Skip to content

feat(quality): let a repo opt into gate-46 blocking on @e2e anchors - #741

Merged
rubenvdlinde merged 1 commit into
mainfrom
feat/hydra-gates-e2e-anchor-blocking-input
Sep 10, 2026
Merged

rubenvdlinde merged 1 commit into
mainfrom
feat/hydra-gates-e2e-anchor-blocking-input

Conversation

@rubenvdlinde

Copy link
Copy Markdown
Contributor

The invitation had no acceptor

Gate-46 resolves @spec and @e2e by identical rules, but only @spec blocks. The @e2e half warns:

Report-only: set HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING=1 for this repo once they are worked down.

Nothing could act on that. The gate step's env: block is fixed in quality.yml, and a reusable workflow does not inherit env from its caller, so an app repo could not set the variable however it wrote its own workflow.

Measured across the organisation today:

$ gh search code --owner ConductionNL 'HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING'
ConductionNL/.github:hydra-gates/scripts/lib/test_gate_46_tests_scope.sh
ConductionNL/.github:hydra-gates/scripts/lib/test_gate_discarded_counts_and_empty_deltas.sh

The runner and the runner's own test. Zero repos had opted in, and none could have.

What this adds

A hydra-gates-e2e-anchor-blocking input, default false, wired to HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING in the gate step.

Default stays false, and must. #726 counted 194 dangling @e2e anchors fleet-wide, 154 of them in four repos (dossiq 46, zaakafhandelapp 43, pipelinq 36, decidesk 29) — the four with the most end-to-end coverage. Blocking on day one turns them red on debt no PR introduced. The count prints on every run either way; this changes the verdict, not the visibility.

Passed as an explicit '1'/'0'. The runner compares against the literal "1". A GitHub boolean interpolates as "true"/"false", which reads as "not 1" — a flag that says it is on, a gate that is off, and nothing in the log to tell them apart.

One input per flag, deliberately. A single passthrough for arbitrary HYDRA_GATE_* variables would also let a caller silently switch OFF a blocking gate, which is the opposite of what these opt-ins are for. Each opt-in stays a visible, named line in the caller's workflow.

Not wired here

Gates 112, 113, 114 and 115 carry the identical invitation and are equally unreachable today. Each is a separate decision about a separate backlog. This one has a repo behind it now: openregister's @e2e count is measured at zero (ConductionNL/openregister#3573), which is what the gate asks for before opting in.

Verification

The runner is unchanged; only the workflow gains the input. The behaviour this exposes was already pinned by the runner's own suite, green on this branch:

$ bash scripts/lib/test_gate_46_tests_scope.sh
  ok   — a dangling @e2e anchor → WARNING
  ok   — a dangling @spec alongside a dangling @e2e → FAIL, not WARNING
  ok   — HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING=1 turns the warning into a failure
test_gate_46_tests_scope.sh: ALL GREEN

quality.yml re-parsed after the edit, rather than the diff being read as evidence: the input is declared with type: boolean, default: false, and the env var resolves on the step with id: gates.

Ordering

This must merge before ConductionNL/openregister sets the input. Callers resolve quality.yml@main, and an input main does not declare is a workflow error, not a no-op.

🤖 Generated with Claude Code

Gate-46 resolves `@spec` and `@e2e` by identical rules, but only `@spec`
blocks. The `@e2e` half is report-only, and its warning ends "set
HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING=1 for this repo once they are worked
down".

That instruction was unreachable. The gate step's `env:` block is fixed in
quality.yml, and a reusable workflow does not inherit `env` from its caller,
so no app repo could set the variable however it wrote its own workflow.
Measured across the organisation today: the string appears nowhere except
the runner and the runner's own test. Zero repos had opted in, and none
could have — the gate has been inviting an opt-in that did not exist.

This adds the input that accepts it. Default false, and it must stay false:
#726 counted 194 dangling @e2e anchors fleet-wide with 154 of them in four
repos, which are the four with the most end-to-end coverage. Blocking on day
one would turn them red on debt no PR introduced.

The value is passed as an explicit '1'/'0' rather than the boolean's own
string form. The runner compares against the literal "1"; a GitHub boolean
interpolates as "true"/"false", which reads as "not 1" and would leave the
gate report-only in a repo that had opted in — a flag that says it is on, a
gate that is off, and nothing in the log to tell them apart.

One input per flag, deliberately. A single passthrough for arbitrary
HYDRA_GATE_* variables would also let a caller silently switch OFF a
blocking gate, which is the opposite of what these opt-ins are for.

Gates 112, 113, 114 and 115 carry the identical invitation and are equally
unreachable today. They are not wired here — each is a separate decision
about a separate backlog, and this one has a repo behind it: openregister's
@e2e count is measured at zero.

Runner side unchanged and already proven: scripts/lib/test_gate_46_tests_scope.sh
asserts "HYDRA_GATE_SPEC_ANCHOR_E2E_BLOCKING=1 turns the warning into a
failure", and the suite is green on this branch.
@rubenvdlinde
rubenvdlinde merged commit 477f930 into main Sep 10, 2026
39 checks passed
@rubenvdlinde
rubenvdlinde deleted the feat/hydra-gates-e2e-anchor-blocking-input branch September 10, 2026 10:15
rubenvdlinde added a commit to ConductionNL/openregister that referenced this pull request Sep 10, 2026
* fix(tests): point the mdm @e2e anchors at their canonical specs

Twenty-eight `@e2e` anchors in tests/e2e/spec-coverage/ named a change
directory rather than the capability's canonical spec. All three changes
they named — mdm-frontend, mdm-merge-ui, mdm-survivorship-override — were
archived on 2026-07-03, so each path has already moved once.

An anchor belongs at `openspec/specs/<capability>/spec.md#<scenario>`. A
spec path is stable; a change path moves on archive, and that move is what
broke the fifteen anchors repaired earlier today. These twenty-eight still
resolve only because gate-46's capability index spans all three homes of a
spec — in flight, archived, canonical — which rescues the reference without
making it honest about where the requirement lives.

Every anchor keeps its fragment. The prefix is the only thing rewritten, so
each line still names the requirement or scenario the test exists for rather
than pointing at a whole file, which resolves while saying nothing.

Verified with the gate's own resolver, scripts/lib/check_spec_anchors.py at
.github@main, over gate-46's full scope of 3,713 files: 0 unresolved
findings across 9,948 @SPEC and 163 @e2e tags. The resolver was
mutation-checked first — an invented capability and an invented fragment
both reported — so the zero is a measurement, not an empty run.

Three anchors are deliberately left naming a change directory. They point
into `mdm-views-route-scoping-e2e`, which is still live and has no spec at
`openspec/specs/mdm-views-route-scoping/spec.md`. Repointing them there
would still pass the gate, because the capability index would resolve it
back to the live change — an anchor naming a file that does not exist and
reading as though it does. They move when that change is archived and its
spec is promoted.

* chore(ci): block on unresolved @e2e anchors in gate-46

Gate-46 resolves `@spec` and `@e2e` by identical rules but blocks only on
`@spec`; the `@e2e` half is report-only until a repo opts in. This is that
opt-in, and it is set only because the count reached zero first.

Measured with the gate's own resolver — scripts/lib/check_spec_anchors.py
from ConductionNL/.github@main, not a vendored copy and not a
reimplementation — over gate-46's full scope of 3,713 files, after #3573
repointed the last 28 change-directory anchors:

    total findings: 0
    @SPEC findings: 0
    @e2e  findings: 0

across 9,948 @SPEC and 163 @e2e openspec tags. The resolver was
mutation-checked before the zero was believed: an invented capability and an
invented fragment were both reported.

One line, its own commit, so it can be reverted without touching the
repoints it depends on.

REQUIRES ConductionNL/.github#741, which adds the input. Callers resolve
quality.yml@main and an input main does not declare is a workflow error, so
this cannot merge first.

---------

Co-authored-by: Conduction Release Bot <release-bot@conduction.nl>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant