CacheLint checks application code for prompt-caching mistakes. Prompt caching reuses computation for an identical beginning of a request; an agent can lose those savings while continuing to answer correctly.
The authors inspect 120 merged fixes across ten agent applications, including Hermes. Of those defects, 108 produce no explicit runtime error.
- CodeQL source queries locate caching and usage-accounting sites. A model checks nearby guards; dated provider documentation supplies capability facts.
- On the same defects used to derive the rules, 94 of 132 warnings identify real issues. The detector recovers 94 of 120 labeled defects.
- Across 571 additional repositories, 32 of 37 warnings identify real defects. The authors report 21 maintainer confirmations and eighteen fixes.
Caveat: Recall is in-sample, and cross-request or cross-turn defects fall outside the detector’s single-site model. The larger scan has no exhaustive defect labels for measuring recall.
The paper claims released artifacts, but the inspected project page supplies no visible implementation download. Its dated provider requirements also need fresh verification before use. Version two corrects an author email; the methods and result tables match version one.