Merge branch 'staging' into feat/custom-llm-provider

.github/workflows/test.yml

@@ -12,6 +12,7 @@ jobs:
   tests:
     name: Tests (${{ matrix.name }})
     runs-on: ubuntu-latest
+    timeout-minutes: 45
     strategy:
       fail-fast: false
       matrix:
@@ -40,11 +41,14 @@ jobs:
       - name: Build WASM channels (for integration tests)
         run: ./scripts/build-wasm-extensions.sh --channels
       - name: Run Tests
-        run: cargo test ${{ matrix.flags }} -- --nocapture
+        run: |
+          timeout --signal=INT --kill-after=30s 40m \
+            cargo test ${{ matrix.flags }} -- --nocapture
 
   heavy-integration-tests:
     name: Heavy Integration Tests
     runs-on: ubuntu-latest
+    timeout-minutes: 20
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
@@ -58,9 +62,13 @@ jobs:
       - name: Build Telegram WASM channel
         run: cargo build --manifest-path channels-src/telegram/Cargo.toml --target wasm32-wasip2 --release
       - name: Run thread scheduling integration tests
-        run: cargo test --no-default-features --features libsql,integration --test e2e_thread_scheduling -- --nocapture
+        run: |
+          timeout --signal=INT --kill-after=30s 15m \
+            cargo test --no-default-features --features libsql,integration --test e2e_thread_scheduling -- --nocapture
       - name: Run Telegram thread-scope regression test
-        run: cargo test --features integration --test telegram_auth_integration test_private_messages_use_chat_id_as_thread_scope -- --exact
+        run: |
+          timeout --signal=INT --kill-after=30s 10m \
+            cargo test --features integration --test telegram_auth_integration test_private_messages_use_chat_id_as_thread_scope -- --exact
 
   telegram-tests:
     name: Telegram Channel Tests
@@ -68,6 +76,7 @@ jobs:
       github.event_name != 'pull_request' ||
       github.base_ref != 'staging'
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
@@ -75,7 +84,9 @@ jobs:
         uses: dtolnay/rust-toolchain@stable
       - uses: Swatinem/rust-cache@v2
       - name: Run Telegram Channel Tests
-        run: cargo test --manifest-path channels-src/telegram/Cargo.toml -- --nocapture
+        run: |
+          timeout --signal=INT --kill-after=30s 10m \
+            cargo test --manifest-path channels-src/telegram/Cargo.toml -- --nocapture
 
   windows-build:
     name: Windows Build (${{ matrix.name }})
@@ -110,6 +121,7 @@ jobs:
       github.event_name != 'pull_request' ||
       github.base_ref != 'staging'
     runs-on: ubuntu-latest
+    timeout-minutes: 30
     steps:
       - name: Checkout repository
         uses: actions/checkout@v6
@@ -125,7 +137,9 @@ jobs:
       - name: Build all WASM extensions against current WIT
         run: ./scripts/build-wasm-extensions.sh
       - name: Instantiation test (host linker compatibility)
-        run: cargo test --all-features wit_compat -- --nocapture
+        run: |
+          timeout --signal=INT --kill-after=30s 20m \
+            cargo test --all-features wit_compat -- --nocapture
 
   bench-compile:
     name: Benchmark Compilation

crates/ironclaw_common/src/event.rs

@@ -7,6 +7,32 @@
 
 use serde::{Deserialize, Serialize};
 
+/// A single tool decision in a reasoning update (SSE DTO).
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ToolDecisionDto {
+    pub tool_name: String,
+    pub rationale: String,
+}
+
+impl ToolDecisionDto {
+    /// Parse a list of tool decisions from a JSON array value.
+    pub fn from_json_array(value: &serde_json::Value) -> Vec<Self> {
+        value
+            .as_array()
+            .map(|arr| {
+                arr.iter()
+                    .filter_map(|d| {
+                        Some(Self {
+                            tool_name: d.get("tool_name")?.as_str()?.to_string(),
+                            rationale: d.get("rationale")?.as_str()?.to_string(),
+                        })
+                    })
+                    .collect()
+            })
+            .unwrap_or_default()
+    }
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(tag = "type")]
 pub enum AppEvent {
@@ -163,6 +189,23 @@ pub enum AppEvent {
         #[serde(skip_serializing_if = "Option::is_none")]
         message: Option<String>,
     },
+
+    /// Agent reasoning update (why it chose specific tools).
+    #[serde(rename = "reasoning_update")]
+    ReasoningUpdate {
+        narrative: String,
+        decisions: Vec<ToolDecisionDto>,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        thread_id: Option<String>,
+    },
+
+    /// Reasoning update for a sandbox job.
+    #[serde(rename = "job_reasoning")]
+    JobReasoning {
+        job_id: String,
+        narrative: String,
+        decisions: Vec<ToolDecisionDto>,
+    },
 }
 
 impl AppEvent {
@@ -191,6 +234,8 @@ impl AppEvent {
             Self::Suggestions { .. } => "suggestions",
             Self::TurnCost { .. } => "turn_cost",
             Self::ExtensionStatus { .. } => "extension_status",
+            Self::ReasoningUpdate { .. } => "reasoning_update",
+            Self::JobReasoning { .. } => "job_reasoning",
         }
     }
 }
@@ -311,6 +356,16 @@ mod tests {
                 status: String::new(),
                 message: None,
             },
+            AppEvent::ReasoningUpdate {
+                narrative: String::new(),
+                decisions: vec![],
+                thread_id: None,
+            },
+            AppEvent::JobReasoning {
+                job_id: String::new(),
+                narrative: String::new(),
+                decisions: vec![],
+            },
         ];
 
         for variant in &variants {

crates/ironclaw_common/src/lib.rs

@@ -3,5 +3,5 @@
 mod event;
 mod util;
 
-pub use event::AppEvent;
+pub use event::{AppEvent, ToolDecisionDto};
 pub use util::truncate_preview;

src/agent/agent_loop.rs

@@ -13,9 +13,10 @@ use futures::StreamExt;
 use uuid::Uuid;
 
 use crate::agent::context_monitor::ContextMonitor;
-use crate::agent::heartbeat::spawn_heartbeat;
+use crate::agent::heartbeat::{spawn_heartbeat, spawn_multi_user_heartbeat};
 use crate::agent::routine_engine::{RoutineEngine, spawn_cron_ticker};
 use crate::agent::self_repair::{DefaultSelfRepair, RepairResult, SelfRepair};
+use crate::agent::session::ThreadState;
 use crate::agent::session_manager::SessionManager;
 use crate::agent::submission::{Submission, SubmissionParser, SubmissionResult};
 use crate::agent::{HeartbeatConfig as AgentHeartbeatConfig, Router, Scheduler, SchedulerDeps};
@@ -84,6 +85,15 @@ fn resolve_owner_scope_notification_user(
     trimmed_option(explicit_user).or_else(|| trimmed_option(owner_fallback))
 }
 
+fn is_single_message_repl(message: &IncomingMessage) -> bool {
+    message.channel == "repl"
+        && message
+            .metadata
+            .get("single_message_mode")
+            .and_then(|value| value.as_bool())
+            .unwrap_or(false)
+}
+
 async fn resolve_channel_notification_user(
     extension_manager: Option<&Arc<ExtensionManager>>,
     channel: Option<&str>,
@@ -172,6 +182,8 @@ pub struct AgentDeps {
     /// Resolved LLM backend identifier (e.g., "nearai", "openai", "groq").
     /// Used by `/model` persistence to determine which env var to update.
     pub llm_backend: String,
+    /// Per-tenant rate limiting registry (lazily creates rate state per user).
+    pub tenant_rates: Arc<crate::tenant::TenantRateRegistry>,
 }
 
 /// The main agent that coordinates all components.
@@ -234,7 +246,10 @@ impl Agent {
             SchedulerDeps {
                 tools: deps.tools.clone(),
                 extension_manager: deps.extension_manager.clone(),
-                store: deps.store.clone(),
+                store: deps
+                    .store
+                    .as_ref()
+                    .map(|db| crate::tenant::AdminScope::new(Arc::clone(db))),
                 hooks: deps.hooks.clone(),
             },
         );
@@ -315,6 +330,50 @@ impl Agent {
         &self.deps.cost_guard
     }
 
+    /// Build a tenant-scoped execution context for the given user.
+    ///
+    /// This is the standard entry point for per-user operations. The returned
+    /// [`TenantCtx`] provides a [`TenantScope`] that auto-binds `user_id` on
+    /// every database operation and a per-user rate limiter.
+    pub(super) async fn tenant_ctx(&self, user_id: &str) -> crate::tenant::TenantCtx {
+        let rate = self.deps.tenant_rates.get_or_create(user_id).await;
+
+        let store = self
+            .deps
+            .store
+            .as_ref()
+            .map(|db| crate::tenant::TenantScope::new(user_id, Arc::clone(db)));
+
+        // Reuse the owner workspace if user matches, otherwise create per-user.
+        let workspace = match &self.deps.workspace {
+            Some(ws) if ws.user_id() == user_id => Some(Arc::clone(ws)),
+            _ => self
+                .deps
+                .store
+                .as_ref()
+                .map(|db| Arc::new(Workspace::new_with_db(user_id, Arc::clone(db)))),
+        };
+
+        crate::tenant::TenantCtx::new(
+            user_id,
+            store,
+            workspace,
+            Arc::clone(&self.deps.cost_guard),
+            rate,
+        )
+    }
+
+    /// Get an admin-scoped database accessor for cross-tenant operations.
+    ///
+    /// Only for system-level components (heartbeat, routine engine, self-repair,
+    /// scheduler). Handler code should use [`tenant_ctx()`](Self::tenant_ctx) instead.
+    pub(super) fn admin_store(&self) -> Option<crate::tenant::AdminScope> {
+        self.deps
+            .store
+            .as_ref()
+            .map(|db| crate::tenant::AdminScope::new(Arc::clone(db)))
+    }
+
     pub(super) fn skill_registry(&self) -> Option<&Arc<std::sync::RwLock<SkillRegistry>>> {
         self.deps.skill_registry.as_ref()
     }
@@ -400,8 +459,8 @@ impl Agent {
             self.config.stuck_threshold,
             self.config.max_repair_attempts,
         );
-        if let Some(ref store) = self.deps.store {
-            self_repair = self_repair.with_store(Arc::clone(store));
+        if let Some(admin) = self.admin_store() {
+            self_repair = self_repair.with_store(admin);
         }
         if let Some(ref builder) = self.deps.builder {
             self_repair = self_repair.with_builder(Arc::clone(builder), Arc::clone(self.tools()));
@@ -508,6 +567,7 @@ impl Agent {
                         .with_interval(std::time::Duration::from_secs(hb_config.interval_secs));
                     config.quiet_hours_start = hb_config.quiet_hours_start;
                     config.quiet_hours_end = hb_config.quiet_hours_end;
+                    config.multi_tenant = hb_config.multi_tenant;
                     config.timezone = hb_config
                         .timezone
                         .clone()
@@ -537,30 +597,52 @@ impl Agent {
                     .await;
                     let notify_user = heartbeat_notify_user;
                     let channels = self.channels.clone();
+                    let is_multi_tenant = hb_config.multi_tenant;
                     tokio::spawn(async move {
                         while let Some(response) = notify_rx.recv().await {
+                            // In multi-tenant mode, extract the owning user_id from
+                            // the response metadata so notifications reach the
+                            // correct user rather than the agent's owner.
+                            // This intentionally overrides the configured notify_target
+                            // because each user's heartbeat should notify that user.
+                            let effective_user = if is_multi_tenant {
+                                response
+                                    .metadata
+                                    .get("owner_id")
+                                    .and_then(|v| v.as_str())
+                                    .map(String::from)
+                            } else {
+                                None
+                            };
+
                             // Try the configured channel first, fall back to
                             // broadcasting on all channels.
-                            let targeted_ok = if let Some(ref channel) = notify_channel
-                                && let Some(ref user) = notify_target
-                            {
-                                channels
-                                    .broadcast(channel, user, response.clone())
-                                    .await
-                                    .is_ok()
+                            let targeted_ok = if let Some(ref channel) = notify_channel {
+                                let target = effective_user.as_deref().or(notify_target.as_deref());
+                                if let Some(user) = target {
+                                    channels
+                                        .broadcast(channel, user, response.clone())
+                                        .await
+                                        .is_ok()
+                                } else {
+                                    false
+                                }
                             } else {
                                 false
                             };
 
-                            if !targeted_ok && let Some(ref user) = notify_user {
-                                let results = channels.broadcast_all(user, response).await;
-                                for (ch, result) in results {
-                                    if let Err(e) = result {
-                                        tracing::warn!(
-                                            "Failed to broadcast heartbeat to {}: {}",
-                                            ch,
-                                            e
-                                        );
+                            if !targeted_ok {
+                                let fallback = effective_user.as_deref().or(notify_user.as_deref());
+                                if let Some(user) = fallback {
+                                    let results = channels.broadcast_all(user, response).await;
+                                    for (ch, result) in results {
+                                        if let Err(e) = result {
+                                            tracing::warn!(
+                                                "Failed to broadcast heartbeat to {}: {}",
+                                                ch,
+                                                e
+                                            );
+                                        }
                                     }
                                 }
                             }
@@ -573,14 +655,29 @@ impl Agent {
                         .map(|h| h.to_workspace_config())
                         .unwrap_or_default();
 
-                    Some(spawn_heartbeat(
-                        config,
-                        hygiene,
-                        workspace.clone(),
-                        self.cheap_llm().clone(),
-                        Some(notify_tx),
-                        self.store().map(Arc::clone),
-                    ))
+                    if config.multi_tenant {
+                        if let Some(admin) = self.admin_store() {
+                            Some(spawn_multi_user_heartbeat(
+                                config,
+                                hygiene,
+                                self.cheap_llm().clone(),
+                                Some(notify_tx),
+                                admin,
+                            ))
+                        } else {
+                            tracing::warn!("Multi-tenant heartbeat requires a database store");
+                            None
+                        }
+                    } else {
+                        Some(spawn_heartbeat(
+                            config,
+                            hygiene,
+                            workspace.clone(),
+                            self.cheap_llm().clone(),
+                            Some(notify_tx),
+                            self.admin_store(),
+                        ))
+                    }
                 } else {
                     tracing::warn!("Heartbeat enabled but no workspace available");
                     None
@@ -602,7 +699,7 @@ impl Agent {
 
                     let engine = Arc::new(RoutineEngine::new(
                         rt_config.clone(),
-                        Arc::clone(store),
+                        crate::tenant::AdminScope::new(Arc::clone(store)),
                         self.llm().clone(),
                         Arc::clone(workspace),
                         notify_tx,
@@ -1140,9 +1237,14 @@ impl Agent {
             && let Submission::UserInput { ref content } = submission
             && let Some(engine) = self.routine_engine().await
         {
+            let single_message_repl = is_single_message_repl(message);
             // Use post-hook content so that BeforeInbound hooks that rewrite
             // input are respected by event trigger matching.
-            let fired = engine.check_event_triggers(message, content).await;
+            let fired = if single_message_repl {
+                engine.check_event_triggers_and_wait(message, content).await
+            } else {
+                engine.check_event_triggers(message, content).await
+            };
             if fired > 0 {
                 tracing::debug!(
                     channel = %message.channel,
@@ -1150,15 +1252,30 @@ impl Agent {
                     fired,
                     "Consumed inbound user message with matching event-triggered routine(s)"
                 );
-                return Ok(Some(String::new()));
+                return if single_message_repl {
+                    Ok(None)
+                } else {
+                    Ok(Some(String::new()))
+                };
             }
         }
 
+        // Build per-tenant execution context once; threaded through all handlers.
+        let tenant = self.tenant_ctx(&message.user_id).await;
+
+        let session_for_empty_exit = Arc::clone(&session);
+
         // Process based on submission type
         let result = match submission {
             Submission::UserInput { content } => {
                 let mut result = self
-                    .process_user_input(message, session.clone(), thread_id, &content)
+                    .process_user_input(
+                        message,
+                        tenant.clone(),
+                        session.clone(),
+                        thread_id,
+                        &content,
+                    )
                     .await;
 
                 // Drain any messages queued during processing.
@@ -1225,7 +1342,13 @@ impl Agent {
                     let mut queued_msg = message.clone();
                     queued_msg.attachments.clear();
                     result = self
-                        .process_user_input(&queued_msg, session.clone(), thread_id, &next_content)
+                        .process_user_input(
+                            &queued_msg,
+                            tenant.clone(),
+                            session.clone(),
+                            thread_id,
+                            &next_content,
+                        )
                         .await;
 
                     // If processing failed, re-queue the drained content so it
@@ -1250,8 +1373,30 @@ impl Agent {
                     command,
                     message.channel
                 );
+                // /reasoning is special-cased here (not in handle_system_command)
+                // because it needs the session + thread_id to read turn reasoning
+                // data, which handle_system_command's signature doesn't provide.
+                if command == "reasoning" {
+                    let result = self
+                        .handle_reasoning_command(&args, &session, thread_id)
+                        .await;
+                    return match result {
+                        SubmissionResult::Response { content } => Ok(Some(content)),
+                        SubmissionResult::Ok { message } => Ok(message),
+                        SubmissionResult::Error { message } => {
+                            Ok(Some(format!("Error: {}", message)))
+                        }
+                        _ => {
+                            if is_single_message_repl(message) {
+                                Ok(None)
+                            } else {
+                                Ok(Some(String::new()))
+                            }
+                        }
+                    };
+                }
                 // Authorization checks (including restart channel check) are enforced in handle_system_command
-                self.handle_system_command(&command, &args, &message.channel)
+                self.handle_system_command(&command, &args, &message.channel, &tenant)
                     .await
             }
             Submission::Undo => self.process_undo(session, thread_id).await,
@@ -1264,12 +1409,9 @@ impl Agent {
             Submission::Summarize => self.process_summarize(session, thread_id).await,
             Submission::Suggest => self.process_suggest(session, thread_id).await,
             Submission::JobStatus { job_id } => {
-                self.process_job_status(&message.user_id, job_id.as_deref())
-                    .await
-            }
-            Submission::JobCancel { job_id } => {
-                self.process_job_cancel(&message.user_id, &job_id).await
+                self.process_job_status(&tenant, job_id.as_deref()).await
             }
+            Submission::JobCancel { job_id } => self.process_job_cancel(&tenant, &job_id).await,
             Submission::Quit => return Ok(None),
             Submission::SwitchThread { thread_id: target } => {
                 self.process_switch_thread(message, target).await
@@ -1309,7 +1451,26 @@ impl Agent {
                     Ok(Some(content))
                 }
             }
-            SubmissionResult::Ok { message } => Ok(message),
+            SubmissionResult::Ok {
+                message: output_message,
+            } => {
+                let should_exit =
+                    if output_message.as_deref() == Some("") && is_single_message_repl(message) {
+                        let sess = session_for_empty_exit.lock().await;
+                        sess.threads
+                            .get(&thread_id)
+                            .map(|thread| thread.state != ThreadState::AwaitingApproval)
+                            .unwrap_or(true)
+                    } else {
+                        false
+                    };
+
+                if should_exit {
+                    Ok(None)
+                } else {
+                    Ok(output_message)
+                }
+            }
             SubmissionResult::Error { message } => Ok(Some(format!("Error: {}", message))),
             SubmissionResult::Interrupted => Ok(Some("Interrupted.".into())),
             SubmissionResult::NeedApproval { .. } => {
@@ -1325,7 +1486,7 @@ impl Agent {
 #[cfg(test)]
 mod tests {
     use super::{
-        chat_tool_execution_metadata, resolve_routine_notification_user,
+        chat_tool_execution_metadata, is_single_message_repl, resolve_routine_notification_user,
         should_fallback_routine_notification, truncate_for_preview,
     };
     use crate::channels::IncomingMessage;
@@ -1487,4 +1648,17 @@ mod tests {
 
         assert!(should_fallback_routine_notification(&error)); // safety: test-only assertion
     }
+
+    #[test]
+    fn single_message_repl_detection_requires_repl_channel_and_metadata_flag() {
+        let repl = IncomingMessage::new("repl", "owner-scope", "hello")
+            .with_metadata(serde_json::json!({ "single_message_mode": true }));
+        let gateway = IncomingMessage::new("gateway", "owner-scope", "hello")
+            .with_metadata(serde_json::json!({ "single_message_mode": true }));
+        let plain_repl = IncomingMessage::new("repl", "owner-scope", "hello");
+
+        assert!(is_single_message_repl(&repl)); // safety: test-only assertion
+        assert!(!is_single_message_repl(&gateway)); // safety: test-only assertion
+        assert!(!is_single_message_repl(&plain_repl)); // safety: test-only assertion
+    }
 }

src/agent/agentic_loop.rs

@@ -10,7 +10,7 @@ use std::borrow::Cow;
 
 use crate::agent::session::PendingApproval;
 use crate::error::Error;
-use crate::llm::{ChatMessage, Reasoning, ReasoningContext, RespondResult};
+use crate::llm::{ChatMessage, FinishReason, Reasoning, ReasoningContext, RespondResult};
 
 /// Signal from the delegate indicating how the loop should proceed.
 pub enum LoopSignal {
@@ -134,6 +134,9 @@ pub async fn run_agentic_loop(
     config: &AgenticLoopConfig,
 ) -> Result<LoopOutcome, Error> {
     let mut consecutive_tool_intent_nudges: u32 = 0;
+    // Accumulates across all iterations (not reset by text responses) so
+    // non-consecutive truncations still escalate to force_text.
+    let mut truncation_count: u32 = 0;
 
     for iteration in 1..=config.max_iterations {
         // Check for external signals (stop, cancellation, user messages)
@@ -215,7 +218,35 @@ pub async fn run_agentic_loop(
                 tool_calls,
                 content,
             } => {
+                // If the response was truncated, tool call parameters are likely
+                // incomplete. Discard them and tell the LLM to try a different
+                // approach rather than executing malformed tool calls.
+                if output.finish_reason == FinishReason::Length {
+                    truncation_count += 1;
+                    let names: Vec<&str> = tool_calls.iter().map(|tc| tc.name.as_str()).collect();
+                    tracing::warn!(
+                        iteration,
+                        tools = ?names,
+                        truncation_count,
+                        "Discarding truncated tool calls (finish_reason=Length)"
+                    );
+                    if let Some(ref text) = content {
+                        reason_ctx.messages.push(ChatMessage::assistant(text));
+                    }
+                    reason_ctx
+                        .messages
+                        .push(ChatMessage::user(crate::llm::TRUNCATED_TOOL_CALL_NOTICE));
+                    // After repeated truncations, force text-only mode so the LLM
+                    // stops attempting tool calls it can't fit in the output budget.
+                    if truncation_count >= 3 {
+                        reason_ctx.force_text = true;
+                    }
+                    delegate.after_iteration(iteration).await;
+                    continue;
+                }
+
                 consecutive_tool_intent_nudges = 0;
+                truncation_count = 0;
 
                 if let Some(outcome) = delegate
                     .execute_tool_calls(tool_calls, content, reason_ctx)
@@ -271,6 +302,7 @@ mod tests {
         RespondOutput {
             result: RespondResult::Text(text.to_string()),
             usage: zero_usage(),
+            finish_reason: FinishReason::Stop,
         }
     }
 
@@ -281,6 +313,7 @@ mod tests {
                 content: None,
             },
             usage: zero_usage(),
+            finish_reason: FinishReason::ToolUse,
         }
     }
 
@@ -414,6 +447,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "echo".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
         let delegate = MockDelegate::new(vec![
             tool_calls_output(vec![tool_call]),
@@ -621,4 +655,95 @@ mod tests {
         let result = truncate_for_preview("café", 4);
         assert_eq!(result, "caf...");
     }
+
+    #[tokio::test]
+    async fn test_truncated_tool_calls_discarded_on_length() {
+        let truncated_tool_call = ToolCall {
+            id: "call_1".to_string(),
+            name: "memory_write".to_string(),
+            arguments: serde_json::json!({}), // empty — truncated
+            reasoning: None,
+        };
+        let truncated_output = RespondOutput {
+            result: RespondResult::ToolCalls {
+                tool_calls: vec![truncated_tool_call],
+                content: Some("I'll write the report.".to_string()),
+            },
+            usage: zero_usage(),
+            finish_reason: FinishReason::Length, // response was truncated
+        };
+        let delegate = MockDelegate::new(vec![truncated_output, text_output("Summarized it.")]);
+        let reasoning = stub_reasoning();
+        let mut ctx = ReasoningContext::new();
+        let config = AgenticLoopConfig {
+            max_iterations: 5,
+            ..Default::default()
+        };
+
+        let outcome = run_agentic_loop(&delegate, &reasoning, &mut ctx, &config)
+            .await
+            .unwrap();
+
+        // Tool calls should NOT have been executed
+        assert_eq!(delegate.tool_exec_count.load(Ordering::SeqCst), 0);
+        // The loop should have continued and returned the text response
+        assert!(matches!(outcome, LoopOutcome::Response(ref t) if t == "Summarized it."));
+        // A truncation notice should have been injected into context
+        assert!(
+            ctx.messages
+                .iter()
+                .any(|m| m.role == crate::llm::Role::User && m.content.contains("truncated")),
+            "Should inject truncation notice into context"
+        );
+        // The partial assistant content should have been preserved
+        assert!(
+            ctx.messages
+                .iter()
+                .any(|m| m.role == crate::llm::Role::Assistant
+                    && m.content.contains("write the report")),
+            "Should preserve partial assistant content"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_repeated_truncations_force_text_mode() {
+        let make_truncated = || RespondOutput {
+            result: RespondResult::ToolCalls {
+                tool_calls: vec![ToolCall {
+                    id: "call_1".to_string(),
+                    name: "memory_write".to_string(),
+                    arguments: serde_json::json!({}),
+                    reasoning: None,
+                }],
+                content: None,
+            },
+            usage: zero_usage(),
+            finish_reason: FinishReason::Length,
+        };
+        // Three truncated responses, then a text response
+        let delegate = MockDelegate::new(vec![
+            make_truncated(),
+            make_truncated(),
+            make_truncated(),
+            text_output("Gave up on tool calls."),
+        ]);
+        let reasoning = stub_reasoning();
+        let mut ctx = ReasoningContext::new();
+        let config = AgenticLoopConfig {
+            max_iterations: 5,
+            ..Default::default()
+        };
+
+        let outcome = run_agentic_loop(&delegate, &reasoning, &mut ctx, &config)
+            .await
+            .unwrap();
+
+        assert!(matches!(outcome, LoopOutcome::Response(_)));
+        assert_eq!(delegate.tool_exec_count.load(Ordering::SeqCst), 0);
+        // After 3 truncations, force_text should be set
+        assert!(
+            ctx.force_text,
+            "Should escalate to force_text after repeated truncations"
+        );
+    }
 }

src/agent/commands.rs

@@ -33,6 +33,7 @@ impl Agent {
         &self,
         intent: MessageIntent,
         message: &IncomingMessage,
+        tenant: &crate::tenant::TenantCtx,
     ) -> Result<SubmissionResult, Error> {
         // Send thinking status for non-trivial operations
         if let MessageIntent::CreateJob { .. } = &intent {
@@ -52,24 +53,18 @@ impl Agent {
                 description,
                 category,
             } => {
-                self.handle_create_job(&message.user_id, title, description, category)
+                self.handle_create_job(tenant, title, description, category)
                     .await?
             }
             MessageIntent::CheckJobStatus { job_id } => {
-                self.handle_check_status(&message.user_id, job_id).await?
-            }
-            MessageIntent::CancelJob { job_id } => {
-                self.handle_cancel_job(&message.user_id, &job_id).await?
-            }
-            MessageIntent::ListJobs { filter } => {
-                self.handle_list_jobs(&message.user_id, filter).await?
-            }
-            MessageIntent::HelpJob { job_id } => {
-                self.handle_help_job(&message.user_id, &job_id).await?
+                self.handle_check_status(tenant, job_id).await?
             }
+            MessageIntent::CancelJob { job_id } => self.handle_cancel_job(tenant, &job_id).await?,
+            MessageIntent::ListJobs { filter } => self.handle_list_jobs(tenant, filter).await?,
+            MessageIntent::HelpJob { job_id } => self.handle_help_job(tenant, &job_id).await?,
             MessageIntent::Command { command, args } => {
                 match self
-                    .handle_command(&command, &args, &message.channel)
+                    .handle_command(&command, &args, &message.channel, tenant)
                     .await?
                 {
                     Some(s) => s,
@@ -83,14 +78,14 @@ impl Agent {
 
     async fn handle_create_job(
         &self,
-        user_id: &str,
+        tenant: &crate::tenant::TenantCtx,
         title: String,
         description: String,
         category: Option<String>,
     ) -> Result<String, Error> {
         let job_id = self
             .scheduler
-            .dispatch_job(user_id, &title, &description, None)
+            .dispatch_job(tenant.user_id(), &title, &description, None)
             .await?;
 
         // Set the dedicated category field (not stored in metadata)
@@ -113,7 +108,7 @@ impl Agent {
 
     async fn handle_check_status(
         &self,
-        user_id: &str,
+        tenant: &crate::tenant::TenantCtx,
         job_id: Option<String>,
     ) -> Result<String, Error> {
         match job_id {
@@ -122,7 +117,8 @@ impl Agent {
                     .map_err(|_| crate::error::JobError::NotFound { id: Uuid::nil() })?;
 
                 // Try DB first for persistent state, fall back to ContextManager.
-                if let Some(store) = self.store()
+                // TenantScope.get_job() auto-filters by ownership — no manual check needed.
+                if let Some(store) = tenant.store()
                     && let Ok(Some(ctx)) = store.get_job(uuid).await
                 {
                     return Ok(format!(
@@ -138,7 +134,7 @@ impl Agent {
                 }
 
                 let ctx = self.context_manager.get_context(uuid).await?;
-                if ctx.user_id != user_id {
+                if ctx.user_id != tenant.user_id() {
                     return Err(crate::error::JobError::NotFound { id: uuid }.into());
                 }
 
@@ -155,7 +151,8 @@ impl Agent {
             }
             None => {
                 // Show summary from DB for consistency with Jobs tab.
-                if let Some(store) = self.store() {
+                // TenantScope methods auto-scope to user — no user_id parameter needed.
+                if let Some(store) = tenant.store() {
                     let mut total = 0;
                     let mut in_progress = 0;
                     let mut completed = 0;
@@ -183,7 +180,7 @@ impl Agent {
                 }
 
                 // Fallback to ContextManager if no DB.
-                let summary = self.context_manager.summary_for(user_id).await;
+                let summary = self.context_manager.summary_for(tenant.user_id()).await;
                 Ok(format!(
                     "Jobs summary: Total: {} In Progress: {} Completed: {} Failed: {} Stuck: {}",
                     summary.total,
@@ -196,19 +193,24 @@ impl Agent {
         }
     }
 
-    async fn handle_cancel_job(&self, user_id: &str, job_id: &str) -> Result<String, Error> {
+    async fn handle_cancel_job(
+        &self,
+        tenant: &crate::tenant::TenantCtx,
+        job_id: &str,
+    ) -> Result<String, Error> {
         let uuid = Uuid::parse_str(job_id)
             .map_err(|_| crate::error::JobError::NotFound { id: Uuid::nil() })?;
 
         let ctx = self.context_manager.get_context(uuid).await?;
-        if ctx.user_id != user_id {
+        if ctx.user_id != tenant.user_id() {
             return Err(crate::error::JobError::NotFound { id: uuid }.into());
         }
 
         self.scheduler.stop(uuid).await?;
 
         // Also update DB so the Jobs tab reflects cancellation immediately.
-        if let Some(store) = self.store()
+        // Use TenantScope — ownership already verified above.
+        if let Some(store) = tenant.store()
             && let Err(e) = store
                 .update_job_status(uuid, JobState::Cancelled, Some("Cancelled by user"))
                 .await
@@ -221,11 +223,12 @@ impl Agent {
 
     async fn handle_list_jobs(
         &self,
-        user_id: &str,
+        tenant: &crate::tenant::TenantCtx,
         _filter: Option<String>,
     ) -> Result<String, Error> {
         // List from DB for consistency with Jobs tab.
-        if let Some(store) = self.store() {
+        // TenantScope methods auto-scope to user.
+        if let Some(store) = tenant.store() {
             let agent_jobs = match store.list_agent_jobs().await {
                 Ok(jobs) => jobs,
                 Err(e) => {
@@ -256,7 +259,7 @@ impl Agent {
         }
 
         // Fallback to ContextManager if no DB.
-        let jobs = self.context_manager.all_jobs_for(user_id).await;
+        let jobs = self.context_manager.all_jobs_for(tenant.user_id()).await;
         if jobs.is_empty() {
             return Ok("No jobs found.".to_string());
         }
@@ -270,12 +273,16 @@ impl Agent {
         Ok(output)
     }
 
-    async fn handle_help_job(&self, user_id: &str, job_id: &str) -> Result<String, Error> {
+    async fn handle_help_job(
+        &self,
+        tenant: &crate::tenant::TenantCtx,
+        job_id: &str,
+    ) -> Result<String, Error> {
         let uuid = Uuid::parse_str(job_id)
             .map_err(|_| crate::error::JobError::NotFound { id: Uuid::nil() })?;
 
         let ctx = self.context_manager.get_context(uuid).await?;
-        if ctx.user_id != user_id {
+        if ctx.user_id != tenant.user_id() {
             return Err(crate::error::JobError::NotFound { id: uuid }.into());
         }
 
@@ -308,11 +315,11 @@ impl Agent {
     /// Show job status inline — either all jobs (no id) or a specific job.
     pub(super) async fn process_job_status(
         &self,
-        user_id: &str,
+        tenant: &crate::tenant::TenantCtx,
         job_id: Option<&str>,
     ) -> Result<SubmissionResult, Error> {
         match self
-            .handle_check_status(user_id, job_id.map(|s| s.to_string()))
+            .handle_check_status(tenant, job_id.map(|s| s.to_string()))
             .await
         {
             Ok(text) => Ok(SubmissionResult::response(text)),
@@ -323,10 +330,10 @@ impl Agent {
     /// Cancel a job by ID.
     pub(super) async fn process_job_cancel(
         &self,
-        user_id: &str,
+        tenant: &crate::tenant::TenantCtx,
         job_id: &str,
     ) -> Result<SubmissionResult, Error> {
-        match self.handle_cancel_job(user_id, job_id).await {
+        match self.handle_cancel_job(tenant, job_id).await {
             Ok(text) => Ok(SubmissionResult::response(text)),
             Err(e) => Ok(SubmissionResult::error(format!("Cancel error: {}", e))),
         }
@@ -465,12 +472,101 @@ impl Agent {
         }
     }
 
+    /// Handle `/reasoning [N|all]` — show reasoning history for the active thread.
+    pub(super) async fn handle_reasoning_command(
+        &self,
+        args: &[String],
+        session: &Arc<Mutex<Session>>,
+        thread_id: Uuid,
+    ) -> SubmissionResult {
+        // Clone the turn data we need, then drop the session lock.
+        let turns_snapshot: Vec<(
+            usize,
+            Option<String>,
+            Vec<crate::agent::session::TurnToolCall>,
+        )>;
+        {
+            let sess = session.lock().await;
+            let thread = match sess.threads.get(&thread_id) {
+                Some(t) => t,
+                None => return SubmissionResult::error("No active thread."),
+            };
+
+            if thread.turns.is_empty() {
+                return SubmissionResult::ok_with_message("No turns yet.");
+            }
+
+            // Parse argument: default=last turn, "all"=all turns, N=specific turn (1-based).
+            let selected: Vec<&crate::agent::session::Turn> = match args.first().map(|s| s.as_str())
+            {
+                Some("all") => thread.turns.iter().collect(),
+                Some(n) => match n.parse::<usize>() {
+                    Ok(0) => return SubmissionResult::error("Turn numbers start at 1."),
+                    Ok(num) if num > thread.turns.len() => {
+                        return SubmissionResult::error(format!(
+                            "Turn {} does not exist (max: {}).",
+                            num,
+                            thread.turns.len()
+                        ));
+                    }
+                    Ok(num) => vec![&thread.turns[num - 1]],
+                    Err(_) => return SubmissionResult::error("Usage: /reasoning [N|all]"),
+                },
+                None => {
+                    // Default: last turn that has tool calls
+                    match thread.turns.iter().rev().find(|t| !t.tool_calls.is_empty()) {
+                        Some(t) => vec![t],
+                        None => {
+                            return SubmissionResult::ok_with_message("No turns with tool calls.");
+                        }
+                    }
+                }
+            };
+
+            turns_snapshot = selected
+                .into_iter()
+                .map(|t| (t.turn_number, t.narrative.clone(), t.tool_calls.clone()))
+                .collect();
+        }
+        // Session lock is now dropped — format output without holding it.
+
+        let mut output = String::new();
+        for (turn_number, narrative, tool_calls) in &turns_snapshot {
+            output.push_str(&format!("--- Turn {} ---\n", turn_number + 1));
+            if let Some(narrative) = narrative {
+                output.push_str(&format!("Reasoning: {}\n", narrative));
+            }
+            if tool_calls.is_empty() {
+                output.push_str("  (no tool calls)\n");
+            } else {
+                for tc in tool_calls {
+                    let status = if tc.error.is_some() {
+                        "error"
+                    } else if tc.result.is_some() {
+                        "ok"
+                    } else {
+                        "pending"
+                    };
+                    output.push_str(&format!("  {} [{}]", tc.name, status));
+                    if let Some(ref rationale) = tc.rationale {
+                        output.push_str(&format!(" — {}", rationale));
+                    }
+                    output.push('\n');
+                }
+            }
+            output.push('\n');
+        }
+
+        SubmissionResult::response(output.trim_end())
+    }
+
     /// Handle system commands that bypass thread-state checks entirely.
     pub(super) async fn handle_system_command(
         &self,
         command: &str,
         args: &[String],
         channel: &str,
+        tenant: &crate::tenant::TenantCtx,
     ) -> Result<SubmissionResult, Error> {
         match command {
             "help" => Ok(SubmissionResult::response(concat!(
@@ -480,6 +576,7 @@ impl Agent {
                 "  /version          Show version info\n",
                 "  /tools            List available tools\n",
                 "  /debug            Toggle debug mode\n",
+                "  /reasoning [N|all] Show agent reasoning for turns\n",
                 "  /ping             Connectivity check\n",
                 "\n",
                 "Jobs:\n",
@@ -663,19 +760,32 @@ impl Agent {
                         }
                     }
 
-                    match self.llm().set_model(requested) {
-                        Ok(()) => {
-                            // Persist the model choice so it survives restarts.
-                            self.persist_selected_model(requested).await;
-                            Ok(SubmissionResult::response(format!(
-                                "Switched model to: {}",
-                                requested
-                            )))
+                    if self.config.multi_tenant {
+                        // Multi-tenant: only persist to per-user DB settings.
+                        // Do NOT call set_model() on the shared provider — that
+                        // would change the default for all users. The per-request
+                        // model_override in the dispatcher reads from the same
+                        // "selected_model" setting and applies it per-user.
+                        self.persist_selected_model(tenant, requested).await;
+                        Ok(SubmissionResult::response(format!(
+                            "Model preference set to: {} (per-user)",
+                            requested
+                        )))
+                    } else {
+                        match self.llm().set_model(requested) {
+                            Ok(()) => {
+                                // Persist the model choice so it survives restarts.
+                                self.persist_selected_model(tenant, requested).await;
+                                Ok(SubmissionResult::response(format!(
+                                    "Switched model to: {}",
+                                    requested
+                                )))
+                            }
+                            Err(e) => Ok(SubmissionResult::error(format!(
+                                "Failed to switch model: {}",
+                                e
+                            ))),
                         }
-                        Err(e) => Ok(SubmissionResult::error(format!(
-                            "Failed to switch model: {}",
-                            e
-                        ))),
                     }
                 }
             }
@@ -817,10 +927,14 @@ impl Agent {
         command: &str,
         args: &[String],
         channel: &str,
+        tenant: &crate::tenant::TenantCtx,
     ) -> Result<Option<String>, Error> {
         // System commands are now handled directly via Submission::SystemCommand,
         // but the router may still send us unknown /commands.
-        match self.handle_system_command(command, args, channel).await? {
+        match self
+            .handle_system_command(command, args, channel, tenant)
+            .await?
+        {
             SubmissionResult::Response { content } => Ok(Some(content)),
             SubmissionResult::Ok { message } => Ok(message),
             SubmissionResult::Error { message } => Ok(Some(format!("Error: {}", message))),
@@ -832,23 +946,33 @@ impl Agent {
     ///
     /// Best-effort: logs warnings on failure but does not propagate errors,
     /// since the in-memory model switch already succeeded.
-    async fn persist_selected_model(&self, model: &str) {
-        // 1. Persist to DB if available.
-        if let Some(store) = self.store() {
+    ///
+    /// In multi-tenant mode, only the per-user DB setting is written — global
+    /// .env and TOML files are shared across users and must not be mutated.
+    async fn persist_selected_model(&self, tenant: &crate::tenant::TenantCtx, model: &str) {
+        // 1. Persist to DB if available (per-user scoped via TenantScope).
+        if let Some(store) = tenant.store() {
             let value = serde_json::Value::String(model.to_string());
-            if let Err(e) = store
-                .set_setting(self.owner_id(), "selected_model", &value)
-                .await
-            {
+            if let Err(e) = store.set_setting("selected_model", &value).await {
                 tracing::warn!("Failed to persist model to DB: {}", e);
             } else {
-                tracing::debug!("Persisted selected_model to DB: {}", model);
+                tracing::debug!(
+                    user_id = tenant.user_id(),
+                    "Persisted selected_model to DB: {}",
+                    model
+                );
             }
         } else {
             tracing::warn!("No database store available — model choice will not persist to DB");
         }
 
-        // 2. Update .env and TOML config file (sync I/O in spawn_blocking).
+        // 2. In multi-tenant mode, skip .env/TOML writes — these are global
+        // files shared by all users. The per-user DB setting is sufficient.
+        if self.config.multi_tenant {
+            return;
+        }
+
+        // 3. Update .env and TOML config file (sync I/O in spawn_blocking).
         let model_owned = model.to_string();
         let backend = self.deps.llm_backend.clone();
         if let Err(e) = tokio::task::spawn_blocking(move || {

src/agent/cost_guard.rs

@@ -21,6 +21,9 @@ pub struct CostGuardConfig {
     pub max_cost_per_day_cents: Option<u64>,
     /// Maximum LLM calls per hour. None = unlimited.
     pub max_actions_per_hour: Option<u64>,
+    /// Maximum spend per user per day in cents. None = unlimited.
+    /// Applied independently per user alongside the global budget.
+    pub max_cost_per_user_per_day_cents: Option<u64>,
 }
 
 /// Error returned when a cost limit is exceeded.
@@ -30,6 +33,12 @@ pub enum CostLimitExceeded {
     DailyBudget { spent_cents: u64, limit_cents: u64 },
     /// Hourly action rate limit reached.
     HourlyRate { actions: u64, limit: u64 },
+    /// Per-user daily spending cap reached.
+    UserDailyBudget {
+        user_id: String,
+        spent_cents: u64,
+        limit_cents: u64,
+    },
 }
 
 impl std::fmt::Display for CostLimitExceeded {
@@ -49,6 +58,17 @@ impl std::fmt::Display for CostLimitExceeded {
                 "Hourly action limit exceeded: {} actions of {} allowed per hour",
                 actions, limit
             ),
+            Self::UserDailyBudget {
+                user_id,
+                spent_cents,
+                limit_cents,
+            } => write!(
+                f,
+                "User '{}' daily cost limit exceeded: spent ${:.2} of ${:.2} allowed",
+                user_id,
+                *spent_cents as f64 / 100.0,
+                *limit_cents as f64 / 100.0
+            ),
         }
     }
 }
@@ -78,6 +98,9 @@ pub struct CostGuard {
 
     /// Per-model token usage since startup.
     model_tokens: Mutex<HashMap<String, ModelTokens>>,
+
+    /// Per-user daily cost tracking. Each entry resets independently at midnight UTC.
+    per_user_daily_cost: Mutex<HashMap<String, DailyCost>>,
 }
 
 struct DailyCost {
@@ -97,6 +120,7 @@ impl CostGuard {
             action_window: Mutex::new(VecDeque::new()),
             budget_exceeded: AtomicBool::new(false),
             model_tokens: Mutex::new(HashMap::new()),
+            per_user_daily_cost: Mutex::new(HashMap::new()),
         }
     }
 
@@ -203,6 +227,11 @@ impl CostGuard {
                 daily.reset_date = today;
                 self.budget_exceeded.store(false, Ordering::Relaxed);
                 tracing::info!("Cost guard: daily counter reset for {}", today);
+
+                // Prune per-user entries from previous days to prevent
+                // unbounded HashMap growth in long-lived deployments.
+                let mut per_user = self.per_user_daily_cost.lock().await;
+                per_user.retain(|_, entry| entry.reset_date == today);
             }
             daily.total += cost;
 
@@ -248,6 +277,85 @@ impl CostGuard {
         cost
     }
 
+    /// Record an LLM call with per-user attribution.
+    ///
+    /// Delegates to `record_llm_call` for global tracking, then additionally
+    /// records the cost against the user's daily budget.
+    #[allow(clippy::too_many_arguments)]
+    pub async fn record_llm_call_for_user(
+        &self,
+        user_id: &str,
+        model: &str,
+        input_tokens: u32,
+        output_tokens: u32,
+        cache_read_input_tokens: u32,
+        cache_creation_input_tokens: u32,
+        cache_read_discount: Decimal,
+        cache_write_multiplier: Decimal,
+        cost_per_token: Option<(Decimal, Decimal)>,
+    ) -> Decimal {
+        let cost = self
+            .record_llm_call(
+                model,
+                input_tokens,
+                output_tokens,
+                cache_read_input_tokens,
+                cache_creation_input_tokens,
+                cache_read_discount,
+                cache_write_multiplier,
+                cost_per_token,
+            )
+            .await;
+
+        // Track per-user daily cost
+        {
+            let today = chrono::Utc::now().date_naive();
+            let mut per_user = self.per_user_daily_cost.lock().await;
+            let entry = per_user
+                .entry(user_id.to_string())
+                .or_insert_with(|| DailyCost {
+                    total: Decimal::ZERO,
+                    reset_date: today,
+                });
+            if today != entry.reset_date {
+                entry.total = Decimal::ZERO;
+                entry.reset_date = today;
+            }
+            entry.total += cost;
+        }
+
+        cost
+    }
+
+    /// Check whether the next action is allowed for a specific user.
+    ///
+    /// Checks the global limits first (via `check_allowed`), then additionally
+    /// checks the per-user daily budget if configured.
+    pub async fn check_allowed_for_user(&self, user_id: &str) -> Result<(), CostLimitExceeded> {
+        // Check global limits first
+        self.check_allowed().await?;
+
+        // Check per-user daily budget
+        if let Some(limit_cents) = self.config.max_cost_per_user_per_day_cents {
+            let today = chrono::Utc::now().date_naive();
+            let per_user = self.per_user_daily_cost.lock().await;
+            if let Some(entry) = per_user.get(user_id)
+                && entry.reset_date == today
+            {
+                let spent_cents = to_cents(entry.total);
+                if spent_cents >= limit_cents {
+                    return Err(CostLimitExceeded::UserDailyBudget {
+                        user_id: user_id.to_string(),
+                        spent_cents,
+                        limit_cents,
+                    });
+                }
+            }
+        }
+
+        Ok(())
+    }
+
     /// Current daily spend in USD (as Decimal).
     pub async fn daily_spend(&self) -> Decimal {
         let daily = self.daily_cost.lock().await;
@@ -259,6 +367,16 @@ impl CostGuard {
         }
     }
 
+    /// Current daily spend for a specific user in USD (as Decimal).
+    pub async fn daily_spend_for_user(&self, user_id: &str) -> Decimal {
+        let today = chrono::Utc::now().date_naive();
+        let per_user = self.per_user_daily_cost.lock().await;
+        match per_user.get(user_id) {
+            Some(entry) if entry.reset_date == today => entry.total,
+            _ => Decimal::ZERO,
+        }
+    }
+
     /// Number of actions in the current hourly window.
     pub async fn actions_this_hour(&self) -> u64 {
         let mut window = self.action_window.lock().await;
@@ -314,7 +432,7 @@ mod tests {
     async fn test_daily_budget_enforcement() {
         let guard = CostGuard::new(CostGuardConfig {
             max_cost_per_day_cents: Some(1), // $0.01 limit
-            max_actions_per_hour: None,
+            ..CostGuardConfig::default()
         });
 
         // First call allowed
@@ -350,8 +468,8 @@ mod tests {
     #[tokio::test]
     async fn test_hourly_rate_enforcement() {
         let guard = CostGuard::new(CostGuardConfig {
-            max_cost_per_day_cents: None,
             max_actions_per_hour: Some(3),
+            ..CostGuardConfig::default()
         });
 
         // First 3 actions allowed
@@ -633,8 +751,8 @@ mod tests {
         // A fresh CostGuard with rate limits should not panic even if
         // checked_sub returns None (simulating short uptime).
         let guard = CostGuard::new(CostGuardConfig {
-            max_cost_per_day_cents: None,
             max_actions_per_hour: Some(100),
+            ..CostGuardConfig::default()
         });
 
         // These must not panic regardless of system uptime
@@ -656,4 +774,119 @@ mod tests {
         let result = Instant::now().checked_sub(std::time::Duration::MAX);
         assert!(result.is_none());
     }
+
+    #[tokio::test]
+    async fn test_per_user_daily_budget_enforcement() {
+        let guard = CostGuard::new(CostGuardConfig {
+            max_cost_per_day_cents: None,
+            max_actions_per_hour: None,
+            max_cost_per_user_per_day_cents: Some(1), // $0.01 per user
+        });
+
+        // Both users initially allowed
+        assert!(guard.check_allowed_for_user("alice").await.is_ok());
+        assert!(guard.check_allowed_for_user("bob").await.is_ok());
+
+        // Alice makes an expensive call
+        guard
+            .record_llm_call_for_user(
+                "alice",
+                "gpt-4o",
+                10_000,
+                10_000,
+                0,
+                0,
+                Decimal::ONE,
+                Decimal::ONE,
+                None,
+            )
+            .await;
+
+        // Alice should be blocked, Bob should still be allowed
+        let result = guard.check_allowed_for_user("alice").await;
+        assert!(result.is_err());
+        match result.unwrap_err() {
+            CostLimitExceeded::UserDailyBudget {
+                user_id,
+                limit_cents,
+                ..
+            } => {
+                assert_eq!(user_id, "alice");
+                assert_eq!(limit_cents, 1);
+            }
+            other => panic!("Expected UserDailyBudget, got {:?}", other),
+        }
+        assert!(guard.check_allowed_for_user("bob").await.is_ok());
+    }
+
+    #[tokio::test]
+    async fn test_per_user_daily_spend_tracking() {
+        let guard = CostGuard::new(CostGuardConfig::default());
+
+        assert_eq!(guard.daily_spend_for_user("alice").await, Decimal::ZERO);
+        assert_eq!(guard.daily_spend_for_user("bob").await, Decimal::ZERO);
+
+        let cost = guard
+            .record_llm_call_for_user(
+                "alice",
+                "gpt-4o",
+                1000,
+                500,
+                0,
+                0,
+                Decimal::ONE,
+                Decimal::ONE,
+                None,
+            )
+            .await;
+
+        assert_eq!(guard.daily_spend_for_user("alice").await, cost);
+        assert_eq!(guard.daily_spend_for_user("bob").await, Decimal::ZERO);
+        // Global spend should also be tracked
+        assert_eq!(guard.daily_spend().await, cost);
+    }
+
+    #[tokio::test]
+    async fn test_per_user_budget_independent_of_global() {
+        let guard = CostGuard::new(CostGuardConfig {
+            max_cost_per_day_cents: Some(100_000), // $1000 global limit
+            max_actions_per_hour: None,
+            max_cost_per_user_per_day_cents: Some(1), // $0.01 per user
+        });
+
+        // User hits their personal limit
+        guard
+            .record_llm_call_for_user(
+                "alice",
+                "gpt-4o",
+                10_000,
+                10_000,
+                0,
+                0,
+                Decimal::ONE,
+                Decimal::ONE,
+                None,
+            )
+            .await;
+
+        // Alice blocked by per-user limit, not global
+        assert!(guard.check_allowed_for_user("alice").await.is_err());
+        // Global limit is far from reached
+        assert!(guard.check_allowed().await.is_ok());
+        // Bob is unaffected
+        assert!(guard.check_allowed_for_user("bob").await.is_ok());
+    }
+
+    #[test]
+    fn test_user_cost_limit_display() {
+        let limit = CostLimitExceeded::UserDailyBudget {
+            user_id: "alice".to_string(),
+            spent_cents: 150,
+            limit_cents: 100,
+        };
+        let msg = limit.to_string();
+        assert!(msg.contains("alice"));
+        assert!(msg.contains("$1.50"));
+        assert!(msg.contains("$1.00"));
+    }
 }

src/agent/dispatcher.rs

@@ -42,6 +42,7 @@ impl Agent {
     pub(super) async fn run_agentic_loop(
         &self,
         message: &IncomingMessage,
+        tenant: crate::tenant::TenantCtx,
         session: Arc<Mutex<Session>>,
         thread_id: Uuid,
         initial_messages: Vec<ChatMessage>,
@@ -63,7 +64,12 @@ impl Agent {
         );
 
         let system_prompt = if let Some(ws) = self.workspace() {
-            match ws
+            let scoped_workspace = if ws.user_id() == message.user_id {
+                Arc::clone(ws)
+            } else {
+                Arc::new(ws.scoped_to_user(&message.user_id))
+            };
+            match scoped_workspace
                 .system_prompt_for_context_tz(is_group_chat, user_tz)
                 .await
             {
@@ -163,6 +169,7 @@ impl Agent {
 
         let delegate = ChatDelegate {
             agent: self,
+            tenant,
             session: session.clone(),
             thread_id,
             message,
@@ -235,6 +242,7 @@ impl Agent {
 /// auth intercept, and cost tracking.
 struct ChatDelegate<'a> {
     agent: &'a Agent,
+    tenant: crate::tenant::TenantCtx,
     session: Arc<Mutex<Session>>,
     thread_id: Uuid,
     message: &'a IncomingMessage,
@@ -298,6 +306,8 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
 
         // Update context for this iteration
         reason_ctx.available_tools = tool_defs;
+        // Preserve force_text if already set (e.g. by truncation escalation).
+        let force_text = force_text || reason_ctx.force_text;
         reason_ctx.system_prompt = Some(if force_text {
             self.cached_prompt_no_tools.clone()
         } else {
@@ -331,8 +341,8 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
         reason_ctx: &mut ReasoningContext,
         iteration: usize,
     ) -> Result<crate::llm::RespondOutput, Error> {
-        // Enforce cost guardrails before the LLM call
-        if let Err(limit) = self.agent.cost_guard().check_allowed().await {
+        // Enforce cost guardrails before the LLM call (global + per-user)
+        if let Err(limit) = self.tenant.check_cost_allowed().await {
             return Err(crate::error::LlmError::InvalidResponse {
                 provider: "agent".to_string(),
                 reason: limit.to_string(),
@@ -340,6 +350,21 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
             .into());
         }
 
+        // Apply per-user model override from settings (first iteration only
+        // to avoid repeated DB lookups within the same agentic loop).
+        // Uses "selected_model" — the same key the /model command persists to
+        // via SettingsStore (per-user scoped via TenantScope).
+        if iteration == 0
+            && let Some(store) = self.tenant.store()
+            && let Ok(Some(value)) = store.get_setting("selected_model").await
+            && let Some(model) = value.as_str()
+        {
+            let model = model.trim();
+            if !model.is_empty() {
+                reason_ctx.model_override = Some(model.to_string());
+            }
+        }
+
         let output = match reasoning.respond_with_tools(reason_ctx).await {
             Ok(output) => output,
             Err(crate::error::LlmError::ContextLengthExceeded { used, limit }) => {
@@ -374,13 +399,22 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
             Err(e) => return Err(e.into()),
         };
 
-        // Record cost and track token usage
-        let model_name = self.agent.llm().active_model_name();
+        // Record cost and track token usage (global + per-user).
+        // When a model override is active, use the override name for attribution
+        // and let CostGuard look up pricing via costs::model_cost() instead of
+        // using the default provider's cost_per_token (which reflects the wrong model).
+        let (model_name, cost_per_token) = if let Some(ref ovr) = reason_ctx.model_override {
+            (ovr.clone(), None)
+        } else {
+            (
+                self.agent.llm().active_model_name(),
+                Some(self.agent.llm().cost_per_token()),
+            )
+        };
         let read_discount = self.agent.llm().cache_read_discount();
         let write_multiplier = self.agent.llm().cache_write_multiplier();
         let call_cost = self
-            .agent
-            .cost_guard()
+            .tenant
             .record_llm_call(
                 &model_name,
                 output.usage.input_tokens,
@@ -389,7 +423,7 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
                 output.usage.cache_creation_input_tokens,
                 read_discount,
                 write_multiplier,
-                Some(self.agent.llm().cost_per_token()),
+                cost_per_token,
             )
             .await;
         tracing::debug!(
@@ -420,6 +454,19 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
         content: Option<String>,
         reason_ctx: &mut ReasoningContext,
     ) -> Result<Option<LoopOutcome>, Error> {
+        // Extract and sanitize the narrative before consuming `content`.
+        let narrative = content
+            .as_deref()
+            .filter(|c| !c.trim().is_empty())
+            .map(|c| {
+                let sanitized = self
+                    .agent
+                    .safety()
+                    .sanitize_tool_output("agent_narrative", c);
+                sanitized.content
+            })
+            .filter(|c| !c.trim().is_empty());
+
         // Add the assistant message with tool_calls to context.
         // OpenAI protocol requires this before tool-result messages.
         reason_ctx
@@ -440,6 +487,41 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
             )
             .await;
 
+        // Build per-tool decisions for the reasoning update.
+        // Sanitize each rationale through SafetyLayer (parity with JobDelegate).
+        let decisions: Vec<crate::channels::ToolDecision> = tool_calls
+            .iter()
+            .filter_map(|tc| {
+                tc.reasoning.as_ref().map(|r| {
+                    let sanitized = self
+                        .agent
+                        .safety()
+                        .sanitize_tool_output("tool_rationale", r)
+                        .content;
+                    crate::channels::ToolDecision {
+                        tool_name: tc.name.clone(),
+                        rationale: sanitized,
+                    }
+                })
+            })
+            .collect();
+
+        // Emit reasoning update to channels.
+        if narrative.is_some() || !decisions.is_empty() {
+            let _ = self
+                .agent
+                .channels
+                .send_status(
+                    &self.message.channel,
+                    StatusUpdate::ReasoningUpdate {
+                        narrative: narrative.clone().unwrap_or_default(),
+                        decisions: decisions.clone(),
+                    },
+                    &self.message.metadata,
+                )
+                .await;
+        }
+
         // Record tool calls in the thread with sensitive params redacted.
         {
             let mut redacted_args: Vec<serde_json::Value> = Vec::with_capacity(tool_calls.len());
@@ -455,8 +537,23 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
             if let Some(thread) = sess.threads.get_mut(&self.thread_id)
                 && let Some(turn) = thread.last_turn_mut()
             {
+                // Set turn-level narrative.
+                if turn.narrative.is_none() {
+                    turn.narrative = narrative;
+                }
                 for (tc, safe_args) in tool_calls.iter().zip(redacted_args) {
-                    turn.record_tool_call(&tc.name, safe_args);
+                    let sanitized_rationale = tc.reasoning.as_ref().map(|r| {
+                        self.agent
+                            .safety()
+                            .sanitize_tool_output("tool_rationale", r)
+                            .content
+                    });
+                    turn.record_tool_call_with_reasoning(
+                        &tc.name,
+                        safe_args,
+                        sanitized_rationale,
+                        Some(tc.id.clone()),
+                    );
                 }
             }
         }
@@ -726,7 +823,7 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
                         if let Some(thread) = sess.threads.get_mut(&self.thread_id)
                             && let Some(turn) = thread.last_turn_mut()
                         {
-                            turn.record_tool_error(error_msg.clone());
+                            turn.record_tool_error_for(&tc.id, error_msg.clone());
                         }
                     }
                     reason_ctx
@@ -852,16 +949,19 @@ impl<'a> LoopDelegate for ChatDelegate<'a> {
                         Err(e) => format!("Tool '{}' failed: {}", tc.name, e),
                     };
 
-                    // Record sanitized result in thread
+                    // Record sanitized result in thread (identity-based matching).
                     {
                         let mut sess = self.session.lock().await;
                         if let Some(thread) = sess.threads.get_mut(&self.thread_id)
                             && let Some(turn) = thread.last_turn_mut()
                         {
                             if is_tool_error {
-                                turn.record_tool_error(result_content.clone());
+                                turn.record_tool_error_for(&tc.id, result_content.clone());
                             } else {
-                                turn.record_tool_result(serde_json::json!(result_content));
+                                turn.record_tool_result_for(
+                                    &tc.id,
+                                    serde_json::json!(result_content),
+                                );
                             }
                         }
                     }
@@ -1234,6 +1334,7 @@ mod tests {
             sandbox_readiness: crate::agent::routine_engine::SandboxReadiness::DisabledByConfig,
             builder: None,
             llm_backend: "nearai".to_string(),
+            tenant_rates: Arc::new(crate::tenant::TenantRateRegistry::new(4, 3)),
         };
 
         Agent::new(
@@ -1249,10 +1350,14 @@ mod tests {
                 allow_local_tools: false,
                 max_cost_per_day_cents: None,
                 max_actions_per_hour: None,
+                max_cost_per_user_per_day_cents: None,
                 max_tool_iterations: 50,
                 auto_approve_tools: false,
                 default_timezone: "UTC".to_string(),
                 max_tokens_per_job: 0,
+                multi_tenant: false,
+                max_llm_concurrent_per_user: None,
+                max_jobs_concurrent_per_user: None,
             },
             deps,
             Arc::new(ChannelManager::new()),
@@ -1462,11 +1567,13 @@ mod tests {
                     id: "call_2".to_string(),
                     name: "http".to_string(),
                     arguments: serde_json::json!({"url": "https://example.com"}),
+                    reasoning: None,
                 },
                 ToolCall {
                     id: "call_3".to_string(),
                     name: "echo".to_string(),
                     arguments: serde_json::json!({"message": "done"}),
+                    reasoning: None,
                 },
             ],
             user_timezone: None,
@@ -1652,6 +1759,7 @@ mod tests {
                     id: "call_1".to_string(),
                     name: "echo".to_string(),
                     arguments: serde_json::json!({"message": "hi"}),
+                    reasoning: None,
                 }],
             ),
             ChatMessage::tool_result("call_1", "echo", "hi"),
@@ -1744,11 +1852,13 @@ mod tests {
                         id: "c1".to_string(),
                         name: "http".to_string(),
                         arguments: serde_json::json!({}),
+                        reasoning: None,
                     },
                     ToolCall {
                         id: "c2".to_string(),
                         name: "echo".to_string(),
                         arguments: serde_json::json!({}),
+                        reasoning: None,
                     },
                 ],
             ),
@@ -1782,6 +1892,7 @@ mod tests {
                     id: "c1".to_string(),
                     name: "echo".to_string(),
                     arguments: serde_json::json!({}),
+                    reasoning: None,
                 }],
             ),
             ChatMessage::tool_result("c1", "echo", "done"),
@@ -1912,6 +2023,7 @@ mod tests {
                     id: crate::llm::generate_tool_call_id(0, 0),
                     name: "echo".to_string(),
                     arguments: serde_json::json!({"message": "looping"}),
+                    reasoning: None,
                 }],
                 input_tokens: 0,
                 output_tokens: 5,
@@ -2065,6 +2177,7 @@ mod tests {
                     id: crate::llm::generate_tool_call_id(0, 0),
                     name: "nonexistent_tool".to_string(),
                     arguments: serde_json::json!({}),
+                    reasoning: None,
                 }],
                 input_tokens: 0,
                 output_tokens: 5,
@@ -2102,6 +2215,7 @@ mod tests {
             sandbox_readiness: crate::agent::routine_engine::SandboxReadiness::DisabledByConfig,
             builder: None,
             llm_backend: "nearai".to_string(),
+            tenant_rates: Arc::new(crate::tenant::TenantRateRegistry::new(4, 3)),
         };
 
         Agent::new(
@@ -2117,10 +2231,14 @@ mod tests {
                 allow_local_tools: false,
                 max_cost_per_day_cents: None,
                 max_actions_per_hour: None,
+                max_cost_per_user_per_day_cents: None,
                 max_tool_iterations,
                 auto_approve_tools: true,
                 default_timezone: "UTC".to_string(),
                 max_tokens_per_job: 0,
+                multi_tenant: false,
+                max_llm_concurrent_per_user: None,
+                max_jobs_concurrent_per_user: None,
             },
             deps,
             Arc::new(ChannelManager::new()),
@@ -2155,13 +2273,14 @@ mod tests {
 
         let message = IncomingMessage::new("test", "test-user", "do something");
         let initial_messages = vec![ChatMessage::user("do something")];
+        let tenant = agent.tenant_ctx("test-user").await;
 
         // The dispatcher must terminate within 5 seconds. If there is an
         // infinite loop bug (e.g., index not advancing on tool failure), the
         // timeout will fire and the test will fail.
         let result = tokio::time::timeout(
             Duration::from_secs(5),
-            agent.run_agentic_loop(&message, session, thread_id, initial_messages),
+            agent.run_agentic_loop(&message, tenant, session, thread_id, initial_messages),
         )
         .await;
 
@@ -2223,6 +2342,7 @@ mod tests {
                 sandbox_readiness: crate::agent::routine_engine::SandboxReadiness::DisabledByConfig,
                 builder: None,
                 llm_backend: "nearai".to_string(),
+                tenant_rates: Arc::new(crate::tenant::TenantRateRegistry::new(4, 3)),
             };
 
             Agent::new(
@@ -2238,10 +2358,14 @@ mod tests {
                     allow_local_tools: false,
                     max_cost_per_day_cents: None,
                     max_actions_per_hour: None,
+                    max_cost_per_user_per_day_cents: None,
                     max_tool_iterations: max_iter,
                     auto_approve_tools: true,
                     default_timezone: "UTC".to_string(),
                     max_tokens_per_job: 0,
+                    multi_tenant: false,
+                    max_llm_concurrent_per_user: None,
+                    max_jobs_concurrent_per_user: None,
                 },
                 deps,
                 Arc::new(ChannelManager::new()),
@@ -2261,13 +2385,14 @@ mod tests {
 
         let message = IncomingMessage::new("test", "test-user", "keep calling tools");
         let initial_messages = vec![ChatMessage::user("keep calling tools")];
+        let tenant = agent.tenant_ctx("test-user").await;
 
         // Even with an LLM that always wants to call tools, the dispatcher
         // must terminate within the timeout thanks to force_text at
         // max_tool_iterations.
         let result = tokio::time::timeout(
             Duration::from_secs(5),
-            agent.run_agentic_loop(&message, session, thread_id, initial_messages),
+            agent.run_agentic_loop(&message, tenant, session, thread_id, initial_messages),
         )
         .await;
 

src/agent/heartbeat.rs

@@ -31,8 +31,8 @@ use chrono_tz::Tz;
 use tokio::sync::mpsc;
 
 use crate::channels::OutgoingResponse;
-use crate::db::Database;
 use crate::llm::{ChatMessage, CompletionRequest, LlmProvider, Reasoning};
+use crate::tenant::AdminScope;
 use crate::workspace::Workspace;
 use crate::workspace::hygiene::HygieneConfig;
 
@@ -57,6 +57,9 @@ pub struct HeartbeatConfig {
     pub quiet_hours_end: Option<u32>,
     /// Timezone for fire_at and quiet hours evaluation (IANA name).
     pub timezone: Option<String>,
+    /// When true, cycle through all users with routines instead of
+    /// running heartbeat for a single user. Requires a database store.
+    pub multi_tenant: bool,
 }
 
 impl Default for HeartbeatConfig {
@@ -71,6 +74,7 @@ impl Default for HeartbeatConfig {
             quiet_hours_start: None,
             quiet_hours_end: None,
             timezone: None,
+            multi_tenant: false,
         }
     }
 }
@@ -178,7 +182,7 @@ pub struct HeartbeatRunner {
     workspace: Arc<Workspace>,
     llm: Arc<dyn LlmProvider>,
     response_tx: Option<mpsc::Sender<OutgoingResponse>>,
-    store: Option<Arc<dyn Database>>,
+    store: Option<AdminScope>,
     consecutive_failures: u32,
 }
 
@@ -207,8 +211,8 @@ impl HeartbeatRunner {
         self
     }
 
-    /// Set the database store for persistent heartbeat conversations.
-    pub fn with_store(mut self, store: Arc<dyn Database>) -> Self {
+    /// Set the admin-scoped database store for persistent heartbeat conversations.
+    pub fn with_store(mut self, store: AdminScope) -> Self {
         self.store = Some(store);
         self
     }
@@ -396,7 +400,7 @@ impl HeartbeatRunner {
     }
 
     /// Send a notification about heartbeat findings.
-    async fn send_notification(&self, message: &str) {
+    pub(crate) async fn send_notification(&self, message: &str) {
         let Some(ref tx) = self.response_tx else {
             tracing::debug!("No response channel configured for heartbeat notifications");
             return;
@@ -493,7 +497,7 @@ pub fn spawn_heartbeat(
     workspace: Arc<Workspace>,
     llm: Arc<dyn LlmProvider>,
     response_tx: Option<mpsc::Sender<OutgoingResponse>>,
-    store: Option<Arc<dyn Database>>,
+    store: Option<AdminScope>,
 ) -> tokio::task::JoinHandle<()> {
     let mut runner = HeartbeatRunner::new(config, hygiene_config, workspace, llm);
     if let Some(tx) = response_tx {
@@ -508,6 +512,179 @@ pub fn spawn_heartbeat(
     })
 }
 
+/// Spawn a multi-user heartbeat runner that cycles through all users that
+/// own routines (enabled or not). Each tick, it queries the DB for distinct
+/// user_ids, creates a per-user workspace, and runs a heartbeat check for
+/// each user concurrently. Per-user failure counts are tracked independently.
+pub fn spawn_multi_user_heartbeat(
+    config: HeartbeatConfig,
+    hygiene_config: HygieneConfig,
+    llm: Arc<dyn LlmProvider>,
+    response_tx: Option<mpsc::Sender<OutgoingResponse>>,
+    store: AdminScope,
+) -> tokio::task::JoinHandle<()> {
+    tokio::spawn(async move {
+        if !config.enabled {
+            tracing::info!("Multi-user heartbeat is disabled");
+            return;
+        }
+
+        let mut tick_interval = if config.fire_at.is_none() {
+            let mut iv = tokio::time::interval(config.interval);
+            iv.tick().await; // skip immediate tick
+            Some(iv)
+        } else {
+            None
+        };
+
+        // Track consecutive failures per user so we can disable heartbeat
+        // for persistently-failing users (same semantics as single-user mode).
+        let mut user_failures: std::collections::HashMap<String, u32> =
+            std::collections::HashMap::new();
+
+        tracing::info!("Starting multi-user heartbeat loop");
+
+        loop {
+            if let Some(fire_at) = config.fire_at {
+                let sleep_dur = duration_until_next_fire(fire_at, config.resolved_tz());
+                tokio::time::sleep(sleep_dur).await;
+            } else if let Some(ref mut iv) = tick_interval {
+                iv.tick().await;
+            }
+
+            if config.is_quiet_hours() {
+                continue;
+            }
+
+            // Get distinct user_ids from routines
+            let user_ids = match store.list_all_routines().await {
+                Ok(routines) => {
+                    let mut ids: Vec<String> = routines
+                        .iter()
+                        .map(|r| r.user_id.clone())
+                        .collect::<std::collections::HashSet<_>>()
+                        .into_iter()
+                        .collect();
+                    ids.sort();
+                    ids
+                }
+                Err(e) => {
+                    tracing::error!("Multi-user heartbeat: failed to list routines: {}", e);
+                    continue;
+                }
+            };
+
+            // Run user heartbeats concurrently so one slow LLM call doesn't
+            // block others. Cap concurrency to avoid flooding the LLM provider.
+            const MAX_CONCURRENT_HEARTBEATS: usize = 8;
+            let mut join_set = tokio::task::JoinSet::new();
+
+            for user_id in &user_ids {
+                // Skip users that have exceeded max_failures
+                let failures = user_failures.get(user_id).copied().unwrap_or(0);
+                if failures >= config.max_failures {
+                    continue;
+                }
+
+                let workspace = Arc::new(Workspace::new_with_db(user_id, Arc::clone(store.db())));
+
+                // Run memory hygiene per user (same as single-user heartbeat).
+                let hygiene_ws = Arc::clone(&workspace);
+                let hygiene_cfg = hygiene_config.clone();
+                let hygiene_user = user_id.clone();
+                tokio::spawn(async move {
+                    let report =
+                        crate::workspace::hygiene::run_if_due(&hygiene_ws, &hygiene_cfg).await;
+                    if report.had_work() {
+                        tracing::info!(
+                            user_id = hygiene_user,
+                            daily_logs_deleted = report.daily_logs_deleted,
+                            conversation_docs_deleted = report.conversation_docs_deleted,
+                            "multi-user heartbeat: memory hygiene deleted stale documents"
+                        );
+                    }
+                });
+
+                // Drain completed tasks to stay within the concurrency cap.
+                while join_set.len() >= MAX_CONCURRENT_HEARTBEATS {
+                    if let Some(join_result) = join_set.join_next().await {
+                        collect_heartbeat_result(join_result, &mut user_failures, &config);
+                    }
+                }
+
+                let uid = user_id.clone();
+                let cfg = config.clone();
+                let hyg = hygiene_config.clone();
+                let llm_clone = llm.clone();
+                let tx = response_tx.clone();
+                let admin = store.clone();
+
+                join_set.spawn(async move {
+                    let mut runner = HeartbeatRunner::new(cfg, hyg, workspace, llm_clone);
+                    if let Some(tx) = tx {
+                        runner = runner.with_response_channel(tx);
+                    }
+                    runner = runner.with_store(admin);
+
+                    let result = runner.check_heartbeat().await;
+                    if let HeartbeatResult::NeedsAttention(msg) = &result {
+                        runner.send_notification(msg).await;
+                    }
+                    (uid, result)
+                });
+            }
+
+            // Collect remaining results and update failure counts
+            while let Some(join_result) = join_set.join_next().await {
+                collect_heartbeat_result(join_result, &mut user_failures, &config);
+            }
+        }
+    })
+}
+
+/// Process a single JoinSet result from the multi-user heartbeat loop.
+fn collect_heartbeat_result(
+    join_result: Result<(String, HeartbeatResult), tokio::task::JoinError>,
+    user_failures: &mut std::collections::HashMap<String, u32>,
+    config: &HeartbeatConfig,
+) {
+    let (uid, result) = match join_result {
+        Ok(pair) => pair,
+        Err(e) => {
+            tracing::error!("Multi-user heartbeat task panicked: {}", e);
+            return;
+        }
+    };
+    match result {
+        HeartbeatResult::Ok => {
+            tracing::trace!(user_id = uid, "Multi-user heartbeat OK");
+            user_failures.remove(&uid);
+        }
+        HeartbeatResult::NeedsAttention(_) => {
+            tracing::info!(user_id = uid, "Multi-user heartbeat needs attention");
+            user_failures.remove(&uid);
+        }
+        HeartbeatResult::Skipped => {}
+        HeartbeatResult::Failed(err) => {
+            let count = user_failures.entry(uid.clone()).or_insert(0);
+            *count += 1;
+            tracing::error!(
+                user_id = uid,
+                consecutive_failures = *count,
+                "Multi-user heartbeat failed: {}",
+                err
+            );
+            if *count >= config.max_failures {
+                tracing::error!(
+                    user_id = uid,
+                    "Multi-user heartbeat disabled for user after {} consecutive failures",
+                    count
+                );
+            }
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -726,7 +903,7 @@ mod tests {
             Arc<crate::workspace::Workspace>,
             Arc<dyn crate::llm::LlmProvider>,
             Option<tokio::sync::mpsc::Sender<crate::channels::OutgoingResponse>>,
-            Option<Arc<dyn crate::db::Database>>,
+            Option<AdminScope>,
         ) -> tokio::task::JoinHandle<()> = spawn_heartbeat;
         let _ = _fn_ptr;
     }

src/agent/mod.rs

@@ -36,7 +36,9 @@ pub(crate) use agent_loop::truncate_for_preview;
 pub use agent_loop::{Agent, AgentDeps};
 pub use compaction::{CompactionResult, ContextCompactor};
 pub use context_monitor::{CompactionStrategy, ContextBreakdown, ContextMonitor};
-pub use heartbeat::{HeartbeatConfig, HeartbeatResult, HeartbeatRunner, spawn_heartbeat};
+pub use heartbeat::{
+    HeartbeatConfig, HeartbeatResult, HeartbeatRunner, spawn_heartbeat, spawn_multi_user_heartbeat,
+};
 pub use router::{MessageIntent, Router};
 pub use routine::{Routine, RoutineAction, RoutineRun, Trigger};
 pub use routine_engine::{RoutineEngine, SandboxReadiness};

src/agent/routine_engine.rs

@@ -18,6 +18,7 @@ use std::time::Duration;
 use chrono::Utc;
 use regex::Regex;
 use tokio::sync::{RwLock, mpsc};
+use tokio::task::JoinHandle;
 use uuid::Uuid;
 
 use crate::agent::Scheduler;
@@ -27,12 +28,12 @@ use crate::agent::routine::{
 use crate::channels::{IncomingMessage, OutgoingResponse};
 use crate::config::RoutineConfig;
 use crate::context::{JobContext, JobState};
-use crate::db::Database;
 use crate::error::RoutineError;
 use crate::extensions::ExtensionManager;
 use crate::llm::{
     ChatMessage, CompletionRequest, FinishReason, LlmProvider, ToolCall, ToolCompletionRequest,
 };
+use crate::tenant::AdminScope;
 use crate::tools::{
     ToolError, ToolRegistry, autonomous_allowed_tool_names, autonomous_unavailable_message,
     prepare_tool_params,
@@ -45,6 +46,11 @@ enum EventMatcher {
     System { routine: Routine },
 }
 
+struct TriggeredRoutine {
+    routine: Routine,
+    detail: String,
+}
+
 /// Distinguishes why sandbox is unavailable so error messages are accurate.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum SandboxReadiness {
@@ -93,7 +99,7 @@ pub(crate) fn routine_matches_message(routine: &Routine, message: &IncomingMessa
 /// The routine execution engine.
 pub struct RoutineEngine {
     config: RoutineConfig,
-    store: Arc<dyn Database>,
+    store: AdminScope,
     llm: Arc<dyn LlmProvider>,
     workspace: Arc<Workspace>,
     /// Sender for notifications (routed to channel manager).
@@ -122,7 +128,7 @@ impl RoutineEngine {
     #[allow(clippy::too_many_arguments)]
     pub fn new(
         config: RoutineConfig,
-        store: Arc<dyn Database>,
+        store: AdminScope,
         llm: Arc<dyn LlmProvider>,
         workspace: Arc<Workspace>,
         notify_tx: mpsc::Sender<OutgoingResponse>,
@@ -202,6 +208,44 @@ impl RoutineEngine {
 
     /// Check incoming message against event triggers. Returns number of routines fired.
     pub async fn check_event_triggers(&self, message: &IncomingMessage, content: &str) -> usize {
+        let triggered = self.matching_event_triggers(message, content).await;
+        let fired = triggered.len();
+        for triggered in triggered {
+            std::mem::drop(self.spawn_fire(triggered.routine, "event", Some(triggered.detail)));
+        }
+        fired
+    }
+
+    /// Fire matching event-triggered routines and wait for them to complete.
+    ///
+    /// Used by single-message REPL mode so the process does not exit before
+    /// background event-triggered routines finish.
+    pub async fn check_event_triggers_and_wait(
+        &self,
+        message: &IncomingMessage,
+        content: &str,
+    ) -> usize {
+        let triggered = self.matching_event_triggers(message, content).await;
+        let fired = triggered.len();
+        let handles: Vec<JoinHandle<()>> = triggered
+            .into_iter()
+            .map(|triggered| self.spawn_fire(triggered.routine, "event", Some(triggered.detail)))
+            .collect();
+
+        for handle in handles {
+            if let Err(e) = handle.await {
+                tracing::warn!(error = %e, "Event-triggered routine task failed");
+            }
+        }
+
+        fired
+    }
+
+    async fn matching_event_triggers(
+        &self,
+        message: &IncomingMessage,
+        content: &str,
+    ) -> Vec<TriggeredRoutine> {
         let cache = self.event_cache.read().await;
 
         // Early return if there are no message matchers at all.
@@ -209,10 +253,9 @@ impl RoutineEngine {
             .iter()
             .any(|m| matches!(m, EventMatcher::Message { .. }))
         {
-            return 0;
+            return Vec::new();
         }
-
-        let mut fired = 0;
+        let mut triggered = Vec::new();
 
         // Collect routine IDs for batch query
         let routine_ids: Vec<Uuid> = cache
@@ -224,13 +267,13 @@ impl RoutineEngine {
             .collect();
 
         if routine_ids.is_empty() {
-            return 0;
+            return Vec::new();
         }
 
         // Single batch query instead of N queries
         let concurrent_counts = match self.batch_concurrent_counts(&routine_ids).await {
             Some(counts) => counts,
-            None => return 0,
+            None => return Vec::new(),
         };
 
         for matcher in cache.iter() {
@@ -285,11 +328,13 @@ impl RoutineEngine {
             }
 
             let detail = truncate(content, 200);
-            self.spawn_fire(routine.clone(), "event", Some(detail));
-            fired += 1;
+            triggered.push(TriggeredRoutine {
+                routine: routine.clone(),
+                detail,
+            });
         }
 
-        fired
+        triggered
     }
 
     /// Emit a structured event to system-event routines.
@@ -737,12 +782,22 @@ impl RoutineEngine {
             });
         }
 
+        // Per-user workspace (same pattern as spawn_fire).
+        let routine_workspace = if routine.user_id == self.workspace.user_id() {
+            self.workspace.clone()
+        } else {
+            Arc::new(Workspace::new_with_db(
+                &routine.user_id,
+                Arc::clone(self.store.db()),
+            ))
+        };
+
         // Execute inline for manual triggers (caller wants to wait)
         let engine = EngineContext {
             config: self.config.clone(),
             store: self.store.clone(),
             llm: self.llm.clone(),
-            workspace: self.workspace.clone(),
+            workspace: routine_workspace,
             notify_tx: self.notify_tx.clone(),
             running_count: self.running_count.clone(),
             scheduler: self.scheduler.clone(),
@@ -845,7 +900,12 @@ impl RoutineEngine {
     }
 
     /// Spawn a fire in a background task.
-    fn spawn_fire(&self, routine: Routine, trigger_type: &str, trigger_detail: Option<String>) {
+    fn spawn_fire(
+        &self,
+        routine: Routine,
+        trigger_type: &str,
+        trigger_detail: Option<String>,
+    ) -> JoinHandle<()> {
         let run = RoutineRun {
             id: Uuid::new_v4(),
             routine_id: routine.id,
@@ -860,11 +920,23 @@ impl RoutineEngine {
             created_at: Utc::now(),
         };
 
+        // Use per-user workspace so each routine executes in the correct
+        // user's context. Fall back to the engine-wide workspace when the
+        // routine belongs to the same user (avoids unnecessary allocation).
+        let routine_workspace = if routine.user_id == self.workspace.user_id() {
+            self.workspace.clone()
+        } else {
+            Arc::new(Workspace::new_with_db(
+                &routine.user_id,
+                Arc::clone(self.store.db()),
+            ))
+        };
+
         let engine = EngineContext {
             config: self.config.clone(),
             store: self.store.clone(),
             llm: self.llm.clone(),
-            workspace: self.workspace.clone(),
+            workspace: routine_workspace,
             notify_tx: self.notify_tx.clone(),
             running_count: self.running_count.clone(),
             scheduler: self.scheduler.clone(),
@@ -882,7 +954,7 @@ impl RoutineEngine {
                 return;
             }
             execute_routine(engine, routine, run).await;
-        });
+        })
     }
 
     fn check_cooldown(&self, routine: &Routine) -> bool {
@@ -917,7 +989,7 @@ impl RoutineEngine {
 /// an active state (Pending/InProgress/Stuck). Maps the final `JobState` to
 /// a `RunStatus` for the routine run.
 struct FullJobWatcher {
-    store: Arc<dyn Database>,
+    store: AdminScope,
     job_id: Uuid,
     routine_name: String,
 }
@@ -928,7 +1000,7 @@ impl FullJobWatcher {
     /// Safety ceiling: 24 hours, derived from POLL_INTERVAL.
     const MAX_POLLS: u32 = (24 * 60 * 60) / Self::POLL_INTERVAL.as_secs() as u32;
 
-    fn new(store: Arc<dyn Database>, job_id: Uuid, routine_name: String) -> Self {
+    fn new(store: AdminScope, job_id: Uuid, routine_name: String) -> Self {
         Self {
             store,
             job_id,
@@ -1000,7 +1072,7 @@ impl FullJobWatcher {
 /// Shared context passed to the execution function.
 struct EngineContext {
     config: RoutineConfig,
-    store: Arc<dyn Database>,
+    store: AdminScope,
     llm: Arc<dyn LlmProvider>,
     workspace: Arc<Workspace>,
     notify_tx: mpsc::Sender<OutgoingResponse>,

src/agent/scheduler.rs

@@ -11,12 +11,12 @@ use uuid::Uuid;
 use crate::agent::task::{Task, TaskContext, TaskOutput};
 use crate::config::AgentConfig;
 use crate::context::{ContextManager, JobContext, JobState};
-use crate::db::Database;
 use crate::error::{Error, JobError};
 use crate::extensions::ExtensionManager;
 use crate::hooks::HookRegistry;
 use crate::llm::LlmProvider;
 use crate::safety::SafetyLayer;
+use crate::tenant::AdminScope;
 use crate::tools::{
     ApprovalContext, ToolRegistry, autonomous_allowed_tool_names, autonomous_unavailable_error,
     prepare_tool_params,
@@ -52,7 +52,7 @@ struct ScheduledSubtask {
 pub struct SchedulerDeps {
     pub tools: Arc<ToolRegistry>,
     pub extension_manager: Option<Arc<ExtensionManager>>,
-    pub store: Option<Arc<dyn Database>>,
+    pub store: Option<AdminScope>,
     pub hooks: Arc<HookRegistry>,
 }
 
@@ -64,7 +64,7 @@ pub struct Scheduler {
     safety: Arc<SafetyLayer>,
     tools: Arc<ToolRegistry>,
     extension_manager: Option<Arc<ExtensionManager>>,
-    store: Option<Arc<dyn Database>>,
+    store: Option<AdminScope>,
     hooks: Arc<HookRegistry>,
     /// SSE manager for live job event streaming.
     sse_tx: Option<Arc<crate::channels::web::sse::SseManager>>,
@@ -780,10 +780,14 @@ mod tests {
             allow_local_tools: true,
             max_cost_per_day_cents: None,
             max_actions_per_hour: None,
+            max_cost_per_user_per_day_cents: None,
             max_tool_iterations: 10,
             auto_approve_tools: true,
             default_timezone: "UTC".to_string(),
             max_tokens_per_job,
+            multi_tenant: false,
+            max_llm_concurrent_per_user: None,
+            max_jobs_concurrent_per_user: None,
         };
         let cm = Arc::new(ContextManager::new(5));
         let llm: Arc<dyn LlmProvider> = Arc::new(StubLlm);

src/agent/self_repair.rs

@@ -8,8 +8,8 @@ use chrono::{DateTime, Utc};
 use uuid::Uuid;
 
 use crate::context::{ContextManager, JobState};
-use crate::db::Database;
 use crate::error::RepairError;
+use crate::tenant::AdminScope;
 use crate::tools::{BuildRequirement, Language, SoftwareBuilder, SoftwareType, ToolRegistry};
 
 /// A job that has been detected as stuck.
@@ -69,7 +69,7 @@ pub struct DefaultSelfRepair {
     /// Jobs in `InProgress` longer than this are treated as stuck.
     stuck_threshold: Duration,
     max_repair_attempts: u32,
-    store: Option<Arc<dyn Database>>,
+    store: Option<AdminScope>,
     builder: Option<Arc<dyn SoftwareBuilder>>,
     tools: Option<Arc<ToolRegistry>>,
 }
@@ -91,8 +91,8 @@ impl DefaultSelfRepair {
         }
     }
 
-    /// Add a Store for tool failure tracking.
-    pub fn with_store(mut self, store: Arc<dyn Database>) -> Self {
+    /// Add an admin-scoped store for tool failure tracking.
+    pub fn with_store(mut self, store: AdminScope) -> Self {
         self.store = Some(store);
         self
     }
@@ -806,7 +806,7 @@ mod tests {
         // Create self-repair with zero threshold (detect immediately),
         // wired with store, builder, and tools.
         let repair = DefaultSelfRepair::new(Arc::clone(&cm), Duration::from_secs(0), 3)
-            .with_store(Arc::clone(&db))
+            .with_store(crate::tenant::AdminScope::new(Arc::clone(&db)))
             .with_builder(
                 Arc::clone(&builder) as Arc<dyn crate::tools::SoftwareBuilder>,
                 tools,

src/agent/session.rs

@@ -449,6 +449,7 @@ impl Thread {
                         id: call_id.clone(),
                         name: tc.name.clone(),
                         arguments: tc.parameters.clone(),
+                        reasoning: None,
                     })
                     .collect();
 
@@ -522,7 +523,12 @@ impl Thread {
                             && let Some(ref tcs) = assistant_msg.tool_calls
                         {
                             for tc in tcs {
-                                turn.record_tool_call(&tc.name, tc.arguments.clone());
+                                turn.record_tool_call_with_reasoning(
+                                    &tc.name,
+                                    tc.arguments.clone(),
+                                    tc.reasoning.clone(),
+                                    Some(tc.id.clone()),
+                                );
                             }
                         }
 
@@ -602,6 +608,10 @@ pub struct Turn {
     pub completed_at: Option<DateTime<Utc>>,
     /// Error message (if failed).
     pub error: Option<String>,
+    /// Agent's reasoning narrative for this turn.
+    /// Cleaned via `clean_response` and sanitized through `SafetyLayer` before storage.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub narrative: Option<String>,
     /// Transient image content parts for multimodal LLM input.
     /// Not serialized — images are only needed for the current LLM call.
     /// The text description in `user_input` persists for compaction/context.
@@ -621,6 +631,7 @@ impl Turn {
             started_at: Utc::now(),
             completed_at: None,
             error: None,
+            narrative: None,
             image_content_parts: Vec::new(),
         }
     }
@@ -656,6 +667,26 @@ impl Turn {
             parameters: params,
             result: None,
             error: None,
+            rationale: None,
+            tool_call_id: None,
+        });
+    }
+
+    /// Record a tool call with reasoning context.
+    pub fn record_tool_call_with_reasoning(
+        &mut self,
+        name: impl Into<String>,
+        params: serde_json::Value,
+        rationale: Option<String>,
+        tool_call_id: Option<String>,
+    ) {
+        self.tool_calls.push(TurnToolCall {
+            name: name.into(),
+            parameters: params,
+            result: None,
+            error: None,
+            rationale,
+            tool_call_id,
         });
     }
 
@@ -672,6 +703,60 @@ impl Turn {
             call.error = Some(error.into());
         }
     }
+
+    /// Record a tool result by tool_call_id, with fallback to first pending call.
+    pub fn record_tool_result_for(&mut self, tool_call_id: &str, result: serde_json::Value) {
+        if let Some(call) = self
+            .tool_calls
+            .iter_mut()
+            .find(|c| c.tool_call_id.as_deref() == Some(tool_call_id))
+        {
+            call.result = Some(result);
+        } else if let Some(call) = self
+            .tool_calls
+            .iter_mut()
+            .find(|c| c.result.is_none() && c.error.is_none())
+        {
+            tracing::debug!(
+                tool_call_id = %tool_call_id,
+                fallback_tool = %call.name,
+                "tool_call_id not found, falling back to first pending call"
+            );
+            call.result = Some(result);
+        } else {
+            tracing::warn!(
+                tool_call_id = %tool_call_id,
+                "Tool result dropped: no matching or pending tool call"
+            );
+        }
+    }
+
+    /// Record a tool error by tool_call_id, with fallback to first pending call.
+    pub fn record_tool_error_for(&mut self, tool_call_id: &str, error: impl Into<String>) {
+        if let Some(call) = self
+            .tool_calls
+            .iter_mut()
+            .find(|c| c.tool_call_id.as_deref() == Some(tool_call_id))
+        {
+            call.error = Some(error.into());
+        } else if let Some(call) = self
+            .tool_calls
+            .iter_mut()
+            .find(|c| c.result.is_none() && c.error.is_none())
+        {
+            tracing::debug!(
+                tool_call_id = %tool_call_id,
+                fallback_tool = %call.name,
+                "tool_call_id not found, falling back to first pending call"
+            );
+            call.error = Some(error.into());
+        } else {
+            tracing::warn!(
+                tool_call_id = %tool_call_id,
+                "Tool error dropped: no matching or pending tool call"
+            );
+        }
+    }
 }
 
 /// Record of a tool call made during a turn.
@@ -685,6 +770,12 @@ pub struct TurnToolCall {
     pub result: Option<serde_json::Value>,
     /// Error from the tool (if failed).
     pub error: Option<String>,
+    /// Agent's reasoning for choosing this tool.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub rationale: Option<String>,
+    /// The tool_call_id from the LLM, for identity-based result matching.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tool_call_id: Option<String>,
 }
 
 #[cfg(test)]
@@ -1309,6 +1400,7 @@ mod tests {
             id: "call_0".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "test"}),
+            reasoning: None,
         };
         let messages = vec![
             ChatMessage::user("Find test"),
@@ -1339,6 +1431,7 @@ mod tests {
             id: "call_0".to_string(),
             name: "http".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
         let messages = vec![
             ChatMessage::user("Fetch URL"),
@@ -1404,11 +1497,13 @@ mod tests {
             id: "call_a".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "data"}),
+            reasoning: None,
         };
         let tc2 = ToolCall {
             id: "call_b".to_string(),
             name: "write".to_string(),
             arguments: serde_json::json!({"path": "out.txt"}),
+            reasoning: None,
         };
         let messages = vec![
             ChatMessage::user("Find and save"),
@@ -1620,4 +1715,100 @@ mod tests {
         let merged = thread.drain_pending_messages().unwrap();
         assert_eq!(merged, "failed batch\nnew msg");
     }
+
+    #[test]
+    fn test_record_tool_result_for_by_id() {
+        let mut turn = Turn::new(0, "test");
+        turn.record_tool_call_with_reasoning(
+            "tool_a",
+            serde_json::json!({}),
+            None,
+            Some("id_a".into()),
+        );
+        turn.record_tool_call_with_reasoning(
+            "tool_b",
+            serde_json::json!({}),
+            None,
+            Some("id_b".into()),
+        );
+
+        // Record result for second tool by ID
+        turn.record_tool_result_for("id_b", serde_json::json!("result_b"));
+        assert!(turn.tool_calls[0].result.is_none());
+        assert_eq!(
+            turn.tool_calls[1].result.as_ref().unwrap(),
+            &serde_json::json!("result_b")
+        );
+    }
+
+    #[test]
+    fn test_record_tool_error_for_by_id() {
+        let mut turn = Turn::new(0, "test");
+        turn.record_tool_call_with_reasoning(
+            "tool_a",
+            serde_json::json!({}),
+            None,
+            Some("id_a".into()),
+        );
+        turn.record_tool_call_with_reasoning(
+            "tool_b",
+            serde_json::json!({}),
+            None,
+            Some("id_b".into()),
+        );
+
+        turn.record_tool_error_for("id_a", "failed");
+        assert_eq!(turn.tool_calls[0].error.as_deref(), Some("failed"));
+        assert!(turn.tool_calls[1].error.is_none());
+    }
+
+    #[test]
+    fn test_record_tool_result_for_fallback_to_pending() {
+        let mut turn = Turn::new(0, "test");
+        turn.record_tool_call_with_reasoning(
+            "tool_a",
+            serde_json::json!({}),
+            None,
+            Some("id_a".into()),
+        );
+        turn.record_tool_call_with_reasoning(
+            "tool_b",
+            serde_json::json!({}),
+            None,
+            Some("id_b".into()),
+        );
+
+        // First tool already has a result
+        turn.tool_calls[0].result = Some(serde_json::json!("done"));
+
+        // Unknown ID should fall back to first pending (tool_b)
+        turn.record_tool_result_for("unknown_id", serde_json::json!("fallback"));
+        assert_eq!(
+            turn.tool_calls[0].result.as_ref().unwrap(),
+            &serde_json::json!("done")
+        );
+        assert_eq!(
+            turn.tool_calls[1].result.as_ref().unwrap(),
+            &serde_json::json!("fallback")
+        );
+    }
+
+    #[test]
+    fn test_record_tool_result_for_no_pending_is_noop() {
+        let mut turn = Turn::new(0, "test");
+        turn.record_tool_call_with_reasoning(
+            "tool_a",
+            serde_json::json!({}),
+            None,
+            Some("id_a".into()),
+        );
+        turn.tool_calls[0].result = Some(serde_json::json!("done"));
+
+        // No pending calls, unknown ID — should be a no-op
+        turn.record_tool_result_for("unknown_id", serde_json::json!("lost"));
+        assert_eq!(
+            turn.tool_calls[0].result.as_ref().unwrap(),
+            &serde_json::json!("done")
+        );
+    }
 }

src/agent/submission.rs

@@ -92,6 +92,17 @@ impl SubmissionParser {
                 args: vec![],
             };
         }
+        if lower == "/reasoning" || lower.starts_with("/reasoning ") {
+            let args: Vec<String> = trimmed
+                .split_whitespace()
+                .skip(1)
+                .map(|s| s.to_string())
+                .collect();
+            return Submission::SystemCommand {
+                command: "reasoning".to_string(),
+                args,
+            };
+        }
         if lower == "/restart" {
             tracing::debug!("[SubmissionParser::parse] Recognized /restart command");
             return Submission::SystemCommand {

src/agent/thread_ops.rs

@@ -175,6 +175,7 @@ impl Agent {
     pub(super) async fn process_user_input(
         &self,
         message: &IncomingMessage,
+        tenant: crate::tenant::TenantCtx,
         session: Arc<Mutex<Session>>,
         thread_id: Uuid,
         content: &str,
@@ -351,7 +352,7 @@ impl Agent {
 
         if let Some(intent) = self.router.route_command(&temp_message) {
             // Explicit command like /status, /job, /list - handle directly
-            return self.handle_job_or_command(intent, message).await;
+            return self.handle_job_or_command(intent, message, &tenant).await;
         }
 
         // Natural language goes through the agentic loop
@@ -462,7 +463,7 @@ impl Agent {
 
         // Run the agentic tool execution loop
         let result = self
-            .run_agentic_loop(message, session.clone(), thread_id, turn_messages)
+            .run_agentic_loop(message, tenant, session.clone(), thread_id, turn_messages)
             .await;
 
         // Re-acquire lock and check if interrupted
@@ -513,10 +514,10 @@ impl Agent {
                 };
 
                 thread.complete_turn(&response);
-                let (turn_number, tool_calls) = thread
+                let (turn_number, tool_calls, narrative) = thread
                     .turns
                     .last()
-                    .map(|t| (t.turn_number, t.tool_calls.clone()))
+                    .map(|t| (t.turn_number, t.tool_calls.clone(), t.narrative.clone()))
                     .unwrap_or_default();
                 let _ = self
                     .channels
@@ -534,6 +535,7 @@ impl Agent {
                     &message.user_id,
                     turn_number,
                     &tool_calls,
+                    narrative.as_deref(),
                 )
                 .await;
                 self.persist_assistant_response(
@@ -725,7 +727,9 @@ impl Agent {
     ///
     /// Stored between the user and assistant messages so that
     /// `build_turns_from_db_messages` can reconstruct the tool call history.
-    /// Content is a JSON array of tool call summaries.
+    /// Content is a JSON object: `{ "calls": [...], "narrative": "..." }`.
+    /// The `calls` array contains tool call summaries with optional `rationale`
+    /// and `tool_call_id` fields. Legacy rows may be plain JSON arrays.
     pub(super) async fn persist_tool_calls(
         &self,
         thread_id: Uuid,
@@ -733,6 +737,7 @@ impl Agent {
         user_id: &str,
         turn_number: usize,
         tool_calls: &[crate::agent::session::TurnToolCall],
+        narrative: Option<&str>,
     ) {
         if tool_calls.is_empty() {
             return;
@@ -767,11 +772,30 @@ impl Agent {
                 if let Some(ref error) = tc.error {
                     obj["error"] = serde_json::Value::String(truncate_preview(error, 200));
                 }
+                if let Some(ref rationale) = tc.rationale {
+                    obj["rationale"] = serde_json::Value::String(truncate_preview(rationale, 500));
+                }
+                if let Some(ref tool_call_id) = tc.tool_call_id {
+                    obj["tool_call_id"] =
+                        serde_json::Value::String(truncate_preview(tool_call_id, 128));
+                }
                 obj
             })
             .collect();
 
-        let content = match serde_json::to_string(&summaries) {
+        // Wrap in an object with optional narrative so it can be reconstructed.
+        // safety: no byte-index slicing here; comment describes JSON shape
+        let wrapper = if let Some(n) = narrative {
+            serde_json::json!({
+                "narrative": truncate_preview(n, 1000),
+                "calls": summaries,
+            })
+        } else {
+            serde_json::json!({
+                "calls": summaries,
+            })
+        };
+        let content = match serde_json::to_string(&wrapper) {
             Ok(c) => c,
             Err(e) => {
                 tracing::warn!("Failed to serialize tool calls: {}", e);
@@ -1104,9 +1128,12 @@ impl Agent {
                     && let Some(turn) = thread.last_turn_mut()
                 {
                     if is_tool_error {
-                        turn.record_tool_error(result_content.clone());
+                        turn.record_tool_error_for(&pending.tool_call_id, result_content.clone());
                     } else {
-                        turn.record_tool_result(serde_json::json!(result_content));
+                        turn.record_tool_result_for(
+                            &pending.tool_call_id,
+                            serde_json::json!(result_content),
+                        );
                     }
                 }
             }
@@ -1358,9 +1385,12 @@ impl Agent {
                         && let Some(turn) = thread.last_turn_mut()
                     {
                         if is_deferred_error {
-                            turn.record_tool_error(deferred_content.clone());
+                            turn.record_tool_error_for(&tc.id, deferred_content.clone());
                         } else {
-                            turn.record_tool_result(serde_json::json!(deferred_content));
+                            turn.record_tool_result_for(
+                                &tc.id,
+                                serde_json::json!(deferred_content),
+                            );
                         }
                     }
                 }
@@ -1444,7 +1474,13 @@ impl Agent {
 
             // Continue the agentic loop (a tool was already executed this turn)
             let result = self
-                .run_agentic_loop(message, session.clone(), thread_id, context_messages)
+                .run_agentic_loop(
+                    message,
+                    self.tenant_ctx(&message.user_id).await,
+                    session.clone(),
+                    thread_id,
+                    context_messages,
+                )
                 .await;
 
             // Handle the result
@@ -1459,10 +1495,10 @@ impl Agent {
                     let (response, suggestions) =
                         crate::agent::dispatcher::extract_suggestions(&response);
                     thread.complete_turn(&response);
-                    let (turn_number, tool_calls) = thread
+                    let (turn_number, tool_calls, narrative) = thread
                         .turns
                         .last()
-                        .map(|t| (t.turn_number, t.tool_calls.clone()))
+                        .map(|t| (t.turn_number, t.tool_calls.clone(), t.narrative.clone()))
                         .unwrap_or_default();
                     // User message already persisted at turn start; save tool calls then assistant response
                     self.persist_tool_calls(
@@ -1471,6 +1507,7 @@ impl Agent {
                         &message.user_id,
                         turn_number,
                         &tool_calls,
+                        narrative.as_deref(),
                     )
                     .await;
                     self.persist_assistant_response(
@@ -1816,7 +1853,20 @@ fn rebuild_chat_messages_from_db(
             "assistant" => result.push(ChatMessage::assistant(&msg.content)),
             "tool_calls" => {
                 // Try to parse the enriched JSON and rebuild tool messages.
-                if let Ok(calls) = serde_json::from_str::<Vec<serde_json::Value>>(&msg.content) {
+                // Supports two formats:
+                // - Old: plain JSON array of tool call summaries
+                // - New: wrapped object { "calls": [...], "narrative": "..." }
+                let calls: Vec<serde_json::Value> =
+                    match serde_json::from_str::<serde_json::Value>(&msg.content) {
+                        Ok(serde_json::Value::Array(arr)) => arr,
+                        Ok(serde_json::Value::Object(obj)) => obj
+                            .get("calls")
+                            .and_then(|v| v.as_array())
+                            .cloned()
+                            .unwrap_or_default(),
+                        _ => Vec::new(),
+                    };
+                {
                     if calls.is_empty() {
                         continue;
                     }
@@ -1839,6 +1889,10 @@ fn rebuild_chat_messages_from_db(
                                     .get("parameters")
                                     .cloned()
                                     .unwrap_or(serde_json::json!({})),
+                                reasoning: c
+                                    .get("rationale")
+                                    .and_then(|v| v.as_str())
+                                    .map(String::from),
                             })
                             .collect();
 

src/app.rs

@@ -897,6 +897,7 @@ impl AppBuilder {
             crate::agent::cost_guard::CostGuardConfig {
                 max_cost_per_day_cents: self.config.agent.max_cost_per_day_cents,
                 max_actions_per_hour: self.config.agent.max_actions_per_hour,
+                max_cost_per_user_per_day_cents: self.config.agent.max_cost_per_user_per_day_cents,
             },
         ));
 

src/channels/channel.rs

@@ -265,6 +265,15 @@ impl OutgoingResponse {
     }
 }
 
+/// A single tool decision within a reasoning update.
+#[derive(Debug, Clone)]
+pub struct ToolDecision {
+    /// Tool name.
+    pub tool_name: String,
+    /// Agent's reasoning for choosing this tool.
+    pub rationale: String,
+}
+
 /// Status update types for showing agent activity.
 #[derive(Debug, Clone)]
 pub enum StatusUpdate {
@@ -333,6 +342,13 @@ pub enum StatusUpdate {
     },
     /// Suggested follow-up messages for the user.
     Suggestions { suggestions: Vec<String> },
+    /// Agent reasoning update (why it chose specific tools).
+    ReasoningUpdate {
+        /// Human-readable summary of the agent's decision.
+        narrative: String,
+        /// Per-tool decisions.
+        decisions: Vec<ToolDecision>,
+    },
     /// Per-turn token usage and cost summary (shown as subtle metadata).
     TurnCost {
         input_tokens: u64,

src/channels/mod.rs

@@ -39,7 +39,7 @@ mod webhook_server;
 
 pub use channel::{
     AttachmentKind, Channel, ChannelSecretUpdater, IncomingAttachment, IncomingMessage,
-    MessageStream, OutgoingResponse, StatusUpdate, routing_target_from_metadata,
+    MessageStream, OutgoingResponse, StatusUpdate, ToolDecision, routing_target_from_metadata,
 };
 pub use http::{HttpChannel, HttpChannelState};
 pub use manager::ChannelManager;

src/channels/repl.rs

@@ -75,6 +75,7 @@ const SLASH_COMMANDS: &[&str] = &[
     "/suggest",
     "/thread",
     "/resume",
+    "/reasoning",
 ];
 
 /// Rustyline helper for slash-command tab completion.
@@ -430,6 +431,18 @@ impl ReplChannel {
             let _ = execute!(stderr, terminal::Clear(terminal::ClearType::FromCursorDown));
         }
     }
+
+    async fn finish_single_message_turn(&self) {
+        if self.single_message.is_none() {
+            return;
+        }
+
+        let tx = self.msg_tx.lock().ok().and_then(|mut guard| guard.take());
+        if let Some(tx) = tx {
+            let msg = IncomingMessage::new("repl", &self.user_id, "/quit");
+            let _ = tx.send(msg).await;
+        }
+    }
 }
 
 impl Default for ReplChannel {
@@ -479,7 +492,9 @@ impl Channel for ReplChannel {
 
     async fn start(&self) -> Result<MessageStream, ChannelError> {
         let (tx, rx) = mpsc::channel(32);
-        // Store tx so send_status can inject approval responses directly
+        // Approval prompts inject responses back through this sender.
+        // In single-message mode we keep it until the turn finishes, then
+        // drop it after enqueuing /quit so the receiver stream can close.
         if let Ok(mut guard) = self.msg_tx.lock() {
             *guard = Some(tx.clone());
         }
@@ -495,11 +510,10 @@ impl Channel for ReplChannel {
 
             // Single message mode: send it and return
             if let Some(msg) = single_message {
-                let incoming = IncomingMessage::new("repl", &user_id, &msg).with_timezone(&sys_tz);
+                let incoming = IncomingMessage::new("repl", &user_id, &msg)
+                    .with_metadata(serde_json::json!({ "single_message_mode": true }))
+                    .with_timezone(&sys_tz);
                 let _ = tx.blocking_send(incoming);
-                // Ensure the agent exits after handling exactly one turn in -m mode,
-                // even when other channels (gateway/http) are enabled.
-                let _ = tx.blocking_send(IncomingMessage::new("repl", &user_id, "/quit"));
                 return;
             }
 
@@ -662,6 +676,7 @@ impl Channel for ReplChannel {
             println!();
             println!();
             self.stdin_locked.store(false, Ordering::Relaxed);
+            self.finish_single_message_turn().await;
             return Ok(());
         }
 
@@ -680,6 +695,7 @@ impl Channel for ReplChannel {
         println!();
         // Unlock stdin so readline can resume
         self.stdin_locked.store(false, Ordering::Relaxed);
+        self.finish_single_message_turn().await;
         Ok(())
     }
 
@@ -779,6 +795,7 @@ impl Channel for ReplChannel {
                 let msg_tx = Arc::clone(&self.msg_tx);
                 let user_id = self.user_id.clone();
                 let lock_flag = Arc::clone(&self.stdin_locked);
+                let single_message_mode = self.single_message.is_some();
                 tokio::task::spawn_blocking(move || {
                     let action = run_approval_selector(allow_always).unwrap_or("n");
                     // Unlock stdin so readline can resume after approval
@@ -787,7 +804,12 @@ impl Channel for ReplChannel {
                         return;
                     };
                     if let Some(tx) = guard.as_ref() {
-                        let msg = IncomingMessage::new("repl", &user_id, action);
+                        let msg = if single_message_mode {
+                            IncomingMessage::new("repl", &user_id, action)
+                                .with_metadata(serde_json::json!({ "single_message_mode": true }))
+                        } else {
+                            IncomingMessage::new("repl", &user_id, action)
+                        };
                         let _ = tx.blocking_send(msg);
                     }
                 });
@@ -841,6 +863,19 @@ impl Channel for ReplChannel {
             StatusUpdate::Suggestions { .. } => {
                 // Suggestions are only rendered by the web gateway
             }
+            StatusUpdate::ReasoningUpdate {
+                narrative,
+                decisions,
+            } => {
+                if !narrative.is_empty() {
+                    let display = truncate_for_preview(&narrative, CLI_STATUS_MAX);
+                    eprintln!("  \x1b[94m\u{25B6} {display}\x1b[0m");
+                }
+                for d in &decisions {
+                    let display = truncate_for_preview(&d.rationale, CLI_STATUS_MAX);
+                    eprintln!("    \x1b[90m\u{2192} {}: {display}\x1b[0m", d.tool_name);
+                }
+            }
             StatusUpdate::TurnCost { .. } => {
                 // Cost display is handled by the TUI channel
             }
@@ -875,6 +910,7 @@ impl Channel for ReplChannel {
 #[cfg(test)]
 mod tests {
     use futures::StreamExt;
+    use tokio::time::{Duration, timeout};
 
     use super::*;
 
@@ -883,16 +919,36 @@ mod tests {
         let repl = ReplChannel::with_message("hi".to_string());
         let mut stream = repl.start().await.expect("repl start should succeed");
 
-        let first = stream.next().await.expect("first message missing");
+        let first = timeout(Duration::from_secs(1), stream.next())
+            .await
+            .expect("timed out waiting for first message")
+            .expect("first message missing");
         assert_eq!(first.channel, "repl");
         assert_eq!(first.content, "hi");
 
-        let second = stream.next().await.expect("quit message missing");
+        assert!(
+            timeout(Duration::from_millis(100), stream.next())
+                .await
+                .is_err(),
+            "single-message mode should wait for the turn to finish before quitting"
+        );
+
+        repl.respond(&first, OutgoingResponse::text("done"))
+            .await
+            .expect("respond should succeed");
+
+        let second = timeout(Duration::from_secs(1), stream.next())
+            .await
+            .expect("timed out waiting for quit message")
+            .expect("quit message missing");
         assert_eq!(second.channel, "repl");
         assert_eq!(second.content, "/quit");
 
         assert!(
-            stream.next().await.is_none(),
+            timeout(Duration::from_secs(1), stream.next())
+                .await
+                .expect("timed out waiting for stream to close")
+                .is_none(),
             "stream should end after /quit"
         );
     }

src/channels/wasm/wrapper.rs

@@ -3061,6 +3061,20 @@ fn status_to_wit(
         },
         // Suggestions and turn cost are web-gateway-only; skip for WASM channels
         StatusUpdate::Suggestions { .. } | StatusUpdate::TurnCost { .. } => return None,
+        StatusUpdate::ReasoningUpdate {
+            narrative,
+            decisions,
+        } => {
+            let mut msg = narrative.clone();
+            for d in decisions {
+                msg.push_str(&format!("\n  → {}: {}", d.tool_name, d.rationale));
+            }
+            wit_channel::StatusUpdate {
+                status: wit_channel::StatusType::Status,
+                message: msg,
+                metadata_json,
+            }
+        }
     })
 }
 

src/channels/web/handlers/chat.rs

@@ -398,8 +398,10 @@ pub async fn chat_history_handler(
                                 truncate_preview(&s, 500)
                             }),
                             error: tc.error.clone(),
+                            rationale: tc.rationale.clone(),
                         })
                         .collect(),
+                    narrative: t.narrative.clone(),
                 })
                 .collect();
 

src/channels/web/handlers/webhooks.rs

@@ -54,10 +54,37 @@ fn validate_webhook_secret(
 ///
 /// This endpoint is **public** (no gateway auth token required) but protected
 /// by the per-routine webhook secret sent via the `X-Webhook-Secret` header.
+///
+/// **Single-user/backward-compatible**: looks up routines by path across all
+/// users. For multi-tenant isolation, use the user-scoped endpoint at
+/// `/api/webhooks/u/{user_id}/{path}` instead.
 pub async fn webhook_trigger_handler(
     State(state): State<Arc<GatewayState>>,
     Path(path): Path<String>,
     headers: HeaderMap,
+) -> Result<Json<serde_json::Value>, (StatusCode, String)> {
+    fire_webhook_inner(state, &path, None, &headers).await
+}
+
+/// Handle incoming webhook POST to `/api/webhooks/u/{user_id}/{path}`.
+///
+/// User-scoped variant for multi-tenant deployments. The `user_id` in the URL
+/// restricts the routine lookup to that user only, preventing cross-user
+/// webhook triggering even when paths collide.
+pub async fn webhook_trigger_user_scoped_handler(
+    State(state): State<Arc<GatewayState>>,
+    Path((user_id, path)): Path<(String, String)>,
+    headers: HeaderMap,
+) -> Result<Json<serde_json::Value>, (StatusCode, String)> {
+    fire_webhook_inner(state, &path, Some(&user_id), &headers).await
+}
+
+/// Shared webhook logic for both scoped and unscoped endpoints.
+async fn fire_webhook_inner(
+    state: Arc<GatewayState>,
+    path: &str,
+    user_id: Option<&str>,
+    headers: &HeaderMap,
 ) -> Result<Json<serde_json::Value>, (StatusCode, String)> {
     // Rate limit check
     if !state.webhook_rate_limiter.check() {
@@ -72,9 +99,9 @@ pub async fn webhook_trigger_handler(
         "Database not available".to_string(),
     ))?;
 
-    // Targeted query instead of loading all routines
+    // Targeted query — when user_id is provided, restrict to that user's routines
     let routine = store
-        .get_webhook_routine_by_path(&path)
+        .get_webhook_routine_by_path(path, user_id)
         .await
         .map_err(|e| (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()))?
         .ok_or((
@@ -99,7 +126,7 @@ pub async fn webhook_trigger_handler(
         ))?
     };
 
-    let run_id = engine.fire_webhook(routine.id, &path).await.map_err(|e| {
+    let run_id = engine.fire_webhook(routine.id, path).await.map_err(|e| {
         let status = match &e {
             crate::error::RoutineError::NotFound { .. } => StatusCode::NOT_FOUND,
             crate::error::RoutineError::Disabled { .. }

src/channels/web/mod.rs

@@ -501,6 +501,20 @@ impl Channel for GatewayChannel {
             },
             StatusUpdate::Suggestions { suggestions } => AppEvent::Suggestions {
                 suggestions,
+                thread_id: thread_id.clone(),
+            },
+            StatusUpdate::ReasoningUpdate {
+                narrative,
+                decisions,
+            } => AppEvent::ReasoningUpdate {
+                narrative,
+                decisions: decisions
+                    .into_iter()
+                    .map(|d| crate::channels::web::types::ToolDecisionDto {
+                        tool_name: d.tool_name,
+                        rationale: d.rationale,
+                    })
+                    .collect(),
                 thread_id,
             },
             StatusUpdate::TurnCost {

src/channels/web/openai_compat.rs

@@ -231,6 +231,7 @@ pub fn convert_messages(messages: &[OpenAiMessage]) -> Result<Vec<ChatMessage>,
                                 name: tc.function.name.clone(),
                                 arguments: serde_json::from_str(&tc.function.arguments)
                                     .unwrap_or(serde_json::Value::Object(Default::default())),
+                                reasoning: None,
                             })
                             .collect();
                         Ok(ChatMessage::assistant_with_tool_calls(
@@ -954,6 +955,7 @@ mod tests {
             id: "call_abc".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "rust"}),
+            reasoning: None,
         }];
 
         let converted = convert_tool_calls_to_openai(&calls);

src/channels/web/server.rs

@@ -421,6 +421,11 @@ pub async fn start_server(
         .route(
             "/api/webhooks/{path}",
             post(crate::channels::web::handlers::webhooks::webhook_trigger_handler),
+        )
+        // User-scoped webhook endpoint for multi-tenant isolation
+        .route(
+            "/api/webhooks/u/{user_id}/{path}",
+            post(crate::channels::web::handlers::webhooks::webhook_trigger_user_scoped_handler),
         );
 
     // Protected routes (require auth)
@@ -1750,8 +1755,10 @@ async fn chat_history_handler(
                             truncate_preview(&s, 500)
                         }),
                         error: tc.error.clone(),
+                        rationale: tc.rationale.clone(),
                     })
                     .collect(),
+                narrative: t.narrative.clone(),
             })
             .collect();
 

src/channels/web/types.rs

@@ -63,6 +63,9 @@ pub struct TurnInfo {
     pub started_at: String,
     pub completed_at: Option<String>,
     pub tool_calls: Vec<ToolCallInfo>,
+    /// Agent's reasoning narrative for this turn.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub narrative: Option<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -74,6 +77,9 @@ pub struct ToolCallInfo {
     pub result_preview: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub error: Option<String>,
+    /// Agent's reasoning for choosing this tool.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub rationale: Option<String>,
 }
 
 #[derive(Debug, Serialize)]
@@ -116,7 +122,7 @@ pub struct ApprovalRequest {
 
 // --- App Event (re-exported from ironclaw_common) ---
 
-pub use ironclaw_common::AppEvent;
+pub use ironclaw_common::{AppEvent, ToolDecisionDto};
 
 // --- Memory ---
 

src/channels/web/util.rs

@@ -4,6 +4,21 @@ use crate::channels::web::types::{ToolCallInfo, TurnInfo};
 
 pub use ironclaw_common::truncate_preview;
 
+/// Parse tool call summary JSON objects into `ToolCallInfo` structs.
+fn parse_tool_call_infos(calls: &[serde_json::Value]) -> Vec<ToolCallInfo> {
+    calls
+        .iter()
+        .map(|c| ToolCallInfo {
+            name: c["name"].as_str().unwrap_or("unknown").to_string(),
+            has_result: c.get("result_preview").is_some_and(|v| !v.is_null()),
+            has_error: c.get("error").is_some_and(|v| !v.is_null()),
+            result_preview: c["result_preview"].as_str().map(String::from),
+            error: c["error"].as_str().map(String::from),
+            rationale: c["rationale"].as_str().map(String::from),
+        })
+        .collect()
+}
+
 /// Build TurnInfo pairs from flat DB messages (user/tool_calls/assistant triples).
 ///
 /// Handles three message patterns:
@@ -27,6 +42,7 @@ pub fn build_turns_from_db_messages(
                 started_at: msg.created_at.to_rfc3339(),
                 completed_at: None,
                 tool_calls: Vec::new(),
+                narrative: None,
             };
 
             // Check if next message is a tool_calls record
@@ -34,18 +50,28 @@ pub fn build_turns_from_db_messages(
                 && next.role == "tool_calls"
             {
                 let tc_msg = iter.next().expect("peeked");
-                match serde_json::from_str::<Vec<serde_json::Value>>(&tc_msg.content) {
-                    Ok(calls) => {
-                        turn.tool_calls = calls
-                            .iter()
-                            .map(|c| ToolCallInfo {
-                                name: c["name"].as_str().unwrap_or("unknown").to_string(),
-                                has_result: c.get("result_preview").is_some(),
-                                has_error: c.get("error").is_some(),
-                                result_preview: c["result_preview"].as_str().map(String::from),
-                                error: c["error"].as_str().map(String::from),
-                            })
-                            .collect();
+                // Parse tool_calls JSON — supports two formats:
+                // safety: no byte-index slicing; comment describes JSON shape
+                match serde_json::from_str::<serde_json::Value>(&tc_msg.content) {
+                    Ok(serde_json::Value::Array(calls)) => {
+                        // Old format: plain array
+                        turn.tool_calls = parse_tool_call_infos(&calls);
+                    }
+                    Ok(serde_json::Value::Object(obj)) => {
+                        // New wrapped format with narrative
+                        turn.narrative = obj
+                            .get("narrative")
+                            .and_then(|v| v.as_str())
+                            .map(String::from);
+                        if let Some(serde_json::Value::Array(calls)) = obj.get("calls") {
+                            turn.tool_calls = parse_tool_call_infos(calls);
+                        }
+                    }
+                    Ok(_) => {
+                        tracing::warn!(
+                            message_id = %tc_msg.id,
+                            "Unexpected tool_calls JSON shape in DB, skipping"
+                        );
                     }
                     Err(e) => {
                         tracing::warn!(
@@ -83,6 +109,7 @@ pub fn build_turns_from_db_messages(
                 started_at: msg.created_at.to_rfc3339(),
                 completed_at: Some(msg.created_at.to_rfc3339()),
                 tool_calls: Vec::new(),
+                narrative: None,
             });
             turn_number += 1;
         }
@@ -201,4 +228,52 @@ mod tests {
         assert!(turns[0].tool_calls.is_empty());
         assert_eq!(turns[0].state, "Completed");
     }
+
+    #[test]
+    fn test_build_turns_with_wrapped_tool_calls_format() {
+        let tc_json = serde_json::json!({
+            "narrative": "Searching memory for context before proceeding.",
+            "calls": [
+                {"name": "memory_search", "result_preview": "found 3 items", "rationale": "consult prior context"},
+                {"name": "shell", "error": "permission denied"}
+            ]
+        });
+        let messages = vec![
+            make_msg("user", "Find info", 0),
+            make_msg("tool_calls", &tc_json.to_string(), 500),
+            make_msg("assistant", "Here's what I found", 1000),
+        ];
+        let turns = build_turns_from_db_messages(&messages);
+        assert_eq!(turns.len(), 1);
+        assert_eq!(
+            turns[0].narrative.as_deref(),
+            Some("Searching memory for context before proceeding.")
+        );
+        assert_eq!(turns[0].tool_calls.len(), 2);
+        assert_eq!(turns[0].tool_calls[0].name, "memory_search");
+        assert_eq!(
+            turns[0].tool_calls[0].rationale.as_deref(),
+            Some("consult prior context")
+        );
+        assert!(turns[0].tool_calls[0].has_result);
+        assert_eq!(turns[0].tool_calls[1].name, "shell");
+        assert!(turns[0].tool_calls[1].has_error);
+        assert_eq!(turns[0].response.as_deref(), Some("Here's what I found"));
+    }
+
+    #[test]
+    fn test_build_turns_wrapped_format_without_narrative() {
+        let tc_json = serde_json::json!({
+            "calls": [{"name": "echo", "result_preview": "hello"}]
+        });
+        let messages = vec![
+            make_msg("user", "Say hi", 0),
+            make_msg("tool_calls", &tc_json.to_string(), 500),
+            make_msg("assistant", "Done", 1000),
+        ];
+        let turns = build_turns_from_db_messages(&messages);
+        assert_eq!(turns.len(), 1);
+        assert!(turns[0].narrative.is_none());
+        assert_eq!(turns[0].tool_calls.len(), 1);
+    }
 }

src/config/agent.rs

@@ -1,6 +1,6 @@
 use std::time::Duration;
 
-use crate::config::helpers::{parse_bool_env, parse_option_env, parse_optional_env};
+use crate::config::helpers::{optional_env, parse_bool_env, parse_option_env, parse_optional_env};
 use crate::error::ConfigError;
 use crate::settings::Settings;
 
@@ -23,6 +23,8 @@ pub struct AgentConfig {
     pub max_cost_per_day_cents: Option<u64>,
     /// Maximum LLM/tool actions per hour. None = unlimited.
     pub max_actions_per_hour: Option<u64>,
+    /// Maximum daily LLM spend per user in cents. None = unlimited.
+    pub max_cost_per_user_per_day_cents: Option<u64>,
     /// Maximum tool-call iterations per agentic loop invocation. Default 50.
     pub max_tool_iterations: usize,
     /// When true, skip tool approval checks entirely. For benchmarks/CI.
@@ -31,6 +33,13 @@ pub struct AgentConfig {
     pub default_timezone: String,
     /// Maximum tokens per job (0 = unlimited).
     pub max_tokens_per_job: u64,
+    /// Whether the deployment is multi-tenant (multiple users sharing one
+    /// instance). Auto-detected from GATEWAY_USER_TOKENS presence.
+    pub multi_tenant: bool,
+    /// Maximum concurrent LLM calls per user. None = use default (4).
+    pub max_llm_concurrent_per_user: Option<usize>,
+    /// Maximum concurrent jobs per user. None = use default (3).
+    pub max_jobs_concurrent_per_user: Option<usize>,
 }
 
 impl AgentConfig {
@@ -49,10 +58,14 @@ impl AgentConfig {
             allow_local_tools: true,
             max_cost_per_day_cents: None,
             max_actions_per_hour: None,
+            max_cost_per_user_per_day_cents: None,
             max_tool_iterations: 10,
             auto_approve_tools: true,
             default_timezone: "UTC".to_string(),
             max_tokens_per_job: 0,
+            multi_tenant: false,
+            max_llm_concurrent_per_user: None,
+            max_jobs_concurrent_per_user: None,
         }
     }
 
@@ -87,6 +100,7 @@ impl AgentConfig {
             allow_local_tools: parse_bool_env("ALLOW_LOCAL_TOOLS", false)?,
             max_cost_per_day_cents: parse_option_env("MAX_COST_PER_DAY_CENTS")?,
             max_actions_per_hour: parse_option_env("MAX_ACTIONS_PER_HOUR")?,
+            max_cost_per_user_per_day_cents: parse_option_env("MAX_COST_PER_USER_PER_DAY_CENTS")?,
             max_tool_iterations: parse_optional_env(
                 "AGENT_MAX_TOOL_ITERATIONS",
                 settings.agent.max_tool_iterations,
@@ -112,6 +126,11 @@ impl AgentConfig {
                 "AGENT_MAX_TOKENS_PER_JOB",
                 settings.agent.max_tokens_per_job,
             )?,
+            // Auto-detected from GATEWAY_USER_TOKENS presence. Not a separate
+            // knob — multi-tenant mode is always implied by configuring user tokens.
+            multi_tenant: optional_env("GATEWAY_USER_TOKENS")?.is_some(),
+            max_llm_concurrent_per_user: parse_option_env("TENANT_MAX_LLM_CONCURRENT")?,
+            max_jobs_concurrent_per_user: parse_option_env("TENANT_MAX_JOBS_CONCURRENT")?,
         })
     }
 }

src/config/heartbeat.rs

@@ -21,6 +21,9 @@ pub struct HeartbeatConfig {
     pub quiet_hours_end: Option<u32>,
     /// Timezone for fire_at and quiet hours evaluation (IANA name).
     pub timezone: Option<String>,
+    /// When true, cycle through all users with routines. Auto-detected from
+    /// GATEWAY_USER_TOKENS or set explicitly via HEARTBEAT_MULTI_TENANT.
+    pub multi_tenant: bool,
 }
 
 impl Default for HeartbeatConfig {
@@ -34,6 +37,7 @@ impl Default for HeartbeatConfig {
             quiet_hours_start: None,
             quiet_hours_end: None,
             timezone: None,
+            multi_tenant: false,
         }
     }
 }
@@ -101,6 +105,12 @@ impl HeartbeatConfig {
                 }
                 tz
             },
+            // Auto-detect multi-tenant mode from GATEWAY_USER_TOKENS presence,
+            // or allow explicit override via HEARTBEAT_MULTI_TENANT.
+            multi_tenant: parse_bool_env(
+                "HEARTBEAT_MULTI_TENANT",
+                optional_env("GATEWAY_USER_TOKENS")?.is_some(),
+            )?,
         })
     }
 }

src/db/libsql/routines.rs

@@ -530,10 +530,24 @@ impl RoutineStore for LibSqlBackend {
     async fn get_webhook_routine_by_path(
         &self,
         path: &str,
+        user_id: Option<&str>,
     ) -> Result<Option<Routine>, DatabaseError> {
         let conn = self.connect().await?;
-        let mut rows = conn
-            .query(
+        let mut rows = if let Some(uid) = user_id {
+            conn.query(
+                &format!(
+                    "SELECT {} FROM routines WHERE enabled = 1 AND trigger_type = 'webhook' \
+                     AND user_id = ?2 \
+                     AND (json_extract(trigger_config, '$.path') = ?1 \
+                     OR (json_extract(trigger_config, '$.path') IS NULL AND CAST(id AS TEXT) = ?1))",
+                    ROUTINE_COLUMNS
+                ),
+                params![path, uid],
+            )
+            .await
+            .map_err(|e| DatabaseError::Query(e.to_string()))?
+        } else {
+            conn.query(
                 &format!(
                     "SELECT {} FROM routines WHERE enabled = 1 AND trigger_type = 'webhook' \
                      AND (json_extract(trigger_config, '$.path') = ?1 \
@@ -543,7 +557,8 @@ impl RoutineStore for LibSqlBackend {
                 params![path],
             )
             .await
-            .map_err(|e| DatabaseError::Query(e.to_string()))?;
+            .map_err(|e| DatabaseError::Query(e.to_string()))?
+        };
 
         match rows
             .next()

src/db/mod.rs

@@ -545,6 +545,7 @@ pub trait RoutineStore: Send + Sync {
     async fn get_webhook_routine_by_path(
         &self,
         path: &str,
+        user_id: Option<&str>,
     ) -> Result<Option<Routine>, DatabaseError>;
 
     /// List routine runs that were dispatched as full_job but have not yet

src/db/postgres.rs

@@ -529,8 +529,9 @@ impl RoutineStore for PgBackend {
     async fn get_webhook_routine_by_path(
         &self,
         path: &str,
+        user_id: Option<&str>,
     ) -> Result<Option<Routine>, DatabaseError> {
-        self.store.get_webhook_routine_by_path(path).await
+        self.store.get_webhook_routine_by_path(path, user_id).await
     }
 
     async fn list_dispatched_routine_runs(&self) -> Result<Vec<RoutineRun>, DatabaseError> {

src/history/store.rs

@@ -1162,15 +1162,25 @@ impl Store {
     pub async fn get_webhook_routine_by_path(
         &self,
         path: &str,
+        user_id: Option<&str>,
     ) -> Result<Option<Routine>, DatabaseError> {
         let conn = self.conn().await?;
-        let row = conn
-            .query_opt(
+        let row = if let Some(uid) = user_id {
+            conn.query_opt(
+                "SELECT * FROM routines WHERE enabled AND trigger_type = 'webhook' \
+                 AND user_id = $2 \
+                 AND (trigger_config->>'path' = $1 OR (trigger_config->>'path' IS NULL AND id::text = $1))",
+                &[&path, &uid],
+            )
+            .await?
+        } else {
+            conn.query_opt(
                 "SELECT * FROM routines WHERE enabled AND trigger_type = 'webhook' \
                  AND (trigger_config->>'path' = $1 OR (trigger_config->>'path' IS NULL AND id::text = $1))",
                 &[&path],
             )
-            .await?;
+            .await?
+        };
         row.as_ref().map(row_to_routine).transpose()
     }
 

src/lib.rs

@@ -69,6 +69,7 @@ pub mod service;
 pub mod settings;
 pub mod setup;
 pub mod skills;
+pub mod tenant;
 pub mod timezone;
 pub mod tools;
 pub mod tracing_fmt;

src/llm/anthropic_oauth.rs

@@ -575,6 +575,7 @@ fn extract_response_content(response: &AnthropicResponse) -> (Option<String>, Ve
                     id: id.clone(),
                     name: name.clone(),
                     arguments: input.clone(),
+                    reasoning: None,
                 });
             }
         }
@@ -623,6 +624,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "test"}),
+            reasoning: None,
         }];
         let messages = vec![
             ChatMessage::user("Search for test"),

src/llm/bedrock.rs

@@ -522,6 +522,7 @@ fn extract_content_blocks(
                     id: tu.tool_use_id().to_string(),
                     name: tu.name().to_string(),
                     arguments: document_to_json(tu.input()),
+                    reasoning: None,
                 });
             }
             // Ignore reasoning, citations, images, etc.
@@ -759,11 +760,13 @@ mod tests {
             id: "call_1".to_string(),
             name: "echo".to_string(),
             arguments: serde_json::json!({"text": "hi"}),
+            reasoning: None,
         };
         let tc2 = crate::llm::provider::ToolCall {
             id: "call_2".to_string(),
             name: "time".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
 
         let messages = vec![
@@ -802,6 +805,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "test"}),
+            reasoning: None,
         };
 
         let messages = vec![
@@ -825,6 +829,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "echo".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
 
         let messages = vec![
@@ -989,11 +994,13 @@ mod tests {
             id: "call_abc".to_string(),
             name: "get_weather".to_string(),
             arguments: serde_json::json!({"city": "NYC"}),
+            reasoning: None,
         };
         let tc2 = crate::llm::provider::ToolCall {
             id: "call_def".to_string(),
             name: "get_time".to_string(),
             arguments: serde_json::json!({"tz": "EST"}),
+            reasoning: None,
         };
 
         let messages = vec![

src/llm/codex_chatgpt.rs

@@ -732,6 +732,7 @@ impl LlmProvider for CodexChatGptProvider {
                     id: tc.call_id,
                     name: tc.name,
                     arguments: args,
+                    reasoning: None,
                 }
             })
             .collect();
@@ -825,6 +826,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "search".to_string(),
             arguments: json!({"query": "rust"}),
+            reasoning: None,
         };
         let msg = ChatMessage::assistant_with_tool_calls(Some("thinking...".into()), vec![tc]);
         let items = CodexChatGptProvider::message_to_input_items(&msg);

src/llm/gemini_oauth.rs

@@ -1898,6 +1898,7 @@ impl GeminiOauthProvider {
                         id,
                         name,
                         arguments: args,
+                        reasoning: None,
                     });
                 }
             }

src/llm/github_copilot.rs

@@ -596,6 +596,7 @@ fn extract_choice_content(choice: &OpenAiChoice) -> (Option<String>, Vec<ToolCal
                     name: tc.function.name.clone(),
                     arguments: serde_json::from_str(&tc.function.arguments)
                         .unwrap_or(serde_json::Value::Object(serde_json::Map::new())),
+                    reasoning: None,
                 })
                 .collect()
         })
@@ -628,6 +629,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "test"}),
+            reasoning: None,
         }];
         let messages = vec![
             ChatMessage::user("Search"),

src/llm/mod.rs

@@ -63,7 +63,8 @@ pub use provider::{
 };
 pub use reasoning::{
     ActionPlan, Reasoning, ReasoningContext, RespondOutput, RespondResult, SILENT_REPLY_TOKEN,
-    TOOL_INTENT_NUDGE, TokenUsage, ToolSelection, is_silent_reply, llm_signals_tool_intent,
+    TOOL_INTENT_NUDGE, TRUNCATED_TOOL_CALL_NOTICE, TokenUsage, ToolSelection, is_silent_reply,
+    llm_signals_tool_intent,
 };
 pub use recording::RecordingLlm;
 pub use registry::{ProviderDefinition, ProviderProtocol, ProviderRegistry};

src/llm/nearai_chat.rs

@@ -587,6 +587,7 @@ impl LlmProvider for NearAiChatProvider {
                     id: tc.id,
                     name: tc.function.name,
                     arguments,
+                    reasoning: None,
                 }
             })
             .collect();
@@ -1180,11 +1181,13 @@ mod tests {
                 id: "call_1".to_string(),
                 name: "list_issues".to_string(),
                 arguments: serde_json::json!({"owner": "foo", "repo": "bar"}),
+                reasoning: None,
             },
             ToolCall {
                 id: "call_2".to_string(),
                 name: "search".to_string(),
                 arguments: serde_json::json!({"query": "test"}),
+                reasoning: None,
             },
         ];
 
@@ -1217,6 +1220,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "test".to_string(),
             arguments: serde_json::json!({"key": "value"}),
+            reasoning: None,
         };
         let msg = ChatMessage::assistant_with_tool_calls(None, vec![tc]);
         let chat_msg: ChatCompletionMessage = msg.into();
@@ -1460,6 +1464,7 @@ mod tests {
                     id: tc.id,
                     name: tc.function.name,
                     arguments,
+                    reasoning: None,
                 }
             })
             .collect();
@@ -1509,6 +1514,7 @@ mod tests {
                     id: tc.id,
                     name: tc.function.name,
                     arguments,
+                    reasoning: None,
                 }
             })
             .collect();
@@ -2131,6 +2137,7 @@ mod tests {
                 id: "call_1".to_string(),
                 name: "test".to_string(),
                 arguments: serde_json::json!({}),
+                reasoning: None,
             }],
         );
         let chat_msg: ChatCompletionMessage = msg.into();

src/llm/openai_codex_provider.rs

@@ -625,6 +625,7 @@ fn parse_sse_response(body: &str) -> Result<ParsedResponse, LlmError> {
                                 id: state.call_id,
                                 name: state.name,
                                 arguments,
+                                reasoning: None,
                             });
                         } else {
                             // Fallback: extract directly from the item
@@ -650,6 +651,7 @@ fn parse_sse_response(body: &str) -> Result<ParsedResponse, LlmError> {
                                 id: call_id,
                                 name,
                                 arguments,
+                                reasoning: None,
                             });
                         }
                     }
@@ -727,6 +729,7 @@ fn parse_sse_response(body: &str) -> Result<ParsedResponse, LlmError> {
                 id: state.call_id,
                 name: state.name,
                 arguments,
+                reasoning: None,
             });
         }
     }
@@ -822,11 +825,13 @@ mod tests {
                 id: "call_1".to_string(),
                 name: "search".to_string(),
                 arguments: serde_json::json!({"query": "test"}),
+                reasoning: None,
             },
             ToolCall {
                 id: "call_2".to_string(),
                 name: "read".to_string(),
                 arguments: serde_json::json!({"path": "/tmp"}),
+                reasoning: None,
             },
         ];
         let msg =

src/llm/provider.rs

@@ -231,6 +231,10 @@ pub struct ToolCall {
     pub id: String,
     pub name: String,
     pub arguments: serde_json::Value,
+    /// Optional reasoning for why this tool was chosen — supplied by the provider
+    /// or derived from the shared response content as a fallback.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub reasoning: Option<String>,
 }
 
 /// Generate a tool-call ID that satisfies all providers.
@@ -637,6 +641,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "echo".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
         let mut messages = vec![
             ChatMessage::user("hello"),
@@ -680,6 +685,7 @@ mod tests {
             id: "call_1".to_string(),
             name: "echo".to_string(),
             arguments: serde_json::json!({}),
+            reasoning: None,
         };
         let mut messages = vec![
             ChatMessage::user("test"),
@@ -705,11 +711,13 @@ mod tests {
             id: "call_sel_1".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "test"}),
+            reasoning: None,
         };
         let tc2 = ToolCall {
             id: "call_sel_2".to_string(),
             name: "http".to_string(),
             arguments: serde_json::json!({"url": "https://example.com"}),
+            reasoning: None,
         };
         let mut messages = vec![
             ChatMessage::system("You are a helpful assistant."),

src/llm/reasoning.rs

@@ -8,8 +8,8 @@ use serde::{Deserialize, Serialize};
 use crate::llm::error::LlmError;
 
 use crate::llm::{
-    ChatMessage, CompletionRequest, LlmProvider, Role, ToolCall, ToolCompletionRequest,
-    ToolDefinition,
+    ChatMessage, CompletionRequest, FinishReason, LlmProvider, Role, ToolCall,
+    ToolCompletionRequest, ToolDefinition,
 };
 
 /// Token the agent returns when it has nothing to say (e.g. in group chats).
@@ -23,6 +23,13 @@ You said you would perform an action, but you did not include any tool calls.\n\
 Do NOT describe what you intend to do — actually call the tool now.\n\
 Use the tool_calls mechanism to invoke the appropriate tool.";
 
+/// Notice injected when the LLM's response was truncated mid-tool-call,
+/// causing incomplete parameters. Tells the LLM to try a different approach.
+pub const TRUNCATED_TOOL_CALL_NOTICE: &str = "\
+Your previous response was truncated while generating tool call parameters. \
+The tool calls were discarded. Please try a different approach — \
+summarize or transform the data instead of echoing it verbatim in a tool call.";
+
 /// Seed value used as the second argument to `generate_tool_call_id` when
 /// recovering tool calls from malformed LLM text responses. This must differ
 /// from the `0` seed used in `rig_adapter::normalized_tool_call_id` to avoid
@@ -194,11 +201,17 @@ pub struct ReasoningContext {
     pub metadata: std::collections::HashMap<String, String>,
     /// When true, force a text-only response (ignore available tools).
     /// Used by the agentic loop to guarantee termination near the iteration limit.
+    /// Sticky: once set, never cleared within a loop invocation. Callers must
+    /// create a fresh `ReasoningContext` per `run_agentic_loop()` call.
     pub force_text: bool,
     /// Pre-built system prompt. When set, `respond_with_tools` uses this directly
     /// instead of calling `build_system_prompt_with_tools`. Allows callers to build
     /// the prompt once and reuse it across iterations.
     pub system_prompt: Option<String>,
+    /// Per-user model override. When set, completion requests use this model
+    /// instead of the provider's default. Only effective with providers that
+    /// support per-request model overrides (e.g. NearAI).
+    pub model_override: Option<String>,
 }
 
 impl ReasoningContext {
@@ -212,6 +225,7 @@ impl ReasoningContext {
             metadata: std::collections::HashMap::new(),
             force_text: false,
             system_prompt: None,
+            model_override: None,
         }
     }
 
@@ -344,6 +358,7 @@ pub enum RespondResult {
 pub struct RespondOutput {
     pub result: RespondResult,
     pub usage: TokenUsage,
+    pub finish_reason: FinishReason,
 }
 
 /// Reasoning engine for the agent.
@@ -525,17 +540,46 @@ impl Reasoning {
 
         let response = self.llm.complete_with_tools(request).await?;
 
-        let reasoning = response.content.unwrap_or_default();
+        // If the response was truncated, tool call parameters are likely incomplete.
+        // Return empty so the caller can fall through to respond_with_tools() which
+        // has a larger output token budget.
+        if response.finish_reason == FinishReason::Length {
+            tracing::warn!(
+                "select_tools response truncated (finish_reason=Length), \
+                 discarding potentially incomplete tool selections"
+            );
+            return Ok(vec![]);
+        }
+
+        let shared_reasoning = response
+            .content
+            .map(|c| {
+                let pre_truncated = truncate_at_tool_tags(&c);
+                clean_response(&pre_truncated)
+            })
+            .unwrap_or_default();
 
         let selections: Vec<ToolSelection> = response
             .tool_calls
             .into_iter()
-            .map(|tool_call| ToolSelection {
-                tool_name: tool_call.name,
-                parameters: tool_call.arguments,
-                reasoning: reasoning.clone(),
-                alternatives: vec![],
-                tool_call_id: tool_call.id,
+            .map(|tool_call| {
+                // Prefer per-tool reasoning if the provider supplied it,
+                // otherwise fall back to the shared response content.
+                let rationale = tool_call
+                    .reasoning
+                    .map(|r| {
+                        let pre_truncated = truncate_at_tool_tags(&r);
+                        clean_response(&pre_truncated)
+                    })
+                    .filter(|r| !r.trim().is_empty())
+                    .unwrap_or_else(|| shared_reasoning.clone());
+                ToolSelection {
+                    tool_name: tool_call.name,
+                    parameters: tool_call.arguments,
+                    reasoning: rationale,
+                    alternatives: vec![],
+                    tool_call_id: tool_call.id,
+                }
             })
             .collect();
 
@@ -653,6 +697,9 @@ Respond in JSON format:
                 .with_temperature(0.7)
                 .with_tool_choice("auto");
             request.metadata = context.metadata.clone();
+            if let Some(ref model) = context.model_override {
+                request.model = Some(model.clone());
+            }
 
             let response = self.llm.complete_with_tools(request).await?;
             let usage = TokenUsage {
@@ -664,15 +711,39 @@ Respond in JSON format:
 
             // If there were tool calls, return them for execution
             if !response.tool_calls.is_empty() {
+                let narrative = response.content.map(|c| {
+                    let pre_truncated = truncate_at_tool_tags(&c);
+                    clean_response(&pre_truncated)
+                });
+                // Populate per-tool reasoning from the shared narrative when the
+                // provider did not supply per-tool rationale.
+                let tool_calls: Vec<ToolCall> = response
+                    .tool_calls
+                    .into_iter()
+                    .map(|mut tc| {
+                        if tc.reasoning.as_ref().is_none_or(|r| r.trim().is_empty()) {
+                            tc.reasoning = narrative.as_ref().filter(|n| !n.is_empty()).cloned();
+                        } else {
+                            // Clean provider-supplied per-tool reasoning the same way
+                            // we clean the shared narrative (strip thinking/tool tags).
+                            tc.reasoning = tc
+                                .reasoning
+                                .map(|r| {
+                                    let pre_truncated = truncate_at_tool_tags(&r);
+                                    clean_response(&pre_truncated)
+                                })
+                                .filter(|r| !r.trim().is_empty());
+                        }
+                        tc
+                    })
+                    .collect();
                 return Ok(RespondOutput {
                     result: RespondResult::ToolCalls {
-                        tool_calls: response.tool_calls,
-                        content: response.content.map(|c| {
-                            let pre_truncated = truncate_at_tool_tags(&c);
-                            clean_response(&pre_truncated)
-                        }),
+                        tool_calls,
+                        content: narrative,
                     },
                     usage,
+                    finish_reason: response.finish_reason,
                 });
             }
 
@@ -700,6 +771,7 @@ Respond in JSON format:
                         },
                     },
                     usage,
+                    finish_reason: response.finish_reason,
                 });
             }
 
@@ -725,6 +797,7 @@ Respond in JSON format:
             Ok(RespondOutput {
                 result: RespondResult::Text(final_text),
                 usage,
+                finish_reason: response.finish_reason,
             })
         } else {
             // No tools, use simple completion
@@ -732,6 +805,9 @@ Respond in JSON format:
                 .with_max_tokens(4096)
                 .with_temperature(0.7);
             request.metadata = context.metadata.clone();
+            if let Some(ref model) = context.model_override {
+                request.model = Some(model.clone());
+            }
 
             let response = self.llm.complete(request).await?;
             let pre_truncated = truncate_at_tool_tags(&response.content);
@@ -753,6 +829,7 @@ Respond in JSON format:
                     cache_read_input_tokens: response.cache_read_input_tokens,
                     cache_creation_input_tokens: response.cache_creation_input_tokens,
                 },
+                finish_reason: response.finish_reason,
             })
         }
     }
@@ -1293,6 +1370,49 @@ fn is_inside_code(pos: usize, regions: &[CodeRegion]) -> bool {
     regions.iter().any(|r| pos >= r.start && pos < r.end)
 }
 
+/// Check whether a byte range overlaps any code region.
+fn overlaps_code_region(start: usize, end: usize, regions: &[CodeRegion]) -> bool {
+    regions.iter().any(|r| start < r.end && end > r.start)
+}
+
+/// Return the byte bounds of the line containing `pos`, excluding the trailing newline.
+fn line_bounds(text: &str, pos: usize) -> (usize, usize) {
+    let start = text[..pos].rfind('\n').map_or(0, |idx| idx + 1);
+    let end = text[pos..].find('\n').map_or(text.len(), |idx| pos + idx);
+    (start, end)
+}
+
+/// Only recover XML-style tool calls when they are isolated content outside
+/// markdown code and quote contexts. This avoids converting code examples or
+/// quoted snippets into executable tool calls.
+fn is_recoverable_tool_call_segment(
+    text: &str,
+    start: usize,
+    end: usize,
+    code_regions: &[CodeRegion],
+) -> bool {
+    if overlaps_code_region(start, end, code_regions) {
+        return false;
+    }
+
+    let (first_line_start, first_line_end) = line_bounds(text, start);
+    let first_line = &text[first_line_start..first_line_end];
+
+    if first_line.trim_start().starts_with('>') {
+        return false;
+    }
+
+    let (_, last_line_end) = line_bounds(text, end.saturating_sub(1));
+    let first_line_prefix = &text[first_line_start..start];
+    let last_line_suffix = &text[end..last_line_end];
+
+    if !first_line_prefix.trim().is_empty() || !last_line_suffix.trim().is_empty() {
+        return false;
+    }
+
+    true
+}
+
 /// Clean up LLM response by stripping model-internal tags and reasoning patterns.
 ///
 /// Some models (GLM-4.7, etc.) emit XML-tagged internal state like
@@ -1312,6 +1432,7 @@ fn recover_tool_calls_from_content(
 ) -> Vec<ToolCall> {
     let tool_names: std::collections::HashSet<&str> =
         available_tools.iter().map(|t| t.name.as_str()).collect();
+    let code_regions = find_code_regions(content);
     let mut calls = Vec::new();
 
     for (open, close) in &[
@@ -1320,15 +1441,23 @@ fn recover_tool_calls_from_content(
         ("<function_call>", "</function_call>"),
         ("<|function_call|>", "<|/function_call|>"),
     ] {
-        let mut remaining = content;
-        while let Some(start) = remaining.find(open) {
+        let mut search_from = 0;
+        while let Some(offset) = content[search_from..].find(open) {
+            let start = search_from + offset;
             let inner_start = start + open.len();
-            let after = &remaining[inner_start..];
-            let Some(end) = after.find(close) else {
+            let after = &content[inner_start..];
+            let Some(end_offset) = after.find(close) else {
                 break;
             };
-            let inner = after[..end].trim();
-            remaining = &after[end + close.len()..];
+            let end = inner_start + end_offset;
+            let segment_end = end + close.len();
+            search_from = segment_end;
+
+            if !is_recoverable_tool_call_segment(content, start, segment_end, &code_regions) {
+                continue;
+            }
+
+            let inner = content[inner_start..end].trim();
 
             if inner.is_empty() {
                 continue;
@@ -1350,6 +1479,7 @@ fn recover_tool_calls_from_content(
                     ),
                     name: name.to_string(),
                     arguments,
+                    reasoning: None,
                 });
                 continue;
             }
@@ -1364,6 +1494,7 @@ fn recover_tool_calls_from_content(
                     ),
                     name: name.to_string(),
                     arguments: serde_json::Value::Object(Default::default()),
+                    reasoning: None,
                 });
             }
         }
@@ -1401,6 +1532,7 @@ fn recover_tool_calls_from_content(
                         ),
                         name: name.to_string(),
                         arguments,
+                        reasoning: None,
                     });
                     remaining = &args_start[bracket_end + 1..];
                     continue;
@@ -1412,6 +1544,7 @@ fn recover_tool_calls_from_content(
                 id: super::provider::generate_tool_call_id(calls.len(), RECOVERED_TOOL_CALL_SEED),
                 name: name.to_string(),
                 arguments: serde_json::Value::Object(Default::default()),
+                reasoning: None,
             });
             remaining = after_name;
         }
@@ -2257,6 +2390,40 @@ That's my plan."#;
         assert_eq!(calls[0].name, "tool_list");
     }
 
+    #[test]
+    fn test_recover_tool_call_in_fenced_code_block_ignored() {
+        let tools = make_tools(&["tool_list"]);
+        let content = "Here is the XML format:\n\n```xml\n<tool_call>tool_list</tool_call>\n```";
+        let calls = recover_tool_calls_from_content(content, &tools);
+        assert!(calls.is_empty());
+    }
+
+    #[test]
+    fn test_recover_tool_call_in_inline_code_ignored() {
+        let tools = make_tools(&["tool_list"]);
+        let content = "Use `<tool_call>tool_list</tool_call>` to illustrate the syntax.";
+        let calls = recover_tool_calls_from_content(content, &tools);
+        assert!(calls.is_empty());
+    }
+
+    #[test]
+    fn test_recover_tool_call_in_blockquote_ignored() {
+        let tools = make_tools(&["tool_list"]);
+        let content = "The page replied:\n> <tool_call>tool_list</tool_call>";
+        let calls = recover_tool_calls_from_content(content, &tools);
+        assert!(calls.is_empty());
+    }
+
+    #[test]
+    fn test_recover_multiline_json_tool_call_on_own_line() {
+        let tools = make_tools(&["memory_search"]);
+        let content = "Let me check.\n\n<tool_call>\n{\"name\": \"memory_search\", \"arguments\": {\"query\": \"test\"}}\n</tool_call>\n\nDone.";
+        let calls = recover_tool_calls_from_content(content, &tools);
+        assert_eq!(calls.len(), 1);
+        assert_eq!(calls[0].name, "memory_search");
+        assert_eq!(calls[0].arguments, serde_json::json!({"query": "test"}));
+    }
+
     // ---- System prompt building tests (issue #565) ----
 
     fn make_test_reasoning() -> Reasoning {
@@ -3145,4 +3312,113 @@ That's my plan."#;
             "Text <function_call>{}</function_call> middle "
         );
     }
+
+    /// Verify that reasoning normalization strips thinking tags and tool tags
+    /// from per-tool reasoning, matching the cleaning applied to shared reasoning.
+    #[test]
+    fn test_reasoning_normalization_strips_thinking_tags() {
+        let raw = "<thinking>Let me consider...</thinking>Search memory for prior context";
+        let pre_truncated = truncate_at_tool_tags(raw);
+        let cleaned = clean_response(&pre_truncated);
+        assert!(!cleaned.contains("<thinking>"));
+        assert!(cleaned.contains("Search memory"));
+    }
+
+    #[test]
+    fn test_reasoning_normalization_strips_tool_tags() {
+        let raw = "Calling search <tool_call>{\"name\": \"search\"}";
+        let pre_truncated = truncate_at_tool_tags(raw);
+        let cleaned = clean_response(&pre_truncated);
+        assert!(!cleaned.contains("<tool_call>"));
+        assert!(cleaned.contains("Calling search"));
+    }
+
+    #[test]
+    fn test_reasoning_normalization_empty_after_cleaning() {
+        let raw = "<thinking>internal only</thinking>";
+        let pre_truncated = truncate_at_tool_tags(raw);
+        let cleaned = clean_response(&pre_truncated);
+        assert!(cleaned.trim().is_empty());
+    }
+
+    // ---- select_tools truncation guard ----
+
+    /// Mock provider that returns tool calls with a configurable finish_reason.
+    struct TruncatingLlm {
+        finish_reason: crate::llm::FinishReason,
+    }
+
+    #[async_trait::async_trait]
+    impl crate::llm::LlmProvider for TruncatingLlm {
+        fn model_name(&self) -> &str {
+            "truncating-stub"
+        }
+        fn cost_per_token(&self) -> (rust_decimal::Decimal, rust_decimal::Decimal) {
+            (rust_decimal::Decimal::ZERO, rust_decimal::Decimal::ZERO)
+        }
+        async fn complete(
+            &self,
+            _request: crate::llm::CompletionRequest,
+        ) -> Result<crate::llm::CompletionResponse, crate::llm::error::LlmError> {
+            unimplemented!()
+        }
+        async fn complete_with_tools(
+            &self,
+            _request: crate::llm::ToolCompletionRequest,
+        ) -> Result<crate::llm::ToolCompletionResponse, crate::llm::error::LlmError> {
+            Ok(crate::llm::ToolCompletionResponse {
+                content: Some("I'll write the report.".to_string()),
+                tool_calls: vec![ToolCall {
+                    id: "call_1".to_string(),
+                    name: "memory_write".to_string(),
+                    arguments: serde_json::json!({}),
+                    reasoning: None,
+                }],
+                input_tokens: 5000,
+                output_tokens: 1024,
+                finish_reason: self.finish_reason,
+                cache_read_input_tokens: 0,
+                cache_creation_input_tokens: 0,
+            })
+        }
+    }
+
+    #[tokio::test]
+    async fn test_select_tools_returns_empty_on_truncation() {
+        let llm = Arc::new(TruncatingLlm {
+            finish_reason: FinishReason::Length,
+        });
+        let reasoning = Reasoning::new(llm);
+        let mut ctx = ReasoningContext::new().with_message(ChatMessage::user("Write a report"));
+        ctx.available_tools.push(ToolDefinition {
+            name: "memory_write".to_string(),
+            description: "Write to memory".to_string(),
+            parameters: serde_json::json!({"type": "object"}),
+        });
+
+        let selections = reasoning.select_tools(&ctx).await.unwrap();
+        assert!(
+            selections.is_empty(),
+            "Truncated tool selections should be discarded (got {} selections)",
+            selections.len()
+        );
+    }
+
+    #[tokio::test]
+    async fn test_select_tools_returns_selections_when_not_truncated() {
+        let llm = Arc::new(TruncatingLlm {
+            finish_reason: FinishReason::ToolUse,
+        });
+        let reasoning = Reasoning::new(llm);
+        let mut ctx = ReasoningContext::new().with_message(ChatMessage::user("Write a report"));
+        ctx.available_tools.push(ToolDefinition {
+            name: "memory_write".to_string(),
+            description: "Write to memory".to_string(),
+            parameters: serde_json::json!({"type": "object"}),
+        });
+
+        let selections = reasoning.select_tools(&ctx).await.unwrap();
+        assert_eq!(selections.len(), 1);
+        assert_eq!(selections[0].tool_name, "memory_write");
+    }
 }

src/llm/rig_adapter.rs

@@ -500,6 +500,7 @@ fn extract_response(
                     id: tc.id.clone(),
                     name: tc.function.name.clone(),
                     arguments: tc.function.arguments.clone(),
+                    reasoning: None,
                 });
             }
             // Reasoning and Image variants are not mapped to IronClaw types
@@ -607,6 +608,30 @@ fn build_rig_request(
     })
 }
 
+/// Inject a per-request model override into the rig request's `additional_params`.
+///
+/// Rig-core bakes the model name at construction time inside each provider's
+/// `CompletionModel` implementation. The actual HTTP request body includes a
+/// `model` field set by the provider. Rig-core's `#[serde(flatten)]` on
+/// `additional_params` emits these fields AFTER the provider's own fields.
+/// Most API servers (Python, Go) use last-key-wins when deserializing
+/// duplicate JSON keys, so the injected `model` value takes effect.
+fn inject_model_override(rig_req: &mut RigRequest, model_override: Option<&str>) {
+    let Some(model) = model_override else {
+        return;
+    };
+    match rig_req.additional_params {
+        Some(ref mut params) => {
+            if let Some(obj) = params.as_object_mut() {
+                obj.insert("model".to_string(), serde_json::json!(model));
+            }
+        }
+        None => {
+            rig_req.additional_params = Some(serde_json::json!({ "model": model }));
+        }
+    }
+}
+
 #[async_trait]
 impl<M> LlmProvider for RigAdapter<M>
 where
@@ -641,15 +666,7 @@ where
         &self,
         mut request: CompletionRequest,
     ) -> Result<CompletionResponse, LlmError> {
-        if let Some(requested_model) = request.model.as_deref()
-            && requested_model != self.model_name.as_str()
-        {
-            tracing::warn!(
-                requested_model = requested_model,
-                active_model = %self.model_name,
-                "Per-request model override is not supported for this provider; using configured model"
-            );
-        }
+        let model_override = request.model.take();
 
         self.strip_unsupported_completion_params(&mut request);
 
@@ -657,7 +674,7 @@ where
         crate::llm::provider::sanitize_tool_messages(&mut messages);
         let (preamble, history) = convert_messages(&messages);
 
-        let rig_req = build_rig_request(
+        let mut rig_req = build_rig_request(
             preamble,
             history,
             Vec::new(),
@@ -667,6 +684,8 @@ where
             self.cache_retention,
         )?;
 
+        inject_model_override(&mut rig_req, model_override.as_deref());
+
         let response =
             self.model
                 .completion(rig_req)
@@ -704,15 +723,7 @@ where
         &self,
         mut request: ToolCompletionRequest,
     ) -> Result<ToolCompletionResponse, LlmError> {
-        if let Some(requested_model) = request.model.as_deref()
-            && requested_model != self.model_name.as_str()
-        {
-            tracing::warn!(
-                requested_model = requested_model,
-                active_model = %self.model_name,
-                "Per-request model override is not supported for this provider; using configured model"
-            );
-        }
+        let model_override = request.model.take();
 
         self.strip_unsupported_tool_params(&mut request);
 
@@ -725,7 +736,7 @@ where
         let tools = convert_tools(&request.tools);
         let tool_choice = convert_tool_choice(request.tool_choice.as_deref());
 
-        let rig_req = build_rig_request(
+        let mut rig_req = build_rig_request(
             preamble,
             history,
             tools,
@@ -735,6 +746,8 @@ where
             self.cache_retention,
         )?;
 
+        inject_model_override(&mut rig_req, model_override.as_deref());
+
         let response =
             self.model
                 .completion(rig_req)
@@ -890,6 +903,7 @@ mod tests {
             id: "Xt7mK9pQ2".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "test"}),
+            reasoning: None,
         };
         let msg = ChatMessage::assistant_with_tool_calls(Some("thinking".to_string()), vec![tc]);
         let messages = vec![msg];
@@ -1007,6 +1021,7 @@ mod tests {
             id: "".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "test"}),
+            reasoning: None,
         };
         let messages = vec![ChatMessage::assistant_with_tool_calls(None, vec![tc])];
         let (_preamble, history) = convert_messages(&messages);
@@ -1038,6 +1053,7 @@ mod tests {
             id: "   ".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "test"}),
+            reasoning: None,
         };
         let messages = vec![ChatMessage::assistant_with_tool_calls(None, vec![tc])];
         let (_preamble, history) = convert_messages(&messages);
@@ -1071,6 +1087,7 @@ mod tests {
             id: "".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"query": "test"}),
+            reasoning: None,
         };
         let assistant_msg = ChatMessage::assistant_with_tool_calls(None, vec![tc]);
         let tool_result_msg = ChatMessage {
@@ -1390,11 +1407,13 @@ mod tests {
             id: "call_a".to_string(),
             name: "search".to_string(),
             arguments: serde_json::json!({"q": "rust"}),
+            reasoning: None,
         };
         let tc2 = IronToolCall {
             id: "call_b".to_string(),
             name: "fetch".to_string(),
             arguments: serde_json::json!({"url": "https://example.com"}),
+            reasoning: None,
         };
         let assistant = ChatMessage::assistant_with_tool_calls(None, vec![tc1, tc2]);
         let result_a = ChatMessage::tool_result("call_a", "search", "search results");

src/main.rs

@@ -917,6 +917,10 @@ async fn async_main() -> anyhow::Result<()> {
         },
         builder: components.builder,
         llm_backend: config.llm.backend.clone(),
+        tenant_rates: Arc::new(ironclaw::tenant::TenantRateRegistry::new(
+            config.agent.max_llm_concurrent_per_user.unwrap_or(4),
+            config.agent.max_jobs_concurrent_per_user.unwrap_or(3),
+        )),
     };
 
     let channels_for_warnings = Arc::clone(&channels);

src/orchestrator/api.rs

@@ -14,6 +14,7 @@ use serde::{Deserialize, Serialize};
 use tokio::sync::{Mutex, broadcast};
 use uuid::Uuid;
 
+use crate::channels::web::types::ToolDecisionDto;
 use crate::db::Database;
 use crate::llm::{CompletionRequest, LlmProvider, ToolCompletionRequest};
 use crate::orchestrator::auth::{TokenStore, worker_auth_middleware};
@@ -344,6 +345,20 @@ async fn job_event_handler(
             // gain context/memory tracking capabilities.
             fallback_deliverable: payload.data.get("fallback_deliverable").cloned(),
         },
+        "reasoning" => {
+            let narrative = payload
+                .data
+                .get("narrative")
+                .and_then(|v| v.as_str())
+                .unwrap_or("")
+                .to_string();
+            let decisions = ToolDecisionDto::from_json_array(&payload.data["decisions"]);
+            AppEvent::JobReasoning {
+                job_id: job_id_str,
+                narrative,
+                decisions,
+            }
+        }
         _ => AppEvent::JobStatus {
             job_id: job_id_str,
             message: payload

src/tenant.rs

@@ -0,0 +1,906 @@
+//! Compile-time tenant isolation.
+//!
+//! Provides two database access tiers:
+//!
+//! - **[`TenantScope`]** (default): All operations are bound to a single user.
+//!   ID-based lookups return `None` if the resource doesn't belong to this user.
+//!   This is the only way handler code should access the database.
+//!
+//! - **[`AdminScope`]**: Cross-tenant access for system-level operations
+//!   (heartbeat, routine engine, self-repair). Must be obtained explicitly via
+//!   [`AgentDeps::admin_store()`](crate::agent::AgentDeps::admin_store).
+//!
+//! [`TenantCtx`] bundles a `TenantScope` with workspace, cost guard, and
+//! per-tenant rate limiting. Constructed once per request at the entry point
+//! where a `user_id` becomes known.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use chrono::{DateTime, Utc};
+use rust_decimal::Decimal;
+use tokio::sync::{Semaphore, SemaphorePermit};
+use uuid::Uuid;
+
+use crate::agent::BrokenTool;
+use crate::agent::cost_guard::{CostGuard, CostLimitExceeded};
+use crate::agent::routine::{Routine, RoutineRun, RunStatus};
+use crate::context::{ActionRecord, JobContext, JobState};
+use crate::db::Database;
+use crate::error::DatabaseError;
+use crate::history::{
+    AgentJobRecord, AgentJobSummary, ConversationMessage, ConversationSummary, LlmCallRecord,
+    SandboxJobRecord, SandboxJobSummary, SettingRow,
+};
+use crate::workspace::Workspace;
+
+// ---------------------------------------------------------------------------
+// TenantScope — scoped database access (default tier)
+// ---------------------------------------------------------------------------
+
+/// Scoped database view. All operations are bound to a single user.
+///
+/// This is the **only** way handler code should access the database.
+/// ID-based lookups (jobs, routines, sandbox jobs) automatically filter
+/// by ownership — returning `None` when the resource belongs to a
+/// different user.
+#[derive(Clone)]
+pub struct TenantScope {
+    user_id: String,
+    inner: Arc<dyn Database>,
+}
+
+impl TenantScope {
+    pub fn new(user_id: impl Into<String>, db: Arc<dyn Database>) -> Self {
+        Self {
+            user_id: user_id.into(),
+            inner: db,
+        }
+    }
+
+    pub fn user_id(&self) -> &str {
+        &self.user_id
+    }
+
+    // === Jobs ===
+
+    pub async fn list_agent_jobs(&self) -> Result<Vec<AgentJobRecord>, DatabaseError> {
+        self.inner.list_agent_jobs_for_user(&self.user_id).await
+    }
+
+    pub async fn agent_job_summary(&self) -> Result<AgentJobSummary, DatabaseError> {
+        self.inner.agent_job_summary_for_user(&self.user_id).await
+    }
+
+    /// Fetch a job by ID, returning `None` if it doesn't belong to this user.
+    pub async fn get_job(&self, id: Uuid) -> Result<Option<JobContext>, DatabaseError> {
+        match self.inner.get_job(id).await? {
+            Some(ctx) if ctx.user_id == self.user_id => Ok(Some(ctx)),
+            _ => Ok(None),
+        }
+    }
+
+    pub async fn get_agent_job_failure_reason(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<String>, DatabaseError> {
+        // Verify ownership first
+        if self.get_job(id).await?.is_none() {
+            return Ok(None);
+        }
+        self.inner.get_agent_job_failure_reason(id).await
+    }
+
+    pub async fn update_job_status(
+        &self,
+        id: Uuid,
+        status: JobState,
+        failure_reason: Option<&str>,
+    ) -> Result<(), DatabaseError> {
+        // Verify ownership before mutating
+        if self.get_job(id).await?.is_none() {
+            return Err(DatabaseError::NotFound {
+                entity: "job".to_string(),
+                id: id.to_string(),
+            });
+        }
+        self.inner
+            .update_job_status(id, status, failure_reason)
+            .await
+    }
+
+    // === Sandbox jobs ===
+
+    pub async fn list_sandbox_jobs(&self) -> Result<Vec<SandboxJobRecord>, DatabaseError> {
+        self.inner.list_sandbox_jobs_for_user(&self.user_id).await
+    }
+
+    pub async fn sandbox_job_summary(&self) -> Result<SandboxJobSummary, DatabaseError> {
+        self.inner.sandbox_job_summary_for_user(&self.user_id).await
+    }
+
+    /// Fetch a sandbox job by ID, returning `None` if it doesn't belong to this user.
+    pub async fn get_sandbox_job(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<SandboxJobRecord>, DatabaseError> {
+        match self.inner.get_sandbox_job(id).await? {
+            Some(job) if job.user_id == self.user_id => Ok(Some(job)),
+            _ => Ok(None),
+        }
+    }
+
+    pub async fn sandbox_job_belongs_to_user(&self, job_id: Uuid) -> Result<bool, DatabaseError> {
+        self.inner
+            .sandbox_job_belongs_to_user(job_id, &self.user_id)
+            .await
+    }
+
+    // === Routines ===
+
+    pub async fn list_routines(&self) -> Result<Vec<Routine>, DatabaseError> {
+        self.inner.list_routines(&self.user_id).await
+    }
+
+    pub async fn get_routine_by_name(&self, name: &str) -> Result<Option<Routine>, DatabaseError> {
+        self.inner.get_routine_by_name(&self.user_id, name).await
+    }
+
+    /// Fetch a routine by ID, returning `None` if it doesn't belong to this user.
+    pub async fn get_routine(&self, id: Uuid) -> Result<Option<Routine>, DatabaseError> {
+        match self.inner.get_routine(id).await? {
+            Some(r) if r.user_id == self.user_id => Ok(Some(r)),
+            _ => Ok(None),
+        }
+    }
+
+    pub async fn create_routine(&self, routine: &Routine) -> Result<(), DatabaseError> {
+        debug_assert_eq!(
+            routine.user_id, self.user_id,
+            "routine.user_id must match TenantScope user"
+        );
+        self.inner.create_routine(routine).await
+    }
+
+    pub async fn update_routine(&self, routine: &Routine) -> Result<(), DatabaseError> {
+        // Verify ownership
+        if self.get_routine(routine.id).await?.is_none() {
+            return Err(DatabaseError::NotFound {
+                entity: "routine".to_string(),
+                id: routine.id.to_string(),
+            });
+        }
+        self.inner.update_routine(routine).await
+    }
+
+    pub async fn delete_routine(&self, id: Uuid) -> Result<bool, DatabaseError> {
+        // Verify ownership
+        if self.get_routine(id).await?.is_none() {
+            return Err(DatabaseError::NotFound {
+                entity: "routine".to_string(),
+                id: id.to_string(),
+            });
+        }
+        self.inner.delete_routine(id).await
+    }
+
+    /// List routine runs, verifying the routine belongs to this user.
+    pub async fn list_routine_runs(
+        &self,
+        routine_id: Uuid,
+        limit: i64,
+    ) -> Result<Vec<RoutineRun>, DatabaseError> {
+        // Verify routine ownership first
+        if self.get_routine(routine_id).await?.is_none() {
+            return Err(DatabaseError::NotFound {
+                entity: "routine".to_string(),
+                id: routine_id.to_string(),
+            });
+        }
+        self.inner.list_routine_runs(routine_id, limit).await
+    }
+
+    pub async fn get_webhook_routine_by_path(
+        &self,
+        path: &str,
+    ) -> Result<Option<Routine>, DatabaseError> {
+        self.inner
+            .get_webhook_routine_by_path(path, Some(&self.user_id))
+            .await
+    }
+
+    // === Settings ===
+
+    pub async fn get_setting(&self, key: &str) -> Result<Option<serde_json::Value>, DatabaseError> {
+        self.inner.get_setting(&self.user_id, key).await
+    }
+
+    pub async fn get_setting_full(&self, key: &str) -> Result<Option<SettingRow>, DatabaseError> {
+        self.inner.get_setting_full(&self.user_id, key).await
+    }
+
+    pub async fn set_setting(
+        &self,
+        key: &str,
+        value: &serde_json::Value,
+    ) -> Result<(), DatabaseError> {
+        self.inner.set_setting(&self.user_id, key, value).await
+    }
+
+    pub async fn delete_setting(&self, key: &str) -> Result<bool, DatabaseError> {
+        self.inner.delete_setting(&self.user_id, key).await
+    }
+
+    pub async fn list_settings(&self) -> Result<Vec<SettingRow>, DatabaseError> {
+        self.inner.list_settings(&self.user_id).await
+    }
+
+    pub async fn get_all_settings(
+        &self,
+    ) -> Result<HashMap<String, serde_json::Value>, DatabaseError> {
+        self.inner.get_all_settings(&self.user_id).await
+    }
+
+    pub async fn set_all_settings(
+        &self,
+        settings: &HashMap<String, serde_json::Value>,
+    ) -> Result<(), DatabaseError> {
+        self.inner.set_all_settings(&self.user_id, settings).await
+    }
+
+    pub async fn has_settings(&self) -> Result<bool, DatabaseError> {
+        self.inner.has_settings(&self.user_id).await
+    }
+
+    // === Conversations ===
+
+    pub async fn create_conversation(
+        &self,
+        channel: &str,
+        thread_id: Option<&str>,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .create_conversation(channel, &self.user_id, thread_id)
+            .await
+    }
+
+    pub async fn ensure_conversation(
+        &self,
+        id: Uuid,
+        channel: &str,
+        thread_id: Option<&str>,
+    ) -> Result<bool, DatabaseError> {
+        self.inner
+            .ensure_conversation(id, channel, &self.user_id, thread_id)
+            .await
+    }
+
+    pub async fn list_conversations_with_preview(
+        &self,
+        channel: &str,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError> {
+        self.inner
+            .list_conversations_with_preview(&self.user_id, channel, limit)
+            .await
+    }
+
+    pub async fn list_conversations_all_channels(
+        &self,
+        limit: i64,
+    ) -> Result<Vec<ConversationSummary>, DatabaseError> {
+        self.inner
+            .list_conversations_all_channels(&self.user_id, limit)
+            .await
+    }
+
+    pub async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .get_or_create_routine_conversation(routine_id, routine_name, &self.user_id)
+            .await
+    }
+
+    pub async fn get_or_create_heartbeat_conversation(&self) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .get_or_create_heartbeat_conversation(&self.user_id)
+            .await
+    }
+
+    pub async fn get_or_create_assistant_conversation(
+        &self,
+        channel: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .get_or_create_assistant_conversation(&self.user_id, channel)
+            .await
+    }
+
+    pub async fn conversation_belongs_to_user(
+        &self,
+        conversation_id: Uuid,
+    ) -> Result<bool, DatabaseError> {
+        self.inner
+            .conversation_belongs_to_user(conversation_id, &self.user_id)
+            .await
+    }
+
+    /// Add a message to a conversation owned by this tenant.
+    ///
+    /// Verifies the conversation belongs to this user before adding.
+    pub async fn add_conversation_message(
+        &self,
+        conversation_id: Uuid,
+        role: &str,
+        content: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .add_conversation_message(conversation_id, role, content)
+            .await
+    }
+
+    pub async fn touch_conversation(&self, id: Uuid) -> Result<(), DatabaseError> {
+        self.inner.touch_conversation(id).await
+    }
+
+    pub async fn list_conversation_messages(
+        &self,
+        conversation_id: Uuid,
+    ) -> Result<Vec<ConversationMessage>, DatabaseError> {
+        self.inner.list_conversation_messages(conversation_id).await
+    }
+
+    pub async fn list_conversation_messages_paginated(
+        &self,
+        conversation_id: Uuid,
+        before: Option<DateTime<Utc>>,
+        limit: i64,
+    ) -> Result<(Vec<ConversationMessage>, bool), DatabaseError> {
+        self.inner
+            .list_conversation_messages_paginated(conversation_id, before, limit)
+            .await
+    }
+
+    pub async fn create_conversation_with_metadata(
+        &self,
+        channel: &str,
+        metadata: &serde_json::Value,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .create_conversation_with_metadata(channel, &self.user_id, metadata)
+            .await
+    }
+
+    pub async fn update_conversation_metadata_field(
+        &self,
+        id: Uuid,
+        key: &str,
+        value: &serde_json::Value,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .update_conversation_metadata_field(id, key, value)
+            .await
+    }
+
+    pub async fn get_conversation_metadata(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<serde_json::Value>, DatabaseError> {
+        self.inner.get_conversation_metadata(id).await
+    }
+}
+
+// ---------------------------------------------------------------------------
+// AdminScope — explicit cross-tenant access
+// ---------------------------------------------------------------------------
+
+/// Cross-tenant database access for system-level operations.
+///
+/// **Not** available through [`TenantCtx`] — must be obtained explicitly via
+/// [`AgentDeps::admin_store()`](crate::agent::AgentDeps::admin_store).
+///
+/// Used by: heartbeat enumeration, routine engine scheduling, self-repair,
+/// scheduler job persistence, worker status updates.
+#[derive(Clone)]
+pub struct AdminScope {
+    inner: Arc<dyn Database>,
+}
+
+impl AdminScope {
+    pub fn new(db: Arc<dyn Database>) -> Self {
+        Self { inner: db }
+    }
+
+    /// Access the raw Database trait object.
+    ///
+    /// Prefer using the typed methods on AdminScope instead. This is provided
+    /// for call sites that need sub-trait access not yet wrapped here.
+    pub fn db(&self) -> &Arc<dyn Database> {
+        &self.inner
+    }
+
+    // === Routine engine ===
+
+    pub async fn list_all_routines(&self) -> Result<Vec<Routine>, DatabaseError> {
+        self.inner.list_all_routines().await
+    }
+
+    pub async fn list_event_routines(&self) -> Result<Vec<Routine>, DatabaseError> {
+        self.inner.list_event_routines().await
+    }
+
+    pub async fn list_due_cron_routines(&self) -> Result<Vec<Routine>, DatabaseError> {
+        self.inner.list_due_cron_routines().await
+    }
+
+    pub async fn list_dispatched_routine_runs(&self) -> Result<Vec<RoutineRun>, DatabaseError> {
+        self.inner.list_dispatched_routine_runs().await
+    }
+
+    pub async fn count_running_routine_runs_batch(
+        &self,
+        routine_ids: &[Uuid],
+    ) -> Result<HashMap<Uuid, i64>, DatabaseError> {
+        self.inner
+            .count_running_routine_runs_batch(routine_ids)
+            .await
+    }
+
+    pub async fn batch_get_last_run_status(
+        &self,
+        routine_ids: &[Uuid],
+    ) -> Result<HashMap<Uuid, RunStatus>, DatabaseError> {
+        self.inner.batch_get_last_run_status(routine_ids).await
+    }
+
+    pub async fn count_running_routine_runs(&self, routine_id: Uuid) -> Result<i64, DatabaseError> {
+        self.inner.count_running_routine_runs(routine_id).await
+    }
+
+    pub async fn update_routine_runtime(
+        &self,
+        id: Uuid,
+        last_run_at: DateTime<Utc>,
+        next_fire_at: Option<DateTime<Utc>>,
+        run_count: u64,
+        consecutive_failures: u32,
+        state: &serde_json::Value,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .update_routine_runtime(
+                id,
+                last_run_at,
+                next_fire_at,
+                run_count,
+                consecutive_failures,
+                state,
+            )
+            .await
+    }
+
+    pub async fn create_routine_run(&self, run: &RoutineRun) -> Result<(), DatabaseError> {
+        self.inner.create_routine_run(run).await
+    }
+
+    pub async fn complete_routine_run(
+        &self,
+        id: Uuid,
+        status: RunStatus,
+        result_summary: Option<&str>,
+        tokens_used: Option<i32>,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .complete_routine_run(id, status, result_summary, tokens_used)
+            .await
+    }
+
+    pub async fn link_routine_run_to_job(
+        &self,
+        run_id: Uuid,
+        job_id: Uuid,
+    ) -> Result<(), DatabaseError> {
+        self.inner.link_routine_run_to_job(run_id, job_id).await
+    }
+
+    pub async fn get_routine(&self, id: Uuid) -> Result<Option<Routine>, DatabaseError> {
+        self.inner.get_routine(id).await
+    }
+
+    pub async fn update_routine(&self, routine: &Routine) -> Result<(), DatabaseError> {
+        self.inner.update_routine(routine).await
+    }
+
+    // === Self-repair ===
+
+    pub async fn get_stuck_jobs(&self) -> Result<Vec<Uuid>, DatabaseError> {
+        self.inner.get_stuck_jobs().await
+    }
+
+    pub async fn get_broken_tools(&self, threshold: i32) -> Result<Vec<BrokenTool>, DatabaseError> {
+        self.inner.get_broken_tools(threshold).await
+    }
+
+    pub async fn record_tool_failure(
+        &self,
+        tool_name: &str,
+        error_message: &str,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .record_tool_failure(tool_name, error_message)
+            .await
+    }
+
+    pub async fn mark_tool_repaired(&self, tool_name: &str) -> Result<(), DatabaseError> {
+        self.inner.mark_tool_repaired(tool_name).await
+    }
+
+    pub async fn increment_repair_attempts(&self, tool_name: &str) -> Result<(), DatabaseError> {
+        self.inner.increment_repair_attempts(tool_name).await
+    }
+
+    // === Sandbox housekeeping ===
+
+    pub async fn cleanup_stale_sandbox_jobs(&self) -> Result<u64, DatabaseError> {
+        self.inner.cleanup_stale_sandbox_jobs().await
+    }
+
+    pub async fn get_sandbox_job(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<SandboxJobRecord>, DatabaseError> {
+        self.inner.get_sandbox_job(id).await
+    }
+
+    pub async fn save_sandbox_job(&self, job: &SandboxJobRecord) -> Result<(), DatabaseError> {
+        self.inner.save_sandbox_job(job).await
+    }
+
+    pub async fn update_sandbox_job_status(
+        &self,
+        id: Uuid,
+        status: &str,
+        success: Option<bool>,
+        message: Option<&str>,
+        started_at: Option<DateTime<Utc>>,
+        completed_at: Option<DateTime<Utc>>,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .update_sandbox_job_status(id, status, success, message, started_at, completed_at)
+            .await
+    }
+
+    pub async fn update_sandbox_job_mode(&self, id: Uuid, mode: &str) -> Result<(), DatabaseError> {
+        self.inner.update_sandbox_job_mode(id, mode).await
+    }
+
+    pub async fn get_sandbox_job_mode(&self, id: Uuid) -> Result<Option<String>, DatabaseError> {
+        self.inner.get_sandbox_job_mode(id).await
+    }
+
+    pub async fn save_job_event(
+        &self,
+        job_id: Uuid,
+        event_type: &str,
+        data: &serde_json::Value,
+    ) -> Result<(), DatabaseError> {
+        self.inner.save_job_event(job_id, event_type, data).await
+    }
+
+    pub async fn list_job_events(
+        &self,
+        job_id: Uuid,
+        limit: Option<i64>,
+    ) -> Result<Vec<crate::history::JobEventRecord>, DatabaseError> {
+        self.inner.list_job_events(job_id, limit).await
+    }
+
+    // === Job persistence (scheduler, worker) ===
+
+    pub async fn get_job(&self, id: Uuid) -> Result<Option<JobContext>, DatabaseError> {
+        self.inner.get_job(id).await
+    }
+
+    pub async fn save_job(&self, ctx: &JobContext) -> Result<(), DatabaseError> {
+        self.inner.save_job(ctx).await
+    }
+
+    pub async fn update_job_status(
+        &self,
+        id: Uuid,
+        status: JobState,
+        failure_reason: Option<&str>,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .update_job_status(id, status, failure_reason)
+            .await
+    }
+
+    pub async fn mark_job_stuck(&self, id: Uuid) -> Result<(), DatabaseError> {
+        self.inner.mark_job_stuck(id).await
+    }
+
+    pub async fn list_agent_jobs(&self) -> Result<Vec<AgentJobRecord>, DatabaseError> {
+        self.inner.list_agent_jobs().await
+    }
+
+    pub async fn get_agent_job_failure_reason(
+        &self,
+        id: Uuid,
+    ) -> Result<Option<String>, DatabaseError> {
+        self.inner.get_agent_job_failure_reason(id).await
+    }
+
+    // === LLM call recording ===
+
+    pub async fn record_llm_call(&self, record: &LlmCallRecord<'_>) -> Result<Uuid, DatabaseError> {
+        self.inner.record_llm_call(record).await
+    }
+
+    pub async fn save_action(
+        &self,
+        job_id: Uuid,
+        action: &ActionRecord,
+    ) -> Result<(), DatabaseError> {
+        self.inner.save_action(job_id, action).await
+    }
+
+    pub async fn get_job_actions(&self, job_id: Uuid) -> Result<Vec<ActionRecord>, DatabaseError> {
+        self.inner.get_job_actions(job_id).await
+    }
+
+    // === Estimation ===
+
+    pub async fn save_estimation_snapshot(
+        &self,
+        job_id: Uuid,
+        category: &str,
+        tool_names: &[String],
+        estimated_cost: Decimal,
+        estimated_time_secs: i32,
+        estimated_value: Decimal,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .save_estimation_snapshot(
+                job_id,
+                category,
+                tool_names,
+                estimated_cost,
+                estimated_time_secs,
+                estimated_value,
+            )
+            .await
+    }
+
+    pub async fn update_estimation_actuals(
+        &self,
+        id: Uuid,
+        actual_cost: Decimal,
+        actual_time_secs: i32,
+        actual_value: Option<Decimal>,
+    ) -> Result<(), DatabaseError> {
+        self.inner
+            .update_estimation_actuals(id, actual_cost, actual_time_secs, actual_value)
+            .await
+    }
+
+    // === Conversations (admin context) ===
+
+    pub async fn add_conversation_message(
+        &self,
+        conversation_id: Uuid,
+        role: &str,
+        content: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .add_conversation_message(conversation_id, role, content)
+            .await
+    }
+
+    pub async fn get_or_create_routine_conversation(
+        &self,
+        routine_id: Uuid,
+        routine_name: &str,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .get_or_create_routine_conversation(routine_id, routine_name, user_id)
+            .await
+    }
+
+    pub async fn get_or_create_heartbeat_conversation(
+        &self,
+        user_id: &str,
+    ) -> Result<Uuid, DatabaseError> {
+        self.inner
+            .get_or_create_heartbeat_conversation(user_id)
+            .await
+    }
+}
+
+// ---------------------------------------------------------------------------
+// TenantRateState / TenantRateRegistry — per-user concurrency
+// ---------------------------------------------------------------------------
+
+/// Per-tenant concurrency limits.
+pub struct TenantRateState {
+    /// Limits concurrent LLM calls for this user.
+    pub llm_semaphore: Arc<Semaphore>,
+    /// Limits concurrent jobs for this user.
+    pub job_semaphore: Arc<Semaphore>,
+}
+
+impl TenantRateState {
+    pub fn new(max_llm_concurrent: usize, max_job_concurrent: usize) -> Self {
+        Self {
+            llm_semaphore: Arc::new(Semaphore::new(max_llm_concurrent)),
+            job_semaphore: Arc::new(Semaphore::new(max_job_concurrent)),
+        }
+    }
+}
+
+/// Registry that lazily creates per-tenant rate state.
+///
+/// Uses `tokio::sync::RwLock<HashMap>` (consistent with the rest of the
+/// codebase — no DashMap dependency).
+pub struct TenantRateRegistry {
+    state: tokio::sync::RwLock<HashMap<String, Arc<TenantRateState>>>,
+    max_llm_concurrent: usize,
+    max_job_concurrent: usize,
+}
+
+impl TenantRateRegistry {
+    pub fn new(max_llm_concurrent: usize, max_job_concurrent: usize) -> Self {
+        Self {
+            state: tokio::sync::RwLock::new(HashMap::new()),
+            max_llm_concurrent,
+            max_job_concurrent,
+        }
+    }
+
+    /// Get or lazily create rate state for a user.
+    pub async fn get_or_create(&self, user_id: &str) -> Arc<TenantRateState> {
+        // Fast path: read lock
+        {
+            let map = self.state.read().await;
+            if let Some(s) = map.get(user_id) {
+                return Arc::clone(s);
+            }
+        }
+        // Slow path: write lock with double-check
+        let mut map = self.state.write().await;
+        if let Some(s) = map.get(user_id) {
+            return Arc::clone(s);
+        }
+        let s = Arc::new(TenantRateState::new(
+            self.max_llm_concurrent,
+            self.max_job_concurrent,
+        ));
+        map.insert(user_id.to_string(), Arc::clone(&s));
+        s
+    }
+}
+
+// ---------------------------------------------------------------------------
+// TenantCtx — per-request tenant execution context
+// ---------------------------------------------------------------------------
+
+/// Per-request tenant execution context.
+///
+/// Bundles a [`TenantScope`] (scoped DB access), workspace, cost guard,
+/// and per-tenant rate limiting. Constructed once per request via
+/// [`AgentDeps::tenant_ctx()`](crate::agent::AgentDeps::tenant_ctx).
+///
+/// `Clone + Send + Sync` — safe to store on `ChatDelegate` without lifetime issues.
+#[derive(Clone)]
+pub struct TenantCtx {
+    user_id: String,
+    store: Option<TenantScope>,
+    workspace: Option<Arc<Workspace>>,
+    cost_guard: Arc<CostGuard>,
+    rate: Arc<TenantRateState>,
+}
+
+impl TenantCtx {
+    pub fn new(
+        user_id: impl Into<String>,
+        store: Option<TenantScope>,
+        workspace: Option<Arc<Workspace>>,
+        cost_guard: Arc<CostGuard>,
+        rate: Arc<TenantRateState>,
+    ) -> Self {
+        Self {
+            user_id: user_id.into(),
+            store,
+            workspace,
+            cost_guard,
+            rate,
+        }
+    }
+
+    pub fn user_id(&self) -> &str {
+        &self.user_id
+    }
+
+    pub fn store(&self) -> Option<&TenantScope> {
+        self.store.as_ref()
+    }
+
+    pub fn workspace(&self) -> Option<&Arc<Workspace>> {
+        self.workspace.as_ref()
+    }
+
+    pub fn cost_guard(&self) -> &CostGuard {
+        &self.cost_guard
+    }
+
+    /// Check cost limits for this tenant (global + per-user).
+    pub async fn check_cost_allowed(&self) -> Result<(), CostLimitExceeded> {
+        self.cost_guard.check_allowed_for_user(&self.user_id).await
+    }
+
+    /// Record an LLM call for this tenant.
+    #[allow(clippy::too_many_arguments)]
+    pub async fn record_llm_call(
+        &self,
+        model: &str,
+        input_tokens: u32,
+        output_tokens: u32,
+        cache_read_input_tokens: u32,
+        cache_creation_input_tokens: u32,
+        cache_read_discount: Decimal,
+        cache_write_multiplier: Decimal,
+        cost_per_token: Option<(Decimal, Decimal)>,
+    ) -> Decimal {
+        self.cost_guard
+            .record_llm_call_for_user(
+                &self.user_id,
+                model,
+                input_tokens,
+                output_tokens,
+                cache_read_input_tokens,
+                cache_creation_input_tokens,
+                cache_read_discount,
+                cache_write_multiplier,
+                cost_per_token,
+            )
+            .await
+    }
+
+    /// Acquire an LLM concurrency permit for this tenant.
+    pub async fn acquire_llm_permit(&self) -> Result<SemaphorePermit<'_>, crate::error::Error> {
+        self.rate.llm_semaphore.acquire().await.map_err(|_| {
+            crate::error::Error::Config(crate::error::ConfigError::InvalidValue {
+                key: "llm_semaphore".to_string(),
+                message: "semaphore closed".to_string(),
+            })
+        })
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn test_rate_registry_returns_same_state_for_same_user() {
+        let registry = TenantRateRegistry::new(4, 3);
+        let a1 = registry.get_or_create("alice").await;
+        let a2 = registry.get_or_create("alice").await;
+        assert!(Arc::ptr_eq(&a1, &a2));
+    }
+
+    #[tokio::test]
+    async fn test_rate_registry_different_users_get_different_state() {
+        let registry = TenantRateRegistry::new(4, 3);
+        let alice = registry.get_or_create("alice").await;
+        let bob = registry.get_or_create("bob").await;
+        assert!(!Arc::ptr_eq(&alice, &bob));
+    }
+}

src/testing/mod.rs

@@ -532,6 +532,7 @@ impl TestHarnessBuilder {
         let cost_guard = Arc::new(CostGuard::new(CostGuardConfig {
             max_cost_per_day_cents: None,
             max_actions_per_hour: None,
+            max_cost_per_user_per_day_cents: None,
         }));
 
         let channel = if self.stub_channel {
@@ -564,6 +565,7 @@ impl TestHarnessBuilder {
             sandbox_readiness: crate::agent::routine_engine::SandboxReadiness::DisabledByConfig,
             builder: None,
             llm_backend: "nearai".to_string(),
+            tenant_rates: std::sync::Arc::new(crate::tenant::TenantRateRegistry::new(4, 3)),
         };
 
         TestHarness {

src/tools/builtin/routine.rs

@@ -915,7 +915,7 @@ fn parse_routine_create_request(
 fn build_routine_trigger(trigger: &NormalizedTriggerRequest) -> Trigger {
     match trigger {
         NormalizedTriggerRequest::Cron { schedule, timezone } => Trigger::Cron {
-            schedule: schedule.clone(),
+            schedule: normalize_cron_expression(schedule),
             timezone: timezone.clone(),
         },
         NormalizedTriggerRequest::Manual => Trigger::Manual,
@@ -1836,6 +1836,20 @@ mod tests {
         assert_eq!(parsed.cooldown_secs, 30);
     }
 
+    #[test]
+    fn build_routine_trigger_normalizes_cron_schedule() {
+        let trigger = build_routine_trigger(&NormalizedTriggerRequest::Cron {
+            schedule: "0 0 9 * * MON-FRI".to_string(),
+            timezone: Some("UTC".to_string()),
+        });
+
+        assert!(matches!(
+            trigger,
+            Trigger::Cron { schedule, timezone }
+                if schedule == "0 0 9 * * MON-FRI *" && timezone.as_deref() == Some("UTC")
+        ));
+    }
+
     #[test]
     fn parses_grouped_message_event_with_tools() {
         let params = serde_json::json!({

src/worker/job.rs

@@ -18,8 +18,8 @@ use crate::agent::agentic_loop::{
 };
 use crate::agent::scheduler::WorkerMessage;
 use crate::agent::task::TaskOutput;
+use crate::channels::web::types::ToolDecisionDto;
 use crate::context::{ContextManager, JobState};
-use crate::db::Database;
 use crate::error::Error;
 use crate::hooks::HookRegistry;
 use crate::llm::{
@@ -27,6 +27,7 @@ use crate::llm::{
     ToolSelection,
 };
 use crate::safety::SafetyLayer;
+use crate::tenant::AdminScope;
 use crate::tools::execute::process_tool_result;
 use crate::tools::rate_limiter::RateLimitResult;
 use crate::tools::{
@@ -44,7 +45,7 @@ pub struct WorkerDeps {
     pub llm: Arc<dyn LlmProvider>,
     pub safety: Arc<SafetyLayer>,
     pub tools: Arc<ToolRegistry>,
-    pub store: Option<Arc<dyn Database>>,
+    pub store: Option<AdminScope>,
     pub hooks: Arc<HookRegistry>,
     pub timeout: Duration,
     pub use_planning: bool,
@@ -93,7 +94,7 @@ impl Worker {
         &self.deps.tools
     }
 
-    fn store(&self) -> Option<&Arc<dyn Database>> {
+    fn store(&self) -> Option<&AdminScope> {
         self.deps.store.as_ref()
     }
 
@@ -200,6 +201,19 @@ impl Worker {
                         .map(|s| s.to_string()),
                     fallback_deliverable: data.get("fallback_deliverable").cloned(),
                 }),
+                "reasoning" => {
+                    let narrative = data
+                        .get("narrative")
+                        .and_then(|v| v.as_str())
+                        .unwrap_or("")
+                        .to_string();
+                    let decisions = ToolDecisionDto::from_json_array(&data["decisions"]);
+                    Some(AppEvent::JobReasoning {
+                        job_id: job_id_str,
+                        narrative,
+                        decisions,
+                    })
+                }
                 _ => None,
             };
             if let Some(event) = event {
@@ -897,6 +911,11 @@ Report when the job is complete or if you encounter issues you cannot resolve."#
                         id: selection.tool_call_id.clone(),
                         name: selection.tool_name.clone(),
                         arguments: selection.parameters.clone(),
+                        reasoning: if action.reasoning.is_empty() {
+                            None
+                        } else {
+                            Some(action.reasoning.clone())
+                        },
                     }],
                 ));
 
@@ -1139,6 +1158,7 @@ impl<'a> JobDelegate<'a> {
         Ok(crate::llm::RespondOutput {
             result: RespondResult::Text(String::new()),
             usage: crate::llm::TokenUsage::default(),
+            finish_reason: crate::llm::FinishReason::Stop,
         })
     }
 }
@@ -1264,6 +1284,7 @@ impl<'a> LoopDelegate for JobDelegate<'a> {
                         content: reasoning_text,
                     },
                     usage: crate::llm::TokenUsage::default(),
+                    finish_reason: crate::llm::FinishReason::ToolUse,
                 });
             }
             Ok(_) => {} // empty selections, fall through
@@ -1357,6 +1378,48 @@ impl<'a> LoopDelegate for JobDelegate<'a> {
             );
         }
 
+        // Emit reasoning event if any tool calls carry reasoning.
+        // Sanitize narrative and per-tool rationale through SafetyLayer
+        // (parity with ChatDelegate in dispatcher.rs).
+        let sanitized_narrative = content
+            .as_deref()
+            .filter(|c| !c.trim().is_empty())
+            .map(|c| {
+                self.worker
+                    .deps
+                    .safety
+                    .sanitize_tool_output("job_narrative", c)
+                    .content
+            })
+            .filter(|c| !c.trim().is_empty())
+            .unwrap_or_default();
+        let decisions: Vec<serde_json::Value> = tool_calls
+            .iter()
+            .filter_map(|tc| {
+                tc.reasoning.as_ref().map(|r| {
+                    let sanitized = self
+                        .worker
+                        .deps
+                        .safety
+                        .sanitize_tool_output("tool_rationale", r)
+                        .content;
+                    serde_json::json!({
+                        "tool_name": tc.name,
+                        "rationale": sanitized,
+                    })
+                })
+            })
+            .collect();
+        if !decisions.is_empty() {
+            self.worker.log_event(
+                "reasoning",
+                serde_json::json!({
+                    "narrative": sanitized_narrative,
+                    "decisions": decisions,
+                }),
+            );
+        }
+
         // Add assistant message with tool_calls (OpenAI protocol)
         reason_ctx
             .messages
@@ -1371,7 +1434,7 @@ impl<'a> LoopDelegate for JobDelegate<'a> {
             .map(|tc| ToolSelection {
                 tool_name: tc.name.clone(),
                 parameters: tc.arguments.clone(),
-                reasoning: String::new(),
+                reasoning: tc.reasoning.clone().unwrap_or_default(),
                 alternatives: vec![],
                 tool_call_id: tc.id.clone(),
             })
@@ -1424,6 +1487,11 @@ fn selections_to_tool_calls(selections: &[ToolSelection]) -> Vec<ToolCall> {
             id: s.tool_call_id.clone(),
             name: s.tool_name.clone(),
             arguments: s.parameters.clone(),
+            reasoning: if s.reasoning.is_empty() {
+                None
+            } else {
+                Some(s.reasoning.clone())
+            },
         })
         .collect()
 }

src/workspace/mod.rs

@@ -149,6 +149,7 @@ fn reject_if_injected(path: &str, content: &str) -> Result<(), WorkspaceError> {
 ///
 /// Allows Workspace to work with either a PostgreSQL `Repository` (the original
 /// path) or any `Database` trait implementation (e.g. libSQL backend).
+#[derive(Clone)]
 enum WorkspaceStorage {
     /// PostgreSQL-backed repository (uses connection pool directly).
     #[cfg(feature = "postgres")]
@@ -576,6 +577,60 @@ impl Workspace {
         self
     }
 
+    /// Clone the workspace configuration for a different primary user scope.
+    ///
+    /// This preserves search config, embeddings, shared read scopes, memory
+    /// layers, and privacy classifier while switching the primary read/write
+    /// scope to `user_id`.
+    pub fn scoped_to_user(&self, user_id: impl Into<String>) -> Self {
+        let user_id = user_id.into();
+
+        let mut memory_layers = self.memory_layers.clone();
+        for layer in &mut memory_layers {
+            if layer.sensitivity == crate::workspace::layer::LayerSensitivity::Private
+                && layer.scope == self.user_id
+            {
+                layer.scope = user_id.clone();
+            }
+        }
+
+        let mut read_user_ids = vec![user_id.clone()];
+        for scope in &self.read_user_ids {
+            if scope != &self.user_id && !read_user_ids.contains(scope) {
+                read_user_ids.push(scope.clone());
+            }
+        }
+        for scope in crate::workspace::layer::MemoryLayer::read_scopes(&memory_layers) {
+            if !read_user_ids.contains(&scope) {
+                read_user_ids.push(scope);
+            }
+        }
+
+        let preserve_flags = user_id == self.user_id;
+        Self {
+            user_id,
+            read_user_ids,
+            agent_id: self.agent_id,
+            storage: self.storage.clone(),
+            embeddings: self.embeddings.clone(),
+            bootstrap_pending: std::sync::atomic::AtomicBool::new(if preserve_flags {
+                self.bootstrap_pending
+                    .load(std::sync::atomic::Ordering::Acquire)
+            } else {
+                false
+            }),
+            bootstrap_completed: std::sync::atomic::AtomicBool::new(if preserve_flags {
+                self.bootstrap_completed
+                    .load(std::sync::atomic::Ordering::Acquire)
+            } else {
+                false
+            }),
+            search_defaults: self.search_defaults.clone(),
+            memory_layers,
+            privacy_classifier: self.privacy_classifier.clone(),
+        }
+    }
+
     /// Get the user ID (primary scope for writes).
     pub fn user_id(&self) -> &str {
         &self.user_id

src/workspace/repository.rs

@@ -15,6 +15,7 @@ use crate::workspace::document::{MemoryChunk, MemoryDocument, WorkspaceEntry};
 use crate::workspace::search::{RankedResult, SearchConfig, SearchResult, fuse_results};
 
 /// Database repository for workspace operations.
+#[derive(Clone)]
 pub struct Repository {
     pool: Pool,
 }

tests/e2e/scenarios/test_telegram_hot_activation.py

@@ -253,6 +253,6 @@ async def test_telegram_hot_activation_transitions_installed_to_active(page):
     assert await card.locator(SEL["ext_pairing_label"]).count() == 0
 
     assert captured_setup_payloads == [
-        {"secrets": {"telegram_bot_token": "123456789:ABCdefGhI"}},
-        {"secrets": {}},
+        {"secrets": {"telegram_bot_token": "123456789:ABCdefGhI"}, "fields": {}},
+        {"secrets": {}, "fields": {}},
     ]

tests/e2e_advanced_traces.rs

@@ -587,6 +587,7 @@ mod advanced {
     async fn mcp_extension_lifecycle() {
         use crate::support::mock_mcp_server::{MockToolResponse, start_mock_mcp_server};
         use ironclaw::extensions::{AuthHint, ExtensionKind, ExtensionSource, RegistryEntry};
+        const TEST_USER_ID: &str = "test-user";
 
         // 1. Start mock MCP server with pre-configured tool responses.
         let mock_server = start_mock_mcp_server(vec![
@@ -654,14 +655,14 @@ mod advanced {
         ext_mgr
             .secrets()
             .create(
-                "default",
+                TEST_USER_ID,
                 ironclaw::secrets::CreateSecretParams::new(secret_name, "mock-access-token")
                     .with_provider("mcp:mock-notion".to_string()),
             )
             .await
             .expect("failed to inject test token");
 
-        let activate_result = ext_mgr.activate("mock-notion", "default").await;
+        let activate_result = ext_mgr.activate("mock-notion", TEST_USER_ID).await;
         assert!(
             activate_result.is_ok(),
             "activation failed: {:?}",

tests/e2e_builtin_tool_coverage.rs

@@ -439,7 +439,7 @@ mod tests {
 
         match &routine.trigger {
             Trigger::Cron { schedule, timezone } => {
-                assert_eq!(schedule, "0 0 9 * * MON-FRI");
+                assert_eq!(schedule, "0 0 9 * * MON-FRI *");
                 assert_eq!(timezone.as_deref(), Some("UTC"));
             }
             other => panic!("expected cron trigger, got {other:?}"),

tests/e2e_routine_heartbeat.rs

@@ -337,14 +337,14 @@ mod tests {
             SchedulerDeps {
                 tools: registry.clone(),
                 extension_manager: extension_manager.clone(),
-                store: Some(db.clone()),
+                store: Some(ironclaw::tenant::AdminScope::new(db.clone())),
                 hooks: Arc::new(HookRegistry::new()),
             },
         ));
 
         Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db,
+            ironclaw::tenant::AdminScope::new(db),
             llm,
             ws,
             notify_tx,
@@ -448,7 +448,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -527,7 +527,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -614,7 +614,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -723,7 +723,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -866,7 +866,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -1049,7 +1049,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            Arc::clone(&db),
+            ironclaw::tenant::AdminScope::new(Arc::clone(&db)),
             llm,
             ws,
             notify_tx,
@@ -1171,7 +1171,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             RoutineConfig::default(),
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,
@@ -1279,7 +1279,7 @@ mod tests {
 
         let engine = Arc::new(RoutineEngine::new(
             config,
-            db.clone(),
+            ironclaw::tenant::AdminScope::new(db.clone()),
             llm,
             ws,
             notify_tx,

tests/e2e_telegram_message_routing.rs

@@ -201,6 +201,7 @@ mod tests {
             sandbox_readiness: ironclaw::agent::SandboxReadiness::DisabledByConfig,
             builder: None,
             llm_backend: "nearai".to_string(),
+            tenant_rates: std::sync::Arc::new(ironclaw::tenant::TenantRateRegistry::new(4, 3)),
         };
 
         let gateway = Arc::new(TestChannel::new());

tests/e2e_workspace_coverage.rs

@@ -12,6 +12,7 @@ mod tests {
 
     use crate::support::test_rig::TestRigBuilder;
     use crate::support::trace_llm::LlmTrace;
+    use ironclaw::workspace::Workspace;
 
     // -----------------------------------------------------------------------
     // Test 1: write_chunk_search
@@ -268,6 +269,7 @@ mod tests {
 
     #[tokio::test]
     async fn identity_in_system_prompt() {
+        const TEST_USER_ID: &str = "test-user";
         let trace = LlmTrace::from_file(concat!(
             env!("CARGO_MANIFEST_DIR"),
             "/tests/fixtures/llm_traces/workspace/identity_prompt.json"
@@ -280,7 +282,7 @@ mod tests {
             .await;
 
         // Seed an IDENTITY.md so the system prompt has real content to inject.
-        let ws = rig.workspace().expect("workspace must be available");
+        let ws = Workspace::new_with_db(TEST_USER_ID, rig.database().clone());
         ws.write(
             "IDENTITY.md",
             "I am TestBot, a helpful testing assistant created for E2E verification.",

tests/multi_tenant_system_prompt.rs

@@ -1,10 +1,10 @@
-//! Tests proving that multi-tenant system prompts are broken.
+//! Regression tests for multi-tenant system prompts.
 //!
-//! Bug: In multi-tenant mode, the agent loop uses `self.workspace()` which
-//! returns a single shared workspace (user_id="default"). Identity files
-//! (IDENTITY.md, SOUL.md, USER.md) seeded under per-user IDs ("alice",
-//! "bob") are invisible to this workspace, so the system prompt is
-//! empty/wrong.
+//! The agent must build the conversational system prompt from a workspace
+//! scoped to the incoming message's user, not from the shared owner-scope
+//! workspace created at startup. Otherwise per-user identity files
+//! (IDENTITY.md, SOUL.md, USER.md) become invisible and different users can
+//! see the same owner-scoped prompt.
 //!
 //! These tests:
 //! 1. Seed identity files for two users (alice, bob) in the database
@@ -13,7 +13,7 @@
 //!    correct user's identity
 //! 4. Verify user A's identity doesn't leak into user B's prompt
 //!
-//! All tests are expected to FAIL until the bug is fixed.
+//! These tests ensure each user's identity is isolated correctly.
 
 #[cfg(feature = "libsql")]
 mod support;

tests/openai_compat_integration.rs

@@ -94,6 +94,7 @@ impl LlmProvider for MockLlmProvider {
                     id: "call_mock_001".to_string(),
                     name: tool.name.clone(),
                     arguments: serde_json::json!({"test": true}),
+                    reasoning: None,
                 }],
                 input_tokens: 15,
                 output_tokens: 8,

tests/support/gateway_workflow_harness.rs

@@ -267,6 +267,7 @@ impl GatewayWorkflowHarness {
                 sandbox_readiness: ironclaw::agent::SandboxReadiness::DisabledByConfig,
                 builder: None,
                 llm_backend: "nearai".to_string(),
+                tenant_rates: std::sync::Arc::new(ironclaw::tenant::TenantRateRegistry::new(4, 3)),
             },
             channels,
             None,

tests/support/test_rig.rs

@@ -642,7 +642,7 @@ impl TestRigBuilder {
                 let (notify_tx, _notify_rx) = tokio::sync::mpsc::channel(16);
                 let engine = Arc::new(RoutineEngine::new(
                     routine_config,
-                    Arc::clone(db_arc),
+                    ironclaw::tenant::AdminScope::new(Arc::clone(db_arc)),
                     components.llm.clone(),
                     Arc::clone(ws),
                     notify_tx,
@@ -762,6 +762,7 @@ impl TestRigBuilder {
             sandbox_readiness: ironclaw::agent::SandboxReadiness::Available, // tests don't use real Docker
             builder: None,
             llm_backend: "nearai".to_string(),
+            tenant_rates: std::sync::Arc::new(ironclaw::tenant::TenantRateRegistry::new(4, 3)),
         };
 
         // 7. Create TestChannel and ChannelManager.

tests/support/trace_llm.rs

@@ -566,6 +566,7 @@ impl LlmProvider for TraceLlm {
                         id: tc.id,
                         name: tc.name,
                         arguments: tc.arguments,
+                        reasoning: None,
                     })
                     .collect();
                 Ok(ToolCompletionResponse {