omp packages

omp-auto-loop0.4.13

General-purpose auto agent loop plugin for omp.

kunweiz92 Published 6 months ago MIT ~210 downloads/mo

Built for omp

The author tagged this package omp-plugin.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

omp-auto-loop

General-purpose agent loop extension for the Oh-My-Pi coding agent. When you send a normal (non-slash) message it becomes the loop goal and the extension runs repeated agent turns until the goal is finished or you stop the loop.

Install

omp plugin install omp-auto-loop

Overview

  • Starts a loop when the user sends a normal message (one that does NOT start with /). That message becomes the loop goal.
  • Injects a short loop-focused system prompt so the agent knows the current iteration and goal.
  • Provides a tool named loop_control which the agent must call to indicate progress: use status: "next" to continue, or status: "done" to finish.
  • Registers a loop-stop command and a Ctrl+Shift+S shortcut to stop the active loop from the UI.

Usage

  • Start a loop: send a normal message like Refactor all test files to use the new assertion API (do NOT prefix with /). The extension will begin iterating on that goal.
  • One-off message: prefix with /once to send a single, non-looping turn (e.g. /once Quick status check).
  • Stop loop: run the loop-stop command or press Ctrl+Shift+S to abort the active loop immediately.

Tool: loop_control

Call this tool from the assistant to signal what happened during the iteration. The tool is only meaningful while a loop is active.

Parameters:

  • status — string: either "next" or "done".
  • summary — string: brief summary of what was accomplished this iteration.
  • reason — optional string: a reason explaining why the goal is complete (used with "done").

Behavior notes:

  • When the model calls loop_control with status: "next", the extension advances the iteration count and schedules the next iteration (it sends a steer message with the updated loop prompt).
  • When the model calls loop_control with status: "done", the extension asks for confirmation (it sets a confirmingDone flag and returns a prompt asking the model to either call loop_control again or finish the response). The loop is finalized if the model finishes the response (skips calling loop_control again) or explicitly confirms by calling loop_control with done again.

Fallbacks:

  • If the agent ends its turn without calling loop_control and without scheduling the next step, the extension increments the iteration count and sends a steer message reminding the agent to call loop_control (so work won't silently stop).

Examples

  • Start by sending a goal (normal message):
Refactor all test files to use the new assertion API
  • Example tool call (assistant -> tool):
{
	"toolName": "loop_control",
	"params": { "status": "next", "summary": "Refactored 3 files; more work remains" }
}
  • Marking completion (assistant -> tool):
{
	"toolName": "loop_control",
	"params": { "status": "done", "summary": "All tests updated and passing" }
}

If the extension asks for confirmation after a done call, finish the response (or call loop_control with done again) to finalize the loop.

UI

  • While a loop is active the status bar and a small widget show the current iteration and provide the Ctrl+Shift+S stop shortcut.

Notes

  • Package name: omp-auto-loop (see package.json).
  • Peer-dep: this extension expects to run within @oh-my-pi/pi-coding-agent.