AI Commit - Automagically generate conventional commit messages with AI

Reads git staged changes via `git diff --staged`, sends the diff content to a pluggable AI generator backend (OpenAI, Moonshot, ERNIE-Bot, or CLI-based tools), and returns a structured Conventional Commits-formatted message with type, scope, subject, and body. The generator system abstracts over both HTTP API clients and subprocess-based CLI tools through a unified interface, enabling swappable backends without code changes.

Implements a generator interface pattern that abstracts over heterogeneous AI backends: HTTP-based APIs (OpenAI Chat/Completions, Moonshot, ERNIE-Bot) and subprocess CLI tools (Bito, GitHub Copilot, GitHub Models). Each generator driver implements a common interface with configurable prompts, retry logic, and error handling. The system loads the appropriate generator based on configuration and CLI flags, enabling runtime backend switching without application code changes.

Built on Laravel Zero, a micro-framework for console applications, providing a structured foundation for command handling, dependency injection, and configuration management. The framework abstracts CLI boilerplate (argument parsing, output formatting, error handling) and provides Laravel's service container for managing generators, configuration, and HTTP clients. This enables rapid feature development and consistent CLI behavior across commands.

Implements an HTTP client layer abstracting over API calls to OpenAI, Moonshot, ERNIE-Bot, and other HTTP-based generators. The client handles request formatting, authentication (API keys), response parsing, error handling, and timeout management. The abstraction allows swapping HTTP clients (e.g., Guzzle, cURL) without changing generator code, and supports custom headers, retry logic, and request/response logging.

Executes CLI-based generators (Bito, GitHub Copilot, GitHub Models) as subprocesses, capturing stdout/stderr and exit codes. The system spawns the CLI tool with appropriate arguments, waits for completion with configurable timeouts, and parses the output into structured commit messages. This enables using local CLI tools without API calls, supporting offline workflows and privacy-sensitive environments.

Formats AI-generated commit messages into Conventional Commits structure (type, scope, subject, body) and validates against specification. The formatter parses raw AI output, extracts components, ensures subject length limits (typically 50 characters), and validates type against configured definitions. Invalid messages are rejected with clear error messages, prompting regeneration or manual editing.

Implements a hierarchical configuration system with three layers: built-in defaults, global user config (~/.ai-commit/.ai-commit.json), and local project config (.ai-commit.json). Configuration covers generator selection, API credentials, prompt templates, commit type definitions, and retry parameters. The system merges layers respecting precedence (local > global > defaults) and provides CLI subcommands (set, get, unset, reset, list, edit) for runtime configuration manipulation without manual file editing.

Presents an interactive prompt allowing users to select or auto-generate the commit type (feat, fix, docs, style, refactor, perf, test, chore, ci, revert) based on the staged diff. The system validates selected types against a configurable commit type definition, ensuring compliance with Conventional Commits specification. Users can accept the AI-suggested type, manually select from a list, or provide a custom type, with validation occurring before message confirmation.

Displays the generated commit message in a formatted table showing type, scope, subject, and body, along with a preview of the staged diff. The system prompts the user to confirm, reject, or regenerate the message before executing `git commit`. Rejection allows the user to re-run the generator with the same diff, and confirmation executes the actual git commit with the approved message. This loop ensures human oversight of AI-generated content before permanent commit.

Executes the full commit message generation pipeline (diff reading, AI backend call, type selection, message formatting) but skips the final `git commit` execution. The generated message is displayed to the user without modifying the repository. This mode is useful for testing generator configurations, previewing messages before committing, and debugging AI backend responses without side effects.

Allows users to pass custom diff content via the --diff flag instead of reading from `git diff --staged`. The diff is sent to the AI generator as-is, enabling commit message generation for arbitrary code changes (e.g., diffs from other tools, historical diffs, or non-git sources). This decouples the tool from git staging, making it usable in non-standard workflows or as a library.

Allows users to customize the prompt sent to the AI generator via configuration files. Prompt templates support variable substitution (e.g., {diff}, {language}, {project_name}) enabling context-aware prompts tailored to specific projects or teams. Templates are stored in configuration and can be overridden globally or per-project, allowing teams to enforce specific commit message styles or instructions without code changes.

Implements configurable retry logic for AI backend calls with exponential backoff and sleep intervals. Users can specify retry count and sleep duration in configuration, allowing the system to gracefully handle transient API failures (rate limits, timeouts, network errors). Each retry attempt uses exponential backoff to avoid overwhelming the backend, and failures after max retries are reported to the user with error details.

Provides a self-update command that downloads the latest ai-commit PHAR binary from GitHub Releases and replaces the current executable. The mechanism checks for new versions, downloads the release artifact, verifies integrity (if checksums are available), and updates the binary in-place. This enables users to stay current with bug fixes and new features without manual re-installation.