diff --git a/packages/coding-agent/src/cli/bench-cli.ts b/packages/coding-agent/src/cli/bench-cli.ts
index 36aedd286..836890615 100644
--- a/packages/coding-agent/src/cli/bench-cli.ts
+++ b/packages/coding-agent/src/cli/bench-cli.ts
@@ -52,6 +52,7 @@ const BENCH_PROMPT = benchPrompt.trim();
 const UTF8_ENCODER = new TextEncoder();
 const UTF8_DECODER = new TextDecoder();
 const CACHE_PREFIX_CHUNK = cachePrefixChunk;
+const CACHE_PREFIX_PLACEHOLDER = "__OMP_CACHE_BENCH_RAW_PREFIX__";
 const CACHE_PREFIX_CHUNK_BYTES = UTF8_ENCODER.encode(CACHE_PREFIX_CHUNK).byteLength;
 const RESPONSE_CACHE_STATUS_HEADERS = ["cf-aig-cache-status"] as const;
 
@@ -341,6 +342,19 @@ function generatedCachePrefix(bytes: number): string {
 	return truncateUtf8(CACHE_PREFIX_CHUNK.repeat(Math.ceil(bytes / CACHE_PREFIX_CHUNK_BYTES)), bytes);
 }
 
+function renderCacheBenchmarkPrefix(prefix: string, namespace: string): string {
+	const rendered = prompt.render(cachePrefixTemplate, {
+		prefix: CACHE_PREFIX_PLACEHOLDER,
+		namespace,
+	});
+	if (!rendered.includes(CACHE_PREFIX_PLACEHOLDER)) {
+		throw new Error("Cache benchmark prefix template is missing its raw prefix placeholder");
+	}
+	// Render the static wrapper first, then inject caller bytes so prompt
+	// normalization cannot trim spaces or collapse blank lines in prefix files.
+	return rendered.replace(CACHE_PREFIX_PLACEHOLDER, prefix);
+}
+
 async function resolveCachePrefix(
 	flags: BenchCommandArgs["flags"],
 	readTextFile: (path: string, maxBytes: number) => Promise<string>,
@@ -821,9 +835,7 @@ export async function runBenchCommand(command: BenchCommandArgs, deps: BenchDepe
 					async (pairIndex): Promise<BenchCachePairReport> => {
 						const cacheNamespace = randomSessionId();
 						const promptCacheKey = `bench-cache:${cacheNamespace}`;
-						const stablePrefix = prompt
-							.render(cachePrefixTemplate, { prefix: cachePrefix!, namespace: cacheNamespace })
-							.trimEnd();
+						const stablePrefix = renderCacheBenchmarkPrefix(cachePrefix!, cacheNamespace);
 						const coldSuffix = prompt.render(cacheSuffixTemplate, { variant: "A" }).trim();
 						const warmSuffix = prompt.render(cacheSuffixTemplate, { variant: "B" }).trim();
 						const coldCapture: CacheRequestCapture = {
diff --git a/packages/coding-agent/test/bench-cache.test.ts b/packages/coding-agent/test/bench-cache.test.ts
index 5c04dbd99..26cad8383 100644
--- a/packages/coding-agent/test/bench-cache.test.ts
+++ b/packages/coding-agent/test/bench-cache.test.ts
@@ -494,6 +494,45 @@ describe("bench cache mode", () => {
 		expect(stablePrefixes[0]).not.toContain("\uFFFD");
 	});
 
+	it("preserves significant whitespace from the default prefix-file reader", async () => {
+		const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), "omp-bench-cache-prefix-whitespace-"));
+		const prefixPath = path.join(tempDir, "prefix.txt");
+		const exactPrefix = "line one  \n\n\nline two\t\n";
+		const stablePrefixes: string[] = [];
+		await Bun.write(prefixPath, exactPrefix);
+		try {
+			await runBenchCommand(
+				{
+					models: ["openai/gpt-cache-test"],
+					flags: { cache: true, cachePrefixFile: prefixPath, cachePrefixBytes: 128, json: true },
+				},
+				{
+					createRuntime: async () => ({ modelRegistry: registry, close: () => {} }),
+					randomSessionId: (() => {
+						let id = 0;
+						return () => `session-${++id}`;
+					})(),
+					writeStdout: () => {},
+					writeStderr: () => {},
+					setExitCode: () => {},
+					streamSimple: (_model, context, options) => {
+						stablePrefixes.push(context.messages[0]?.content as string);
+						void options?.onPayload?.({ input: context.messages });
+						return streamWithMessage(successfulMessage(0, 0));
+					},
+					stdoutIsTTY: false,
+				},
+			);
+		} finally {
+			await fs.rm(tempDir, { recursive: true, force: true });
+		}
+
+		expect(stablePrefixes).toHaveLength(2);
+		expect(stablePrefixes[0]).toBe(stablePrefixes[1]);
+		expect(stablePrefixes[0]?.startsWith(exactPrefix)).toBe(true);
+		expect(stablePrefixes[0]?.slice(exactPrefix.length)).toStartWith("\n\nPrompt-cache benchmark namespace:");
+	});
+
 	it("does not turn zero cache counters into a miss", async () => {
 		const summary = await runBenchCommand(
 			{ models: ["openai/gpt-cache-test"], flags: { cache: true, json: true } },