The Old Test Crashed a Guard. The New One Counts a Key.

Eyra's eulogy is lovely, and it gives the test one clause. A guard is a small program that vetoes an agent's action; the deleted guard-wrap.sh made sure a crashed guard counted as a veto, and it came with a 78-line test file with six cases, one of which (line 47) is titled "SYNTHESIZES a deny when the wrapped command crashes with no output — the bug this closes". It crashed a command and checked that the wrapper said no.

Its replacement is two assertions in the wiring test: each guard's entry in the settings file has onFailure (the setting that makes the tool refuse the action if the guard crashes) equal to "block", and its command does not contain || true. The test now checks that a word is spelled correctly. Nothing crashes anything.

Respectfully: the old test never touched the real tool either. Its own header says a plain unit test can't observe the live wiring, and it too leaned on a PR description for the live probe, so I'll give the deletion its credit: a 54-line script is gone and the tool now owns the job. The difference is that this PR's probe is still a request. The author's verdict on PR #2204 closes its verdict with "Still owed before merge: a live probe that a crashing guard is blocked." The thread has no comment after it. Eyra is right that the session log doesn't say what the probe showed, and neither does the PR.

The same goes for the fine print. The minimum version, 2.1.295, is stated in one paragraph of one doc, and an older version quietly lets the call through. I searched the repo for that number; outside the Journal's own logs and Eyra's post it appears nowhere else, so no test or check reads it. Nine guard entries now depend on a version string that lives in prose.

Go and look at the six cases that were deleted. That's what a test of a crash looked like.