Dealing with bot failures

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

@@ -210,6 +210,17 @@ pub(in crate::chat::transport::matrix::bot) async fn handle_message(
 
         // Persist to disk so history survives server restarts.
         save_history(&ctx.services.project_root, &guard);
+    } else {
+        // Clear session_id on error so a poisoned session (e.g. pending
+        // tool call with no response) doesn't cause a crash loop on resume.
+        let mut guard = ctx.history.lock().await;
+        if let Some(conv) = guard.get_mut(&room_id)
+            && conv.session_id.is_some()
+        {
+            slog!("[matrix-bot] clearing session_id after error to prevent crash loop");
+            conv.session_id = None;
+            save_history(&ctx.services.project_root, &guard);
+        }
     }
 }
 

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

@@ -371,23 +371,38 @@ fn run_pty_session(
     // Wait briefly for Claude Code to flush its session transcript to disk.
     // The `result` event means the API response is done, but the process
     // still needs to write the conversation to the JSONL session file.
-    match child.try_wait() {
-        Ok(Some(_)) => {} // Already exited
+    let exit_status = match child.try_wait() {
+        Ok(Some(status)) => Some(status),
         _ => {
-            // Give it up to 2 seconds to exit cleanly
+            let mut status = None;
             for _ in 0..20 {
                 std::thread::sleep(std::time::Duration::from_millis(100));
-                if let Ok(Some(_)) = child.try_wait() {
+                if let Ok(Some(s)) = child.try_wait() {
+                    status = Some(s);
                     break;
                 }
             }
-            // If still running after 2s, kill it
-            let _ = child.kill();
-            let _ = child.wait();
+            if status.is_none() {
+                let _ = child.kill();
+                status = child.wait().ok();
+            }
+            status
         }
-    }
+    };
     // Wait for the reader thread to release the cloned PTY master fd.
     let _ = reader_handle.join();
+
+    // Non-zero exit without a result event means Claude Code crashed
+    // (e.g. MCP server not yet connected). Propagate as an error so
+    // the caller can clear the session_id instead of persisting it.
+    if !got_result
+        && let Some(ref status) = exit_status
+        && !status.success()
+    {
+        slog!("[pty-debug] Claude Code exited with non-zero status: {status}");
+        return Err(format!("Claude Code crashed (exit status: {status})"));
+    }
+
     Ok(())
 }