Content-Addressed Caching Can Cut Redundant AI Model Calls in CI Pipelines

Repeated identical prompts in CI/CD pipelines consume free model quota unnecessarily, even when inputs and templates have not changed between runs. A content-addressed caching approach uses SHA-256 hashes of normalized request components — including the prompt, model route, and sampling parameters — as cache keys to detect duplicate requests before making a new model call. A shell script stores responses locally in a .model-cache directory and validates them against a configurable TTL, skipping the model call entirely on a cache hit. GitLab CI's native cache mechanism can restore these files across jobs, with the cache key tied to the prompt template file so stale entries are automatically invalidated when prompts change. Extending the cache to a shared HTTP endpoint further multiplies savings across branches, runners, and pipelines.