fix: streaming Activity loss during scrolling

api/models.py

@@ -3950,13 +3950,22 @@ def _session_message_merge_key(msg: dict):
     message_identity = msg.get("id") or msg.get("message_id")
     if message_identity:
         return ("message_id", str(message_identity))
+    # Include tool_calls so assistant messages that invoke different tools
+    # (but share identical empty content and same-second timestamp) are not
+    # collapsed by the merge-key guard at line ~4216.  Without this,
+    # all tool-calling messages map to the same legacy key and the
+    # timestamp<=max_sidecar_timestamp blanket-skip at line ~4218 drops
+    # every state.db tool-call after the first one registered by the sidecar.
+    _tc = msg.get("tool_calls")
+    _tc_key = json.dumps(_tc, sort_keys=True, default=str) if _tc else ""
     return (
         "legacy",
         str(msg.get("role") or ""),
         str(msg.get("content") or ""),
         _normalized_message_timestamp_for_key(msg.get("timestamp")),
         str(msg.get("tool_call_id") or ""),
         str(msg.get("tool_name") or msg.get("name") or ""),
+        _tc_key,
     )
 
 
@@ -3973,13 +3982,19 @@ def _session_message_dedup_key(msg: dict):
     message_identity = msg.get("id") or msg.get("message_id")
     if message_identity:
         return ("message_id", str(message_identity))
+    # Include tool_calls in the key so assistant messages that carry
+    # different tool invocations (but identical empty content/timestamp)
+    # are never collapsed into one.  (#3346 regression)
+    _tc = msg.get("tool_calls")
+    _tc_key = json.dumps(_tc, sort_keys=True, default=str) if _tc else ""
     return (
         "legacy",
         str(msg.get("role") or ""),
         str(msg.get("content") or ""),
         str(msg.get("timestamp") or ""),
         str(msg.get("tool_call_id") or ""),
         str(msg.get("tool_name") or msg.get("name") or ""),
+        _tc_key,
     )
 
 
@@ -4007,9 +4022,16 @@ def _session_message_content_key(msg: dict):
 def _session_message_visible_key(msg: dict):
     if not isinstance(msg, dict):
         return ("non_dict", repr(msg))
+    # Include tool_calls so assistant messages that invoke different tools
+    # (but share identical empty content) are not collapsed by sidecar
+    # prefix matching.  Without this, all tool-calling messages map to
+    # ("assistant", "") and the merge treats state.db rows as replays.
+    _tc = msg.get("tool_calls")
+    _tc_key = json.dumps(_tc, sort_keys=True, default=str) if _tc else ""
     return (
         str(msg.get("role") or ""),
         _normalized_session_message_content(msg),
+        _tc_key,
     )
 
 
@@ -4018,8 +4040,9 @@ def _build_visible_duplicate_lookup(visible_keys: set[tuple]) -> dict:
     loose_by_key = {}
     for key in visible_keys:
         try:
-            role, content = key
-        except (TypeError, ValueError):
+            role = key[0]
+            content = key[1]
+        except (TypeError, IndexError):
             continue
         if not content:
             continue
@@ -4031,24 +4054,27 @@ def _build_visible_duplicate_lookup(visible_keys: set[tuple]) -> dict:
 def _matching_visible_duplicate(visible_key: tuple, visible_keys: set[tuple], lookup: dict | None = None):
     if visible_key in visible_keys:
         return visible_key
-    role, content = visible_key
+    role = visible_key[0]
+    content = visible_key[1] if len(visible_key) > 1 else ""
     if not content:
         return None
     if lookup is None:
         lookup = _build_visible_duplicate_lookup(visible_keys)
     loose_content = None
-    for existing_role, existing_content in lookup.get("by_role", {}).get(role, []):
+    for existing_key in lookup.get("by_role", {}).get(role, []):
+        existing_role = existing_key[0]
+        existing_content = existing_key[1] if len(existing_key) > 1 else ""
         if role != existing_role or not existing_content:
             continue
         if content in existing_content or existing_content in content:
-            return (existing_role, existing_content)
+            return existing_key
         if loose_content is None:
             loose_content = _loose_session_message_content(content)
-        loose_existing = lookup.get("loose_by_key", {}).get((existing_role, existing_content), "")
+        loose_existing = lookup.get("loose_by_key", {}).get(existing_key, "")
         if loose_content and loose_existing and (
             loose_content in loose_existing or loose_existing in loose_content
         ):
-            return (existing_role, existing_content)
+            return existing_key
     return None
 
 
@@ -4250,7 +4276,13 @@ def merge_session_messages_append_only(
             if key in seen_message_keys and key[0] == "message_id":
                 continue
             if not (isinstance(key, tuple) and key[:1] == ("message_id",)):
-                continue
+                # Legacy key within sidecar timestamp range — only skip if
+                # this exact merge_key was already registered by the sidecar.
+                # Different tool_calls produce different merge_keys even with
+                # identical content/timestamp, so an unchecked continue here
+                # would drop legitimately distinct turns.  (#3346 / PR #3665)
+                if key in seen_message_keys:
+                    continue
         if key in seen_message_keys and key[0] == "message_id":
             continue
         matched_visible_key = _matching_visible_duplicate(
@@ -4280,7 +4312,20 @@ def merge_session_messages_append_only(
             and timestamp is not None
             and timestamp <= max_sidecar_timestamp
         ):
-            continue
+            # Legacy key within sidecar timestamp range.  Normally skip — the
+            # sidecar already has this message.  Exception: if the state.db
+            # message has tool_calls that DIFFER from the sidecar version
+            # (same content_key but different dedup_key because tool_calls
+            # differ), preserve it — distinct tool_calls must not be collapsed.
+            _tc = msg.get("tool_calls")
+            if _tc:
+                _ck = _session_message_content_key(msg)
+                if _ck in seen_content_keys and dedup_key not in seen_dedup_keys:
+                    pass  # different tool_calls from sidecar — preserve
+                else:
+                    continue
+            else:
+                continue
         seen_message_keys.add(key)
         seen_dedup_keys.add(dedup_key)
         seen_content_keys.add(_session_message_content_key(msg))

api/routes.py

@@ -5158,16 +5158,12 @@ def handle_get(handler, parsed) -> bool:
                         _threshold_tokens = 0
                     _persisted_cl = _fb_cl
             _session_tool_calls = getattr(s, "tool_calls", []) if load_messages else []
-            if (
-                load_messages
-                and msg_limit is not None
-                and _messages_include_tool_metadata(_truncated_msgs)
-            ):
-                # The browser ignores session-level tool_calls when the returned
-                # messages already carry per-message tool metadata. Avoid sending
-                # the full historical list with a small tail window.
-                _session_tool_calls = []
-            elif _windowed_messages:
+            # Always include session-level tool_calls so the browser can merge
+            # them with per-message tool_calls for messages that lack the
+            # per-message variant (older messages whose tool_calls live only
+            # in the session-level list).  The browser-side
+            # _syncToolCallsForLoadedMessages handles deduplication by tid.
+            if _windowed_messages:
                 _session_tool_calls = _tool_calls_for_message_window(
                     _session_tool_calls,
                     _messages_offset,

static/messages.js

@@ -2157,6 +2157,7 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
           S.session=d.session;S.messages=_carryForwardEphemeralTurnFields(S.messages||[], d.session.messages||[]);if(typeof _messagesTruncated!=='undefined')_messagesTruncated=!!d.session._messages_truncated;
           S.messages=_filterRecoveryControlMessages(S.messages || []);
           if(typeof _hydrateTodosFromSession==='function') _hydrateTodosFromSession(S.session);
+          if(typeof clearVisibleMessageRowCache==='function') clearVisibleMessageRowCache();
           if(S.session&&S.session.session_id){
             try{localStorage.setItem('hermes-webui-session',S.session.session_id);}catch(_){}
             if(typeof _setActiveSessionUrl==='function') _setActiveSessionUrl(S.session.session_id);
@@ -2259,6 +2260,16 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
           if(!S.messages.some(m=>m.role==='assistant'&&String(m.content||'').trim())&&!assistantText){removeThinking();S.messages.push({role:'assistant',content:'**No response received.** Check your API key and model selection.'});}
           if(_markerOnlyAssistantError&&typeof showToast==='function') showToast('No response received after context compression. Please retry.',5000,'error');
           if(isSessionViewed) _markSessionViewed(completedSid, completedSession.message_count ?? S.messages.length);
+          // Cooldown: prevent refreshActiveSessionIfExternallyUpdated from
+          // force-reloading immediately after "done" — the event already
+          // delivered the final messages and tool calls.
+          if(typeof window!=='undefined') window._streamJustFinished=true;
+          setTimeout(()=>{ if(typeof window!=='undefined') window._streamJustFinished=false; }, 5000);
+          // Expand render window to cover all messages so the done render
+          // doesn't hide Activity behind a tiny window (winSize=50).
+          if(typeof _messageRenderableMessageCount==='function'&&typeof _messageRenderWindowSize!=='undefined'){
+            _messageRenderWindowSize=Math.max(typeof _currentMessageRenderWindowSize==='function'?_currentMessageRenderWindowSize():50, _messageRenderableMessageCount());
+          }
           syncTopbar();renderMessages({preserveScroll:true});
           if(shouldFollowOnDone&&typeof scrollToBottom==='function') scrollToBottom();
           if(typeof noteWorkspaceMutationsFromToolCalls==='function') noteWorkspaceMutationsFromToolCalls(S.toolCalls);
@@ -2718,6 +2729,10 @@ function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
           S.toolCalls=[];
         }
         if(isSessionViewed) _markSessionViewed(completedSid, session.message_count ?? S.messages.length);
+        // Expand render window so the settled render doesn't hide Activity.
+        if(typeof _messageRenderableMessageCount==='function'&&typeof _messageRenderWindowSize!=='undefined'){
+          _messageRenderWindowSize=Math.max(typeof _currentMessageRenderWindowSize==='function'?_currentMessageRenderWindowSize():50, _messageRenderableMessageCount());
+        }
         syncTopbar();renderMessages({preserveScroll:true});
       }
       if(_isActiveSession()) _queueDrainSid=activeSid;

static/sessions.js

@@ -1543,6 +1543,9 @@ function _messageReloadLimitForSession(sid){
 
 function _syncToolCallsForLoadedMessages(messages, sessionToolCalls){
   const msgs=Array.isArray(messages)?messages:[];
+  // During active streaming, skip — clearing S.toolCalls would lose Activity
+  // and the renderMessages fallback is blocked by S.busy=true.
+  if(S.busy||S.activeStreamId) return;
   const hasMessageToolMetadata=msgs.some(m=>{
     if(!m) return false;
     const hasTc=Array.isArray(m.tool_calls)&&m.tool_calls.length>0;
@@ -1585,7 +1588,13 @@ async function _ensureMessagesLoaded(sid) {
   // toast on every mobile message (SSE/visibility events trigger this reload path
   // more aggressively on mobile).
   let msgs = (data.session.messages || []).filter(m => m && m.role);
-  _syncToolCallsForLoadedMessages(msgs, data.session.tool_calls);
+  // Skip _syncToolCalls when INFLIGHT exists — the INFLIGHT restore path
+  // (loadSession line ~871) will overwrite S.toolCalls from INFLIGHT[sid].toolCalls.
+  // Clearing here and then overwriting is wasteful, and if S.busy becomes true
+  // before the next render, the fallback can't re-derive from messages.
+  if(!(typeof INFLIGHT !== 'undefined' && INFLIGHT && INFLIGHT[sid])){
+    _syncToolCallsForLoadedMessages(msgs, data.session.tool_calls);
+  }
   clearLiveToolCards();
   // #3018: preserve client-side ephemeral turn fields (_turnUsage, _turnDuration,
   // _turnTps, _gatewayRouting, _statusCard) across the loadSession replace.
@@ -1601,6 +1610,11 @@ async function _ensureMessagesLoaded(sid) {
   }
   if(typeof clearVisibleMessageRowCache==='function') clearVisibleMessageRowCache();
   S.messages = msgs;
+  // Expand render window to cover all loaded messages so the next
+  // renderMessages() doesn't hide most of them behind a tiny window.
+  if(typeof _messageRenderableMessageCount==='function'&&typeof _currentMessageRenderWindowSize==='function'){
+    _messageRenderWindowSize=Math.max(_currentMessageRenderWindowSize(), _messageRenderableMessageCount());
+  }
   if(S.session&&S.session.session_id===sid){
     S.session.message_count=Number(data.session.message_count || msgs.length);
     S.lastUsage={...(data.session.last_usage||S.lastUsage||{})};
@@ -2965,6 +2979,10 @@ async function refreshActiveSessionIfExternallyUpdated(reason){
   if(_activeSessionExternalRefreshInFlight) return;
   if(!S.session || !S.session.session_id) return;
   if(S.busy || S.activeStreamId) return;
+  // Cooldown: don't force-reload immediately after streaming ends — the
+  // "done" event already delivered the final messages. Reloading here would
+  // clear S.toolCalls and lose Activity.
+  if(typeof window !== 'undefined' && window._streamJustFinished) return;
   if(typeof document !== 'undefined' && document.hidden) return;
   const sid = S.session.session_id;
   const localCount = Number(S.session.message_count || (Array.isArray(S.messages)?S.messages.length:0) || 0);

static/ui.js

@@ -307,9 +307,11 @@ let _messageRenderWindowSize=MESSAGE_RENDER_WINDOW_DEFAULT;
 // Cached visWithIdx array — invalidated when S.messages.length changes.
 let _visWithIdxCache=null;
 let _visWithIdxCacheLen=0;
+let _visWithIdxCacheSrc=null;  // S.messages reference — detects wholesale replacement with same length
 function clearVisibleMessageRowCache(){
   _visWithIdxCache=null;
   _visWithIdxCacheLen=0;
+  _visWithIdxCacheSrc=null;
 }
 function _resetMessageRenderWindow(sid){
   _messageRenderWindowSid=sid||null;
@@ -355,9 +357,10 @@ function _messageRenderableMessageCount(){
   for(const m of (S.messages||[])){
     if(!m||!m.role||m.role==='tool') continue;
     if(_isContextCompactionMessage(m)||_isPreservedCompressionTaskListMessage(m)) continue;
+    if(_isRecoveryControlMessage(m)) continue;
     const hasTc=Array.isArray(m.tool_calls)&&m.tool_calls.length>0;
     const hasTu=Array.isArray(m.content)&&m.content.some(p=>p&&p.type==='tool_use');
-    if(msgContent(m)||m.attachments?.length||(m.role==='assistant'&&(hasTc||hasTu||_messageHasReasoningPayload(m)))) count++;
+    if(msgContent(m)||m._statusCard||m.attachments?.length||(m.role==='assistant'&&(hasTc||hasTu||_messageHasReasoningPayload(m)||_assistantMessageHasVisibleContent(m)))) count++;
   }
   return count;
 }
@@ -416,9 +419,18 @@ async function jumpToSessionStart(){
   _messageUserUnpinned=true;
   _programmaticScroll=true;
   try{
-    if(typeof _ensureAllMessagesLoaded==='function') await _ensureAllMessagesLoaded();
+    // During active streaming, skip full message load — API response won't
+    // include live messages from the current turn, and replacing S.messages
+    // would lose user/assistant/tool messages.
+    if(!(S.busy||S.activeStreamId)){
+      if(typeof _ensureAllMessagesLoaded==='function') await _ensureAllMessagesLoaded();
+    }
     _messageRenderWindowSize=Math.max(_currentMessageRenderWindowSize(),_messageRenderableMessageCount());
-    renderMessages({ preserveScroll:true });
+    // During streaming, skip renderMessages — it rebuilds the DOM but tool card
+    // insertion is blocked by !S.busy, losing Activity until "done" fires.
+    if(!(S.busy||S.activeStreamId)){
+      renderMessages({ preserveScroll:true });
+    }
     requestAnimationFrame(()=>{
       container.scrollTop=0;
       _updateSessionStartJumpButton();
@@ -6965,7 +6977,7 @@ function renderMessages(options){
   // Cache visWithIdx so expanding the render window (Load earlier) doesn't
   // re-scan S.messages from scratch.  Invalidate only when the message array
   // length changes — i.e. new messages arrived or session was truncated.
-  if(!_visWithIdxCache || _visWithIdxCacheLen !== S.messages.length){
+  if(!_visWithIdxCache || _visWithIdxCacheLen !== S.messages.length || _visWithIdxCacheSrc !== S.messages){
     const rebuilt=[];
     let ri=0;
     for(const m of S.messages){
@@ -6980,6 +6992,7 @@ function renderMessages(options){
     }
     _visWithIdxCache=rebuilt;
     _visWithIdxCacheLen=S.messages.length;
+    _visWithIdxCacheSrc=S.messages;
   }
   const visWithIdx=_visWithIdxCache;
   const preservedCompressionRawIdxs=[];
@@ -7434,7 +7447,11 @@ function renderMessages(options){
     });
     if(derived.length) S.toolCalls=derived;
   }
-  if(!S.busy){
+  // Render tool cards: allow during streaming when S.toolCalls is already
+  // populated (e.g. from INFLIGHT restore or SSE events). Only the fallback
+  // derivation above is blocked by S.busy — DOM insertion should proceed
+  // whenever tool cards exist.
+  if(!S.busy || (S.toolCalls&&S.toolCalls.length)){
     inner.querySelectorAll('.tool-call-group:not([data-compression-card]),.tool-card-row:not([data-compression-card]),.agent-activity-thinking:not([data-live-thinking="1"])').forEach(el=>el.remove());
     const byAssistant = {};
     for(const tc of (S.toolCalls||[])){

tests/test_issue3306_loadsession_carry_forward.py

@@ -31,7 +31,7 @@ def _load_session_clear_block() -> str:
 
 def _ensure_messages_loaded_body() -> str:
     start = SESSIONS_JS.index("async function _ensureMessagesLoaded")
-    return SESSIONS_JS[start: start + 2500]
+    return SESSIONS_JS[start: start + 3000]
 
 
 def test_pending_carry_forward_snapshot_declared_at_module_scope():