diff --git a/frontend/src/app/pages/AgentChat/tool-ui/hasWorkAfterLatestShowUi.test.ts b/frontend/src/app/pages/AgentChat/tool-ui/hasWorkAfterLatestShowUi.test.ts
new file mode 100644
index 00000000..73ff2f05
--- /dev/null
+++ b/frontend/src/app/pages/AgentChat/tool-ui/hasWorkAfterLatestShowUi.test.ts
@@ -0,0 +1,46 @@
+// Run: node --test (via frontend/scripts/run-tests.mjs)
+//
+// ENG-272. The collapsed pill hid a plan/progress widget whenever the session was running, on the
+// assumption that a mid-turn plan is stale. That is true when work followed it and false when the
+// widget IS the newest thing the agent said, which is exactly what a long-running skill does when it
+// re-emits its tracker. Result: the minimized view showed older state all day. Staleness is now
+// "did real work land after this widget", which these pin.
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { hasWorkAfterLatestShowUi } from './showUiPayload.ts';
+
+const showUi = (id: string) => ({ role: 'tool_call', content: { id, tool: 'mcp__openswarm-core__ShowUI', input: { component: 'plan', props: {} } } });
+const bash = (id: string) => ({ role: 'tool_call', content: { id, tool: 'Bash', input: { command: 'ls' } } });
+const say = (text: string) => ({ role: 'assistant', content: text });
+const user = (text: string) => ({ role: 'user', content: text });
+
+test('a widget that is the newest thing is NOT stale', () => {
+  assert.equal(hasWorkAfterLatestShowUi([user('go'), bash('b1'), showUi('u1')]), false);
+});
+
+test('a re-emitted widget is not stale either, which is the whole bug', () => {
+  // The send-text case: work, widget, more work, widget again. The newest widget stands.
+  assert.equal(hasWorkAfterLatestShowUi([showUi('u1'), bash('b1'), say('sent the batch'), showUi('u2')]), false);
+});
+
+test('a tool call after the widget DOES make it stale', () => {
+  assert.equal(hasWorkAfterLatestShowUi([showUi('u1'), bash('b1')]), true);
+});
+
+test('a spoken answer after the widget makes it stale', () => {
+  assert.equal(hasWorkAfterLatestShowUi([showUi('u1'), say('all done')]), true);
+});
+
+test('an empty assistant message is not work', () => {
+  assert.equal(hasWorkAfterLatestShowUi([showUi('u1'), say('   ')]), false);
+});
+
+test('a user message after the widget is not agent work', () => {
+  // The user typing does not make the agent's own widget stale; the agent has not acted yet.
+  assert.equal(hasWorkAfterLatestShowUi([showUi('u1'), user('any update?')]), false);
+});
+
+test('no widget at all reports no work-after', () => {
+  assert.equal(hasWorkAfterLatestShowUi([user('go'), bash('b1')]), true);
+  assert.equal(hasWorkAfterLatestShowUi([]), false);
+});
diff --git a/frontend/src/app/pages/AgentChat/tool-ui/showUiPayload.ts b/frontend/src/app/pages/AgentChat/tool-ui/showUiPayload.ts
index 3f9c0f5f..7645a6f9 100644
--- a/frontend/src/app/pages/AgentChat/tool-ui/showUiPayload.ts
+++ b/frontend/src/app/pages/AgentChat/tool-ui/showUiPayload.ts
@@ -125,6 +125,26 @@ export function extractPendingAskUi(messages: Array<{ id: string; role: string;
   return null;
 }
 
+/** True when real agent work landed AFTER the newest ShowUI, which is what makes that widget stale.
+ *
+ * A tool call or a spoken answer after the widget means the plan it drew has moved on. A widget that
+ * is still the last thing in the transcript is the agent's CURRENT statement, however it is named,
+ * and a long-running skill re-emitting its tracker lands exactly there (ENG-272).
+ */
+export function hasWorkAfterLatestShowUi(messages: Array<{ role: string; content: unknown }>): boolean {
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const msg = messages[i];
+    if (msg.role === 'tool_call') {
+      const body = (typeof msg.content === 'object' && msg.content !== null ? msg.content : {}) as { tool?: unknown };
+      // The ShowUI itself ends the scan: anything below it is older than the widget.
+      if (/(^|__)ShowUI$/.test(String(body.tool || ''))) return false;
+      return true;
+    }
+    if (msg.role === 'assistant' && typeof msg.content === 'string' && msg.content.trim()) return true;
+  }
+  return false;
+}
+
 export function extractLatestShowUi(messages: Array<{ role: string; content: unknown }>): ShowUiPayload | null {
   for (let i = messages.length - 1; i >= 0; i--) {
     const msg = messages[i];
diff --git a/frontend/src/app/pages/Dashboard/cards/AgentCard.tsx b/frontend/src/app/pages/Dashboard/cards/AgentCard.tsx
index 360fe5cb..ae63bd2c 100644
--- a/frontend/src/app/pages/Dashboard/cards/AgentCard.tsx
+++ b/frontend/src/app/pages/Dashboard/cards/AgentCard.tsx
@@ -43,7 +43,7 @@ import { openCardContextMenu, isNativeMenuTarget } from '../desktop/openCardCont
 import { agentCardMenuRows } from './agentCardMenuRows';
 import { extractLatestTodos } from '../desktop/agentTodos';
 import { extractLiveSteps } from '../desktop/agentLiveSteps';
-import { extractLatestShowUi, extractPendingAskUi, freezeIfDone, artifactName } from '@/app/pages/AgentChat/tool-ui/showUiPayload';
+import { extractLatestShowUi, extractPendingAskUi, freezeIfDone, artifactName, hasWorkAfterLatestShowUi } from '@/app/pages/AgentChat/tool-ui/showUiPayload';
 import { useDragEndBackstops } from '../hooks/interaction/useDragEndBackstops';
 import { useBrowserPillShot } from '../desktop/useBrowserPillShot';
 import { useAppDispatch, useAppSelector } from '@/shared/hooks';
@@ -724,11 +724,18 @@ const AgentCard: React.FC<Props> = ({
     [session.messages, session.status],
   );
   const pillArtifact = useMemo(() => {
-    const artifact = extractLatestShowUi(session.messages || []);
+    const msgs = session.messages || [];
+    const artifact = extractLatestShowUi(msgs);
     if (!artifact) return null;
-    // A plan/progress widget posted mid-turn goes stale the moment work continues; while running the
-    // pill prefers living surfaces (browser shot, live steps, todos). Answer widgets still win.
-    if (session.status === 'running' && /plan|progress/i.test(artifactName(artifact))) return null;
+    // A plan/progress widget posted mid-turn goes stale the moment work continues, so while running
+    // the pill prefers living surfaces (browser shot, live steps, todos). But "stale" means NOTHING
+    // HAPPENED SINCE, not "it is named plan": a skill whose whole job is a live-updating tracker
+    // re-emits it as the newest thing in the transcript, and hiding that left the collapsed view
+    // showing older state all day (ENG-272). So only suppress it when real work followed it.
+    if (session.status === 'running' && /plan|progress/i.test(artifactName(artifact))
+        && hasWorkAfterLatestShowUi(msgs)) {
+      return null;
+    }
     return freezeIfDone(artifact, session.status === 'running');
   }, [session.messages, session.status]);
   const pillAskPair = useMemo(