Why your AI agent costs more after midnight

Prompt caching lets an AI provider reuse work on the unchanged start of your prompt, but only if that prefix stays identical. Neuralwatt found coding agents like OpenCode 1.x and Kilo Code put today’s date in the system prompt, so crossing midnight forced the model to re-read almost the entire conversation. The date isn’t the only trigger. Switching reasoning effort, connecting an MCP server, or adding a skill mid-session can break the cache too, and deep in a long session, one full re-read can cost as much as hundreds of normal turns. The fix: keep the system prompt fixed, leave out dates and git status, keep the tool list stable, and append changes at the end. #PromptCaching #AIAgents #LLM #AIEngineering #CodingAgents