opencode: hardcoded 4Gi memory limit causes OOMKilled / silent session deaths (manifest declares limitedMemory 24Gi) #3395

Summary

The OpenCode app container ships with a hardcoded resources.limits.memory: 4Gi in opencode/templates/opencode.yaml, while the app's own OlaresManifest.yaml declares limitedMemory: 24Gi / limitedCpu: 13. On an Olares One (96 GB RAM) this 4 GiB cap is routinely exceeded by memory-heavy subprocesses that OpenCode legitimately runs inside its sessions (dev servers, vitest --max-old-space-size=8192, etc.), producing OOMKilled restarts that silently kill every active session.

Evidence (live diagnosis, 2026-08-08)

• Pod opencode container: reason: OOMKilled, exitCode: 137 — 31 restarts in 23 h.
• cgroup memory reached 4.26 GiB (over the 4 GiB limit) during a npx vitest --max-old-space-size=8192 run, ~40 s before the next OOM kill.
• Every restart kills all in-flight sessions "silently"; reloading afterwards fails with envoy upstream connect error or disconnect/reset before headers ... Connection refused until the container is back up.
• Hardcoded block in opencode/templates/opencode.yaml:

resources:
  limits:
    cpu: "4"
    memory: 4Gi
  requests:
    cpu: 200m
    memory: 1Gi

vs. the manifest:

requiredMemory: 2304Mi
limitedMemory: 24Gi
limitedCpu: 13

Suggested fix

• Raise the hardcoded container memory limit (e.g. 8–12 GiB) to match the declared limitedMemory: 24Gi, and/or
• Make the resource limits configurable via values.yaml / an env var instead of hardcoding them.

Workaround (applied successfully)

Patching the Deployment to limits.memory: 12Gi resolves the OOM loop (0 restarts after rollout).

Environment

• Olares app: opencode 1.0.18 (image beclab/opencode:1.18.9.1), source market.olares
• Olares One, 96 GB RAM / 24 cores, NVIDIA RTX 5090