Copilot CLI crashes on exit (Windows) — libuv `uv_async_send` on a closing handle (FAST_FAIL_FATAL_APP_EXIT)
- Vorherrschende Sprache
- Shell
- Sterne
- 11.2k
- Forks
- 1.9k
- Ø Merge
- 14 Std. 16 Min.
- Gemergte PRs (30 T.)
- 6
Beschreibung
### Describe the bug
On Windows x64, `copilot.exe` **consistently crashes at process exit** with a fatal
fail-fast (`0xc0000409`, subcode `0x7 FAST_FAIL_FATAL_APP_EXIT`). The session's work
completes normally; the crash happens only during teardown.
Live-debugger analysis (WinDbg attached to the crashing process) shows this is the
well-known **Windows-only Node.js/libuv shutdown race** — `uv_async_send()` is called
on a `uv_async_t` whose `UV_HANDLE_CLOSING` flag is already set, tripping the libuv
assertion:
```
Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), file src\win\async.c
```
which calls `abort()` → `__fastfail(FAST_FAIL_FATAL_APP_EXIT)`. The bundled Node
(v24.16.0) ships libuv with this assertion **enabled**, so the otherwise-benign
teardown race becomes a hard crash.
Upstream tracking issue: https://github.com/nodejs/node/issues/56645
(Node 23/24/25 lines on Windows; Linux/macOS unaffected.)
### Affected version
- **GitHub Copilot CLI:** 1.0.74-0 (crash first captured on 1.0.73.0; reproduces on 1.0.74-0) - **Bundled Node.js:** v24.16.0 (win-x64, single-executable build), ICU 78 - **OS:** Windows x64 (build 10.0.29635) - **Onset:** started ~1 month ago (consistent with a CLI update that bumped the bundled Node)
### Steps to reproduce the behavior
1. Launch `copilot` on Windows x64.
2. Use it normally (any session involving network/async activity).
3. Exit the CLI.
4. The process aborts on exit with a fail-fast crash / WER entry (exit is not clean).
Reproduces every time on the affected machine.
### Expected behavior
The CLI should exit cleanly (exit code 0, no fail-fast/WER crash) after the session ends.
### Additional context
Two threads show both sides of the shutdown race:
**Crashing thread — "V8Worker" (a V8 background platform worker):**
```
copilot!uv_async_send+0x3a
→ _wassert("!(handle->flags & UV_HANDLE_CLOSING)", "src\win\async.c", 94)
→ abort()
→ __fastfail(FAST_FAIL_FATAL_APP_EXIT) ; int 29h, rcx=7
```
Disassembly of `uv_async_send` confirms the fatal branch is the top-of-function
guard (`test byte ptr [handle+0x58],1` / `UV_HANDLE_CLOSING`), not a failed
`PostQueuedCompletionStatus`.
**Main thread — in Node's process-exit teardown, joining workers:**
```
node::DefaultProcessExitHandler
→ (V8 platform shutdown)
→ uv_thread_join
→ WaitForSingleObjectEx ; waiting for the V8 worker to exit
```
So during `DefaultProcessExitHandler`, the main thread closes the per-isolate
"flush-tasks" `uv_async_t` (sets `UV_HANDLE_CLOSING`) and joins the V8 workers; a
V8 worker drains one last task and calls `uv_async_send()` on that now-closing
handle → assertion → abort.
`!analyze -v` summary:
- `ExceptionCode: c0000409` (STATUS_STACK_BUFFER_OVERRUN), `Subcode: 0x7 FAST_FAIL_FATAL_APP_EXIT`
- `FAILURE_BUCKET_ID: FAIL_FAST_FATAL_APP_EXIT_c0000409_copilot.exe!Unknown`
- Faulting thread is the V8 worker; instruction is `int 29h` (`__fastfail`).
## Suggested fixes (for maintainers)
- Avoid `process.exit()` during teardown when V8/libuv async handles are still
live; prefer setting `process.exitCode` and letting the loop drain, or add a
small deferral before forced exit (common community workaround for #56645).
- Ship a Node build with the libuv fix for nodejs/node#56645 once available, or a
Node line where the shutdown ordering doesn't hit this path on Windows.
- Impact is at-exit only (no data loss), but it produces a non-zero exit and a WER
crash entry on every exit, which is noisy and can break scripts that check the
CLI's exit code.
Beitragsleitfaden
Rechercherichtung
Beginne damit, den im Bericht beschriebenen Node DefaultProcessExitHandler und das Herunterfahren der V8-Plattform nachzuverfolgen, wobei du dich auf den Pfad von process.exit und das per-Isolate-Flush-Tasks-uv_async_t konzentrierst. Reproduziere den Exit-Crash unter Windows x64 und vergleiche das Verhalten des gebündelten Node v24.16.0 mit dem Tracking-Issue nodejs/node#56645 im Upstream. Als erledigt gilt die Aufgabe, wenn copilot sauber mit Code 0 beendet wird und kein Fail-Fast- oder WER-Crash auftritt.
Vom Indexierungsmodell aus dem Issue-Text verfasst.
Bewertung
- Tech-Stack
- node.js
- Bereich
- cli
- Issue-Typ
- Bug
- Schwierigkeit
- 4/5
- Geschätzter Aufwand
- 3-5 Tage
- Aktivitätsstatus
- Ruhig
- Klarheit
- Größtenteils klar
- Anfängerfreundlichkeit
- 48/100