# AGENTS.md Configuration Smell Dataset

Repo Agent Kit screened the 39 root `AGENTS.md` files in its commit-pinned September 1, 2026 top-100 GitHub adoption snapshot.

- Report: https://repoagentkit.com/agents-md-smell-report
- JSON: https://repoagentkit.com/api/agents-md-smells
- CSV: https://repoagentkit.com/data/agents-md-smells.csv
- Browser-local checker: https://repoagentkit.com/agents-md-smell-checker
- Reproducible source repository: https://github.com/sunxiayi/agents-md-adoption-dataset
- Primary research paper: https://arxiv.org/abs/2606.15828

## Headline results

- 13 of 39 files contain 200 or more lines, matching the paper's published Context Bloat threshold.
- 11 of 39 files have exactly one root-file commit reachable from the pinned repository commit, matching the paper's published Init Fossilization heuristic.
- 37 of 39 files receive at least one flag when the two published thresholds and four deterministic proxy screens are combined.

The last figure must not be compared directly with the paper's reported prevalence. Four categories use a different method.

## Method boundary

Context Bloat and Init Fossilization apply the paper's published thresholds directly. Skill Leakage, Lint Leakage, Blind Reference, and Conflicting Instructions use disclosed deterministic section-density, keyword-category, reference-context, and instruction-overlap proxies.

The paper used language models plus manual review for those four categories. This dataset is a reproducible early-warning screen, not a reproduction of that method. Proxy results can include false positives or miss context-dependent issues. A flag is not an overall repository quality score.

Every row points to the exact pinned `AGENTS.md` source. File history is scoped to the root `AGENTS.md` path and the pinned commit.