StrawPot OpenHands CLI

A Go wrapper that translates StrawPot protocol arguments into OpenHands CLI flags. It acts as a pure translation layer — process management, sessions, and infrastructure are handled by StrawPot core.

Prerequisites

• OpenHands CLI (pip install openhands-ai)
• An LLM API key for your chosen model provider

Installation

curl -fsSL https://raw.githubusercontent.com/strawpot/strawpot_openhands_cli/main/strawpot_openhands/install.sh | sh

This downloads a pre-built binary for your platform (macOS/Linux, amd64/arm64) to /usr/local/bin. Override the install directory with INSTALL_DIR:

INSTALL_DIR=~/.local/bin curl -fsSL ... | sh

Usage

The wrapper exposes two subcommands:

setup

Runs openhands login to authenticate.

strawpot_openhands setup

build

Translates StrawPot protocol flags into an OpenHands CLI command and outputs it as JSON.

strawpot_openhands build \
  --agent-workspace-dir /path/to/workspace \
  --working-dir /path/to/project \
  --task "fix the bug" \
  --config '{"model":"gpt-4o"}'

Output:

{
  "cmd": ["openhands", "--headless", "--json", "-t", "fix the bug", "--override-with-envs"],
  "cwd": "/path/to/project",
  "env": {"LLM_MODEL": "gpt-4o"}
}

Build flags

Flag	Required	Description
--agent-workspace-dir	Yes	Workspace directory for instructions and symlinks
--working-dir	No	Working directory for the command (cwd in output)
--task	No	Task prompt (passed as -t, or combined into instructions file when prompts are present)
--config	No	JSON config object (default: {})
--role-prompt	No	Role prompt text (written to instructions.md)
--memory-prompt	No	Memory/context prompt (appended to instructions.md)
--skills-dir	No	Directory with skill subdirectories (symlinked to skills/)
--roles-dir	No	Directory with role subdirectories (repeatable, symlinked to roles/)
--agent-id	No	Agent identifier

Configuration

Config JSON

Pass via --config:

Key	Type	Default	Description
model	string	(none)	Model to use (passed via LLM_MODEL env var with --override-with-envs)

Environment variables

Variable	Description
LLM_API_KEY	LLM API key for the configured model provider

Notes

• OpenHands headless mode always runs in auto-approve mode — there is no opt-out.
• Since OpenHands CLI does not support a separate --system-prompt flag, role/memory prompts and task are combined into a single instructions.md file and passed via -f. When only a task is provided (no prompts), it is passed directly via -t.
• Model selection uses the LLM_MODEL environment variable (returned in the env field of the JSON output) rather than a CLI flag.

Development

cd openhands/wrapper
go test -v ./...

Releases are built with GoReleaser and published automatically via GitHub Actions.

License

See LICENSE for details.