fix(sdk): cleared async singleton after startup failures

If createAgentSession failed after installing a newly created AsyncJobManager but before AgentSession took ownership, the process-global singleton stayed installed. The new singleton guard then caused the next top-level session to skip constructing a scoped manager, disabling async bash/task support.

The startup-error cleanup now clears the singleton only when it still points at the newly created manager, disposes that manager, and then continues the existing registry/kernel cleanup. The regression test forces a startup failure after singleton installation and verifies the next top-level session can create and use its own async manager.

packages/coding-agent/CHANGELOG.md

@@ -4,7 +4,7 @@
 
 ### Fixed
 
-- Fixed `AsyncJobManager.instance()` being cleared while the owning top-level session was still live, which broke the `task` async path with "Async execution is enabled but no async job manager is available" until process restart. Any in-process secondary top-level `createAgentSession()` call (e.g. the Agent Control Center's create flow in `agent-dashboard.ts`) constructed a fresh `AsyncJobManager`, overwrote the singleton, and then cleared it on its own dispose. Secondary sessions now leave the live singleton untouched, and their dispose-time cleanup is scoped so it can no longer cancel the primary session's running bash/task jobs. `bash` / `task` / `job` tools and session job snapshots now resolve the manager through session-scoped async manager wiring rather than `AsyncJobManager.instance()`, so a secondary in-process top-level session cannot accidentally register background work on the owning session's manager or report the owning session's jobs; subagents still inherit the parent's manager via their scoped async manager ([#1923](https://github.com/can1357/oh-my-pi/issues/1923)).
+- Fixed `AsyncJobManager.instance()` being cleared while the owning top-level session was still live, which broke the `task` async path with "Async execution is enabled but no async job manager is available" until process restart. Any in-process secondary top-level `createAgentSession()` call (e.g. the Agent Control Center's create flow in `agent-dashboard.ts`) constructed a fresh `AsyncJobManager`, overwrote the singleton, and then cleared it on its own dispose. Secondary sessions now leave the live singleton untouched, and their dispose-time cleanup is scoped so it can no longer cancel the primary session's running bash/task jobs. `bash` / `task` / `job` tools and session job snapshots now resolve the manager through session-scoped async manager wiring rather than `AsyncJobManager.instance()`, so a secondary in-process top-level session cannot accidentally register background work on the owning session's manager or report the owning session's jobs; subagents still inherit the parent's manager via their scoped async manager. Startup failures after a top-level session installs its manager now clear and dispose that manager before the next session decides whether it can create its own ([#1923](https://github.com/can1357/oh-my-pi/issues/1923)).
 
 ## [15.9.1] - 2026-06-04
 

packages/coding-agent/src/sdk.ts

@@ -2276,6 +2276,12 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 				await session.dispose();
 			} else {
 				if (hasRegistered) agentRegistry.unregister(resolvedAgentId);
+				if (asyncJobManager) {
+					if (AsyncJobManager.instance() === asyncJobManager) {
+						AsyncJobManager.setInstance(undefined);
+					}
+					await asyncJobManager.dispose({ timeoutMs: 3_000 });
+				}
 				await disposeKernelSessionsByOwner(evalKernelOwnerId);
 			}
 		} catch (cleanupError) {

packages/coding-agent/test/sdk-async-job-manager-singleton.test.ts

@@ -133,4 +133,40 @@ describe("AsyncJobManager singleton across concurrent top-level sessions", () =>
 			await primary.dispose();
 		}
 	});
+
+	it("clears a manager installed before a top-level session startup failure takes ownership", async () => {
+		const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), `pi-sdk-async-startup-failure-${Snowflake.next()}-`));
+		tempDirs.push(tempDir);
+		const cwd = path.join(tempDir, `project-${Snowflake.next()}`);
+		const agentDir = path.join(tempDir, "agent");
+		fs.mkdirSync(cwd, { recursive: true });
+
+		await expect(
+			createAgentSession({
+				cwd,
+				agentDir,
+				settings: Settings.isolated({ "bash.autoBackground.enabled": true }),
+				disableExtensionDiscovery: true,
+				skills: [],
+				contextFiles: [],
+				promptTemplates: [],
+				slashCommands: [],
+				enableMCP: false,
+				enableLsp: false,
+				systemPrompt: () => {
+					throw new Error("forced startup failure");
+				},
+			}),
+		).rejects.toThrow("forced startup failure");
+
+		expect(AsyncJobManager.instance()).toBeUndefined();
+
+		const replacement = await spawnTopLevelSession();
+		try {
+			expect(AsyncJobManager.instance()).toBeDefined();
+			expect(replacement.getAsyncJobSnapshot()).not.toBeNull();
+		} finally {
+			await replacement.dispose();
+		}
+	});
 });