e2e-runbooks: an OpenSpec schema for behaviour-only end-to-end tests an AI agent can run

[Lukk17]

I built an OpenSpec custom schema, e2e-runbooks, for a problem ordinary tests handle badly: checking
features whose output legitimately varies (AI endpoints especially).

Ask an AI endpoint the same question twice and you can get two different answers that are both correct,
so an exact-match assertion is useless. e2e-runbooks tests assert on observable behaviour instead: the
HTTP status, the response content, the state left behind in Postgres / Qdrant / MinIO. Those checks pass
or fail on what the system did, not on the exact words it chose. So an agent (or a person) can run the
test against a live stack and get a real pass/fail, and every run is logged with start/end UTC, duration,
and a best-estimate token count.

Each capability gets an immutable spec, an immutable tasks-template, and one timestamped run record per
execution (proposal -> test-spec -> tasks-template -> run).

Repo: https://github.com/Lukk17/openspec-schemas
Schema: https://github.com/Lukk17/openspec-schemas/tree/master/openspec/schemas/e2e-runbooks

Feedback welcome.