AI PR Review, Fix & Security Scan

An autonomous GitHub Action that reviews pull requests, fixes bugs, and runs a dedicated security scan — all governed by node9.

---

What It Does

On every push to a non-main branch, the agent runs a four-step pipeline:

Step	What happens
1. Fix loop	Reads changed files, fixes real bugs, runs tests (up to 6 turns)
2. Safety check	Reverts AI changes if tests break after fixes
3. Code review	Reviews the agent's own fixes — posts as 🔍 node9 Code Review on your PR
4. Security review	Data-flow pass on the original diff — posts as 🔒 node9 Security Review on your PR

Two possible outcomes:

• Agent found nothing to fix — one PR (yours), two review comments on it
• Agent fixed bugs — your PR gets review comments + a separate [node9] AI fixes PR with the agent's changes for you to review

---

Quick Start

1. Add the workflow

Create .github/workflows/node9-review.yml in your repo:

name: node9 AI PR Reviewon:
push:
branches: ['**', '!main', '!master', '!node9/**']jobs:
review:
runs-on: ubuntu-latestpermissions:
contents: writepull-requests: writesteps:
- uses: actions/checkout@v4with:
fetch-depth: 0
- name: Configure gitrun: | git config user.name "node9[bot]" git config user.email "bot@node9.ai" git fetch origin main - uses: node9-ai/node9-pr-agent@v1with:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}node9_api_key: ${{ secrets.NODE9_API_KEY }}github_token: ${{ secrets.GITHUB_TOKEN }}test_cmd: 'npm test'

2. Set secrets

In your repo: Settings → Secrets and variables → Actions

Secret	Required	Description
ANTHROPIC_API_KEY	Yes	Anthropic API key — console.anthropic.com
NODE9_API_KEY	No	node9 SaaS key for audit trail — node9.ai

That's it. Push to any feature branch and the agent runs automatically.

---

Inputs

Input	Required	Default	Description
anthropic_api_key	Yes	—	Anthropic API key
github_token	Yes	—	GitHub token for PR operations
node9_api_key	No	—	node9 SaaS key for audit trail
test_cmd	No	npm test	Command to run tests
base_ref	No	main	Base branch to diff against

Examples for test_cmd: pytest, cargo test, go test ./..., npm run build && npm test

---

What You'll See

After each push to a feature branch, the agent posts two comments on your PR:

🔍 node9 Code Review — logic and correctness review of the agent's own fixes (or the original diff if no fixes were needed)

🔒 node9 Security Review — focused data-flow analysis of the original PR diff, looking for:

• User-controlled inputs reaching filesystem sinks (path.join, open())
• Execution sinks (exec, spawn, subprocess)
• Network sinks (URLs constructed from input)
• Deserialization of untrusted input
• Validation gaps (blocklist vs allowlist, unanchored regex)

Findings are rated HIGH / MEDIUM / LOW.

---

Security Model

The agent runs with node9 in standard mode — every tool call (file read, file write, bash command) is logged to the node9 audit trail and dangerous operations are blocked.

Prompt injection mitigations:

• Security review instructions live in the system message, separate from the untrusted diff
• The diff is explicitly labelled as untrusted data in the user message
• Model output is scanned for GITHUB_TOKEN, ANTHROPIC_API_KEY, and NODE9_API_KEY before posting to GitHub

Tool safety:

• File reads and writes are sandboxed to GITHUB_WORKSPACE via safe_path()
• Tool inputs are validated against an allowlist before dispatch
• Branch names are sanitized to prevent shell injection

---

License

Apache 2.0 — see LICENSE.