ebpro/ai-review

Canonical reusable AI code-review workflow for the ebpro org.

Call .github/workflows/ai-review-reusable.yml from any repository — public or private — to get an automated review of every pull request by 11 specialized opencode agents (architecture, security, performance, testing, docs, and more) running on self-hosted runners.

This repo lives here (public) because GitHub forbids a public consumer repo from calling a private repo's reusable workflow. The runner image, agent definitions, and their build pipeline remain in the private ebpro/quarkus-ci-runner repo — consumers never need them.

Consumer quickstart

Add this file to your repo at .github/workflows/ai-review.yml:

# AI Review — thin caller for the ebpro reusable AI review workflow.
name: AI Review

on:
  pull_request:
    types: [opened, synchronize, reopened]

permissions:
  contents: read
  pull-requests: write

jobs:
  ai-review:
    # Skip fork PRs (they cannot receive org secrets anyway).
    if: github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name
    uses: ebpro/ai-review/.github/workflows/ai-review-reusable.yml@main
    with:
      pr_number: ${{ github.event.pull_request.number }}
      repo: ${{ github.repository }}
      agents: all          # "all" | "code" | "lecture" | '["review-security",...]'
    secrets:
      LLM_REVIEW_ENDPOINT: ${{ secrets.LLM_REVIEW_ENDPOINT }}
      LLM_REVIEW_TOKEN: ${{ secrets.LLM_REVIEW_TOKEN }}

Why explicit named secrets instead of secrets: inherit: least privilege — the review job receives only these two values, not every repository/org secret an inherit would forward to the called workflow. On ref pinning: @main is used here for teaching convenience; the hub is org-owned. If you want stronger guarantees, pin a commit SHA (@<full-sha>) and/or enable the org-level "require SHA pinning" setting.

Inputs

Input	Required	Default	Description
pr_number	yes	—	PR number to review (${{ github.event.pull_request.number }})
repo	yes	—	owner/name of the consuming repo (${{ github.repository }})
agents	no	all	all, code, lecture, or a JSON array of agent names

Prerequisites

For the workflow to execute (not just parse), the consuming repo needs:

1. Org secrets shared to it: LLM_REVIEW_ENDPOINT and LLM_REVIEW_TOKEN (GitHub org/user secrets → Secrets and variables → Actions → repository access, or dependabot-style selected access). Set at the org level, the caller passes them through with an explicit named-secret mapping (LLM_REVIEW_ENDPOINT: ${{ secrets.LLM_REVIEW_ENDPOINT }}, etc.), which resolves org-level secrets in the caller context.
   • Public repos need the secrets' visibility widened to public repos — by default org secrets are visible only to private repositories.
2. Runner access: the workflow targets self-hosted runners labeled ebpro-ai-review (plus linux, self-hosted, best-practices). For organization runner groups / runner scale sets, grant the consuming repo access in the runner group settings.

If either is missing, jobs will queue forever or the LLM connectivity pre-check will fail — everything else still works and each agent posts a clear fallback comment rather than silently disappearing.

What you get

Each selected agent posts one issue comment on the PR, identified by an <!-- ai-review:<agent> --> marker (updated in place on later pushes). If an agent errors (bad model config, API outage), a fallback comment records the failure so the review is visibly incomplete instead of silently missing.

Fork PRs

Fork pull requests are skipped by the if: guard above. GitHub never exposes org/repo secrets to workflows from forks, and running third-party code on your self-hosted runners is a well-known supply-chain risk. Internal-branch PRs always run.

Security posture

• This repository contains no secrets, no internal URLs, no credentials. It is 100% safe to read, and public by design.
• The review model, endpoint, and agent definitions live in the private runner image built by ebpro/quarkus-ci-runner; nothing sensitive is evaluated from this repo's contents.
• Org secrets reach the runner only through the explicit named-secret passthrough above, on PRs from same-repo branches (fork PRs are skipped, see above).
• The reusable workflow only reads your PR diff and posts comments; it never writes to your repository.

Related

• ebpro/quarkus-ci-runner — the private runner image + 11 opencode agents + the dogfooding caller for this hub.