fix: live reasoning, tool progress, in-flight session recovery (#367)

* fix: preserve live session output across chat switches

(cherry picked from commit 401e3b643d25e8dad8c06883b478b3c3073f07a5)

* fix: preserve todo state after session reload

(cherry picked from commit 7ee093ba19978af23b79148df2f2347e2f1e5bde)

* fix: preserve live assistant anchor across rerenders

* fix: stream live reasoning and tool progress

* fix: recover inflight session state after reload

* fix: add loadInflightState stub + CHANGELOG v0.50.21

- static/ui.js: add loadInflightState() function (currently returns null —
  the typeof guard in sessions.js means reload recovery works via the
  else-path attachLiveStream call; this stub satisfies the guard cleanly
  and documents the extension point for future localStorage-backed state)
- CHANGELOG.md: v0.50.21 entry; 960 tests (up from 949)

---------

Co-authored-by: Jordan SkyLF <jordan@skylinkfiber.net>
Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>

CHANGELOG.md

@@ -10,6 +10,17 @@
 - **Workspace file downloads no longer crash for Unicode filenames** (`api/routes.py`): Clicking a PDF or other file with Chinese, Japanese, Arabic, or other non-ASCII characters in its name caused a `UnicodeEncodeError` because Python's HTTP server requires header values to be latin-1 encodable. A new `_content_disposition_value(disposition, filename)` helper centralises `Content-Disposition` generation: it strips CR/LF (injection guard), builds an ASCII fallback for the legacy `filename=` parameter (non-ASCII chars replaced with `_`), and preserves the full UTF-8 name in `filename*=UTF-8''...` per RFC 5987. Both `attachment` and `inline` responses use it.
   - 2 new integration tests in `tests/test_sprint29.py` covering Chinese filenames for both download and inline responses, verifying the header is latin-1 encodable and `filename*=UTF-8''` is present; 924 tests total (up from 922)
 
+## [v0.50.21] Live reasoning, tool progress, and in-flight session recovery (PR #367)
+
+- **Live reasoning cards during streaming** (`static/ui.js`, `static/messages.js`): The generic thinking spinner now upgrades to a live reasoning card when the backend streams reasoning text. `_thinkingMarkup(text)` and `updateThinking(text)` centralize the markup so the spinner and card share the same DOM slot. Works with models that emit reasoning via the agent's `reasoning_callback` or `tool_progress_callback`.
+- **`tool_complete` SSE events** (`api/streaming.py`, `static/messages.js`): Tool progress callback now accepts the current agent signature `on_tool(*cb_args, **cb_kwargs)` — handles both the old 3-arg `(name, preview, args)` form and the new 4-arg `(event_type, name, preview, args)` form. `tool.completed` events transition live tool cards from running to done cleanly.
+- **In-flight session state stable across switches** (`static/messages.js`, `static/sessions.js`): `attachLiveStream` refactored out of `send()` into a standalone function; partial assistant text mirrored into `INFLIGHT` state on every token; `data-live-assistant` DOM anchor preserved across `renderMessages()` calls so switching away and back doesn't lose or duplicate live output.
+- **Reload recovery** (`api/models.py`, `api/routes.py`, `api/streaming.py`, `static/sessions.js`): `active_stream_id`, `pending_user_message`, `pending_attachments`, and `pending_started_at` now persisted on the session object before streaming starts and cleared on completion (or exception). `/api/session` returns these fields. After a page reload or session switch, `loadSession()` detects `active_stream_id` and calls `attachLiveStream(..., {reconnecting:true})` to reattach to the live SSE stream.
+- **Session-scoped message queue** (`static/ui.js`, `static/messages.js`): Global `MSG_QUEUE` replaced with `SESSION_QUEUES` keyed by session ID. Queued follow-up messages are associated with the session they were typed in and only drained when that session becomes idle — no cross-session bleed.
+- **`newSession()` idle reset** (`static/sessions.js`): Sets `S.busy=false`, `S.activeStreamId=null`, clears the cancel button, resets composer status — ensures a fresh chat is immediately usable even if another session's stream is still running.
+- **Todos survive session reload** (`static/panels.js`): `loadTodos()` now reads from `S.session.messages` (raw, includes tool-role messages) rather than `S.messages` (filtered display), so todo state reconstructed from tool outputs survives reloads.
+  - 12 new regression tests in `tests/test_regressions.py`; 960 tests total (up from 949)
+
 ## [v0.50.20] Silent error fix, stale model cleanup, live model fetching (fixes #373, #374, #375)
 
 ### Fix: Chat no longer silently swallows agent failures (fixes #373)

api/models.py

@@ -44,6 +44,10 @@ class Session:
                  project_id: str=None, profile=None,
                  input_tokens: int=0, output_tokens: int=0, estimated_cost=None,
                  personality=None,
+                 active_stream_id: str=None,
+                 pending_user_message: str=None,
+                 pending_attachments=None,
+                 pending_started_at=None,
                  **kwargs):
         self.session_id = session_id or uuid.uuid4().hex[:12]
         self.title = title
@@ -61,6 +65,10 @@ class Session:
         self.output_tokens = output_tokens or 0
         self.estimated_cost = estimated_cost
         self.personality = personality
+        self.active_stream_id = active_stream_id
+        self.pending_user_message = pending_user_message
+        self.pending_attachments = pending_attachments or []
+        self.pending_started_at = pending_started_at
 
     @property
     def path(self):

api/routes.py

@@ -365,6 +365,10 @@ def handle_get(handler, parsed) -> bool:
             raw = s.compact() | {
                 "messages": s.messages,
                 "tool_calls": getattr(s, "tool_calls", []),
+                "active_stream_id": getattr(s, "active_stream_id", None),
+                "pending_user_message": getattr(s, "pending_user_message", None),
+                "pending_attachments": getattr(s, "pending_attachments", []),
+                "pending_started_at": getattr(s, "pending_started_at", None),
             }
             return j(handler, {"session": redact_session_data(raw)})
         except KeyError:
@@ -1683,11 +1687,15 @@ def _handle_chat_start(handler, body):
     attachments = [str(a) for a in (body.get("attachments") or [])][:20]
     workspace = str(Path(body.get("workspace") or s.workspace).expanduser().resolve())
     model = body.get("model") or s.model
+    stream_id = uuid.uuid4().hex
     s.workspace = workspace
     s.model = model
+    s.active_stream_id = stream_id
+    s.pending_user_message = msg
+    s.pending_attachments = attachments
+    s.pending_started_at = time.time()
     s.save()
     set_last_workspace(workspace)
-    stream_id = uuid.uuid4().hex
     q = queue.Queue()
     with STREAMS_LOCK:
         STREAMS[stream_id] = q

api/streaming.py

@@ -172,12 +172,47 @@ def _run_agent_streaming(session_id, msg_text, model, workspace, stream_id, atta
                 _token_sent = True
                 put('token', {'text': text})
 
-            def on_tool(name, preview, args):
+            def on_reasoning(text):
+                if text is None:
+                    return
+                put('reasoning', {'text': str(text)})
+
+            def on_tool(*cb_args, **cb_kwargs):
+                event_type = None
+                name = None
+                preview = None
+                args = None
+
+                if len(cb_args) >= 4:
+                    event_type, name, preview, args = cb_args[:4]
+                elif len(cb_args) == 3:
+                    name, preview, args = cb_args
+                    event_type = 'tool.started'
+                elif len(cb_args) == 2:
+                    event_type, name = cb_args
+                elif len(cb_args) == 1:
+                    name = cb_args[0]
+                    event_type = 'tool.started'
+
+                if event_type in ('reasoning.available', '_thinking'):
+                    reason_text = preview if event_type == 'reasoning.available' else name
+                    if reason_text:
+                        put('reasoning', {'text': str(reason_text)})
+                    return
+
                 args_snap = {}
                 if isinstance(args, dict):
                     for k, v in list(args.items())[:4]:
-                        s2 = str(v); args_snap[k] = s2[:120]+('...' if len(s2)>120 else '')
-                put('tool', {'name': name, 'preview': preview, 'args': args_snap})
+                        s2 = str(v)
+                        args_snap[k] = s2[:120] + ('...' if len(s2) > 120 else '')
+
+                if event_type in (None, 'tool.started'):
+                    put('tool', {
+                        'event_type': event_type or 'tool.started',
+                        'name': name,
+                        'preview': preview,
+                        'args': args_snap,
+                    })
                     # Fallback: poll for pending approval in case notify_cb wasn't
                     # registered (e.g. older approval module without gateway support).
                     try:
@@ -189,6 +224,18 @@ def _run_agent_streaming(session_id, msg_text, model, workspace, stream_id, atta
                                 put('approval', p)
                     except ImportError:
                         pass
+                    return
+
+                if event_type == 'tool.completed':
+                    put('tool_complete', {
+                        'event_type': event_type,
+                        'name': name,
+                        'preview': preview,
+                        'args': args_snap,
+                        'duration': cb_kwargs.get('duration'),
+                        'is_error': bool(cb_kwargs.get('is_error', False)),
+                    })
+                    return
 
             _AIAgent = _get_ai_agent()
             if _AIAgent is None:
@@ -252,6 +299,7 @@ def _run_agent_streaming(session_id, msg_text, model, workspace, stream_id, atta
                 session_id=session_id,
                 session_db=_session_db,
                 stream_delta_callback=on_token,
+                reasoning_callback=on_reasoning,
                 tool_progress_callback=on_tool,
             )
 
@@ -458,6 +506,10 @@ def _run_agent_streaming(session_id, msg_text, model, workspace, stream_id, atta
                         'assistant_msg_idx': asst_idx, 'args': args_snap,
                     })
             s.tool_calls = tool_calls
+            s.active_stream_id = None
+            s.pending_user_message = None
+            s.pending_attachments = []
+            s.pending_started_at = None
             # Tag the matching user message with attachment filenames for display on reload
             # Only tag a user message whose content relates to this turn's text
             # (msg_text is the full message including the [Attached files: ...] suffix)
@@ -516,6 +568,15 @@ def _run_agent_streaming(session_id, msg_text, model, workspace, stream_id, atta
 
     except Exception as e:
         print('[webui] stream error:\n' + traceback.format_exc(), flush=True)
+        if s is not None:
+            s.active_stream_id = None
+            s.pending_user_message = None
+            s.pending_attachments = []
+            s.pending_started_at = None
+            try:
+                s.save()
+            except Exception:
+                pass
         err_str = str(e)
         # Detect rate limit errors specifically so the client can show a helpful card
         # rather than the generic "Connection lost" message

static/messages.js

@@ -10,10 +10,12 @@ async function send(){
   // If busy, queue the message instead of dropping it
   if(S.busy){
     if(text){
-      MSG_QUEUE.push(text);
+      if(!S.session){await newSession();await renderSessionList();}
+      queueSessionMessage(S.session.session_id,{text,files:[...S.pendingFiles]});
       $('msg').value='';autoResize();
-      updateQueueBadge();
-      showToast(`Queued: "${text.slice(0,40)}${text.length>40?'\u2026':''}"`,2000);
+      S.pendingFiles=[];renderTray();
+      updateQueueBadge(S.session.session_id);
+      showToast(`Queued: "${text.slice(0,40)}${text.length>40?'…':''}"`,2000);
     }
     return;
   }
@@ -37,7 +39,7 @@ async function send(){
   S.toolCalls=[];  // clear tool calls from previous turn
   clearLiveToolCards();  // clear any leftover live cards from last turn
   S.messages.push(userMsg);renderMessages();appendThinking();setBusy(true);
-  INFLIGHT[activeSid]={messages:[...S.messages],uploaded};
+  INFLIGHT[activeSid]={messages:[...S.messages],uploaded,toolCalls:[]};
   startApprovalPolling(activeSid);
   S.activeStreamId = null;  // will be set after stream starts
 
@@ -81,7 +83,32 @@ async function send(){
   }
 
   // Open SSE stream and render tokens live
+  attachLiveStream(activeSid, streamId, uploaded);
+
+}
+
+const LIVE_STREAMS={};
+
+function closeLiveStream(sessionId, streamId){
+  const live=LIVE_STREAMS[sessionId];
+  if(!live) return;
+  if(streamId&&live.streamId!==streamId) return;
+  try{live.source.close();}catch(_){ }
+  delete LIVE_STREAMS[sessionId];
+}
+
+function attachLiveStream(activeSid, streamId, uploaded=[], options={}){
+  if(!activeSid||!streamId) return;
+  const reconnecting=!!options.reconnecting;
+  closeLiveStream(activeSid);
+  if(!INFLIGHT[activeSid]) INFLIGHT[activeSid]={messages:[...S.messages],uploaded:[...uploaded],toolCalls:[]};
+  else {
+    if(uploaded.length) INFLIGHT[activeSid].uploaded=[...uploaded];
+    if(!Array.isArray(INFLIGHT[activeSid].toolCalls)) INFLIGHT[activeSid].toolCalls=[];
+  }
+
   let assistantText='';
+  let reasoningText='';
   let assistantRow=null;
   let assistantBody=null;
   // Thinking tag patterns for streaming display
@@ -90,8 +117,45 @@ async function send(){
     {open:'<|channel>thought\n',close:'<channel|>'}
   ];
 
+  function _isActiveSession(){
+    return !!(S.session&&S.session.session_id===activeSid);
+  }
+  function _closeSource(){
+    closeLiveStream(activeSid, streamId);
+  }
+  function syncInflightAssistantMessage(){
+    const inflight=INFLIGHT[activeSid];
+    if(!inflight) return;
+    if(!Array.isArray(inflight.messages)) inflight.messages=[];
+    let assistantIdx=-1;
+    for(let i=inflight.messages.length-1;i>=0;i--){
+      const msg=inflight.messages[i];
+      if(msg&&msg.role==='assistant'&&msg._live){assistantIdx=i;break;}
+    }
+    const ts=Date.now()/1000;
+    if(assistantIdx>=0){
+      inflight.messages[assistantIdx].content=assistantText;
+      inflight.messages[assistantIdx].reasoning=reasoningText||undefined;
+      inflight.messages[assistantIdx]._ts=inflight.messages[assistantIdx]._ts||ts;
+      return;
+    }
+    inflight.messages.push({role:'assistant',content:assistantText,reasoning:reasoningText||undefined,_live:true,_ts:ts});
+  }
   function ensureAssistantRow(){
-    if(assistantRow)return;
+    if(!_isActiveSession()) return;
+    if(assistantRow&&!assistantRow.isConnected){assistantRow=null;assistantBody=null;}
+    if(!assistantRow){
+      const existing=$('msgInner').querySelector('.msg-row[data-live-assistant="1"]');
+      if(existing){
+        assistantRow=existing;
+        assistantBody=existing.querySelector('.msg-body');
+      }
+    }
+    if(assistantRow){
+      if(typeof placeLiveToolCardsHost==='function') placeLiveToolCardsHost();
+      return;
+    }
+
     removeThinking();
     const tr=$('toolRunningRow');if(tr)tr.remove();
     $('emptyState').style.display='none';
@@ -115,6 +179,7 @@ async function send(){
   // and hiding content still inside an open thinking block.
   function _streamDisplay(){
     const raw=assistantText;
+    if(reasoningText) return raw;
     for(const {open,close} of _thinkPairs){
       // Trim leading whitespace before checking for the open tag — some models
       // (e.g. MiniMax) emit newlines before <think>.
@@ -134,15 +199,52 @@ async function send(){
     }
     return raw;
   }
+  function _parseStreamState(){
+    const raw=assistantText;
+    if(reasoningText){
+      return {thinkingText:reasoningText, displayText:_streamDisplay(), inThinking:false};
+    }
+    for(const {open,close} of _thinkPairs){
+      const trimmed=raw.trimStart();
+      if(trimmed.startsWith(open)){
+        const ci=trimmed.indexOf(close,open.length);
+        if(ci!==-1){
+          return {
+            thinkingText: trimmed.slice(open.length, ci).trim(),
+            displayText: trimmed.slice(ci+close.length).replace(/^\s+/,''),
+            inThinking:false,
+          };
+        }
+        return {
+          thinkingText: trimmed.slice(open.length).trim(),
+          displayText:'',
+          inThinking:true,
+        };
+      }
+      if(open.startsWith(trimmed)){
+        return {thinkingText:'', displayText:'', inThinking:true};
+      }
+    }
+    return {thinkingText:'', displayText:raw, inThinking:false};
+  }
+  function _renderLiveThinking(parsed){
+    const text=(parsed&&parsed.thinkingText)||'';
+    if(text||(parsed&&parsed.inThinking)){
+      if(typeof updateThinking==='function') updateThinking(text||'Thinking…');
+      else appendThinking();
+      return;
+    }
+    removeThinking();
+  }
   function _scheduleRender(){
     if(_renderPending) return;
     _renderPending=true;
     requestAnimationFrame(()=>{
       _renderPending=false;
+      const parsed=_parseStreamState();
+      _renderLiveThinking(parsed);
       if(assistantBody){
-        const txt=_streamDisplay();
-        const isThinking=!txt&&assistantText.length>0;
-        assistantBody.innerHTML=txt?renderMd(txt):(isThinking?'<span style="color:var(--muted);font-size:13px">Thinking\u2026</span>':'');
+        assistantBody.innerHTML=parsed.displayText?renderMd(parsed.displayText):'';
       }
       scrollIfPinned();
     });
@@ -153,17 +255,59 @@ async function send(){
       if(!S.session||S.session.session_id!==activeSid) return;
       const d=JSON.parse(e.data);
       assistantText+=d.text;
+      syncInflightAssistantMessage();
+      if(!S.session||S.session.session_id!==activeSid) return;
+
       ensureAssistantRow();
       _scheduleRender();
     });
 
+    source.addEventListener('reasoning',e=>{
+      const d=JSON.parse(e.data);
+      reasoningText += d.text || '';
+      syncInflightAssistantMessage();
+      if(!S.session||S.session.session_id!==activeSid) return;
+      _scheduleRender();
+    });
+
     source.addEventListener('tool',e=>{
       const d=JSON.parse(e.data);
+      const tc={name:d.name, preview:d.preview||'', args:d.args||{}, snippet:'', done:false, tid:d.tid||`live-${Date.now()}-${Math.random().toString(36).slice(2,8)}`};
+      if(!Array.isArray(INFLIGHT[activeSid].toolCalls)) INFLIGHT[activeSid].toolCalls=[];
+      INFLIGHT[activeSid].toolCalls.push(tc);
+      S.toolCalls=INFLIGHT[activeSid].toolCalls;
+
       if(!S.session||S.session.session_id!==activeSid) return;
       removeThinking();
       const oldRow=$('toolRunningRow');if(oldRow)oldRow.remove();
-      const tc={name:d.name, preview:d.preview||'', args:d.args||{}, snippet:'', done:false};
-      S.toolCalls.push(tc);
+      appendLiveToolCard(tc);
+      scrollIfPinned();
+    });
+
+    source.addEventListener('tool_complete',e=>{
+      const d=JSON.parse(e.data);
+      const inflight=INFLIGHT[activeSid];
+      if(!inflight) return;
+      if(!Array.isArray(inflight.toolCalls)) inflight.toolCalls=[];
+      let tc=null;
+      for(let i=inflight.toolCalls.length-1;i>=0;i--){
+        const cur=inflight.toolCalls[i];
+        if(cur&&cur.done===false&&(!d.name||cur.name===d.name)){
+          tc=cur;
+          break;
+        }
+      }
+      if(!tc){
+        tc={name:d.name||'tool', preview:d.preview||'', args:d.args||{}, snippet:'', done:true};
+        inflight.toolCalls.push(tc);
+      }
+      tc.preview=d.preview||tc.preview||'';
+      tc.args=d.args||tc.args||{};
+      tc.done=true;
+      tc.is_error=!!d.is_error;
+      if(d.duration!==undefined) tc.duration=d.duration;
+      S.toolCalls=inflight.toolCalls;
+      if(!S.session||S.session.session_id!==activeSid) return;
       appendLiveToolCard(tc);
       scrollIfPinned();
     });

static/panels.js

@@ -308,10 +308,11 @@ async function cronDelete(id) {
 function loadTodos() {
   const panel = $('todoPanel');
   if (!panel) return;
+  const sourceMessages = (S.session && Array.isArray(S.session.messages) && S.session.messages.length) ? S.session.messages : S.messages;
   // Parse the most recent todo state from message history
   let todos = [];
-  for (let i = S.messages.length - 1; i >= 0; i--) {
-    const m = S.messages[i];
+  for (let i = sourceMessages.length - 1; i >= 0; i--) {
+    const m = sourceMessages[i];
     if (m && m.role === 'tool') {
       try {
         const d = JSON.parse(typeof m.content === 'string' ? m.content : JSON.stringify(m.content));

static/sessions.js

@@ -11,7 +11,7 @@ const ICONS={
 };
 
 async function newSession(flash){
-  MSG_QUEUE.length=0;updateQueueBadge();
+  updateQueueBadge();
   S.toolCalls=[];
   clearLiveToolCards();
   // Use profile default workspace for new sessions after a profile switch (one-shot),
@@ -20,9 +20,19 @@ async function newSession(flash){
   S._profileDefaultWorkspace=null; // consume — only applies to the first new session after switch
   const data=await api('/api/session/new',{method:'POST',body:JSON.stringify({model:$('modelSelect').value,workspace:inheritWs})});
   S.session=data.session;S.messages=data.session.messages||[];
+  S.lastUsage={...(data.session.last_usage||{})};
   if(flash)S.session._flash=true;
   localStorage.setItem('hermes-webui-session',S.session.session_id);
-  syncTopbar();await loadDir('.');renderMessages();
+  // Reset per-session visual state: a fresh chat is idle even if another
+  // conversation is still streaming in the background.
+  S.busy=false;
+  S.activeStreamId=null;
+  updateSendBtn();
+  const _cb=$('btnCancel');if(_cb)_cb.style.display='none';
+  setStatus('');
+  setComposerStatus('');
+  updateQueueBadge(S.session.session_id);
+  syncTopbar();renderMessages();loadDir('.');
   // don't call renderSessionList here - callers do it when needed
 }
 
@@ -30,40 +40,74 @@ async function loadSession(sid){
   stopApprovalPolling();hideApprovalCard();
   const data=await api(`/api/session?session_id=${encodeURIComponent(sid)}`);
   S.session=data.session;
+  S.lastUsage={...(data.session.last_usage||{})};
   localStorage.setItem('hermes-webui-session',S.session.session_id);
-  // B9: sanitize empty assistant messages that can appear when agent only ran tool calls
-  data.session.messages=(data.session.messages||[]).filter(m=>{
-    if(!m||!m.role)return false;
-    if(m.role==='tool')return false;
-    if(m.role==='assistant'){let c=m.content||'';if(Array.isArray(c))c=c.filter(p=>p&&p.type==='text').map(p=>p.text||'').join('');return String(c).trim().length>0;}
-    return true;
-  });
+  const activeStreamId=data.session.active_stream_id||null;
+  if(!INFLIGHT[sid]&&activeStreamId&&typeof loadInflightState==='function'){
+    const stored=loadInflightState(sid, activeStreamId);
+    if(stored){
+      INFLIGHT[sid]={
+        messages:Array.isArray(stored.messages)&&stored.messages.length?stored.messages:[...(data.session.messages||[])],
+        uploaded:Array.isArray(stored.uploaded)?stored.uploaded:[...(data.session.pending_attachments||[])],
+        toolCalls:Array.isArray(stored.toolCalls)?stored.toolCalls:[],
+        reattach:true,
+      };
+    }
+  }
+  // Keep raw session.messages intact so side panels (e.g. Todos) can still
+  // reconstruct state from tool outputs after reload. Visible transcript rows
+  // are filtered later by renderMessages().
   if(INFLIGHT[sid]){
     S.messages=INFLIGHT[sid].messages;
-    // Restore live tool cards for this in-flight session
+    S.toolCalls=(INFLIGHT[sid].toolCalls||[]);
+    S.busy=true;
+    syncTopbar();renderMessages();appendThinking();loadDir('.');
     clearLiveToolCards();
+    if(typeof placeLiveToolCardsHost==='function') placeLiveToolCardsHost();
     for(const tc of (S.toolCalls||[])){
       if(tc&&tc.name) appendLiveToolCard(tc);
     }
-    syncTopbar();await loadDir('.');renderMessages();appendThinking();
     setBusy(true);setComposerStatus('');
     startApprovalPolling(sid);
+    S.activeStreamId=activeStreamId;
+    const _cb=$('btnCancel');if(_cb&&activeStreamId)_cb.style.display='inline-flex';
+    if(INFLIGHT[sid].reattach&&activeStreamId&&typeof attachLiveStream==='function'){
+      INFLIGHT[sid].reattach=false;
+      attachLiveStream(sid, activeStreamId, data.session.pending_attachments||[], {reconnecting:true});
+    }
   }else{
-    MSG_QUEUE.length=0;updateQueueBadge();  // clear queue for the viewed session
+    updateQueueBadge(sid);
     S.messages=data.session.messages||[];
+    const pendingMsg=typeof getPendingSessionMessage==='function'?getPendingSessionMessage(data.session):null;
+    if(pendingMsg) S.messages.push(pendingMsg);
     S.toolCalls=(data.session.tool_calls||[]).map(tc=>({...tc,done:true}));
+    clearLiveToolCards();
+    if(activeStreamId){
+      S.busy=true;
+      S.activeStreamId=activeStreamId;
+      updateSendBtn();
+      const _cb=$('btnCancel');if(_cb)_cb.style.display='inline-flex';
+      setStatus('');
+      setComposerStatus('');
+      syncTopbar();renderMessages();appendThinking();loadDir('.');
+      updateQueueBadge(sid);
+      startApprovalPolling(sid);
+      if(typeof attachLiveStream==='function') attachLiveStream(sid, activeStreamId, data.session.pending_attachments||[], {reconnecting:true});
+      else if(typeof watchInflightSession==='function') watchInflightSession(sid, activeStreamId);
+    }else{
       // Reset per-session visual state: the viewed session is idle even if another
       // session's stream is still running in the background.
       // We directly update the DOM instead of calling setBusy(false), because
-    // setBusy(false) drains MSG_QUEUE which we don't want here.
+      // setBusy(false) drains the viewed session's queued follow-up turns.
       S.busy=false;
       S.activeStreamId=null;
       updateSendBtn();
       const _cb=$('btnCancel');if(_cb)_cb.style.display='none';
       setStatus('');
       setComposerStatus('');
-    clearLiveToolCards();
-    syncTopbar();await loadDir('.');renderMessages();highlightCode();
+      updateQueueBadge(sid);
+      syncTopbar();renderMessages();highlightCode();loadDir('.');
+    }
   }
   // Sync context usage indicator from session data
   const _s=S.session;

static/ui.js

@@ -1,7 +1,28 @@
 const S={session:null,messages:[],entries:[],busy:false,pendingFiles:[],toolCalls:[],activeStreamId:null,currentDir:'.',activeProfile:'default'};
 const INFLIGHT={};  // keyed by session_id while request in-flight
-const MSG_QUEUE=[];  // messages queued while a request is in-flight
+const SESSION_QUEUES={};  // keyed by session_id for queued follow-up turns
 const $=id=>document.getElementById(id);
+function _getSessionQueue(sid, create=false){
+  if(!sid) return [];
+  if(!SESSION_QUEUES[sid]&&create) SESSION_QUEUES[sid]=[];
+  return SESSION_QUEUES[sid]||[];
+}
+function queueSessionMessage(sid, payload){
+  if(!sid||!payload) return 0;
+  const q=_getSessionQueue(sid,true);
+  q.push(payload);
+  return q.length;
+}
+function shiftQueuedSessionMessage(sid){
+  const q=_getSessionQueue(sid,false);
+  if(!q.length) return null;
+  const next=q.shift();
+  if(!q.length) delete SESSION_QUEUES[sid];
+  return next;
+}
+function getQueuedSessionCount(sid){
+  return _getSessionQueue(sid,false).length;
+}
 const esc=s=>String(s??'').replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
 
 // Dynamic model labels -- populated by populateModelDropdown(), fallback to static map
@@ -513,28 +534,37 @@ function setBusy(v){
     setComposerStatus('');
     // Always hide Cancel button when not busy
     const _cb=$('btnCancel');if(_cb)_cb.style.display='none';
-    updateQueueBadge();
-    // Drain one queued message after UI settles
-    if(MSG_QUEUE.length>0){
-      const next=MSG_QUEUE.shift();
-      updateQueueBadge();
-      setTimeout(()=>{ $('msg').value=next; send(); }, 120);
+    const sid=S.session&&S.session.session_id;
+    updateQueueBadge(sid);
+    // Drain one queued message for the currently viewed session after UI settles
+    const next=sid?shiftQueuedSessionMessage(sid):null;
+    if(next){
+      updateQueueBadge(sid);
+      setTimeout(()=>{
+        $('msg').value=next.text||'';
+        S.pendingFiles=Array.isArray(next.files)?[...next.files]:[];
+        autoResize();
+        renderTray();
+        send();
+      },120);
     }
   }
 }
 
-function updateQueueBadge(){
+function updateQueueBadge(sessionId){
+  const sid=sessionId||(S.session&&S.session.session_id);
+  const count=sid?getQueuedSessionCount(sid):0;
   let badge=$('queueBadge');
-  if(MSG_QUEUE.length>0){
+  if(count>0){
     if(!badge){
       badge=document.createElement('div');
       badge.id='queueBadge';
       badge.style.cssText='position:fixed;bottom:80px;right:24px;background:rgba(124,185,255,.18);border:1px solid rgba(124,185,255,.4);color:var(--blue);font-size:12px;font-weight:600;padding:6px 14px;border-radius:20px;z-index:50;pointer-events:none;backdrop-filter:blur(8px);';
       document.body.appendChild(badge);
     }
-    badge.textContent=MSG_QUEUE.length===1?'1 message queued':`${MSG_QUEUE.length} messages queued`;
-  } else {
-    if(badge) badge.remove();
+    badge.textContent=count===1?'1 message queued':`${count} messages queued`;
+  } else if(badge) {
+    badge.remove();
   }
 }
 function showToast(msg,ms){const el=$('toast');el.textContent=msg;el.classList.add('show');clearTimeout(el._t);el._t=setTimeout(()=>el.classList.remove('show'),ms||2800);}
@@ -714,11 +744,11 @@ async function refreshSession() {
   try {
     const data = await api(`/api/session?session_id=${encodeURIComponent(S.session.session_id)}`);
     S.session = data.session;
-    S.messages = (data.session.messages || []).filter(m => {
-      if (!m || !m.role || m.role === 'tool') return false;
-      if (m.role === 'assistant') { let c = m.content || ''; if (Array.isArray(c)) c = c.map(p => p.text||'').join(''); return String(c).trim().length > 0; }
-      return true;
-    });
+    S.messages = data.session.messages || [];
+    const pendingMsg=getPendingSessionMessage(data.session);
+    if(pendingMsg) S.messages.push(pendingMsg);
+    S.activeStreamId=data.session.active_stream_id||null;
+
     syncTopbar(); renderMessages();
     showToast('Conversation refreshed');
   } catch(e) { setStatus('Refresh failed: ' + e.message); }
@@ -764,12 +794,46 @@ async function applyUpdates(){
   }
 }
 
+function getPendingSessionMessage(session){
+  const text=String(session?.pending_user_message||'').trim();
+  if(!text) return null;
+  const attachments=Array.isArray(session?.pending_attachments)?session.pending_attachments.filter(Boolean):[];
+  const messages=Array.isArray(session?.messages)?session.messages:[];
+  const lastUser=[...messages].reverse().find(m=>m&&m.role==='user');
+  if(lastUser){
+    const lastText=String(msgContent(lastUser)||'').trim();
+    if(lastText===text){
+      if(attachments.length&&!lastUser.attachments?.length) lastUser.attachments=attachments;
+      return null;
+    }
+  }
+  return {
+    role:'user',
+    content:text,
+    attachments:attachments.length?attachments:undefined,
+    _ts:session?.pending_started_at||Date.now()/1000,
+    _pending:true,
+  };
+}
+// loadInflightState — retrieve in-memory inflight state for a session.
+// Called by loadSession() when active_stream_id is set on the server session
+// but no INFLIGHT[sid] entry exists (e.g. after a session switch back).
+// Returns the stored state dict or null. The else-path in loadSession handles
+// page reloads directly via attachLiveStream when this returns null.
+function loadInflightState(sid, streamId) {
+  // In-memory store: only survives within the same page load.
+  // If INFLIGHT[sid] exists but the caller already checked !INFLIGHT[sid],
+  // this won't be reached. Return null — the else path handles page reloads.
+  return null;
+}
+
 async function checkInflightOnBoot(sid) {
   const raw = localStorage.getItem(INFLIGHT_KEY);
   if (!raw) return;
   try {
     const {sid: inflightSid, streamId, ts} = JSON.parse(raw);
     if (inflightSid !== sid) { clearInflight(); return; }
+    if (S.activeStreamId && S.activeStreamId === streamId) return;
     // Only show banner if the in-flight entry is less than 10 minutes old
     if (Date.now() - ts > 10 * 60 * 1000) { clearInflight(); return; }
     // Check if stream is still active
@@ -915,6 +979,7 @@ function renderMessages(){
     }
     const row=document.createElement('div');row.className='msg-row';
     row.dataset.msgIdx=rawIdx;row.dataset.role=m.role||'assistant';
+    if(m._live) row.setAttribute('data-live-assistant','1');
     let filesHtml='';
     if(m.attachments&&m.attachments.length)
       filesHtml=`<div class="msg-files">${m.attachments.map(f=>`<div class="msg-file-badge">${li('paperclip',12)} ${esc(f)}</div>`).join('')}</div>`;
@@ -1357,12 +1422,29 @@ function renderKatexBlocks(){
   });
 }
 
-function appendThinking(){
-  $('emptyState').style.display='none';
-  const row=document.createElement('div');row.className='msg-row';row.id='thinkingRow';
-  row.innerHTML=`<div class="msg-role assistant"><div class="role-icon assistant">H</div>Hermes</div><div class="thinking"><div class="dot"></div><div class="dot"></div><div class="dot"></div></div>`;
-  $('msgInner').appendChild(row);scrollToBottom();
+function _thinkingMarkup(text=''){
+  const _bn=window._botName||'Hermes';
+  const icon=esc(_bn.charAt(0).toUpperCase());
+  const label=esc(_bn);
+  const body=(text&&String(text).trim())
+    ? `<div class="thinking-card open"><div class="thinking-card-header"><span class="thinking-card-icon">${li('lightbulb',14)}</span><span class="thinking-card-label">${t('thinking')}</span></div><div class="thinking-card-body"><pre>${esc(String(text).trim())}</pre></div></div>`
+    : `<div class="thinking"><div class="dot"></div><div class="dot"></div><div class="dot"></div></div>`;
+  return `<div class="msg-role assistant"><div class="role-icon assistant">${icon}</div>${label}</div>${body}`;
 }
+function appendThinking(text=''){
+  $('emptyState').style.display='none';
+  let row=$('thinkingRow');
+  if(!row){
+    row=document.createElement('div');
+    row.className='msg-row';
+    row.id='thinkingRow';
+    $('msgInner').appendChild(row);
+  }
+  row.className=(text&&String(text).trim())?'msg-row thinking-card-row':'msg-row';
+  row.innerHTML=_thinkingMarkup(text);
+  scrollToBottom();
+}
+function updateThinking(text=''){appendThinking(text);}
 function removeThinking(){const el=$('thinkingRow');if(el)el.remove();}
 
 function fileIcon(name, type){

tests/test_regressions.py

@@ -241,6 +241,24 @@ def test_done_handler_guards_setbusy_with_inflight_check(cleanup_test_sessions):
     assert "INFLIGHT[S.session.session_id]" in src,         "messages.js must guard setBusy(false) with INFLIGHT check for current session"
 
 
+def test_refresh_handler_does_not_drop_tool_messages_needed_by_todos(cleanup_test_sessions):
+    """Todo panel state must survive session reload/refresh.
+    The UI can hide tool-role messages from the visible transcript, but it must not
+    destroy the raw session messages because loadTodos reconstructs state from the
+    latest todo tool output.
+    """
+    sessions_src = (REPO_ROOT / "static/sessions.js").read_text()
+    ui_src = (REPO_ROOT / "static/ui.js").read_text()
+    panels_src = (REPO_ROOT / "static/panels.js").read_text()
+
+    assert "data.session.messages=(data.session.messages||[]).filter(" not in sessions_src, \
+        "sessions.js must not overwrite raw session.messages when filtering transcript display"
+    assert "S.messages = (data.session.messages || []).filter(" not in ui_src, \
+        "ui.js refreshSession must not rebuild S.messages by discarding tool messages from the raw session payload"
+    assert "const sourceMessages = (S.session && Array.isArray(S.session.messages) && S.session.messages.length) ? S.session.messages : S.messages;" in panels_src, \
+        "loadTodos must prefer raw S.session.messages so todo state survives reloads"
+
+
 def test_cancel_button_not_cleared_across_sessions(cleanup_test_sessions):
     """R7c: The Cancel button and activeStreamId must only be cleared when the
     done/error event belongs to the currently viewed session.
@@ -440,7 +458,166 @@ def test_newSession_clears_live_tool_cards(cleanup_test_sessions):
     assert "clearLiveToolCards" in new_sess_body,         "newSession() must call clearLiveToolCards() to clear stale live cards"
 
 
-# ── R16: Stack traces must not leak to clients in 500 responses ────────────
+def test_newSession_resets_busy_state_for_fresh_chat(cleanup_test_sessions):
+    """R15b: newSession() must reset the viewed chat to idle state.
+    Without this, starting a second chat while another session is streaming leaves
+    S.busy=true, so the first send in the new chat gets incorrectly queued.
+    """
+    src = (REPO_ROOT / "static/sessions.js").read_text()
+    new_sess_idx = src.find("async function newSession(")
+    assert new_sess_idx >= 0
+    next_fn = src.find("async function ", new_sess_idx + 10)
+    new_sess_body = src[new_sess_idx:next_fn]
+    assert "S.busy=false;" in new_sess_body, \
+        "newSession() must clear S.busy so a fresh chat is immediately sendable"
+    assert "S.activeStreamId=null;" in new_sess_body, \
+        "newSession() must clear the active stream id for the newly viewed chat"
+    assert "updateQueueBadge(S.session.session_id);" in new_sess_body, \
+        "newSession() must refresh the badge for the new session rather than leaving the old session's queue badge visible"
+
+
+def test_session_scoped_message_queue_frontend_wiring(cleanup_test_sessions):
+    """R15bb: queued follow-ups must stay attached to their originating session.
+    The frontend should use a session-keyed queue store and drain only the active
+    session's queued messages when that session becomes idle.
+    """
+    ui_src = (REPO_ROOT / "static/ui.js").read_text()
+    messages_src = (REPO_ROOT / "static/messages.js").read_text()
+    sessions_src = (REPO_ROOT / "static/sessions.js").read_text()
+    assert "const SESSION_QUEUES" in ui_src
+    assert "function queueSessionMessage" in ui_src
+    assert "function shiftQueuedSessionMessage" in ui_src
+    assert "const sid=S.session&&S.session.session_id;" in ui_src
+    assert "const next=sid?shiftQueuedSessionMessage(sid):null;" in ui_src
+    assert "queueSessionMessage(S.session.session_id" in messages_src
+    assert "updateQueueBadge(S.session.session_id);" in messages_src
+    assert "updateQueueBadge(sid);" in sessions_src
+
+
+def test_chat_start_persists_pending_turn_metadata_for_reload_recovery(cleanup_test_sessions):
+    """R15c: chat/start must expose enough pending-turn metadata for a reload to
+    rebuild the in-flight conversation instead of showing a blank session.
+    """
+    routes_src = (REPO_ROOT / "api/routes.py").read_text()
+    assert 's.active_stream_id = stream_id' in routes_src
+    assert 's.pending_user_message = msg' in routes_src
+    assert 's.pending_attachments = attachments' in routes_src
+    assert '"active_stream_id": getattr(s, "active_stream_id", None)' in routes_src
+    assert '"pending_user_message": getattr(s, "pending_user_message", None)' in routes_src
+
+
+def test_reload_path_restores_pending_message_and_reattaches_live_stream(cleanup_test_sessions):
+    """R15d: the frontend reload path must show the pending user turn and
+    reattach to the live SSE stream after loadSession().
+    """
+    sessions_src = (REPO_ROOT / "static/sessions.js").read_text()
+    ui_src = (REPO_ROOT / "static/ui.js").read_text()
+    messages_src = (REPO_ROOT / "static/messages.js").read_text()
+    assert 'getPendingSessionMessage' in ui_src
+    assert 'pending_user_message' in ui_src
+    assert 'function attachLiveStream' in messages_src
+    assert 'const pendingMsg=typeof getPendingSessionMessage' in sessions_src
+    assert 'const activeStreamId=data.session.active_stream_id||null;' in sessions_src
+    assert 'attachLiveStream(sid, activeStreamId' in sessions_src
+    assert 'if (S.activeStreamId && S.activeStreamId === streamId) return;' in ui_src
+
+
+# ── R16: Switching away/back must preserve live partial assistant output ─────
+
+
+def test_live_stream_tokens_persist_partial_assistant_for_session_switch(cleanup_test_sessions):
+    """R16: in-flight assistant text must be mirrored into INFLIGHT session state,
+    and the live stream must rebind to the rebuilt DOM after switching away and back.
+    Without this, partial assistant output disappears until the final done payload lands.
+    """
+    messages_src = (REPO_ROOT / "static/messages.js").read_text()
+    ui_src = (REPO_ROOT / "static/ui.js").read_text()
+
+    assert "content:assistantText" in messages_src, \
+        "messages.js must persist the partial assistant text into INFLIGHT state"
+    assert "_live:true" in messages_src, \
+        "messages.js must mark the persisted in-flight assistant row so renderMessages can re-anchor it"
+    assert "syncInflightAssistantMessage();" in messages_src, \
+        "token handler must update INFLIGHT state before checking the active session"
+    assert "assistantRow&&!assistantRow.isConnected" in messages_src, \
+        "live stream must drop stale detached assistant DOM references after session switches"
+    assert "data-live-assistant" in ui_src, \
+        "renderMessages must preserve a live-assistant DOM anchor when rebuilding the thread"
+
+
+def test_inflight_session_state_tracks_live_tool_cards_per_session(cleanup_test_sessions):
+    """R16b: live tool cards must be stored on the in-flight session, not only in the
+    global S.toolCalls array, so switching chats does not lose or misattach them.
+    """
+    messages_src = (REPO_ROOT / "static/messages.js").read_text()
+    sessions_src = (REPO_ROOT / "static/sessions.js").read_text()
+
+    assert "INFLIGHT[activeSid].toolCalls.push(tc);" in messages_src, \
+        "tool SSE handler must persist live tool calls onto the in-flight session"
+    assert "S.toolCalls=(INFLIGHT[sid].toolCalls||[]);" in sessions_src, \
+        "loadSession() must restore live tool calls from the in-flight session state"
+
+
+def test_loadSession_inflight_sets_busy_before_renderMessages(cleanup_test_sessions):
+    """R16c: loading an in-flight session must mark it busy before renderMessages().
+    Otherwise renderMessages() treats S.toolCalls as settled history cards and the
+    same tool call appears once inline and once in the live tool host after a
+    session switch.
+    """
+    src = (REPO_ROOT / "static/sessions.js").read_text()
+    inflight_idx = src.find("if(INFLIGHT[sid]){")
+    assert inflight_idx >= 0, "INFLIGHT branch not found in loadSession"
+    inflight_block = src[inflight_idx:inflight_idx+700]
+    busy_pos = inflight_block.find("S.busy=true;")
+    render_pos = inflight_block.find("renderMessages();appendThinking();")
+    assert busy_pos >= 0, "loadSession INFLIGHT branch must set S.busy=true"
+    assert render_pos >= 0, "loadSession INFLIGHT branch must call renderMessages()"
+    assert busy_pos < render_pos, \
+        "loadSession must set S.busy=true before renderMessages() to avoid duplicate tool cards"
+
+
+def test_streaming_bridge_accepts_current_tool_progress_callback_signature(cleanup_test_sessions):
+    """R17: api/streaming.py must accept the current Hermes agent callback contract.
+    The agent now calls tool_progress_callback(event_type, name, preview, args, **kwargs).
+    If the WebUI bridge only accepts (name, preview, args), live tool updates silently vanish.
+    """
+    src = (REPO_ROOT / "api/streaming.py").read_text()
+    assert "def on_tool(*cb_args, **cb_kwargs):" in src, \
+        "streaming.py must accept variable callback args for tool progress events"
+    assert "reasoning_callback=on_reasoning" in src, \
+        "streaming.py must wire the agent's reasoning callback into the SSE bridge"
+    assert "put('tool_complete'" in src or 'put("tool_complete"' in src, \
+        "streaming.py must emit live tool completion SSE events"
+
+
+def test_messages_js_supports_live_reasoning_and_tool_completion(cleanup_test_sessions):
+    """R18: messages.js must render live reasoning and react to tool completion events.
+    Without these handlers, the operator only sees generic Thinking… or nothing
+    until the final done snapshot redraws the whole turn.
+    """
+    src = (REPO_ROOT / "static/messages.js").read_text()
+    assert "let reasoningText=''" in src, \
+        "messages.js must track streamed reasoning text separately from assistant text"
+    assert "source.addEventListener('reasoning'" in src or 'source.addEventListener("reasoning"' in src, \
+        "messages.js must listen for live reasoning SSE events"
+    assert "source.addEventListener('tool_complete'" in src or 'source.addEventListener("tool_complete"' in src, \
+        "messages.js must listen for live tool completion SSE events"
+    assert "function _parseStreamState()" in src, \
+        "messages.js must parse live stream state into reasoning + visible answer"
+
+
+def test_ui_js_can_upgrade_thinking_spinner_into_live_reasoning_card(cleanup_test_sessions):
+    """R19: ui.js must be able to replace the placeholder thinking spinner with
+    streamed reasoning text while a turn is in progress.
+    """
+    src = (REPO_ROOT / "static/ui.js").read_text()
+    assert "function _thinkingMarkup(text='')" in src or 'function _thinkingMarkup(text="")' in src, \
+        "ui.js must centralize thinking row markup so it can switch between spinner and live text"
+    assert "function updateThinking(text=''){appendThinking(text);}" in src or 'function updateThinking(text=""){appendThinking(text);}' in src, \
+        "ui.js must expose an updateThinking helper for live reasoning rendering"
+
+
+# ── R17: Stack traces must not leak to clients in 500 responses ────────────
 
 def test_500_response_has_no_trace_field():
     """R16: HTTP 500 responses must not include a 'trace' field.