v1.0.80: event-storage exhaustion retry storm drives long-running session into GC/compaction loop and Node OOM

[justinchuby]

Describe the bug

A long-running active Copilot CLI session eventually reached remote event-storage exhaustion. After that, the exporter kept attempting 500-event flushes, while the process repeatedly reported memory pressure, forced GC/emergency compaction, and thousands of bridge acknowledgement timeouts. The CLI ultimately terminated with a Node out-of-memory error shown in the terminal.

This appears to connect two existing failure modes that are currently reported separately:

• Long-running agent sessions exhaust event storage and appear cancelled while CLI remains active #4467: session_event_storage_exhausted
• Memory-pressure watchdog force-compacts the conversation at 23% context usage, recovers 0.003% of tokens, then loops until OOM #4506: memory-pressure GC/compaction loop ending in OOM
• Also related to Resume of a large session OOMs / grinds one CPU core for ~70 min in 1.0.74 (regression vs 1.0.73; ~3–4× memory) #4251 for unbounded memory growth in large sessions

The new observation is that remote storage exhaustion does not merely make session status unreliable: the repeated exporter failures/backlog correlate with sustained local memory pressure and eventual process death.

Environment

• Copilot CLI at failure: 1.0.80
• Embedded Node.js: v24.18.1
• OS: Linux x86_64, kernel 6.17.0-1022-azure
• Host memory: 125 GiB RAM, no swap
• Session created: 2026-08-10
• Process/resume started: 2026-08-18
• Failure: 2026-08-27 (about 8.8 days of this process, about 16.9 days of session lifetime)
• Workload: long-running interactive/autonomous software-engineering session with many tool calls and subagents

Session size at failure

• events.jsonl: 934,639,321 bytes (~892 MiB)
• Event records: 282,912
• Checkpoints: 32

Log evidence

Across the process log:

Event	Count
session_event_storage_exhausted	18,909
Memory pressure detected - requesting garbage collection	21,352
Memory pressure persists after GC - triggering emergency compaction	473
timed out waiting for bridge event ack	3,001

Representative final minutes:

[WARNING] Failed to submit events to Mission Control session …: 409
{"code":"session_event_storage_exhausted","message":"session event storage exhausted"}
[WARNING] remote session batch flush failed {"event_count":500,...}
[WARNING] remote session exporter circuit opened after batch flush failure
[WARNING] Memory pressure detected - requesting garbage collection
[WARNING] Memory pressure persists after GC - triggering emergency compaction
[WARNING] timed out waiting for bridge event ack {"timeout_ms":30000}
[INFO] Compacted 17 messages, saved ~9267 tokens
[WARNING] Memory pressure detected - requesting garbage collection
...
[INFO] Compacted 38 messages, saved ~23339 tokens
[WARNING] Memory pressure detected - requesting garbage collection

The application log ends in repeated memory-pressure warnings. The fatal Node OOM text was printed by the terminating runtime to the terminal and was not persisted into the Copilot process log. No kernel OOM-killer event was recorded, so this was a Node/V8 process failure rather than the Linux kernel killing the process.

The host currently has ample free memory after restart; I did not capture /proc/<pid>/status immediately before the crash, so I cannot claim whether the final allocation failure was at the V8 heap limit or in native/external memory.

Steps to reproduce

1. Keep one Copilot CLI session active/resumable for more than a week with frequent tool calls and subagent events.
2. Allow events.jsonl to grow toward 1 GB and remote Mission Control event storage to fill.
3. Observe repeated 409 session_event_storage_exhausted responses for 500-event batches.
4. Continue using the session.
5. Observe repeated memory-pressure GC, emergency compaction, bridge ack timeouts, and eventually a Node OOM termination.

Expected behavior

• Rotate/continue/compact remote event storage before its limit is reached.
• Once the server permanently rejects a stream as exhausted, stop retaining/retrying the rejected backlog in a way that grows local memory.
• Keep local session operation and checkpointing healthy when remote export is unavailable.
• Apply bounded backoff and expose a clear degraded-export state.
• Do not repeatedly perform lossy conversation compaction unless it measurably relieves the memory pressure.
• Persist the fatal Node diagnostic report path or equivalent crash telemetry in the process log.

Actual behavior

A permanent remote quota condition produced 18,909 failed submissions, concurrent memory-pressure/compaction loops, degraded bridge responsiveness, and eventual Node OOM. Recovery required starting a new session and manually reading the old checkpoint and session directory.

[vicondoa]

Same here, it also seems to get very slow over time while it grinds non stop doing GC. I also have it segfault with OOM when there is plenty of free on the host.