High CPU (110%) and memory (2GB) with low context usage (16%) after long session

Environment

OS: macOS 26.5.1 (ARM64, MacBook)
opencode: installed via Homebrew
Provider: deepseek/deepseek-v4-pro
Project: non-git Android Java project (~30 files)

Summary

After a ~3 hour coding session, opencode consumes 110% CPU and 2.0 GB physical memory while the context window is only 16% full.

Symptoms

CPU fluctuates 66%–110% even when idle between messages
Physical memory footprint: 2.0 GB (peak 2.3 GB)
opencode.db grew to 180 MB
Only one opencode process, no zombie subprocesses
Context usage shown in TUI: ~16%

Sample output (1ms interval)

Physical footprint:         2.0G
Physical footprint (peak):  2.3G

799 Thread_2169788   DispatchQueue_1: com.apple.main-thread
  799 start  (in dyld) + 6992
    799 ???  (in opencode)  load address 0x102bd0000 + 0x9bdd0c
      ...
      642 kevent64  (in libsystem_kernel.dylib) + 8       ← event loop idle (80%)
      131 ???  (in opencode)  ... multiple internal frames  ← active work (16%)

80% of samples are in kevent64 (event loop idle), but the 2GB memory footprint suggests significant GC pressure that causes CPU spikes.

Steps to reproduce

Start opencode in a moderate-sized project
Run a long session (~3 hours) with many tool calls (file edits, bash, subagents)
Observe CPU and memory creep over time even when idle

Workaround

Quitting and restarting opencode resolves the issue (memory drops and CPU normalizes).

Notes

No plugins installed
Session persistence via opencode.db ensures no data loss on restart
This could be a GC issue from accumulated context objects not being cleaned up

Update: this is reproducible across multiple sessions, not just a one-off. The computer gets noticeably hot every time opencode runs, even in shorter sessions.

This issue might be a duplicate of or related to existing issues. Please check:

Memory Megathread #20695: Memory Megathread — a dedicated thread for tracking and resolving memory growth issues in opencode. The maintainers are actively collecting heap snapshots there to diagnose exactly this kind of problem (high memory after long sessions). Consider following the instructions in that thread to capture and upload a heap snapshot, which would directly help the team investigate your report.

I've captured heap snapshots following the Memory Megathread instructions (manual Ctrl+P flow).

Files generated:

server.heapsnapshot (100 MB)
tui.heapsnapshot (58 MB)

Heap analysis summary:

Metric	server	tui
Total nodes	1,670,688	916,586
Self size	154.7 MB	99.4 MB
Objects	924k	342k
Closures	388k	246k
Strings	90k	106k

Notable findings:

2.6M total objects across both processes — each GC mark-sweep cycle must traverse this entire graph
634k closures — closures retain enclosing scope references, a common leak source
33.5 MB ArrayBuffer — single large buffer potentially leaked
Duplicate system prompts — same prompt strings stored multiple times (15-32 KB each)
Multiple 500KB+ WeakMaps — large key-value stores not being released

This directly explains the CPU spikes: the massive object graph (2.6M nodes) means V8's concurrent GC threads are constantly working even when idle, and each GC pause grows longer as the heap accumulates.

Will upload the snapshots to the Dropbox link shortly (network restrictions on my end). Let me know if there's an alternative upload method available.

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

Navigation Menu

Repository navigation

High CPU (110%) and memory (2GB) with low context usage (16%) after long session #34226

Environment

Summary

Symptoms

Sample output (1ms interval)

Steps to reproduce

Workaround

Notes

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

High CPU (110%) and memory (2GB) with low context usage (16%) after long session #34226

Description

Environment

Summary

Symptoms

Sample output (1ms interval)

Steps to reproduce

Workaround

Notes

Activity

aaabyf commented on Jun 27, 2026

github-actions commented on Jun 27, 2026

aaabyf commented on Jun 27, 2026

github-actions commented on Aug 27, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions