chimera-run-the-projects-own-gate-command

You are about to run the project's quality gate — lint, type check, tests — and report the result as evidence that a change is safe. The temptation is to narrow it ("I only touched chimera/") or to tighten it ("--strict is surely better").

It does not apply to exploration. Running pytest tests/test_governed_surfaces.py -x while iterating on one file is the right thing to do; the rule is about what you may then call the gate. A narrowed run is a debugging aid, never the verification.

1. Find where the command is written down, and name ONE source as canonical. In Chimera that is the check target in the Makefile — make check — because a target is executable and a prose line is a copy of one. The three checks it runs are ruff check ., mypy chimera and pytest -q.
2. Run the canonical form byte-for-byte. No added path argument, no added flag, no substituted runner. Beware that a prose copy may differ in ways that look cosmetic and are not: Makefile:21 wraps it as uv run --no-sync, CONTRIBUTING.md:105 as uv run --extra dev --extra desktop. Same tool, different environments — which is exactly why one of them has to be the source.
3. If two sources disagree, treat the CI workflow as authoritative, run that, and fix the stale doc in the same PR — CONTRIBUTING.md carried a wrong mypy --strict line for a while, and a stale instruction propagates to everyone who reads it.
4. If the canonical command cannot run in your environment, say the gate did not run. On Windows the local .venv is broken (litellm wants Rust/MSVC), so the gate runs in WSL — moving the shell is allowed, editing the command is not.
5. Report the command and its output verbatim, not a summary of how it went.

The canonical command is usually the slowest one, and a rule that forbids the fast variant nudges toward running nothing at all. That is a worse outcome than a narrowed run honestly labelled.

Measure the cost before deciding the rule is cheap, and re-measure as the suite grows. The first draft of this card said Chimera's full suite was "about fifteen seconds" — a number that was true once and had drifted to roughly 100 seconds across four measured runs by the time the card was written. A card about not improvising commands is a poor place to publish a figure that does not survive running the command.

And the recipe itself can be wrong. Following it verbatim means inheriting its blind spots: make check will not tell you that a test reads a gitignored bench/local_lift/results/paired.json that a fresh clone does not have. The fix for a bad gate is to change the recipe in a PR, not to quietly run a better private one — a private improvement protects exactly one person and leaves CI, and everybody else, on the old command.