Mem watchdog ops

Mem Watchdog Ops Skill

Use this skill when working in constrained Crostini environments where VS Code, Chrome/Playwright, and extension host memory spikes can trigger OOM.

Ground Rules

• Never read or reason from SwapFree values on Crostini.
• Prefer systemctl --user only — never sudo systemctl.
• Treat the mem-watchdog daemon as independent runtime authority.
• For tuning, use extension settings (memWatchdog.*) so the extension writes config and restarts service safely.

Fast Workflows

1. Status snapshot

   • Use chat participant command: /memwatchdog status
   • Or run helper: ~/.copilot/skills/mem-watchdog-ops/watchdog-snapshot.sh

2. Recent actions / diagnosis

   • /memwatchdog logs
   • Look for action markers in journal:
     • ACTION(SIGTERM):
     • ACTION(SIGKILL):
     • RECOVERY(SIGTERM):

3. Apply tuning profile

   • /memwatchdog tune balanced — general dev use (warn 3.4 GB, emerg 3.8 GB)
   • /memwatchdog tune conservative — earlier intervention (warn 3.0 GB, emerg 3.4 GB)
   • /memwatchdog tune playwright — more headroom for automation (warn 3.8 GB, emerg 4.2 GB)

4. Manual protective action

   • /memwatchdog act kill chrome
   • /memwatchdog act restart service

Tuning Guidance

• balanced: general dev use — recommended for Copilot Chat multi-agent workloads that peak at 3.0–3.5 GB
• conservative: earlier intervention, safer on tight RAM — may cause more Chrome kills
• playwright: more headroom for headed browser automation sessions

When recommending profile changes, explain expected trade-off between stability and interruption frequency.

Kill Hierarchy (daemon v20260329.1+)

Condition	Action
MemAvailable ≤ 15%	SIGKILL Chrome/Playwright
MemAvailable ≤ 25%	SIGTERM Chrome/Playwright (deferred if Playwright active)
PSI full avg10 > 25%	SIGTERM Chrome (deferred if Playwright active)
VS Code RSS ≥ 3.8 GB	SIGKILL Chrome (always — even during Playwright sessions); if none → kill_vscode_main()
VS Code RSS ≥ 3.4 GB	SIGTERM Chrome (deferred if Playwright active); if none → kill helper (Shared Process, File Watcher, Network Service); if none → cgroup throttle/reclaim
RSS velocity spike	Kill lowest-value helper — never a language server or Extension Host at normal severity
Chrome PIDs > 3	SIGKILL oldest excess (skipped when Playwright is active)

Playwright Awareness (daemon v20260329.1+, Issue #109)

The daemon detects active Playwright sessions via pgrep -f 'node.*playwright'.

When Playwright is active:

• CHROME-EXCESS cap enforcement is skipped — Playwright legitimately spawns 5-10 Chrome PIDs
• Non-critical Chrome kills (WARN, ACCEL, Stage 2-3) are deferred — callers fall through to helper kills or cgroup throttle/reclaim
• EMERGENCY and Stage 4 always kill Chrome regardless — genuine OOM must be prevented

Before launching Playwright automation:

1. Check memory status: /memwatchdog status
2. If RSS is already above 3.0 GB, consider applying the playwright profile: /memwatchdog tune playwright
3. Close any idle Chrome browsers from previous sessions

If the watchdog kills Playwright Chrome during automation:

• This only happens at EMERGENCY level (≥3.8 GB RSS or ≤15% free memory)
• Consider: closing other VS Code windows, reducing open editor tabs, or restarting the Extension Host to reclaim memory before retrying

Key Repo References

• Snapshot helper — the in-repo memory snapshot script shipped with this skill.

The watchdog daemon, extension activation, dashboard actions, config bridge, and the system-stability design doc live in the upstream VS Code extension repository (the canonical home of this skill), not in this harness checkout.