type: Schema Reference title: prgenius description: PR Genius — 提交前改进顾问，stdlib-first CLI + stdio MCP shell version: 1.1.1 created: 2026-07-04 updated: 2026-07-09 author: zsxh1990

prgenius

PR Genius — 提交前改进顾问。Local-only. Stdlib-first.

A pure-Python (zero hard deps) library + CLI for PR contribution quality analysis.

Install

pip install prgenius-core

Or run from a checkout:

cd prgenius
PYTHONPATH=src python3 -m prgenius --version

Important: PyPI package is the interface, not the data

The PyPI wheel ships only the Python code (prgenius/ package). The knowledge base (profile markdown, case studies, OKF schemas) lives in the GitHub repo at https://github.com/zsxh1990/pr-genius and is not bundled into the wheel. To use prgenius-core after pip install, point it at a checkout of the knowledge base:

git clone https://github.com/zsxh1990/pr-genius
prgenius-core --repo-root ./pr-genius profile get astral-sh/uv

If you cloned a specific tag/commit, the package's __version__ should match (e.g. v0.7.7 ↔ prgenius-core==0.7.7).

Optional: MCP server

The MCP entry point (prgenius-core mcp serve) requires the mcp package:

pip install "prgenius-core[mcp]"

Quick Start

# 1) Schema reference
PYTHONPATH=src python3 -m prgenius schema info

# 2) Look up one repo
PYTHONPATH=src python3 -m prgenius profile get astral-sh/uv

# 3) List currently-open case studies
PYTHONPATH=src python3 -m prgenius case list --status=open

# 4) NDJSON dump of every case study (for benchmarks / agent context)
PYTHONPATH=src python3 -m prgenius dump > cases.ndjson

# 5) Run as stdio MCP server (for Cursor/Cline/Claude Code)
PYTHONPATH=src python3 -m prgenius mcp serve

Programmatic use

from prgenius.parser import profile_get, iter_case_studies, schema_info

prof = profile_get("path/to/repo_root", "astral-sh/uv")
print(prof["frontmatter"]["title"])

for case in iter_case_studies("path/to/repo_root"):
    fm = case["frontmatter"]
    if fm.get("final_status") == "open":
        print(fm.get("pr_number"), fm.get("repo"))

What's exposed

MCP surface (when mcp is installed)

6 tools for local agents:

• analyze_pr(title, repo, body, ...) — 结构化信号 + 建议 + 三档风险
• coach_pr(title, repo, body, ...) — pass/fail + checklist
• get_repo_profile(repo) — 仓库画像
• list_open_prs() — open PR 列表
• get_case_study(repo, pr_number) — PR 案例
• schema_info() — schema 版本

No network, no auth. Stdio only.

--repo-root flag

The MCP server defaults to a path computed from its install location (<package>/../.. × 3 to reach the knowledge base). When that path is wrong (editable install, venv, worktree, fork), pass --repo-root:

python3 -m prgenius --repo-root /path/to/big-repo-pr-knowledge mcp serve

Wiring into Cursor / Claude Code / Cline

These editors all read MCP server config from JSON. Point command at python3 -m prgenius and args at mcp serve (add --repo-root if the auto-detected path doesn't match your checkout). No env keys, no auth.

Claude Code (~/.claude/mcp.json or project-local .mcp.json):

{
  "mcpServers": {
    "pr-genius": {
      "command": "python3",
      "args": ["-m", "prgenius", "--repo-root", "/abs/path/to/big-repo-pr-knowledge", "mcp", "serve"]
    }
  }
}

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "pr-genius": {
      "command": "python3 -m prgenius",
      "args": ["--repo-root", "/abs/path/to/big-repo-pr-knowledge", "mcp", "serve"]
    }
  }
}

Cline (VS Code cline_mcp_settings.json):

{
  "mcpServers": {
    "pr-genius": {
      "command": "python3",
      "args": ["-m", "prgenius", "--repo-root", "/abs/path/to/big-repo-pr-knowledge", "mcp", "serve"],
      "disabled": false
    }
  }
}

Replace /abs/path/to/big-repo-pr-knowledge with the actual checkout. Omit --repo-root only if you installed prgenius from this exact checkout (then the default path resolves correctly).

Schema we honor

• rounds v0.5.0 — action enum + delta object + case-level close_decision
• rounds v0.7.0 — adds optional verified_at, evidence_urls, confidence to round + case level (BC over v0.5.0)

See ../ROUNDS_SCHEMA.md for the canonical schema.

Why this exists

The repo is meant to be human-readable (<org>-<repo>/index.md) AND machine-readable via this package. People get the narrative; agents get structured tool calls. Both views share one source of truth.

License

MIT (same as parent repo).