diff --git a/navi/core/agent.py b/navi/core/agent.py
index cccc8d1..f2a2072 100644
--- a/navi/core/agent.py
+++ b/navi/core/agent.py
@@ -451,6 +451,7 @@
                 return
 
             accumulated_text = ""
+            accumulated_thinking = ""
             turn_tool_calls: list[ToolCallRequest] | None = None
             thinking_active = False
             context_tokens: int | None = None
@@ -490,6 +491,7 @@
                 if chunk.prompt_tokens is not None or chunk.completion_tokens is not None:
                     context_tokens = (chunk.prompt_tokens or 0) + (chunk.completion_tokens or 0)
                 if chunk.thinking:
+                    accumulated_thinking += chunk.thinking
                     if not thinking_active:
                         thinking_active = True
                     yield ThinkingDelta(delta=chunk.thinking)
@@ -519,6 +521,7 @@
             if not turn_tool_calls:
                 # Final response — text already streamed above
                 assistant_msg = Message(role="assistant", content=accumulated_text or None,
+                                        thinking=accumulated_thinking or None,
                                         created_at=datetime.now(timezone.utc))
                 session.messages.append(assistant_msg)
                 session.context.append(assistant_msg)
@@ -539,6 +542,7 @@
             assistant_msg = Message(
                 role="assistant",
                 content=accumulated_text or None,
+                thinking=accumulated_thinking or None,
                 tool_calls=turn_tool_calls,
             )
             session.messages.append(assistant_msg)
@@ -743,8 +747,10 @@
         if not _re.search(r"(TOOL:|AGENT:|→\s*SELF)", plan_text):
             log.warning("agent.planning_no_executors", hint="plan lacks TOOL/AGENT/SELF assignments")
 
-        # Inject plan so the main loop continues from it
+        # Inject plan into context so the main loop continues from it,
+        # and into messages (with is_plan flag) so the UI can render a plan card after reload.
         session.context.append(Message(role="assistant", content=plan_text))
+        session.messages.append(Message(role="assistant", content=plan_text, is_plan=True))
         log.debug("agent.plan_ready", phases=2, length=len(plan_text))
         return [PlanReady(plan=plan_text)]
 
diff --git a/navi/llm/base.py b/navi/llm/base.py
index 8e13c03..56fe714 100644
--- a/navi/llm/base.py
+++ b/navi/llm/base.py
@@ -42,6 +42,10 @@
     created_at: datetime | None = None
     # marks a compressed history block injected by the context compressor
     is_summary: bool = False
+    # reasoning text produced during this turn (not sent to LLM, display only)
+    thinking: str | None = None
+    # marks a planning block — stored in messages for UI rendering, not as LLM context
+    is_plan: bool = False
 
 
 class LLMResponse(BaseModel):