Minimal Agentic Coding Runner

A minimal coding runner that uses a single tool: bash.

It runs an LLM-driven loop with strict phases and can be interrupted by humans at any time.

• Single execution tool: bash(cmd)
• Strict agent protocol: plan -> act -> verify -> chat -> done
• Mandatory TASK.md updates per step
• Step logs in logs/run_YYYYMMDD_HHMMSS.log
• Works with OpenAI-compatible APIs (OpenAI / Qwen DashScope)
• Human-in-the-loop controls:
  • Ctrl+C to pause and enter control mode
  • Post-verify gate: only finishes after user confirms done
  • Bare text in interactive prompts = feedback + resume

• Python 3.10+
• OPENAI_API_KEY or API_KEY
• Optional: rg (ripgrep), used for faster file inventory

python3 vibe_runner.py "创建一个 hello.py 并写入打印语句"

export PROVIDER=openai
export BASE_URL="https://api.openai.com/v1"
export API_KEY="<your_api_key>"
export MODEL="gpt-4.1-mini"
python3 vibe_runner.py "your goal"

export PROVIDER=qwen
export BASE_URL="https://dashscope.aliyuncs.com/compatible-mode/v1"
export API_KEY="<your_dashscope_key>"
export MODEL="qwen3-max"
python3 vibe_runner.py "your goal"

• resume / r: continue agent loop
• feedback <text>: add feedback to context
• run <bash>: run one manual command and inject output back to context
• quit: exit runner (with confirmation)
• bare text: treated as feedback and auto-resume

• done: finish task (with confirmation)
• resume: continue improving
• feedback <text> / run <bash> / quit
• bare text: treated as feedback and auto-resume

• used when agent chooses phase=chat for consulting/explaining tasks
• supports resume, feedback <text>, quit
• bare text: treated as feedback and auto-resume

• Denylist blocks risky commands (rm -rf, mkfs, dd, shutdown, reboot, curl, wget, nc, ssh)
• One command per agent step
• Command timeout controlled by CMD_TIMEOUT (default 20s)
• Max steps controlled by MAX_STEPS (default 30)

• vibe_runner.py: runner implementation
• TASK.md: per-task working memory (overwritten each step)
• logs/run_*.log: execution audit logs