ReviewFlow AI Documentation

Architecture, installation guides, review policies, prompt injection defenses, and API reference.

1. Installing the GitHub App

ReviewFlow AI integrates via a native GitHub App rather than asking developers to paste personal access tokens. Once installed on your personal account or organization, GitHub delivers signed webhook events on pull_request.opened and synchronize.

# Required GitHub App Permissions (Least Privilege):

• Pull Requests: Read & Write (for line-level inline review comments)

• Checks: Read & Write (for check runs and status badges)

• Repository Metadata: Read-only

2. Prompt-Injection Defense & Sandboxing

Because pull requests can originate from external or untrusted contributors, diff contents may contain adversarial jailbreaks (e.g., hidden comments attempting to command the AI to approve the PR).

•Quarantine Delimiters: All diff lines are strictly enclosed in <UNTRUSTED_DIFF> tags.
•Adversarial Flagging: If an injection command is identified, ReviewFlow AI halts and creates a CRITICAL security alert.
•Zero Code Execution: ReviewFlow AI never runs the untrusted pull request code. Static AST scanning and tokenized diff parsing run in a read-only environment.

3. REST API Reference

ReviewFlow AI provides an OpenAPI 3.0 documented REST API. Authenticate using Bearer tokens in the Authorization header.

GET /api/v1/healthService health check

GET /api/v1/pull-requestsList PRs with severity filter & pagination

GET /api/v1/reviews/:idFetch review job, summary, & findings

POST /api/v1/webhooks/githubGitHub App HMAC webhook endpoint