CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

PR-Helper is a self-hosted web service that auto-generates PR descriptions from Git history and performs AI-powered code review. Designed for internal/local use only (no auth).

Tech Stack

Backend: Go + Gin + SQLite (go-sqlite3) + go-git + chromedp (PDF generation)
Frontend: Go html/template + HTMX + D3.js + diff2html + Tailwind CSS
LLM: OpenAI-compatible API (SSE streaming)
Deploy: Docker (includes Chromium for PDF)

Build & Run

# Build (CGO required for SQLite)
CGO_ENABLED=1 go build -o pr-helper .

# Run
./pr-helper
# Server listens on :8080

# Docker
docker compose up --build

Architecture

handlers/   → HTTP handlers (pages + JSON API + SSE endpoints)
services/   → Business logic (git ops, LLM calls, PDF gen, cache management)
models/     → Data models (repository, settings, analysis)
database/   → SQLite init and migrations
config/     → Configuration loading
templates/  → Go HTML templates (layouts/, pages/, partials/, reports/)
static/     → CSS (Tailwind output), JS (graph, diff-viewer, sse), vendor libs

Data flow: Browser ↔ Gin handlers → services (git/llm/pdf) → SQLite + filesystem (data/)

Key service layer responsibilities:

services/git.go — clone, diff, graph data extraction via go-git
services/llm.go — OpenAI-compatible API calls with SSE streaming
services/generate.go — PR description generation (commits + diff → LLM → structured output)
services/review.go — AI code review (per-file analysis + summary, Top-N strategy for large diffs)
services/pdf.go — chromedp HTML→PDF conversion
services/cache.go — repository cache lifecycle (clone, expiry cleanup)

Key Patterns

SSE streaming: Clone progress, PR generation, and AI review all use Server-Sent Events. Handlers write event: / data: lines; frontend consumes via static/js/sse.js.
Large diff handling: Diffs are split per-file, sorted by change size, truncated to Top-N (configurable, default 20). Each file analyzed independently by LLM, then a summary prompt aggregates results.
Review notes: Three-level scoping — overall, file, suggestion — stored in review_notes table. AI output is read-only; users append notes separately.
Diff rendering: diff2html in Split view by default. AI review suggestions are inline-embedded next to code lines via line + side fields from SSE.

Frontend Libraries (vendored in static/lib/)

HTMX 2.x, D3.js 7.x, diff2html 3.x, highlight.js 11.x
Tailwind CSS compiled to static/css/style.css
Custom JS: sse.js (SSE client), graph.js (D3 git graph with click selection), diff-viewer.js (diff2html + file tree sidebar), review-inline.js (inline AI suggestions)

Data Storage

SQLite DB at data/pr-helper.db — settings (KV), repositories, analyses, review_notes
Cloned repos cached at data/repos/ with configurable expiry (default 7 days)

Development Notes

Refer to PLAN.md for full specification including API routes, SSE event formats, SQL schema, and prompt templates.
No authentication — do not expose to public internet.
All LLM prompts are in PLAN.md §8; model/endpoint/key are configurable via the settings page.

3.3 KiB Raw Blame History