huskies: merge 1213 story Chat "stop" command that immediately aborts the in-flight LLM turn

server/src/chat/commands/mod.rs

@@ -229,6 +229,11 @@ pub fn commands() -> &'static [BotCommand] {
             description: "Distill the session transcript into a seed, then reset (unlike `reset`, keeps distilled context for the next message)",
             handler: handle_compact_fallback,
         },
+        BotCommand {
+            name: "stop",
+            description: "Cancel the in-flight LLM turn for this room (synonyms: `halt`, `abort`)",
+            handler: handle_stop_fallback,
+        },
         BotCommand {
             name: "timer",
             description: "Schedule a deferred agent start: `timer <story_id> <HH:MM>`, `timer list`, `timer cancel <story_id>`",
@@ -439,6 +444,16 @@ fn handle_compact_fallback(_ctx: &CommandContext) -> Option<String> {
     None
 }
 
+/// Fallback handler for the `stop` command when it is not intercepted by the
+/// async handler in `on_room_message`.  In practice this is never called —
+/// stop is detected and handled before `try_handle_command` is invoked.
+/// The entry exists in the registry only so `help` lists it.
+///
+/// Returns `None` to prevent the LLM from receiving "stop" as a prompt.
+fn handle_stop_fallback(_ctx: &CommandContext) -> Option<String> {
+    None
+}
+
 /// Fallback handler for the `cleanup_worktrees` command when it is not
 /// intercepted by the async handler in `on_room_message`.  In practice this is
 /// never called — cleanup_worktrees is detected and handled before

server/src/chat/dispatcher.rs

@@ -19,6 +19,7 @@
 use crate::slog;
 use std::collections::HashMap;
 use std::pin::Pin;
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::{Arc, Mutex};
 use std::time::Duration;
 use tokio::sync::{mpsc, watch};
@@ -47,6 +48,20 @@ enum SessionMsg {
 
 struct SessionHandle {
     tx: mpsc::UnboundedSender<SessionMsg>,
+    /// `true` whenever the session task is coalescing, running, or draining a
+    /// batch — i.e. there is something for [`ChatDispatcher::stop`] to cancel.
+    /// `false` while the task is idle in Phase 1, waiting for a message.
+    active: Arc<AtomicBool>,
+}
+
+/// Result of a [`ChatDispatcher::stop`] call.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StopOutcome {
+    /// A turn was actively coalescing, running, or draining pending messages
+    /// for the session, and a cancellation was sent.
+    Cancelled,
+    /// The session doesn't exist, or exists but has no turn in flight.
+    NothingRunning,
 }
 
 /// Coalescing, serialising dispatcher for chat-to-LLM message routing.
@@ -79,8 +94,14 @@ impl ChatDispatcher {
         let coalesce_ms = self.coalesce_ms;
         let handle = guard.entry(session_key.clone()).or_insert_with(|| {
             let (tx, rx) = mpsc::unbounded_channel();
-            tokio::spawn(session_task(session_key.clone(), rx, coalesce_ms));
-            SessionHandle { tx }
+            let active = Arc::new(AtomicBool::new(false));
+            tokio::spawn(session_task(
+                session_key.clone(),
+                rx,
+                coalesce_ms,
+                Arc::clone(&active),
+            ));
+            SessionHandle { tx, active }
         });
         let _ = handle.tx.send(SessionMsg::UserMessage {
             text: message,
@@ -90,15 +111,19 @@ impl ChatDispatcher {
 
     /// Stop the active LLM run for `session_key` and clear its pending queue.
     ///
-    /// Returns `true` if the session existed (whether or not anything was
-    /// actually running), `false` if no session for that key has been created.
-    pub fn stop(&self, session_key: &str) -> bool {
+    /// Returns [`StopOutcome::Cancelled`] and sends the cancellation only when
+    /// a turn is actually coalescing, running, or draining pending messages;
+    /// otherwise returns [`StopOutcome::NothingRunning`] without sending
+    /// anything (covers both "no session was ever created" and "session
+    /// exists but is idle").
+    pub fn stop(&self, session_key: &str) -> StopOutcome {
         let guard = self.sessions.lock().unwrap();
-        if let Some(handle) = guard.get(session_key) {
+        match guard.get(session_key) {
+            Some(handle) if handle.active.load(Ordering::SeqCst) => {
                 let _ = handle.tx.send(SessionMsg::Stop);
-            true
-        } else {
-            false
+                StopOutcome::Cancelled
+            }
+            _ => StopOutcome::NothingRunning,
         }
     }
 }
@@ -119,6 +144,7 @@ async fn session_task(
     session_key: String,
     mut rx: mpsc::UnboundedReceiver<SessionMsg>,
     coalesce_ms: u64,
+    active: Arc<AtomicBool>,
 ) {
     let coalesce_dur = Duration::from_millis(coalesce_ms);
 
@@ -131,6 +157,9 @@ async fn session_task(
                 Some(SessionMsg::UserMessage { text, factory }) => break (text, factory),
             }
         };
+        // From here until we loop back to Phase 1, there is something in
+        // flight (coalescing, running, or draining) for `stop()` to cancel.
+        active.store(true, Ordering::SeqCst);
 
         // ── Phase 2: coalesce window (debounce) ──────────────────────────────
         let mut batch: Vec<String> = vec![first_text];
@@ -160,6 +189,7 @@ async fn session_task(
         }
 
         if batch.is_empty() {
+            active.store(false, Ordering::SeqCst);
             continue; // Stop received during coalesce — restart
         }
 
@@ -209,6 +239,7 @@ async fn session_task(
             }
 
             if stopped || pending_texts.is_empty() {
+                active.store(false, Ordering::SeqCst);
                 break; // back to Phase 1
             }
 
@@ -351,7 +382,12 @@ mod tests {
         );
 
         // Stop immediately.
-        dispatcher.stop(&session);
+        let outcome = dispatcher.stop(&session);
+        assert_eq!(
+            outcome,
+            StopOutcome::Cancelled,
+            "a run was active, so stop must report Cancelled"
+        );
 
         // Wait longer than the run would have taken if not stopped.
         tokio::time::sleep(Duration::from_millis(700)).await;
@@ -364,4 +400,35 @@ mod tests {
             "stop should discard pending; got {count} spawns"
         );
     }
+
+    /// Stopping a session key that was never submitted reports NothingRunning
+    /// rather than silently no-opping.
+    #[tokio::test]
+    async fn stop_on_unknown_session_reports_nothing_running() {
+        let dispatcher = ChatDispatcher::new(30);
+        assert_eq!(
+            dispatcher.stop("never-seen-room"),
+            StopOutcome::NothingRunning
+        );
+    }
+
+    /// Stopping a session that exists but has no active turn (its one run
+    /// already completed) reports NothingRunning, not Cancelled.
+    #[tokio::test]
+    async fn stop_on_idle_session_reports_nothing_running() {
+        let spawn_count = Arc::new(AtomicUsize::new(0));
+        let dispatcher = ChatDispatcher::new(20);
+        let session = "room4".to_string();
+
+        dispatcher.submit(
+            session.clone(),
+            "hello".to_string(),
+            make_factory(Arc::clone(&spawn_count), 20),
+        );
+
+        // Wait long enough for the coalesce window and the run to finish.
+        tokio::time::sleep(Duration::from_millis(200)).await;
+
+        assert_eq!(dispatcher.stop(&session), StopOutcome::NothingRunning);
+    }
 }

server/src/chat/transport/matrix/bot/messages/handle_message.rs

@@ -3,7 +3,7 @@
 
 use crate::chat::ChatTransport;
 use crate::chat::util::drain_complete_paragraphs;
-use crate::llm::providers::claude_code::{ClaudeCodeProvider, ClaudeCodeResult};
+use crate::llm::providers::claude_code::{CANCELLED, ClaudeCodeProvider, ClaudeCodeResult};
 use crate::slog;
 use matrix_sdk::ruma::OwnedRoomId;
 use std::sync::Arc;
@@ -184,6 +184,7 @@ pub(in crate::chat::transport::matrix::bot) async fn handle_message(
     let remaining = buffer.lock().unwrap().trim().to_string();
     let did_send_any = sent_any_chunk.load(Ordering::Relaxed);
 
+    let mut was_cancelled = false;
     let (assistant_reply, new_session_id, turn_usage) = match result {
         Ok(ClaudeCodeResult {
             messages,
@@ -213,6 +214,15 @@ pub(in crate::chat::transport::matrix::bot) async fn handle_message(
             slog!("[matrix-bot] session_id from chat_stream: {:?}", session_id);
             (reply, session_id, usage)
         }
+        Err(e) if e == CANCELLED => {
+            // A user-initiated "stop" — not a crash. The stop handler already
+            // sent a confirmation, so don't post a second message here, and
+            // don't run the crash-recovery path below (which would otherwise
+            // clear the room's session_id as if the session were poisoned).
+            slog!("[matrix-bot] LLM turn cancelled via stop for room {room_id}");
+            was_cancelled = true;
+            (String::new(), None, None)
+        }
         Err(e) => {
             slog!("[matrix-bot] LLM error: {e}");
             let err_msg = if let Some(url) = crate::llm::oauth::extract_login_url_from_error(&e) {
@@ -230,6 +240,10 @@ pub(in crate::chat::transport::matrix::bot) async fn handle_message(
     drop(msg_tx);
     let _ = post_task.await;
 
+    if was_cancelled {
+        return;
+    }
+
     // Record this exchange in the per-room conversation history and persist
     // the session ID so the next turn resumes with structured API messages.
     let mut compact_suggestion: Option<String> = None;

server/src/chat/transport/matrix/bot/messages/on_room_message.rs

@@ -295,6 +295,47 @@ async fn try_handle_compact_command(
     true
 }
 
+/// Attempt to handle an addressed message as a bare `stop`/`halt`/`abort`
+/// command.
+///
+/// Returns `true` when `user_message` was recognised as a stop command and
+/// handled: the in-flight turn (if any) was cancelled via
+/// [`crate::chat::dispatcher::ChatDispatcher::stop`] and a confirmation was
+/// sent via `ctx.transport`. Returns `false` so the caller falls through to
+/// the dispatcher submit path — this must run before `chat_dispatcher.submit`
+/// so a bare "stop" is never itself queued as a prompt.
+async fn try_handle_stop_command(
+    ctx: &BotContext,
+    sender: &str,
+    user_message: &str,
+    room_id_str: &str,
+) -> bool {
+    if super::super::super::stop::extract_stop_command(
+        user_message,
+        &ctx.services.bot_name,
+        ctx.matrix_user_id.as_str(),
+    )
+    .is_none()
+    {
+        return false;
+    }
+    slog!("[matrix-bot] stop command from {sender} for session {room_id_str}");
+    let response = match ctx.services.chat_dispatcher.stop(room_id_str) {
+        crate::chat::dispatcher::StopOutcome::Cancelled => "Stopped.",
+        crate::chat::dispatcher::StopOutcome::NothingRunning => "Nothing is currently running.",
+    };
+    let html = markdown_to_html(response);
+    if let Ok(msg_id) = ctx
+        .transport
+        .send_message(room_id_str, response, &html)
+        .await
+        && let Ok(event_id) = msg_id.parse()
+    {
+        ctx.bot_sent_event_ids.lock().await.insert(event_id);
+    }
+    true
+}
+
 pub(in crate::chat::transport::matrix::bot) async fn on_room_message(
     ev: OriginalSyncRoomMessageEvent,
     room: Room,
@@ -490,6 +531,13 @@ pub(in crate::chat::transport::matrix::bot) async fn on_room_message(
             // entry silently swallows the command (bug: compact no-op in
             // gateway mode, distinct from the 1192/1205 registry-ordering bug).
             "compact",
+            // `stop` is a gateway-local session command (sibling of `compact`):
+            // it cancels THIS gateway's in-flight LLM turn via the local
+            // `chat_dispatcher`. It must be listed here for the same reason
+            // as `compact` above — once registered in `commands()` (so `help`
+            // lists it), the gateway proxy would otherwise forward it to the
+            // active project's sled instead of executing it locally.
+            "stop",
             "switch",
             "all_status",
             "new",
@@ -1403,28 +1451,11 @@ pub(in crate::chat::transport::matrix::bot) async fn on_room_message(
         return;
     }
 
-    // "stop" — cancel the running LLM turn for this session and clear pending queue.
-    {
-        let stripped = crate::chat::util::strip_bot_mention(
-            &user_message,
-            &ctx.services.bot_name,
-            ctx.matrix_user_id.as_str(),
-        )
-        .trim()
-        .to_ascii_lowercase();
-        if stripped == "stop" {
-            slog!("[matrix-bot] stop command from {sender} for session {room_id_str}");
-            ctx.services.chat_dispatcher.stop(&room_id_str);
-            let msg = "Stopped.";
-            let html = markdown_to_html(msg);
-            if let Ok(msg_id) = ctx.transport.send_message(&room_id_str, msg, &html).await
-                && let Ok(event_id) = msg_id.parse()
-            {
-                ctx.bot_sent_event_ids.lock().await.insert(event_id);
-            }
+    // "stop"/"halt"/"abort" — cancel the running LLM turn for this session
+    // and clear its pending queue.
+    if try_handle_stop_command(&ctx, &sender, &user_message, &room_id_str).await {
         return;
     }
-    }
 
     // Hand the message to the protocol-agnostic dispatcher instead of spawning
     // directly.  The dispatcher applies a coalesce window and a per-session
@@ -1465,7 +1496,7 @@ pub(in crate::chat::transport::matrix::bot) async fn on_room_message(
 mod tests {
     use super::{
         eval_gateway_overview_command, eval_gateway_status_command, eval_switch_command,
-        try_handle_compact_command,
+        try_handle_compact_command, try_handle_stop_command,
     };
     use crate::chat::{ChatTransport, MessageId};
     use crate::service::gateway::config::ProjectEntry;
@@ -1987,4 +2018,97 @@ mod tests {
             "no reply should be sent for a message that isn't compact"
         );
     }
+
+    // ── stop (story 1213) ──────────────────────────────────────────────
+
+    /// AC1: a bare "stop" cancels an active run and replies "Stopped.".
+    #[tokio::test]
+    async fn stop_command_cancels_active_run_and_confirms() {
+        use std::sync::Arc;
+
+        let room_id: matrix_sdk::ruma::OwnedRoomId = "!test:example.com".parse().unwrap();
+        let project_root_dir = tempfile::tempdir().unwrap();
+        let services = crate::services::Services::new_test(
+            project_root_dir.path().to_path_buf(),
+            "Huskies".to_string(),
+        );
+        let transport = Arc::new(CapturingTransport::new());
+        let ctx = make_test_ctx(services, transport.clone());
+
+        // Start a long-running "turn" on the dispatcher for this room.
+        let factory: crate::chat::dispatcher::SpawnFn = Arc::new(|_prompt, _cancel_rx| {
+            Box::pin(async move {
+                tokio::time::sleep(std::time::Duration::from_millis(500)).await;
+            })
+        });
+        ctx.services.chat_dispatcher.submit(
+            room_id.to_string(),
+            "long-running".to_string(),
+            factory,
+        );
+        // Let the coalesce window fire so the run is actually active.
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+
+        let handled =
+            try_handle_stop_command(&ctx, "@alice:example.com", "stop", room_id.as_str()).await;
+
+        assert!(handled, "bare stop must be recognized and handled");
+        let sent = transport.sent.lock().unwrap().clone();
+        assert_eq!(sent.len(), 1, "exactly one reply must be sent");
+        assert_eq!(sent[0].1, "Stopped.");
+    }
+
+    /// AC6: stop with nothing running replies that nothing is running,
+    /// instead of silently no-opping.
+    #[tokio::test]
+    async fn stop_command_with_nothing_running_reports_nothing_running() {
+        use std::sync::Arc;
+
+        let room_id: matrix_sdk::ruma::OwnedRoomId = "!test:example.com".parse().unwrap();
+        let project_root_dir = tempfile::tempdir().unwrap();
+        let services = crate::services::Services::new_test(
+            project_root_dir.path().to_path_buf(),
+            "Huskies".to_string(),
+        );
+        let transport = Arc::new(CapturingTransport::new());
+        let ctx = make_test_ctx(services, transport.clone());
+
+        let handled =
+            try_handle_stop_command(&ctx, "@alice:example.com", "stop", room_id.as_str()).await;
+
+        assert!(handled, "bare stop must be recognized and handled");
+        let sent = transport.sent.lock().unwrap().clone();
+        assert_eq!(sent.len(), 1, "exactly one reply must be sent");
+        assert_eq!(sent[0].1, "Nothing is currently running.");
+    }
+
+    /// AC4: a non-bare message that merely contains the word "stop" must
+    /// fall through untouched, never treated as an abort.
+    #[tokio::test]
+    async fn stop_with_trailing_text_falls_through() {
+        use std::sync::Arc;
+
+        let room_id: matrix_sdk::ruma::OwnedRoomId = "!test:example.com".parse().unwrap();
+        let project_root_dir = tempfile::tempdir().unwrap();
+        let services = crate::services::Services::new_test(
+            project_root_dir.path().to_path_buf(),
+            "Huskies".to_string(),
+        );
+        let transport = Arc::new(CapturingTransport::new());
+        let ctx = make_test_ctx(services, transport.clone());
+
+        let handled = try_handle_stop_command(
+            &ctx,
+            "@alice:example.com",
+            "stop the deployment and redeploy",
+            room_id.as_str(),
+        )
+        .await;
+
+        assert!(!handled, "non-bare 'stop' text must fall through");
+        assert!(
+            transport.sent.lock().unwrap().is_empty(),
+            "no reply should be sent for a message that isn't a bare stop"
+        );
+    }
 }

server/src/chat/transport/matrix/mod.rs

@@ -52,6 +52,9 @@ pub mod rmtree;
 pub mod sled_upgrade;
 /// Start command — handles `!start` bot commands to launch agents on stories.
 pub mod start;
+/// Stop/halt/abort command — cancels the in-flight LLM turn for this room
+/// and clears any coalesced/pending messages.
+pub mod stop;
 /// Matrix `ChatTransport` implementation wrapping the Matrix SDK client.
 pub mod transport_impl;
 

server/src/chat/transport/matrix/stop.rs

@@ -0,0 +1,103 @@
+//! Stop command: cancel the in-flight LLM turn for the current room's chat
+//! session.
+//!
+//! `{bot_name} stop` (or `halt`/`abort`) cancels the actively-running Claude
+//! Code turn for this room via [`crate::chat::dispatcher::ChatDispatcher`]
+//! and clears any coalesced/pending messages. Unlike [`super::reset`] and
+//! [`crate::chat::compact`], which match on the first word and ignore
+//! trailing text, this requires an EXACT match (after mention-stripping) —
+//! an instruction like "stop X and do Y" must not be misread as an abort.
+
+use crate::chat::util::strip_bot_mention;
+
+/// A parsed stop/halt/abort command.
+#[derive(Debug, PartialEq)]
+pub struct StopCommand;
+
+/// Parse a stop command from a raw message body.
+///
+/// Strips the bot mention prefix and requires the *entire* remaining,
+/// trimmed message to equal `stop`, `halt`, or `abort` (case-insensitive).
+/// Returns `None` for anything else, including a recognised word followed by
+/// trailing text (e.g. "stop the deployment").
+pub fn extract_stop_command(
+    message: &str,
+    bot_name: &str,
+    bot_user_id: &str,
+) -> Option<StopCommand> {
+    let stripped = strip_bot_mention(message, bot_name, bot_user_id);
+    let trimmed = stripped.trim();
+
+    if trimmed.eq_ignore_ascii_case("stop")
+        || trimmed.eq_ignore_ascii_case("halt")
+        || trimmed.eq_ignore_ascii_case("abort")
+    {
+        Some(StopCommand)
+    } else {
+        None
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Tests
+// ---------------------------------------------------------------------------
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn extract_bare_stop() {
+        let cmd = extract_stop_command("stop", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_with_display_name() {
+        let cmd = extract_stop_command("Timmy stop", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_with_full_user_id() {
+        let cmd = extract_stop_command("@timmy:home.local stop", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_halt_synonym() {
+        let cmd = extract_stop_command("Timmy halt", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_abort_synonym() {
+        let cmd = extract_stop_command("abort", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_case_insensitive() {
+        let cmd = extract_stop_command("Timmy STOP", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, Some(StopCommand));
+    }
+
+    #[test]
+    fn extract_non_stop_returns_none() {
+        let cmd = extract_stop_command("Timmy help", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, None);
+    }
+
+    #[test]
+    fn extract_with_trailing_text_returns_none() {
+        // "stop X and do Y" is an instruction, not an abort request.
+        let cmd = extract_stop_command("Timmy stop X and do Y", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, None);
+    }
+
+    #[test]
+    fn extract_stop_followed_by_words_returns_none() {
+        let cmd = extract_stop_command("stop the deployment please", "Timmy", "@timmy:home.local");
+        assert_eq!(cmd, None);
+    }
+}

server/src/llm/providers/claude_code/mod.rs

@@ -41,6 +41,12 @@ mod parse;
 
 use events::process_json_event;
 
+/// Sentinel error string `chat_stream` returns when a turn was cancelled via
+/// `cancel_rx` (e.g. a user-initiated chat "stop"), as opposed to a genuine
+/// Claude Code crash. Callers must check for this before running crash-
+/// recovery logic (e.g. clearing a stored session_id).
+pub const CANCELLED: &str = "Cancelled";
+
 /// Orchestrates Claude Code CLI sessions via a PTY for streaming agent chat.
 pub struct ClaudeCodeProvider;
 
@@ -354,7 +360,7 @@ fn run_pty_session(
             let _ = child.kill();
             let _ = child.wait();
             let _ = reader_handle.join();
-            return Err("Cancelled".to_string());
+            return Err(CANCELLED.to_string());
         }
 
         match line_rx.recv_timeout(std::time::Duration::from_millis(500)) {