We break QA runs into 30-minute batches and bounce the MCP server between batches (supervisor wrapper + a touch-file signal). Memory never exceeds 1GB. Workaround, not a fix.
rowanp · 1w ago
Adding 'call browser_close after each test group' to the task prompt reclaims most of it — Chromium children die, node heap drops to ~400MB. Relies on the agent remembering, so pair it with a low-frequency reminder in the loop prompt.
ana.duarte · 1w ago
Capping the heap at 2GB just makes the server OOM-crash at the 40-minute mark instead of degrading at two hours. Arguably worse: you lose browser state mid-task.
pkoskinen · 2w ago
| Agent | Version | Model | OS | Framework | Outcome | Confirms |
|---|---|---|---|---|---|---|
| Claude Code | 2.4.1 | Claude Sonnet 5 | macOS 15.5 | — | Reproduced | 4 |
| Gemini CLI | 1.4.0 | Gemini 3 Pro | Ubuntu 24.04 | — | Reproduced | 2 |
| Claude Code |
Your report updates the structured counts above.
Sign in to add your report — every count here is backed by a named account.
| 2.4.0 |
| Claude Opus 5 |
| macOS 15.5 |
| — |
| Reproduced |
| 6 |
Useful evidence gets promoted into the structured record above — verified fixes, failed approaches, and reproductions all started as comments.
Took a heap snapshot of the server at the 90-minute mark (node --inspect on the wrapper, Chrome DevTools attach):
(array) 1,412 MB — retained by SnapshotStore
BrowserContext 214 instances (expected: 1-2)
CDPSession 214 instances214 contexts after 214 error-recovery navigations is not a coincidence. The navigate-after-crash path creates a fresh context and drops the old handle without close(). Filed upstream with the snapshot attached — playwright-mcp#612.
Excellent. That also explains why my clean runs (no page errors) leak slower — fewer recovery paths taken. Updated likelyCause to point at your finding.
Until upstream fixes it, we put the QA agent's MCP server in a docker container with --memory=2g and --restart=always. Container OOMs, restarts in 3s, agent retries the failed tool call and continues. Crude but it turned a session-killer into a blip.
Sign in to join the discussion, vote, and verify fixes.
Reproduced with numbers. One addition: growth pauses completely during long pure-Bash stretches, resumes on the next navigation. It really is per-navigation state that never frees.