@vellumai/credential-executor 0.10.3-dev.202606292229.153e9b2 → 0.10.3-dev.202606292327.80883ab

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/credential-executor",
-  "version": "0.10.3-dev.202606292229.153e9b2",
+  "version": "0.10.3-dev.202606292327.80883ab",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/command-workspace.test.ts

@@ -112,6 +112,30 @@ describe("stageInputs", () => {
     }
   });
 
+  test("concurrent stagings get isolated scratch directories", () => {
+    // run_authenticated_command relies on each invocation getting its own
+    // scratch dir so that interleaved commands cannot read or clobber each
+    // other's staged inputs/outputs. Same config, two stagings → distinct dirs.
+    const config: WorkspaceStageConfig = {
+      workspaceDir,
+      inputs: [{ workspacePath: "input.txt" }],
+      outputs: [],
+      secrets: new Set(),
+    };
+
+    const a = stageInputs(config);
+    const b = stageInputs(config);
+
+    try {
+      expect(a.scratchDir).not.toBe(b.scratchDir);
+      expect(existsSync(a.scratchDir)).toBe(true);
+      expect(existsSync(b.scratchDir)).toBe(true);
+    } finally {
+      cleanupScratchDir(a.scratchDir);
+      cleanupScratchDir(b.scratchDir);
+    }
+  });
+
   test("staged inputs are read-only", () => {
     const config: WorkspaceStageConfig = {
       workspaceDir,

package/src/__tests__/manage-secure-command-tool.test.ts

@@ -0,0 +1,134 @@
+/**
+ * Tests for the manage_secure_command_tool handler's operation serialization.
+ *
+ * The register path awaits a bundle download mid-handler. Without
+ * serialization, a concurrent unregister could run its "still in use?" check
+ * and bundle delete during that await — against a registry that doesn't yet
+ * reflect the in-flight registration — transiently deleting a bundle another
+ * caller is publishing or executing. The handler runs operations one-at-a-time
+ * to close that window.
+ */
+
+import { describe, expect, test } from "bun:test";
+
+import type { ManageSecureCommandTool } from "@vellumai/service-contracts/credential-rpc";
+
+import {
+  createManageSecureCommandToolHandler,
+  type ManageSecureCommandToolHandlerDeps,
+} from "../server.js";
+
+const CTX = { sessionId: "test-session" };
+
+function registerRequest(toolName: string): ManageSecureCommandTool {
+  return {
+    action: "register",
+    toolName,
+    bundleId: "bundle-1",
+    version: "1.0.0",
+    sourceUrl: "https://example.com/bundle.tgz",
+    sha256: "deadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeefdeadbeef0",
+    credentialHandle: "local_static:svc/key",
+    description: "a tool",
+    // publishBundle is mocked in these tests, so the manifest contents are
+    // irrelevant — only its presence matters for the required-field check.
+    secureCommandManifest: {} as unknown as ManageSecureCommandTool["secureCommandManifest"],
+  };
+}
+
+function unregisterRequest(toolName: string): ManageSecureCommandTool {
+  return { action: "unregister", toolName };
+}
+
+describe("manage_secure_command_tool serialization", () => {
+  test("a slow register does not let a concurrent unregister interleave", async () => {
+    const events: string[] = [];
+
+    let releaseDownload!: () => void;
+    const downloadGate = new Promise<void>((resolve) => {
+      releaseDownload = resolve;
+    });
+
+    const deps: ManageSecureCommandToolHandlerDeps = {
+      downloadBundle: async () => {
+        events.push("download:start");
+        await downloadGate;
+        events.push("download:end");
+        return Buffer.from("bundle-bytes");
+      },
+      publishBundle: () => {
+        events.push("publish");
+        return { success: true, deduplicated: false, bundlePath: "/tmp/bundle" };
+      },
+      registerTool: () => {
+        events.push("register");
+      },
+      unregisterTool: (toolName: string) => {
+        events.push(`unregister:${toolName}`);
+        return true;
+      },
+    };
+
+    const handler = createManageSecureCommandToolHandler(deps);
+
+    // Fire a register (which blocks in downloadBundle) then an unregister.
+    const registerPromise = handler(registerRequest("tool-a"), CTX);
+    const unregisterPromise = handler(unregisterRequest("tool-b"), CTX);
+
+    // Let the event loop run: the register has reached its download await, and
+    // the unregister must be queued behind it — its delete must not have run.
+    await new Promise((resolve) => setTimeout(resolve, 0));
+    expect(events).toEqual(["download:start"]);
+
+    // Release the download; both operations complete in order.
+    releaseDownload();
+    const [registerResult, unregisterResult] = await Promise.all([
+      registerPromise,
+      unregisterPromise,
+    ]);
+
+    expect(registerResult.success).toBe(true);
+    expect(unregisterResult.success).toBe(true);
+
+    // The unregister ran only after the register fully completed.
+    expect(events).toEqual([
+      "download:start",
+      "download:end",
+      "publish",
+      "register",
+      "unregister:tool-b",
+    ]);
+  });
+
+  test("a rejected operation does not break serialization for later ones", async () => {
+    const events: string[] = [];
+
+    const deps: ManageSecureCommandToolHandlerDeps = {
+      downloadBundle: async () => {
+        throw new Error("network down");
+      },
+      publishBundle: () => ({
+        success: true,
+        deduplicated: false,
+        bundlePath: "/tmp/bundle",
+      }),
+      registerTool: () => {},
+      unregisterTool: (toolName: string) => {
+        events.push(`unregister:${toolName}`);
+        return true;
+      },
+    };
+
+    const handler = createManageSecureCommandToolHandler(deps);
+
+    // First op fails inside the handler (download error → structured failure),
+    // second op must still run.
+    const first = await handler(registerRequest("tool-a"), CTX);
+    const second = await handler(unregisterRequest("tool-b"), CTX);
+
+    expect(first.success).toBe(false);
+    expect(first.error?.code).toBe("DOWNLOAD_FAILED");
+    expect(second.success).toBe(true);
+    expect(events).toEqual(["unregister:tool-b"]);
+  });
+});

package/src/managed-main.ts

@@ -122,7 +122,7 @@ function buildHandlers(
   apiKeyRef: ApiKeyRef,
   assistantIdRef: AssistantIdRef,
   secureKeyBackend: SecureKeyBackend,
-): { handlers: RpcHandlerRegistry; temporaryGrantStore: TemporaryGrantStore } {
+): RpcHandlerRegistry {
   // -- Grant stores ----------------------------------------------------------
   const persistentGrantStore = new PersistentGrantStore(
     getCesGrantsDir("managed"),
@@ -411,7 +411,7 @@ function buildHandlers(
     return { results };
   }) as (typeof handlers)[string];
 
-  return { handlers, temporaryGrantStore };
+  return handlers;
 }
 
 // ---------------------------------------------------------------------------
@@ -668,10 +668,12 @@ async function main(): Promise<void> {
   // `unregister` miss a tool registered in an earlier session and orphan its
   // bundle.
   //
-  // The in-memory temporary-grant store is the exception: `allow_once` /
-  // `allow_10m` grants are keyed by proposal hash only (not session), so they
-  // would otherwise leak ephemeral approvals across sessions. It is cleared at
-  // the end of every session below so a reconnecting assistant must re-prompt.
+  // The in-memory temporary-grant store is also process-scoped, and (unlike
+  // before) is NOT cleared between sessions: ephemeral approvals are shared
+  // across all of a daemon's connections for the process lifetime. Its own
+  // semantics keep this safe — `allow_once` is consumed on first use,
+  // `allow_10m` expires by wall-clock TTL, `allow_conversation` is cleared when
+  // its conversation ends (see the serve loop below).
   //
   // The mutable refs carry the handshake-provided API key and assistant ID;
   // handlers read them at call time. These don't vary across a daemon's
@@ -680,7 +682,7 @@ async function main(): Promise<void> {
   // at call time for audit attribution).
   const apiKeyRef: ApiKeyRef = { current: "" };
   const assistantIdRef: AssistantIdRef = { current: "" };
-  const { handlers, temporaryGrantStore } = buildHandlers(
+  const handlers = buildHandlers(
     apiKeyRef,
     assistantIdRef,
     secureKeyBackend,
@@ -780,14 +782,16 @@ async function main(): Promise<void> {
 
     rpcConnected = false;
 
-    // Drop all ephemeral approvals when the session ends. `allow_once` /
-    // `allow_10m` grants are keyed by proposal hash only, so reusing the
-    // store across a reconnect would let a pre-disconnect approval be
-    // consumed by a later session without re-prompting. Clearing here
-    // restores the prior behavior, where the process exited on stream end
-    // and these grants never survived.
-    temporaryGrantStore.clear();
-
+    // Temporary grants are process-shared: they are NOT cleared when a session
+    // ends. CES is moving to a model where the assistant daemon's multiple
+    // processes each hold their own connection, so an ephemeral approval
+    // (`allow_once` / `allow_10m` / `allow_conversation`) granted on one
+    // connection must remain usable by the others rather than being scoped to a
+    // single session. The store's own semantics keep this safe across a
+    // reconnect: `allow_once` is consumed on first use, `allow_10m` is bounded
+    // by its wall-clock TTL, and `allow_conversation` is cleared when its
+    // conversation ends.
+    //
     // A signal-driven end means the process is shutting down; exit the loop.
     // Any other end reason (the assistant disconnected, its stream closed,
     // or the transport errored) means we keep the sidecar up and await a

package/src/server.ts

@@ -678,7 +678,19 @@ export interface ManageSecureCommandToolHandlerDeps {
 export function createManageSecureCommandToolHandler(
   deps: ManageSecureCommandToolHandlerDeps,
 ): RpcMethodHandler<ManageSecureCommandTool, ManageSecureCommandToolResponse> {
-  return async (request) => {
+  // Serialize all manage_secure_command_tool operations. The register path
+  // awaits a bundle download mid-handler; during that await a concurrent
+  // unregister would run its "still in use?" check + bundle delete against a
+  // registry that doesn't yet reflect the in-flight registration — transiently
+  // deleting a bundle another caller is publishing or executing. Running these
+  // operations one-at-a-time closes that window. The registry and toolstore are
+  // process-global, so this single chain serializes tool management across
+  // every connection.
+  let tail: Promise<unknown> = Promise.resolve();
+
+  const handle = async (
+    request: ManageSecureCommandTool,
+  ): Promise<ManageSecureCommandToolResponse> => {
     if (request.action === "unregister") {
       const removed = deps.unregisterTool(request.toolName);
       if (!removed) {
@@ -784,6 +796,18 @@ export function createManageSecureCommandToolHandler(
 
     return { success: true };
   };
+
+  return (request) => {
+    // Chain each operation onto the previous one so they never interleave.
+    const result = tail.then(() => handle(request));
+    // Keep the chain alive regardless of this op's outcome — a rejection must
+    // not break serialization for subsequent operations.
+    tail = result.then(
+      () => undefined,
+      () => undefined,
+    );
+    return result;
+  };
 }
 
 /**