PR kirodotdev#11946 created the Real Adapter Contract Tests lane, whose selection is
`-m real_adapter` with a measured floor of 4: three marked codex tests plus
one opencode. Six minutes later PR kirodotdev#11960's rewrite of
test_codex_session_mcp.py replaced those three markers (and the
_require_codex_acp gate behind them) with plain skipif guards -- the exact
skip-on-absent shape kirodotdev#11946 existed to end. The lane now selects 1 test
repo-wide and fails its own floor on every PR:
AssertionError: selected 1 of at least 4:
['test_real_opencode_acp_accepts_the_crew_stdio_element']
Restore the marker and the require gate on the same three tests:
test_real_codex_acp_accepts_the_crew_stdio_element,
test_real_codex_acp_session_close_evicts, and
test_the_installed_adapter_still_builds_the_frames_the_refusal_reads.
The gate keeps skipif's local semantics (absent adapter skips) but FAILS
under KIROCREW_E2E_REQUIRE=1, so the lane cannot green by measuring
nothing. test_real_codex_acp_load_after_close_restores stays unmarked:
it needs a live codex build (CODEX_PATH) and spends real model tokens,
which the lane's runners do not have.
Verified with the lane's exact contract: npm ci of the pinned adapters,
then KIROCREW_E2E_REQUIRE=1 pytest -m real_adapter over both files --
4 selected, 4 passed, none skipped.
What
Restores the three
@pytest.mark.real_adaptermarkers (and the_require_codex_acpgate behind them) that #11960's rewrite oftest/test_codex_session_mcp.pydropped, six minutes after #11946 created the lane that selects on exactly that marker.Why
The Real Adapter Contract Tests lane (#11946) runs
pytest -m real_adapterover the codex and opencode test files and asserts a measured floor of 4 selected tests: three marked codex tests plus one opencode. #11960 replaced those three markers with plainskipif(_ENTRY is None)guards — the exact skip-on-absent shape #11946's own commit message describes as "the one place drift is invisible". The lane now selects 1 test repo-wide and fails its floor on every PR:First observed on #9969 (head cda0737); reproduces on pristine main tip
1c7f96370.How
real_adapter_gateimport and the_require_codex_acp(*, with_node)helper, byte-identical to the pre-feat(acp): continue a codex subagent from its persisted thread #11960 version.@pytest.mark.real_adapter+ the require call on the same three tests that carried them before:test_real_codex_acp_accepts_the_crew_stdio_element,test_real_codex_acp_session_close_evicts,test_the_installed_adapter_still_builds_the_frames_the_refusal_reads.test_real_codex_acp_load_after_close_restores(new in feat(acp): continue a codex subagent from its persisted thread #11960) deliberately stays unmarked: it needs a live codex build (CODEX_PATH) and opts into spending real model tokens (KIROCREW_LIVE_CODEX_PROMPT_TESTS=1), neither of which the lane's runners have.The gate keeps skipif's local semantics — absent adapter skips on a developer box — but FAILS under
KIROCREW_E2E_REQUIRE=1, so the lane cannot report green having measured nothing.Tested
Ran the lane's exact contract locally:
npm ciof the pinned adapters fromtest/real_adapters/package-lock.json, thenKIROCREW_E2E_REQUIRE=1 pytest -v -n0 --no-cov --timeout=1200 -m real_adapterover both files withCODEX_ACP_BIN/OPENCODE_BINpointing at the installed entries: 4 selected, 4 passed, none skipped (codex-acp 1.11.0, the pinned release). Without the adapters installed, the same selection skips all 4 — unchanged developer-box behavior. black / isort / flake8 / brand gate clean.Pattern harvest
Same shape as #11876's census collision: a guard's contract (here the marker set + floor) lives in ci.yml while the guarded surface lives in a test file, and two PRs merged 6 minutes apart — same author — were each green alone and red together. Any rewrite of a file carrying
pytest.markselections that a workflow floor counts should re-run that lane's selection before merge.