Guardrails outlineExperimental — needs a TypeSafe key; calibrate the confidence threshold on your own repos

Stray file pre-commit gate

Who: Teams whose agents leave scratch scripts and dumps in the working tree. Steps: On pre-commit, run local deny rules first; for other new paths ask Jev to pick belongs versus scratch, dump, build output or local-only; block when confidence that it does not belong clears your threshold. Expected effect: fewer accidental junk commits without maintaining an endless ignore list.

Source: https://github.com/Eve2512/Strayguard ↗ Docs / source ↗

Treat this as an outline — adapt state and questions to your data.

Implement in code — call System One / Jev; compose answers yourself.

Gate on confidence — act, confirm, or escalate before side effects.

Outline sketch

Prompt

# Stray file pre-commit gate
State: path, size, and a short preview of each newly staged file.
Ask: Choice among belongs / scratch / dump / build output / local-only.
Then: block the commit when non-belong confidence clears your threshold; never send files matched by local secret rules.

Needs access to: gitpre-commit

Who it's for

Teams whose coding agents leave scratch files in the working tree

Steps / how it's set up

Steps (experimental)

Hook pre-commit so only newly added paths are inspected.
Apply local deny rules for secrets and known junk names without calling Jev.
Ask Jev one Choice question per remaining file and gate on confidence in code.

Sources

https://github.com/Eve2512/Strayguard

This pattern is experimental. Author-reported behavior — calibrate on your own commits.

Expected effect

Catches agent leftover files that ignore rules miss

Unofficial outline for learning. Paraphrased from public docs and tutorials — not a production recipe. Review sources before you automate anything.