deslop

A deletion-first Agent Skill for agent-maintained codebases

简体中文 · English

Evidence-backed cleanup that reduces accumulated machinery while preserving real behavior.

deslop audits and, when explicitly authorized, removes complexity accumulated through repeated coding-agent implementation and correction cycles. Those cycles often leave overlapping regression tests, producer-verifies-producer checks, and fallback layers that hide failures instead of handling a current contract.

This is semantic subtraction, not source beautification. deslop is not a formatter, style humanizer, test-count minimizer, blanket ban on defensive code, or automatic permission to edit a repository. It follows justification chains to independent evidence and preserves behavior whose contract remains real or uncertain.

Reduce test surface, not behavior surface.

What it targets

The percentages below are design priorities, not measured prevalence.

Generic dead code, wrappers, abstractions, and comments are secondary. They matter only when they belong to one of these clusters or have direct high-confidence deletion evidence.

Subtract machinery. Preserve behavior.

Resemblance to a smell is a lead, not a verdict. Security and trust boundaries, supported callers, persisted formats, and numerical constraints are preserved by default when evidence is incomplete.

Quick Start

Codex: install v0.3.2 as a standalone Skill

Invoke the bundled installer with this GitHub Skill URL:

$skill-installer
Install the Skill from:
https://github.com/MrZoyo/deslop-GPT/tree/v0.3.2/skills/deslop

For a reviewable local checkout, symlink the runtime directory into Codex's canonical user Skill location:

git clone --branch v0.3.2 --depth 1 https://github.com/MrZoyo/deslop-GPT.git "$HOME/.local/share/deslop-GPT"
mkdir -p "$HOME/.agents/skills"
ln -s "$HOME/.local/share/deslop-GPT/skills/deslop" "$HOME/.agents/skills/deslop"

Codex supports symlinked Skill directories and detects changes automatically. The tagged v0.3.2 path is the current released, pinned standalone Skill; main is the development branch and may contain unreleased changes.

Claude Code: install the Plugin from GitHub

Inside Claude Code, add this repository as a marketplace and install the Plugin:

/plugin marketplace add MrZoyo/deslop-GPT
/plugin install deslop@deslop

The canonical Plugin command is /deslop:deslop. For a local checkout, load the repository directly with claude --plugin-dir . from the repository root. The marketplace catalog is read from main, but its Plugin source uses an explicit HTTPS Git URL pinned to the v0.3.2 tag and release commit 0cc15c036b07691c600bda1219b8cc5c197ca3f1. This patch release makes missing-evidence limits explicit in the closed-loop rule. The v0.3.1 evaluation evidence remains tied to its exact released payload.

One checkout, standalone discovery on both hosts

The same released runtime payload can be linked into each host's user Skill directory:

mkdir -p "$HOME/.agents/skills" "$HOME/.claude/skills"
ln -s "$HOME/.local/share/deslop-GPT/skills/deslop" "$HOME/.agents/skills/deslop"
ln -s "$HOME/.local/share/deslop-GPT/skills/deslop" "$HOME/.claude/skills/deslop"

Use only the link for the host you need, and run each ln command only when its destination does not already exist. A standalone Claude Code installation invokes the Skill as /deslop. See Getting Started for installation scope, v0.1.0 migration, updates, removal, and a safer review-first workflow. deslop is an independent community project, not an OpenAI or Anthropic product.

Distribution status

The shared skills/deslop/ payload follows the open Agent Skills structure and is used unchanged by Codex and Claude Code. .claude-plugin/plugin.json and .claude-plugin/marketplace.json provide Claude Code packaging. Codex Plugin distribution remains withheld because the tested Codex host installed and cached a Skills-only Plugin without registering its bundled Skill; Codex standalone installation remains supported. See the distribution compatibility note.

Invoke it explicitly

Append the same mode and scope arguments to the command name for each host:

Only apply authorizes edits. Staging, commits, pushes, branch changes, resets, and fetching still require separate permission.

Example workflow

Start with evidence, not edits:

$deslop deep

HIGH
- two fallback layers handle the same internal parse failure;
  current callers and history show no supported legacy input
- a local receipt is produced and verified by the same workflow;
  no external consumer or persisted trust boundary exists

PRESERVE
- a persisted readback detects truncated output across a write/read boundary
- a compatibility branch is required by a documented external protocol

Review each evidence chain and preservation decision. Apply only the supported scope:

$deslop deep apply

The example is schematic; it does not represent a benchmark fixture or performance claim.

How deslop decides

• Independent evidence roots: current requirements, real callers, public contracts, protocols, trust boundaries, persistence boundaries, or scientific invariants.
• Closed justification loops: production code and tests do not become necessary merely by justifying each other.
• Production reachability and edge closure: prove the current input-to-consumer path, not only isolated callers or test-injected branches.
• Production/test asymmetry: redundant test evidence can be removed without deleting the behavior it observes.
• Fail-visible bias: unexpected internal failures should surface unless a concrete recovery or translation contract exists.
• Subtraction without redesign: dependencies, abstractions, wrappers, compatibility layers, and replacement scaffolding have a default budget of zero.

The full decision model is documented in Design. The self-contained runtime SKILL.md remains authoritative for agent behavior.

Safety model

Codex enforces explicit invocation through allow_implicit_invocation: false. Claude Code does not read that OpenAI-specific metadata; the shared standards-compatible frontmatter instead tells Claude to invoke deslop explicitly. Claude Code may still select the Skill from its description — a 2026-09-03 control run recorded it doing exactly that — but such an invocation remains read-only unless the user includes apply. Default and audit modes are read-only, and suspicious constructs can be recorded as deliberate preservation decisions. Code is not removable merely because it looks defensive, was written by an agent, or has a test that could be deleted.

Read-only verification should redirect caches or generated output when practical and disclose incidental residue. Apply authorization permits scoped edits; it does not resolve uncertainty in favor of deletion. See Getting Started for the review sequence and Design for confidence classes and preserved boundaries.

Evidence

Validation status

Version-bound forward smokes are published under evals/release-smoke/. The 2026-09-04 cross-version smoke is the v0.3.2 release gate; it uses known cases and two runs per payload, so it supports only the narrow release decision recorded above. The earlier Claude Code CLI runs remain under evals/runtime-controls/results/ and evals/dev-v2-focused/results/. Those Haiku runs exposed the t02b ambiguity but remain secondary host diagnostics, not target-model release evidence. None of these records is held-out model-effect evidence. The older rc3 micro pilot measured 63.1% more total tokens and 16.5% more wall time with its then-current Skill; that one-run result does not predict v0.3.x cost, but it supports using deslop for deliberate accumulated-slop work rather than routine tiny diffs.

Focused development evaluation

dev-v2-focused tests preservation and simplification decisions across paired micro cases and three end-to-end miniature repositories. Behavior gates run before reduction metrics. Micro and mini-repository results remain separate, and the repository publishes no project-level performance score.

The follow-up dev-v3-evidence-edges draft records 19 anonymized field observations and implements 7 new executable pairs. It is validated as a draft, not reported as model-performance evidence.

See Evaluation for interpretation limits and evals/README.md for the canonical protocol.

Real-world field trials

The first field trial records both accepted cleanups and deliberate preservation decisions with public before/after provenance. It had no independent baseline run from the same frozen state, so it is not a controlled A/B comparison and does not establish general superiority, 100% precision, or production-proven correctness.

Future cases can be added without becoming Skill-tuning inputs; see Field Trials.

Documentation

Repository structure

.claude-plugin/                 Claude Code Plugin and marketplace metadata
skills/deslop/                   Self-contained runtime Skill payload
docs/                            User, design, evidence, and development guides
evals/dev-v2-focused/            Active focused development evaluation
evals/dev-v3-evidence-edges/     Follow-up evidence-edge draft
evals/runtime-controls/          Authorization and host/runtime controls
evals/release-smoke/             Version-bound forward-smoke records
evals/real-world/                Manually adjudicated real-world evidence
evals/archive/                   Retired historical evaluation material
scripts/                         Validation and evaluation tooling
assets/                          README and project presentation assets

Project status and contributing

Public releases use semantic versioning, beginning with v0.1.0. A 0.x release is usable but still evolving; it is not a stable, production-ready, or 1.0-quality claim. Immutable Git tags identify released runtime and distribution states. Benchmark candidates retain their separate evaluation tags.

The most useful contribution is an evidence-backed case with a nearby preservation counterexample and an independent behavioral oracle—not an isolated snippet that merely looks verbose. Read Development before proposing a Skill policy or evaluation change.

License

MIT