AgentWorks PR Review Pipeline

A workshop-teachable PR review pipeline for spring-projects/spring-ai built with the AgentWorks stack.

Prerequisites

1. Java 21+

java -version   # must be 21 or later

2. Claude Code

The AI assessment steps delegate to Claude Code. Install it globally:

npm install -g @anthropic-ai/claude-code

Then set your API key:

export ANTHROPIC_API_KEY=sk-ant-...

Workshop attendees: get a key at pollack.ai/key

Without Claude Code or an API key, run with --workshop.skip-ai=true to exercise the deterministic pipeline only.

3. Local clone of spring-ai

The pipeline rebases PRs and runs tests locally, so it needs a clone of the target repository:

git clone https://github.com/spring-projects/spring-ai.git ../spring-ai

The default location is ../spring-ai (sibling directory). Override with:

# Environment variable
export SPRING_AI_REPO_DIR=/path/to/spring-ai

# Or command-line property
--workshop.repo-dir=/path/to/spring-ai

4. GitHub token (recommended)

Without a token the GitHub API rate limit is 60 requests/hour. Set a personal access token for 5,000 req/hr:

export GITHUB_TOKEN=ghp_...

Quick Start

# Pre-flight check — validates all prerequisites
./mvnw spring-boot:run -Dspring-boot.run.arguments="--check"

# Review a specific PR (default: 5774)
./mvnw spring-boot:run -Dspring-boot.run.arguments="5774"

# Deterministic-only mode (no AI, no API key needed)
./mvnw spring-boot:run -Dspring-boot.run.arguments="5774 --workshop.skip-ai=true"

When complete, open the generated report:

open reports/review-pr-5774.html     # macOS
xdg-open reports/review-pr-5774.html # Linux

Output

• Report: reports/review-pr-{N}.html and reports/review-pr-{N}.md
• Journal diary: journal/experiments/pr-review/runs/{uuid}/events.jsonl

The journal diary is the "let me show you what the agent did" artifact — a machine-readable trace of every step, judge verdict, state transition, and LLM call.

Configuration

All properties can be set in application.yml, as environment variables, or as command-line arguments:

Property	Default	Description
workshop.default-pr	5774	PR number when none specified on command line
workshop.skip-ai	false	Skip AI assessment steps
workshop.fix-tests	false	Enable AI fix-tests step when tests fail
workshop.use-dsl	true	Use DSL workflow (false = manual orchestrator)
workshop.journal-dir	./journal	Directory for journal output
workshop.repo-dir	../spring-ai	Local clone of the target repository
github.repo	spring-projects/spring-ai	GitHub owner/repo
github.base-url	https://api.github.com	GitHub API base URL
github.token	(none)	GitHub personal access token

Pipeline Phases

1. Phase 1: Deterministic Context Gathering — fetch PR metadata, rebase, detect conflicts, run tests
2. T0 Gate: Build Judge — deterministic pass/fail on rebase, conflicts, and test results
3. T1 Gate: Version Pattern Judge — deterministic scan for Boot 3-to-4 migration anti-patterns (FAIL sets verdict but does not block AI)
4. Phase 2: AI Assessment — Claude-powered code quality and backport assessment (skipped only if T0 fails or skip-ai=true)
5. T2 Gate: Quality Judge — LLM meta-judge cross-checking AI assessments
6. Phase 3: Report Generation — markdown report summarizing all findings

Workflow Modes

The pipeline ships with two workflow implementations:

• DSL workflow (default) — composed using the Workflow.define() DSL from agent-workflow
• Manual workflow — explicit step-by-step orchestration for maximum readability

Switch between them:

# DSL (default)
./mvnw spring-boot:run -Dspring-boot.run.arguments="5774"

# Manual orchestrator
./mvnw spring-boot:run -Dspring-boot.run.arguments="5774 --workshop.use-dsl=false"

Both produce identical reports. Previous manual reports are preserved as reports/review-pr-*-manual.* for comparison.

Build

./mvnw verify          # full build with tests and quality checks
./mvnw test            # tests only