From c85f32b4794bceacb813c9cec165fc7892db69dd Mon Sep 17 00:00:00 2001
From: Kyle Brown <kyleb@openai.com>
Date: Wed, 5 Aug 2026 16:16:49 +0000
Subject: [PATCH] Validate knowledge-base documents during dry-run scans

Reuse full scan document validation during SDK preflight so dry-run scans reject missing, unsupported, malformed, empty, and symlinked knowledge-base inputs before runtime initialization.
---
 sdk/typescript/src/api.ts           |  7 +++++++
 sdk/typescript/tests-ts/api.test.ts | 26 ++++++++++++++++++++++++++
 2 files changed, 33 insertions(+)

diff --git a/sdk/typescript/src/api.ts b/sdk/typescript/src/api.ts
index 8cabdbdf..ae6b0811 100644
--- a/sdk/typescript/src/api.ts
+++ b/sdk/typescript/src/api.ts
@@ -322,6 +322,13 @@ export class CodexSecurity {
       await realpath(tmpdir()),
       "temporary",
     );
+    if (options.knowledgeBasePaths?.length) {
+      const knowledgeBase = await prepareKnowledgeBase(
+        options.knowledgeBasePaths,
+        options.signal,
+      );
+      await knowledgeBase.cleanup();
+    }
     const configuration = await mergedCodexConfig(this.config);
     const model = scanModelConfiguration(configuration);
     validateScanCostLimit(options.maxCostUsd, model.model);
diff --git a/sdk/typescript/tests-ts/api.test.ts b/sdk/typescript/tests-ts/api.test.ts
index 996c6a2b..4ee5fbe4 100644
--- a/sdk/typescript/tests-ts/api.test.ts
+++ b/sdk/typescript/tests-ts/api.test.ts
@@ -847,9 +847,13 @@ describe("CodexSecurity orchestration", () => {
     const repository = join(root, "repository");
     const knowledgeBase = join(root, "threat-model.md");
     const invalidDocument = join(root, "broken.pdf");
+    const unsupportedDocument = join(root, "unsupported.exe");
+    const emptyDirectory = join(root, "empty");
     await mkdir(repository);
+    await mkdir(emptyDirectory);
     await writeFile(knowledgeBase, "# Threat model\nPublic API is in scope.\n");
     await writeFile(invalidDocument, "not a PDF");
+    await writeFile(unsupportedDocument, "not a supported document");
     let runtimeStarted = false;
     const client = new TestClient(
       {},
@@ -865,6 +869,28 @@ describe("CodexSecurity orchestration", () => {
     await expect(
       client.preflight(repository, { knowledgeBasePaths: [knowledgeBase] }),
     ).resolves.toMatchObject({ knowledgeBasePaths: [knowledgeBase] });
+    const invalidDocuments: Array<[string, string]> = [
+      [join(root, "missing.md"), "ENOENT"],
+      [unsupportedDocument, "Unsupported knowledge base document"],
+      [invalidDocument, "Cannot extract text from knowledge base PDF"],
+      [
+        emptyDirectory,
+        "Knowledge base directory contains no supported documents",
+      ],
+    ];
+    if (process.platform !== "win32") {
+      const linkedDocument = join(root, "linked.md");
+      await symlink(knowledgeBase, linkedDocument);
+      invalidDocuments.push([
+        linkedDocument,
+        "Knowledge base paths cannot be symbolic links",
+      ]);
+    }
+    for (const [path, message] of invalidDocuments) {
+      await expect(
+        client.preflight(repository, { knowledgeBasePaths: [path] }),
+      ).rejects.toThrow(message);
+    }
     await expect(
       client.run(repository, {
         knowledgeBasePaths: [join(root, "missing.md")],