Multi-Agent Software Engineering System

English | 简体中文

A local-first coding-agent workbench for turning a coding request into a guarded software-engineering loop: understand the repo, plan the work, draft code, review the result, preview the diff, run tests, and only then apply changes.

This project is not about adding more agents for show. The main goal is to make multi-agent coding behave more like a careful engineering workflow.

What It Does

• Builds task context from project docs and source files.
• Uses planner, coder, reviewer, integrator, and knowledge roles when they are useful.
• Keeps generated drafts in a separate workspace before touching real source files.
• Reviews drafts for path scope, missing sections, risky changes, secrets, and test instructions.
• Supports dry-run diff preview before applying changes.
• Runs tests before local commits when changes are applied.
• Keeps long-term knowledge writes behind explicit user confirmation.
• Uses a local mock provider by default, so a fresh run does not call an external API.

Current Status

Current version: v0.17.15

Current direction: Stable Coding Loop -> Hierarchical Context -> Evaluation.

The project is still an experimental prototype, but the main loop is intentionally conservative: draft first, review, preview diff, test, then apply.

Quick Start

Requires Python 3.11 or newer.

git clone https://github.com/shinji00222/code-agent-collab.git
cd code-agent-collab
$env:PYTHONPATH="src"
python -m code_agent_collab.cli provider
python -m code_agent_collab.cli coding-loop "add a tiny test"

The command above runs in dry-run mode and previews the proposed change. To apply a reviewed draft, run:

python -m code_agent_collab.cli coding-loop "add a tiny test" --apply

By default, the project uses the local mock provider. Configure DeepSeek, OpenAI, or an OpenAI-compatible provider only when you want real model calls.

Main Commands

$env:PYTHONPATH="src"
python -m code_agent_collab.cli provider
python -m code_agent_collab.cli start "task goal"
python -m code_agent_collab.cli run-adaptive "task goal"
python -m code_agent_collab.cli plans
python -m code_agent_collab.cli approve "task id or keyword"
python -m code_agent_collab.cli apply-draft "task id or keyword"
python -m code_agent_collab.cli coding-loop "task goal" [--apply]
python -m code_agent_collab.webui
python -m code_agent_collab.desktop

Safety Model

The project separates AI output from real project files by default:

• generated drafts go to dev-vault/projects;
• knowledge candidates go to dev-vault/pending;
• real source changes require an explicit apply step;
• public config examples are safe to commit, while real local config stays ignored;
• secrets should be passed through environment variables, never committed.

Project Layout

src/code_agent_collab/   source code
tests/                   automated tests
product-docs/            product notes and design docs
dev-vault/               generated drafts and isolated project knowledge
logs/                    local run logs and progress snapshots
scripts/                 packaging and setup scripts

Documentation

• 中文 README
• Version plan
• Changelog
• Maintenance guide
• Project skills
• Research route
• Code simplification plan

Tests

$env:PYTHONPATH="src"
python scripts/run-tests.py

You can also run the standard unittest discovery command:

$env:PYTHONPATH="src"
python -m unittest discover -s tests