How is fuzzy linting different from ESLint or grep?

Classic linters match patterns; patdown asks whether the file violates the written rule and returns yes or no with probability. One heading is one rule with globs and thresholds, and a failing file costs one more call to locate evidence.

Will this burn API budget on large repos?

Yes if you lint wide trees with packs. Files are read once per run but judged per matching rule, nothing is cached across runs, and max-judgments caps the run before it spends too much.

Can I split rules across files or reuse packs?

Yes. Frontmatter include loads other paths, directories load every markdown file except README, and packs for TypeScript, Effect, and anti-slop ship as packages to include by path.

How do I stop secrets from being sent to the judge?

Built-in skips cover git, dist, node_modules, env files, keys, and credential names, and piped input warns against secrets. Stdin reads UTF-8 text and cannot combine with input-text; ask mode never fails on yes.

How do I install and run it, and what does it need?

npx patdown --help or pnpm add -D patdown, with Node.js at least 22.22.2 and pnpm 11.8.0. For GitHub Actions there is a composite action, and the library entry is patdown while rule adapters import @patdown/rules.

Is the judge swappable, or locked to TypeSafe?

Swappable. The default backend uses Effect Decision/DecisionModel with TypeSafe, while rules and CLI commands use a provider-neutral interface, so you can replace the rule source with a module exporting loadPatdownRules or an Effect Layer called PatdownRuleSourceLive.

What is the ask command for, and can I pipe input to it?

ask answers a yes/no question with no files involved, printing yes or no, and --verbose adds a P(yes) shade bar, the cutoff and the judge call time. --stdin reads piped UTF-8 text through EOF and cannot be combined with --input-text; do not pipe secrets, because the content goes to the configured judge.

CommunityJev Agents & MCP

Patdown

Node CLI that lints a file tree against fuzzy natural-language rules in one Markdown file, judging each matched file with a swappable TypeSafe-backed judge.

See highlights

Open github.com

Stars: 15
Language: TypeScript
Published by: Community
Author: tyler-dot-earth
Last verified: Oct 2026

tyler-dot-earth/patdown

User repository on GitHub

View on GitHub

Block, steer, and "fuzzy lint" with Jev to make agents follow your rules and conventions. CLI, github action, pi package, claude extension, and more. Built with Effect + TypeScript.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Standalone CLI that lints a tree against fuzzy rules in one markdown file. Wrap it as a hook, plugin, or extension. The judge is swappable. The default backend uses Effect Decision / DecisionModel with TypeSafe; rules and CLI commands use a provider-neutral interface. Requires Node.js >=22.22.2. The library entry is patdown; rule adapters import @patdown/rules. Optional agent packages: @patdown/pi (Pi write steering) and @patdown/claude (Claude Code Write/Edit hooks). See Pi write steering and Claude Code plugin.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Moderation Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

One # heading per rule with stacked globs, per-rule yes-threshold, and annotation lines.
Frontmatter include composes packs and project rules; duplicate titles and cycles fail the run.
ask and doctor commands answer single questions and check keys without linting the tree.
Pi write steering and Claude Code hooks reuse the same rules to block or warn before disk writes.

Watch out

Each matched file is judged once per matching rule, so packs multiply call count with no result cache across runs. Credential files and build dirs are always skipped. Failures, missing rules, and budget overruns exit 1. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

stepwarden

A Claude Code plugin that verifies every tool call before it runs, asking Jev five parallel questions about the plan, scope, prior action and stated intent, then allowing, asking or blocking. Fresh installs ship in audit mode.