broneq Test command(s) for this project, one block per tier:
!python3 ${CLAUDE_PLUGIN_ROOT}/scripts/get_settings.py test-tools
If the line above reads run the project's test suite, .bdk/settings.json is not configured — emit one warning line [bdk] .bdk/settings.json not configured — run /bdk:setup then detect from project files (package.json scripts, Makefile, pyproject.toml, go.mod, Cargo.toml, etc.) and proceed under the same policy below.
Pick the form, not just the command
Each block names a tier and up to four forms of the same command. {files} is a placeholder: replace it with the space-separated paths you were given, quoted.
| Form | Run it when |
|---|---|
scoped |
Default. You were given file paths, or you know which test files matter. |
related |
You were given source paths and need the tests covering them. Use this instead of asking another agent which tests cover a change — the runner computes it in a second. |
failed |
Re-running after a fix, when the previous run in this session already failed. |
full |
Only when the caller explicitly asked for a full suite. |
If a block has no scoped form, derive one from full by appending the paths the runner accepts (<full> -- {files} for an npm/yarn/pnpm script, a bare path list for most direct runners). Say in your report that you derived it, so the settings can be fixed once instead of re-derived forever.
Tier policy
- If every path you were given is non-executable content (yaml/md/json/plain config not feeding build or codegen), do not run any tier at all - report
nothing to verify for these paths: <list>instead. Build-feeding config (tsconfig, lockfiles, codegen schemas) counts as source. - A
fasttier is cheap: run it scoped whenever you have paths. - An
e2etier is the most expensive thing in the pipeline. Run it only when either the caller passed you e2e spec paths (because the work touched those specs), or the caller explicitly asked for the end-of-run full gate. Never reach for an e2e tier on your own initiative because a change "might" affect a flow. - The unscoped
fullform of any tier runs once per plan, at the end-of-plan gate. Everywhere else — per task, per group, per fix cycle — is scoped, related, or failed. If you cannot scope a tier you were asked to run and the caller did not ask for a full run, report that rather than silently running everything.
Reporting
Always name the exact command you ran (form and substituted paths) alongside the pass/fail counts. A caller deciding whether a later result still stands needs to know what was actually covered, and "tests passed" without a scope is not an answer.