Reduce execution and maintenance cost while preserving meaningful failure detection. Investigate aggressively; delete only with evidence. Retaining an area after review is a valid result. This skill is self-contained; upstream attribution is in NOTICE.md.
Choose the scope
- Focused cleanup: follow this entry point and read the candidate evidence guide before editing.
- Multiple owners, agents, or PRs: also read CAMPAIGN.md. Declare whether the work is a targeted sweep or an exhaustive subsystem audit.
- Validation and cost claims: read references/validation.md.
- Paperclip commands, invariants, and delivery: read references/paperclip.md alongside repository instructions.
Decision gate
For every proposed removal, answer:
- Which observable contract and plausible regression do the assertions cover?
- Which retained file/case detects that regression through the responsible production code? Otherwise, why is the assertion vacuous or contract retired?
- Does this candidate uniquely cover a transport, permission, lifecycle, storage boundary, platform, default, or other independent failure mode?
- What improves: executions, fixtures, processes, maintenance, or runtime?
Prefer the smallest test that reaches the code responsible for the contract. A broader suite is not automatically stronger proof. Similarity, size, slowness, or suspected AI authorship alone never justify deletion.
Prevent new duplication
Apply this gate to tests added or changed during gardening:
- Name the regression that existing proof misses. Extend an appropriate owner or table before adding another suite; another layer needs a distinct risk.
- Exercise the responsible production behavior. Do not add a mock that supplies the result being asserted or a production hook solely for convenient testing.
- Check observable outcomes that survive behavior-preserving refactors. A negative case must reach its named guard with otherwise valid inputs.
Focused workflow
- Inspect: pin the base; read candidates, keepers, fixtures, production callers, history, and CI routing. Preserve unrelated work.
- Plan: mark retain, repair, consolidate, delete, or defer. Name remaining proof and assertions to move; look for redundant suites as well as cases.
- Prove: establish the affected baseline locally or from matching-revision CI. Diagnose failures before treating them as cleanup candidates.
- Edit: move unique assertions first, then retire duplicates and unused support. Production-seam deletion needs its own caller/contract evidence.
- Validate: run keepers and affected siblings, verify selection, and apply the validation guide's intended-failure controls where needed.
- Review: reconcile deleted contracts with retained proof. Publish or merge only within the task's authorization.
Completion and reporting
Report removals and keepers, retained/deferred candidates, checks and limitations, distinct cases versus repeated executions, and test/support versus production LOC. Count documentation separately. Label cost evidence as measured, estimated, or historical; a proven simplification does not require a measurable speedup.
State the actual endpoint: findings, validated local patch, published PR, intermediate merge, or landed in the intended target. Stop at the requested scope/time limit or completed delivery, not a deletion quota. For another pass, refresh the base and prior findings; revisit retained tests only with new evidence.