diff --git a/src/adapters/openai-chat/summary-budget.ts b/src/adapters/openai-chat/summary-budget.ts
index 2e044a46421..783e65b8f4c 100644
--- a/src/adapters/openai-chat/summary-budget.ts
+++ b/src/adapters/openai-chat/summary-budget.ts
@@ -37,6 +37,16 @@ function isTinyCap(value: unknown): value is number {
   return typeof value === "number" && Number.isInteger(value) && value >= 1 && value <= 1024;
 }
 
+function hasConversationEnvelope(transcript: string): boolean {
+  // Case-insensitive tag search in place: request bodies can reach hundreds of MiB, so
+  // lowercasing a copy would roughly double peak memory for an already large request.
+  const opening = /<conversation>/i.exec(transcript);
+  if (opening === null) return false;
+  const closing = /<\/conversation>/gi;
+  closing.lastIndex = opening.index + opening[0].length;
+  return closing.test(transcript);
+}
+
 /**
  * Aside's emergency checkpoint is a standalone summary, not an ordinary short answer (#5465).
  * Runs at the physical Chat destination, after all combo effort overrides, so `effort` is the
@@ -64,7 +74,7 @@ export function protectGlmSummaryBudget(
   const transcript = textContent(user.content);
   if (instruction === undefined || transcript === undefined) return false;
   if (!/\b(?:summari[sz](?:e|ation|er|ing)|summary|checkpoint)\b/i.test(instruction)) return false;
-  if (!/<conversation>[\s\S]*<\/conversation>/i.test(transcript)) return false;
+  if (!hasConversationEnvelope(transcript)) return false;
   if (transcript.length < MIN_CHECKPOINT_TRANSCRIPT_CHARS) return false;
   if (isTinyCap(body.max_tokens)) body.max_tokens = PROTECTED_SUMMARY_CAP;
   if (isTinyCap(body.max_completion_tokens)) body.max_completion_tokens = PROTECTED_SUMMARY_CAP;
diff --git a/tests/adapters/openai/openai-chat-glm-summary.test.ts b/tests/adapters/openai/openai-chat-glm-summary.test.ts
index b892d3a04f5..ea8ebcb7dac 100644
--- a/tests/adapters/openai/openai-chat-glm-summary.test.ts
+++ b/tests/adapters/openai/openai-chat-glm-summary.test.ts
@@ -1,5 +1,6 @@
-import { describe, expect, test } from "bun:test";
+import { describe, expect, spyOn, test } from "bun:test";
 import { buildOpenAIChatPassthroughRequest, createOpenAIChatAdapter } from "../../../src/adapters/openai-chat";
+import { protectGlmSummaryBudget } from "../../../src/adapters/openai-chat/summary-budget";
 import { chatCompletionsToResponsesBody } from "../../../src/chat/inbound";
 import { concreteComboRequestBody } from "../../../src/combos/request";
 import { parseRequest } from "../../../src/responses/parser";
@@ -84,6 +85,51 @@ describe("GLM summary mitigation stays inside its boundary (#5953 review)", () =
     const short = [messages[0], { role: "user", content: "<conversation>User: hi.</conversation>" }];
     for (const body of bodies({ messages: short })) untouched(body);
   });
+  /** The only patterns detection may scan the transcript with: the two fixed tags. */
+  const FIXED_TAG_SCANS = ["<conversation>", "<\\/conversation>"];
+  /** Runs detection while counting scans/copies targeted at the transcript — deterministic, unlike a wall-clock bound. */
+  const detectWithProbe = (userContent: string) => {
+    const scans: string[] = [];
+    const copies: string[] = [];
+    const origExec = RegExp.prototype.exec;
+    const execProbe = spyOn(RegExp.prototype, "exec").mockImplementation(function (this: RegExp, str: string) {
+      if (str === userContent) scans.push(this.source);
+      return origExec.call(this, str);
+    });
+    const copyProbes = (["toLowerCase", "toUpperCase", "toLocaleLowerCase", "toLocaleUpperCase"] as const).map(name => {
+      const orig = String.prototype[name];
+      return spyOn(String.prototype, name).mockImplementation(function (this: unknown) {
+        if (String(this) === userContent) copies.push(name);
+        return orig.call(this as string);
+      });
+    });
+    try {
+      const result = protectGlmSummaryBudget({
+        model, max_tokens: 512,
+        messages: [messages[0], { role: "user", content: userContent }],
+      }, provider.baseUrl, "max");
+      // A call count cannot bound work inside one scan — pinning every transcript scan to a
+      // fixed tag is what fails the original greedy `[\s\S]*` pass, not the count alone.
+      for (const source of scans) expect(FIXED_TAG_SCANS).toContain(source);
+      return { result, scans, copies };
+    } finally {
+      execProbe.mockRestore();
+      for (const probe of copyProbes) probe.mockRestore();
+    }
+  };
+  test("rejects many unmatched conversation openings with a bounded number of transcript scans", () => {
+    const { result, scans, copies } = detectWithProbe("<conversation>".repeat(32_000));
+    expect(result).toBeFalse();
+    expect(scans.length).toBeLessThanOrEqual(2);
+    expect(copies).toEqual([]);
+  });
+  test("detects an uppercase checkpoint in place, without a transcript-sized normalized copy", () => {
+    const transcript = `<CONVERSATION>${"LOREM IPSUM DOLOR SIT AMET.\n".repeat(200)}</CONVERSATION>`;
+    const { result, scans, copies } = detectWithProbe(transcript);
+    expect(result).toBeTrue();
+    expect(scans.length).toBeLessThanOrEqual(2);
+    expect(copies).toEqual([]);
+  });
   test.each(["low", "medium"])("an effective %s effort is not overridden", effort => {
     for (const body of bodies({ reasoning_effort: effort })) untouched(body, effort);
   });