agentsclimarketplace

Spec correctness

Skill agent-ix/quoin/skills/spec-correctness

Turn classified acceptance criteria into runnable property tests. Consumes `quire properties --json` per-criterion records, grounds each criterion's domain, precondition and oracle in the spec and the code, and emits property tests in the repo's own harness (proptest / fast-check / hypothesis) keyed on `row_id`. Emits a validated SpecReview recording every criterion it could not ground and why.From its SKILL.md

Install
npx -y skills add agent-ix/quoin --skill spec-correctness

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

One thing to look at

  • 20 stars20 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

4.8 KB, as published. Nobody here has run it

Spec Correctness

Use this skill to generate property tests from acceptance criteria you already wrote.

quire properties classifies every binding acceptance criterion by property shape and stops there — it names no test framework and its deterministic recall has a measured ceiling. This skill closes the three gaps that leaves:

  1. The harness. The shape→harness mapping is this skill's, not the engine's.
  2. The clauses. Domain/precondition/oracle spans reach only a small minority of criteria, so this skill grounds them from the spec and the code itself.
  3. The residue. Criteria the deterministic pass labelled example or unclassified go to an LLM second pass — review-gated, so it can afford recall the engine cannot.

Specified by FR-028.

The one rule that is not negotiable

A low extractable ratio is data, never a verdict. This skill reports counts. It emits no threshold, no grade, no pass/fail, and it never suggests rewording a criterion to score better. A criterion that describes one concrete scenario is a good criterion; it just isn't a property. (quire-rs FR-052-CON-1.)

Two more, inherited:

  • Never write a framework name into spec/**. (CON-2.)
  • Never propose a new regex or lexicon entry for quire-rs. 31.3% recall at 93.3% precision is a measured ceiling, not a bug. (quire-rs#45.)

When to use

  • After acceptance criteria are written and validated, to give them verification.
  • Before gap-analysis, so the matrix rows it reconciles have real backing tests.
  • On an existing repo, to find which criteria have no test and why.

Not for: authoring criteria (specify), building the matrix (spec-matrix), or judging whether existing tests are good (gap-analysis).

Inputs

  • The target repo and its spec glob (default spec/**/*.md).
  • quire properties --scope <repo> --json '<glob>' output — quire-cli ≥ 0.12.0.
  • The repo's source and test trees.

Steps

  1. Scope and harness: resolve the repo and glob, run quire properties --json, detect the harness, test dir, and the repo's existing tracking-tag style.
  2. Census: bucket records by property × extraction. Data only — no verdict, no threshold, no rewording suggestion.
  3. Clause grounding: derive domain, precondition and oracle for each record, each cited to a file:line. No citation, no test.
  4. Strategy selection: pick the assertion skeleton from property, and the routing lane from extraction.
  5. Generate tests: emit into the repo's harness with the row_id tracking tags and the provenance line.
  6. Second pass: the LLM pass over example, unclassified, and anything step 2 refused to ground.
  7. Review artifact: the SpecReview at reviews/YY-MM-DD-<slug>.md recording what could not be grounded, and why.
  8. Report and handoff: the run report, the Property rows for spec-matrix, and the reconciliation check against gap-analysis.

Steps 0–4 and 6–7 always run. Step 5 is an expensive LLM pass — ask before running it when the residue is large (say, more than 30 records).

Outputs

PathWhat
tests/props/<per-harness naming>Property tests, all of them runnable
reviews/YY-MM-DD-<slug>.mdA SpecReview (analysis: spec-correctness) recording what could not be grounded

Nothing else is written to the tree (FR-028-CON-1). No skipped tests, no _review/ directory, no ad-hoc report file. A generated test arrives in a pull request, and that is what puts it under review; the artifact carries the one thing a PR does not — why a criterion was left ungrounded.

Two facts that will bite if you assume otherwise

  • shape is not property. shape is the FR-047 grammar axis (assertion | obligation | given-when-then | unstructured). property is the 10-value taxonomy. Strategy selects on property. Routing selects on extraction.
  • StR scores highest, not lowest — 29.1% against FR's 19.2%. Do not down-weight stakeholder requirements (quire-rs CR-029 falsified the CR-020 prediction).

Keep looking

Skills are one crate of 611,919. 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.