Node.js OOM crash after ~37 min — 31,965 leaked async libuv handles (SEA ignores NODE_OPTIONS)

Bug: Node.js OOM crash after ~37 minutes — 31,965 leaked async libuv handles

Version: 1.0.82
Platform: Linux x86_64 (Amazon EC2, Node v24.20.0 embedded in SEA)
Observed: 2026-09-01, crash at ~37 min session uptime

Summary

Every session crashes with FATAL ERROR: Reached heap limit Allocation failed - JavaScript heap out of memory after roughly 30–60 minutes of use. The crash is caused by a monotonic leak of async libuv handles (~32k by crash time), not by conversation-context growth. /compact and increasing NODE_OPTIONS (which the SEA ignores) do not prevent the crash.

Heap report metrics (from report.20260901.153402.2817455.0.001.json)

Metric	Value
Session uptime at crash	2,203 s (~37 min)
old_space used / capacity	3,899 MB / 3,899 MB (100% saturated)
externalMemory	36 MB
mallocedMemory	~1 MB
memoryLimit	4,298 MB (~4.0 GB)
RSS at crash	4,634 MB
GC current mu	0.173 (thrashing — normally >0.5 is healthy)
CPU at crash	~292% (GC threads saturating 3 cores)

Growth rate: ~100 MB/min of old_space objects throughout the session.

libuv handle leak

At crash time, the report shows:

libuv handle count: 31,996
  type=async, is_active=true, is_referenced=false: 31,965
  type=signal, is_active=true:                         19

• 31,965 active-but-unreferenced async handles — every one holds a JS closure/context in old_space. With no detail string (details: ""), these are likely AsyncResource / AsyncLocalStorage contexts that were created but never destroyed.
• 19 active signal handles (vs. the ~17 signals that exist): duplicate SIGABRT, SIGIO registrations suggest repeated process.on('signal', ...) calls without corresponding removeListener — consistent with code that re-registers handlers per turn or per tool call.

The combination is a textbook listener/async-context leak: each model turn or tool execution allocates one or more async handles that are never cleaned up. After ~32k turns/tool calls the heap saturates.

Why NODE_OPTIONS and heap increases appear to help but don't

The copilot binary is a Node Single Executable Application (SEA) (NODE_SEA_BLOB present; 159 MB on disk). SEAs ignore NODE_OPTIONS entirely — empirically verified: NODE_OPTIONS="--this-is-not-a-real-flag" copilot --version exits 0 with no error.

The commandLine in the crash report confirms only Copilot's own flags are applied:

['copilot','--no-warnings','--report-on-fatalerror','--optimize-for-size','--expose-gc','copilot']

Increasing the (ignored) NODE_OPTIONS heap size is a placebo — the 4 GB cap is always in effect.

Reproduction pattern

The crash is reproducible by running a session of moderate length (>30 min) with regular tool calls (bash, view, grep). The leak appears to be per tool execution or per model turn rather than per session.

The --report-on-fatalerror flag already embedded in the binary's commandLine will produce a report.YYYYMMDD.*.json in the CWD on every crash — these reports contain full diagnostics.

Suggested investigation areas

1. AsyncResource / AsyncLocalStorage contexts created per turn/tool that escape cleanup (search for new AsyncResource or AsyncLocalStorage usages in app.js)
2. process.on(signal, ...) calls without process.removeListener — each session re-registration without cleanup would explain the 19 duplicate signal handles
3. Any setInterval/setTimeout created per tool call that is never clearInterval-ed — these register as async handles too

Workaround (until fixed)

Run the Copilot JS dist under a real node with a larger heap:

~/.nvm/versions/node/v24.20.0/bin/node \
    --max-old-space-size=12288 \
    --expose-gc --no-warnings --optimize-for-size \
    ~/.cache/copilot/pkg/linux-x64/1.0.82/index.js \
    "$@"

This buys ~2 hours before the same leak exhausts 12 GB — same leak, bigger bucket.

Start by reproducing a session with regular tool calls and inspect the generated report.YYYYMMDD.*.json, then read the async-context and signal-handling code in app.js. Search the mentioned AsyncResource, AsyncLocalStorage, process.on, timer, and cleanup paths; done means the handle count no longer grows monotonically and extended sessions avoid the reported OOM.