JevX for GitHub 0.3.1:

Fixes "no Jev needed" on big repos. Before, only ~6 tiny files reached the AI.

• Scans up to 400 source files (Standard) or 2,500 (Whole repo) on your computer, for free
• Only files with rules or AI calls go to your AI, in parts, two at a time
• Files named like decisions (classify, filter, moderation, route, prompt…) are read first; tiny barrel files are skipped
• New Standard / Whole repo picker
• When only part of a repo was scanned, it now says "No Jev fits found yet" and offers "Scan the whole repo"

Install: download jevx-for-github-0.3.1-store.zip, unzip it, open chrome://extensions, then Load unpacked. Chrome Web Store version coming soon.
CLI unchanged: npx @vij-sameerb5/jevx (0.4.1).

JevX 0.4.1

Install: npm i -g @vij-sameerb5/jevx (the command is jevx) · Website: https://jevx.live

Downloads

• Claude Desktop: jevx-0.4.1.mcpb. Double-click it, pick your project folder, and optionally add your TypeSafe key.
• Chrome extension (preview): jevx-for-github-0.3.0.zip. Unzip it, open chrome://extensions, turn on Developer mode, click Load unpacked and pick the folder. Then add your Gemini (free tier), xAI or OpenRouter key in its settings.

What's new in 0.4.1

• Safer by default: if a project has no passing tests or typecheck, JevX previews but writes nothing. Use --allow-untested to opt out.
• Fewer failed edits: small whitespace and indentation differences no longer stop a change.
• --only <paths>: scan just part of a repo, e.g. jevx --only src/api.
• Version check: jevx --version and the MCP server now always match the npm version.

New: JevX for GitHub (Chrome extension preview)

• Open any GitHub repo and see exactly which lines should use Jev: file, line numbers, the current code, and the question Jev would ask.
• Every spot gets a fit score from your AI, real Jev patterns and TypeSafe (optional), plus a plain verdict: Jev fits here, could help, mixed signals or not needed.
• Choose Gemini, xAI or OpenRouter. Read-only: nothing changes on GitHub, and your keys stay in your browser.
• Copy a ready prompt for Claude, or the JevX command, to apply the change locally with your tests running before and after.

Learned from real projects

We studied 82 Jev decisions in 22 open-source projects (see docs/COMMUNITY-PATTERNS.md). More than half replace an LLM call that only picks a label, which shapes the 2.0 plan.

Full changelog: apps/jevx/CHANGELOG.md

v0.4.0

v0.4.0

Install: npm i -g @vij-sameerb5/jevx

Claude Desktop: download jevx-0.4.0.mcpb below and double-click it.

Website: https://jevx.live

Changelog

All notable changes to jevx. Versions follow semver; before 1.0, minor versions may change behaviour.

0.4.0 — 2026-09-23 · first public release candidate

• Learns where and why Jev fits. Every finding now carries a generic pattern (e.g. error-message-regex-classifier: input kind, rule kind, code shape, why Jev beats the rule). With --share, these go to the JevX dataset — scrubbed in code of every file, function, identifier and string literal. Learned outcomes feed back into the patterns score in later releases.
• --min-fit <percent>: write fits from 50–100% (default 70 = strong fits only). Anything under 70 prints a warning; under 50 is refused. --include-disagree also writes fits the sources disagree on. --include-possible = --min-fit 50.
• Every AI app: jevx mcp install now configures Claude Code, Claude Desktop, Cursor, Windsurf, VS Code (Copilot), Gemini CLI and Codex (backs up each config; jevx mcp uninstall removes it). jevx.mcpb: one-click Claude Desktop extension with a project-folder picker and a secure TypeSafe key field.
• jevx_apply / jevx_undo MCP tools for apps that can't edit files: backup, tests before/after, auto-revert, never under 50%. The MCP server refuses to index / or your home folder.
• Scorecard calibrated on real data (docs/SCORECARD.md): TypeSafe's "bounded" is now a gate instead of a free ~0.25 (it was ≈ 0.97 for every spot, controls included); the AI gets an explicit 0.1–0.9 rubric tied to the verdict bands. Real Jev sites still reach 84–90%; deterministic controls drop to WEAK. The 70% bar is unchanged.
• Fewer approval prompts: jevx_read reads up to 20 files per call; read-only tools say so. The scan reports whether TypeSafe is configured.
• Privacy verified by tests: read-only tools write nothing and send nothing, even with sharing on; only jevx_share sends. The .mcpb build refuses personal notes, .env, research data, local paths or key-like strings.
• Claude Code / Cursor (MCP) now works like the terminal: balanced guide, a reading order so the editor's AI reads the files itself, the generic pattern on every scorecard, the minimum-fit rule, and an opt-in jevx_share tool.
• First-run welcome: a JEVX logo the first time you run it (jevx --welcome shows it again).
• --report-json <file>: the full result as JSON, for comparing runs (stays on your machine).
• Packaging for npm: MIT license, Node ≥ 20.10, release check script.

0.3.0 — 2026-09-23 · the AI reads your code

• The AI reads the source itself, part by part (logic folders first, UI last), instead of only the spots static analysis picked. Static analysis now only sets the reading order, and is the fallback if reading fails.
• Balanced second check: an added network call is not a reason to reject; small spots get low scores instead of a "no".
• Fixed: AI calls were cut off after 30 seconds (now 4 minutes).
• --dry-run previews POSSIBLE fits too; --include-possible writes them.
• --fast: less reasoning while reading (cheaper).
• Keys from an env file (JEVX_ENV_FILE or ~/.jevx/.env); your shell always wins.
• Opt-in anonymous outcomes (--share) to the JevX dataset (Supabase, insert-only, no code).
• .jevx/debug/ explains every run; .jevx/ ignores itself in git.

0.2.0 — 2026-09-22 · one command

• jevx finds, judges and changes strong fits with no picking or accepting.
• Safety in the tool: the old rule stays as the fallback, callers are updated, files with your uncommitted changes are skipped, your tests run before and after (a change that breaks them is reverted automatically), backups + jevx undo.
• jevx --dry-run, jevx mcp install (Claude Code + Cursor), scorecards and red/green diffs in the terminal (#c15f3c).

0.1.0 — 2026-09-22 · MCP server

• jevx-mcp: JevX as tools for Claude Code / Cursor — scan, read, related, search, a three-source scorecard (AI + TypeSafe + patterns), red/green preview and an HTML report.

Before 0.1 — research (not released)

Sessions 1–26 (2026-09-19 → 09-21): a static decision detector, TypeSafe validation and calibration, Gemini / OpenRouter / xAI code analysts, and a study of 30 real Jev projects that produced the first pattern profile (Layer E). That research is frozen; its profile is one of the three scorecard inputs.