PR-Agent — Quick Capabilities Summary

Overview:

PR-Agent is a CLI and integration tool that automates pull request (PR) review tasks using LLMs. It provides automated PR reviews, descriptions, code suggestions, documentation helpers, and related workflows for multiple Git providers.

Core Capabilities:

/review (PR review): Generate a reviewer-style audit of a PR including security, tests, effort, and best-practice suggestions.
/describe (PR description): Produce a clear PR description and changelog updates from the diff and user notes.
/improve (Code suggestions): Suggest code improvements, refactorings, and non-functional improvements (style, tests, docs).
/add_docs (Docs generation): Create or update documentation for changed code or new APIs.
/custom_prompt: Run a custom prompt against the PR context.
/ask (Interactive questions): Ask follow-up questions about a PR and receive focused answers.
CLI: pr-agent command-line tool to run review/describe/improve locally or in CI.
GitHub Actions & App: Integrations for automated runs in workflows and as a GitHub App.

Integration / Providers:

Git providers: GitHub, GitLab, Gitea, Bitbucket, Bitbucket Server, Gerrit, local git.
LLM Providers via LiteLLM: OpenAI-compatible endpoints (including Blackbox.ai via OPENAI_API_BASE), Anthropic, Cohere, HuggingFace, and others supported by LiteLLM. Configuration uses provider-prefixed model strings (e.g., openai/gpt-4o).
Secret providers: AWS Secrets Manager and Google Cloud Storage secret provider are supported for secure secret management.

Configuration & Secrets:

Settings are managed with dynaconf and loaded from pr_agent/settings/* and optional repository-level pyproject.toml under tool.pr-agent.
Local secret file: pr_agent/settings/.secrets.toml (DO NOT commit real tokens). The template is pr_agent/settings/.secrets_template.toml.
Key settings:
- openai.key and openai.api_base for OpenAI-compatible endpoints.
- config.model controls default model (use provider-prefixed strings LiteLLM expects).

Extensibility:

The codebase provides modular AI handler layers (algo/ai_handlers) so new providers/adapters can be added.
Token handling and patch generation are pluggable to tune prompt size and chunking.

Development / Quick Start:

Create a Python virtualenv and install:

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Provide secrets locally in pr_agent/settings/.secrets.toml or export env vars (preferred for CI):

export OPENAI_API_KEY="<YOUR_KEY>"
export OPENAI_API_BASE="https://api.blackbox.ai"  # if using Blackbox

Example local run (review):

python cli.py --pr_url="https://github.com/<org>/<repo>/pull/123" review

Security & Best Practices:

Never commit API keys or .secrets.toml to version control.
If secrets are accidentally pushed, rotate and remove from history (BFG or git filter-repo) immediately.
Use a secrets manager for CI and production deployments.

Where to get help / docs:

Project docs are under docs/ and pr_agent/settings/*.toml contain configuration docs and prompts.
For LiteLLM provider formats, see: https://docs.litellm.ai/docs/providers

Files of interest:

If you want, I can:

Add this content into the main README.md replacing or merging with the existing one.
Generate a one-page CONTRIBUTING snippet for local development and secret handling.
Commit the new README file and create a small PR with the change.