ruleherd
A linter tool for auditing AI coding agent instruction files, such as CLAUDE.md and AGENTS.md, to detect configuration drift. It performs inspection rather than migration and supports detection of seven categories of issues.
Timeline (last 90 days)
Developer princejain756 has open-sourced ruleherd on GitHub, a linter tool for auditing whether AI coding agent instruction files have drifted, positioned as doing no migration, only health checks.
ruleherd is run with the command npx ruleherd, and can scan all instruction files in a repository and report seven categories of issues: CONFLICT, DEAD REF, DEAD CMD, STALE, DEAD GLOB, DUPLICATE, and BLOAT.
ruleherd's BLOAT detection rule is: a single instruction file exceeding about 4k tokens is considered bloated, because it is loaded into the prompt every time and wastes the context window.
ruleherd gives a score out of 100, supports strict mode, and exits with a non-zero code in CI; it has zero dependencies and uses the MIT license.
The author ran ruleherd on his own machine (including scanning the global configuration) and got GRADE B (85/100), scanning out 9 instruction files, about 4,757 tokens of agent context, and three duplicated passages each appearing in 6 files.
ruleherd's contradiction detection is heuristic, based on keyword pairing and never/always polarity judgment, not semantic understanding, so it may miss subtle contradictions or falsely flag cases with different wording but the same meaning.
ruleherd's stale-reference check is deliberately conservative, preferring to miss detections rather than produce false positives; its first version falsely flagged @types/node as a stale file reference and pnpm install as a missing npm script.
The 'clean' rating given by ruleherd only means no drift was detected, not that the rules are well written.
ruleherd is suitable for integration into CI pipelines to check configuration drift on every commit.
All Facts (9)
The 'clean' rating given by ruleherd only means no drift was detected, not that the rules are well written.
50%Unverifiedruleherd is suitable for integration into CI pipelines to check configuration drift on every commit.
50%Unverifiedruleherd's BLOAT detection rule is: a single instruction file exceeding about 4k tokens is considered bloated, because it is loaded into the prompt every time and wastes the context window.
50%Unverifiedruleherd gives a score out of 100, supports strict mode, and exits with a non-zero code in CI; it has zero dependencies and uses the MIT license.
50%Unverifiedruleherd is run with the command npx ruleherd, and can scan all instruction files in a repository and report seven categories of issues: CONFLICT, DEAD REF, DEAD CMD, STALE, DEAD GLOB, DUPLICATE, and BLOAT.
50%Unverifiedruleherd's contradiction detection is heuristic, based on keyword pairing and never/always polarity judgment, not semantic understanding, so it may miss subtle contradictions or falsely flag cases with different wording but the same meaning.
50%Unverifiedruleherd's stale-reference check is deliberately conservative, preferring to miss detections rather than produce false positives; its first version falsely flagged @types/node as a stale file reference and pnpm install as a missing npm script.
50%UnverifiedThe author ran ruleherd on his own machine (including scanning the global configuration) and got GRADE B (85/100), scanning out 9 instruction files, about 4,757 tokens of agent context, and three duplicated passages each appearing in 6 files.
50%UnverifiedDeveloper princejain756 has open-sourced ruleherd on GitHub, a linter tool for auditing whether AI coding agent instruction files have drifted, positioned as doing no migration, only health checks.
50%