From a7661cb4f45ee2d9a989f8a61fe1bea5597a0ddf Mon Sep 17 00:00:00 2001
From: Jaakko Husso <jaakko@n8n.io>
Date: Thu, 27 Aug 2026 12:47:15 +0000
Subject: [PATCH] test(core): Support tool-result cleanup in the eval stub
 workspace (no-changelog) (#37202)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../__tests__/stub-workspace.test.ts          | 29 +++++++++++++++++++
 .../evaluations/harness/stub-workspace.ts     | 25 ++++++++++++++--
 2 files changed, 52 insertions(+), 2 deletions(-)

diff --git a/packages/@n8n/instance-ai/evaluations/__tests__/stub-workspace.test.ts b/packages/@n8n/instance-ai/evaluations/__tests__/stub-workspace.test.ts
index 78a810fdb5d..70d19deddf5 100644
--- a/packages/@n8n/instance-ai/evaluations/__tests__/stub-workspace.test.ts
+++ b/packages/@n8n/instance-ai/evaluations/__tests__/stub-workspace.test.ts
@@ -28,6 +28,35 @@ describe('createStubWorkspace', () => {
 		await expect(filesystem().readFile('missing.ts')).rejects.toThrow('No such file');
 	});
 
+	// The runtime offloads oversized tool results to `tool-results/runs/<hash>/`
+	// and clears that directory at the end of every run, calling the filesystem
+	// directly rather than through the exposed tools.
+	describe('tool-result cleanup', () => {
+		const runDirectory = 'tool-results/runs/run-hash';
+
+		it('reports a directory as existing once a file sits under it', async () => {
+			const fs = filesystem();
+			await fs.writeFile(`${runDirectory}/call.result.json`, '{}');
+
+			await expect(fs.exists(runDirectory)).resolves.toBe(true);
+		});
+
+		it('reports a directory that was never written as missing', async () => {
+			await expect(filesystem().exists(runDirectory)).resolves.toBe(false);
+		});
+
+		it('removes every file under the directory it is given, and nothing else', async () => {
+			const fs = filesystem();
+			await fs.writeFile(`${runDirectory}/call.result.json`, '{}');
+			await fs.writeFile('tool-results/runs/other-hash/call.result.json', '{}');
+
+			await fs.rmdir(runDirectory);
+
+			await expect(fs.exists(runDirectory)).resolves.toBe(false);
+			await expect(fs.exists('tool-results/runs/other-hash')).resolves.toBe(true);
+		});
+	});
+
 	it('offers only the tools production exposes', () => {
 		expect(
 			createStubWorkspace()
diff --git a/packages/@n8n/instance-ai/evaluations/harness/stub-workspace.ts b/packages/@n8n/instance-ai/evaluations/harness/stub-workspace.ts
index 1df3ef8d61f..49a859d511e 100644
--- a/packages/@n8n/instance-ai/evaluations/harness/stub-workspace.ts
+++ b/packages/@n8n/instance-ai/evaluations/harness/stub-workspace.ts
@@ -4,6 +4,11 @@
 // Without a workspace, `build-workflow` fails its source read with a
 // `code_fixable` remediation telling the agent to write the file with
 // `workspace_write_file` — a tool that only exists when a workspace is attached.
+//
+// Tool exposure is filtered to CORE_WORKSPACE_TOOL_NAMES, but the runtime also
+// calls the filesystem directly for its own bookkeeping — offloading oversized
+// tool results and clearing them at the end of every run — so those operations
+// are implemented rather than rejected.
 // ---------------------------------------------------------------------------
 
 import {
@@ -58,13 +63,29 @@ class InMemoryWorkspaceFilesystem implements WorkspaceFilesystem {
 		await Promise.resolve();
 	}
 
+	async exists(path: string): Promise<boolean> {
+		const key = relativePath(path);
+		if (this.files.has(key)) return await Promise.resolve(true);
+		const prefix = `${key}/`;
+		for (const file of this.files.keys()) {
+			if (file.startsWith(prefix)) return await Promise.resolve(true);
+		}
+		return await Promise.resolve(false);
+	}
+
+	async rmdir(path: string): Promise<void> {
+		const prefix = `${relativePath(path)}/`;
+		for (const file of this.files.keys()) {
+			if (file.startsWith(prefix)) this.files.delete(file);
+		}
+		await Promise.resolve();
+	}
+
 	appendFile = unreachable;
 	deleteFile = unreachable;
 	copyFile = unreachable;
 	moveFile = unreachable;
-	rmdir = unreachable;
 	readdir = unreachable;
-	exists = unreachable;
 	stat = unreachable;
 }