Jev-Reviewer

Your agent wrote the code. You make the call.

Jev-Reviewer turns each changed GitHub file into a behavioral comparison:

• Old logic — what the committed code used to do.
• New logic — what it does after the change.
• What changed — the difference in behavior.
• Human review question — the decision or assumption to inspect.

TypeSafe Jev assigns attention priorities. OpenAI writes the explanations. The extension keeps GitHub's PR header, tabs, file rows, filenames, and native collapse controls. It replaces only each file's code table with Old logic and New logic columns, plus a concise change note. P0 files open automatically; P1 and P2 start collapsed. The original code remains one popup toggle away. Priorities suggest where to spend attention, not whether code is correct.

This is an early, open-source demo for reviewing your own coding agent's PRs on the same computer. It does not publish PR comments or send a report to teammates.

Watch the short recording · Open the demonstration PR

The recording uses real Jev classifications and clearly labeled prepared explanation copy. Live OpenAI explanations are implemented but verification is pending funded API access; see demo provenance.

Quick demo

Requirements: Node.js 22 or newer.

git clone https://github.com/egma-ai/jev-reviewer.git
cd jev-reviewer
npm install
npm run demo

Open http://127.0.0.1:4731/demo for the standalone replay, or load the extension and open the demonstration PR's Files changed page to see the native GitHub integration. The bundled replay is explicitly labeled and makes no provider calls. Its provenance is included in demo/README.md.

Use on a real PR

Install GitHub CLI and sign in with gh auth login. Keep a local clone of the PR's base repository, with origin pointing to that GitHub repository. Historical PRs work too; the CLI fetches missing commits without checking out another branch.

# Optional command alias; otherwise use node /path/to/jev-reviewer/bin/jev-reviewer.mjs
npm link

# Enter keys in your own terminal, never in an agent chat.
jev-reviewer setup

# Recommended: local static context graph, no extra model key.
uv tool install graphifyy

# Start the local browser bridge in a terminal and leave it running.
jev-reviewer serve

In a second terminal or through the installed skill:

jev-reviewer analyze \
  --pr https://github.com/OWNER/REPOSITORY/pull/123 \
  --repo /path/to/local/clone

The CLI uses TYPESAFE_API_KEY and OPENAI_API_KEY from the local setup file, falling back to environment variables. Optional model overrides: JEV_MODEL and OPENAI_MODEL. Live analysis sends the selected committed code, diff, and context to TypeSafe and OpenAI. Graphify runs locally in --code-only mode. It is a structural aid, not a complete runtime dependency map.

Keys entered through setup are saved to ~/.config/jev-reviewer/credentials.json with owner-only permissions. This file is outside the repository and not encrypted. Saved keys take precedence over inherited environment variables. To replace an exhausted OpenAI key, run node scripts/setup-keys.mjs --replace-openai locally. Reports, which contain code excerpts, are stored in ~/.cache/jev-reviewer/reviews with owner-only file permissions.

Load the Chrome extension

1. Open chrome://extensions and enable Developer mode.
2. Click Load unpacked and select this repository's extension/ folder. No build or store publication is needed.
3. Run jev-reviewer token locally and copy the pairing token. This is a separate local token, not either model API key.
4. Click the Jev-Reviewer extension icon, then follow Connection → Local pairing token → Save.
5. Open the PR's Files changed page. In the popup, enable Show logic in place of code and use Refresh report after regenerating analysis.

The extension fetches the report from 127.0.0.1:4731. It does not access your filesystem or model credentials. All Jev-Reviewer controls—including connection, refresh, the code/logic toggle, and expand-by-default preferences—live in the extension popup. There is no on-page dashboard, toolbar, or "Analyze PR" button; the CLI/agent generates the report.

When enabled with a current report, the extension replaces only the native diff table inside each matching GitHub file row and adds its P0/P1/P2 badge to the native file header. Turning the view off restores the original tables. If the server is missing or the report is stale or cannot be verified, native code stays visible; a warning badge and the popup explain why.

Install the agent skill

See installation instructions for Codex and Claude Code. The skill instructs your coding agent to run the review step after creating or updating a PR. This is agent guidance, not a guaranteed PR event hook. For other agents that support SKILL.md, use the canonical skill/jev-reviewer/ directory.

Configure priorities

Copy config/policy.json to .jev-reviewer.json in the repository you review, or pass --policy /path/to/policy.json.

Level	Default meaning	Default display
P0	Human judgment required: security, permissions, destructive behavior, consequential contracts	Expanded
P1	Human review recommended: bounded behavioral changes	Collapsed
P2	Low attention: documentation, mechanical changes, clear behavior preservation	Collapsed

Edit the priority descriptions and instructions to steer Jev. alwaysReviewPaths applies deterministic P0 overrides. Incomplete changed code is escalated to uncertainPriority (P0 or P1, never P2). Display preferences can also be changed in the extension. Priority levels are review attention, not bug severity.

Scope and limitations

• Units are diff hunks with surrounding old/new committed source. Related source and Graphify neighbors provide bounded context. Cross-hunk behavior can still be missed.
• Up to 12 units are analyzed by default. Excess units are recorded as P0 / not analyzed; the extension leaves affected files as native code. Increase --max-units explicitly for larger PRs.
• The Graphify snapshot is head-only, bounded to 500 supported code files / about 5 MB. Missing Graphify or unsupported source is surfaced in the report.
• The report records base/head commits. The extension refuses to replace code when the report is stale or GitHub does not expose enough information to verify freshness. Native code stays visible, with a warning badge on the extension icon and status in the popup.
• The demo supports github.com and a local clone. No GitHub App is required. Remote/cloud agents need an additional transport and are outside v0.1.
• Private repositories use your existing local GitHub access. This is local report delivery, not fully offline analysis: provider requests transmit source context.
• API failures are explicit; the live command never silently substitutes a fixture for a provider result.

Development

npm test
node scripts/test-extension.mjs

See architecture, demo guide, and extension details. Runtime uses Node's built-in modules; Playwright is a development dependency for browser verification.

The extension browser test needs Chromium (npx playwright install chromium, or set CHROMIUM_PATH) and port 4731 free. It uses an isolated browser profile and synthetic fixtures; real-provider evidence is documented separately in demo/README.md.

MIT licensed. Independent project; not affiliated with TypeSafe, OpenAI, GitHub, or Graphify.