OOM crash (`JavaScript heap out of memory`) on long `--resume` sessions; crash dumps written into the user's cwd

Comece reproduzindo uma sessão longa de --resume com --report-on-fatalerror e --optimize-for-size, depois inspecione o relatório de diagnóstico do Node anexado e o comportamento do histórico de sessões. Considera-se concluído quando sessões longas não continuarem crescendo até causar um OOM irrecuperável, e os relatórios de erro fatal forem gravados em um local pertencente à ferramenta em vez do cwd do usuário.

Describe the bug

Summary

Copilot CLI 1.0.82 repeatedly dies with a V8 heap OOM during long resumed sessions. It crashed 3 times in ~14 hours for me, always at the 4 GiB heap cap.

Separately, the resulting Node diagnostic reports are written into the current working directory, so they land inside whatever git repo I'm working in and show up as untracked files.

Environment

Copilot CLI	1.0.82
Node	v24.18.1
Platform	linux x64 (glibc 2.39)
Invocation	copilot --no-warnings --report-on-fatalerror --optimize-for-size --expose-gc copilot --resume

What happens

"event":   "Allocation failed - JavaScript heap out of memory"
"trigger": "OOMError"

Three crashes on 2 Sep 2026, all in resumed sessions:

Time (UTC)	Heap used / limit	RSS	CPU (user s)
00:51:37	3.98 / 4.00 GiB	4.63 GiB	7833
03:08:30	3.99 / 4.00 GiB	4.75 GiB	4065
14:39:00	3.99 / 4.00 GiB	4.69 GiB	2785

Diagnostic detail

At crash, old_space alone accounted for 3.96 GiB of the 3.99 GiB used:

old_space:                3.96 GiB
code_space:               0.01 GiB
trusted_space:            0.01 GiB
large_object_space:       0.00 GiB

Nearly all of it is long-lived, survived-GC data — consistent with unbounded retention/a leak rather than a transient allocation spike. javascriptStack is "No stack." / Unavailable., as expected for an OOM abort.

All three crashes were --resume sessions with substantial CPU time behind them (46m–2h11m user CPU), which points at per-session state (conversation/tool history?) accumulating without bound. Heap is pinned right at the cap each time, and RSS exceeds it by ~0.6–0.75 GiB.

Impact

1. Data/work loss — long sessions die abruptly once they get big enough. The practical ceiling appears to be session length, not task complexity.
2. Repo pollution — because --report-on-fatalerror is set and Node writes reports relative to cwd, files like report.20260902.143900.310882.0.001.json appear inside the user's repository. They show as untracked in git status and are easy to commit by accident. A CLI shouldn't drop crash artifacts into arbitrary project directories.

Notes for Copilot debugging

The 4 GiB ceiling is not user-imposed:

• NODE_OPTIONS is unset.
• Host has 30 GiB RAM with ~9 GiB free at crash time — no system memory pressure.
• Plain node on this host reports a 4.05 GiB default heap limit; node --optimize-for-size reports exactly 4.00 GiB, matching javascriptHeap.memoryLimit in every dump.

So the cap comes from the --optimize-for-size flag the CLI passes itself. The process aborts at a self-imposed ceiling while the machine still has memory to spare — worth considering whether that flag is right for long-lived sessions, though the underlying retention growth looks like the real bug.

Affected version

1.0.82

Steps to reproduce the behavior

1. Start a Copilot CLI session in a git repo.
2. Use it heavily / resume it over an extended period (mine ran for hours of CPU time).
3. Session eventually aborts with the OOM above; a report.*.json appears in the repo root.

Expected behavior

• Session memory should be bounded (evict/compact old history), or degrade gracefully instead of a hard abort.
• Crash dumps should go somewhere tool-owned — e.g. ~/.copilot/logs/ or $TMPDIR — not the user's cwd. If cwd is intentional, it should be configurable.

Additional context

report.20260902.005137.422602.0.001.json