claude-human-review или weilliptic-receipts

claude-human-review adds a human approval gate at the end of every Claude Code turn that edits files. When Claude tries to stop, the plugin pauses it and has Claude do two things before the turn ends. First, drop a one-line doc comment on any class, function, or non-obvious config block it created or materially changed, using the right style per language (JSDoc for TS/JS, docstrings for Python, // for Go, and so on). Second, write a short plain-English review of what changed and why, flagging risks the diff alone won't show.\n You reply with 1 (Approve) to keep the changes or 2 (Undo) to roll them back. Undo is deterministic: filesystem snapshots are taken before each edit, so it works in git repos, non-repos, and folders full of untracked files. State lives on disk, so it survives Claude's context compaction on long sessions.\n The review is tuned to stay around 60 to 120 words per turn and skips sections when they would be filler, so the extra output-token cost is small. Inline comments are capped to one li

Claude Code plugin that audits AI-assisted code changes. Automatically records every file edit, session, and commit as a verifiable receipt providing tamper-proof provenance for AI-generated code.