tidy-skill

Case: synthetic dirty repo → clean layout

Kind: reproducible lab fixture (temp directories)
Scripts: score_repo_hygiene.py, audit_agent_artifacts.py
Source: tools/run_evals.py cases dirty_repo_score_below_80 / clean_repo_score_at_least_90
Run date: 2026-08-05 (author machine)

Before

Fixture layout:

dirty/
├─ plan.md
├─ todo.md
└─ README.md

Observed:

Metric	Value
Hygiene score	45 / 100
Suspicious root files	plan.md, todo.md
.agent_tmp/	missing
.agent_reports/	missing

Agent behavior this models: mid-task process Markdown dumped at repo root with no lifecycle.

After

Fixture layout:

clean/
├─ README.md
├─ LICENSE
├─ CHANGELOG.md
├─ .gitignore          # includes .agent_tmp/ and .agent_reports/
├─ docs/guide.md
├─ .agent_tmp/.gitkeep
└─ .agent_reports/.gitkeep

Observed:

Metric	Value
Hygiene score	100 / 100
Suspicious root files	0
.gitkeep counted as artifacts?	no

Delta

	Before	After
Score	45	100
Root process Markdown	2	0
Placement dirs	0	2

Reproduce

uv run python tools/run_evals.py
# see docs/evals/latest.md

Limits

• Synthetic paths only; not a customer monorepo.
• Scoring model is this project’s own rubric.
• Does not measure agent behavior over multi-hour sessions.

This site is open source. Improve this page.