sloppoke
Blazing-fast AI-slop firewall for your git workflow. Sub-10ms
verdict per patch. Learns from every correction you ship.
slop poke # working tree vs HEAD (default)
slop poke --staged # the staged index
slop poke --range main..HEAD # everything diverged from main
slop poke --since main # shorthand for the same
slop poke --patch foo.patch # raw unified-diff file
slop poke --gh org/repo --range X..Y # any public github repo, no clone needed
slop poke --repo URL --range X..Y # any git URL (gitlab, bitbucket, self-hosted)
slop apply # apply the cached patch, amend HEAD
slop learn "false positive on src/foo.rs"
Output goes to stdout as a unified-diff patch — apply hint and
verdict line on stderr — so piping is trivial:
slop poke > slop.patch # save the patch
slop poke | git apply --unidiff-zero # apply directly
slop poke | delta # view in a pager
How do we characterize slop?
Slop is what writing-by-suggestion leaves in source. Three flavours,
easier to feel than to define:
- Wordy nothing — comments that restate the next line, names
that say less than nothing, prose that drifts.
- Defensive theatre — guards for impossible cases, empty
catches, retries with no upstream, asserts that can't fire.
- Unfinished work shipped — placeholder brackets, untested
branches, stub markers, AI-authorship trailers committed instead
of staying in metadata.
We don't publish the catalog, and it isn't static. Every slop learn "…" you submit calibrates the engine for your account; the
set firing on your repo on day 30 isn't the set firing today.
What it does
slop sweeps your diff through a blazing-fast machine-learning engine
that knows what AI-generated code looks like — and pulls it out before
you ship.
- Catches the artefacts an LLM leaves behind. Scaffolding
comments, placeholder identifiers, defensive crud, half-finished
markers, untested branches — the residue that survives "looks fine"
review but rots six months later.
- Adapts to your codebase. Every
slop learn you submit tunes the
engine for your account and your project. The catalogue you scan
against on day 30 isn't the catalogue you started with — it's the
one calibrated to your team's idioms.
- Fixes what's safe, flags what isn't. Mechanical noise is
stripped automatically; anything that could change behaviour gets a
TODO comment spliced into your diff for you to decide on. The CLI
refuses to mutate your tree unless
git apply agrees the change is
clean.
Detection runs server-side and updates continuously — no re-install
needed. Works on every language; the deepest analysis lights up first
for Rust, TypeScript/JavaScript, Python, and Go.
Install
Homebrew (macOS + Linux)
brew install peeramid-labs/tap/slop
Prebuilt binaries for macOS (arm64 + x86_64) and Linux (arm64 + x86_64).
APT repo still pending.
Build from source
git clone https://github.com/peeramid-labs/sloppoke.git
cd sloppoke
cargo install --path crates/sloppoke-cli
Needs rust 1.86+. One binary, no runtime daemons, no native deps
beyond ssh-keygen for request signing. cargo install places slop
in ~/.cargo/bin/ — make sure that's on your $PATH.
Claude Code plugin
Two-step install in any Claude Code session:
/plugin marketplace add peeramid-labs/sloppoke
/plugin install sloppoke@peeramid-labs
Ships /slop:poke, /slop:apply, /slop:learn slash commands plus
matching skills, so Claude Code calls the CLI on its own before
committing. Requires slop on PATH (Homebrew or source install above)
and slop login once.
Get started
slop login # SSH-key handshake; cache identity
slop poke # scan; 402 returns a Stripe Checkout URL
slop apply # auto-strip flagged lines, amend HEAD
slop learn "false positive" # shapes future scans
slop apply runs git apply + git commit --amend locally. Use
--no-commit if you want to inspect the staged diff first.
Learns as you go
