health

name

health

description

Take the target repo's health pulse on demand — run its test/lint/typecheck (and any extra) checks and emit one pass/fail summary. Auto-detects the repo's checks (package.json scripts or Makefile) with explicit --cmd override. Opt-in; not wired into the default pipeline. Output streams live but is never captured (secret-safe). Use to answer "is the repo green right now?" before shipping or after a major change.

/health — is the repo green right now?

Goal: one command that runs the repo's checks (test, lint, typecheck, …) and tells you PASS or FAIL — on any project, with a secret-free record. This is a green/red pulse, not a coverage gate (the pipeline /harness-claude:test skill owns the 80% floor).

Opt-in. No default-pipeline skill or hook invokes /health. It shell-executes the repo's own check commands with your privileges — run it deliberately. Git boundary: it only reads and runs; it never commits, pushes, or branches.

When to run

• Before /harness-claude:ship, or after a major change, to confirm the repo is still green.
• On a fresh checkout / unfamiliar repo, to see what passes with one command.
• As the --score half of a /benchmark run, or any time you want a quick health check.

Secret-safety (hard invariant)

Each check's output streams live to your terminal (same as running the command yourself) but is never captured — it cannot reach the summary, the JSON artifact, or any log. The artifact records only name / command / pass / durationMs + the gate. A secret printed by a check stays on your tty and out of every persisted file.

Discovery

With no --cmd, the runner auto-detects checks:

• package.json scripts — test, lint, typecheck (run as <pm> run <name>; the package manager is detected from the lockfile — pnpm / yarn / npm). Non-standard scripts (e.g. build) are ignored.
• else Makefile — test and lint targets (make test / make lint).

--cmd '<shell>' (repeatable) replaces the detected set when present.

Do this

1. Inspect (optional): see what will run without running it.

   node scripts/eval/continuous.js --list
   

2. Run the pulse:

   node scripts/eval/continuous.js                       # auto-detect
   node scripts/eval/continuous.js --cmd 'pytest -q' --cmd 'ruff check .'   # explicit
   

   Exit 0 = all checks PASS · 1 = any check FAIL · 2 = nothing to run (no --cmd and nothing detectable). Per-check timeout defaults to 300s (--timeout-ms overrides); a timed-out check counts as FAIL and the run continues.
3. Read the result. The summary lists each check (PASS|FAIL <name> <dur>, with a → re-run: <command> hint on failure) and a gate line. The full machine record is written to .claude/eval/continuous/<timestamp>.json (and latest.json) — metrics only, no output.
4. Act on it. On FAIL, re-run the named command yourself to see the captured output (it was streamed live, never stored), fix, and re-run /health.

Security & trust

• Detected / --cmd commands are author-authored shell — same trust boundary as the repo's own npm test / build scripts and the eval fences in checkpoint.js. Don't point /health at a repo whose package.json scripts / Makefile you don't trust.
• The artifact never contains output, env, or secrets — only the commands you ran and pass/fail.

Notes

• Self-contained: pure Node + shell, reuses _lib.js; no companions (mgrep/graph/context7), no network, zero new deps.
• Distinct from the pipeline /harness-claude:test skill: it enforces coverage during Verify; /health is an on-demand pass/fail pulse you can run against any repo.

Exit criterion

A continuous-eval run has produced a pass/fail summary (and .claude/eval/continuous/latest.json), and you've acted on any FAIL — repo confirmed green, or the failing checks identified for fixing.