goodpr 0.1.1

Add your description here

goodpr

Generate a professional pull request description from a local git repository using LangChain Deep Agents and Gemini.

Requirements

• Python 3.13+
• uv
• A GOOGLE_API_KEY for the Gemini API

Setup

# Clone and install
uv sync

# Add your Gemini API key
echo GOOGLE_API_KEY=your_key_here > .env

Usage

uv run goodpr <path/to/repo> --commit-offset <N>

Argument	Description
path/to/repo	Absolute or relative path to a local git repository
--commit-offset N	Number of commits back from HEAD to include (default: 5)
--log-file PATH	Log file path (default: goodpr.log)

Example — describe the last 10 commits:

uv run goodpr C:/projects/myapp --commit-offset 10

Output is written to stdout as Markdown, suitable for pasting directly into a GitHub / GitLab PR.

How it works

flowchart TD
    CLI["CLI\nmain.py"]
    Git["git format-patch\nHEAD~N..HEAD"]
    File["patch_context.txt\nraw patch"]
    BuildIndex["build_patch_index()\nchunk by file, BM25"]
    Index["BM25 index\nin-memory"]
    Main["Main Agent\nGemini 3 Flash"]
    SumAgent["summary-agent\nsearch_patch() only"]
    ImplAgent["implications-agent\nsearch_patch() only"]
    SumOut["TITLE / SUMMARY\nFILES / CHANGE_TYPES"]
    ImplOut["BREAKING / MIGRATIONS\nDEPS_CONFIG / TESTING / RISK"]
    PR["Final PR\nMarkdown output"]

    CLI --> Git
    Git --> File
    File --> BuildIndex
    BuildIndex --> Index
    Main -->|"task() STEP 1"| SumAgent
    SumAgent -->|"search_patch(query)"| Index
    SumAgent --> SumOut
    SumOut --> Main
    Main -->|"task() STEP 2"| ImplAgent
    ImplAgent -->|"search_patch(query)"| Index
    ImplAgent --> ImplOut
    ImplOut --> Main
    Main -->|"STEP 3 compose"| PR

Key design decisions

• Raw patch on disk — the full patch is written to patch_context.txt (capped at ~500KB from git format-patch). Subagents do not read this file directly; they only see patch content via search results.
• BM25 search only — at startup, the raw patch is chunked per-file-per-commit and indexed with BM25. Subagents use search_patch(query) to retrieve relevant chunks (e.g. by file name, function, or keyword) and can analyze the full patch without loading it all at once.
• File-based handoff — the patch path is passed in the user message and to build_pr_agent() so the BM25 index can be built. After that, all patch access is through the search_patch tool, not direct file reads.
• Structured subagent output — each subagent returns a fixed labeled format (TITLE:, SUMMARY:, BREAKING:, RISK:, etc.) so the main agent can compose the final PR deterministically.
• Skills — PR writing guidelines are loaded from skills/pr/SKILL.md at runtime.