islo-reviewer

Composite GitHub Actions for automated PR review and CI babysitting. Runs inside Islo sandboxes using the Claude Agent SDK.

Quick Start

Add an ISLO_API_KEY secret to your repo, then create two workflow files:

.github/workflows/islo-review.yml — reviews PRs on open:

name: PR Review
on:
  pull_request:
    types: [opened, reopened]
jobs:
  review:
    runs-on: ubuntu-latest
    timeout-minutes: 15
    steps:
      - uses: islo-labs/islo-reviewer/review@v1
        with:
          pr_number: ${{ github.event.pull_request.number }}
        env:
          ISLO_API_KEY: ${{ secrets.ISLO_API_KEY }}

.github/workflows/islo-babysit.yml — fixes CI failures automatically:

name: Babysit CI
on:
  workflow_run:
    workflows: ["CI"]  # replace with your CI workflow name(s)
    types: [completed]
jobs:
  babysit:
    runs-on: ubuntu-latest
    timeout-minutes: 15
    if: >
      github.event.workflow_run.conclusion == 'failure' &&
      github.event.workflow_run.event == 'pull_request' &&
      github.event.workflow_run.head_repository.full_name == github.event.workflow_run.repository.full_name
    steps:
      - uses: islo-labs/islo-reviewer/babysit@v1
        with:
          run_id: ${{ github.event.workflow_run.id }}
        env:
          ISLO_API_KEY: ${{ secrets.ISLO_API_KEY }}

That's it. Both actions handle sandbox creation, script execution, and cleanup.

Actions

islo-labs/islo-reviewer/review@v1

Automated PR code review. Reads the diff, explores the codebase, optionally runs tests, and posts a GitHub review with inline comments.

islo-labs/islo-reviewer/babysit@v1

Automated CI failure fixer. Reads CI logs, fixes mechanical issues (lint, formatting, type errors, test updates), and pushes a fix commit. Does not change logic.

Inputs

Both actions share the same configuration inputs (except pr_number vs run_id):

Input	Required	Default	Description
pr_number	yes (review)	—	PR number to review
run_id	yes (babysit)	—	Failed workflow run ID
islo_config	no	''	Path to an islo.yaml for sandbox config (e.g., .github/islo-review.yaml). Triggers a repo checkout so the file is available.
snapshot	no	''	Sandbox snapshot name. Use to restore a pre-built environment.
cpu	no	4	CPU cores for the sandbox
memory	no	4096	Memory in MB for the sandbox
model	no	claude-opus-4-6	Claude model to use (e.g., claude-sonnet-4 for lower cost)
max_turns	no	50	Maximum agentic turns (tool-use round trips)
max_budget_usd	no	10	Cost cap in USD.

Customizing Review Context

Create a REVIEW.md file at the root of your repo to provide extra context to the reviewer. This file is automatically read and injected into the review/babysit prompt.

Example:

## Architecture

React frontend using Vite, React Router, and TanStack Query.

## Cross-repo context

The full stack is available at `/workspace/`:
- `/workspace/backend` -- Python/FastAPI backend (this app consumes its API)

## Review focus areas

- Component structure and React patterns
- API client changes (must match backend contract)
- Accessibility and responsive design

Advanced Configuration

Using an islo.yaml config

If your sandbox needs sources, setup scripts, or other config from an islo.yaml, point the action at it:

- uses: islo-labs/islo-reviewer/review@v1
  with:
    pr_number: ${{ github.event.pull_request.number }}
    islo_config: .github/islo-review.yaml
  env:
    ISLO_API_KEY: ${{ secrets.ISLO_API_KEY }}

This checks out your repo on the runner so the config file is available, then passes --config .github/islo-review.yaml to islo use.

Cost control

Use model and max_budget_usd to control spend:

- uses: islo-labs/islo-reviewer/review@v1
  with:
    pr_number: ${{ github.event.pull_request.number }}
    model: claude-sonnet-4
    max_budget_usd: '2.00'
  env:
    ISLO_API_KEY: ${{ secrets.ISLO_API_KEY }}

Manual trigger

Both workflow examples above support workflow_dispatch for manual reruns. For review, add:

on:
  pull_request:
    types: [opened, reopened]
  workflow_dispatch:
    inputs:
      pr_number:
        description: "PR number to review"
        required: true

Then use ${{ github.event.pull_request.number || inputs.pr_number }} as the pr_number input.

How It Works

1. GitHub Action triggers on PR open (review) or CI failure (babysit)
2. Action installs the Islo CLI and creates an ephemeral sandbox
3. Inside the sandbox, it clones this repo and runs the appropriate script
4. The script uses the Claude Agent SDK to analyze code and take action
5. Sandbox is destroyed after the script completes

Safety

• Babysit only fixes mechanical issues — lint, formatting, type errors, test updates
• Bot commit counter prevents infinite fix-fail-fix loops (max 3 attempts)
• Review triggers once per PR open, not on every push
• max_budget_usd provides a hard cost cap per invocation