Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 40 additions & 0 deletions .claude/skills/ci-gates/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -1572,3 +1572,43 @@ example in front of it.
State: 13 gates, 3 with boundary survivors, 8 remaining — two of them proven
equivalences that will never go away and now say so.

## 40. A verification that checked the adjacent thing

A gate written yesterday for a question nobody was asking went red on real data
today, and what it found was eleven blog posts absent from a live site for two
days.

A deploy had replaced the site's application bundle with one built from a
different repository. Eleven posts existed **only** in the old bundle — they had
never been migrated into the shared source — so the swap removed them. They had
no static pages either, so nothing served them at all.

That deploy was verified before pushing. Its commit message says so:

> Verified before pushing: deck/, .claude/, blog/ (29) and ru/ (11) untouched

**Every word of that is true, and it is the wrong question.** It checked the
static trees. The loss was entirely inside the bundle. This is not carelessness —
it is a verification aimed one step to the left of the thing being changed, which
is much harder to notice than no verification at all, because the commit reads as
checked.

**The reusable rule: name what your change REPLACES, and verify that.** The
deploy replaced `assets/`; the check covered `blog/` and `ru/`. Nothing in the
sentence "blog/ untouched" is false and nothing in it is relevant.

**And a gate is not tested until it is red on data you did not plant.** The new
gate had four planted control cases, all passing, before it had ever seen a real
deploy. Replaying it over 160 commits of history found three drops — two healed
by the next deploy, one not. That replay is now a mode of the tool (`--history`)
rather than a script I ran once.

**Which needed its own two directions.** On its first real run `--history`
printed "still missing today: 0" for all three drops — correct, because the loss
had just been repaired, and a mode whose only observed output is green has not
been shown to go red. It now has three planted cases in a real git repository:
an unhealed drop reds, a healed drop stays green and says why, a clean history
does not invent one. The middle one is load-bearing: a drop a later deploy healed
is history, not damage, and a gate that reddens forever over a fixed incident
gets run once.

Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
# NOW — a verification that checked the adjacent thing (2026-08-23)

A gate written yesterday went red on real data today. What it found: eleven blog posts absent from a live site for two days.

- A deploy on 2026-08-21 replaced the site's application bundle with one built from a different repository. Eleven posts existed **only** in the old bundle — never migrated into the shared source — so the swap removed them. No static pages either, so nothing served them at all.
- That deploy **was** verified before pushing: *"Verified before pushing: deck/, .claude/, blog/ (29) and ru/ (11) untouched."* Every word true, and the wrong question. It checked the static trees; the loss was entirely inside the bundle.
- A verification aimed one step to the left of the thing being changed is harder to notice than no verification, because the commit reads as checked.

**The rule:** name what your change REPLACES, and verify that. The deploy replaced `assets/`; the check covered `blog/` and `ru/`.

**And a gate is not tested until it is red on data you did not plant.** Four planted cases all passed before it had seen a real deploy. Replaying it over 160 commits of history found three drops — two healed by the next deploy, one not. The replay is now a mode of the tool, not a script run once.

**Which needed its own two directions.** On its first real run the replay printed "still missing today: 0" for all three — correct, because the loss had just been repaired. A mode whose only observed output is green has not been shown to go red. Three planted cases now: an unhealed drop reds, a healed drop stays green and says why, a clean history does not invent one.

All eleven restored and live, both languages, verified by content on production rather than by exit code.
Loading