trash-truck

Finds and removes code slop without changing behavior. Opens small, reviewable PRs. Inspired by the real cost of AI-generated entropy: teams that skip cleanup spend 20% of their week on slop.

Invocation

/trash-truck [org/repo]
/trash-truck [org/repo] focus:dead-code
/trash-truck [org/repo] focus:duplicates
/trash-truck [org/repo] focus:drift
/trash-truck [org/repo] focus:misplaced
/trash-truck [org/repo] focus:committed-by-error
/trash-truck [org/repo] focus:unused-functions

Targets

Category	What to look for
dead-code	Unused exports, orphaned imports, debug console.log/print, commented-out code blocks
duplicates	Same concept implemented multiple ways across the codebase
inconsistent	Same operation done differently in different modules (error handling, fetch patterns, config access)
drift	Deviations from golden principles documented in CLAUDE.md, docs/, or established skill files
misplaced	Files in wrong directories — test files outside test dirs, configs buried in source trees
committed-by-error	Files that shouldn't be in version control — .env, .DS_Store, swap files, compiled bytecode, IDE configs, database files
unused-functions	Functions/methods defined but never called or referenced anywhere in the codebase (Python via ast, JS/TS via export/import analysis)

How It Works

1. Clone or enter the target repo
2. Run pre-scan for cheap static analysis before spending tokens on LLM review
3. Review pre-scan results — filter false positives, group into cleanup units
4. Open each PR with a focused, reviewable diff

Pre-scan (Step 2)

The pre-scan.sh script in this skill's directory uses rg + Python ast to find candidates cheaply before Claude reviews them. This cuts token cost by 5-10x compared to having Claude grep the entire codebase.

# Run from the repo root — outputs structured JSON
SKILL_DIR="$(dirname "$(readlink -f "$0")")"  # or wherever the skill was synced
bash "$SKILL_DIR/pre-scan.sh" [focus] [repo-root]

# Examples:
bash pre-scan.sh all /path/to/repo          # everything except unused-functions
bash pre-scan.sh dead-code /path/to/repo     # just debug stmts + commented code
bash pre-scan.sh unused-functions /path/to/repo  # deeper analysis, slower

The script outputs JSON with findings grouped by category. Each finding has file, kind, and usually line and detail. Use this output to guide which files to read and what to clean up — don't re-scan what the script already found.

What pre-scan does NOT do: It finds candidates, not confirmed slop. Claude still needs to verify each finding (is the "unused" function actually called via reflection? is the "debug" print actually a user-facing log?). The script is intentionally aggressive — false positives are filtered by Claude, not suppressed at scan time.

PR Rules

• One PR per logical cleanup — keep diffs small and reviewable
• Title format: chore: [what was cleaned] in [file/module]
• No behavior changes — if a change could alter runtime behavior, skip it and open an issue instead
• CTO reviews process (was the right pattern followed?), not the code itself

What NOT to Do

• Do not refactor logic — only remove or consolidate dead/duplicate patterns
• Do not open a single giant PR — small diffs or nothing
• Do not touch tests unless the test itself is dead or duplicated