Two separate false-positive alerts, same root cause: monitoring conjoined liveness with conditional behavior. Separating health signals quiets noise under load.
Stop guessing. We profile a realistic Node.js API under load, surgically find the bottlenecks, fix them with code — not platitudes — and verify every improvement with hard numbers.
After cutting Claude Code tokens by 96% with the Skill Vault, I went looking for what was still eating tokens. Three more sinks, ~5K saved per session, zero capability lost.
Your tool returned 50KB. The agent saw 700 chars. Then it confidently summarized what wasn't there. How tool-result truncation hides in production agents.