cache-guard

CI guard for LLM prompt-cache prefixes — fails the build when a system prompt, tool list, or first message changes unintentionally.

Why: providers (Anthropic, OpenAI, Gemini, Grok) serve repeated prompt prefixes from cache at ~10% of list price. One "harmless" edit to a system prompt silently invalidates that cache in production and 10x-es the token bill. cache-guard pins the cache-relevant prefix blocks of your prompt fixtures and turns that silent regression into a red build.

Usage

Keep one or more request-body fixtures (Anthropic Messages API shape) in your repo, then:

npx cache-guard snapshot fixtures/*.json   # write .cacheguard.json baseline
npx cache-guard check    fixtures/*.json   # exit 1 if any prefix hash changed

In CI:

- run: npx cache-guard check fixtures/*.json

When a change is intentional, refresh the baseline with snapshot and commit .cacheguard.json.

• CACHE_GUARD_FILE — override the baseline path (default .cacheguard.json)
• Hashes cover the blocks that decide cache hits: tools, system, and the first message. Nothing else in the fixture matters to the guard.

Made by Caching.ai — a drop-in proxy that watches, protects, and keeps your LLM prompt cache warm.