Add abort/stop feature and agentui CLI command

Adds cooperative abort via threading.Event so users can stop running agents.
Backend abort registry tracks parent-child relationships for recursive abort.
Frontend toggles SEND→STOP during generation, with AbortController on fetch.
Command center STOP aborts all child agents. Aborted agents produce a result
message so parent agents are aware. Also adds `agentui` CLI entry point.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/agent.py

@@ -93,7 +93,8 @@ def stream_agent_execution(
     model: str,
     messages: List[Dict],
     serper_key: str,
-    extra_params: Optional[Dict] = None
+    extra_params: Optional[Dict] = None,
+    abort_event=None
 ):
     """
     Run the agent tool-calling loop.
@@ -115,6 +116,11 @@ def stream_agent_execution(
     has_result = False
 
     while not done and turns < MAX_TURNS:
+        # Check abort before each turn
+        if abort_event and abort_event.is_set():
+            yield {"type": "aborted"}
+            return
+
         turns += 1
 
         # --- LLM call with retry ---
@@ -145,7 +151,13 @@ def stream_agent_execution(
                         "delay": delay,
                         "message": error_info["message"],
                     }
-                    time.sleep(delay)
+                    if abort_event:
+                        abort_event.wait(delay)
+                        if abort_event.is_set():
+                            yield {"type": "aborted"}
+                            return
+                    else:
+                        time.sleep(delay)
                 else:
                     yield {"type": "error", "content": error_info["message"]}
                     return
@@ -182,6 +194,11 @@ def stream_agent_execution(
         # --- Handle tool calls ---
         if tool_calls:
             for tool_call in tool_calls:
+                # Check abort between tool calls
+                if abort_event and abort_event.is_set():
+                    yield {"type": "aborted"}
+                    return
+
                 func_name = tool_call.function.name
 
                 # Parse arguments

backend/agents.py

@@ -79,7 +79,11 @@ AGENT_REGISTRY = {
             "- Do NOT ask agents to save or create files unless the user explicitly requests it.\n"
             "- NEVER overwrite existing files without explicit user permission.\n"
             "- Each agent has a task_id. Reuse the task_id when a new message relates to an existing agent "
-            "(this preserves context and the Jupyter kernel for code agents)."
+            "(this preserves context and the Jupyter kernel for code agents).\n\n"
+            "## Handling Aborted Agents\n\n"
+            "If an agent's result is 'Generation aborted by user.', the user deliberately stopped it. "
+            "Do NOT automatically re-launch the same task. Instead, briefly acknowledge the abort and "
+            "ask the user what went wrong or how they'd like to proceed."
         ),
         "tool": None,
         "tool_arg": None,

backend/code.py

@@ -204,7 +204,7 @@ def download_files_from_sandbox(sbx: Sandbox, files: List[Dict], files_root: str
     return "\n".join(results)
 
 
-def stream_code_execution(client, model: str, messages: List[Dict], sbx: Sandbox, files_root: str = None, extra_params: Optional[Dict] = None):
+def stream_code_execution(client, model: str, messages: List[Dict], sbx: Sandbox, files_root: str = None, extra_params: Optional[Dict] = None, abort_event=None):
     """
     Stream code execution results
 
@@ -226,6 +226,11 @@ def stream_code_execution(client, model: str, messages: List[Dict], sbx: Sandbox
     has_result = False
 
     while not done and turns < MAX_TURNS:
+        # Check abort before each turn
+        if abort_event and abort_event.is_set():
+            yield {"type": "aborted"}
+            return
+
         turns += 1
 
         # Retry loop for LLM calls
@@ -261,7 +266,13 @@ def stream_code_execution(client, model: str, messages: List[Dict], sbx: Sandbox
                         "error_type": error_info["type"]
                     }
                     import time
-                    time.sleep(delay)
+                    if abort_event:
+                        abort_event.wait(delay)
+                        if abort_event.is_set():
+                            yield {"type": "aborted"}
+                            return
+                    else:
+                        time.sleep(delay)
                 else:
                     # Final attempt failed or non-retryable error
                     yield {"type": "error", "content": error_info["message"]}
@@ -298,6 +309,11 @@ def stream_code_execution(client, model: str, messages: List[Dict], sbx: Sandbox
         # Handle tool calls
         if tool_calls:
             for tool_call in tool_calls:
+                # Check abort between tool calls
+                if abort_event and abort_event.is_set():
+                    yield {"type": "aborted"}
+                    return
+
                 if tool_call.function.name == "execute_code":
                     # Parse arguments
                     try:

backend/command.py

@@ -56,7 +56,7 @@ def parse_llm_error(error: Exception) -> dict:
     }
 
 
-def stream_command_center(client, model: str, messages: List[Dict], extra_params: dict = None):
+def stream_command_center(client, model: str, messages: List[Dict], extra_params: dict = None, abort_event=None):
     """
     Stream command center responses with agent launching capabilities
 
@@ -67,6 +67,11 @@ def stream_command_center(client, model: str, messages: List[Dict], extra_params
     done = False
 
     while not done and turns < MAX_TURNS:
+        # Check abort before each turn
+        if abort_event and abort_event.is_set():
+            yield {"type": "aborted"}
+            return
+
         turns += 1
 
         # Retry loop for LLM calls
@@ -105,7 +110,13 @@ def stream_command_center(client, model: str, messages: List[Dict], extra_params
                         "error_type": error_info["type"]
                     }
                     import time
-                    time.sleep(delay)
+                    if abort_event:
+                        abort_event.wait(delay)
+                        if abort_event.is_set():
+                            yield {"type": "aborted"}
+                            return
+                    else:
+                        time.sleep(delay)
                 else:
                     # Final attempt failed or non-retryable error
                     yield {"type": "error", "content": error_info["message"]}
@@ -127,6 +138,11 @@ def stream_command_center(client, model: str, messages: List[Dict], extra_params
         # Handle tool calls (agent launches)
         if tool_calls:
             for tool_call in tool_calls:
+                # Check abort between tool calls
+                if abort_event and abort_event.is_set():
+                    yield {"type": "aborted"}
+                    return
+
                 function_name = tool_call.function.name
 
                 # Parse arguments

backend/image.py

@@ -194,7 +194,8 @@ def stream_image_execution(
     hf_token: str,
     image_gen_model: Optional[str] = None,
     image_edit_model: Optional[str] = None,
-    extra_params: Optional[Dict] = None
+    extra_params: Optional[Dict] = None,
+    abort_event=None
 ):
     """
     Run the image agent tool-calling loop.
@@ -218,6 +219,11 @@ def stream_image_execution(
     result_sent = False
 
     while not done and turns < MAX_TURNS:
+        # Check abort before each turn
+        if abort_event and abort_event.is_set():
+            yield {"type": "aborted"}
+            return
+
         turns += 1
 
         # --- LLM call with retry ---
@@ -248,7 +254,13 @@ def stream_image_execution(
                         "delay": delay,
                         "message": error_info["message"],
                     }
-                    time.sleep(delay)
+                    if abort_event:
+                        abort_event.wait(delay)
+                        if abort_event.is_set():
+                            yield {"type": "aborted"}
+                            return
+                    else:
+                        time.sleep(delay)
                 else:
                     yield {"type": "error", "content": error_info["message"]}
                     return
@@ -286,6 +298,11 @@ def stream_image_execution(
         # --- Handle tool calls ---
         if tool_calls:
             for tool_call in tool_calls:
+                # Check abort between tool calls
+                if abort_event and abort_event.is_set():
+                    yield {"type": "aborted"}
+                    return
+
                 func_name = tool_call.function.name
 
                 # Parse arguments

backend/main.py

@@ -11,6 +11,7 @@ import logging
 import asyncio
 import signal
 import sys
+import threading
 from datetime import datetime
 from concurrent.futures import ThreadPoolExecutor
 from .agents import AGENT_REGISTRY, get_default_counters, get_registry_for_frontend
@@ -34,6 +35,39 @@ def signal_handler(signum, frame):
 signal.signal(signal.SIGINT, signal_handler)
 signal.signal(signal.SIGTERM, signal_handler)
 
+# ============================================
+# Abort Registry
+# ============================================
+# Per-agent abort flags for cooperative cancellation
+_abort_flags: Dict[str, threading.Event] = {}
+_agent_children: Dict[str, set] = {}  # parent_id -> set of child_ids
+
+def register_agent(agent_id: str, parent_id: str = None) -> threading.Event:
+    """Register an agent and return its abort event."""
+    event = threading.Event()
+    _abort_flags[agent_id] = event
+    if parent_id:
+        _agent_children.setdefault(parent_id, set()).add(agent_id)
+    return event
+
+def unregister_agent(agent_id: str):
+    """Clean up abort flag and parent-child mappings."""
+    _abort_flags.pop(agent_id, None)
+    # Remove from parent's children set
+    for children in _agent_children.values():
+        children.discard(agent_id)
+    _agent_children.pop(agent_id, None)
+
+def abort_agent_tree(agent_id: str) -> list:
+    """Recursively abort an agent and all its children. Returns list of aborted IDs."""
+    aborted = []
+    if agent_id in _abort_flags:
+        _abort_flags[agent_id].set()
+        aborted.append(agent_id)
+    for child_id in list(_agent_children.get(agent_id, [])):
+        aborted.extend(abort_agent_tree(child_id))
+    return aborted
+
 # Configure logging to match uvicorn's format
 logging.basicConfig(
     level=logging.INFO,
@@ -213,9 +247,14 @@ class ChatRequest(BaseModel):
     research_parallel_workers: Optional[int] = None  # Number of parallel workers for research
     research_max_websites: Optional[int] = None  # Max websites to analyze per research session
     agent_id: Optional[str] = None  # Unique agent/tab ID for session management
+    parent_agent_id: Optional[str] = None  # Parent agent ID for abort propagation
     frontend_context: Optional[FrontendContext] = None  # Dynamic context from frontend
 
 
+class AbortRequest(BaseModel):
+    agent_id: str
+
+
 class TitleRequest(BaseModel):
     query: str
     endpoint: str  # User's configured LLM endpoint
@@ -240,11 +279,19 @@ async def stream_code_agent(
     e2b_key: str,
     session_id: str,
     tab_id: str = "default",
+    parent_agent_id: Optional[str] = None,
     frontend_context: Optional[Dict] = None,
     extra_params: Optional[Dict] = None
 ):
     """Handle code agent with execution capabilities"""
+    abort_event = register_agent(tab_id, parent_agent_id)
+    try:
+        async for chunk in _stream_code_agent_inner(messages, endpoint, token, model, e2b_key, session_id, tab_id, frontend_context, extra_params, abort_event):
+            yield chunk
+    finally:
+        unregister_agent(tab_id)
 
+async def _stream_code_agent_inner(messages, endpoint, token, model, e2b_key, session_id, tab_id, frontend_context, extra_params, abort_event):
     if not E2B_AVAILABLE:
         yield f"data: {json.dumps({'type': 'error', 'content': 'E2B not available. Install with: pip install e2b-code-interpreter'})}\n\n"
         return
@@ -279,7 +326,7 @@ async def stream_code_agent(
 
         def run_sync_generator():
             try:
-                for update in stream_code_execution(client, model, full_messages, sbx, files_root=FILES_ROOT, extra_params=extra_params):
+                for update in stream_code_execution(client, model, full_messages, sbx, files_root=FILES_ROOT, extra_params=extra_params, abort_event=abort_event):
                     loop.call_soon_threadsafe(queue.put_nowait, update)
             finally:
                 loop.call_soon_threadsafe(queue.put_nowait, None)  # Signal completion
@@ -329,7 +376,7 @@ async def stream_code_agent(
 
                 def run_retry_generator():
                     try:
-                        for update in stream_code_execution(client, model, full_messages, sbx, files_root=FILES_ROOT, extra_params=extra_params):
+                        for update in stream_code_execution(client, model, full_messages, sbx, files_root=FILES_ROOT, extra_params=extra_params, abort_event=abort_event):
                             loop.call_soon_threadsafe(retry_queue.put_nowait, update)
                     finally:
                         loop.call_soon_threadsafe(retry_queue.put_nowait, None)
@@ -360,13 +407,21 @@ async def stream_research_agent(
     parallel_workers: Optional[int] = None,
     max_websites: Optional[int] = None,
     tab_id: str = "default",
+    parent_agent_id: Optional[str] = None,
     sub_agent_endpoint: Optional[str] = None,
     sub_agent_token: Optional[str] = None,
     extra_params: Optional[Dict] = None,
     sub_agent_extra_params: Optional[Dict] = None
 ):
     """Handle research agent with web search"""
+    abort_event = register_agent(tab_id, parent_agent_id)
+    try:
+        async for chunk in _stream_research_agent_inner(messages, endpoint, token, model, serper_key, sub_agent_model, parallel_workers, max_websites, tab_id, sub_agent_endpoint, sub_agent_token, extra_params, sub_agent_extra_params, abort_event):
+            yield chunk
+    finally:
+        unregister_agent(tab_id)
 
+async def _stream_research_agent_inner(messages, endpoint, token, model, serper_key, sub_agent_model, parallel_workers, max_websites, tab_id, sub_agent_endpoint, sub_agent_token, extra_params, sub_agent_extra_params, abort_event):
     if not RESEARCH_AVAILABLE:
         yield f"data: {json.dumps({'type': 'error', 'content': 'Research dependencies not available. Install with: pip install trafilatura requests'})}\n\n"
         return
@@ -413,7 +468,7 @@ async def stream_research_agent(
 
         def run_sync_generator():
             try:
-                for update in stream_research(client, model, question, serper_key, max_websites=max_sites, system_prompt=system_prompt, sub_agent_model=analysis_model, parallel_workers=workers, sub_agent_client=sub_agent_client, extra_params=extra_params, sub_agent_extra_params=sub_agent_extra_params):
+                for update in stream_research(client, model, question, serper_key, max_websites=max_sites, system_prompt=system_prompt, sub_agent_model=analysis_model, parallel_workers=workers, sub_agent_client=sub_agent_client, extra_params=extra_params, sub_agent_extra_params=sub_agent_extra_params, abort_event=abort_event):
                     loop.call_soon_threadsafe(queue.put_nowait, update)
             finally:
                 loop.call_soon_threadsafe(queue.put_nowait, None)  # Signal completion
@@ -447,7 +502,14 @@ async def stream_command_center_handler(
     extra_params: Optional[Dict] = None
 ):
     """Handle command center with tool-based agent launching"""
+    abort_event = register_agent(tab_id)
+    try:
+        async for chunk in _stream_command_center_inner(messages, endpoint, token, model, tab_id, extra_params, abort_event):
+            yield chunk
+    finally:
+        unregister_agent(tab_id)
 
+async def _stream_command_center_inner(messages, endpoint, token, model, tab_id, extra_params, abort_event):
     if not COMMAND_AVAILABLE:
         # Fallback to regular chat if command tools not available
         async for chunk in stream_chat_response(messages, endpoint, token, model, "command", tab_id, extra_params):
@@ -474,7 +536,7 @@ async def stream_command_center_handler(
 
         def run_sync_generator():
             try:
-                for update in stream_command_center(client, model, full_messages, extra_params=extra_params):
+                for update in stream_command_center(client, model, full_messages, extra_params=extra_params, abort_event=abort_event):
                     loop.call_soon_threadsafe(queue.put_nowait, update)
             finally:
                 loop.call_soon_threadsafe(queue.put_nowait, None)  # Signal completion
@@ -506,10 +568,18 @@ async def stream_web_agent(
     model: str,
     serper_key: str,
     tab_id: str = "default",
+    parent_agent_id: Optional[str] = None,
     extra_params: Optional[Dict] = None
 ):
     """Handle web agent with tools (search, read, screenshot)"""
+    abort_event = register_agent(tab_id, parent_agent_id)
+    try:
+        async for chunk in _stream_web_agent_inner(messages, endpoint, token, model, serper_key, tab_id, extra_params, abort_event):
+            yield chunk
+    finally:
+        unregister_agent(tab_id)
 
+async def _stream_web_agent_inner(messages, endpoint, token, model, serper_key, tab_id, extra_params, abort_event):
     if not AGENT_AVAILABLE:
         async for chunk in stream_chat_response(messages, endpoint, token, model, "agent", tab_id, extra_params):
             yield chunk
@@ -528,7 +598,7 @@ async def stream_web_agent(
 
         def run_sync_generator():
             try:
-                for update in stream_agent_execution(client, model, full_messages, serper_key, extra_params=extra_params):
+                for update in stream_agent_execution(client, model, full_messages, serper_key, extra_params=extra_params, abort_event=abort_event):
                     loop.call_soon_threadsafe(queue.put_nowait, update)
             finally:
                 loop.call_soon_threadsafe(queue.put_nowait, None)
@@ -559,10 +629,18 @@ async def stream_image_agent(
     image_gen_model: Optional[str] = None,
     image_edit_model: Optional[str] = None,
     tab_id: str = "default",
+    parent_agent_id: Optional[str] = None,
     extra_params: Optional[Dict] = None
 ):
     """Handle image agent with HuggingFace image generation tools"""
+    abort_event = register_agent(tab_id, parent_agent_id)
+    try:
+        async for chunk in _stream_image_agent_inner(messages, endpoint, token, model, hf_token, image_gen_model, image_edit_model, tab_id, extra_params, abort_event):
+            yield chunk
+    finally:
+        unregister_agent(tab_id)
 
+async def _stream_image_agent_inner(messages, endpoint, token, model, hf_token, image_gen_model, image_edit_model, tab_id, extra_params, abort_event):
     if not IMAGE_AVAILABLE:
         yield f"data: {json.dumps({'type': 'error', 'content': 'Image agent not available. Install with: pip install huggingface_hub Pillow'})}\n\n"
         return
@@ -584,7 +662,7 @@ async def stream_image_agent(
 
         def run_sync_generator():
             try:
-                for update in stream_image_execution(client, model, full_messages, hf_token, image_gen_model=image_gen_model, image_edit_model=image_edit_model, extra_params=extra_params):
+                for update in stream_image_execution(client, model, full_messages, hf_token, image_gen_model=image_gen_model, image_edit_model=image_edit_model, extra_params=extra_params, abort_event=abort_event):
                     loop.call_soon_threadsafe(queue.put_nowait, update)
             finally:
                 loop.call_soon_threadsafe(queue.put_nowait, None)
@@ -775,6 +853,14 @@ async def generate_title(request: TitleRequest):
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@app.post("/api/abort")
+async def abort_agent(request: AbortRequest):
+    """Abort a running agent and all its children."""
+    aborted = abort_agent_tree(request.agent_id)
+    logger.info(f"Aborted agents: {aborted}")
+    return {"aborted": aborted}
+
+
 @app.post("/api/chat/stream")
 async def chat_stream(request: ChatRequest):
     """Proxy streaming chat to user's configured LLM endpoint"""
@@ -819,6 +905,7 @@ async def chat_stream(request: ChatRequest):
                 e2b_key or "",
                 session_id,
                 tab_id,
+                request.parent_agent_id,
                 frontend_context,
                 request.extra_params
             ),
@@ -847,6 +934,7 @@ async def chat_stream(request: ChatRequest):
                 request.research_parallel_workers,
                 None,
                 tab_id,
+                request.parent_agent_id,
                 sub_agent_endpoint,
                 sub_agent_token,
                 request.extra_params,
@@ -872,6 +960,7 @@ async def chat_stream(request: ChatRequest):
                 request.image_gen_model,
                 request.image_edit_model,
                 tab_id,
+                request.parent_agent_id,
                 request.extra_params
             ),
             media_type="text/event-stream",
@@ -892,6 +981,7 @@ async def chat_stream(request: ChatRequest):
                 request.model or "gpt-4",
                 serper_key or "",
                 tab_id,
+                request.parent_agent_id,
                 request.extra_params
             ),
             media_type="text/event-stream",

backend/research.py

@@ -233,6 +233,7 @@ def stream_research(
     sub_agent_model: Optional[str] = None,
     parallel_workers: int = 8,
     max_tool_calls: int = 20,
+    abort_event=None,
     **kwargs
 ):
     """
@@ -264,6 +265,12 @@ def stream_research(
     iterations_without_tool_calls = 0
 
     while tool_call_count < max_tool_calls:
+        # Check abort before each iteration
+        if abort_event and abort_event.is_set():
+            yield {"type": "aborted"}
+            yield {"type": "done"}
+            return
+
         iteration += 1
 
         # Call DR-TULU
@@ -348,6 +355,12 @@ def stream_research(
         search_idx = 0  # Track which search query we're on within this batch
 
         for i, tc in enumerate(tool_calls):
+            # Check abort between tool executions
+            if abort_event and abort_event.is_set():
+                yield {"type": "aborted"}
+                yield {"type": "done"}
+                return
+
             tool_call_count += 1
 
             if tc["name"] == "google_search":

frontend/index.html

@@ -7,7 +7,7 @@
     <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@300;400;500;700&display=swap" rel="stylesheet">
     <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/prism/1.29.0/themes/prism.min.css">
     <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.9/dist/katex.min.css">
-    <link rel="stylesheet" href="style.css?v=89">
+    <link rel="stylesheet" href="style.css?v=90">
 </head>
 <body>
     <div class="app-container">
@@ -479,6 +479,6 @@
     <script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
     <script src="https://cdn.jsdelivr.net/npm/katex@0.16.9/dist/katex.min.js"></script>
     <script src="research-ui.js?v=23"></script>
-    <script src="script.js?v=75"></script>
+    <script src="script.js?v=79"></script>
 </body>
 </html>

frontend/script.js

@@ -92,6 +92,9 @@ let agentCounters = getDefaultCounters();
 // Debounce timer for workspace saving
 let saveWorkspaceTimer = null;
 
+// Abort controllers for in-flight fetch requests (tabId -> AbortController)
+const activeAbortControllers = {};
+
 // Timeline data structure for sidebar
 // Maps tabId -> { type, title, events: [{type: 'user'|'assistant'|'agent', content, childTabId?}], parentTabId?, isGenerating }
 const timelineData = {
@@ -482,7 +485,7 @@ function renderAgentTimeline(tabId, notebook, isNested = false) {
                                         ${childIsGenerating ? `
                                             <div class="agent-progress"><span></span><span></span><span></span></div>
                                         ` : `
-                                            <div class="agent-done"></div>
+                                            <div class="agent-done${childNotebook.aborted ? ' aborted' : ''}"></div>
                                         `}
                                     </div>
                                 </div>
@@ -1516,7 +1519,62 @@ function scrollToTimelineEvent(tabId, eventIndex) {
     });
 }
 
+async function abortAgent(tabId) {
+    // Abort the frontend fetch
+    const controller = activeAbortControllers[tabId];
+    if (controller) {
+        controller.abort();
+    }
+
+    // For command center (tab 0): also abort all generating child tabs
+    if (tabId === 0) {
+        for (const [childId, td] of Object.entries(timelineData)) {
+            if (td.parentTabId === 0 && td.isGenerating) {
+                // Abort frontend fetch
+                const childController = activeAbortControllers[childId];
+                if (childController) {
+                    childController.abort();
+                }
+                // Abort backend agent
+                try {
+                    await fetch('/api/abort', {
+                        method: 'POST',
+                        headers: { 'Content-Type': 'application/json' },
+                        body: JSON.stringify({ agent_id: childId.toString() })
+                    });
+                } catch (e) { /* ignore */ }
+            }
+        }
+        // Unblock command center input and restore SEND button
+        commandInputBlocked = false;
+        pendingAgentLaunches = 0;
+        setCommandCenterStopState(false);
+        const commandInput = document.getElementById('input-command');
+        if (commandInput) {
+            commandInput.disabled = false;
+        }
+    }
+
+    // Tell the backend to set the abort flag for this agent
+    try {
+        await fetch('/api/abort', {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ agent_id: tabId.toString() })
+        });
+    } catch (e) {
+        // Ignore abort endpoint errors
+    }
+}
+
 async function sendMessage(tabId) {
+    // If tab is currently generating, abort instead of sending
+    // For command center: also abort if input is blocked (sub-agents still running)
+    if (timelineData[tabId]?.isGenerating || (tabId === 0 && commandInputBlocked)) {
+        abortAgent(tabId);
+        return;
+    }
+
     const content = document.querySelector(`[data-content-id="${tabId}"]`);
     if (!content) return;
 
@@ -1587,6 +1645,9 @@ async function sendMessage(tabId) {
         );
         if (anyChildGenerating || pendingAgentLaunches > 0) {
             commandInputBlocked = true;
+            // Keep STOP button visible and input disabled while children run
+            input.disabled = true;
+            setCommandCenterStopState(true);
         } else {
             input.disabled = false;
             input.focus();
@@ -1600,6 +1661,21 @@ async function sendMessage(tabId) {
     saveWorkspaceDebounced();
 }
 
+function setCommandCenterStopState(showStop) {
+    const content = document.querySelector('[data-content-id="0"]');
+    if (!content) return;
+    const sendBtn = content.querySelector('.input-container button');
+    if (!sendBtn) return;
+    if (showStop) {
+        sendBtn.textContent = 'STOP';
+        sendBtn.classList.add('stop-btn');
+        sendBtn.disabled = false;
+    } else {
+        sendBtn.textContent = 'SEND';
+        sendBtn.classList.remove('stop-btn');
+    }
+}
+
 async function continueCommandCenter() {
     // Called when all launched agents finish — re-run command center with actual results in history
     const chatContainer = document.getElementById('messages-command');
@@ -1619,6 +1695,8 @@ async function continueCommandCenter() {
     );
     if (anyChildGenerating || pendingAgentLaunches > 0) {
         commandInputBlocked = true;
+        if (commandInput) commandInput.disabled = true;
+        setCommandCenterStopState(true);
     } else if (commandInput) {
         commandInput.disabled = false;
         commandInput.focus();
@@ -1802,10 +1880,22 @@ async function streamChatResponse(messages, chatContainer, agentType, tabId) {
         ? currentSettings.models?.[currentSettings.imageEditModel]?.modelId || null
         : null;
 
+    // Set up AbortController for this request
+    const abortController = new AbortController();
+    if (tabId !== undefined) {
+        activeAbortControllers[tabId] = abortController;
+    }
+
     try {
+        // Determine parent agent ID for abort propagation
+        const parentAgentId = timelineData[tabId]?.parentTabId != null
+            ? timelineData[tabId].parentTabId.toString()
+            : null;
+
         const response = await fetch(`${backendEndpoint}/chat/stream`, {
             method: 'POST',
             headers: { 'Content-Type': 'application/json' },
+            signal: abortController.signal,
             body: JSON.stringify({
                 messages: messages,
                 agent_type: agentType,
@@ -1826,6 +1916,7 @@ async function streamChatResponse(messages, chatContainer, agentType, tabId) {
                 research_parallel_workers: currentSettings.researchParallelWorkers || null,
                 research_max_websites: currentSettings.researchMaxWebsites || null,
                 agent_id: tabId.toString(),  // Send unique tab ID for sandbox sessions
+                parent_agent_id: parentAgentId,  // Parent agent ID for abort propagation
                 frontend_context: getFrontendContext()  // Dynamic context for system prompts
             })
         });
@@ -2225,6 +2316,29 @@ async function streamChatResponse(messages, chatContainer, agentType, tabId) {
                         // Reset current message element so any subsequent thinking starts fresh
                         currentMessageEl = null;
 
+                    } else if (data.type === 'aborted') {
+                        // Agent was aborted by user
+                        hideProgressWidget(chatContainer);
+                        removeRetryIndicator(chatContainer);
+
+                        // Mark the action widget as aborted (show × instead of ✓)
+                        const widget = actionWidgets[tabId];
+                        if (widget) {
+                            const orbitIndicator = widget.querySelector('.orbit-indicator');
+                            if (orbitIndicator) {
+                                const abortedIndicator = document.createElement('div');
+                                abortedIndicator.className = 'done-indicator aborted';
+                                orbitIndicator.replaceWith(abortedIndicator);
+                            }
+                        }
+
+                        // For timeline agent boxes, mark as aborted
+                        if (timelineData[tabId]) {
+                            timelineData[tabId].aborted = true;
+                        }
+
+                        break;  // Stop reading stream
+
                     } else if (data.type === 'done') {
                         // Remove retry indicator on success
                         removeRetryIndicator(chatContainer);
@@ -2299,13 +2413,45 @@ async function streamChatResponse(messages, chatContainer, agentType, tabId) {
         }
     } catch (error) {
         hideProgressWidget(chatContainer);
-        const errorDiv = document.createElement('div');
-        errorDiv.className = 'message assistant';
-        errorDiv.innerHTML = `<div class="message-content" style="color: #c62828;">Connection error: ${escapeHtml(error.message)}</div>`;
-        chatContainer.appendChild(errorDiv);
+        if (error.name === 'AbortError') {
+            // User-initiated abort — show as a result block
+            const abortResultText = 'Generation aborted by user.';
+            const resultDiv = document.createElement('div');
+            resultDiv.className = 'research-report';
+            resultDiv.innerHTML = `
+                <div class="report-header">Result</div>
+                <div class="report-content"><p>${abortResultText}</p></div>
+            `;
+            chatContainer.appendChild(resultDiv);
+
+            // Send abort result to parent action widget (so command center knows it was aborted)
+            updateActionWidgetWithResult(tabId, abortResultText, {}, {});
+
+            // Override the done indicator to show × instead of ✓
+            const widget = actionWidgets[tabId];
+            if (widget) {
+                const doneIndicator = widget.querySelector('.done-indicator');
+                if (doneIndicator) {
+                    doneIndicator.classList.add('aborted');
+                }
+            }
+
+            // Mark timeline data as aborted
+            if (timelineData[tabId]) {
+                timelineData[tabId].aborted = true;
+            }
+        } else {
+            const errorDiv = document.createElement('div');
+            errorDiv.className = 'message assistant';
+            errorDiv.innerHTML = `<div class="message-content" style="color: #c62828;">Connection error: ${escapeHtml(error.message)}</div>`;
+            chatContainer.appendChild(errorDiv);
+        }
         if (tabId) {
             setTabGenerating(tabId, false);
         }
+    } finally {
+        // Clean up abort controller
+        delete activeAbortControllers[tabId];
     }
 }
 
@@ -2697,6 +2843,22 @@ function setTabGenerating(tabId, isGenerating) {
         }, 300);
     }
 
+    // Toggle SEND/STOP button
+    const content = document.querySelector(`[data-content-id="${tabId}"]`);
+    if (content) {
+        const sendBtn = content.querySelector('.input-container button');
+        if (sendBtn) {
+            if (isGenerating) {
+                sendBtn.textContent = 'STOP';
+                sendBtn.classList.add('stop-btn');
+                sendBtn.disabled = false;  // Keep enabled so user can click STOP
+            } else {
+                sendBtn.textContent = 'SEND';
+                sendBtn.classList.remove('stop-btn');
+            }
+        }
+    }
+
     // Update timeline to reflect generating state
     setTimelineGenerating(tabId, isGenerating);
 
@@ -2712,6 +2874,7 @@ function setTabGenerating(tabId, isGenerating) {
         );
         if (!anyStillGenerating && pendingAgentLaunches === 0) {
             commandInputBlocked = false;
+            setCommandCenterStopState(false);
             // Auto-continue: call command center again with agent results now in history
             continueCommandCenter();
         }

frontend/style.css

@@ -765,6 +765,18 @@ body {
     transform: rotate(45deg) translate(-1px, -1px);
 }
 
+/* Aborted agent - cross instead of checkmark */
+.agent-done.aborted::before {
+    content: '×';
+    width: auto;
+    height: auto;
+    border: none;
+    transform: none;
+    font-size: 10px;
+    line-height: 1;
+    color: var(--text-secondary);
+}
+
 /* Simple turn - natural height from label, min-height fallback */
 .tl-row.turn {
     min-height: 12px;
@@ -1772,6 +1784,18 @@ pre code [class*="token"] {
     transform: rotate(45deg) translate(-1px, -1px);
 }
 
+/* Aborted action widget - cross instead of checkmark */
+.action-widget .done-indicator.aborted::before {
+    content: '×';
+    width: auto;
+    height: auto;
+    border: none;
+    transform: none;
+    font-size: 12px;
+    line-height: 1;
+    color: var(--bg-primary);
+}
+
 .action-widget-body {
     padding: 12px;
     background: var(--bg-tertiary);

pyproject.toml

@@ -16,6 +16,9 @@ dependencies = [
     "markdownify",
 ]
 
+[project.scripts]
+agentui = "backend.main:start"
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"