repofix 0.1.2

RepoFix — run any GitHub repo locally with one command; auto-detects stack, installs deps, self-heals errors

RepoFix

Clone any GitHub repo, detect the stack, install dependencies, run it, and recover from failures — without a three-hour README scavenger hunt.

Installation · Quick start · Examples · Bring your own LLM · CLI reference · Contributing · llms.txt (LLM / discovery summary)

---

Why RepoFix exists

Every developer knows the loop: star a repo → clone → guess the Node version → fight pnpm vs yarn → discover a missing .env → port already in use → give up. READMEs are written for humans, not machines — and machines are better at grinding through setup errors than you are at 11pm.

RepoFix automates that grind: one command takes you from a URL (or local path) to a running app, with rule-based fixes, optional on-device LLM, and optional cloud LLM fallbacks (Gemini, OpenAI, Anthropic) when heuristics run out. It remembers what worked so the next repo costs less.

Pain points we optimize for

You’ve been here before	What RepoFix does
“Works on my machine” in the README, not on yours	Detects stack and picks install/run commands from manifests (package.json, Makefile, Procfile, docker-compose, README hints)
Cryptic errors after install	Classifies failures, applies safe fixes (deps, ports, env), retries
Novel or project-specific failures	Escalates to local LLM and/or cloud APIs, with provider fallback
Re-cloning the same repo for every branch	Branch cache — reuses installed environments per branch
Long-running dev servers you lose track of	repofix ps / logs / stop / start / restart — process registry with log files
Global npm install -g in a README	Prompts for isolated vs system install; documents where binaries land

---

Features

• One command to run a repo — GitHub URL or local directory.
• Multi-stack stack detection — Node, Python, Go, Rust, Java/Kotlin, PHP, Ruby, Docker / Compose, and more.
• Command discovery — package.json, Makefile, Procfile, docker-compose.yml, README-driven deploy hints.
• Self-healing pipeline — Retries with fixes for common classes of failures (missing dependencies, ports, environment).
• AI layer (optional) — On-device Qwen2.5-Coder-3B via llama-cpp-python, plus Gemini / OpenAI / Anthropic with automatic fallback when configured.
• Bring your own LLM — Pin any model id each vendor supports, or point the OpenAI integration at an OpenAI-compatible base URL (local gateway, proxy, or enterprise endpoint); see Bring your own LLM.
• Fix memory — Persists successful fixes and run history for faster repeat runs.
• Safety controls — Command allow/block lists and assist mode for approval before applying fixes.
• Deploy mode hints — When a README describes both self-hosted and local dev paths, RepoFix can prompt (or use --prod / --dev to skip).
• Release binaries — Optional install from GitHub Releases when available (--binary / --source).
• Branch environment cache — repofix branches, repofix branch-clean.
• Process lifecycle — ps, logs, stop, start, restart with persisted metadata and log paths.

Default clone and config locations: ~/.repofix/repos/, ~/.repofix/config.toml, local model under ~/.repofix/models/.

---

Installation

Always required: git on your PATH.

Sometimes required: Docker when the project you run is Docker-based.

---

1. One-line install (recommended)

Installs the standalone binary from GitHub Releases. No Python is needed for RepoFix itself.

curl -sSfL https://raw.githubusercontent.com/sriramnarendran/RepoFix/main/scripts/install_binary.sh | bash

---

2. Install with pip

Requires Python 3.10+.

pip install repofix

Isolated CLI (recommended if you use pipx for tools):

pipx install repofix

Using a virtual environment:

python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install repofix

repofix --help

---

Other ways to get the binary

Download from Releases or CI artifacts; filenames are repofix-linux-x86_64, repofix-macos-arm64, repofix-macos-x86_64, repofix-windows-x86_64.exe. Build locally: ./scripts/build_binary.sh.

Linux (glibc): the x86_64 binary is built on Ubuntu 22.04 so it runs on typical glibc-based distros from that era onward. If you see GLIBC_2.xx not found, your OS is older than what that build targets — use pip install repofix (or a venv) instead, or build the binary on your own machine with ./scripts/build_binary.sh.

---

repofix not found

• pip install --user: add Python’s user script directory to your PATH (often ~/.local/bin on Linux/macOS).
• Virtualenv: activate the venv before running repofix.
• Binary: ensure /usr/local/bin (or your INSTALL_PREFIX) is on PATH.

---

Install from source (contributors)

git clone https://github.com/sriramnarendran/RepoFix.git
cd RepoFix
pip install -e ".[dev]"

---

Quick start

# Run from a GitHub URL (first run may prompt for local LLM + optional cloud API key)
repofix run https://github.com/user/repo

# Specific branch
repofix run https://github.com/user/repo --branch develop

# Local checkout
repofix run ./my-project

# Safer: confirm each fix before applying
repofix run https://github.com/user/repo --mode assist

# Inject env vars from a file
repofix run https://github.com/user/repo --env-file ./secrets.env

First-run setup

On the first repofix run, you may be offered:

1. On-device AI — Installs llama-cpp-python (prebuilt wheel when available) and downloads ~2 GB model weights to ~/.repofix/models/. You can skip and rely on cloud keys only, or disable later via repofix config.
2. Cloud API key (optional) — Gemini, OpenAI, or Anthropic for harder errors.

Manage these anytime:

repofix config show
repofix config set-key --provider gemini
repofix config set-default --local-llm    # or --no-local-llm
repofix model download                  # pre-fetch model before going offline
repofix model status

---

Examples

# Prefer production/self-hosting path from README (skip prompt)
repofix run https://github.com/org/app --prod

# Force local development path
repofix run https://github.com/org/app --dev

# Use a GitHub Release binary when detected (skip source build)
repofix run https://github.com/org/cli-tool --binary

# Always build from source
repofix run https://github.com/org/cli-tool --source

# Override port and cap retry cycles
repofix run https://github.com/org/api --port 8080 --retries 8

# No automated fixes (observe raw errors)
repofix run ./local-service --no-fix

# CI / non-interactive: skip confirmation prompts
repofix run https://github.com/org/app --auto-approve

# Override install or run commands explicitly
repofix run ./legacy-app --install "pip install -r requirements.txt" --command "uvicorn main:app --reload"

After the app is running

repofix ps                          # List processes started by RepoFix
repofix logs my-app --lines 100     # Tail log file
repofix logs my-app --follow        # Stream logs
repofix stop my-app                 # Graceful stop (SIGTERM)
repofix stop my-app --force         # SIGKILL
repofix start                       # Resume last stopped process
repofix start my-app
repofix restart my-app              # Stop + relaunch same command (no reinstall)

Branch cache

repofix branches                              # List cached branch environments
repofix branches https://github.com/user/app
repofix branch-clean https://github.com/user/app --branch feature-x
repofix branch-clean --yes                    # Clear all branch caches (destructive)

---

Execution modes

Mode	Behaviour
auto (default)	Applies fixes automatically within safety rules
assist	Prompts before applying each fix
debug	Verbose output for detection and fix decisions

---

CLI reference

repofix run

repofix run <REPO> [OPTIONS]

Arguments:
  REPO                     GitHub URL or local path

Options:
  -b, --branch TEXT        Git branch
  -m, --mode TEXT          auto | assist | debug (default: auto)
  -p, --port INT           Override listening port
  -r, --retries INT        Max fix/retry cycles (default: from config, else 5)
  -e, --env-file PATH      .env file to load
  --no-fix                 Disable automated fixes
  --auto-approve           Skip confirmation prompts (e.g. npm global install choice)
  -c, --command TEXT       Override run command
  -i, --install TEXT       Override install command
  --binary                 Prefer prebuilt release binary when available
  --source                 Always run from source (skip binary)
  --prod                   Use production/self-hosting path when README offers both
  --dev                    Use local development path when README offers both

Other commands

Command	Purpose
repofix config show	Current settings (keys shown as set/not set)
repofix config set-key [--provider gemini|openai|anthropic]	Store API key
repofix config set-default ...	Defaults: AI provider, models, OPENAI_BASE_URL, retries, local LLM, etc.
repofix history [--limit N]	Recent runs and fixes
repofix clear-memory [--yes]	Wipe fix memory + run history
repofix ps	Processes + PIDs + log paths
repofix logs <name> [--lines N] [--follow]	Show or follow logs
repofix stop <name> [--force]	Stop a process
repofix start [name]	Start stopped/crashed process (last one if name omitted)
repofix restart <name>	Stop + start same command
repofix branches [repo]	List cached branch environments
repofix branch-clean [repo] [--branch B] [--yes]	Clear branch cache
repofix model download	Install local LLM deps + download weights
repofix model status	Local model status
repofix model remove [--yes]	Delete local weights (~2 GB)

Run repofix --help and repofix <command> --help for the full Typer-generated help.

---

Configuration

• File: ~/.repofix/config.toml
• Environment variables: GEMINI_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, optional OPENAI_BASE_URL for OpenAI-compatible endpoints.

Provider preference (e.g. try Gemini then OpenAI vs a single primary) is configurable via repofix config set-default (--ai-provider, --ai-fallback / --no-ai-fallback).

Bring your own LLM

RepoFix does not lock you to one cloud model or one vendor.

Per-provider model ids — Set whichever checkpoint your org uses (as long as the provider’s API accepts that id):

repofix config set-default \
  --gemini-model gemini-2.0-flash-lite \
  --openai-model gpt-4o-mini \
  --anthropic-model claude-3-5-haiku-20241022

OpenAI-compatible endpoints — The OpenAI path uses the chat-completions JSON shape. You can aim it at a local or self-hosted server (Ollama’s OpenAI shim, LM Studio, vLLM, LiteLLM, many corporate gateways) by setting a base URL. Env var wins over config:

export OPENAI_BASE_URL="http://127.0.0.1:11434/v1"   # example: Ollama
export OPENAI_API_KEY="ollama"                        # required field; use dummy if the server ignores it
repofix config set-default --openai-model llama3.2 --ai-provider openai

Or persist in config:

repofix config set-default --openai-base-url http://127.0.0.1:11434/v1 --openai-model your-model-name

Pin the provider so traffic doesn’t bounce to Gemini/Anthropic when you intend to use only that gateway: --ai-provider openai (and disable cross-provider fallback with --no-ai-fallback if you want a single stack).

On-device default — The bundled local weights are fixed to Qwen2.5-Coder-3B (GGUF) for a consistent offline path; custom GGUF paths are not wired in config today—use cloud or OpenAI-compatible for a fully custom model choice there.

---

Supported stacks (high level)

Ecosystem	Notes
Node.js	Next.js, React, Express, Vue, Angular, NestJS, etc.
Python	FastAPI, Flask, Django, Streamlit, etc.
Go	go run / module workflows
Rust	Cargo
Java / Kotlin	Maven, Gradle
PHP	Composer, Laravel
Ruby	Rails, Sinatra
Docker	Dockerfile, docker-compose

Exact behaviour depends on project layout and detection heuristics. If something mis-detects, --install / --command overrides are the escape hatch.

---

How it works (short)

1. Resolve Git URL or local path → clone or validate.
2. Detect language, framework, and entry commands.
3. Resolve environment (e.g. .env.example hints, missing vars).
4. Run install → build → run with live log streaming.
5. On failure: classify → rule-based fix → retry up to limit.
6. If still stuck: local and/or cloud AI suggests next steps (JSON-structured, safety-filtered).
7. Record outcome in fix memory; register long-lived processes for ps / logs / lifecycle commands.

---

Security & privacy

• RepoFix executes shell commands in your checked-out repos. Use assist mode or --no-fix when you don’t want automatic remediation.
• Cloud AI sends error context and project snippets to the provider you configure — use local-only mode if that’s unacceptable for your org.
• Review repofix fixing/safety.py and the allow/block behaviour before running on untrusted codebases.
• sudo global npm installs are never silent; they require explicit user choice.

---

Development

pip install -e ".[dev]"
pytest
ruff check src tests

---

Contributing

We welcome issues and PRs. A few norms:

1. Open an issue first for large features (detection rules, new stack support, AI behaviour) so we align on design.
2. Small, focused PRs — one logical change per branch; match existing style (ruff, type hints where used).
3. Tests — Add or extend tests under tests/ for detection, fixing, safety, or CLI behaviour you change.
4. No drive-by refactors — Keep diffs tight; unrelated formatting churn makes review harder.
5. Document user-visible flags — If you add CLI options, update this README and Typer help strings.

License: MIT — see LICENSE.

---

Support

• Bug reports & feature requests: GitHub Issues
• Questions & show-and-tell: GitHub Discussions

When reporting bugs, include: OS, Python version, RepoFix version, the repo URL or stack, full --mode debug output (redact secrets), and whether local LLM / cloud AI was enabled.

---

Roadmap & status

Project status is alpha (Development Status :: 3 - Alpha on PyPI). Expect breaking CLI or config changes until v1.0. Priorities include broader stack coverage, sharper safety defaults, and more regression tests for real-world READMEs.

---

RepoFix — paste the URL. Run the code.

MIT License · Built for developers who’d rather ship than configure.