RPC mode: resuming an existing session on Windows crashes with libuv assertion (0xC0000409) during shutdown #9507

What happened?

On Windows, pi in RPC mode exits with 3221226505 (0xC0000409) instead of 0, ~20–40 ms after agent_settled, even though the turn completed and the session JSONL was fully written. stderr:

Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), file src\win\async.c, line 94

Reproduced 4/5 resume runs using an ordinary auto-exit extension. This looks like the known Node.js bug nodejs/node#56645, fixed by nodejs/node#61999 ("Ignore PostDelayedTask after Stop"): the fix is not in Node v24.19.0 but is in v24.20.0+ (v24.20.0 changelog lists PR #61999). pi's RPC host shutdown calls process.exit(exitCode) right after runtimeHost.dispose(); on Node v24.19.0 that can hit the assertion. Same class as the #5805 workaround for pi update --self (return from main() and let Node drain instead of immediate process.exit(0)).

Environment: Windows 11 Pro, version 10.0.26200 (build 26200); Node v24.19.0 (libuv 1.52.1); pi 0.85.1 (npm global install); Git Bash, headless spawn.

Steps to reproduce

1. Save autoexit.ts (all it does is shut the RPC process down after the turn):

import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
export default function (pi: ExtensionAPI) {
  pi.on("agent_end", (_e, ctx) => { ctx.shutdown(); });
}

2. Run pi --mode rpc --model <model> --session <existing-session.jsonl> -e autoexit.ts.
3. Send {"id":"1","type":"prompt","message":"say hi"} on stdin and wait for agent_settled; the process exits on its own.

Expected behavior: exit code 0.

Actual: exit code 3221226505; assertion above on stderr, ~20–40 ms after agent_settled; the session file already contains the final assistant message at crash time.

Evidence

• One raw run: agent_settled at t+2574 ms → assertion at t+2598 ms → exit 3221226505.
• Node v24.19.0: 4/5 resume runs crashed with the minimal extension. Node v24.21.0: 0/8 crashes with the identical repro.
• v24.21.0's src/node_platform.cc contains the #61999 guard (if (has_shut_down_) return; in PostDelayedTask); v24.19.0 does not.
• Not extension-caused beyond triggering pi's own shutdown: the repro extension contains no third-party code, and both RPC exit paths reach the same shutdown() → process.exit() (the ctx.shutdown() flag is consumed on agent_settled; stdin end calls the same shutdown).
• Without any extension I could not trigger it in 16 EOF runs today, but earlier runs on this machine (same pi/Node) crashed on EOF too; the timing window is narrow and an auto-exit hook hits it consistently.

Version: pi 0.85.1

Impact: on Node v24.19 (LTS at the time) a Windows RPC session that shuts down via auto-exit/EOF can be reported to orchestrators as a crashed child despite a completed turn; resume is the common path for sub-agent-style usage.

Related: #5805 (closed) · nodejs/node#56645 · nodejs/node#61999 (merged 2026-07-24, backported in Node v24.20.0)

AI disclosure: filed by an automated agent on behalf of a user; every measurement above comes from local runs with the repro shown.