Pruning tests

Report first, edit second, prove every batch. A deleted test changes what the suite protects.

• The candidate list goes to the person before any test is edited — "clean it up" asks for the list, not permission to skip it.
• The pruner does not grade itself: the proof is numbers, or a separate reviewer.
• What makes a test low-value is decided in vvkit:writing-tests. This skill covers removing it without losing protection.

Audit

1. Baseline. Record every in-scope test's pass/fail before touching anything. A failing test is a bug report until shown otherwise — in one subsystem-wide prune, all three baseline failures were real delivery bugs. Fix them at the owner; never delete them.
2. Read and mark each test by its assertions, not its name — one named for clearing a window asserted it was not cleared. Read the production owner, callers and history (did a bug fix add it?).

3. Report and stop, as one table:

   Test

   Mark

   Bug it claims to catch

   Proof that remains

   Seam freed

   Marks: retain; fix (keep the contract, repair the assertion); consolidate (name the absorbing test); delete (name the remaining proof, or why no contract exists). No evidence, no delete.

What stays

• Race and hang proofs; the only test of a filesystem or database failure mode; regressions from real bugs; observable ordering; public API, protocol, config, migration, security and default-value contracts. Slow or static is no reason to delete.
• A weak test that is the only one reaching a path is rewritten, not deleted. Deleting one that slept and asserted loosely took its path from 65% coverage to 0%.

While pruning

• Audit doubles too. Replace a hand-written double with the shared one, keep it with a comment naming the control the shared one lacks, or merge copies. The third copy gets promoted.
• Duplicated production code behind a cluster of tests: extract it and test it once.
• Never weaken a test to improve the deletion count. One coherent batch per change.
• A whole subsystem splits by production owner: one read-only agent marks each area, then one pass across the marks looks for a redundant layer.

Prove the prune

Copy into the work for every batch that deletes, merges or rewrites tests:

- [ ] Mutant per removed claim: the bug each removed test claimed to catch, applied as a one-line
      mutant; its killer named before and after
- [ ] Coverage per function, before (worktree at the merge base) and after; report non-empty
- [ ] Every function dropped to 0%, or newly over CRAP 30, has a rewritten test or a stated reason
- [ ] Up to ~6 more mutants where coverage dropped or tests changed; all old kills still killed
- [ ] Independent review of deleted coverage against what remains (batches beyond a handful)
- [ ] Report: mutant table with killers before and after, coverage drops with reasons, seams removed

• The first item is never skipped. "The remaining tests already catch it" is an argument, not a run. Apply the mutant.
• Coverage before mutants. Mutants only reach paths some test still runs — twelve missed the 65%-to-0% drop that coverage caught.
• Each mutant is tied to the deletion: a killer before the prune is a test this batch removes or rewrites. One at a time; restore by reversing the edit, confirm by file hash.
• A survivor means restore or rewrite a test. Surviving on both sides is an old gap — add a test only if it clears the acceptance filter, watched failing under the mutant.
• Some runners return an empty coverage report, without warning, when coverage was not enabled.
• An independent reviewer finds what the pruner misses — on one prune, nine contracts that lost their only proof and one new assertion that could not fail.
• Tests written after seeing survivors: say so; the score only fits that sample.

CRAP = complexity² × (1 − coverage)³ + complexity. Where the project has tools for coverage, CRAP, duplication or mutation testing, use them over hand-run mutants. Coverage commands for a stack are in its testing skill — for Swift, vvkit-swift:swift-testing.

Adapted from sergdort's "Pragmatic Testing" agent skill and post (2026), and OpenClaw's test-audit skill, the source of the baseline and review figures.