Tool call hangs after extension startup fails

Describe the bug

After I resumed a large session, an extension failed inside joinSession() before reporting ready, and its process exited.

The CLI continued to offer the extension's custom tool. I asked the CLI to call it. The CLI started the call, but the handler did not run and the call reported no error. I aborted it after 9 minutes 38 seconds.

This issue concerns cleanup and dispatch after extension startup failure. A separate large-session attachment defect triggered the failure in this run.

Environment

GitHub Copilot CLI 1.0.82
@github/copilot-sdk 1.0.9-preview.2
Node.js v24.18.1
Linux x86_64

Confirmed startup trigger

I traced the installed 1.0.82 app.js bundle. joinSession() sends a session.resume request containing registration options rather than session history. The host still serializes the resident session's full event array while attaching the extension:

1. handleSessionResumeCore() handles session.resume and calls the host's initializeSession() with replayEvents: false.
2. The host's initializeSession() ignores that option and passes session.getEvents() to seamHost.initializeSession().
3. seamHost.initializeSession() calls JSON.stringify(events).

Node.js v24.18.1 rejects JavaScript strings longer than 536,870,888 UTF-16 code units and throws Invalid string length. I estimated each serialized event array's length from its session event log. Extension attachment succeeded below the limit and failed above it.

In the run that reproduced the stale registration, joinSession() failed with Invalid string length. That session's events.jsonl file was about 1.2 GB.

Honoring replayEvents: false would remove this trigger. Registration cleanup must still handle every startup failure.

Reduced extension

import { writeFile } from "node:fs/promises";
import { joinSession } from "@github/copilot-sdk/extension";

await joinSession({
    tools: [{
        name: "stale_tool_probe",
        description: "Verify extension tool lifecycle.",
        parameters: {
            type: "object",
            properties: {},
            additionalProperties: false,
        },
        skipPermission: true,
        handler: async () => {
            await writeFile(
                new URL("handler-ran.marker", import.meta.url),
                "handler ran\n",
            );
            return "handler-ran";
        },
    }],
});

Observed sequence

1. I stored the reduced extension under ~/.copilot/session-state/<session-id>/extensions/stale-tool-probe/extension.mjs.
2. I resumed the large session.
3. joinSession() failed with Invalid string length before the extension reported ready, and the extension process exited.
4. I asked the CLI to call stale_tool_probe. The CLI started the call instead of rejecting it as an unknown tool.
5. The call was still pending when I aborted it. handler-ran.marker did not appear.

Expected behavior

• The CLI offers an extension's registrations only while the extension is ready and connected.
• Startup failure or process exit removes every registration owned by that extension.
• A tool call fails promptly if its owning extension is unavailable at dispatch or exits while the call is pending.

Actual behavior

• The failed extension's tool remained available after its process exited.
• The call did not reach the handler or report an error.

Related issues

#4590 reports a different session.resume failure involving hook-processor disposal during extension reconnects.

#3856 reports the inverse symptom: Extension tools disappear after a resume.

从 joinSession()、handleSessionResumeCore()、initializeSession() 和 seamHost.initializeSession() 周围的扩展生命周期入手，以复现的 stale_tool_probe 扩展作为切入点。在进程不可用期间，跟踪启动失败、注册清理和 dispatch。完成标准是：失败或已退出的扩展不再暴露注册信息，待处理或新的工具调用会及时失败，而不是一直挂起。