@tifan/pi-recap

Re-enter a session without rereading the transcript.

pi-recap shows a one-line recap on demand or after you have been away. It starts with why you opened the session, then adds the current state, important decisions, relevant files or commands, and the likely next action.

Install

pi install npm:@tifan/pi-recap

This package requires Pi 0.84.2 or newer.

How it works

• /recap generates a fresh, goal-first recap and shows it above the editor.
• After the agent settles, pi-recap waits 5 minutes. If you stay idle, it generates one automatic recap.
• On resume, pi-recap shows the saved recap if it is current. If it is stale or missing, it generates a fresh recap.
• The recap clears when you send a non-/recap message.

Recaps use pi's current session context, so they follow the active branch and respect compaction. They do not scrape the full session file or terminal history. The latest recap is stored outside LLM context.

A good recap should answer "what was I trying to do here?". For example:

Deciding whether pi-inline-skills should switch from $skill to /skill. Recommendation is / only with commands winning; next decide whether leading /skill should expand.

Commands

• /recap: Generate and show a fresh recap.
• /recap status: Show the selected model, active model, recap freshness, and whether the recap is visible. A recap is current when it still matches the latest session state; otherwise it is stale.
• /recap config: Choose a recap model.
• /recap help: List recap commands.

Configuration

Out of the box, pi-recap uses this default model: openai-codex/gpt-5.6-luna.

Run /recap config to choose a different model.

After you choose a model, pi-recap uses only that model. Choose Use default in /recap config to return to the default.

You can also edit $PI_CODING_AGENT_DIR/extensions/pi-recap.json manually:

{
  "model": "openai-codex/gpt-5.6-luna"
}

Release notes

See CHANGELOG.md

License

MIT