Test tiers

./check.sh runs at the end of every response. A gate that takes an hour is either skipped (nothing is checked) or blocks the session for an hour: both are bugs. The default gate stays under ~5 minutes; everything slower moves to a full tier that runs less often, never away.

What to run when

A full-tier run is a background job: start it once with the shell tool's maximum timeout, write its log to a unique file (/tmp/<project>-full-$$.log — parallel sessions share /tmp), keep working on something that doesn't touch the code under test, and read the result before calling the work done. Never start a second copy while one runs, and never rerun it to see whether a flaky test passes — fix or quarantine the flaky test.

Split the suite (once per project)

1. Measure first. A few tests are usually most of the hour: pytest --durations=20, go test -json ./... (the Elapsed field), cargo nextest run (it prints slow tests), Jest --verbose. Typical causes, cheapest fix first: real sleeps and timeouts (fake the clock), a fresh DB or container per test (one per session, a transaction per test), live network calls (recorded fixtures — AGENTS.md Testing rule), a serial runner (workers), e2e tests repeating what unit tests already prove (delete the duplicate).
2. Fast tier = lint, type check, code-health, unit tests: no network, no real browser, DB server or GPU, no sleeps. Full tier = fast tier + integration, e2e, browser, GPU or model, large fixtures, performance.
3. Mark slow tests where they live — the framework's marker or a folder (tests/unit/ vs tests/integration/), never a list of file names inside check.sh.
4. One gate, one flag. check.sh takes --full (its check.ps1 twin the same flag):

   #!/usr/bin/env bash
   set -euo pipefail
   full=0; [ "${1:-}" = "--full" ] && full=1
   # lint, type check, code-health …
   if [ "$full" = 1 ]; then uv run pytest -n auto; else uv run pytest -n auto -m "not slow"; fi

5. Write the durations down in PROJECT_STATE.md Key Commands, so the next agent knows what it is starting: ./check.sh # fast tier, ~2 min and ./check.sh --full # + integration/e2e, ~50 min — CI, end of feature. Name the full tier in the AGENTS.md Testing rule and point CI (.github/workflows/check.yml) at it.

Per framework

No marker mechanism fits → two scripts called by check.sh (tests/fast.sh, tests/full.sh), still behind the one --full flag.