The web app

slop-chop.com runs the same rules engine as the CLI, compiled to WebAssembly. One codebase, two front ends, byte-identical output. Nothing is uploaded and there is no server: the page is static files on GitHub Pages.

How the pieces fit

The Go engine in internal/sanitize compiles to slop-chop.wasm through the shim in wasm/main.go. The page never touches the engine directly: a Web Worker owns the WASM instance so a giant paste chops off the main thread and typing never freezes.

Piece	Job
wasm/main.go	Registers the engine calls on the worker's global scope.
docs/assets/worker.js	Boots the WASM, answers {id, fn, arg} messages from the page.
docs/assets/app.js	The whole UI: panes, marks, drawer, share links, connectors.
overrides/main.html	Social card metadata on every page.

The shim exposes four calls. slopChop takes text, a full profile, and preset names, and returns the output, the findings, and the score. slopDefaults and slopPresets feed the settings panel from the same source of truth the CLI uses. slopRewritePrompt and slopJudgePrompt return the model instructions from internal/rewrite/prompt, a package split off so the WASM build shares the CLI's exact prompts without pulling the HTTP client into the binary. Adding a shim export means adding it to the worker's allowlist too.

Profile semantics

The page mirrors the CLI. The settings panel builds one profile object: the built-in defaults merged under the user's entries, with the user winning on any key both set. Chosen presets merge on top through the same ApplyPresets the CLI uses, profile winning on conflicts. Copy profile JSON exports exactly what the page runs, and the file works verbatim with --profile.

The marks

Both panes sit on mirror layers. The input mirror places a mark under every finding, cut on byte offsets from the engine. The output mirror runs a Myers token diff between input and output and marks what changed, bailing to a plain mirror past 1,500 edits. The mirrors match the textarea's font metrics and compensate for scrollbar width, so marks line up with characters on every platform.

Files in and out

A text file dropped on the input pane loads and chops in place, and the Download button saves the output pane under the dropped file's name, so a file round-trips through the chopper without a copy and paste. Text with no file behind it downloads as chopped.txt. A file past two megabytes, or one that looks binary, is refused with a message and the panes keep what they had.

Model connectors

The rewrite pass is optional and browser-direct. Anthropic calls go straight from the page with the user's own key and the CORS opt-in header. Any OpenAI-compatible endpoint works for local models: Ollama, LM Studio, vLLM. Keys live in localStorage and are sent only to the chosen endpoint. The reply streams into the output pane as the model writes it, and a stream that dies puts the rules output back instead of leaving half a reply. After a rewrite, the same provider judges the result against the original with the CLI's verify prompt, the page reports whether meaning held, and a Restore button returns the pane to the rules output when the model's version loses.

Share links

Copy link packs the settings into the URL hash as base64 JSON. API keys are stripped before encoding. On load, a valid hash applies the settings and cleans itself from the URL. A mangled hash degrades to a normal visit.

Offline

A service worker caches the page, the engine, and the theme's hashed bundles on the first visit, so the chopper works with no network at all. Same-origin requests are served from the cache and refreshed in the background, which means a deploy reaches a returning visitor one load later. Calls to model providers never touch the worker. A small manifest makes the page installable as an app. This is the privacy pitch made literal: after the first load, nothing requires the network but the optional rewrite.

Verifying changes

make wasm builds the engine into docs/assets, mkdocs build assembles the site, and the suite in e2e/ drives the result in Chromium, Firefox, and WebKit with model providers mocked at the network layer. The e2e workflow runs it on every push and pull request that touches the site, the engine, or the shim. See the e2e readme for the local recipe.

Releasing

Push a v* tag on the commit you want stamped. The release workflow cross-builds the CLI, publishes the archives, and bumps the Homebrew formula. The docs workflow stamps the wasm with git describe, so tagging the deployed commit itself yields a clean version in the settings panel footer.