Skip to content
项目

can1357/llm-git has 62 GitHub stars and 5 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-21 01:21 UTC. Last push 2026-11-15.

这些数字如何统计

可以做什么

  • Conventional commits — Generates properly formatted commit messages with type, scope, and past-tense summary (≤72 chars)
  • Automatic changelogs — Maintains CHANGELOG.md following Keep a Changelog format with monorepo support
  • Compose mode — Splits large staged changes into multiple logical atomic commits
  • Rewrite mode — Converts entire git history to conventional commits (with automatic backup)

实测星标增长

24小时
0
7天
0
15天
正在收集历史
1个月
正在收集历史

各周期内的星标净变化,周期相互重叠,并非每日新增。缺失的测量保持空白。星标代表关注度,不代表软件质量。

关于

AI commit message generator. Conventional commits, compose mode, and history rewrite via Claude/GPT APIs.

开发工具

README.md

原文

lgit

LLM-powered git commit message generator

CI PyPI License Python

Generates conventional commits from git diffs using Claude AI or any OpenAI-compatible API.
Automatic changelog maintenance, multi-commit composition, and full history rewriting.


Features

  • Conventional commits — Generates properly formatted commit messages with type, scope, and past-tense summary (≤72 chars)
  • Automatic changelogs — Maintains CHANGELOG.md following Keep a Changelog format with monorepo support
  • Compose mode — Splits large staged changes into multiple logical atomic commits
  • Rewrite mode — Converts entire git history to conventional commits (with automatic backup)
  • Map-reduce analysis — Parallel per-file analysis for large commits without truncation
  • Any LLM provider — Works with Anthropic, OpenAI, OpenRouter, or any OpenAI-compatible API

Quick Start

# Install
uv tool install lgit-cli

# Configure (pick one)
export LLM_GIT_API_KEY=your_anthropic_key                    # Direct Anthropic
export LLM_GIT_API_URL=https://openrouter.ai/api/v1          # OpenRouter
litellm --port 4000                                           # Local proxy (default)

# Use
git add .
lgit                    # Analyze, update changelog, commit
lgit --dry-run          # Preview without committing
lgit --compose          # Split into multiple commits

Usage

Basic Commands

lgit                                # Analyze staged changes and commit
lgit --dry-run                      # Preview message without committing
lgit --copy                         # Copy message to clipboard
lgit -p                             # Commit and push
lgit -S                             # GPG sign the commit
lgit -s                             # Add Signed-off-by trailer
lgit --amend                        # Amend the previous commit
lgit > msg.txt                      # Save raw message to file (auto-detected pipe mode)
lgit --dry-run | git commit -F -    # Generate message and commit with custom git flags

# Modes
lgit --mode=unstaged                # Preview unstaged changes (no commit)
lgit --mode=commit --target=HEAD~1  # Analyze a specific commit

# Models
lgit -m opus                        # Use Opus for analysis (more capable)
lgit -m sonnet                      # Use Sonnet (default)

# Context
lgit Fixed regression from PR #123  # Add context via trailing text
lgit --fixes 123 456                # Add "Fixes #123, #456" to body
lgit --breaking                     # Mark as breaking change

Compose Mode

Split staged changes into multiple logical commits:

lgit --compose                      # Propose and create atomic commits
lgit --compose --compose-preview    # Preview splits without committing
lgit --compose --compose-max-commits 5
lgit --compose --compose-test-after-each

Rewrite Mode

Convert repository history to conventional commits:

lgit --rewrite                      # Rewrite full history (creates backup)
lgit --rewrite --rewrite-preview 10 # Preview first 10 commits
lgit --rewrite --rewrite-dry-run    # Show all changes without applying
lgit --rewrite --rewrite-start main~50  # Rewrite last 50 commits only
lgit --rewrite --rewrite-parallel 20    # 20 concurrent API calls

Profiling Trace

Write detailed tracing spans and HTTP timing events to a JSONL file:

lgit --trace-output /tmp/lgit-profile.jsonl --dry-run
LLM_GIT_TRACE_FILE=/tmp/lgit-profile.jsonl lgit --compose

Each span close event includes busy/idle timing from tracing-subscriber; API events also record TTFT, total response time, status, and response size.

Automatic Changelog

lgit automatically maintains CHANGELOG.md files when committing:

  • Auto-detection — Finds all CHANGELOG.md files in your repository
  • Monorepo support — Routes changes to the correct changelog based on file paths
  • Deduplication — Skips entries semantically similar to existing ones
  • Category mapping — Maps commit types to sections (Added, Fixed, Changed, etc.)
project/
├── CHANGELOG.md              ← covers: src/, docs/
├── packages/
│   ├── core/
│   │   └── CHANGELOG.md      ← covers: packages/core/**
│   └── cli/
│       └── CHANGELOG.md      ← covers: packages/cli/**

Disable with --no-changelog or changelog_enabled = false in config.

When staged changes have a changelog boundary, one synthesis call generates the commit analysis and all applicable changelog updates together. Compose uses the same synthesis for each already-planned group; grouping and staging are unchanged. Large inputs reuse per-file map observations, and valid summaries avoid a separate summary call.

Summary repair/rewrite, authored-entry protection, existing-entry revisions, and staged/worktree isolation still apply. Malformed shared output falls back to separate generation. --dry-run, commit/unstaged analysis, and --no-changelog retain their existing generation paths.

Prompt overrides remain explicit: a relevant analysis.md, fast.md, map.md, reduce.md, or changelog.md override keeps separate calls unless the override directory also supplies synthesis.md.

Configuration

Create ~/.config/llm-git/config.toml:

# API
api_base_url = "http://localhost:4000"    # Default: LiteLLM proxy
api_key = "sk-..."                        # Or use LLM_GIT_API_KEY env var

# Model — any role accepts a `;`-separated fallback chain, tried left to right
analysis_model = "claude-sonnet-4.5"      # Diff analysis / classification
summary_model = "claude-haiku-4-5"        # Summary line generation
map_model = ""                            # Per-file map phase; empty = summary_model
fast_model = "flash-lite;haiku"           # Fast mode; falls back when the primary stalls

# Commit message limits
summary_guideline = 72                    # Target length
summary_soft_limit = 96                   # Triggers retry
summary_hard_limit = 128                  # Absolute max

# Features
changelog_enabled = true
auto_fast_threshold_lines = 200           # Auto-use fast mode for small diffs; 0 disables
map_reduce_enabled = true                 # Parallel analysis for large commits
disable_git_background_features = true    # Disables fsmonitor/untrackedCache for lgit subprocesses

# Typed judgments (TypeSafe jev, ~$0.04 per million input tokens, ~1s per round trip).
# With a key: mechanical follow-through hunks (imports, call sites, formatting, lockfiles)
# are pruned from analysis prompts and the rest get a per-hunk type hint; compose binds
# hunks to groups by judgment instead of the analysis model; changelog generation is
# skipped for changes judged not user-visible and paraphrased duplicate entries are
# dropped. Shared commit/changelog synthesis does not require a judge key.
judgment_api_key = ""                     # Or TYPESAFE_API_KEY env var; `!cmd` runs a command
judgment_model = "jev-latest"
mechanical_hunk_threshold = 0.7           # Prune hunks judged mechanical with >= this probability
mechanical_prune_min_tokens = 5000        # Skip pruning below this diff size (round trip not worth it)
mechanical_prune_min_model_rate = 1.0     # ...and unless the analysis model bills >= this $/M (judging costs ~$0.08/M of diff)
changelog_gate_threshold = 0.3            # Skip the changelog model below this user-visibility; 0 disables

# Commit signing
gpg_sign = false                          # GPG sign commits by default (-S)
signoff = false                           # Add Signed-off-by trailer by default (-s)

Provider Examples

Anthropic Direct:

api_base_url = "https://api.anthropic.com/v1"
api_key = "sk-ant-..."

OpenRouter:

api_base_url = "https://openrouter.ai/api/v1"
api_key = "sk-or-..."
analysis_model = "anthropic/claude-sonnet-4.5"
summary_model = "anthropic/claude-haiku-4-5"

OpenAI:

api_base_url = "https://api.openai.com/v1"
api_key = "sk-..."
analysis_model = "gpt-4o"
summary_model = "gpt-4o-mini"

Judgments (optional, any provider above):

judgment_api_key = "!cat ~/.secrets/typesafe"   # key from https://console.typesafe.ai/

Set LLM_GIT_JUDGMENT_DISABLED=1 to switch judgments off without removing the key.

The client uses the same response formats across OpenRouter, LiteLLM, Anthropic, and OpenAI endpoints: markdown/plain text for individual operations and a validated JSON object for shared commit/changelog synthesis.

Commit Types

Customize commit type classification:

[types.feat]
description = "New public API or user-observable behavior change"
diff_indicators = ["pub fn", "pub struct", "export function"]

[types.fix]
description = "Fixes incorrect behavior"
diff_indicators = ["unwrap() → ?", "bounds check", "error handling"]

[types.refactor]
description = "Internal restructuring with unchanged behavior"
hint = "If behavior changes, use feat instead."

Changelog Categories

[[categories]]
name = "Breaking"
header = "Breaking Changes"
match.body_contains = ["breaking", "incompatible"]

[[categories]]
name = "Added"
match.types = ["feat"]

[[categories]]
name = "Fixed"
match.types = ["fix"]

[[categories]]
name = "Changed"
default = true

Environment Variables

Variable Description Default
LLM_GIT_API_URL API endpoint http://localhost:4000
LLM_GIT_API_KEY API key none
LLM_GIT_CONFIG Config file path ~/.config/llm-git/config.toml
LLM_GIT_VERBOSE Debug output false
LLM_GIT_TRACE_FILE JSONL profiling trace output path none

Installation

From PyPI

uv tool install lgit-cli    # recommended
pipx install lgit-cli       # or
pip install lgit-cli

From source

git clone https://github.com/can1357/llm-git.git
cd llm-git
uv tool install .

Prerequisites

  • Python 3.14+
  • Git
  • API access (Anthropic, OpenAI, OpenRouter, or local LiteLLM proxy)

License

MIT