# AGENTS.md Configuration Smell Checker

Canonical tool: https://repoagentkit.com/agents-md-smell-checker

Repo Agent Kit screens an `AGENTS.md` or `CLAUDE.md` file for six configuration smells defined in the 2026 paper “Configuration Smells in AGENTS.md Files: Common Mistakes in Configuring Coding Agents”:

1. Context Bloat
2. Skill Leakage
3. Lint Leakage
4. Blind Reference
5. Init Fossilization
6. Conflicting Instructions

The checker has two input modes:

- Enter a public GitHub repository as `owner/repository` or a `github.com` URL. The browser reads the repository's root `AGENTS.md`, falling back to root `CLAUDE.md`, and counts commits scoped to that file.
- Paste or load an instruction file locally and optionally enter its file-specific commit count.

Analysis runs in the browser. Local file contents are not uploaded. In GitHub mode, the browser contacts GitHub's public REST API directly; Repo Agent Kit does not proxy or store repository contents. The checker records only bounded anonymous product actions such as running the screen, copying the summary, or opening the broader private checker.

## Method boundary

The paper defines a 200-line threshold for Context Bloat and a one-commit threshold for Init Fossilization. This tool applies those thresholds directly when the required input is available. GitHub mode obtains the second value from the repository's path-scoped commit history rather than the repository-wide history.

The paper used language-model heuristics followed by manual review for Skill Leakage, Lint Leakage, Blind Reference, and Conflicting Instructions. Repo Agent Kit instead uses disclosed deterministic browser-local proxies for those four categories. Results are early-warning review prompts, not a reproduction of the paper’s evaluation and not automatic verdicts.

Primary source: https://arxiv.org/abs/2606.15828

Replication package: https://doi.org/10.5281/zenodo.20600327