I built a local verifier for when Claude Code skips required Skills

Hello! I built Miko after repeatedly running into the same problem while using Claude Code for product work. My projects only had five or six Skills, but once a conversation became long enough, Claude would sometimes stop loading or following them. I ended up pasting the Skill README into the chat or asking “did you actually use the required Skill?” before every important change. Miko is a small local verifier for Claude Code hooks. If Claude tries to edit a protected file before the required Skill or reference has been observed, Miko can pause the action, tell Claude exactly what is missing, and let it retry after reloading the Skill. It also has a REVIEW state for uncertain cases where asking the user is more appropriate than automatically denying everything. I t does not inspect hidden context, prove that Claude understood a Skill, or override Claude Code’s own permission system. It only verifies observable events such as a Skill being loaded, a reference being read, a tool succeeding, or a test actually passing. It keeps a privacy-minimized local JSONL ledger without storing prompts, source code, or tool output. Claude and Haiku helped me work through the early contract and UX design, and I used Claude Code hook runs to test the denial, Skill reload, retry, and completion flow. The verifier itself is deterministic TypeScript rather than another LLM call. Miko is open source, MIT licensed, and completely free to try. The isolated demo does not need an API key: npx --yes koma-miko@alpha probe --host claude Repo: github.com/swnotmetal/Project-Koma I’d especially appreciate feedback on the REVIEW behavior. When required evidence is missing, would you prefer Claude to recover automatically whenever possible, or should Miko ask before every retry?