diff --git a/studio/frontend/src/components/assistant-ui/code-token-cache.ts b/studio/frontend/src/components/assistant-ui/code-token-cache.ts
index 69339cb13f..59d4d489a0 100644
--- a/studio/frontend/src/components/assistant-ui/code-token-cache.ts
+++ b/studio/frontend/src/components/assistant-ui/code-token-cache.ts
@@ -17,7 +17,8 @@
 //                    storing frame N drops frames 1..N-1 of the same fence immediately. A
 //                    streaming reply occupies ONE entry while it streams, not one per window.
 //                    A size cap alone would not do this: the prefixes would sit in the cache
-//                    until later replies pushed them out one at a time.
+//                    until later replies pushed them out one at a time. It is ONE-DIRECTIONAL on
+//                    purpose; see the comment at the eviction itself.
 //   A CHARACTER CAP  tokens cost roughly 17x their source in retained heap, so a budget in
 //                    characters of source is a budget in megabytes. A count cap alone would let
 //                    64 large fences pin far more than 64 small ones.
@@ -87,7 +88,20 @@ export function createTokenCache<T>(options: TokenCacheOptions): TokenCache<T> {
       // the rest, so this walk is safe.
       for (const [otherKey, other] of entries) {
         if (other.group !== group) continue;
-        if (code.startsWith(other.code) || other.code.startsWith(code)) {
+        // ONE DIRECTION ONLY: drop what the new code EXTENDS, never what extends it.
+        //
+        // Evicting both ways live-locks the app. Two fences can be on screen at once where one is
+        // a prefix of the other, which is what a page showing the same snippet at two lengths
+        // does. Storing the long one would evict the short one, the short one's next render would
+        // miss and evict the long one, and so on: a miss returns null, which schedules an
+        // asynchronous tokenisation, whose callback re-renders, which misses again. Measured on
+        // the pair below before this was one-directional: two misses per round, forever, with the
+        // cache stuck at one entry. It hung a CI job for thirty minutes.
+        //
+        // The direction kept here is the one the memory win needs. Every earlier frame of a
+        // GROWING fence is a prefix of the current one, so this still collapses a streamed reply
+        // onto one entry. The reverse case, a source that SHRINKS, is left to the size cap.
+        if (code.startsWith(other.code)) {
           drop(otherKey, other);
         }
       }
diff --git a/studio/frontend/tests/code-token-cache.test.ts b/studio/frontend/tests/code-token-cache.test.ts
index 84edebfb5d..889df8bdcc 100644
--- a/studio/frontend/tests/code-token-cache.test.ts
+++ b/studio/frontend/tests/code-token-cache.test.ts
@@ -61,6 +61,31 @@ test("a streamed fence occupies one entry, not one per refresh window", () => {
   }
 });
 
+test("two fences on screen at once, one extending the other, both stay cached", () => {
+  // The live-lock this exists to stop. A miss returns null to the renderer, which schedules an
+  // asynchronous tokenisation, whose callback re-renders and asks again. If storing the longer
+  // fence evicted the shorter one AND storing the shorter one evicted the longer, two fences on
+  // screen would miss forever and the page would never go idle. That hung a CI job for thirty
+  // minutes before eviction was made one-directional.
+  const cache = roomy();
+  const short = "def f():\n    return 1\n";
+  const long = `${short}def g():\n    return 2\n`;
+  let misses = 0;
+  for (let round = 0; round < 8; round += 1) {
+    for (const code of [short, long]) {
+      if (cache.get(GROUP, code) === null) {
+        misses += 1;
+        cache.set(GROUP, code, `tokens-${code.length}`);
+      }
+    }
+  }
+  // Three at most: each fence missing once, plus one re-fetch if the first store evicted it.
+  assert.ok(misses <= 3, `the cache never settled: ${misses} misses over 8 rounds`);
+  assert.equal(cache.stats().entries, 2);
+  assert.equal(cache.get(GROUP, short), `tokens-${short.length}`);
+  assert.equal(cache.get(GROUP, long), `tokens-${long.length}`);
+});
+
 test("prefix eviction does not reach across groups", () => {
   const cache = roomy();
   cache.set(OTHER_GROUP, "const a", "ts-tokens");
diff --git a/studio/frontend/tests/mutants/code-token-cache-mutants.mjs b/studio/frontend/tests/mutants/code-token-cache-mutants.mjs
index 756b397128..dce50a8494 100644
--- a/studio/frontend/tests/mutants/code-token-cache-mutants.mjs
+++ b/studio/frontend/tests/mutants/code-token-cache-mutants.mjs
@@ -39,6 +39,14 @@ const MUTANTS = [
     from: "        if (code.startsWith(other.code) || other.code.startsWith(code)) {\n          drop(otherKey, other);\n        }",
     to: "        if (false) {\n          drop(otherKey, other);\n        }",
   },
+  {
+    // The shipped bug: evicting in BOTH directions live-locks two fences that are on screen at
+    // once when one extends the other.
+    name: "prefix eviction runs in both directions",
+    test: "two fences on screen at once, one extending the other, both stay cached",
+    from: "        if (code.startsWith(other.code)) {",
+    to: "        if (code.startsWith(other.code) || other.code.startsWith(code)) {",
+  },
   {
     name: "prefix eviction ignores the group",
     test: "prefix eviction does not reach across groups",
diff --git a/tests/studio/test_playwright_suites_run_in_ci.py b/tests/studio/test_playwright_suites_run_in_ci.py
index 4b607b9750..bfa7181993 100644
--- a/tests/studio/test_playwright_suites_run_in_ci.py
+++ b/tests/studio/test_playwright_suites_run_in_ci.py
@@ -40,6 +40,19 @@ NOT_IN_CI = {
     # can go wrong silently, the verdict in harness_failures, is driven without a
     # browser by test_autoscroll_harness_contract.py, which CI does run.
     "playwright_thread_weight.py",
+    # A measurement harness rather than a gate: it prints the retained-heap slope per streamed
+    # code fence that #9228 was sized from, and sets no budget. It needs a forced GC between
+    # samples, so it is Chromium-with-CDP only, and the sizes that make the slope mean anything
+    # (six 32 KB fences per arm across four arms, plus a tick-rate ladder) cost tens of minutes.
+    # Run by hand when that slope needs re-measuring. What can go wrong silently in it, a settle
+    # predicate that reports "finished" while highlighting is still queued, is covered without a
+    # browser by studio/frontend/tests/code-token-cache.test.ts, which CI does run.
+    "playwright_shiki_retention.py",
+    # The same, for frame timing: it accumulates 40 streamed replies to separate the arms by
+    # hundreds of megabytes of retained heap before recording a single frame, which is minutes per
+    # arm, and it reports a comparison rather than gating on it. Its result was a NULL (the
+    # retained heap costs bytes, not frames), so there is no budget here to defend.
+    "playwright_shiki_jank.py",
 }