MANUAL RUN: set the account-level spend caps (Cloudflare alert, LLM key limits, Railway hard limit, Actions $0 budget) #798

Code can't stop an account from billing. These switches can, and only the owner can flip them. COSTS.md has the full checklist. This is the short list, given what's live now.

• Cloudflare (codeSplitters runs on Workers Paid, which has no spend cap). Under Notifications, add a Usage Based Billing alert at a low threshold. Also check which CODESPLITTERS_* flags production was deployed with. Containers for checks, agents and deploys are capped by maxInstances (about $41 a month worst case for checks). Workers AI and Vectorize are only on with CODESPLITTERS_VECTORS=1.
• Anthropic and OpenAI consoles: set a monthly spend limit on the API keys hosted agents use (the keys passed in at src/sandbox.ts). Hosted agent runs have no token or dollar cap in code (codeSplitters: cost caps still missing in code (hosted agents, site-wide container budget, unruled Artifacts writes) #797).
• Railway (the hippo and park test boxes): Workspace → Usage → set a hard usage limit at the plan's included usage.
• GitHub: Settings → Billing → Budgets and alerts. Set the Actions budget to $0 with Stop usage when budget limit is reached. It's free while the repo is public, and this holds if that changes. Cost guards: MCP tool calls count against rate limits; GitHub Actions stay free #796 adds a CI guard against paid runners and missing job timeouts.
• R2 (spa-example's druids-curse-assets): look at usage now and then, since storage grows quietly.

🤖 Generated with Claude Code