prompt-cache-skills

This repository is a collection of drop-in fixes for a specific problem: most popular open-source AI coding agents (tools like Cline, Roo Code, Continue, Aider, and OpenCode) are not taking full advantage of prompt caching, a feature offered by AI providers like Anthropic and OpenAI that can cut API costs by 30 to 90 percent. The bugs are subtle: some agents have caching turned off by default, some have their cache breakpoints placed in the wrong spots, and some handle certain cloud providers (like AWS Bedrock or Google Cloud Vertex) incorrectly.

The project calls its fixes "skills," meaning self-contained instruction files each named SKILL.md that an AI coding agent can read and apply on its own. You point your agent at this repository, tell it to apply every skill that matches the tools you use, and the agent reads each file, checks whether the fix is relevant to your project, applies the code change, and then verifies the result by checking that cached token counts actually appear in API responses. You do not need to understand the underlying code changes yourself.

The repository covers 13 specific, verified bugs across five coding tools. Cline gets fixes for volatile message ordering and a missing OpenAI cache key. Roo Code gets a volatile message fix and a correction for custom AWS Bedrock ARNs. Continue gets default caching enabled, volatile message handling, and explicit Gemini caching. Aider gets caching turned on by default and a longer cache TTL. OpenCode gets proxy detection fixes and a Bedrock document-block correction. Several other agents including Claude Desktop Code, Codex CLI, and Grok CLI were audited and found to already work correctly.

The folder also includes audit write-ups for every tool reviewed, reference documentation on how each AI provider implements caching, a list of 16 common pitfalls, and scripts you can run to confirm that caching is actually working in your setup.