Deslop

Deslop

Quick start

/deslop                  # full codebase
/deslop src/api          # directory
/deslop 1234             # PR number (gh)
/deslop feat/new-auth    # branch (diff vs default branch)

Workflow

1. Resolve scope — turn the argument into a concrete file list:

   • Empty → full codebase
   • All digits or #N → gh pr view N --json files → changed files
   • Branch name → git diff --name-only <default>...<branch>
   • Path → use as-is (dir or glob)

2. Compute excludes — build an ignore-list and filter the file list through it:

   • Every pattern in .gitignore and .git/info/exclude
   • Build output: dist/, build/, out/, .next/, target/, __pycache__/
   • Vendored / generated: node_modules/, vendor/, .venv/, *.generated.*, *.gen.go, *_pb2.py, __snapshots__/, *.snap
   • Lockfiles

3. Size check — if scope > 500 files (tunable), stop and prompt the user to narrow: by directory, by churn (git log --since='3 months' --name-only --pretty=format:), or by largest N files. Agents blow their context on huge scopes — don't push through.

4. Detect tooling — read package.json, tsconfig*.json, pyproject.toml, go.mod, Cargo.toml. Note installed tools (knip, ts-prune, ruff, vulture, staticcheck, clippy). For polyglot repos, bucket files by language; each language gets its own set of 7 agents.

5. Launch all 7 agents in parallel — a single message with 7 Agent calls, subagent_type: "general-purpose". Each agent's prompt is its section of AGENTS.md with {scope}, {excludes}, {tools} filled in, plus the shared output format. This pass is analysis-only — no file edits, so parallel agents can't conflict. Print [N/7 done: <category>] as each agent returns.

6. Consolidate findings — merge the reports so every finding from all 7 appears exactly once: deduplicate overlap (agents 1 and 2 often flag the same shared-type issue), group by file, sort by severity. Quarantine findings below 0.7 confidence in a "needs review" section — never auto-applied. Write the full report to deslop-report.md at the repo root.

7. Review with user — summarise from the report: count per category, a few top examples each, total files touched. Point at deslop-report.md for the full list. Ask which categories or files to apply.

8. Apply changes (branch + worktree fan-out) — when the user approves:

   1. Confirm clean working tree. Create branch deslop/<yyyy-mm-dd> off the current branch. Refuse to edit on main/master.
   2. For each approved category, spawn an Agent with isolation: "worktree" and that category's findings. Each worktree edits independently — no conflicts.
   3. Merge each worktree back to the cleanup branch in precedence order: deletion (3, 6) → simplification (5, 7) → consolidation (1, 2) → typing (4). Run build + typecheck + tests between each merge; stop and surface on regressions.
   4. Done when the deslop/<date> branch holds one commit per applied category and deslop-report.md records every finding as applied, deferred, or quarantined.

The 7 agents

1. Deduplication
2. Type consolidation
3. Unused code
4. Weak types
5. Defensive programming
6. Legacy / deprecated code
7. AI slop — comments, over-nesting, style drift

Per-agent prompts: AGENTS.md.

Guardrails

• Behavior unchanged. Cleanups are refactors, not rewrites — no semantic changes smuggled in unless fixing a clear bug.
• PR scope stays in scope. Don't propose changes outside the diff unless the user opts in.
• Bias to deletion over rewriting (agents 3, 5, 6, 7). Surface anything ambiguous.