Fix Multi Agent Mode

ui/qvac/index.js

@@ -281,12 +281,31 @@ export function createQvacView(opts) {
     })
   }
 
-  const swarm = createSwarmPanel(opts.els.agentBar, {
+  // Swarm mounts into each assistant message (transcript), not a sticky chrome bar
+  const swarm = createSwarmPanel(null, {
     onStop: () => stopInference(),
   })
 
-  function clearAgentBar() {
-    swarm.hide()
+  /** Release live swarm controller only; past message nodes stay in the transcript. */
+  function releaseLiveSwarm() {
+    if (typeof swarm.release === 'function') swarm.release()
+    else swarm.hide()
+  }
+
+  /**
+   * Mount a swarm host inside the assistant message (above answer body).
+   * @param {{ div: HTMLElement, body: HTMLElement }|null|undefined} assistantUi
+   * @returns {HTMLElement|null}
+   */
+  function ensureSwarmSlot(assistantUi) {
+    if (!assistantUi?.div) return null
+    let slot = assistantUi.div.querySelector('.qvac-swarm-slot')
+    if (slot) return /** @type {HTMLElement} */ (slot)
+    slot = document.createElement('div')
+    slot.className = 'qvac-swarm-slot'
+    // Insert before message body so the answer streams below the swarm card
+    assistantUi.div.insertBefore(slot, assistantUi.body || assistantUi.div.firstChild)
+    return slot
   }
 
   /** Finish onboarding in tools-only mode (no model download). */
@@ -814,7 +833,10 @@ export function createQvacView(opts) {
     let acc = ''
     let thinkingAcc = ''
     setStatus('Thinking…', 'busy')
-    clearAgentBar()
+    releaseLiveSwarm()
+    // Hide legacy sticky bar if still in DOM
+    opts.els.agentBar?.classList.add('hidden')
+    opts.els.agentBar && (opts.els.agentBar.innerHTML = '')
     try {
       // History for the model: clean answers only (no think tags)
       const hist = messages
@@ -833,18 +855,19 @@ export function createQvacView(opts) {
         subAgents: settings().qvacSubAgents !== false,
         maxSubAgents: Number(settings().qvacMaxSubAgents) || 3,
       }
-
       if (
         shouldUseSubAgents(text, prefs) &&
         opts.isConnected?.() &&
         !engine.isAborted?.()
       ) {
         setStatus('Multi-agent swarm…', 'busy')
+        const swarmSlot = ensureSwarmSlot(assistantUi)
         if (streamBody) {
           streamBody.innerHTML = formatMdLite(
-            '_Swarm is running — watch specialists above while tools gather live data…_'
+            '_Specialists are gathering live data in this turn…_'
           )
         }
+        followMessagesScroll({ force: true })
         const pack = await runSubAgents({
           tools,
           query: text,
@@ -857,8 +880,10 @@ export function createQvacView(opts) {
                 label: s.label,
                 description: s.description,
               })),
-              text
+              text,
+              { mount: swarmSlot }
             )
+            followMessagesScroll({ force: true })
           },
           onAgent: (ev) => {
             if (engine.isAborted?.()) return
@@ -866,6 +891,7 @@ export function createQvacView(opts) {
             if (ev.status === 'start' || ev.status === 'running') {
               setStatus(`${ev.label}…`, 'busy')
             }
+            followMessagesScroll()
           },
         })
         if (engine.isAborted?.()) {
@@ -876,6 +902,10 @@ export function createQvacView(opts) {
           return
         }
         swarm.complete('All agents done — synthesizing answer…')
+        // Collapse into transcript so the answer can stream below without covering chat
+        swarm.collapse(
+          `${pack.agents?.length || 0} specialist${(pack.agents?.length || 0) === 1 ? '' : 's'} · complete`
+        )
         completeOpts.agentBriefing = pack.contextText
         completeOpts.agentFallbackText = synthesizeFromAgents(text, pack)
         completeOpts.agentToolCalls = pack.agents.map((a) => ({
@@ -896,6 +926,7 @@ export function createQvacView(opts) {
           assistantUi.div.appendChild(renderToolChips(toolLog))
         }
         setStatus('Synthesizing…', 'busy')
+        followMessagesScroll({ force: true })
       }
 
       if (engine.isAborted?.()) {
@@ -1005,8 +1036,11 @@ export function createQvacView(opts) {
   }
 
   function newChat() {
+    releaseLiveSwarm()
     messages = []
     if (opts.els.messages) opts.els.messages.innerHTML = ''
+    opts.els.agentBar?.classList.add('hidden')
+    if (opts.els.agentBar) opts.els.agentBar.innerHTML = ''
     appendMsg(
       'assistant',
       'New chat. Ask about host health, metrics, anomalies, or processes.'