code-review-ai-cli 1.3.0

Automated AI-powered code review CLI for Azure DevOps / TFS Pull Requests

AI Code Review CLI

Automated code review tool with Pull Request integration for Azure DevOps/TFS and support for multiple LLM providers. Documentation where docs/index.md for complete guides on CLI usage, LLM configuration, and architecture.

Features

• AI Pull Request Review — Automated code analysis with configurable LLM providers
• Structured Comments — Inline suggestions + general summary comments
• Dry-run Mode — Validate reviews before posting
• Multiple LLM Providers — OpenAI, Azure OpenAI, Gemini, Claude, Ollama, GitHub Copilot, AWS Bedrock
• Smart Filtering — Filter by file extensions, limit diff size
• Project-aware PR Context — Sends repository and linked work item context while restricting findings to modified PR lines
• RAG Context — Enriches the review with related code snippets found via git grep in the local repository
• Single Reviewer Context — One Markdown context file with local override support
• Usage Tracking — Store per-PR token usage and optional cost estimates
• Interactive CLI — Menu-driven selection and confirmation

Local Repository Requirement

The CLI must be run from inside the local clone of the repository being reviewed.

The PR diff is fetched from Azure DevOps/TFS, but several features depend on the local git state:

Feature	Local dependency
PR diff	git fetch + three-dot diff against origin/<branch>
RAG context	git grep on the working tree
Branch validation	git branch --show-current

Branch requirement when RAG is enabled

When review.rag.enabled: true, the local repository must be checked out on the PR target branch (e.g. development). If the local branch differs from the target, the review is blocked:

Local branch mismatch: you are on 'main' but this PR targets 'development'.
RAG context would be built from the wrong branch, which may corrupt the review.
Please run:  git checkout development

Example:

# PR: merge 'feature/my-feature' → 'development'
cd /path/to/my-repo          # must be inside the repo
git checkout development     # must match PR target branch
ai-review pr-review 42

To skip the branch check entirely, disable RAG:

review:
  rag:
    enabled: false

Installation & Quick Start

1. Install Package

From PyPI:

pip install code-review-ai-cli

With optional LLM SDK extras:

pip install "code-review-ai-cli[bedrock]"    # AWS Bedrock
pip install "code-review-ai-cli[openai]"     # OpenAI SDK
pip install "code-review-ai-cli[gemini]"     # Google Gemini SDK
pip install "code-review-ai-cli[claude]"     # Anthropic Claude SDK
pip install "code-review-ai-cli[all]"        # All optional SDKs

For development:

pip install -r requirements.txt
pip install "code-review-ai-cli[dev]"        # Test suite + linting

2. Initialize Configuration

Generate config templates in your project directory:

ai-review init

Creates:

• config.yaml — LLM and review settings
• review_context.example.md — Canonical example reviewer context
• review_context.local.md — Local reviewer context override, ignored by git
• .gitignore entry for review_context.local.md

3. Run Your First Review

# List active PRs
ai-review list-prs

# Review a specific PR
ai-review pr-review 42

# Dry-run (preview without posting)
ai-review pr-review 42 --dry-run

# Check stored token/cost usage
ai-review usage

Configuration File

After ai-review init, edit config.yaml:

llm:
  provider: bedrock                    # or: openai, gemini, claude, etc.
  model: anthropic.claude-3-5-sonnet-20240620-v1:0

bedrock:
  region: us-east-1

tfs:
  base_url: https://dev.azure.com/your-org
  project: YourProject
  pat: xxxxxxxxx

review:
  language: pt
  verbosity: detailed                  # or: quick, security
  scope: diff_with_context             # diff_with_context, diff_only
  file_extensions_filter: [".cs", ".ts", ".py"]
  max_diff_files: 50
  max_comments_to_post: 20
  custom_prompt_file: review_context.local.md
  rag:
    enabled: true                      # requires local branch == PR target
    max_chars: 40000
  project_context:
    enabled: true
    mode: on_demand                      # on_demand, full
    manifest_max_chars: 60000
    retrieval_max_rounds: 2
    retrieval_max_files: 20
    retrieval_max_chars: 120000
    retrieval_file_max_chars: 30000
  work_item_context:
    enabled: true
    max_items: 20

By default, repository context is loaded on demand: the prompt includes explicit source-branch change packets, full changed-file contents as read-only context, linked work item documentation, and a repository manifest, then the model requests any extra files it needs. Inline PR comments must still be grounded in actual changed lines from the change packets. Set review.project_context.mode: full to send the full eligible repository snapshot instead. Bedrock uses a default estimated prompt budget of 180000 tokens; override it with llm.max_prompt_tokens if your model supports more or less.

Reviewer context is loaded from exactly one Markdown file. By default the CLI uses review_context.local.md when it exists, otherwise it falls back to the packaged src/prompts/review_context.example.md. Keep local team tweaks in review_context.local.md; it is gitignored so the canonical context cannot drift across machines.

For Copilot-backed reviews, Claude Sonnet models are often strong choices for large PR validation, for example llm.provider: copilot with a Claude Sonnet model available to your organization. The tool still enforces the same source-branch grounding, duplicate checks, and comment cap regardless of model.

Development & Testing

This is a standard Python project with the following structure:

src/
  ai_review.py        # CLI entry point
  config.py           # Configuration management
  llm_client.py       # LLM provider abstraction
  tfs_client.py       # Azure DevOps integration
  git_utils.py        # Git diff processing
  formatter.py        # Output formatting (terminal, markdown, JSON)
  rag_engine.py       # RAG context via git grep

tests/
  test_*.py           # Unit and integration tests

RAG Context

Current Implementation

The RAG engine enriches the LLM prompt with related code snippets found in the local repository:

1. Extract identifiers — function and class names are parsed from the PR diff
2. Search — git grep finds files containing those identifiers
3. Extract snippets — ±10 lines around each match are included as read-only context

All operations run locally with no extra dependencies. The quality of RAG context depends entirely on the local repository state, which is why the local branch must match the PR target branch.

Recommended Stack for Enhanced RAG (Local & Open Source)

For teams wanting semantic similarity instead of keyword search, the recommended local stack is:

Component	Library	Reason
Vector database	ChromaDB	Runs in-memory or persists to a local SQLite file — no server needed, pip install chromadb
Embeddings	sentence-transformers	Generates vectors locally on CPU — no API calls, no cost

Example integration pattern:

from sentence_transformers import SentenceTransformer
import chromadb

model = SentenceTransformer("all-MiniLM-L6-v2")   # ~80 MB, CPU-friendly
client = chromadb.Client()                          # in-memory
collection = client.create_collection("repo-index")

# Index
collection.add(
    documents=[snippet_text],
    embeddings=model.encode([snippet_text]).tolist(),
    ids=["file:line"],
)

# Query
results = collection.query(
    query_embeddings=model.encode([query]).tolist(),
    n_results=5,
)

This stack keeps the CLI lightweight (pip install) and requires no external services or paid APIs.

Run Tests

python -m pytest --cov=src --cov-report=term

Code Quality

Project follows PEP8 with Black formatter and Ruff linter:

# Format code
black src/ tests/

# Lint
ruff check src/ tests/

# Type checking (if using mypy)
mypy src/

VS Code Integration

Predefined tasks for quick execution in VS Code:

Available Tasks

1. AI Review: Pull Request (Interactive)

• Interactive mode with main menu
• Run with: Ctrl+Shift+B → Select task

2. AI Review: PR (Dry-Run)

• Dry-run for a specific PR
• Prompts for PR ID

3. AI Review: List Active PRs

• Lists active PRs
• Quick diagnostics

4. AI Review: Interactive Mode

• Full tool menu
• Runs in background

How to run tasks

In VS Code:

1. Ctrl+Shift+P → "Tasks: Run Task"
2. Select the desired task
3. Fill in parameters if needed