diff --git a/backend/apps/agents/manager/Messaging.py b/backend/apps/agents/manager/Messaging.py
index fa67bfee..bd6cdd09 100644
--- a/backend/apps/agents/manager/Messaging.py
+++ b/backend/apps/agents/manager/Messaging.py
@@ -17,6 +17,7 @@ from backend.apps.agents.manager.session.session_store import load_session_data
 from backend.apps.agents.manager.session.apply_context_window import apply_context_window
 from backend.apps.agents.manager.prompt.tool_catalog import get_all_tool_names
 from backend.apps.agents.manager.prompt.prompt_context import resolve_mode
+from backend.apps.agents.manager.ttft_probe import ttft_probe
 
 logger = logging.getLogger(__name__)
 
@@ -44,6 +45,7 @@ class Messaging(AgentManagerProtocol):
         client_message_id: Optional[str] = None,
     ):
         """Send a follow-up message to an existing session."""
+        ttft_probe(session_id, "msg_received")
         session = self.sessions.get(session_id)
         if not session:
             data = load_session_data(session_id)
diff --git a/backend/apps/agents/manager/run/TurnRunner.py b/backend/apps/agents/manager/run/TurnRunner.py
index c457f97c..c69ce89b 100644
--- a/backend/apps/agents/manager/run/TurnRunner.py
+++ b/backend/apps/agents/manager/run/TurnRunner.py
@@ -113,6 +113,7 @@ class TurnRunner(AgentManagerProtocol):
                         message, session, session_id, turn, thinking, self.live_partial, self.sessions
                     )
                 elif isinstance(message, ResultMessage):
+                    ttft_probe(session_id, "result", chars=turn.assistant_text_chars)
                     await handle_result_message(
                         message, session, session_id, turn, thinking, self.sessions,
                         resolved_model, api_type, global_settings,
diff --git a/backend/apps/agents/manager/streaming/handle_stream_event.py b/backend/apps/agents/manager/streaming/handle_stream_event.py
index 17d5a626..9de30ea7 100644
--- a/backend/apps/agents/manager/streaming/handle_stream_event.py
+++ b/backend/apps/agents/manager/streaming/handle_stream_event.py
@@ -14,6 +14,7 @@ from backend.apps.agents.core.models import AgentSession
 from backend.apps.agents.core.ws_manager import ws_manager
 from backend.apps.agents.manager.streaming.state import ThinkingState, TurnState
 from backend.apps.agents.manager.streaming.PartialReply import PartialReply
+from backend.apps.agents.manager.ttft_probe import ttft_probe
 
 try:
     from claude_agent_sdk.types import StreamEvent
@@ -44,6 +45,7 @@ async def handle_stream_event(
 
         if block_type == "text":
             if turn.stream_text_msg_id is None:
+                ttft_probe(session_id, "first_text_block")
                 turn.stream_text_msg_id = uuid4().hex
                 await ws_manager.send_to_session(session_id, "agent:stream_start", {
                     "session_id": session_id,
@@ -54,6 +56,8 @@ async def handle_stream_event(
 
         elif block_type == "thinking":
             # Reasoning trace from thinking-capable models (GPT-5.3 Codex, Gemini 3 Pro/Flash, Claude with extended thinking). Rendered as a collapsible "thinking" message in the UI via the existing stream infrastructure, the frontend already handles role="thinking" for the DynamicIsland/agent card rendering.
+            if not thinking.block_starts:
+                ttft_probe(session_id, "first_thinking_block")
             thinking_msg_id = uuid4().hex
             turn.stream_block_index_map[index] = thinking_msg_id
             # Server-stamp start so we can accumulate per-turn elapsed_ms across multiple thinking blocks (think → tool → think → answer turns sum correctly).
@@ -85,6 +89,8 @@ async def handle_stream_event(
 
         if msg_id and delta_type == "text_delta":
             text_chunk = delta.get("text", "")
+            if turn.assistant_text_chars == 0:
+                ttft_probe(session_id, "first_text_delta", chars=len(text_chunk))
             turn.assistant_text_chars += len(text_chunk)
             turn.stream_text_accum += text_chunk
             live_partial[session_id] = PartialReply(