load-shedding

Cria/Atualiza:

• Patches no target_path
• supabase/functions/_shared/load-shedder.ts (se não existe)
• supabase/functions/_shared/token-bucket.ts (se rate-limit aplicado)

Após: handler retorna 503 + Retry-After quando saturated, em vez de aceitar e cair. Defesa central contra cascade.

Exemplos:

/load-shedding supabase/functions/process-orders/index.ts
/load-shedding src/handlers/webhook.ts --max-concurrent 500
/load-shedding supabase/functions/api/index.ts --patterns concurrency-limit,deadline-aware

1. Parsear argumentos

TARGET_PATH=$(echo "$ARGUMENTS" | awk '{print $1}')
MAX_CONCURRENT=$(echo "$ARGUMENTS" | grep -oE -- '--max-concurrent [0-9]+' | awk '{print $2}')
QUEUE_MAX_SIZE=$(echo "$ARGUMENTS" | grep -oE -- '--queue-max-size [0-9]+' | awk '{print $2}')
PATTERNS=$(echo "$ARGUMENTS" | grep -oE -- '--patterns [^ ]+' | awk '{print $2}')

[ -z "$MAX_CONCURRENT" ] && MAX_CONCURRENT=1000
[ -z "$QUEUE_MAX_SIZE" ] && QUEUE_MAX_SIZE=10000

if [ -z "$TARGET_PATH" ]; then
  echo "ERROR: target_path obrigatório"
  exit 1
fi
[ ! -f "$TARGET_PATH" ] && { echo "ERROR: arquivo não encontrado"; exit 1; }

2. Dispatch para load-shedding-instrumenter

Task(
  subagent_type="load-shedding-instrumenter",
  prompt="
target_path: ${TARGET_PATH}
max_concurrent: ${MAX_CONCURRENT}
queue_max_size: ${QUEUE_MAX_SIZE}
${PATTERNS:+patterns: ${PATTERNS}}

Aplicar skill load-shedding-graceful-degradation. Patches:
1. Concurrency limit + 503 graceful (LoadShedder em _shared)
2. Deadline-aware handler (x-deadline-ms header parsing)
3. Queue bound + drop oldest (FIFO drop)
4. Server-side rate limit (token bucket per-client)
5. Slow start state machine (5 min ramp em recovery)

Verify compilação verde após patches. Output curto com lista de patches aplicados.
"
)