trust-guard

Post-edit verification that prevents silent edit failures. After every Write, Edit, or MultiEdit, re-reads files to confirm changes actually saved to disk across all call sites. Assigns a 0-100 trust score per edit. Catches: silent failures (tool reports success but file unchanged), partial application (some call sites missed), ghost writes (subagent claims write but file empty), worktree confusion, MultiEdit partials. Use whenever editing, writing, modifying, refactoring, or creating files — especially multi-file changes, production code, payments, auth, security-sensitive code, or when using subagents. Works with Claude Code, Cursor, Copilot, Codex, and any AI coding agent. Zero executable code, zero dependencies, pure instructions.

By sdcrd schedule Updated 6/5/2026