_cost_command_r_plus() {
+        let cost = MeteringEngine::estimate_cost("command-r-plus", 1_000_000, 1_000_000);
+        assert!((cost - 12.50).abs() < 0.01); // $2.50 + $10.00
+    }
+
+    #[test]
+    fn test_estimate_cost_unknown() {
+        let cost = MeteringEngine::estimate_cost("my-custom-model", 1_000_000, 1_000_000);
+        assert!((cost - 4.0).abs() < 0.01); // $1.00 + $3.00
+    }
+
+    #[test]
+    fn test_estimate_cost_grok() {
+        let cost = MeteringEngine::estimate_cost("grok-2", 1_000_000, 1_000_000);
+        assert!((cost - 12.0).abs() < 0.01); // $2.00 + $10.00
+    }
+
+    #[test]
+    fn test_estimate_cost_grok_mini() {
+        let cost = MeteringEngine::estimate_cost("grok-2-mini", 1_000_000, 1_000_000);
+        assert!((cost - 0.80).abs() < 0.01); // $0.30 + $0.50
+    }
+
+    #[test]
+    fn test_estimate_cost_sonar_pro() {
+        let cost = MeteringEngine::estimate_cost("sonar-pro", 1_000_000, 1_000_000);
+        assert!((cost - 18.0).abs() < 0.01); // $3.00 + $15.00
+    }
+
+    #[test]
+    fn test_estimate_cost_jamba() {
+        let cost = MeteringEngine::estimate_cost("jamba-1.5-large", 1_000_000, 1_000_000);
+        assert!((cost - 10.0).abs() < 0.01); // $2.00 + $8.00
+    }
+
+    #[test]
+    fn test_estimate_cost_cerebras() {
+        let cost = MeteringEngine::estimate_cost("cerebras/llama3.3-70b", 1_000_000, 1_000_000);
+        assert!((cost - 0.12).abs() < 0.01); // $0.06 + $0.06
+    }
+
+    #[test]
+    fn test_estimate_cost_with_catalog() {
+        let catalog = openfang_runtime::model_catalog::ModelCatalog::new();
+        // Sonnet: $3/M input, $15/M output
+        let cost = MeteringEngine::estimate_cost_with_catalog(
+            &catalog,
+            "claude-sonnet-4-20250514",
+            1_000_000,
+            1_000_000,
+        );
+        assert!((cost - 18.0).abs() < 0.01);
+    }
+
+    #[test]
+    fn test_estimate_cost_with_catalog_alias() {
+        let catalog = openfang_runtime::model_catalog::ModelCatalog::new();
+        // "sonnet" alias should resolve to same pricing
+        let cost =
+            MeteringEngine::estimate_cost_with_catalog(&catalog, "sonnet", 1_000_000, 1_000_000);
+        assert!((cost - 18.0).abs() < 0.01);
+    }
+
+    #[test]
+    fn test_estimate_cost_with_catalog_unknown_uses_default() {
+        let catalog = openfang_runtime::model_catalog::ModelCatalog::new();
+        // Unknown model falls back to $1/$3
+        let cost = MeteringEngine::estimate_cost_with_catalog(
+            &catalog,
+            "totally-unknown-model",
+            1_000_000,
+            1_000_000,
+        );
+        assert!((cost - 4.0).abs() < 0.01);
+    }
+
+    #[test]
+    fn test_get_summary() {
+        let engine = setup();
+        let agent_id = AgentId::new();
+
+        engine
+            .record(&UsageRecord {
+                agent_id,
+                model: "haiku".to_string(),
+                input_tokens: 500,
+                output_tokens: 200,
+                cost_usd: 0.005,
+                tool_calls: 3,
+            })
+            .unwrap();
+
+        let summary = engine.get_summary(Some(agent_id)).unwrap();
+        assert_eq!(summary.call_count, 1);
+        assert_eq!(summary.total_input_tokens, 500);
+    }
+}

crates/openfang-kernel/src/pairing.rs

@@ -0,0 +1,510 @@
+//! Device pairing — QR-code flow for mobile/desktop clients.
+//!
+//! Supports pairing via short-lived tokens, device management, and
+//! push notifications via ntfy.sh or gotify.
+
+use dashmap::DashMap;
+use openfang_types::config::PairingConfig;
+use serde::{Deserialize, Serialize};
+use tracing::{debug, warn};
+
+/// Maximum concurrent pairing requests (prevent token flooding).
+const MAX_PENDING_REQUESTS: usize = 5;
+
+/// A paired device record.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct PairedDevice {
+    pub device_id: String,
+    pub display_name: String,
+    pub platform: String,
+    pub paired_at: chrono::DateTime<chrono::Utc>,
+    pub last_seen: chrono::DateTime<chrono::Utc>,
+    #[serde(skip_serializing)]
+    pub push_token: Option<String>,
+}
+
+/// Pairing request (short-lived, for QR code flow).
+#[derive(Debug, Clone)]
+pub struct PairingRequest {
+    pub token: String,
+    pub created_at: chrono::DateTime<chrono::Utc>,
+    pub expires_at: chrono::DateTime<chrono::Utc>,
+}
+
+/// Persistence callback — kernel injects this so PairingManager can save without
+/// taking a direct dependency on openfang-memory.
+pub type PersistFn = Box<dyn Fn(&PairedDevice, PersistOp) + Send + Sync>;
+
+/// Persistence operation kind.
+#[derive(Debug, Clone, Copy)]
+pub enum PersistOp {
+    Save,
+    Remove,
+}
+
+/// Device pairing manager.
+pub struct PairingManager {
+    config: PairingConfig,
+    pending: DashMap<String, PairingRequest>,
+    devices: DashMap<String, PairedDevice>,
+    persist: Option<PersistFn>,
+}
+
+impl PairingManager {
+    pub fn new(config: PairingConfig) -> Self {
+        Self {
+            config,
+            pending: DashMap::new(),
+            devices: DashMap::new(),
+            persist: None,
+        }
+    }
+
+    /// Attach a persistence callback (called after pair/unpair operations).
+    pub fn set_persist(&mut self, f: PersistFn) {
+        self.persist = Some(f);
+    }
+
+    /// Bulk-load devices from persistence (call once at boot).
+    pub fn load_devices(&self, devices: Vec<PairedDevice>) {
+        for d in devices {
+            self.devices.insert(d.device_id.clone(), d);
+        }
+        debug!(
+            count = self.devices.len(),
+            "Loaded paired devices from database"
+        );
+    }
+
+    /// Generate a new pairing request. Returns token for QR encoding.
+    pub fn create_pairing_request(&self) -> Result<PairingRequest, String> {
+        if !self.config.enabled {
+            return Err("Device pairing is disabled".into());
+        }
+
+        // Enforce max pending limit
+        if self.pending.len() >= MAX_PENDING_REQUESTS {
+            // Clean expired first
+            self.clean_expired();
+            if self.pending.len() >= MAX_PENDING_REQUESTS {
+                return Err("Too many pending pairing requests. Try again later.".into());
+            }
+        }
+
+        // Generate secure random token (32 bytes = 64 hex chars)
+        let mut token_bytes = [0u8; 32];
+        use rand::RngCore;
+        rand::thread_rng().fill_bytes(&mut token_bytes);
+        let token = hex::encode(token_bytes);
+
+        let now = chrono::Utc::now();
+        let expires_at = now + chrono::Duration::seconds(self.config.token_expiry_secs as i64);
+
+        let request = PairingRequest {
+            token: token.clone(),
+            created_at: now,
+            expires_at,
+        };
+
+        self.pending.insert(token, request.clone());
+
+        Ok(request)
+    }
+
+    /// Complete pairing — device submits token + device info.
+    pub fn complete_pairing(
+        &self,
+        token: &str,
+        device_info: PairedDevice,
+    ) -> Result<PairedDevice, String> {
+        // SECURITY: Constant-time token comparison
+        let found = self.pending.iter().find(|entry| {
+            use subtle::ConstantTimeEq;
+            let stored = entry.value().token.as_bytes();
+            let provided = token.as_bytes();
+            if stored.len() != provided.len() {
+                return false;
+            }
+            stored.ct_eq(provided).into()
+        });
+
+        let entry = found.ok_or("Invalid or expired pairing token")?;
+        let request = entry.value().clone();
+        let key = entry.key().clone();
+        drop(entry);
+
+        // Check expiry
+        if chrono::Utc::now() > request.expires_at {
+            self.pending.remove(&key);
+            return Err("Pairing token has expired".into());
+        }
+
+        // Check max devices
+        if self.devices.len() >= self.config.max_devices {
+            return Err(format!(
+                "Maximum paired devices ({}) reached. Remove a device first.",
+                self.config.max_devices
+            ));
+        }
+
+        // Remove the used token
+        self.pending.remove(&key);
+
+        // Store the device
+        let device_id = device_info.device_id.clone();
+        self.devices.insert(device_id.clone(), device_info.clone());
+
+        // Persist to database
+        if let Some(ref persist) = self.persist {
+            persist(&device_info, PersistOp::Save);
+        }
+
+        debug!(device_id = %device_id, "Device paired successfully");
+
+        Ok(device_info)
+    }
+
+    /// List paired devices.
+    pub fn list_devices(&self) -> Vec<PairedDevice> {
+        self.devices.iter().map(|e| e.value().clone()).collect()
+    }
+
+    /// Remove a paired device.
+    pub fn remove_device(&self, device_id: &str) -> Result<(), String> {
+        let removed = self
+            .devices
+            .remove(device_id)
+            .ok_or_else(|| format!("Device '{device_id}' not found"))?;
+
+        // Persist removal to database
+        if let Some(ref persist) = self.persist {
+            persist(&removed.1, PersistOp::Remove);
+        }
+
+        Ok(())
+    }
+
+    /// Send push notification to all paired devices.
+    pub async fn notify_devices(
+        &self,
+        title: &str,
+        body: &str,
+    ) -> Vec<(String, Result<(), String>)> {
+        let mut results = Vec::new();
+
+        match self.config.push_provider.as_str() {
+            "ntfy" => {
+                let url = self.config.ntfy_url.as_deref().unwrap_or("https://ntfy.sh");
+                let topic = match &self.config.ntfy_topic {
+                    Some(t) => t.clone(),
+                    None => {
+                        results.push(("ntfy".to_string(), Err("ntfy_topic not configured".into())));
+                        return results;
+                    }
+                };
+
+                let full_url = format!("{}/{}", url.trim_end_matches('/'), topic);
+
+                let client = reqwest::Client::new();
+                match client
+                    .post(&full_url)
+                    .header("Title", title)
+                    .body(body.to_string())
+                    .timeout(std::time::Duration::from_secs(10))
+                    .send()
+                    .await
+                {
+                    Ok(resp) if resp.status().is_success() => {
+                        for device in self.devices.iter() {
+                            results.push((device.device_id.clone(), Ok(())));
+                        }
+                    }
+                    Ok(resp) => {
+                        let status = resp.status();
+                        results.push((
+                            "ntfy".to_string(),
+                            Err(format!("ntfy returned HTTP {status}")),
+                        ));
+                    }
+                    Err(e) => {
+                        results
+                            .push(("ntfy".to_string(), Err(format!("ntfy request failed: {e}"))));
+                    }
+                }
+            }
+            "gotify" => {
+                // Gotify requires an app token
+                let app_token = match std::env::var("GOTIFY_APP_TOKEN") {
+                    Ok(t) => t,
+                    Err(_) => {
+                        results
+                            .push(("gotify".to_string(), Err("GOTIFY_APP_TOKEN not set".into())));
+                        return results;
+                    }
+                };
+
+                let server_url = match std::env::var("GOTIFY_SERVER_URL") {
+                    Ok(u) => u,
+                    Err(_) => {
+                        results.push((
+                            "gotify".to_string(),
+                            Err("GOTIFY_SERVER_URL not set".into()),
+                        ));
+                        return results;
+                    }
+                };
+
+                let url = format!("{}/message", server_url.trim_end_matches('/'));
+                let body_json = serde_json::json!({
+                    "title": title,
+                    "message": body,
+                    "priority": 5,
+                });
+
+                let client = reqwest::Client::new();
+                match client
+                    .post(&url)
+                    .header("X-Gotify-Key", &app_token)
+                    .json(&body_json)
+                    .timeout(std::time::Duration::from_secs(10))
+                    .send()
+                    .await
+                {
+                    Ok(resp) if resp.status().is_success() => {
+                        for device in self.devices.iter() {
+                            results.push((device.device_id.clone(), Ok(())));
+                        }
+                    }
+                    Ok(resp) => {
+                        let status = resp.status();
+                        results.push((
+                            "gotify".to_string(),
+                            Err(format!("gotify returned HTTP {status}")),
+                        ));
+                    }
+                    Err(e) => {
+                        results.push((
+                            "gotify".to_string(),
+                            Err(format!("gotify request failed: {e}")),
+                        ));
+                    }
+                }
+            }
+            "none" | "" => {
+                // No push provider configured — silent
+            }
+            other => {
+                warn!(provider = other, "Unknown push notification provider");
+            }
+        }
+
+        results
+    }
+
+    /// Clean expired pairing requests.
+    pub fn clean_expired(&self) {
+        let now = chrono::Utc::now();
+        self.pending.retain(|_, req| req.expires_at > now);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn default_config() -> PairingConfig {
+        PairingConfig::default()
+    }
+
+    fn enabled_config() -> PairingConfig {
+        PairingConfig {
+            enabled: true,
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn test_manager_creation() {
+        let mgr = PairingManager::new(default_config());
+        assert!(mgr.devices.is_empty());
+        assert!(mgr.pending.is_empty());
+    }
+
+    #[test]
+    fn test_create_request_disabled() {
+        let mgr = PairingManager::new(default_config());
+        let result = mgr.create_pairing_request();
+        assert!(result.is_err());
+        assert!(result.unwrap_err().contains("disabled"));
+    }
+
+    #[test]
+    fn test_create_request_success() {
+        let mgr = PairingManager::new(enabled_config());
+        let req = mgr.create_pairing_request().unwrap();
+        assert_eq!(req.token.len(), 64); // 32 bytes = 64 hex chars
+        assert!(req.expires_at > req.created_at);
+    }
+
+    #[test]
+    fn test_max_pending_requests() {
+        let mgr = PairingManager::new(enabled_config());
+        for _ in 0..MAX_PENDING_REQUESTS {
+            mgr.create_pairing_request().unwrap();
+        }
+        let result = mgr.create_pairing_request();
+        assert!(result.is_err());
+        assert!(result.unwrap_err().contains("Too many"));
+    }
+
+    #[test]
+    fn test_complete_pairing_invalid_token() {
+        let mgr = PairingManager::new(enabled_config());
+        let device = PairedDevice {
+            device_id: "dev-1".to_string(),
+            display_name: "My Phone".to_string(),
+            platform: "android".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+        let result = mgr.complete_pairing("invalid-token", device);
+        assert!(result.is_err());
+        assert!(result.unwrap_err().contains("Invalid"));
+    }
+
+    #[test]
+    fn test_complete_pairing_success() {
+        let mgr = PairingManager::new(enabled_config());
+        let req = mgr.create_pairing_request().unwrap();
+
+        let device = PairedDevice {
+            device_id: "dev-1".to_string(),
+            display_name: "My Phone".to_string(),
+            platform: "android".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+
+        let result = mgr.complete_pairing(&req.token, device);
+        assert!(result.is_ok());
+        assert_eq!(mgr.devices.len(), 1);
+        assert!(mgr.pending.is_empty()); // Token consumed
+    }
+
+    #[test]
+    fn test_max_devices_enforced() {
+        let config = PairingConfig {
+            enabled: true,
+            max_devices: 1,
+            ..Default::default()
+        };
+        let mgr = PairingManager::new(config);
+
+        // Pair first device
+        let req1 = mgr.create_pairing_request().unwrap();
+        let d1 = PairedDevice {
+            device_id: "dev-1".to_string(),
+            display_name: "Phone 1".to_string(),
+            platform: "ios".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+        mgr.complete_pairing(&req1.token, d1).unwrap();
+
+        // Try second device
+        let req2 = mgr.create_pairing_request().unwrap();
+        let d2 = PairedDevice {
+            device_id: "dev-2".to_string(),
+            display_name: "Phone 2".to_string(),
+            platform: "android".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+        let result = mgr.complete_pairing(&req2.token, d2);
+        assert!(result.is_err());
+        assert!(result.unwrap_err().contains("Maximum"));
+    }
+
+    #[test]
+    fn test_list_devices() {
+        let mgr = PairingManager::new(enabled_config());
+        let req = mgr.create_pairing_request().unwrap();
+        let device = PairedDevice {
+            device_id: "dev-1".to_string(),
+            display_name: "My Phone".to_string(),
+            platform: "android".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+        mgr.complete_pairing(&req.token, device).unwrap();
+
+        let devices = mgr.list_devices();
+        assert_eq!(devices.len(), 1);
+        assert_eq!(devices[0].display_name, "My Phone");
+    }
+
+    #[test]
+    fn test_remove_device() {
+        let mgr = PairingManager::new(enabled_config());
+        let req = mgr.create_pairing_request().unwrap();
+        let device = PairedDevice {
+            device_id: "dev-1".to_string(),
+            display_name: "My Phone".to_string(),
+            platform: "android".to_string(),
+            paired_at: chrono::Utc::now(),
+            last_seen: chrono::Utc::now(),
+            push_token: None,
+        };
+        mgr.complete_pairing(&req.token, device).unwrap();
+
+        assert!(mgr.remove_device("dev-1").is_ok());
+        assert!(mgr.devices.is_empty());
+    }
+
+    #[test]
+    fn test_remove_nonexistent_device() {
+        let mgr = PairingManager::new(enabled_config());
+        assert!(mgr.remove_device("nonexistent").is_err());
+    }
+
+    #[test]
+    fn test_clean_expired() {
+        let config = PairingConfig {
+            enabled: true,
+            token_expiry_secs: 0, // Expire immediately
+            ..Default::default()
+        };
+        let mgr = PairingManager::new(config);
+        mgr.create_pairing_request().unwrap();
+        assert_eq!(mgr.pending.len(), 1);
+
+        // Wait a tiny bit for expiry
+        std::thread::sleep(std::time::Duration::from_millis(10));
+        mgr.clean_expired();
+        assert!(mgr.pending.is_empty());
+    }
+
+    #[test]
+    fn test_token_length() {
+        let mgr = PairingManager::new(enabled_config());
+        let req = mgr.create_pairing_request().unwrap();
+        // 32 random bytes = 64 hex chars
+        assert_eq!(req.token.len(), 64);
+    }
+
+    #[test]
+    fn test_config_defaults() {
+        let config = PairingConfig::default();
+        assert!(!config.enabled);
+        assert_eq!(config.max_devices, 10);
+        assert_eq!(config.token_expiry_secs, 300);
+        assert_eq!(config.push_provider, "none");
+        assert!(config.ntfy_url.is_none());
+        assert!(config.ntfy_topic.is_none());
+    }
+}

crates/openfang-kernel/src/registry.rs

@@ -0,0 +1,464 @@
+//! Agent registry — tracks all agents, their state, and indexes.
+
+use dashmap::DashMap;
+use openfang_types::agent::{AgentEntry, AgentId, AgentMode, AgentState};
+use openfang_types::error::{OpenFangError, OpenFangResult};
+
+/// Registry of all agents in the kernel.
+pub struct AgentRegistry {
+    /// Primary index: agent ID → entry.
+    agents: DashMap<AgentId, AgentEntry>,
+    /// Name index: human-readable name → agent ID.
+    name_index: DashMap<String, AgentId>,
+    /// Tag index: tag → list of agent IDs.
+    tag_index: DashMap<String, Vec<AgentId>>,
+}
+
+impl AgentRegistry {
+    /// Create a new empty registry.
+    pub fn new() -> Self {
+        Self {
+            agents: DashMap::new(),
+            name_index: DashMap::new(),
+            tag_index: DashMap::new(),
+        }
+    }
+
+    /// Register a new agent.
+    pub fn register(&self, entry: AgentEntry) -> OpenFangResult<()> {
+        if self.name_index.contains_key(&entry.name) {
+            return Err(OpenFangError::AgentAlreadyExists(entry.name.clone()));
+        }
+        let id = entry.id;
+        self.name_index.insert(entry.name.clone(), id);
+        for tag in &entry.tags {
+            self.tag_index.entry(tag.clone()).or_default().push(id);
+        }
+        self.agents.insert(id, entry);
+        Ok(())
+    }
+
+    /// Get an agent entry by ID.
+    pub fn get(&self, id: AgentId) -> Option<AgentEntry> {
+        self.agents.get(&id).map(|e| e.value().clone())
+    }
+
+    /// Find an agent by name.
+    pub fn find_by_name(&self, name: &str) -> Option<AgentEntry> {
+        self.name_index
+            .get(name)
+            .and_then(|id| self.agents.get(id.value()).map(|e| e.value().clone()))
+    }
+
+    /// Update agent state.
+    pub fn set_state(&self, id: AgentId, state: AgentState) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.state = state;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update agent operational mode.
+    pub fn set_mode(&self, id: AgentId, mode: AgentMode) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.mode = mode;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Remove an agent from the registry.
+    pub fn remove(&self, id: AgentId) -> OpenFangResult<AgentEntry> {
+        let (_, entry) = self
+            .agents
+            .remove(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        self.name_index.remove(&entry.name);
+        for tag in &entry.tags {
+            if let Some(mut ids) = self.tag_index.get_mut(tag) {
+                ids.retain(|&agent_id| agent_id != id);
+            }
+        }
+        Ok(entry)
+    }
+
+    /// List all agents.
+    pub fn list(&self) -> Vec<AgentEntry> {
+        self.agents.iter().map(|e| e.value().clone()).collect()
+    }
+
+    /// Add a child agent ID to a parent's children list.
+    pub fn add_child(&self, parent_id: AgentId, child_id: AgentId) {
+        if let Some(mut entry) = self.agents.get_mut(&parent_id) {
+            entry.children.push(child_id);
+        }
+    }
+
+    /// Count of registered agents.
+    pub fn count(&self) -> usize {
+        self.agents.len()
+    }
+
+    /// Update an agent's session ID (for session reset).
+    pub fn update_session_id(
+        &self,
+        id: AgentId,
+        new_session_id: openfang_types::agent::SessionId,
+    ) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.session_id = new_session_id;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's workspace path.
+    pub fn update_workspace(
+        &self,
+        id: AgentId,
+        workspace: Option<std::path::PathBuf>,
+    ) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.workspace = workspace;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's visual identity (emoji, avatar, color).
+    pub fn update_identity(
+        &self,
+        id: AgentId,
+        identity: openfang_types::agent::AgentIdentity,
+    ) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.identity = identity;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's model configuration.
+    pub fn update_model(&self, id: AgentId, new_model: String) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.model.model = new_model;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's model AND provider together.
+    pub fn update_model_and_provider(
+        &self,
+        id: AgentId,
+        new_model: String,
+        new_provider: String,
+    ) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.model.model = new_model;
+        entry.manifest.model.provider = new_provider;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's skill allowlist.
+    pub fn update_skills(&self, id: AgentId, skills: Vec<String>) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.skills = skills;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's MCP server allowlist.
+    pub fn update_mcp_servers(&self, id: AgentId, servers: Vec<String>) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.mcp_servers = servers;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's system prompt (hot-swap, takes effect on next message).
+    pub fn update_system_prompt(&self, id: AgentId, new_prompt: String) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.model.system_prompt = new_prompt;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Update an agent's name (also updates the name index).
+    pub fn update_name(&self, id: AgentId, new_name: String) -> OpenFangResult<()> {
+        if self.name_index.contains_key(&new_name) {
+            return Err(OpenFangError::AgentAlreadyExists(new_name));
+        }
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        let old_name = entry.name.clone();
+        entry.name = new_name.clone();
+        entry.manifest.name = new_name.clone();
+        entry.last_active = chrono::Utc::now();
+        // Update name index
+        drop(entry);
+        self.name_index.remove(&old_name);
+        self.name_index.insert(new_name, id);
+        Ok(())
+    }
+
+    /// Update an agent's description.
+    pub fn update_description(&self, id: AgentId, new_desc: String) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.manifest.description = new_desc;
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+
+    /// Mark an agent's onboarding as complete.
+    pub fn mark_onboarding_complete(&self, id: AgentId) -> OpenFangResult<()> {
+        let mut entry = self
+            .agents
+            .get_mut(&id)
+            .ok_or_else(|| OpenFangError::AgentNotFound(id.to_string()))?;
+        entry.onboarding_completed = true;
+        entry.onboarding_completed_at = Some(chrono::Utc::now());
+        entry.last_active = chrono::Utc::now();
+        Ok(())
+    }
+}
+
+impl Default for AgentRegistry {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use chrono::Utc;
+    use openfang_types::agent::*;
+    use std::collections::HashMap;
+
+    fn test_entry(name: &str) -> AgentEntry {
+        AgentEntry {
+            id: AgentId::new(),
+            name: name.to_string(),
+            manifest: AgentManifest {
+                name: name.to_string(),
+                version: "0.1.0".to_string(),
+                description: "test".to_string(),
+                author: "test".to_string(),
+                module: "test".to_string(),
+                schedule: ScheduleMode::default(),
+                model: ModelConfig::default(),
+                fallback_models: vec![],
+                resources: ResourceQuota::default(),
+                priority: Priority::default(),
+                capabilities: ManifestCapabilities::default(),
+                profile: None,
+                tools: HashMap::new(),
+                skills: vec![],
+                mcp_servers: vec![],
+                metadata: HashMap::new(),
+                tags: vec![],
+                routing: None,
+                autonomous: None,
+                pinned_model: None,
+                workspace: None,
+                generate_identity_files: true,
+                exec_policy: None,
+            },
+            state: AgentState::Created,
+            mode: AgentMode::default(),
+            created_at: Utc::now(),
+            last_active: Utc::now(),
+            parent: None,
+            children: vec![],
+            session_id: SessionId::new(),
+            tags: vec![],
+            identity: Default::default(),
+            onboarding_completed: false,
+            onboarding_completed_at: None,
+        }
+    }
+
+    #[test]
+    fn test_register_and_get() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("test-agent");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+        assert!(registry.get(id).is_some());
+    }
+
+    #[test]
+    fn test_find_by_name() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("my-agent");
+        registry.register(entry).unwrap();
+        assert!(registry.find_by_name("my-agent").is_some());
+    }
+
+    #[test]
+    fn test_duplicate_name() {
+        let registry = AgentRegistry::new();
+        registry.register(test_entry("dup")).unwrap();
+        assert!(registry.register(test_entry("dup")).is_err());
+    }
+
+    #[test]
+    fn test_remove() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("removable");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+        registry.remove(id).unwrap();
+        assert!(registry.get(id).is_none());
+    }
+
+    #[test]
+    fn test_set_state() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("state-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        registry.set_state(id, AgentState::Running).unwrap();
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.state, AgentState::Running);
+    }
+
+    #[test]
+    fn test_set_mode() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("mode-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        registry.set_mode(id, AgentMode::Autonomous).unwrap();
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.mode, AgentMode::Autonomous);
+    }
+
+    #[test]
+    fn test_update_identity() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("identity-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        let identity = AgentIdentity {
+            emoji: Some("🤖".to_string()),
+            avatar_url: None,
+            color: Some("#ff0000".to_string()),
+            archetype: None,
+            vibe: None,
+            greeting_style: None,
+        };
+        registry.update_identity(id, identity).unwrap();
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.identity.emoji, Some("🤖".to_string()));
+        assert_eq!(updated.identity.color, Some("#ff0000".to_string()));
+    }
+
+    #[test]
+    fn test_update_name() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("old-name");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        registry.update_name(id, "new-name".to_string()).unwrap();
+
+        // Should be findable by new name
+        assert!(registry.find_by_name("new-name").is_some());
+        // Should not be findable by old name
+        assert!(registry.find_by_name("old-name").is_none());
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.name, "new-name");
+    }
+
+    #[test]
+    fn test_update_description() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("desc-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        registry.update_description(id, "New description".to_string()).unwrap();
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.manifest.description, "New description");
+    }
+
+    #[test]
+    fn test_update_model() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("model-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        registry.update_model(id, "gpt-4".to_string()).unwrap();
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.manifest.model.model, "gpt-4");
+    }
+
+    #[test]
+    fn test_update_skills() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("skills-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        let skills = vec!["web-search".to_string(), "code-execution".to_string()];
+        registry.update_skills(id, skills.clone()).unwrap();
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.manifest.skills, skills);
+    }
+
+    #[test]
+    fn test_update_mcp_servers() {
+        let registry = AgentRegistry::new();
+        let entry = test_entry("mcp-test");
+        let id = entry.id;
+        registry.register(entry).unwrap();
+
+        let servers = vec!["filesystem".to_string(), "database".to_string()];
+        registry.update_mcp_servers(id, servers.clone()).unwrap();
+
+        let updated = registry.get(id).unwrap();
+        assert_eq!(updated.manifest.mcp_servers, servers);
+    }
+}

crates/openfang-kernel/src/scheduler.rs

@@ -0,0 +1,170 @@
+//! Agent scheduler — manages agent execution and resource tracking.
+
+use dashmap::DashMap;
+use openfang_types::agent::{AgentId, ResourceQuota};
+use openfang_types::error::{OpenFangError, OpenFangResult};
+use openfang_types::message::TokenUsage;
+use std::time::Instant;
+use tokio::task::JoinHandle;
+use tracing::debug;
+
+/// Tracks resource usage for an agent with a rolling hourly window.
+#[derive(Debug)]
+pub struct UsageTracker {
+    /// Total tokens consumed within the current window.
+    pub total_tokens: u64,
+    /// Total tool calls made within the current window.
+    pub tool_calls: u64,
+    /// Start of the current usage window.
+    pub window_start: Instant,
+}
+
+impl Default for UsageTracker {
+    fn default() -> Self {
+        Self {
+            total_tokens: 0,
+            tool_calls: 0,
+            window_start: Instant::now(),
+        }
+    }
+}
+
+impl UsageTracker {
+    /// Reset counters if the current window has expired (1 hour).
+    fn reset_if_expired(&mut self) {
+        if self.window_start.elapsed() >= std::time::Duration::from_secs(3600) {
+            self.total_tokens = 0;
+            self.tool_calls = 0;
+            self.window_start = Instant::now();
+        }
+    }
+}
+
+/// The agent scheduler manages execution ordering and resource quotas.
+pub struct AgentScheduler {
+    /// Resource quotas per agent.
+    quotas: DashMap<AgentId, ResourceQuota>,
+    /// Usage tracking per agent.
+    usage: DashMap<AgentId, UsageTracker>,
+    /// Active task handles per agent.
+    tasks: DashMap<AgentId, JoinHandle<()>>,
+}
+
+impl AgentScheduler {
+    /// Create a new scheduler.
+    pub fn new() -> Self {
+        Self {
+            quotas: DashMap::new(),
+            usage: DashMap::new(),
+            tasks: DashMap::new(),
+        }
+    }
+
+    /// Register an agent with its resource quota.
+    pub fn register(&self, agent_id: AgentId, quota: ResourceQuota) {
+        self.quotas.insert(agent_id, quota);
+        self.usage.insert(agent_id, UsageTracker::default());
+    }
+
+    /// Record token usage for an agent.
+    pub fn record_usage(&self, agent_id: AgentId, usage: &TokenUsage) {
+        if let Some(mut tracker) = self.usage.get_mut(&agent_id) {
+            tracker.reset_if_expired();
+            tracker.total_tokens += usage.total();
+        }
+    }
+
+    /// Check if an agent has exceeded its quota.
+    pub fn check_quota(&self, agent_id: AgentId) -> OpenFangResult<()> {
+        let quota = match self.quotas.get(&agent_id) {
+            Some(q) => q.clone(),
+            None => return Ok(()), // No quota = no limit
+        };
+        let mut tracker = match self.usage.get_mut(&agent_id) {
+            Some(t) => t,
+            None => return Ok(()),
+        };
+
+        // Reset the window if an hour has passed
+        tracker.reset_if_expired();
+
+        if quota.max_llm_tokens_per_hour > 0
+            && tracker.total_tokens > quota.max_llm_tokens_per_hour
+        {
+            return Err(OpenFangError::QuotaExceeded(format!(
+                "Token limit exceeded: {} / {}",
+                tracker.total_tokens, quota.max_llm_tokens_per_hour
+            )));
+        }
+
+        Ok(())
+    }
+
+    /// Abort an agent's active task.
+    pub fn abort_task(&self, agent_id: AgentId) {
+        if let Some((_, handle)) = self.tasks.remove(&agent_id) {
+            handle.abort();
+            debug!(agent = %agent_id, "Aborted agent task");
+        }
+    }
+
+    /// Remove an agent from the scheduler.
+    pub fn unregister(&self, agent_id: AgentId) {
+        self.abort_task(agent_id);
+        self.quotas.remove(&agent_id);
+        self.usage.remove(&agent_id);
+    }
+
+    /// Get usage stats for an agent.
+    pub fn get_usage(&self, agent_id: AgentId) -> Option<(u64, u64)> {
+        self.usage
+            .get(&agent_id)
+            .map(|t| (t.total_tokens, t.tool_calls))
+    }
+}
+
+impl Default for AgentScheduler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_record_usage() {
+        let scheduler = AgentScheduler::new();
+        let id = AgentId::new();
+        scheduler.register(id, ResourceQuota::default());
+        scheduler.record_usage(
+            id,
+            &TokenUsage {
+                input_tokens: 100,
+                output_tokens: 50,
+            },
+        );
+        let (tokens, _) = scheduler.get_usage(id).unwrap();
+        assert_eq!(tokens, 150);
+    }
+
+    #[test]
+    fn test_quota_check() {
+        let scheduler = AgentScheduler::new();
+        let id = AgentId::new();
+        let quota = ResourceQuota {
+            max_llm_tokens_per_hour: 100,
+            ..Default::default()
+        };
+        scheduler.register(id, quota);
+        scheduler.record_usage(
+            id,
+            &TokenUsage {
+                input_tokens: 60,
+                output_tokens: 50,
+            },
+        );
+        assert!(scheduler.check_quota(id).is_err());
+    }
+}

crates/openfang-kernel/src/supervisor.rs

@@ -0,0 +1,227 @@
+//! Process supervision — graceful shutdown, signal handling, and health monitoring.
+
+use dashmap::DashMap;
+use openfang_types::agent::AgentId;
+use std::sync::atomic::{AtomicU64, Ordering};
+use tokio::sync::watch;
+use tracing::{info, warn};
+
+/// Shutdown signal manager with health monitoring.
+pub struct Supervisor {
+    /// Send side of the shutdown signal.
+    shutdown_tx: watch::Sender<bool>,
+    /// Receive side of the shutdown signal (clonable).
+    shutdown_rx: watch::Receiver<bool>,
+    /// Restart count (how many times agents have been restarted).
+    restart_count: AtomicU64,
+    /// Total panics caught across all agents.
+    panic_count: AtomicU64,
+    /// Per-agent restart counts for enforcing max_restarts.
+    agent_restarts: DashMap<AgentId, u32>,
+}
+
+impl Supervisor {
+    /// Create a new supervisor.
+    pub fn new() -> Self {
+        let (tx, rx) = watch::channel(false);
+        Self {
+            shutdown_tx: tx,
+            shutdown_rx: rx,
+            restart_count: AtomicU64::new(0),
+            panic_count: AtomicU64::new(0),
+            agent_restarts: DashMap::new(),
+        }
+    }
+
+    /// Get a receiver that will be notified on shutdown.
+    pub fn subscribe(&self) -> watch::Receiver<bool> {
+        self.shutdown_rx.clone()
+    }
+
+    /// Trigger a graceful shutdown.
+    pub fn shutdown(&self) {
+        info!("Supervisor: initiating graceful shutdown");
+        let _ = self.shutdown_tx.send(true);
+    }
+
+    /// Check if shutdown has been requested.
+    pub fn is_shutting_down(&self) -> bool {
+        *self.shutdown_rx.borrow()
+    }
+
+    /// Record that a panic was caught during agent execution.
+    pub fn record_panic(&self) {
+        self.panic_count.fetch_add(1, Ordering::Relaxed);
+        warn!(
+            total_panics = self.panic_count.load(Ordering::Relaxed),
+            "Agent panic recorded"
+        );
+    }
+
+    /// Record that an agent was restarted.
+    pub fn record_restart(&self) {
+        self.restart_count.fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Get the total number of panics caught.
+    pub fn panic_count(&self) -> u64 {
+        self.panic_count.load(Ordering::Relaxed)
+    }
+
+    /// Get the total number of restarts.
+    pub fn restart_count(&self) -> u64 {
+        self.restart_count.load(Ordering::Relaxed)
+    }
+
+    /// Record a restart for a specific agent and check if limit is exceeded.
+    ///
+    /// Returns Ok(restart_count) if within limit, or Err(count) if limit exceeded.
+    pub fn record_agent_restart(&self, agent_id: AgentId, max_restarts: u32) -> Result<u32, u32> {
+        let mut count = self.agent_restarts.entry(agent_id).or_insert(0);
+        *count += 1;
+        self.record_restart();
+
+        if max_restarts > 0 && *count > max_restarts {
+            warn!(
+                agent = %agent_id,
+                restarts = *count,
+                max = max_restarts,
+                "Agent exceeded max restart limit"
+            );
+            Err(*count)
+        } else {
+            Ok(*count)
+        }
+    }
+
+    /// Get the restart count for a specific agent.
+    pub fn agent_restart_count(&self, agent_id: AgentId) -> u32 {
+        self.agent_restarts.get(&agent_id).map(|r| *r).unwrap_or(0)
+    }
+
+    /// Reset restart counter for an agent (e.g., on manual intervention).
+    pub fn reset_agent_restarts(&self, agent_id: AgentId) {
+        self.agent_restarts.remove(&agent_id);
+    }
+
+    /// Get a health summary.
+    pub fn health(&self) -> SupervisorHealth {
+        SupervisorHealth {
+            is_shutting_down: self.is_shutting_down(),
+            panic_count: self.panic_count(),
+            restart_count: self.restart_count(),
+        }
+    }
+}
+
+impl Default for Supervisor {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Health report from the supervisor.
+#[derive(Debug, Clone)]
+pub struct SupervisorHealth {
+    pub is_shutting_down: bool,
+    pub panic_count: u64,
+    pub restart_count: u64,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_shutdown() {
+        let supervisor = Supervisor::new();
+        assert!(!supervisor.is_shutting_down());
+        supervisor.shutdown();
+        assert!(supervisor.is_shutting_down());
+    }
+
+    #[test]
+    fn test_subscribe() {
+        let supervisor = Supervisor::new();
+        let rx = supervisor.subscribe();
+        assert!(!*rx.borrow());
+        supervisor.shutdown();
+        assert!(rx.has_changed().unwrap());
+    }
+
+    #[test]
+    fn test_panic_tracking() {
+        let supervisor = Supervisor::new();
+        assert_eq!(supervisor.panic_count(), 0);
+        supervisor.record_panic();
+        supervisor.record_panic();
+        assert_eq!(supervisor.panic_count(), 2);
+    }
+
+    #[test]
+    fn test_restart_tracking() {
+        let supervisor = Supervisor::new();
+        assert_eq!(supervisor.restart_count(), 0);
+        supervisor.record_restart();
+        assert_eq!(supervisor.restart_count(), 1);
+    }
+
+    #[test]
+    fn test_health() {
+        let supervisor = Supervisor::new();
+        let health = supervisor.health();
+        assert!(!health.is_shutting_down);
+        assert_eq!(health.panic_count, 0);
+        assert_eq!(health.restart_count, 0);
+    }
+
+    #[test]
+    fn test_agent_restart_within_limit() {
+        let supervisor = Supervisor::new();
+        let agent_id = AgentId::new();
+
+        // Allow up to 3 restarts
+        assert!(supervisor.record_agent_restart(agent_id, 3).is_ok());
+        assert_eq!(supervisor.agent_restart_count(agent_id), 1);
+        assert!(supervisor.record_agent_restart(agent_id, 3).is_ok());
+        assert!(supervisor.record_agent_restart(agent_id, 3).is_ok());
+        assert_eq!(supervisor.agent_restart_count(agent_id), 3);
+    }
+
+    #[test]
+    fn test_agent_restart_exceeds_limit() {
+        let supervisor = Supervisor::new();
+        let agent_id = AgentId::new();
+
+        assert!(supervisor.record_agent_restart(agent_id, 2).is_ok());
+        assert!(supervisor.record_agent_restart(agent_id, 2).is_ok());
+        // 3rd restart exceeds max_restarts=2
+        let result = supervisor.record_agent_restart(agent_id, 2);
+        assert!(result.is_err());
+        assert_eq!(result.unwrap_err(), 3);
+    }
+
+    #[test]
+    fn test_agent_restart_zero_limit_unlimited() {
+        let supervisor = Supervisor::new();
+        let agent_id = AgentId::new();
+
+        // max_restarts=0 means unlimited
+        for _ in 0..100 {
+            assert!(supervisor.record_agent_restart(agent_id, 0).is_ok());
+        }
+    }
+
+    #[test]
+    fn test_reset_agent_restarts() {
+        let supervisor = Supervisor::new();
+        let agent_id = AgentId::new();
+
+        supervisor.record_agent_restart(agent_id, 10).unwrap();
+        supervisor.record_agent_restart(agent_id, 10).unwrap();
+        assert_eq!(supervisor.agent_restart_count(agent_id), 2);
+
+        supervisor.reset_agent_restarts(agent_id);
+        assert_eq!(supervisor.agent_restart_count(agent_id), 0);
+    }
+}

crates/openfang-kernel/src/triggers.rs

@@ -0,0 +1,511 @@
+//! Event-driven agent triggers — agents auto-activate when events match patterns.
+//!
+//! Agents register triggers that describe which events should wake them.
+//! When a matching event arrives on the EventBus, the trigger system
+//! sends the event content as a message to the subscribing agent.
+
+use chrono::{DateTime, Utc};
+use dashmap::DashMap;
+use openfang_types::agent::AgentId;
+use openfang_types::event::{Event, EventPayload, LifecycleEvent, SystemEvent};
+use serde::{Deserialize, Serialize};
+use tracing::{debug, info};
+use uuid::Uuid;
+
+/// Unique identifier for a trigger.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
+pub struct TriggerId(pub Uuid);
+
+impl TriggerId {
+    pub fn new() -> Self {
+        Self(Uuid::new_v4())
+    }
+}
+
+impl Default for TriggerId {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl std::fmt::Display for TriggerId {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
+/// What kind of events a trigger matches on.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TriggerPattern {
+    /// Match any lifecycle event (agent spawned, started, terminated, etc.).
+    Lifecycle,
+    /// Match when a specific agent is spawned.
+    AgentSpawned { name_pattern: String },
+    /// Match when any agent is terminated.
+    AgentTerminated,
+    /// Match any system event.
+    System,
+    /// Match a specific system event by keyword.
+    SystemKeyword { keyword: String },
+    /// Match any memory update event.
+    MemoryUpdate,
+    /// Match memory updates for a specific key pattern.
+    MemoryKeyPattern { key_pattern: String },
+    /// Match all events (wildcard).
+    All,
+    /// Match custom events by content substring.
+    ContentMatch { substring: String },
+}
+
+/// A registered trigger definition.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct Trigger {
+    /// Unique trigger ID.
+    pub id: TriggerId,
+    /// Which agent owns this trigger.
+    pub agent_id: AgentId,
+    /// The event pattern to match.
+    pub pattern: TriggerPattern,
+    /// Prompt template to send when triggered. Use `{{event}}` for event description.
+    pub prompt_template: String,
+    /// Whether this trigger is currently active.
+    pub enabled: bool,
+    /// When this trigger was created.
+    pub created_at: DateTime<Utc>,
+    /// How many times this trigger has fired.
+    pub fire_count: u64,
+    /// Maximum number of times this trigger can fire (0 = unlimited).
+    pub max_fires: u64,
+}
+
+/// The trigger engine manages event-to-agent routing.
+pub struct TriggerEngine {
+    /// All registered triggers.
+    triggers: DashMap<TriggerId, Trigger>,
+    /// Index: agent_id → list of trigger IDs belonging to that agent.
+    agent_triggers: DashMap<AgentId, Vec<TriggerId>>,
+}
+
+impl TriggerEngine {
+    /// Create a new trigger engine.
+    pub fn new() -> Self {
+        Self {
+            triggers: DashMap::new(),
+            agent_triggers: DashMap::new(),
+        }
+    }
+
+    /// Register a new trigger.
+    pub fn register(
+        &self,
+        agent_id: AgentId,
+        pattern: TriggerPattern,
+        prompt_template: String,
+        max_fires: u64,
+    ) -> TriggerId {
+        let trigger = Trigger {
+            id: TriggerId::new(),
+            agent_id,
+            pattern,
+            prompt_template,
+            enabled: true,
+            created_at: Utc::now(),
+            fire_count: 0,
+            max_fires,
+        };
+        let id = trigger.id;
+        self.triggers.insert(id, trigger);
+        self.agent_triggers.entry(agent_id).or_default().push(id);
+
+        info!(trigger_id = %id, agent_id = %agent_id, "Trigger registered");
+        id
+    }
+
+    /// Remove a trigger.
+    pub fn remove(&self, trigger_id: TriggerId) -> bool {
+        if let Some((_, trigger)) = self.triggers.remove(&trigger_id) {
+            if let Some(mut list) = self.agent_triggers.get_mut(&trigger.agent_id) {
+                list.retain(|id| *id != trigger_id);
+            }
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Remove all triggers for an agent.
+    pub fn remove_agent_triggers(&self, agent_id: AgentId) {
+        if let Some((_, trigger_ids)) = self.agent_triggers.remove(&agent_id) {
+            for id in trigger_ids {
+                self.triggers.remove(&id);
+            }
+        }
+    }
+
+    /// Enable or disable a trigger. Returns true if the trigger was found.
+    pub fn set_enabled(&self, trigger_id: TriggerId, enabled: bool) -> bool {
+        if let Some(mut t) = self.triggers.get_mut(&trigger_id) {
+            t.enabled = enabled;
+            true
+        } else {
+            false
+        }
+    }
+
+    /// List all triggers for an agent.
+    pub fn list_agent_triggers(&self, agent_id: AgentId) -> Vec<Trigger> {
+        self.agent_triggers
+            .get(&agent_id)
+            .map(|ids| {
+                ids.iter()
+                    .filter_map(|id| self.triggers.get(id).map(|t| t.clone()))
+                    .collect()
+            })
+            .unwrap_or_default()
+    }
+
+    /// List all registered triggers.
+    pub fn list_all(&self) -> Vec<Trigger> {
+        self.triggers.iter().map(|e| e.value().clone()).collect()
+    }
+
+    /// Evaluate an event against all triggers. Returns a list of
+    /// (agent_id, message_to_send) pairs for matching triggers.
+    pub fn evaluate(&self, event: &Event) -> Vec<(AgentId, String)> {
+        let event_description = describe_event(event);
+        let mut matches = Vec::new();
+
+        for mut entry in self.triggers.iter_mut() {
+            let trigger = entry.value_mut();
+
+            if !trigger.enabled {
+                continue;
+            }
+
+            // Check max fires
+            if trigger.max_fires > 0 && trigger.fire_count >= trigger.max_fires {
+                trigger.enabled = false;
+                continue;
+            }
+
+            if matches_pattern(&trigger.pattern, event, &event_description) {
+                let message = trigger
+                    .prompt_template
+                    .replace("{{event}}", &event_description);
+                matches.push((trigger.agent_id, message));
+                trigger.fire_count += 1;
+
+                debug!(
+                    trigger_id = %trigger.id,
+                    agent_id = %trigger.agent_id,
+                    fire_count = trigger.fire_count,
+                    "Trigger fired"
+                );
+            }
+        }
+
+        matches
+    }
+
+    /// Get a trigger by ID.
+    pub fn get(&self, trigger_id: TriggerId) -> Option<Trigger> {
+        self.triggers.get(&trigger_id).map(|t| t.clone())
+    }
+}
+
+impl Default for TriggerEngine {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Check if an event matches a trigger pattern.
+fn matches_pattern(pattern: &TriggerPattern, event: &Event, description: &str) -> bool {
+    match pattern {
+        TriggerPattern::All => true,
+        TriggerPattern::Lifecycle => {
+            matches!(event.payload, EventPayload::Lifecycle(_))
+        }
+        TriggerPattern::AgentSpawned { name_pattern } => {
+            if let EventPayload::Lifecycle(LifecycleEvent::Spawned { name, .. }) = &event.payload {
+                name.contains(name_pattern.as_str()) || name_pattern == "*"
+            } else {
+                false
+            }
+        }
+        TriggerPattern::AgentTerminated => matches!(
+            event.payload,
+            EventPayload::Lifecycle(LifecycleEvent::Terminated { .. })
+                | EventPayload::Lifecycle(LifecycleEvent::Crashed { .. })
+        ),
+        TriggerPattern::System => {
+            matches!(event.payload, EventPayload::System(_))
+        }
+        TriggerPattern::SystemKeyword { keyword } => {
+            if let EventPayload::System(se) = &event.payload {
+                let se_str = format!("{:?}", se).to_lowercase();
+                se_str.contains(&keyword.to_lowercase())
+            } else {
+                false
+            }
+        }
+        TriggerPattern::MemoryUpdate => {
+            matches!(event.payload, EventPayload::MemoryUpdate(_))
+        }
+        TriggerPattern::MemoryKeyPattern { key_pattern } => {
+            if let EventPayload::MemoryUpdate(delta) = &event.payload {
+                delta.key.contains(key_pattern.as_str()) || key_pattern == "*"
+            } else {
+                false
+            }
+        }
+        TriggerPattern::ContentMatch { substring } => description
+            .to_lowercase()
+            .contains(&substring.to_lowercase()),
+    }
+}
+
+/// Create a human-readable description of an event for use in prompts.
+fn describe_event(event: &Event) -> String {
+    match &event.payload {
+        EventPayload::Message(msg) => {
+            format!("Message from {:?}: {}", msg.role, msg.content)
+        }
+        EventPayload::ToolResult(tr) => {
+            format!(
+                "Tool '{}' {} ({}ms): {}",
+                tr.tool_id,
+                if tr.success { "succeeded" } else { "failed" },
+                tr.execution_time_ms,
+                &tr.content[..tr.content.len().min(200)]
+            )
+        }
+        EventPayload::MemoryUpdate(delta) => {
+            format!(
+                "Memory {:?} on key '{}' for agent {}",
+                delta.operation, delta.key, delta.agent_id
+            )
+        }
+        EventPayload::Lifecycle(le) => match le {
+            LifecycleEvent::Spawned { agent_id, name } => {
+                format!("Agent '{name}' (id: {agent_id}) was spawned")
+            }
+            LifecycleEvent::Started { agent_id } => {
+                format!("Agent {agent_id} started")
+            }
+            LifecycleEvent::Suspended { agent_id } => {
+                format!("Agent {agent_id} suspended")
+            }
+            LifecycleEvent::Resumed { agent_id } => {
+                format!("Agent {agent_id} resumed")
+            }
+            LifecycleEvent::Terminated { agent_id, reason } => {
+                format!("Agent {agent_id} terminated: {reason}")
+            }
+            LifecycleEvent::Crashed { agent_id, error } => {
+                format!("Agent {agent_id} crashed: {error}")
+            }
+        },
+        EventPayload::Network(ne) => {
+            format!("Network event: {:?}", ne)
+        }
+        EventPayload::System(se) => match se {
+            SystemEvent::KernelStarted => "Kernel started".to_string(),
+            SystemEvent::KernelStopping => "Kernel stopping".to_string(),
+            SystemEvent::QuotaWarning {
+                agent_id,
+                resource,
+                usage_percent,
+            } => format!("Quota warning: agent {agent_id}, {resource} at {usage_percent:.1}%"),
+            SystemEvent::HealthCheck { status } => {
+                format!("Health check: {status}")
+            }
+            SystemEvent::QuotaEnforced {
+                agent_id,
+                spent,
+                limit,
+            } => {
+                format!("Quota enforced: agent {agent_id}, spent ${spent:.4} / ${limit:.4}")
+            }
+            SystemEvent::ModelRouted {
+                agent_id,
+                complexity,
+                model,
+            } => {
+                format!("Model routed: agent {agent_id}, complexity={complexity}, model={model}")
+            }
+            SystemEvent::UserAction {
+                user_id,
+                action,
+                result,
+            } => {
+                format!("User action: {user_id} {action} -> {result}")
+            }
+            SystemEvent::HealthCheckFailed {
+                agent_id,
+                unresponsive_secs,
+            } => {
+                format!(
+                    "Health check failed: agent {agent_id}, unresponsive for {unresponsive_secs}s"
+                )
+            }
+        },
+        EventPayload::Custom(data) => {
+            format!("Custom event ({} bytes)", data.len())
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use openfang_types::event::*;
+
+    #[test]
+    fn test_register_trigger() {
+        let engine = TriggerEngine::new();
+        let agent_id = AgentId::new();
+        let id = engine.register(
+            agent_id,
+            TriggerPattern::All,
+            "Event occurred: {{event}}".to_string(),
+            0,
+        );
+        assert!(engine.get(id).is_some());
+    }
+
+    #[test]
+    fn test_evaluate_lifecycle() {
+        let engine = TriggerEngine::new();
+        let watcher = AgentId::new();
+        engine.register(
+            watcher,
+            TriggerPattern::Lifecycle,
+            "Lifecycle: {{event}}".to_string(),
+            0,
+        );
+
+        let event = Event::new(
+            AgentId::new(),
+            EventTarget::Broadcast,
+            EventPayload::Lifecycle(LifecycleEvent::Spawned {
+                agent_id: AgentId::new(),
+                name: "new-agent".to_string(),
+            }),
+        );
+
+        let matches = engine.evaluate(&event);
+        assert_eq!(matches.len(), 1);
+        assert_eq!(matches[0].0, watcher);
+        assert!(matches[0].1.contains("new-agent"));
+    }
+
+    #[test]
+    fn test_evaluate_agent_spawned_pattern() {
+        let engine = TriggerEngine::new();
+        let watcher = AgentId::new();
+        engine.register(
+            watcher,
+            TriggerPattern::AgentSpawned {
+                name_pattern: "coder".to_string(),
+            },
+            "Coder spawned: {{event}}".to_string(),
+            0,
+        );
+
+        // This should match
+        let event = Event::new(
+            AgentId::new(),
+            EventTarget::Broadcast,
+            EventPayload::Lifecycle(LifecycleEvent::Spawned {
+                agent_id: AgentId::new(),
+                name: "coder".to_string(),
+            }),
+        );
+        assert_eq!(engine.evaluate(&event).len(), 1);
+
+        // This should NOT match
+        let event2 = Event::new(
+            AgentId::new(),
+            EventTarget::Broadcast,
+            EventPayload::Lifecycle(LifecycleEvent::Spawned {
+                agent_id: AgentId::new(),
+                name: "researcher".to_string(),
+            }),
+        );
+        assert_eq!(engine.evaluate(&event2).len(), 0);
+    }
+
+    #[test]
+    fn test_max_fires() {
+        let engine = TriggerEngine::new();
+        let agent_id = AgentId::new();
+        engine.register(
+            agent_id,
+            TriggerPattern::All,
+            "Event: {{event}}".to_string(),
+            2, // max 2 fires
+        );
+
+        let event = Event::new(
+            AgentId::new(),
+            EventTarget::Broadcast,
+            EventPayload::System(SystemEvent::HealthCheck {
+                status: "ok".to_string(),
+            }),
+        );
+
+        // First two should match
+        assert_eq!(engine.evaluate(&event).len(), 1);
+        assert_eq!(engine.evaluate(&event).len(), 1);
+        // Third should not
+        assert_eq!(engine.evaluate(&event).len(), 0);
+    }
+
+    #[test]
+    fn test_remove_trigger() {
+        let engine = TriggerEngine::new();
+        let agent_id = AgentId::new();
+        let id = engine.register(agent_id, TriggerPattern::All, "msg".to_string(), 0);
+        assert!(engine.remove(id));
+        assert!(engine.get(id).is_none());
+    }
+
+    #[test]
+    fn test_remove_agent_triggers() {
+        let engine = TriggerEngine::new();
+        let agent_id = AgentId::new();
+        engine.register(agent_id, TriggerPattern::All, "a".to_string(), 0);
+        engine.register(agent_id, TriggerPattern::System, "b".to_string(), 0);
+        assert_eq!(engine.list_agent_triggers(agent_id).len(), 2);
+
+        engine.remove_agent_triggers(agent_id);
+        assert_eq!(engine.list_agent_triggers(agent_id).len(), 0);
+    }
+
+    #[test]
+    fn test_content_match() {
+        let engine = TriggerEngine::new();
+        let agent_id = AgentId::new();
+        engine.register(
+            agent_id,
+            TriggerPattern::ContentMatch {
+                substring: "quota".to_string(),
+            },
+            "Alert: {{event}}".to_string(),
+            0,
+        );
+
+        let event = Event::new(
+            AgentId::new(),
+            EventTarget::System,
+            EventPayload::System(SystemEvent::QuotaWarning {
+                agent_id: AgentId::new(),
+                resource: "tokens".to_string(),
+                usage_percent: 85.0,
+            }),
+        );
+        assert_eq!(engine.evaluate(&event).len(), 1);
+    }
+}

crates/openfang-kernel/src/whatsapp_gateway.rs

@@ -0,0 +1,345 @@
+//! WhatsApp Web gateway — embedded Node.js process management.
+//!
+//! Embeds the gateway JS at compile time, extracts it to `~/.openfang/whatsapp-gateway/`,
+//! runs `npm install` if needed, and spawns `node index.js` as a managed child process
+//! that auto-restarts on crash.
+
+use crate::config::openfang_home;
+use std::path::PathBuf;
+use std::sync::Arc;
+use tracing::{info, warn};
+
+/// Gateway source files embedded at compile time.
+const GATEWAY_INDEX_JS: &str =
+    include_str!("../../../packages/whatsapp-gateway/index.js");
+const GATEWAY_PACKAGE_JSON: &str =
+    include_str!("../../../packages/whatsapp-gateway/package.json");
+
+/// Default port for the WhatsApp Web gateway.
+const DEFAULT_GATEWAY_PORT: u16 = 3009;
+
+/// Maximum restart attempts before giving up.
+const MAX_RESTARTS: u32 = 3;
+
+/// Restart backoff delays in seconds: 5s, 10s, 20s.
+const RESTART_DELAYS: [u64; 3] = [5, 10, 20];
+
+/// Get the gateway installation directory.
+fn gateway_dir() -> PathBuf {
+    openfang_home().join("whatsapp-gateway")
+}
+
+/// Compute a simple hash of content for change detection.
+fn content_hash(content: &str) -> String {
+    // Use a simple FNV-style hash — no crypto needed, just change detection.
+    let mut hash: u64 = 0xcbf29ce484222325;
+    for byte in content.as_bytes() {
+        hash ^= *byte as u64;
+        hash = hash.wrapping_mul(0x100000001b3);
+    }
+    format!("{hash:016x}")
+}
+
+/// Write a file only if its content hash differs from the existing file.
+/// Returns `true` if the file was written (content changed).
+fn write_if_changed(path: &std::path::Path, content: &str) -> std::io::Result<bool> {
+    let hash_path = path.with_extension("hash");
+    let new_hash = content_hash(content);
+
+    // Check existing hash
+    if let Ok(existing_hash) = std::fs::read_to_string(&hash_path) {
+        if existing_hash.trim() == new_hash {
+            return Ok(false); // No change
+        }
+    }
+
+    std::fs::write(path, content)?;
+    std::fs::write(&hash_path, &new_hash)?;
+    Ok(true)
+}
+
+/// Ensure the gateway files are extracted and npm dependencies installed.
+///
+/// Returns the gateway directory path on success, or an error message.
+async fn ensure_gateway_installed() -> Result<PathBuf, String> {
+    let dir = gateway_dir();
+    std::fs::create_dir_all(&dir).map_err(|e| format!("Failed to create gateway dir: {e}"))?;
+
+    let index_path = dir.join("index.js");
+    let package_path = dir.join("package.json");
+
+    // Write files only if content changed (avoids unnecessary npm install)
+    let index_changed =
+        write_if_changed(&index_path, GATEWAY_INDEX_JS).map_err(|e| format!("Write index.js: {e}"))?;
+    let package_changed = write_if_changed(&package_path, GATEWAY_PACKAGE_JSON)
+        .map_err(|e| format!("Write package.json: {e}"))?;
+
+    let node_modules = dir.join("node_modules");
+    let needs_install = !node_modules.exists() || package_changed;
+
+    if needs_install {
+        info!("Installing WhatsApp gateway npm dependencies...");
+
+        // Determine npm command (npm.cmd on Windows, npm elsewhere)
+        let npm_cmd = if cfg!(windows) { "npm.cmd" } else { "npm" };
+
+        let output = tokio::process::Command::new(npm_cmd)
+            .arg("install")
+            .arg("--production")
+            .current_dir(&dir)
+            .stdout(std::process::Stdio::piped())
+            .stderr(std::process::Stdio::piped())
+            .output()
+            .await
+            .map_err(|e| format!("npm install failed to start: {e}"))?;
+
+        if !output.status.success() {
+            let stderr = String::from_utf8_lossy(&output.stderr);
+            return Err(format!("npm install failed: {stderr}"));
+        }
+
+        info!("WhatsApp gateway npm dependencies installed");
+    } else if index_changed {
+        info!("WhatsApp gateway index.js updated (binary upgrade)");
+    }
+
+    Ok(dir)
+}
+
+/// Check if Node.js is available on the system.
+async fn node_available() -> bool {
+    let node_cmd = if cfg!(windows) { "node.exe" } else { "node" };
+    tokio::process::Command::new(node_cmd)
+        .arg("--version")
+        .stdout(std::process::Stdio::null())
+        .stderr(std::process::Stdio::null())
+        .status()
+        .await
+        .map(|s| s.success())
+        .unwrap_or(false)
+}
+
+/// Start the WhatsApp Web gateway as a managed child process.
+///
+/// This function:
+/// 1. Checks if Node.js is available
+/// 2. Extracts and installs the gateway files
+/// 3. Spawns `node index.js` with appropriate env vars
+/// 4. Sets `WHATSAPP_WEB_GATEWAY_URL` so the daemon finds it
+/// 5. Monitors the process and restarts on crash (up to 3 times)
+///
+/// The PID is stored in the kernel's `whatsapp_gateway_pid` for shutdown cleanup.
+pub async fn start_whatsapp_gateway(kernel: &Arc<super::kernel::OpenFangKernel>) {
+    // Only start if WhatsApp is configured
+    let wa_config = match &kernel.config.channels.whatsapp {
+        Some(cfg) => cfg.clone(),
+        None => return,
+    };
+
+    // Check for Node.js
+    if !node_available().await {
+        warn!(
+            "WhatsApp Web gateway requires Node.js >= 18 but `node` was not found. \
+             Install Node.js to enable WhatsApp Web integration."
+        );
+        return;
+    }
+
+    // Extract and install
+    let gateway_path = match ensure_gateway_installed().await {
+        Ok(p) => p,
+        Err(e) => {
+            warn!("WhatsApp Web gateway setup failed: {e}");
+            return;
+        }
+    };
+
+    let port = DEFAULT_GATEWAY_PORT;
+    let api_listen = &kernel.config.api_listen;
+    let openfang_url = format!("http://{api_listen}");
+    let default_agent = wa_config
+        .default_agent
+        .as_deref()
+        .unwrap_or("assistant")
+        .to_string();
+
+    // Auto-set the env var so the rest of the system finds the gateway
+    std::env::set_var("WHATSAPP_WEB_GATEWAY_URL", format!("http://127.0.0.1:{port}"));
+    info!("WHATSAPP_WEB_GATEWAY_URL set to http://127.0.0.1:{port}");
+
+    // Spawn with crash monitoring
+    let kernel_weak = Arc::downgrade(kernel);
+    let gateway_pid = Arc::clone(&kernel.whatsapp_gateway_pid);
+
+    tokio::spawn(async move {
+        let mut restarts = 0u32;
+
+        loop {
+            let node_cmd = if cfg!(windows) { "node.exe" } else { "node" };
+
+            info!("Starting WhatsApp Web gateway (attempt {})", restarts + 1);
+
+            let child = tokio::process::Command::new(node_cmd)
+                .arg("index.js")
+                .current_dir(&gateway_path)
+                .env("WHATSAPP_GATEWAY_PORT", port.to_string())
+                .env("OPENFANG_URL", &openfang_url)
+                .env("OPENFANG_DEFAULT_AGENT", &default_agent)
+                .stdout(std::process::Stdio::inherit())
+                .stderr(std::process::Stdio::inherit())
+                .spawn();
+
+            let mut child = match child {
+                Ok(c) => c,
+                Err(e) => {
+                    warn!("Failed to spawn WhatsApp gateway: {e}");
+                    return;
+                }
+            };
+
+            // Store PID for shutdown cleanup
+            if let Some(pid) = child.id() {
+                if let Ok(mut guard) = gateway_pid.lock() {
+                    *guard = Some(pid);
+                }
+                info!("WhatsApp Web gateway started (PID {pid})");
+            }
+
+            // Wait for process exit
+            match child.wait().await {
+                Ok(status) => {
+                    // Clear stored PID
+                    if let Ok(mut guard) = gateway_pid.lock() {
+                        *guard = None;
+                    }
+
+                    // Check if kernel is still alive (not shutting down)
+                    let kernel = match kernel_weak.upgrade() {
+                        Some(k) => k,
+                        None => {
+                            info!("WhatsApp gateway exited (kernel dropped)");
+                            return;
+                        }
+                    };
+
+                    if kernel.supervisor.is_shutting_down() {
+                        info!("WhatsApp gateway stopped (daemon shutting down)");
+                        return;
+                    }
+
+                    if status.success() {
+                        info!("WhatsApp gateway exited cleanly");
+                        return;
+                    }
+
+                    warn!(
+                        "WhatsApp gateway crashed (exit: {status}), restart {}/{MAX_RESTARTS}",
+                        restarts + 1
+                    );
+                }
+                Err(e) => {
+                    if let Ok(mut guard) = gateway_pid.lock() {
+                        *guard = None;
+                    }
+                    warn!("WhatsApp gateway wait error: {e}");
+                }
+            }
+
+            restarts += 1;
+            if restarts >= MAX_RESTARTS {
+                warn!(
+                    "WhatsApp gateway exceeded max restarts ({MAX_RESTARTS}), giving up"
+                );
+                return;
+            }
+
+            // Backoff before restart
+            let delay = RESTART_DELAYS
+                .get(restarts as usize - 1)
+                .copied()
+                .unwrap_or(20);
+            info!("Restarting WhatsApp gateway in {delay}s...");
+            tokio::time::sleep(std::time::Duration::from_secs(delay)).await;
+        }
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_embedded_files_not_empty() {
+        assert!(!GATEWAY_INDEX_JS.is_empty());
+        assert!(!GATEWAY_PACKAGE_JSON.is_empty());
+        assert!(GATEWAY_INDEX_JS.contains("WhatsApp"));
+        assert!(GATEWAY_PACKAGE_JSON.contains("@openfang/whatsapp-gateway"));
+    }
+
+    #[test]
+    fn test_content_hash_deterministic() {
+        let h1 = content_hash("hello world");
+        let h2 = content_hash("hello world");
+        assert_eq!(h1, h2);
+    }
+
+    #[test]
+    fn test_content_hash_changes_on_different_input() {
+        let h1 = content_hash("version 1");
+        let h2 = content_hash("version 2");
+        assert_ne!(h1, h2);
+    }
+
+    #[test]
+    fn test_gateway_dir_under_openfang_home() {
+        let dir = gateway_dir();
+        assert!(dir.ends_with("whatsapp-gateway"));
+        assert!(dir
+            .parent()
+            .unwrap()
+            .to_string_lossy()
+            .contains(".openfang"));
+    }
+
+    #[test]
+    fn test_write_if_changed_creates_new_file() {
+        let tmp = std::env::temp_dir().join("openfang_test_gateway");
+        let _ = std::fs::create_dir_all(&tmp);
+        let path = tmp.join("test_write.js");
+        let hash_path = path.with_extension("hash");
+
+        // Clean up any previous runs
+        let _ = std::fs::remove_file(&path);
+        let _ = std::fs::remove_file(&hash_path);
+
+        // First write should return true (new file)
+        let changed = write_if_changed(&path, "console.log('v1')").unwrap();
+        assert!(changed);
+        assert!(path.exists());
+        assert!(hash_path.exists());
+
+        // Same content should return false
+        let changed = write_if_changed(&path, "console.log('v1')").unwrap();
+        assert!(!changed);
+
+        // Different content should return true
+        let changed = write_if_changed(&path, "console.log('v2')").unwrap();
+        assert!(changed);
+
+        // Clean up
+        let _ = std::fs::remove_file(&path);
+        let _ = std::fs::remove_file(&hash_path);
+        let _ = std::fs::remove_dir(&tmp);
+    }
+
+    #[test]
+    fn test_default_gateway_port() {
+        assert_eq!(DEFAULT_GATEWAY_PORT, 3009);
+    }
+
+    #[test]
+    fn test_restart_backoff_delays() {
+        assert_eq!(RESTART_DELAYS, [5, 10, 20]);
+        assert_eq!(MAX_RESTARTS, 3);
+    }
+}

crates/openfang-kernel/src/wizard.rs

@@ -0,0 +1,435 @@
+//! NL Auto-Bootstrap Wizard — generates agent configs from natural language.
+//!
+//! The wizard takes a user's natural language description of what they want
+//! an agent to do, extracts structured intent, and generates a complete
+//! agent manifest (TOML config) ready to spawn.
+
+use openfang_types::agent::{
+    AgentManifest, ManifestCapabilities, ModelConfig, Priority, ResourceQuota, ScheduleMode,
+};
+use serde::{Deserialize, Serialize};
+use std::collections::HashMap;
+
+/// The extracted intent from a user's natural language description.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AgentIntent {
+    /// Agent name (slug-style).
+    pub name: String,
+    /// Short description.
+    pub description: String,
+    /// What the agent should do (summarized task).
+    pub task: String,
+    /// What skills/tools it needs.
+    pub skills: Vec<String>,
+    /// Suggested model tier (simple, medium, complex).
+    pub model_tier: String,
+    /// Whether it runs on a schedule.
+    pub scheduled: bool,
+    /// Schedule expression (cron or interval).
+    pub schedule: Option<String>,
+    /// Suggested capabilities.
+    pub capabilities: Vec<String>,
+}
+
+/// A generated setup plan from the wizard.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct SetupPlan {
+    /// The extracted intent.
+    pub intent: AgentIntent,
+    /// Generated agent manifest (ready to write as TOML).
+    pub manifest: AgentManifest,
+    /// Skills to install (if not already installed).
+    pub skills_to_install: Vec<String>,
+    /// Human-readable summary of what will be created.
+    pub summary: String,
+}
+
+/// The setup wizard builds agent configurations from natural language.
+pub struct SetupWizard;
+
+impl SetupWizard {
+    /// Build a setup plan from an extracted intent.
+    ///
+    /// This maps the intent into a concrete agent manifest with appropriate
+    /// model configuration, capabilities, and schedule.
+    pub fn build_plan(intent: AgentIntent) -> SetupPlan {
+        // Map model tier to provider/model
+        let (provider, model) = match intent.model_tier.as_str() {
+            "simple" => ("groq", "llama-3.3-70b-versatile"),
+            "complex" => ("anthropic", "claude-sonnet-4-20250514"),
+            _ => ("groq", "llama-3.3-70b-versatile"), // medium default
+        };
+
+        // Build capabilities from intent
+        let mut caps = ManifestCapabilities::default();
+        for cap in &intent.capabilities {
+            match cap.as_str() {
+                "web" | "network" => caps.network.push("*".to_string()),
+                "file_read" => caps.tools.push("file_read".to_string()),
+                "file_write" => caps.tools.push("file_write".to_string()),
+                "file" | "files" => {
+                    for t in &["file_read", "file_write", "file_list"] {
+                        let s = t.to_string();
+                        if !caps.tools.contains(&s) {
+                            caps.tools.push(s);
+                        }
+                    }
+                }
+                "shell" => caps.shell.push("*".to_string()),
+                "memory" => {
+                    caps.memory_read.push("*".to_string());
+                    caps.memory_write.push("*".to_string());
+                    for t in &["memory_store", "memory_recall"] {
+                        let s = t.to_string();
+                        if !caps.tools.contains(&s) {
+                            caps.tools.push(s);
+                        }
+                    }
+                }
+                "browser" | "browse" => {
+                    caps.network.push("*".to_string());
+                    for t in &[
+                        "browser_navigate",
+                        "browser_click",
+                        "browser_type",
+                        "browser_read_page",
+                        "browser_screenshot",
+                        "browser_close",
+                    ] {
+                        let s = t.to_string();
+                        if !caps.tools.contains(&s) {
+                            caps.tools.push(s);
+                        }
+                    }
+                }
+                other => caps.tools.push(other.to_string()),
+            }
+        }
+
+        // Add web_search + web_fetch if web/network capability is needed
+        if caps.network.contains(&"*".to_string()) {
+            for t in &["web_search", "web_fetch"] {
+                let s = t.to_string();
+                if !caps.tools.contains(&s) {
+                    caps.tools.push(s);
+                }
+            }
+        }
+
+        // Build schedule
+        let schedule = if intent.scheduled {
+            if let Some(ref cron) = intent.schedule {
+                ScheduleMode::Periodic { cron: cron.clone() }
+            } else {
+                ScheduleMode::default()
+            }
+        } else {
+            ScheduleMode::default()
+        };
+
+        // Build system prompt — rich enough to guide the agent on its task.
+        // The prompt_builder will wrap this with tool descriptions, memory protocol,
+        // safety guidelines, etc. at execution time.
+        let tool_hints = Self::tool_hints_for(&caps.tools);
+        let system_prompt = format!(
+            "You are {name}, an AI agent running inside the OpenFang Agent OS.\n\
+             \n\
+             YOUR TASK: {task}\n\
+             \n\
+             APPROACH:\n\
+             - Understand the request fully before acting.\n\
+             - Use your tools to accomplish the task rather than just describing what to do.\n\
+             - If you need information, search for it. If you need to read a file, read it.\n\
+             - Be concise in your responses. Lead with results, not process narration.\n\
+             {tool_hints}",
+            name = intent.name,
+            task = intent.task,
+            tool_hints = tool_hints,
+        );
+
+        let manifest = AgentManifest {
+            name: intent.name.clone(),
+            version: "0.1.0".to_string(),
+            description: intent.description.clone(),
+            author: "wizard".to_string(),
+            module: "builtin:chat".to_string(),
+            schedule,
+            model: ModelConfig {
+                provider: provider.to_string(),
+                model: model.to_string(),
+                max_tokens: 4096,
+                temperature: 0.7,
+                system_prompt,
+                api_key_env: None,
+                base_url: None,
+            },
+            resources: ResourceQuota::default(),
+            priority: Priority::default(),
+            capabilities: caps,
+            tools: HashMap::new(),
+            skills: intent.skills.clone(),
+            mcp_servers: vec![],
+            metadata: HashMap::new(),
+            tags: vec![],
+            routing: None,
+            autonomous: None,
+            pinned_model: None,
+            workspace: None,
+            generate_identity_files: true,
+            profile: None,
+            fallback_models: vec![],
+            exec_policy: None,
+        };
+
+        let skills_to_install: Vec<String> = intent
+            .skills
+            .iter()
+            .filter(|s| !s.is_empty())
+            .cloned()
+            .collect();
+
+        let summary = format!(
+            "Agent '{}': {}\n  Model: {}/{}\n  Skills: {}\n  Schedule: {}",
+            intent.name,
+            intent.description,
+            provider,
+            model,
+            if skills_to_install.is_empty() {
+                "none".to_string()
+            } else {
+                skills_to_install.join(", ")
+            },
+            if intent.scheduled {
+                intent.schedule.as_deref().unwrap_or("on-demand")
+            } else {
+                "on-demand"
+            }
+        );
+
+        SetupPlan {
+            intent,
+            manifest,
+            skills_to_install,
+            summary,
+        }
+    }
+
+    /// Build a short tool usage hint block for the system prompt based on granted tools.
+    fn tool_hints_for(tools: &[String]) -> String {
+        let mut hints = Vec::new();
+        let has = |name: &str| tools.iter().any(|t| t == name);
+
+        if has("web_search") {
+            hints.push("- Use web_search to find current information on any topic.");
+        }
+        if has("web_fetch") {
+            hints.push("- Use web_fetch to read the full content of a specific URL as markdown.");
+        }
+        if has("browser_navigate") {
+            hints.push("- Use browser_navigate/click/type/read_page to interact with websites.");
+        }
+        if has("file_read") {
+            hints.push("- Use file_read to examine files before modifying them.");
+        }
+        if has("shell_exec") {
+            hints.push(
+                "- Use shell_exec to run commands. Explain destructive commands before running.",
+            );
+        }
+        if has("memory_store") {
+            hints.push(
+                "- Use memory_store/memory_recall to persist and retrieve important context.",
+            );
+        }
+
+        if hints.is_empty() {
+            String::new()
+        } else {
+            format!("\nKEY TOOLS:\n{}", hints.join("\n"))
+        }
+    }
+
+    /// Generate a TOML string from an agent manifest.
+    pub fn manifest_to_toml(manifest: &AgentManifest) -> Result<String, toml::ser::Error> {
+        toml::to_string_pretty(manifest)
+    }
+
+    /// Parse an intent from a JSON string (typically LLM output).
+    pub fn parse_intent(json: &str) -> Result<AgentIntent, serde_json::Error> {
+        serde_json::from_str(json)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn sample_intent() -> AgentIntent {
+        AgentIntent {
+            name: "research-bot".to_string(),
+            description: "Researches topics and provides summaries".to_string(),
+            task: "Search the web for information and provide concise summaries".to_string(),
+            skills: vec!["web-summarizer".to_string()],
+            model_tier: "medium".to_string(),
+            scheduled: false,
+            schedule: None,
+            capabilities: vec!["web".to_string(), "memory".to_string()],
+        }
+    }
+
+    #[test]
+    fn test_build_plan_basic() {
+        let intent = sample_intent();
+        let plan = SetupWizard::build_plan(intent);
+
+        assert_eq!(plan.manifest.name, "research-bot");
+        assert_eq!(plan.manifest.model.provider, "groq");
+        assert!(plan
+            .manifest
+            .capabilities
+            .network
+            .contains(&"*".to_string()));
+        assert!(plan.summary.contains("research-bot"));
+    }
+
+    #[test]
+    fn test_build_plan_complex_tier() {
+        let mut intent = sample_intent();
+        intent.model_tier = "complex".to_string();
+        let plan = SetupWizard::build_plan(intent);
+
+        assert_eq!(plan.manifest.model.provider, "anthropic");
+        assert!(plan.manifest.model.model.contains("sonnet"));
+    }
+
+    #[test]
+    fn test_build_plan_scheduled() {
+        let mut intent = sample_intent();
+        intent.scheduled = true;
+        intent.schedule = Some("0 */6 * * *".to_string());
+        let plan = SetupWizard::build_plan(intent);
+
+        match &plan.manifest.schedule {
+            ScheduleMode::Periodic { cron } => {
+                assert_eq!(cron, "0 */6 * * *");
+            }
+            _ => panic!("Expected periodic schedule mode"),
+        }
+    }
+
+    #[test]
+    fn test_parse_intent_json() {
+        let json = r#"{
+            "name": "code-reviewer",
+            "description": "Reviews code and suggests improvements",
+            "task": "Analyze pull requests and provide feedback",
+            "skills": [],
+            "model_tier": "complex",
+            "scheduled": false,
+            "schedule": null,
+            "capabilities": ["file_read"]
+        }"#;
+
+        let intent = SetupWizard::parse_intent(json).unwrap();
+        assert_eq!(intent.name, "code-reviewer");
+        assert_eq!(intent.model_tier, "complex");
+    }
+
+    #[test]
+    fn test_manifest_to_toml() {
+        let intent = sample_intent();
+        let plan = SetupWizard::build_plan(intent);
+        let toml = SetupWizard::manifest_to_toml(&plan.manifest);
+        assert!(toml.is_ok());
+        let toml_str = toml.unwrap();
+        assert!(toml_str.contains("research-bot"));
+    }
+
+    #[test]
+    fn test_web_tools_auto_added() {
+        let intent = AgentIntent {
+            name: "test".to_string(),
+            description: "test".to_string(),
+            task: "test".to_string(),
+            skills: vec![],
+            model_tier: "simple".to_string(),
+            scheduled: false,
+            schedule: None,
+            capabilities: vec!["web".to_string()],
+        };
+        let plan = SetupWizard::build_plan(intent);
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"web_fetch".to_string()));
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"web_search".to_string()));
+    }
+
+    #[test]
+    fn test_memory_tools_auto_added() {
+        let intent = AgentIntent {
+            name: "test".to_string(),
+            description: "test".to_string(),
+            task: "test".to_string(),
+            skills: vec![],
+            model_tier: "simple".to_string(),
+            scheduled: false,
+            schedule: None,
+            capabilities: vec!["memory".to_string()],
+        };
+        let plan = SetupWizard::build_plan(intent);
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"memory_store".to_string()));
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"memory_recall".to_string()));
+    }
+
+    #[test]
+    fn test_browser_tools_auto_added() {
+        let intent = AgentIntent {
+            name: "test".to_string(),
+            description: "test".to_string(),
+            task: "test".to_string(),
+            skills: vec![],
+            model_tier: "simple".to_string(),
+            scheduled: false,
+            schedule: None,
+            capabilities: vec!["browser".to_string()],
+        };
+        let plan = SetupWizard::build_plan(intent);
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"browser_navigate".to_string()));
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"browser_click".to_string()));
+        assert!(plan
+            .manifest
+            .capabilities
+            .tools
+            .contains(&"browser_read_page".to_string()));
+    }
+
+    #[test]
+    fn test_wizard_system_prompt_has_task() {
+        let intent = sample_intent();
+        let plan = SetupWizard::build_plan(intent);
+        assert!(plan.manifest.model.system_prompt.contains("YOUR TASK:"));
+        assert!(plan.manifest.model.system_prompt.contains("Search the web"));
+    }
+}

crates/openfang-kernel/tests/integration_test.rs

@@ -0,0 +1,163 @@
+//! Integration test: boot kernel -> spawn agent -> send message via Groq API.
+//!
+//! Run with: GROQ_API_KEY=gsk_... cargo test -p openfang-kernel --test integration_test -- --nocapture
+
+use openfang_kernel::OpenFangKernel;
+use openfang_types::agent::AgentManifest;
+use openfang_types::config::{DefaultModelConfig, KernelConfig};
+
+fn test_config() -> KernelConfig {
+    let tmp = std::env::temp_dir().join("openfang-integration-test");
+    let _ = std::fs::remove_dir_all(&tmp);
+    std::fs::create_dir_all(&tmp).unwrap();
+
+    KernelConfig {
+        home_dir: tmp.clone(),
+        data_dir: tmp.join("data"),
+        default_model: DefaultModelConfig {
+            provider: "groq".to_string(),
+            model: "llama-3.3-70b-versatile".to_string(),
+            api_key_env: "GROQ_API_KEY".to_string(),
+            base_url: None,
+        },
+        ..KernelConfig::default()
+    }
+}
+
+#[tokio::test]
+async fn test_full_pipeline_with_groq() {
+    if std::env::var("GROQ_API_KEY").is_err() {
+        eprintln!("GROQ_API_KEY not set, skipping integration test");
+        return;
+    }
+
+    // Boot kernel
+    let config = test_config();
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    // Spawn agent
+    let manifest: AgentManifest = toml::from_str(
+        r#"
+name = "test-agent"
+version = "0.1.0"
+description = "Integration test agent"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are a test agent. Reply concisely in one sentence."
+
+[capabilities]
+tools = ["file_read"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+
+    let agent_id = kernel.spawn_agent(manifest).expect("Agent should spawn");
+
+    // Send message
+    let result = kernel
+        .send_message(agent_id, "Say hello in exactly 5 words.")
+        .await
+        .expect("Message should get a response");
+
+    println!("\n=== AGENT RESPONSE ===");
+    println!("{}", result.response);
+    println!(
+        "=== USAGE: {} tokens in, {} tokens out, {} iterations ===",
+        result.total_usage.input_tokens, result.total_usage.output_tokens, result.iterations
+    );
+
+    assert!(!result.response.is_empty(), "Response should not be empty");
+    assert!(
+        result.total_usage.input_tokens > 0,
+        "Should have used tokens"
+    );
+
+    // Kill agent
+    kernel.kill_agent(agent_id).expect("Agent should be killed");
+    kernel.shutdown();
+}
+
+#[tokio::test]
+async fn test_multiple_agents_different_models() {
+    if std::env::var("GROQ_API_KEY").is_err() {
+        eprintln!("GROQ_API_KEY not set, skipping integration test");
+        return;
+    }
+
+    let config = test_config();
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    // Spawn agent 1: llama 70b
+    let manifest1: AgentManifest = toml::from_str(
+        r#"
+name = "agent-llama70b"
+version = "0.1.0"
+description = "Llama 70B agent"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are Agent A. Always start your reply with 'A:'."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+
+    // Spawn agent 2: llama 8b (faster, smaller)
+    let manifest2: AgentManifest = toml::from_str(
+        r#"
+name = "agent-llama8b"
+version = "0.1.0"
+description = "Llama 8B agent"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "groq"
+model = "llama-3.1-8b-instant"
+system_prompt = "You are Agent B. Always start your reply with 'B:'."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+
+    let id1 = kernel.spawn_agent(manifest1).expect("Agent 1 should spawn");
+    let id2 = kernel.spawn_agent(manifest2).expect("Agent 2 should spawn");
+
+    // Send messages to both
+    let r1 = kernel
+        .send_message(id1, "What model are you?")
+        .await
+        .expect("Agent 1 response");
+    let r2 = kernel
+        .send_message(id2, "What model are you?")
+        .await
+        .expect("Agent 2 response");
+
+    println!("\n=== AGENT 1 (llama-70b) ===");
+    println!("{}", r1.response);
+    println!("\n=== AGENT 2 (llama-8b) ===");
+    println!("{}", r2.response);
+
+    assert!(!r1.response.is_empty());
+    assert!(!r2.response.is_empty());
+
+    // Cleanup
+    kernel.kill_agent(id1).unwrap();
+    kernel.kill_agent(id2).unwrap();
+    kernel.shutdown();
+}

crates/openfang-kernel/tests/multi_agent_test.rs

@@ -0,0 +1,201 @@
+//! Multi-agent integration test: spawn 6 agents, send messages, verify all respond.
+//!
+//! Run with: GROQ_API_KEY=gsk_... cargo test -p openfang-kernel --test multi_agent_test -- --nocapture
+
+use openfang_kernel::OpenFangKernel;
+use openfang_types::agent::AgentManifest;
+use openfang_types::config::{DefaultModelConfig, KernelConfig};
+
+fn test_config() -> KernelConfig {
+    let tmp = std::env::temp_dir().join("openfang-multi-agent-test");
+    let _ = std::fs::remove_dir_all(&tmp);
+    std::fs::create_dir_all(&tmp).unwrap();
+
+    KernelConfig {
+        home_dir: tmp.clone(),
+        data_dir: tmp.join("data"),
+        default_model: DefaultModelConfig {
+            provider: "groq".to_string(),
+            model: "llama-3.3-70b-versatile".to_string(),
+            api_key_env: "GROQ_API_KEY".to_string(),
+            base_url: None,
+        },
+        ..KernelConfig::default()
+    }
+}
+
+fn load_manifest(toml_str: &str) -> AgentManifest {
+    toml::from_str(toml_str).expect("Should parse manifest")
+}
+
+#[tokio::test]
+async fn test_six_agent_fleet() {
+    if std::env::var("GROQ_API_KEY").is_err() {
+        eprintln!("GROQ_API_KEY not set, skipping multi-agent test");
+        return;
+    }
+
+    let kernel = OpenFangKernel::boot_with_config(test_config()).expect("Kernel should boot");
+
+    // Define all 6 agents with different roles and models
+    let agents = vec![
+        (
+            "coder",
+            r#"
+name = "coder"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are Coder. Reply with 'CODER:' prefix. Be concise."
+[capabilities]
+tools = ["file_read", "file_write"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "Write a one-line Rust function that adds two numbers.",
+        ),
+        (
+            "researcher",
+            r#"
+name = "researcher"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are Researcher. Reply with 'RESEARCHER:' prefix. Be concise."
+[capabilities]
+tools = ["web_fetch"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "What is Rust's primary advantage over C++? One sentence.",
+        ),
+        (
+            "writer",
+            r#"
+name = "writer"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are Writer. Reply with 'WRITER:' prefix. Be concise."
+[capabilities]
+tools = ["file_read", "file_write"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "Write a one-sentence tagline for an Agent Operating System.",
+        ),
+        (
+            "ops",
+            r#"
+name = "ops"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.1-8b-instant"
+system_prompt = "You are Ops. Reply with 'OPS:' prefix. Be concise."
+[capabilities]
+tools = ["shell_exec"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "What would you check first if a server is running slowly?",
+        ),
+        (
+            "analyst",
+            r#"
+name = "analyst"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are Analyst. Reply with 'ANALYST:' prefix. Be concise."
+[capabilities]
+tools = ["file_read"]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "What are the top 3 metrics to track for an API service?",
+        ),
+        (
+            "hello-world",
+            r#"
+name = "hello-world"
+module = "builtin:chat"
+[model]
+provider = "groq"
+model = "llama-3.1-8b-instant"
+system_prompt = "You are a friendly greeter. Reply with 'HELLO:' prefix. Be concise."
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+            "Greet the user in a fun way.",
+        ),
+    ];
+
+    println!("\n{}", "=".repeat(60));
+    println!("  OPENFANG MULTI-AGENT FLEET TEST");
+    println!("  Spawning {} agents...", agents.len());
+    println!("{}\n", "=".repeat(60));
+
+    // Spawn all agents
+    let mut agent_ids = Vec::new();
+    for (name, manifest_str, _) in &agents {
+        let manifest = load_manifest(manifest_str);
+        let id = kernel
+            .spawn_agent(manifest)
+            .unwrap_or_else(|e| panic!("Failed to spawn {name}: {e}"));
+        println!("  Spawned: {name:<12} -> {id}");
+        agent_ids.push(id);
+    }
+
+    assert_eq!(kernel.registry.count(), 6, "Should have 6 agents");
+    println!(
+        "\n  All {} agents spawned. Sending messages...\n",
+        agents.len()
+    );
+
+    // Send messages to each agent sequentially (to respect Groq rate limits)
+    let mut results = Vec::new();
+    for (i, (name, _, message)) in agents.iter().enumerate() {
+        let result = kernel
+            .send_message(agent_ids[i], message)
+            .await
+            .unwrap_or_else(|e| panic!("Failed to message {name}: {e}"));
+
+        println!("--- {name} ---");
+        println!("  Q: {message}");
+        println!("  A: {}", result.response);
+        println!(
+            "  [{} tokens in, {} tokens out, {} iters]",
+            result.total_usage.input_tokens, result.total_usage.output_tokens, result.iterations
+        );
+        println!();
+
+        assert!(
+            !result.response.is_empty(),
+            "{name} response should not be empty"
+        );
+        results.push(result);
+    }
+
+    // Summary
+    let total_input: u64 = results.iter().map(|r| r.total_usage.input_tokens).sum();
+    let total_output: u64 = results.iter().map(|r| r.total_usage.output_tokens).sum();
+    println!("============================================================");
+    println!("  FLEET SUMMARY");
+    println!("  Agents:       {}", agents.len());
+    println!("  Total input:  {} tokens", total_input);
+    println!("  Total output: {} tokens", total_output);
+    println!("  All responded: YES");
+    println!("============================================================");
+
+    // Cleanup
+    for id in agent_ids {
+        kernel.kill_agent(id).unwrap();
+    }
+    kernel.shutdown();
+}

crates/openfang-kernel/tests/wasm_agent_integration_test.rs

@@ -0,0 +1,410 @@
+//! WASM agent integration tests.
+//!
+//! Tests the full pipeline: boot kernel → spawn agent with `module = "wasm:..."`
+//! → send message → verify WASM module executes and returns response.
+//!
+//! These tests use real WASM execution — no mocks.
+
+use openfang_kernel::OpenFangKernel;
+use openfang_types::agent::AgentManifest;
+use openfang_types::config::{DefaultModelConfig, KernelConfig};
+use std::sync::Arc;
+
+/// Minimal echo module: returns input JSON wrapped as `{"response": "..."}`.
+///
+/// Reads the "message" field from input and echoes it back as the response.
+/// Since WAT can't do real string manipulation, this module echoes the
+/// entire input JSON as-is (which the kernel extracts via serde).
+const ECHO_WAT: &str = r#"
+    (module
+        (memory (export "memory") 1)
+        (global $bump (mut i32) (i32.const 1024))
+
+        (func (export "alloc") (param $size i32) (result i32)
+            (local $ptr i32)
+            (local.set $ptr (global.get $bump))
+            (global.set $bump (i32.add (global.get $bump) (local.get $size)))
+            (local.get $ptr)
+        )
+
+        (func (export "execute") (param $ptr i32) (param $len i32) (result i64)
+            ;; Echo: return the input as-is (kernel will extract from JSON)
+            (i64.or
+                (i64.shl
+                    (i64.extend_i32_u (local.get $ptr))
+                    (i64.const 32)
+                )
+                (i64.extend_i32_u (local.get $len))
+            )
+        )
+    )
+"#;
+
+/// Module that always returns a fixed JSON response.
+/// Writes `{"response":"hello from wasm"}` at offset 0 and returns it.
+const HELLO_WAT: &str = r#"
+    (module
+        (memory (export "memory") 1)
+        (global $bump (mut i32) (i32.const 4096))
+
+        ;; Fixed response bytes: {"response":"hello from wasm"}
+        (data (i32.const 0) "{\"response\":\"hello from wasm\"}")
+
+        (func (export "alloc") (param $size i32) (result i32)
+            (local $ptr i32)
+            (local.set $ptr (global.get $bump))
+            (global.set $bump (i32.add (global.get $bump) (local.get $size)))
+            (local.get $ptr)
+        )
+
+        (func (export "execute") (param $ptr i32) (param $len i32) (result i64)
+            ;; Return pointer=0, length=30 (the fixed response)
+            (i64.const 30)  ;; low 32 = len=30, high 32 = ptr=0
+        )
+    )
+"#;
+
+/// Module with infinite loop — tests fuel exhaustion enforcement.
+const INFINITE_LOOP_WAT: &str = r#"
+    (module
+        (memory (export "memory") 1)
+        (global $bump (mut i32) (i32.const 1024))
+
+        (func (export "alloc") (param $size i32) (result i32)
+            (local $ptr i32)
+            (local.set $ptr (global.get $bump))
+            (global.set $bump (i32.add (global.get $bump) (local.get $size)))
+            (local.get $ptr)
+        )
+
+        (func (export "execute") (param $ptr i32) (param $len i32) (result i64)
+            (loop $inf
+                (br $inf)
+            )
+            (i64.const 0)
+        )
+    )
+"#;
+
+/// Host-call proxy: forwards input to host_call and returns the response.
+const HOST_CALL_PROXY_WAT: &str = r#"
+    (module
+        (import "openfang" "host_call" (func $host_call (param i32 i32) (result i64)))
+        (memory (export "memory") 2)
+        (global $bump (mut i32) (i32.const 1024))
+
+        (func (export "alloc") (param $size i32) (result i32)
+            (local $ptr i32)
+            (local.set $ptr (global.get $bump))
+            (global.set $bump (i32.add (global.get $bump) (local.get $size)))
+            (local.get $ptr)
+        )
+
+        (func (export "execute") (param $input_ptr i32) (param $input_len i32) (result i64)
+            (call $host_call (local.get $input_ptr) (local.get $input_len))
+        )
+    )
+"#;
+
+fn test_config(tmp: &tempfile::TempDir) -> KernelConfig {
+    KernelConfig {
+        home_dir: tmp.path().to_path_buf(),
+        data_dir: tmp.path().join("data"),
+        default_model: DefaultModelConfig {
+            provider: "ollama".to_string(),
+            model: "test".to_string(),
+            api_key_env: "OLLAMA_API_KEY".to_string(),
+            base_url: None,
+        },
+        ..KernelConfig::default()
+    }
+}
+
+fn wasm_manifest(name: &str, module: &str) -> AgentManifest {
+    let toml_str = format!(
+        r#"
+name = "{name}"
+version = "0.1.0"
+description = "WASM test agent"
+author = "test"
+module = "wasm:{module}"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "WASM agent."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#
+    );
+    toml::from_str(&toml_str).unwrap()
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+/// Test that a WASM agent can be spawned and returns a response.
+#[tokio::test]
+async fn test_wasm_agent_hello_response() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("hello.wat"), HELLO_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest = wasm_manifest("wasm-hello", "hello.wat");
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let result = kernel
+        .send_message(agent_id, "Hi there!")
+        .await
+        .expect("WASM agent should execute");
+
+    assert_eq!(result.response, "hello from wasm");
+    assert_eq!(result.iterations, 1);
+
+    kernel.shutdown();
+}
+
+/// Test that a WASM echo module returns input data.
+#[tokio::test]
+async fn test_wasm_agent_echo() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("echo.wat"), ECHO_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest = wasm_manifest("wasm-echo", "echo.wat");
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let result = kernel
+        .send_message(agent_id, "test message")
+        .await
+        .expect("Echo agent should execute");
+
+    // Echo returns the entire input JSON, so the response should contain our message
+    assert!(
+        result.response.contains("test message"),
+        "Response should contain the input message, got: {}",
+        result.response
+    );
+
+    kernel.shutdown();
+}
+
+/// Test that WASM fuel exhaustion is caught and reported as an error.
+#[tokio::test]
+async fn test_wasm_agent_fuel_exhaustion() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("loop.wat"), INFINITE_LOOP_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest = wasm_manifest("wasm-loop", "loop.wat");
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let result = kernel.send_message(agent_id, "go").await;
+    assert!(
+        result.is_err(),
+        "Infinite loop should fail with fuel exhaustion"
+    );
+    let err_msg = format!("{}", result.unwrap_err());
+    assert!(
+        err_msg.contains("Fuel exhausted") || err_msg.contains("fuel") || err_msg.contains("WASM"),
+        "Error should mention fuel exhaustion, got: {err_msg}"
+    );
+
+    kernel.shutdown();
+}
+
+/// Test that a missing WASM module produces a clear error.
+#[tokio::test]
+async fn test_wasm_agent_missing_module() {
+    let tmp = tempfile::tempdir().unwrap();
+    // Don't write any .wat file
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest = wasm_manifest("wasm-missing", "nonexistent.wasm");
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let result = kernel.send_message(agent_id, "hello").await;
+    assert!(result.is_err(), "Missing module should fail");
+    let err_msg = format!("{}", result.unwrap_err());
+    assert!(
+        err_msg.contains("Failed to read") || err_msg.contains("nonexistent"),
+        "Error should mention the missing file, got: {err_msg}"
+    );
+
+    kernel.shutdown();
+}
+
+/// Test that host_call time_now works end-to-end through the kernel.
+#[tokio::test]
+async fn test_wasm_agent_host_call_time() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("proxy.wat"), HOST_CALL_PROXY_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    // Proxy module forwards input to host_call — send a time_now request
+    let toml_str = r#"
+name = "wasm-proxy"
+version = "0.1.0"
+description = "Host call proxy"
+author = "test"
+module = "wasm:proxy.wat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "Proxy."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#;
+    let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    // The proxy module expects JSON like {"method":"time_now","params":{}}
+    // But our kernel wraps it as {"message":"...", "agent_id":"...", "agent_name":"..."}
+    // So the proxy will try to dispatch with method=null which returns "Unknown"
+    // This still proves the full pipeline works end-to-end
+    let result = kernel
+        .send_message(agent_id, r#"{"method":"time_now","params":{}}"#)
+        .await
+        .expect("Proxy agent should execute");
+
+    // The response will contain the host_call dispatch result
+    assert!(!result.response.is_empty(), "Response should not be empty");
+
+    kernel.shutdown();
+}
+
+/// Test WASM agent with streaming (falls back to single event).
+#[tokio::test]
+async fn test_wasm_agent_streaming_fallback() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("hello.wat"), HELLO_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+    let kernel = Arc::new(kernel);
+
+    let manifest = wasm_manifest("wasm-stream", "hello.wat");
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let (mut rx, handle) = kernel
+        .send_message_streaming(agent_id, "Hi!", None)
+        .expect("Streaming should start");
+
+    // Collect all stream events
+    let mut events = vec![];
+    while let Some(event) = rx.recv().await {
+        events.push(event);
+    }
+
+    // Should have gotten a TextDelta + ContentComplete
+    assert!(
+        events.len() >= 2,
+        "Expected at least 2 stream events, got {}",
+        events.len()
+    );
+
+    let final_result = handle.await.unwrap().expect("Task should complete");
+    assert_eq!(final_result.response, "hello from wasm");
+
+    kernel.shutdown();
+}
+
+/// Test that spawning multiple WASM agents works concurrently.
+#[tokio::test]
+async fn test_multiple_wasm_agents() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("hello.wat"), HELLO_WAT).unwrap();
+    std::fs::write(tmp.path().join("echo.wat"), ECHO_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let hello_id = kernel
+        .spawn_agent(wasm_manifest("hello-agent", "hello.wat"))
+        .unwrap();
+    let echo_id = kernel
+        .spawn_agent(wasm_manifest("echo-agent", "echo.wat"))
+        .unwrap();
+
+    // Execute both
+    let hello_result = kernel.send_message(hello_id, "hi").await.unwrap();
+    let echo_result = kernel.send_message(echo_id, "test data").await.unwrap();
+
+    assert_eq!(hello_result.response, "hello from wasm");
+    assert!(echo_result.response.contains("test data"));
+
+    // Verify agent list shows both
+    let agents = kernel.registry.list();
+    assert_eq!(agents.len(), 2);
+
+    kernel.shutdown();
+}
+
+/// Test WASM agent alongside LLM agent (mixed fleet).
+#[tokio::test]
+async fn test_mixed_wasm_and_llm_agents() {
+    let tmp = tempfile::tempdir().unwrap();
+    std::fs::write(tmp.path().join("hello.wat"), HELLO_WAT).unwrap();
+
+    let config = test_config(&tmp);
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    // Spawn a WASM agent
+    let wasm_id = kernel
+        .spawn_agent(wasm_manifest("wasm-agent", "hello.wat"))
+        .unwrap();
+
+    // Spawn a regular LLM agent (won't actually call LLM since ollama isn't running,
+    // but it should spawn fine and coexist)
+    let llm_toml = r#"
+name = "llm-agent"
+version = "0.1.0"
+description = "LLM test agent"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "You are a test agent."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#;
+    let llm_manifest: AgentManifest = toml::from_str(llm_toml).unwrap();
+    let llm_id = kernel.spawn_agent(llm_manifest).unwrap();
+
+    // Verify both agents exist
+    let agents = kernel.registry.list();
+    assert_eq!(agents.len(), 2);
+
+    // WASM agent should work
+    let result = kernel.send_message(wasm_id, "hello").await.unwrap();
+    assert_eq!(result.response, "hello from wasm");
+
+    // LLM agent exists but we won't send it a message (no real LLM)
+    assert!(kernel.registry.get(llm_id).is_some());
+
+    // Kill WASM agent
+    kernel.kill_agent(wasm_id).unwrap();
+    assert_eq!(kernel.registry.list().len(), 1);
+
+    kernel.shutdown();
+}

crates/openfang-kernel/tests/workflow_integration_test.rs

@@ -0,0 +1,404 @@
+//! End-to-end workflow integration tests.
+//!
+//! Tests the full pipeline: boot kernel → spawn agents → create workflow →
+//! execute workflow → verify outputs flow through the pipeline.
+//!
+//! LLM tests require GROQ_API_KEY. Non-LLM tests verify the kernel-level
+//! workflow wiring without making real API calls.
+
+use openfang_kernel::workflow::{
+    ErrorMode, StepAgent, StepMode, Workflow, WorkflowId, WorkflowStep,
+};
+use openfang_kernel::OpenFangKernel;
+use openfang_types::agent::AgentManifest;
+use openfang_types::config::{DefaultModelConfig, KernelConfig};
+use std::sync::Arc;
+
+fn test_config(provider: &str, model: &str, api_key_env: &str) -> KernelConfig {
+    let tmp = tempfile::tempdir().unwrap();
+    KernelConfig {
+        home_dir: tmp.path().to_path_buf(),
+        data_dir: tmp.path().join("data"),
+        default_model: DefaultModelConfig {
+            provider: provider.to_string(),
+            model: model.to_string(),
+            api_key_env: api_key_env.to_string(),
+            base_url: None,
+        },
+        ..KernelConfig::default()
+    }
+}
+
+fn spawn_test_agent(
+    kernel: &OpenFangKernel,
+    name: &str,
+    system_prompt: &str,
+) -> openfang_types::agent::AgentId {
+    let manifest_str = format!(
+        r#"
+name = "{name}"
+version = "0.1.0"
+description = "Workflow test agent: {name}"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "{system_prompt}"
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#
+    );
+    let manifest: AgentManifest = toml::from_str(&manifest_str).unwrap();
+    kernel.spawn_agent(manifest).expect("Agent should spawn")
+}
+
+// ---------------------------------------------------------------------------
+// Kernel-level workflow wiring tests (no LLM needed)
+// ---------------------------------------------------------------------------
+
+/// Test that workflow registration and agent resolution work at the kernel level.
+#[tokio::test]
+async fn test_workflow_register_and_resolve() {
+    let config = test_config("ollama", "test-model", "OLLAMA_API_KEY");
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+    let kernel = Arc::new(kernel);
+
+    // Spawn agents
+    let manifest: AgentManifest = toml::from_str(
+        r#"
+name = "agent-alpha"
+version = "0.1.0"
+description = "Alpha"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "Alpha."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+    let alpha_id = kernel.spawn_agent(manifest).unwrap();
+
+    let manifest2: AgentManifest = toml::from_str(
+        r#"
+name = "agent-beta"
+version = "0.1.0"
+description = "Beta"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "Beta."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+    let beta_id = kernel.spawn_agent(manifest2).unwrap();
+
+    // Create a 2-step workflow referencing agents by name
+    let workflow = Workflow {
+        id: WorkflowId::new(),
+        name: "alpha-beta-pipeline".to_string(),
+        description: "Tests agent resolution by name".to_string(),
+        steps: vec![
+            WorkflowStep {
+                name: "step-alpha".to_string(),
+                agent: StepAgent::ByName {
+                    name: "agent-alpha".to_string(),
+                },
+                prompt_template: "Analyze: {{input}}".to_string(),
+                mode: StepMode::Sequential,
+                timeout_secs: 30,
+                error_mode: ErrorMode::Fail,
+                output_var: Some("alpha_out".to_string()),
+            },
+            WorkflowStep {
+                name: "step-beta".to_string(),
+                agent: StepAgent::ByName {
+                    name: "agent-beta".to_string(),
+                },
+                prompt_template: "Summarize: {{input}} (alpha said: {{alpha_out}})".to_string(),
+                mode: StepMode::Sequential,
+                timeout_secs: 30,
+                error_mode: ErrorMode::Fail,
+                output_var: None,
+            },
+        ],
+        created_at: chrono::Utc::now(),
+    };
+
+    let wf_id = kernel.register_workflow(workflow).await;
+
+    // Verify workflow is registered
+    let workflows = kernel.workflows.list_workflows().await;
+    assert_eq!(workflows.len(), 1);
+    assert_eq!(workflows[0].name, "alpha-beta-pipeline");
+
+    // Verify agents can be found by name
+    let alpha = kernel.registry.find_by_name("agent-alpha");
+    assert!(alpha.is_some());
+    assert_eq!(alpha.unwrap().id, alpha_id);
+
+    let beta = kernel.registry.find_by_name("agent-beta");
+    assert!(beta.is_some());
+    assert_eq!(beta.unwrap().id, beta_id);
+
+    // Verify workflow run can be created
+    let run_id = kernel
+        .workflows
+        .create_run(wf_id, "test input".to_string())
+        .await;
+    assert!(run_id.is_some());
+
+    let run = kernel.workflows.get_run(run_id.unwrap()).await.unwrap();
+    assert_eq!(run.input, "test input");
+
+    kernel.shutdown();
+}
+
+/// Test workflow with agent referenced by ID.
+#[tokio::test]
+async fn test_workflow_agent_by_id() {
+    let config = test_config("ollama", "test-model", "OLLAMA_API_KEY");
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest: AgentManifest = toml::from_str(
+        r#"
+name = "id-agent"
+version = "0.1.0"
+description = "Test"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "Test."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    let workflow = Workflow {
+        id: WorkflowId::new(),
+        name: "by-id-test".to_string(),
+        description: "".to_string(),
+        steps: vec![WorkflowStep {
+            name: "step1".to_string(),
+            agent: StepAgent::ById {
+                id: agent_id.to_string(),
+            },
+            prompt_template: "{{input}}".to_string(),
+            mode: StepMode::Sequential,
+            timeout_secs: 30,
+            error_mode: ErrorMode::Fail,
+            output_var: None,
+        }],
+        created_at: chrono::Utc::now(),
+    };
+
+    let wf_id = kernel.register_workflow(workflow).await;
+
+    // Can create run (agent resolution happens at execute time)
+    let run_id = kernel
+        .workflows
+        .create_run(wf_id, "hello".to_string())
+        .await;
+    assert!(run_id.is_some());
+
+    kernel.shutdown();
+}
+
+/// Test trigger registration and listing at kernel level.
+#[tokio::test]
+async fn test_trigger_registration_with_kernel() {
+    use openfang_kernel::triggers::TriggerPattern;
+
+    let config = test_config("ollama", "test-model", "OLLAMA_API_KEY");
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+
+    let manifest: AgentManifest = toml::from_str(
+        r#"
+name = "trigger-agent"
+version = "0.1.0"
+description = "Trigger test"
+author = "test"
+module = "builtin:chat"
+
+[model]
+provider = "ollama"
+model = "test"
+system_prompt = "Test."
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#,
+    )
+    .unwrap();
+    let agent_id = kernel.spawn_agent(manifest).unwrap();
+
+    // Register triggers
+    let t1 = kernel
+        .register_trigger(
+            agent_id,
+            TriggerPattern::Lifecycle,
+            "Lifecycle event: {{event}}".to_string(),
+            0,
+        )
+        .unwrap();
+
+    let t2 = kernel
+        .register_trigger(
+            agent_id,
+            TriggerPattern::SystemKeyword {
+                keyword: "deploy".to_string(),
+            },
+            "Deploy event: {{event}}".to_string(),
+            5,
+        )
+        .unwrap();
+
+    // List all triggers
+    let all = kernel.list_triggers(None);
+    assert_eq!(all.len(), 2);
+
+    // List triggers for specific agent
+    let agent_triggers = kernel.list_triggers(Some(agent_id));
+    assert_eq!(agent_triggers.len(), 2);
+
+    // Remove one
+    assert!(kernel.remove_trigger(t1));
+    let remaining = kernel.list_triggers(None);
+    assert_eq!(remaining.len(), 1);
+    assert_eq!(remaining[0].id, t2);
+
+    kernel.shutdown();
+}
+
+// ---------------------------------------------------------------------------
+// Full E2E with real LLM (skip if no GROQ_API_KEY)
+// ---------------------------------------------------------------------------
+
+/// End-to-end: boot kernel → spawn 2 agents → create 2-step workflow →
+/// run it through the real Groq LLM → verify output flows from step 1 to step 2.
+#[tokio::test]
+async fn test_workflow_e2e_with_groq() {
+    if std::env::var("GROQ_API_KEY").is_err() {
+        eprintln!("GROQ_API_KEY not set, skipping E2E workflow test");
+        return;
+    }
+
+    let config = test_config("groq", "llama-3.3-70b-versatile", "GROQ_API_KEY");
+    let kernel = OpenFangKernel::boot_with_config(config).expect("Kernel should boot");
+    let kernel = Arc::new(kernel);
+    kernel.set_self_handle();
+
+    // Spawn two agents with distinct roles
+    let _analyst_id = spawn_test_agent(
+        &kernel,
+        "wf-analyst",
+        "You are an analyst. When given text, respond with exactly: ANALYSIS: followed by a one-sentence analysis.",
+    );
+    let _writer_id = spawn_test_agent(
+        &kernel,
+        "wf-writer",
+        "You are a writer. When given text, respond with exactly: SUMMARY: followed by a one-sentence summary.",
+    );
+
+    // Create a 2-step pipeline: analyst → writer
+    let workflow = Workflow {
+        id: WorkflowId::new(),
+        name: "analyst-writer-pipeline".to_string(),
+        description: "E2E integration test workflow".to_string(),
+        steps: vec![
+            WorkflowStep {
+                name: "analyze".to_string(),
+                agent: StepAgent::ByName {
+                    name: "wf-analyst".to_string(),
+                },
+                prompt_template: "Analyze the following: {{input}}".to_string(),
+                mode: StepMode::Sequential,
+                timeout_secs: 60,
+                error_mode: ErrorMode::Fail,
+                output_var: None,
+            },
+            WorkflowStep {
+                name: "summarize".to_string(),
+                agent: StepAgent::ByName {
+                    name: "wf-writer".to_string(),
+                },
+                prompt_template: "Summarize this analysis: {{input}}".to_string(),
+                mode: StepMode::Sequential,
+                timeout_secs: 60,
+                error_mode: ErrorMode::Fail,
+                output_var: None,
+            },
+        ],
+        created_at: chrono::Utc::now(),
+    };
+
+    let wf_id = kernel.register_workflow(workflow).await;
+
+    // Run the workflow
+    let result = kernel
+        .run_workflow(
+            wf_id,
+            "The Rust programming language is growing rapidly.".to_string(),
+        )
+        .await;
+
+    assert!(
+        result.is_ok(),
+        "Workflow should complete: {:?}",
+        result.err()
+    );
+    let (run_id, output) = result.unwrap();
+
+    println!("\n=== WORKFLOW OUTPUT ===");
+    println!("{output}");
+    println!("======================\n");
+
+    assert!(!output.is_empty(), "Workflow output should not be empty");
+
+    // Verify the workflow run record
+    let run = kernel.workflows.get_run(run_id).await.unwrap();
+    assert!(matches!(
+        run.state,
+        openfang_kernel::workflow::WorkflowRunState::Completed
+    ));
+    assert_eq!(run.step_results.len(), 2);
+    assert_eq!(run.step_results[0].step_name, "analyze");
+    assert_eq!(run.step_results[1].step_name, "summarize");
+
+    // Both steps should have used tokens
+    assert!(run.step_results[0].input_tokens > 0);
+    assert!(run.step_results[0].output_tokens > 0);
+    assert!(run.step_results[1].input_tokens > 0);
+    assert!(run.step_results[1].output_tokens > 0);
+
+    // List runs
+    let runs = kernel.workflows.list_runs(None).await;
+    assert_eq!(runs.len(), 1);
+
+    kernel.shutdown();
+}