jupyter-ai-acp-hub 0.1.2

Jupyter AI ACP personas with JupyterHub-safe session handling.

jupyter-ai-acp-hub

Jupyter AI ACP personas that treat ACP sessions as ephemeral runtime state.

The visible Jupyter chat history remains in .chat files, but live ACP session IDs are not persisted into chat metadata and are not loaded after restart. This avoids stale session/load failures in JupyterHub pods where chat files survive but agent subprocesses do not.

Personas

• @HubClaude: wraps claude-agent-acp
• @HubOpenCode: wraps opencode acp

These personas are intentionally separate from upstream @Claude and @OpenCode so they can be tested side by side.

Install

For a running JupyterHub notebook pod, avoid pulling Jupyter server dependencies into ~/.local:

python -m pip install --user --no-deps jupyter-ai-acp-hub

For a local development checkout:

python -m pip install -e ".[test]"

Restart the Jupyter server process after installation so Jupyter AI reloads the entry points.

jupyter-ai-acp-hub intentionally has no required runtime dependencies. It is meant to use the Jupyter AI ACP packages already present in the notebook image. If you are installing into a clean environment, use:

python -m pip install "jupyter-ai-acp-hub[runtime]"

The underlying agents still need to be available in the notebook image:

• claude-agent-acp for @HubClaude
• opencode for @HubOpenCode

Behavior

• always creates a fresh ACP session for each live persona instance
• never writes metadata.acp_session_ids into .chat
• never loads metadata.acp_session_ids from .chat
• returns the live session ID from the current session future
• injects recent chat history once after creating a fresh session, using the bounded recovery context from jupyter-ai-acp-client

Verify

python - <<'PY'
import importlib.metadata as metadata

for ep in metadata.entry_points(group="jupyter_ai.personas"):
    if ep.name.startswith("hub-"):
        print(ep.name, ep.value)
PY