Code Deleter

Every run must end with fewer or the same code lines as it started with. Net growth is forbidden. A verified no-op is a successful result. Preserve observable behavior; a smaller broken program fails this skill.

The budget

Measure against the worktree at invocation, including staged, unstaged, and untracked code, never against HEAD or a merge base. Preserve the user's existing edits. The budget covers every code change this run leaves behind, including tests, helpers, scripts, and code outside the initially targeted module.

• Require both production code alone and all code including tests to have a delta <= 0. Necessary new tests are allowed only within actual code savings.
• Count physical and nonblank lines after the project's normal formatting. Neither measure may grow. Review the diff for actual code reduction too: deleting comments, blank lines, documentation, fixtures, or generated/vendor content cannot fund new implementation. Do not remove tests or weaken checks to make the budget pass.
• No minification, semicolon packing, unreadable expressions, moving logic into strings/config, new dependencies to hide implementation, or counting moves and renames as deletions. Keep the accounting scope and test classification fixed throughout the run. No unrelated cleanup to manufacture deletion credit.
• If verification or the budget fails, shrink the proposed change or undo only this run's edits. Never leave net-positive code as a partial result. If a safe refactor needs more lines, report it as a separate recommendation.

Workflow

1. Establish the baseline before editing. Read local instructions and the relevant entry points, callers, tests, and contracts. Record the starting Git status; save original contents for files you may edit so rollback preserves dirty work. Keep scratch probes and the budget snapshot outside the project. Use the shipped gate below. If a complete baseline cannot be measured, make no code changes and report the limitation.
2. Find candidates, then prove them. Search for repeated logic, unused private code, pass-through layers, duplicated validation, and abstractions with no current purpose. Use references/safe-deletions.md for the evidence required. Search absence, type annotations, and green tests alone do not establish safety. For each edit, state what becomes unnecessary and which enforced invariant or caller evidence proves it.
3. Make the smallest justified simplification. Prefer an existing primitive or direct flow to a new abstraction. Preserve results, exception types and messages, ordering, mutation, side effects, cleanup, public interfaces, and supported compatibility. Do not silently fix bugs or change requirements. In review-only requests, report candidates and leave the code unchanged.
4. Verify behavior and size. Establish relevant behavior before editing; reuse existing tests and add focused characterization cases only where needed and affordable. Exercise the removed branch's boundary and failure paths. Run the applicable tests, type/build checks, and formatter, then inspect the final diff and rerun the gate. Missing tools or unresolved baseline failures are limitations, never a passing verification. Undo edits whose safety cannot be established. Recheck after any further edit, including test changes.
5. Deliver the measured result. Give before/after/net code and production counts, the safety evidence for each retained simplification, checks run and limitations, and valuable candidates skipped. With no safe reduction, report zero change. Do not create a report file or commit unless requested.

Mechanical gate

Needs Git and Python 3.11+. Let SKILL be this skill's installed directory and BASELINE a fresh path in a temporary directory outside the project. Commands run from the project's Git root. Adapt variable syntax to the current shell.

Inventory the project's authored code first. Repeat --include for every source type/path (including templates and executable scripts) and --tests for every test convention. These are case-sensitive, repository-relative globs; *.py includes nested files because * also matches /. The following is a Python example, not a default scope for other languages:

python "$SKILL/scripts/line_budget.py" snapshot --root . --baseline "$BASELINE" --include '*.py' --tests 'tests/*' --tests 'test_*.py' --tests '*/test_*.py'
# Perform the justified refactor, tests, and normal formatting.
python "$SKILL/scripts/line_budget.py" check --root . --baseline "$BASELINE"

The snapshot records current tracked and nonignored untracked source. The check includes newly created matching files and remembers original paths even if they become ignored. Exit 0 means all four budgets pass, 1 means growth, 2 means measurement failed; only 0 permits retaining a verified refactor. Never overwrite/recreate the starting snapshot, narrow its patterns, reclassify production as tests, change ignore rules, edit excluded code, or introduce an unmeasured source type. The gate refuses directory boundaries that hide nested Git repositories or submodules. Work wholly inside a measurable worktree or leave code unchanged; do not exclude a boundary to force a passing measurement. An aggregate gate cannot prove behavioral safety or detect comment-only credit; the contract and diff review remain mandatory.