FAQ
General
What operating systems does Peen-SWE support?
Peen-SWE runs on Linux, macOS, and Windows. It works anywhere a bash shell is available or inside container runtimes (Docker, Podman, Singularity/Apptainer, Bubblewrap).
How does Peen-SWE differ from older agents like SWE-agent?
Peen-SWE strips away bloated scaffold abstractions in favor of radical simplicity:
- Stateless execution: Actions run via clean
subprocess.runcalls rather than keeping an interactive shell process open. This prevents state contamination, pseudo-terminal hangs, and shell crashes. - Native Bash + MCP tools: Instead of proprietary internal tools with custom prompt engines, Peen-SWE pairs bare shell commands with standard Model Context Protocol (MCP) servers.
- Linear history: Context is an append-only log of messages with zero magic prompt rewriting, making debugging predictable and straightforward.
Where is the global configuration stored?
Global configuration resides in the .env file within your user config directory (the full path is printed when you run peen). You can also inspect or edit it using peen-extra config edit.
Models
Which models are supported?
Peen-SWE supports any model available through LiteLLM, OpenRouter, Portkey, or direct local providers (Ollama, vLLM, LM Studio).
How do I configure API keys?
Set them in your shell environment:
export ANTHROPIC_API_KEY="sk-ant-..."
Or persist them across all sessions with the config helper:
peen-extra config set ANTHROPIC_API_KEY "sk-ant-..."
How do I set the default model?
Set the PEEN_MODEL_NAME environment variable:
peen-extra config set PEEN_MODEL_NAME "anthropic/claude-3-5-sonnet-20241022"
Architecture
Why avoid a persistent background shell session?
Traditional coding agents keep a persistent pseudo-terminal session open across all steps. This causes several critical problems: 1. Termination ambiguity: Determining when an arbitrary shell command has finished requires fragile heuristics (watching prompt escapes, monitoring PIDs, or polling buffers). 2. Shell corruption: Misbehaving commands can leave the shell in raw mode, corrupt environment variables, or orphan background child processes. 3. Session death: If a command triggers a shell segfault or kills the parent terminal, the entire agent run crashes.
Peen-SWE executes each command independently using standard subprocess.run (or docker exec in containerized environments). The exit code and exact output are captured deterministically with zero heuristics.