Eval

/eval — acceptance-criteria checkpoint

Goal: prove the work satisfies the spec it was built against, criterion by criterion. Green tests show the code runs; this shows it meets what was promised.

Opt-in. This skill is not invoked by any default-pipeline skill or hook. Running it is always explicit; it changes no baseline behavior. (Fulfills v0.4 AC-6 / v0.3 AC-E5.)

Do this

1. Locate the spec. Default to the latest docs/specs/*.md; let the user name one if the change targets an older spec. Confirm it's the spec this change was built against.
2. Run the deterministic gate:

   node scripts/eval/checkpoint.js [spec.md]
   

   It prints a per-criterion table (PASS / FAIL / MANUAL) and exits 0 (all PASS, none manual) · 1 (a check FAILED) · 2 (no failures, but MANUAL criteria remain).
   • Each criterion may carry an inline ```eval fenced shell command — that's its machine-check. Add one to an AC when a deterministic check exists (port it from the spec's own test commands).
3. On FAIL (exit 1): the criterion is objectively unmet. Fix the implementation (not the criterion) and re-run. Do not proceed while any check FAILs.
4. Adjudicate MANUAL (exit 2): for each prose criterion the script can't decide, gather concrete evidence (run the app, read the diff, check .claude/traces/ for what fired) and judge PASS or FAIL yourself. The script never auto-passes prose — closing these is your job. If a MANUAL criterion is genuinely verifiable, prefer adding an ```eval fence so it's deterministic next time.
5. Gate. Ship only when every criterion is PASS — checks green and every MANUAL adjudicated PASS with evidence. Report the verdict criterion by criterion.

Security & trust

• ```eval fences are shell-executed** (via _lib.sh → execSync) with your privileges. They are author-authored, committed markdown — trust them exactly like the repo's own test scripts or build hooks. **Review specs from untrusted branches/PRs before running /eval.
• checkpoint.js --list is the safe inspect-before-run path: it parses and shows which ACs carry a check, but executes nothing.
• Secret-safe: a check's output is never printed. On FAIL the report shows only the exit code + the command (committed text), so you re-run it yourself to inspect output.
• Each check is bounded by a 60s timeout.

Notes

• Self-contained: the runner is pure Node and needs no companions (mgrep/graph/context7).
• Complements harness-claude:test (suite + coverage) and harness-claude:verify (observe real behavior); this one ties the change back to its acceptance criteria.

Exit criterion

Every acceptance criterion in the spec is PASS (machine-checked or evidence-adjudicated). Then harness-claude:ship.