Stop Me Check

stop-me-check

Run the stop-me gates backwards over code that already exists. Report only what would have failed all three.

Method

1. Read .stop-me.yml if present, else infer horizon from the repo (deploy targets, tenancy columns, contributor count) and say which you assumed.
2. Read DECISIONS.md if present. Anything already logged there is settled — never report it.
3. Sweep the irreversible surfaces from the stop-me skill — that list lives there and only there; never keep a second copy here. For each, gather evidence before you claim anything:
   • migrations/ — drops, renames, narrowing types, missing down-path
   • money paths — charge/refund/quota, and whether an idempotency key and a test exist
   • secrets — literals in tracked files; git log -S for the commit that introduced them
   • auth — password hashing, token expiry, permission checks on write paths
   • tenancy — tables missing a tenant column that siblings have
   • published shapes — unversioned endpoints, any on exported types, CLI flags
   • deletes — DELETE/drop/rm -rf without a restore path
   • data not collected — events and audit rows with no actor or timestamp
4. Rank by blast radius, not by count.

Output

stop-me check · <repo> · horizon: <h> (<inferred|from .stop-me.yml>)

<n> findings on irreversible surfaces

🛑 <file:line> — <one line>
   [evidence] <what you actually read>
   [cost] <what undoing this now requires>

...

Clean: <surfaces you checked and found nothing on>

End with the Clean: line every time. A list of what you checked and cleared is what makes the findings above it believable.

Rules

• Cite file:line for every finding. No citation, no finding.
• Report what the code does, never what the author was thinking.
• Reversible problems are not findings. A slow query, an ugly name, a missing type on an internal function — all silence.
• Zero findings is a real result. Print the Clean: line and stop. Do not pad.