The SLOBAC Audit - So AI wrote your code a ton of tests. Are they good tests? This can answer!

Showcase Built for Cursor
, ,
1

SLOBAC - The Suite Life of Bobs and Code

What would you say your tests do here?
What would you say your tests do here?1000×542 121 KB

AI is great at writing tests. It is also great at writing tests that look great and don’t protect anything. It’s also great at writing the same test 10 times. Do you know what’s in your test suite?

If you’ve got a suite that’s six to sixty months into “still passes, but…” - sprint labels in the names, assertions pinned to copy that shipped once, mocks that re-encode the implementation, two files testing the same behavior under different titles - that’s the niche. It can happen overnight while a beefy agent chugs through your ask, or over years as traditional development lets “good enough” into the codebase over and over and over again…

SLOBAC is a manifesto about what tests should be, a taxonomy of the named ways they go wrong, and a Cursor Agent Skill (/slobac-audit) that walks an existing suite and emits a portable, read-only audit report. Semantic judgment, not syntactic smell counts.

What You Get

  1. A manifesto you can read without installing anything — principles, workflows, glossary, and a polyglot taxonomy (docs, llms.txt if you want to @-docs it).
  2. An Agent Skill that audits against that manifesto and writes a report you can hand-execute, hand off to another agent, paste into a PR, or ignore.
  3. Scoping - one file, one tree, or the whole suite; one smell or all. Maybe just the tests you touched in this PR?

Smells it knows about include (among others): tautology-theatre, deliverable-fossils, naming-lies, semantic-redundancy, over-specified-mock, pseudo-tested, shared-state, wrong-level. Some of those might sound familiar, but some I made up! Full catalog: taxonomy.

What it is not

  • Not a linter (it defers where pattern-match tools already win)
  • Not a test generator, mutation engine, or codemod runner
  • Not a smell-count scoreboard - (optimizing raw smell counts is an explicit anti-goal (EMSE 2023))
  • Not Cursor-only - it’s AgentSkills.io-shaped and also runs on Claude Code / anything that can load the plugin.

Install in Cursor

  1. Cursor Settings → Plugins
  2. Paste the marketplace URL: https://github.com/Texarkanine/txrk9-agent-plugins
  3. Install the slobac plugin from that marketplace

That gives you /slobac-audit.

Details: Running the SLOBAC audit

Run it

Use a frontier / high-effort model and the largest context window you can (MAX mode if you’ve got it). Lighter models under-find and invent soft diagnoses - I saw this clearly when comparing Auto / Composer against Opus-class / GPT-5.5 / Grok-class runs on the same suites. The new Cursor Grok works swimmingly with it, by the way!

/slobac-audit src/__tests__/** all smells - 1M context window

The audit is read-only with respect to your code. It reports; it does not rewrite your tests. It claims a .slobac/ working directory where it writes intermediate artifacts and its final report. You decide what to apply, and how.

Honest Caveats

  • Big suites can take a while (on the order of ~30 minutes in my dogfood runs) because the skill shards work to subagents and tries not to compact away the suite.
  • One pass is often incomplete on a large codebase - run again until it stabilizes or you’re happy. That’s expected, not a failure mode.
  • It works at ~200K context, but you’ll trade recall on cross-suite smells (semantic-redundancy, wrong-level, …) for safety as it shards harder.

Links

If you run it on a real suite, I’d love to hear where the taxonomy misfires, which smells are noise, and which findings you’d actually merge. Especially false positives - those are the interesting ones.