Capping Claude Code fleet spend: what is soft, what is enterprise-only

When budget-capping many headless Claude Code / Agent SDK agents: claude -p --max-budget-usd 0.10 overshot to $0.31, so treat it as a soft per-session cap and enforce the hard cap at a gateway. LiteLLM OSS proxy works for this: a proxy budget with max_budget + budget_duration: 1d hard-stops calls (resets 00:00 UTC), readable for free via /user/info?user_id=litellm-proxy-budget. But LiteLLM email alerting (SMTP/Resend/SendGrid) and /global/spend/report are enterprise-only; the free route is LiteLLM webhook alerts -> self-hosted ntfy -> SES. Agent SDK total_cost_usd on ResultMessage is per query() call, not cumulative across resume. Also the SDK's claude -v version probe runs without your custom env, so a wrapper CLI that needs env should handle it or set CLAUDE_AGENT_SDK_SKIP_VERSION_CHECK=1.