repopilot-agent 0.2.1

Local-first AI code agent in your terminal. Inspired by Claude Code and Codex CLI.

RepoPilot

Local-first AI code agent in your terminal.

English | 中文

What is RepoPilot?

RepoPilot is a CLI coding agent inspired by Claude Code and Codex CLI. Navigate to any project, run repopilot, and describe what you want done in natural language — RepoPilot reads, searches, edits, runs tests, and fixes bugs autonomously in a sandboxed environment.

$ cd your-project
$ repopilot

────────────────────────────── RepoPilot ──────────────────────────────
  Directory: /your-project
  Model:     doubao-seed-evolving
  Sandbox:   local
  Approval:  auto

Type /help for commands, /exit to quit.

repopilot> fix the failing test in test_auth.py
> read_file(path=test_auth.py)
> bash(command=python -m pytest test_auth.py -v)
> edit_file(path=auth.py, ...)
> bash(command=python -m pytest test_auth.py -v)

  All tests pass. Fixed the token validation bug in auth.py line 42.

Features

• Claude Code / Codex CLI style REPL — cd project && repopilot, start chatting immediately
• Pure ReAct agent loop — single model does all thinking, no multi-agent overhead
• Streaming with waiting feedback — shows Thinking... immediately, then renders assistant text incrementally
• Session diff and undo — inspect changes with /diff and roll them back step-by-step with /undo
• Interactive approvals — confirm writes/commands with y/n/a/d; dangerous operations remain hard-denied
• Tool timing — each tool execution displays its elapsed time
• Persistent multi-turn conversation with automatic context compaction
• Layered memory system — global + project REPOPILOT.md files (like CLAUDE.md)
• Cross-session resume — /resume to continue where you left off
• Docker sandbox with CPU/memory limits and optional network isolation
• 4 approval modes: auto / confirm / edit-only / deny (default: confirm — you approve writes/executions)
• Dangerous command blacklist (path traversal, rm -rf /, curl|sh, force push, credential theft)
• 10 built-in tools: read/write/edit/grep/glob/list_dir/bash/run_python/repo_tree/finish
• tree-sitter repo map — code structure overview without reading every file
• Circuit breaker + exponential backoff for reliable LLM calls
• Cross-platform — Windows / Linux / macOS with automatic Unix→Windows command translation
• Any OpenAI-compatible LLM — use your own API key (Doubao, DeepSeek, OpenAI, vLLM, local models, etc.)
• No RAG / no vector database — deterministic grep/glob/tree-sitter retrieval is faster and more accurate for code

Installation

pip install repopilot-agent

Or install the latest version directly from GitHub:

pip install git+https://github.com/ZhangYang2297/repopilot.git

For an isolated install (recommended for CLI tools):

pipx install repopilot-agent

Requirements: Python 3.10+

Windows note: if pip install fails with "Cargo, the Rust package manager, is not installed", run pip install repopilot-agent --only-binary :all: to force pre-built wheels, or use pipx.

First Run

On first run you will be prompted for your LLM configuration:

1. Model name (e.g. openai/doubao-seed-evolving, openai/gpt-4o, openai/deepseek-chat)
2. API key (sk-...)
3. Base URL (for providers other than OpenAI, e.g. https://ark.cn-beijing.volces.com/api/v3)

You can also configure via environment variables:

export REPOPILOT_MODEL=openai/doubao-seed-evolving
export REPOPILOT_API_KEY=sk-your-key
export REPOPILOT_BASE_URL=https://ark.cn-beijing.volces.com/api/v3

Usage

Interactive Mode (Recommended)

cd your-project
repopilot                          # current directory, local sandbox, confirm approval
repopilot -r ../other-proj         # specify a different project directory
repopilot --sandbox docker         # run inside a Docker container
repopilot --approval-mode auto     # skip confirmations (trust the agent)
repopilot -m openai/gpt-4o         # override model

One-shot Task Mode

repopilot chat "fix the bug in auth.py"
repopilot chat "add --verbose flag to cli.py" -r ./myproj

Slash Commands

Command	Description
/exit, /quit	Exit (Ctrl+C / Ctrl+D also supported)
/help	Show help
/model [name]	Show or switch model
/approval [mode]	Switch approval mode
/compact	Trigger context compaction
/clear	Start fresh conversation
/cd [path]	Switch working directory
/memory [note]	Show or add memory notes
/resume [id]	Resume a previous session
/sessions	List recent sessions
/cost	Show token usage/cost
/status	Show current configuration
/diff	Show all file changes made during this session
/undo	Revert the most recent file change

Project Memory (REPOPILOT.md)

Create a REPOPILOT.md in your project root to give RepoPilot persistent instructions:

# Project Memory

## Build/Test
- Test: python -m pytest tests/ -v
- Lint: ruff check .

## Conventions
- Use type hints on all functions
- Never modify files in migrations/

Global memory lives at ~/.repopilot/REPOPILOT.md and applies across all projects.

Configuration

Config file: ~/.repopilot/config.toml

[core]
model = "openai/doubao-seed-evolving"
api_key = "sk-..."
base_url = "https://ark.cn-beijing.volces.com/api/v3"
sandbox_type = "local"
approval_mode = "auto"
max_steps = 200
budget_tokens = 500000
tool_timeout = 120

Manage config via CLI:

repopilot config show
repopilot config set model openai/gpt-4o
repopilot config init  # re-run setup wizard
repopilot models       # list recommended models

Architecture

┌─────────────────────────────────┐
│  CLI (Typer + Rich)    REPL     │
├─────────────────────────────────┤
│  Shared AgentLoopCore (ReAct)   │
├─────────────────────────────────┤
│  Context Manager  L0-L5 memory  │
├─────────────────────────────────┤
│  Tool Registry + Permission     │
├─────────────────────────────────┤
│  Sandbox (Local / Docker)       │
├─────────────────────────────────┤
│  LLM Service (LiteLLM)          │
└─────────────────────────────────┘

Both one-shot chat tasks and the interactive REPL use the same UI-independent AgentLoopCore for response normalization and tool execution. The REPL adds pre-first-event Thinking... feedback, streaming rendering, interactive approval, persistent context, /diff, and /undo. The indicator only communicates that a request is active; it does not expose or fabricate hidden model reasoning.

Development Status

• Current package version: v0.2.1
• Phase 2A is complete: streaming, first-event waiting feedback, TTFT/total-duration metrics, tool timing, interactive approval, /diff, and /undo
• 492 project tests pass (excluding fixture and E2E data projects from collection)

Built-in Tools

Tool	Description
read_file	Read a file (with optional line range and offset limit)
write_file	Write content to a file (creates or overwrites)
edit_file	Find-and-replace edit (string replacement)
grep_search	Search file contents with regex
glob	Find files by glob pattern
list_dir	List directory contents
repo_tree	Show tree-sitter generated repository map
bash	Execute a shell command (sandboxed)
run_python	Execute Python code in an isolated temp file
finish	Signal task completion and return to user

Supported LLM Providers

Any OpenAI-compatible endpoint works out of the box via LiteLLM:

• Volcengine ARK (Doubao) — recommended, tested extensively
• OpenAI (GPT-4o, GPT-4, o1, etc.)
• DeepSeek (deepseek-chat, deepseek-reasoner)
• Alibaba Qwen (qwen2.5-coder series)
• Zhipu GLM (glm-4, glm-5 series)
• Local models via vLLM / Ollama / llama.cpp (any OpenAI-compatible server)
• Anthropic Claude (via LiteLLM)

License

MIT — see LICENSE for details.

Acknowledgements

Built after studying Claude Code (Anthropic), Codex CLI (OpenAI), and the SWE-bench / SWE-agent research.