The CLI resolves each client's config home from its own environment
variable before falling back to $HOME — cbm_codex_config_dir() honours
$CODEX_HOME and ignores the home_dir it is handed. A test that redirects
only HOME therefore does not isolate those clients: it resolves, reads
and WRITES the developer's real configuration.
On a machine with $CODEX_HOME exported (any Codex or Orca user) 27 of
269 cli tests failed for that reason alone. Every one of them asserted a
successful rc while the CLI was reporting genuine failures against the
ambient config: an unbalanced managed marker left in the real
config.toml, an AGENTS.md that is a symlink and is refused by the
O_NOFOLLOW writer, and user-modified agent profiles that are correctly
preserved. The behaviour under test passed in all of them; only the rc
was poisoned. Worse than the red: the suite mutated that real config,
installing and removing SKILL.md under the user's Codex home.
The individual tests that already neutralize CODEX_HOME are exactly the
ones that pass, and the file carries a comment admitting one suite was
"green only via env leaked from earlier suites". Rather than extend those
ad-hoc per-test lists again, neutralize the whole inventory once in the
runner, and reuse the inventory the shell fixtures are already required
to satisfy.
tests/test_smoke_fixture_contract.sh grows one check so the two cannot
drift: a client isolated in the smoke scripts but forgotten in the C
runner now fails the contract.
Verified on macOS arm64 with $CODEX_HOME and $OPENCODE_CONFIG_DIR
exported: cli went from 241 passed / 28 failed to 268 passed / 0 failed,
with no change outside tests/.
Signed-off-by: Gabriel Gonzalez Pineda <64bits@gmail.com>
Gabriel Gonzalez Pineda
6d09c19e11