pi-prompt-queue

Run many prompts unattended against a single local model (e.g. llama-swap). Type /queue <prompt> in any pi session — from any project folder, any terminal — and prompts run one at a time, globally, in FIFO order. Each prompt runs in the session that queued it (that session's context, project dir, and model).

Install

pi install npm:pi-prompt-queue

or from git:

pi install git:github.com/e-mend/pi-prompt-queue

Restart pi (or start a new session) to load it.

Usage

/queue [model:] <prompt>    enqueue a prompt; runs when it reaches the head
/queue                      show the queue
/queue-cancel <id|mine|all> cancel items

• model: is optional — unc-q6k: fix the parser (or llama-swap/unc-q6k: …) switches that one prompt to another model before running it.
• The session stays interactive while queued items wait; a status line shows your queue position, and you get a notification when each item starts and finishes.

How it works

• All pi sessions on the machine share one queue file: ~/.pi/agent/prompt-queue/ (override with PI_QUEUE_DIR).
• Exactly one item runs at a time; the rest wait in FIFO order by enqueue time.
• Self-healing: if a session is closed, its pending items are canceled and the queue keeps going. A running item whose owner died is canceled immediately; a stalled owner is canceled after 90 s without heartbeat.
• If the model errors mid-run (e.g. llama-swap reloading the model), the item is retried up to 3 times with a 15 s cooldown.

Notes

• Closing a terminal cancels its pending items — re-queue to retry.
• No runtime dependencies; state is plain JSON on disk.