crank-test-prune

Test Prune

Goal

A leaner suite judged against the rewrite test: tests exist to catch behavior regressions, not to pin the shape of the code. Prune aggressively; fewer, sharper tests beat a thicker suite that resists refactoring.

Hard Rules

• Every test in scope gets exactly one verdict — KEEP, DELETE, REFACTOR, or MERGE. No sampling.
• Expected values come from an oracle — a hand-verified literal from a real fixture or committed snapshot — never recomputed with the same formula the code under test uses.
• Redundancy is judged suite-wide, not per file — two tests pinning the same behavior at the same seam are one KEEP and one DELETE, even when they live in different files. Which one keeps: the test at the truer seam; at equal seam fidelity, the cheaper, faster one — a slow end-to-end test duplicating behavior already pinned lower survives only as one of a handful of happy-path journeys.
• Tests only. This skill edits and deletes test code; production code is never touched. A pruning pass that surfaces a production bug reports it, it doesn't fix it.
• Baselines come from a throwaway worktree, never git stash. Any clean-tree comparison (formatter, lint, the suite at HEAD) runs in a disposable git worktree add checkout of HEAD, removed after. git stash stays banned for the whole run, coordinator and subagents alike: a later drop destroys the uncommitted user work it swept up.
• Green gate. The full suite runs after applying; the pass isn't done until it's green.

Verdicts

• KEEP — passes the rewrite test: pins a distinct observable behavior at a seam.
• DELETE — an implementation-detail test, a tautological assertion (expected value recomputed with the code's own formula, so it passes by construction), contract-shaped (asserts structure, types, or wiring rather than behavior — including re-checking what the type system or schema validation already guarantees), copy-pinning (asserts that incidental prose appears — a description, label, warning, or log message; a kept test pins the behavior or a stable structured contract, not the wording), or a redundant test beside a kept one.
• REFACTOR — behavior worth pinning, assertion too weak to pin it. The common weak shape: scanning all values for a property instead of asserting one known concrete case — replace "no value in the map is undefined" with "this specific known key resolves to this hand-verified literal."
• MERGE — behavior worth pinning, test not worth its own setup. The common shapes: one-assertion fragments of a single workflow, or an isolated test pinning an incidental transition state a broader workflow passes through anyway. Fold the assertions into the journey test that walks the workflow end-to-end and delete the shell; the verdict row names the destination test.

Flow

1. Scope

From the argument, settle the test tree under review (default: the whole suite) and capture the test-file inventory plus the suite's run command. Done when: the file list and run command are pinned and stated.

2. Verdict

Fan out standard subagents — one per directory or module on a large suite — each returning one verdict row per test: file, test name, verdict, one-line reason (a MERGE row's reason names its destination test). Merge the rows, then reconcile redundancy and consolidation suite-wide yourself: each subagent sees only its slice, so cross-slice duplicates — and workflow fragments whose destination test lives in another slice — are yours to catch. Done when: every test in the inventory has exactly one verdict.

3. Apply

Apply the DELETEs, REFACTORs, and MERGEs. Every apply agent's instruction carries three standing constraints: touch only the test files in your assigned verdict rows; never move, rename, or delete repository configuration (formatter configs, hooks, manifests) to work around tooling — report the interference and stop instead; and siblings are editing the same tree concurrently, so unexplained git status entries are expected — leave them alone, never investigate or revert them. Done when: every DELETE, REFACTOR, and MERGE row is applied.

4. Verify

Run the full suite and repair any breakage the pruning itself caused. Done when: the suite is green.

5. Report

Render the verdict table (every test in scope), counts per verdict, the net line delta, and any production bug the pass surfaced. Done when: the table is rendered.

References

Subagents

This skill spawns verdict subagents at the standard tier — resolve it per SUBAGENT-TIERS.md: a loaded instruction file's preference first, its harness models only as the fallback.

Vocabulary

Read VOCABULARY.md before the verdicts in step 2: this skill leans on the seam, the oracle, the implementation-detail test, the rewrite test, the journey test, and the redundant test.