@amolith/pi-handoff

Get to the end of the thing you're doing, then use a subagent to extract relevant bits from the current session and create a new one from them with /handoff thorough goal description. You can tell Pi to hand something off because it has a handoff tool too.

After extraction, whether triggered by you or Pi, the draft handoff goes in the new session's prompt editor and an auto-submit countdown starts that's cancelled by pressing most any key.

Install

pi install npm:@amolith/pi-handoff

If you try it and find that it doesn't behave as expected, please send me your session file, the goal, and the resulting handoff so I can debug. Contact me via the methods listed at the bottom of my website, or if you know me elsewhere, you're welcome to DM me there.

Do it yourself

Run /handoff with whatever the next session should do:

/handoff finish the release notes and publish both packages

If you omit the goal, Pi asks for one. Before the new session starts, the extension opens the generated handoff draft in the editor so you can review or tweak it. After the session switch, it starts an auto-submit countdown and pressing most any key cancels it.

To use a different model in the new session, pass -model provider/modelId:

/handoff -model neuralwatt/glm-5.1 finish the release notes

Let the agent do it

The package also registers a handoff tool. The agent should only use it when you explicitly ask for a handoff. It accepts:

• goal: the goal for the new session
• model: optional provider/modelId for the new session

The handoff happens after the current agent turn completes and still opens the editor so you can tweak before starting the new session.

Configure the extraction model

It uses the current session model by default. Interactively select a cheaper/faster model by running:

/handoff:set-extraction-model

It opens Pi's model picker and saves the selected model in $PI_CODING_AGENT_DIR/settings.json as handoffExtractionModel.

You can also set a provider/modelId with one of these:

1. Startup flag: --handoff-extraction-model provider/modelId
2. Env var: $PI_HANDOFF_EXTRACTION_MODEL
3. Settings value: handoffExtractionModel in $PI_CODING_AGENT_DIR/settings.json

If you set more than one, that list is the precedence order.

Example settings:

{
  "handoffExtractionModel": "neuralwatt/qwen3.6-35b-fast"
}

If the configured value is malformed or doesn't match a registered model, the extension falls back to the current session model.

Ask about the old session

New handoff prompts include the parent session path when available. The package registers a session_query tool so the next agent can ask focused questions about that prior session when the handoff summary isn't enough.