fix(runner): contain protocol-failure cleanup rejection

Co-Authored-By: Paperclip <noreply@paperclip.ing>

doc/execution-semantics.md

@@ -1231,3 +1231,9 @@ An asynchronous remote signal failure, including a sandbox already removed by
 the operator, must not crash the controller. Logging that failure must also be
 contained. A rejected signal does not prove termination: existing process and
 provider monitoring still own stop acknowledgement and cleanup proof.
+
+Protocol-failure handling can begin transport cleanup before the owning runtime
+awaits it. That background invocation observes rejection immediately, including
+when a remote sandbox has already disappeared. The owner's awaited close still
+receives the original failure; containment never fabricates a successful close
+or permission to reuse an unverified execution.

packages/paperclip-runner/src/drivers/codex/codex-protocol-cleanup.test.ts

@@ -0,0 +1,62 @@
+import { execFileSync } from "node:child_process";
+import { describe, expect, it } from "vitest";
+
+describe("protocol-failure cleanup", () => {
+  it.each(["immediate", "delayed", "successful"])(
+    "contains %s background cleanup without hiding its outcome from the owner",
+    (outcome) => {
+      // Use Node's fatal unhandled-rejection policy in a separate process.
+      // A test-runner rejection listener would conceal the controller crash.
+      const source = `
+        import assert from "node:assert/strict";
+        import { CodexHarnessSession } from ${JSON.stringify(new URL("./codex-harness-session.ts", import.meta.url).href)};
+        const cleanupError = new Error("Sandbox not found during cleanup");
+        let closePromise;
+        let cleanupAttempts = 0;
+        const transport = {
+          setServerRequestHandler() {},
+          async *notifications() { throw new Error("Sandbox not found during monitoring"); },
+          close() {
+            if (!closePromise) {
+              cleanupAttempts++;
+              closePromise = ${JSON.stringify(outcome)} === "successful"
+                ? Promise.resolve()
+                : ${JSON.stringify(outcome)} === "delayed"
+                  ? new Promise((_, reject) => setTimeout(() => reject(cleanupError), 10))
+                  : Promise.reject(cleanupError);
+            }
+            return closePromise;
+          },
+        };
+        const session = new CodexHarnessSession({
+          transport, runId: "run-cleanup", normalizedSessionId: "session-cleanup",
+          opened: { lineage: { threadId: "thread-cleanup" }, context: {} }, taskEnvelope: {},
+          conversationMode: "task", resumed: false, activeTurnId: "turn-cleanup",
+          sourceSequence: 0, now: () => new Date(), runnerInstanceId: "runner-cleanup",
+          driverKind: "codex", capabilities: {}, goalCapability: "disabled",
+          goalAvailability: "unavailable", goalReasonCode: null, goalReason: null,
+          dynamicTools: [],
+        });
+        // The ordinary owner may only join cleanup on a later event-loop turn.
+        await new Promise(resolve => setTimeout(resolve, 40));
+        assert.equal(session.protocolFailed, true);
+        assert.equal(session.protocolFailureCode, "notification_transport_failed");
+        assert.equal(session.activeTurnId, null);
+        const events = [];
+        for await (const event of session.eventQueue) events.push(event);
+        assert.equal(events.filter(event => event.eventType === "session.failed").length, 1);
+        assert.equal(events.filter(event => event.eventType === "turn.failed").length, 1);
+        assert.equal(events.some(event => event.eventType === "turn.completed"), false);
+        session.failProtocol("duplicate_failure", "must not retry cleanup");
+        if (${JSON.stringify(outcome)} === "successful") await session.close();
+        else await assert.rejects(session.close(), error => error === cleanupError);
+        assert.equal(cleanupAttempts, 1);
+        process.stdout.write("HOST_ALIVE_CLEANUP_OUTCOME_PRESERVED");
+      `;
+      expect(execFileSync(process.execPath, [
+        "--unhandled-rejections=strict", "--import", import.meta.resolve("tsx"),
+        "--input-type=module", "--eval", source,
+      ], { encoding: "utf8", timeout: 10_000 })).toBe("HOST_ALIVE_CLEANUP_OUTCOME_PRESERVED");
+    },
+  );
+});

packages/paperclip-runner/src/drivers/codex/codex-session-state.ts

@@ -415,7 +415,12 @@ export class CodexSessionState {
     }
     this.terminal = true;
     this.eventQueue.close();
-    void this.transport.close(`protocol_failure:${code}`);
+    // Notification failure can initiate cleanup before the owning runtime joins
+    // it. Observe this background rejection immediately so a deleted remote
+    // sandbox cannot crash the controller. The transport retains its original
+    // close promise: the owner's awaited session.close still receives any
+    // cleanup failure and must not treat it as confirmed termination.
+    void this.transport.close(`protocol_failure:${code}`).catch(() => undefined);
   }
 
   emit(