v0.4.2

What's changed

Container build refactor for faster, rate-limit-free local builds:

• Published GHCR base image. The slow-changing runtime layers (Python + Node.js + Claude Code CLI + uv + git + non-root user) now live in ghcr.io/skillberry-ai/runspace-agent-base (multi-arch: amd64 + arm64). The runtime image starts FROM that base and only layers Python deps + the package source on top, so runspace-srv builds are fast and avoid apt/npm registry rate limits.
• uv replaces pip for the dependency install layer.
• git added to the image (remote skills are cloned via npx skills add at container startup).
• New scripts/publish-base-image.sh to (re)publish the base; README documents it.

No public API changes.

v0.4.1

Container mode can now reach host services.

• fix(container): add extra_hosts host.docker.internal=host-gateway to the container run so an agent in container mode can reach MCP servers (and other services) running on the host — e.g. a Skillberry Store control MCP at http://host.docker.internal:<port>/control_sse. Required on Linux; harmless on Docker Desktop. (#13)

v0.4.0

Breaking change

preinstalled_skills has been removed from RunspaceSession, the server RunRequest, and prepare_skills. Skills are now provided only via:

• remote_skills — installed at run setup via npx skills add (GitHub URLs, owner/repo, or repo subpaths), or
• skills_dir — a user folder of skills.

Bundled skills (skill-creator, mcp-builder) are loaded via remote_skills using their GitHub sources, e.g. https://github.com/anthropics/skills/tree/main/skills/skill-creator.

Callers using preinstalled_skills must migrate to remote_skills or skills_dir.

(#12)

v0.3.1

remote_skills — install skills from remote repos per run

Adds a third skill source alongside preinstalled_skills and skills_dir: remote_skills, a per-run list of sources installed at run setup via the skills CLI (owner/repo slugs, GitHub URLs, repo subpaths).

RunspaceSession(..., remote_skills=["vercel-labs/agent-skills"])

• Agents declare npx_agent_name (the skills CLI -a value; claude-code for Claude Code).
• Works in local and container modes — container installs run inside the container using its own Node, so the host never needs npx.
• A source that fails to install fails the run; npx scaffolding is cleaned up afterward.

This is the first published cut of the feature (0.3.0 was held back and never released due to a wrong skills-CLI agent name). Closes #7, #9.

v0.2.0

First correctly-versioned PyPI release (earlier tags all published as 0.1.0 due to a static version in the build).

Highlights since 0.1.0:

• Removed persistent container mode and the container_mode option; container runs always use a fresh, auto-removed container (breaking change).
• Session workspaces now live under <home>/sessions/<id>; RUNSPACE_DATA_DIR sets the home (defaults to {tmp}/runspace).
• Server port configurable via RUNSPACE_PORT; python -m runspace_agent works as an alias for runspace-srv.
• Session UI shows the execution mode; agent-agnostic credential docs (pass creds via agent_settings.env).
• Shared build_claude_env helper for examples/tests.
• Publish workflow now derives the version from the release tag.

v0.1.4

v0.1.4

v0.1.3

v0.1.3

v0.1.2

v0.1.2

v0.1.0

Initial Release