`session.disconnect()` is cooperative; doesn't abort `sendAndWait` or close the transport

从 nodejs/src/session.ts 开始，重点关注 disconnect()、abort()、sendAndWait() 以及它们对 MessageConnection 的使用。跟踪正在进行的请求和传输关闭是如何处理的，然后明确取消操作应当保留 session 的有效性，还是强制 teardown；当所选行为得到明确规定，并且正在进行的调用能够及时以干净的错误结束时，即视为完成。

Summary

session.disconnect() sends session.destroy over RPC and clears local handlers, but does not close the underlying MessageConnection. Any in-flight sendAndWait() continues to await events from the CLI for up to TCP-timeout duration. The method name implies abortive semantics; the behavior is cooperative signaling.

Repro

const client = new CopilotClient();
await client.start();
const session = await client.createSession({ /* ... */ });

// Start a long-running call
const inflight = session.sendAndWait({ prompt: "Write a very long story..." });

// Try to abort after 5 seconds
setTimeout(async () => {
  await session.disconnect();
  console.log("disconnect() returned");
}, 5000);

// `inflight` does NOT reject after disconnect returns; it continues to
// pend for up to 30–90 minutes (until the CLI's natural completion or
// TCP timeout).
await inflight;

Expected

disconnect() returns → underlying transport closed → in-flight sendAndWait() rejects with a clean SessionAborted (or similar) error → consumer can release resources immediately.

Actual

disconnect() returns immediately but the JSON-RPC transport stays open. The in-flight sendAndWait() is unaffected and continues running until the CLI naturally returns or TCP times out (often 30+ minutes).

Evidence (SDK source)

nodejs/src/session.ts (v0.3.0): disconnect() calls sendRequest("session.destroy", { sessionId }) and clears local eventHandlers. No call to connection.close() or transport-level termination. session.abort() (same file, line ~1073) sends session.abort to cancel the current message at the CLI but also doesn't close the SDK-side transport.

Workaround (consumer-side)

Layer an AbortController on the polling layer plus a "session result ignored" flag so the eventually-resolving sendAndWait doesn't pollute downstream state. ~50–100 LOC of scaffolding that every consumer with reliability requirements has to reinvent.

Suggested fix (pick one)

• Enhance abort() to also reject in-flight sendAndWait with a clean SessionAborted error, preserving session validity for new messages. Suits "cancel and retry" consumers.
• Add session.disconnect({ force: true }) (or session.kill()) that closes the underlying JSON-RPC connection immediately. Suits "tear down for good" consumers — the watchdog scenario above. Keep current cooperative disconnect() behavior as the default.

Related

• #590 — "Long-running prompt suddenly hangs forever with no error or exception" — describes the symptom from the consumer side. The disconnect-doesn't-abort behavior is one mechanism that contributes to this symptom: even when the consumer notices the hang and tries to recover, disconnect() doesn't free the in-flight call cleanly.

Environment

- SDK: @github/copilot-sdk@0.3.0
- CLI: @github/copilot@1.0.45
- Node: 22 LTS
- OS: Windows 11 (primary), macOS 14 (secondary)
- Model: claude-sonnet-4-6 (CLI default routing)