agentsclimarketplace

Verify

Skill Deadlymind/nanolama/skills/verify

Drives a change end to end and observes real behavior — exercising the actual flow through the API, UI, or logs, not just tests or typecheck — before it is called done. Use when about to claim a task is finished, working, or fixed, when the user asks to verify or confirm a change, or before opening a PR. Enforces a gate — no completion claim without a fresh proving-command run and its pasted output. Not for writing the tests themselves (see write-tests) or root-causing a failure (see debug).From its SKILL.md

Install
npx -y skills add Deadlymind/nanolama --skill verify

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 24 days oldThe repository was created 24 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.6 KB, 759 tokens by cl100k_base, as published. Nobody here has run it

Verify (prove it actually works)

When to use

Right before you say "done", "fixed", or "it works" — and before opening a PR. A change is not verified because it compiles, typechecks, or has tests; it is verified when you have watched the real behavior it was supposed to produce.

Pattern

The gate: no completion claim without a fresh run of a proving command and its actual output pasted in. Decide up front what observable behavior would prove the change, produce that observation on the real system, then report it. If you cannot run the proving command, say so plainly instead of implying success.

Steps / idioms

  1. Name the proving behavior. What should a user (or caller) now see that they couldn't before? Turn it into a concrete command or click-path.
  2. Run it fresh, end to end, against a running system — not a dry run:
    • API: hit the endpoint with a real auth cookie and inspect the status + body.

      # cookie-JWT flow: prove the endpoint returns the tenant's data, not a 500
      curl -sS -b cookies.txt http://localhost:8000/api/invoices/ | head -c 400
      
    • UI: drive the actual page/flow and watch the result (see run if available).

    • Async/websocket: trigger it and tail the worker/consumer logs for the effect.

    • Authorization: exercise the NEGATIVE case. A successful request from an authorized user proves nothing about whether an unauthorized one is actually blocked — send the same request as a user without the role and confirm 403.

  3. Read the evidence, don't assume it. A 200 with the wrong body is still a fail.
  4. Paste the observed output into your report, then state the claim.
  5. If a regression is possible, add a test that would have caught it (see write-tests).

The claim → requires → not-sufficient table

ClaimRequiresNot sufficient
"Endpoint works"fresh request + observed 2xx + correct bodytypecheck / unit test passing
"Bug fixed"reproduce the original failure, then show it gone"the code looks right"
"Tenant-scoped"user B gets 404/empty for user A's rowfilter present in the diff
"Authorized only"user without the role gets 403, one with it gets 2xxpermission_class present in the diff
"Deploys"health check green on the target envbuild succeeded locally

Adapt to your repo

Define your proving commands: the dev-server start command, the login/cookie step, the health-check URL, where worker and websocket logs live. Capture them once (a project verify note or Makefile target) so every change is proven the same way.

Gotchas

  • Green CI/typecheck is necessary, not sufficient — it proves shape, not behavior.
  • Mocked tests passing says nothing about the real integration; exercise the seam.
  • "Should work" is a hypothesis, not a result — never report it as done.
  • Verify on data that resembles production (multiple tenants, empty states, RTL text).

See also

  • write-tests
  • debug
  • code-review
  • security-review
<!-- Discipline adapted from obra/superpowers "verification-before-completion" (MIT). See THIRD_PARTY_NOTICES.md. -->

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 326,852. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.