pi-auto-continue

A pi extension that automatically sends a "continue" message every time the agent finishes a turn, so you don't have to babysit long-running work.

Why

The motivating use case is autonomous overnight research loops, in the spirit of Andrej Karpathy's autoresearch: you give an agent a program.md-style brief, walk away, and let it iterate — edit code, run experiments, evaluate, keep or discard, repeat — until morning.

The problem with doing that on a local coding agent is that the agent politely stops after every turn and waits for you to type "continue." If you're asleep, it sits there doing nothing. Ralph-style shell loops around the CLI work but are fiddly to set up and don't share state with the running session.

pi-auto-continue solves that at the extension layer: pi itself pushes a user message back into the loop as soon as the agent stops, bounded by a safety counter. Set your continue message, flip it on, and leave. The next morning you have a log of iterations instead of an idle prompt.

What it does

Whenever the agent loop ends, this extension calls pi.sendUserMessage(...) to push a user message back into the loop, just as if you had typed it. A hard cap of 100 iterations prevents runaway loops. Typing anything yourself resets the counter. Hitting Escape to abort a turn disables auto-continue — if you want out, you want out.

Install

pi install npm:@latent-variable/pi-auto-continue

Or directly from GitHub:

pi install git:github.com/latent-variable/pi-auto-continue

Or from a local clone:

pi install /path/to/pi-auto-continue

Then /reload inside pi if it's already running.

Usage

Four commands, all discoverable via /auto tab-completion:

Command	What it does
/auto-continue-on	Enable (uses the current continue text, default "continue")
/auto-continue-off	Disable
/auto-continue-set <text>	Set the continue message AND enable
/auto-continue-max <n>	Set the iteration cap (default: 100)

When enabled, the footer shows auto-continue N/<max>.

Example

/auto-continue-set keep iterating on train.py until val_bpb improves

Flips auto-continue on with that as the message. Walk away. Come back and /auto-continue-off (or hit Escape mid-turn) to stop.

How it works

• Subscribes to agent_end. Every time the agent loop ends, it calls pi.sendUserMessage(text), which re-enters the agent loop as if you had typed it.
• Subscribes to input and resets the counter whenever event.source === "interactive" — so your own typing doesn't count against the iteration cap.
• Checks ctx.signal?.aborted in agent_end. If the turn was aborted (Escape), it disables itself instead of firing another continue.

Notes / gotchas

• agent_end fires on normal completion and on errors, so auto-continue will keep trying even after a failed turn. Usually what you want in a research loop, but a persistently broken turn will burn your 100-iteration cap fast.
• The counter and enabled flag live in memory per pi process. Restarting pi, /new, or /fork clears them — re-enable with /auto-continue-set <your text> after.
• The iteration cap defaults to 100 and can be changed at runtime with /auto-continue-max <n>. It resets to 100 on restart.