jevgrep 0.2.0

package/package.json

@@ -0,0 +1,25 @@
+{
+  "name": "jevgrep",
+  "version": "0.2.0",
+  "description": "Semantic grep: describe the code you want in English, get file:line hits. Powered by TypeSafe Jev. No index, no embeddings, no dependencies.",
+  "keywords": ["grep", "semantic-search", "code-search", "jev", "typesafe", "lint", "cli"],
+  "license": "MIT",
+  "repository": { "type": "git", "url": "https://github.com/kyu1204/jgrep.git" },
+  "homepage": "https://github.com/kyu1204/jgrep",
+  "type": "module",
+  "bin": {
+    "jgrep": "dist/jgrep.js"
+  },
+  "files": ["dist", "skill", "README.md", "LICENSE"],
+  "engines": {
+    "node": ">=18"
+  },
+  "scripts": {
+    "build": "bun build src/cli.ts --target=node --outfile=dist/jgrep.js && chmod +x dist/jgrep.js",
+    "test": "bun test src/",
+    "prepublishOnly": "bun run test && bun run build"
+  },
+  "devDependencies": {
+    "@clack/prompts": "^1.8.1"
+  }
+}

package/skill/SKILL.md

@@ -0,0 +1,78 @@
+---
+name: jgrep
+description: >-
+  Semantic grep for code. Use `jgrep` when you need to locate code by what it
+  DOES rather than by an identifier you already know ("where do we retry
+  failed requests", "code that swallows errors", "anything that shells out"),
+  before opening many files to look for something, and as a cheap self-check
+  of your own diff before committing. Returns file:line ranges with a
+  probability, in about 2 seconds for a whole src/ tree. Keep using plain
+  grep/rg for exact names and strings.
+---
+
+# jgrep
+
+`jgrep "<description in English>" [paths]` asks a fast decision model (TypeSafe
+Jev) one yes/no question per 5-60 line chunk and prints the chunks that match.
+It never reads files into your context: you get a short list, then you Read
+only the ranges you need.
+
+## When to use it instead of grep or reading files
+
+- You know the behavior, not the name: "validates the webhook signature",
+  "builds the SQL string by concatenation", "catches and ignores errors".
+- You would otherwise open 5+ files hunting for something.
+- You want a second opinion on your own change: run it on the diff.
+
+Do not use it for exact identifiers, strings, or paths. `rg` is free and instant.
+
+## Commands
+
+```
+jgrep "description" src/                 # hits with p >= 0.7, file order
+jgrep -t 0.85 "description" src/         # fewer, higher-precision hits
+jgrep -C "description" src/              # print the matching chunk bodies
+jgrep --json "description" src/          # [{file,start,end,p,text}]
+jgrep -a -t 0 "description" src/ | head  # everything, best first (when 0 hits)
+jgrep --diff --staged "rule"             # lint your staged change
+jgrep --diff origin/main "rule"          # lint the branch against main
+```
+
+Exit status: 0 hits found, 1 none, 2 error. Every run prints a summary line
+on stderr: `N hits / M chunks · tokens · $cost · seconds`.
+
+## Reading results
+
+```
+src/loop/state.ts:108-115  p=0.96  export function readRun(...)
+```
+
+`p` is the probability the chunk matches. Treat >= 0.9 as reliable, 0.7-0.9
+as worth a look, below 0.5 as no. After a run, Read only the listed ranges
+(`offset`/`limit`), not whole files.
+
+## Writing the description
+
+- English, one concrete behavior per query. Split compound questions.
+- Describe the code, not the feature name: "decides whether to send an
+  alert based on OCR confidence" beats "alert feature".
+- Chunks are 5-60 lines seen in isolation, so cross-file flow ("does this
+  eventually write to the DB") will not match; ask about the local code.
+
+## Self-review before committing
+
+Run 2-4 rules on the staged diff; each rule is one sentence about a mistake
+you might have made. Exit 0 means a hit, so investigate those chunks.
+
+```
+jgrep --diff --staged "leaves debug output such as console.log or print"
+jgrep --diff --staged "changes behavior without a corresponding test change"
+jgrep --diff --staged "adds an endpoint or handler with no input validation"
+```
+
+## Requirements
+
+Installed globally as `jgrep`; the key lives in `~/.config/jgrep/env`. If it
+reports "No TypeSafe API key", tell the user rather than working around it.
+Run it inside a project directory or pass the project path: it refuses to
+walk a non-git directory with more than 5000 files.