opencode2: RSS grows ~530 MB/min during active agent turn (OOM crash)

Summary

opencode2 (@opencode-ai/cli@0.0.0-next-17088) leaks memory at a steady ~530 MB/min RSS during an active agent turn, climbing from ~3.8 GB to ~6.2 GB in ~4.5 minutes and eventually crashing the process. Related: #20695 (Memory Megathread).

Environment

• opencode version: 0.0.0-next-17088 (opencode2 from @opencode-ai/cli)
• OS: Linux 7.0.0-29-generic (Ubuntu, x86_64) — Linux tuxell 7.0.0-29-generic #29-Ubuntu SMP PREEMPT_DYNAMIC Fri Jul 17 20:52:35 UTC 2026 x86_64 GNU/Linux
• Terminal: TERM=xterm-256color, COLORTERM=truecolor (no TERM_PROGRAM)
• Shell: /bin/bash
• Install/channel: next — binary opencode2.exe via npm global @opencode-ai/cli@0.0.0-next-17088 (~/n/bin/opencode2 → @opencode-ai/cli/bin/opencode2.exe)
• Active plugins (config + local):
  • @sveltejs/opencode (configured)
  • local: ~/.config/opencode/plugins/notify.ts, worktree.ts
  • skill snapshot also listed: @knikolov/opencode-plugin-simple-memory (verify if still loaded)
• MCP (enabled): codegraph (local), mdn (remote), axiom, posthog, paddle-sandbox, paddle-live, netlify
• Compaction: auto: false, prune: false

Reproduction

1. Start opencode2 in a project session (managed/background service default).
2. Run an active agent turn (model streaming + tool calls; not idle).
3. In another terminal, sample RSS every 30s:

while true; do date; ps -o pid,rss,vsz,comm -p $(pgrep -f "opencode2.exe"); echo "---"; sleep 30; done | tee /tmp/opencode_mem.log

4. Observe linear RSS growth until the process dies / becomes unusable.

Expected Behavior

RSS should stay roughly stable or grow slowly with context, not hundreds of MB per minute during a normal agent turn. The process should not OOM/crash from unbounded growth.

Actual Behavior

Near-linear RSS growth while an agent turn is active:

Derived from full /tmp/opencode_mem.log (10 samples):

• Start RSS: ~3.8 GB → end: ~6.2 GB
• Total growth: ~2.4 GB in ~4.5 min
• Average: ~267 MB / 30s ≈ ~533 MB/min
• VSZ also climbs in lockstep (~80 GB range; large virtual map)

Process eventually crashes / becomes unusable (continuous crash loop from user POV). Host has 29 GiB RAM; swap was already under pressure (free: ~4.9 GiB swap used at capture time).

Additional Context

• Workload during leak: active agent turn (streaming + tools), not idle.
• Session had multiple MCP servers and subagent use available; leak observed under normal interactive agent work.
• Heap snapshot not yet captured — will follow megathread flow and upload:
  • Manual: Ctrl+P → “heap snapshot” while RSS is elevated, upload to the Dropbox request in #20695
  • Auto: OPENCODE_AUTO_HEAP_SNAPSHOT=1 opencode2 then use until RSS > ~2 GB; snapshots land in the log dir as .heapsnapshot
• Happy to upload a heap snapshot and add a short note on #20695 once captured.
• No secrets included; plugin/MCP list redacted of tokens.