dstl-code-review 0.1.0

Distill code review patterns from GitHub PRs into Claude Code skill files

dstl-code-review

CLI tool that distills GitHub PR review comments into reusable skill files for Claude Code. Powered by pydstl.

The Problem

Your team's best code review feedback is buried across hundreds of PRs. New contributors don't benefit from it, and AI assistants don't know your team's standards.

The Solution

dstl-code-review extracts review comments from your repos, filters out noise (bots, "LGTM", short comments), and distills the meaningful ones into a structured skill file that Claude Code can reference during development.

Install

You need uv — a fast Python package manager. If you don't have it:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

That's it. No cloning, no virtual environments — uvx handles everything.

Setup

Before running, export your API keys:

# GitHub — use a fine-grained token (https://github.com/settings/tokens?type=beta)
# "Public repositories (read-only)" scope is enough for public repos
export GITHUB_TOKEN=ghp_your_token_here

# LLM — set the key for whichever provider you're using
export GEMINI_API_KEY=your_key       # Google Gemini
export OPENAI_API_KEY=your_key       # OpenAI
export ANTHROPIC_API_KEY=your_key    # Anthropic

Usage

Navigate to your project directory and run the command — the skill file will be created relative to where you run it:

cd ~/projects/my-app
uvx dstl-code-review \
  --repo your-org/your-repo \
  --model google-gla:gemini-3-flash-preview

This creates .claude/skills/code-review/SKILL.md inside your project, ready for Claude Code to pick up.

Quick example

Try it with real public repos:

uvx dstl-code-review \
  --repo encode/httpx \
  --repo Textualize/rich \
  --model google-gla:gemini-3-flash-preview \
  --limit 20

Single repo

uvx dstl-code-review \
  --repo your-org/your-repo \
  --model google-gla:gemini-3-flash-preview

Fetches all PR review comments, filters noise, and distills them into .claude/skills/code-review/SKILL.md.

Multiple repos

Pass --repo multiple times to distill comments from several repos into a single consolidated skill file:

uvx dstl-code-review \
  --repo your-org/frontend \
  --repo your-org/backend \
  --repo your-org/shared-lib \
  --model google-gla:gemini-3-flash-preview

Each repo is distilled independently first, then the results are consolidated into one unified document. Patterns from each repo are preserved without being diluted.

Incremental runs

Re-running the same command only fetches new comments since the last run. Evidence accumulates in a local SQLite DB, so you can run it periodically to keep the skill file up to date:

# First run — fetches everything
uvx dstl-code-review --repo your-org/api --model google-gla:gemini-3-flash-preview

# Later — only fetches new comments, re-distills with all evidence
uvx dstl-code-review --repo your-org/api --model google-gla:gemini-3-flash-preview

Custom output location

uvx dstl-code-review \
  --repo your-org/your-repo \
  --model google-gla:gemini-3-flash-preview \
  --output ./my-project/.claude/skills/code-review

Output is always SKILL.md inside the specified directory.

How It Works

GitHub PRs ──► Fetch comments ──► Filter noise ──► Store as evidence
                                                         │
                                              ┌──────────┴──────────┐
                                              ▼                     ▼
                                        Per-repo distill    Per-repo distill
                                              │                     │
                                              └──────────┬──────────┘
                                                         ▼
                                                   Consolidate
                                                         │
                                                         ▼
                                          .claude/skills/code-review/SKILL.md

Filtering — automatically drops bot comments (dependabot, copilot, etc.), low-signal phrases ("LGTM", "+1", "looks good"), and comments shorter than 30 characters.

Options

--repo         GitHub repo (owner/name), repeatable [required]
--model        LLM model string [required]
--limit        Max comments per repo (default: 100)
--min-length   Min comment length to keep (default: 30)
--db-path      SQLite database path (default: .claude/dstl-code-review.db)
--output       Output skill directory (default: .claude/skills/code-review)
--token        GitHub token (or set GITHUB_TOKEN env var)

Supported LLM models

Any model supported by pydantic-ai:

--model google-gla:gemini-3-flash-preview    # Google Gemini
--model openai:gpt-4o                         # OpenAI
--model anthropic:claude-sonnet-4-20250514    # Anthropic

Development

uv sync --all-extras
uv run pytest tests/ -v

License

MIT