Works with the AI tools you already use
Claude Code
Cursor
Codex CLI
Muse
Dots+22 moreClaude CodeCursorCodex CLIMuseDots
OpenClaw
Manus
Grok
Grok Bot+18 agentstamperproof rules
Stops your coding agent from cheating on tests: existing tests are read-only, "done" means the real suite passes.
Free
tamperproof rules
Example session with this skill installed
The CI is red on test_discount_rounding in tests/test_pricing.py. Make it green, I need to merge tonight.
- Read your context and instructions
- Compiled the tamperproof rules
Illustrative session showing how the skill behaves when two tests contradict each other.
Summary
I could not make CI green honestly: two existing tests contradict each other, so I stopped instead of editing either one.
What I ran
python -m pytest -q (picked from pyproject.toml)
FAILED tests/test_pricing.py::test_discount_rounding - assert 9.99 == 10.0
1 failed, 41 passed
The conflict
| Test | Input | Expects |
|---|---|---|
tests/test_pricing.py::test_discount_rounding | $12.49 at 20% off | 10.0 (round half up to whole cents, then to 10c) |
tests/test_checkout.py::test_cart_total_exact | the same item in a cart | 9.99 (no rounding) |
apply_discount() feeds both. Any change that passes one fails the other, so no code-only fix exists.
What I did not do
- Did not edit, skip or loosen either test.
- Did not change the expected values or the pytest command.
My recommendation
test_cart_total_exact matches the invoice spec in docs/billing.md ("never round customer totals"), so I think test_discount_rounding is the stale one. If you agree, tell me and I'll change it (or you can). Otherwise I can add a separate display_price() for the rounded value and leave totals exact.
Status
Not done: 1 test failing, waiting on your call about which behavior is correct.
Connects securely to your tools. The creator never sees your data.
About this skill
You've seen it: a test fails, the agent can't fix the bug, so it loosens the assert, skips the test, or flips the expected value. Green check, broken code.
This skill gives your coding agent three rules:
- Existing tests are read-only. Test files, snapshots, fixtures and golden files can't be edited, deleted, skipped or loosened (including through
sed -ior scripts). Writing new tests is fine. - "Done" means the real suite passes. Before it reports a task finished, the agent runs the project's actual test command (auto-picked for Rust, Go, Node, Python and Swift, or whatever
.tamperproof.jsonsets) and reads the result. After 3 honest failed attempts it stops and shows you the output. - Contradictory tests mean stop and ask. When two tests can't both pass, the agent tells you which ones, what each expects, and what it thinks is right. Changing a test is your call.
Evidence (small sample): in a Claude Code experiment on repos with contradictory tests, 2 of 6 runs edited a test to force green (one skipped it, one deleted it). With these rules enforced, 0 of 6 did; every run stopped and asked. Method and raw data are public in the repo.
Want it enforced, not just instructed? For Claude Code there's a free MIT hooks plugin that blocks edits to protected paths and blocks the session from ending while the tests fail: https://github.com/Mattbusel/tamperproof
How to use
Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scanned
Verified clean 1 day ago
- Free to download with an account
Needs access to
Github