Production Safeguards

Monitors memory + CPU during heavy operations (indexing, vector DB). Intervenes when thresholds are exceeded via a graduated recovery ladder before falling back to a hard emergency stop.

Monitored metrics

Every N seconds:

Metric	Source
Heap used	process.memoryUsage()
Heap total	process.memoryUsage()
RSS	process.memoryUsage()
External	process.memoryUsage()
CPU user	process.cpuUsage()
CPU system	process.cpuUsage()

Limits

Limit	Default	Purpose
maxMemoryMB	1024	Total RSS ceiling
maxHeapMB	512	V8 heap ceiling
maxCpuPercent	80	CPU ceiling
gcThresholdMB	256	Heap size that triggers GC suggestion
alertThresholdMB	400	Heap size that triggers first recovery

Recovery ladder

Each action has a cooldown and max retries to prevent thrashing. Actions run in priority order — the first applicable one fires.

Priority	Action	Trigger	Cooldown	Max retries
1	Clear cache	Heap > alertThresholdMB	30 s	3
3	Reduce batch size	Heap > 80% of alert AND indexing active	60 s	2
4	Pause indexing	Heap > 90% of maxHeapMB AND indexing	2 min	1
5	Restart worker	Heap > maxHeapMB AND indexing	5 min	1
6	Emergency stop	RSS > maxMemoryMB	—	1

Context-awareness: REDUCE_BATCH_SIZE, PAUSE_INDEXING, RESTART_WORKER only fire when indexing is actually in progress (checked via ServiceStatusChecker). CLEAR_CACHE always fires — safe regardless of state.

Circuit breaker

Operations wrapped in executeWithSafeguards() go through a circuit breaker:

graph LR A[CLOSED<br/>Normal] -->|Failures ≥ threshold| B[OPEN<br/>All rejected] B -->|Timeout expires| C[HALF_OPEN<br/>Single probe] C -->|Success| A C -->|Failure| B

Wrapped operations also get:

• Timeout — killed after a deadline
• Retry — configurable count
• Bypass flag — skip the circuit breaker for critical ops

Emergency stop

When RSS > maxMemoryMB:

1. emergencyStopActive = true
2. All subsequent executeWithSafeguards() calls reject immediately
3. Requires manual recovery (CodeBuddy: Reset Safeguards) or extension restart

Prevents the extension from OOM-killing the editor.

Related

• Performance profiler — the metrics this reacts to
• Self-healing — how these safeguards interact with agent-level safety