diff --git a/packages/chat-ui/src/streaming-reply.ts b/packages/chat-ui/src/streaming-reply.ts
index 730c6e958..fe352d013 100644
--- a/packages/chat-ui/src/streaming-reply.ts
+++ b/packages/chat-ui/src/streaming-reply.ts
@@ -198,9 +198,17 @@ export function hydrateStreamingReplyFromTurn(
   return awaiting(runningTurn.textSnapshot ?? "");
 }
 
-/** How long an empty pending reply may sit with no tokens before the
- * indicator clears itself — the backstop for a turn whose stream events
- * never arrive (agent down, SSE dropped mid-reconnect). */
+/** How long a turn may go without a single new token before it's declared
+ * dead — the backstop for both a turn whose stream events never arrive at
+ * all (agent down, SSE dropped mid-reconnect) and one that starts streaming
+ * and then stalls (model OOM, dropped Ollama connection, sidecar crash: all
+ * routine with local models, CL-6486). This measures the gap *since the
+ * last token*, not total turn duration — a healthy local model can
+ * legitimately run 200s+ end to end (round 1 measured ~216s on
+ * `qwen3.8:27b`), so a total-elapsed timeout would fire on working replies.
+ * 120s of dead air with zero new output, on the other hand, is never a
+ * healthy sign even for a slow model — it's long past any single decode
+ * step, tool round-trip, or reconnect a client is expected to ride out. */
 const PENDING_REPLY_CLEAR_MS = 120_000;
 
 /** Floor on how long the empty typing pulse stays up after it first
@@ -244,7 +252,16 @@ export function useStreamingReply(
   }, [workbenchId]);
 
   useEffect(() => {
-    if (!isPendingReply(streamingReply)) return;
+    // Arm for the whole "awaiting" phase, not just its tokenless prefix
+    // (CL-6486): a token still growing the reply is not evidence the turn
+    // is alive forever, only that it was alive as of that token. Every
+    // token produces a new `streamingReply` object (see `awaiting`), so
+    // this effect's own dependency below tears down the previous timer and
+    // arms a fresh one on each token — the window this constructs is
+    // therefore inter-token silence, never total elapsed time.
+    if (streamingReply === null || streamingReply.phase !== "awaiting") {
+      return;
+    }
     const timer = setTimeout(() => {
       // The dependency below re-arms this effect (clearing this exact
       // timer) the instant `streamingReply` changes, so this callback only
diff --git a/packages/chat-ui/test/use-streaming-reply.test.tsx b/packages/chat-ui/test/use-streaming-reply.test.tsx
index 7e4a90878..eb174ed7f 100644
--- a/packages/chat-ui/test/use-streaming-reply.test.tsx
+++ b/packages/chat-ui/test/use-streaming-reply.test.tsx
@@ -163,17 +163,35 @@ describe("useStreamingReply's reply-timeout backstop (CL-6252 #6)", () => {
     harness.unmount();
   });
 
-  test("a token arriving before the backstop never marks it timed out", async () => {
+  test("a token resets the backstop window instead of clearing it (CL-6486)", async () => {
     const harness = mount("chan_a", 30);
     harness.awaitReply();
+    await harness.settle(20);
+    // A token arrives just before the original deadline would have fired —
+    // it must push the deadline out from here, not merely have prevented
+    // the earlier one.
     harness.send("chat.agent", delta("Hi"));
-
-    await harness.settle(60);
+    await harness.settle(20);
     expect(harness.get()).toEqual({ phase: "awaiting", text: "Hi" });
     expect(harness.timedOut()).toBe(false);
     harness.unmount();
   });
 
+  test("silence after a token still times out a mid-stream stall (CL-6486)", async () => {
+    const harness = mount("chan_a", 30);
+    harness.awaitReply();
+    harness.send("chat.agent", delta("Hi"));
+    expect(harness.get()).toEqual({ phase: "awaiting", text: "Hi" });
+
+    // The model died mid-sentence: no further tokens, no terminal event.
+    // The backstop must still fire off the last token, not stay armed
+    // forever just because *some* text streamed.
+    await harness.settle(60);
+    expect(harness.get()).toBeNull();
+    expect(harness.timedOut()).toBe(true);
+    harness.unmount();
+  });
+
   test("switching workbenches clears a stale timed-out flag from the one just left", async () => {
     const harness = mount("chan_a", 30);
     harness.awaitReply();