pi-goal-hermes

Goal-driven autonomous continuation extension for Pi. Set a goal and let the agent work autonomously until done, evaluated by a secondary LLM judge.

Background

This extension is inspired by and ported from Hermes's goal continuation architecture. The core idea -- an LLM judge loop that evaluates progress and decides whether to continue or stop -- is adapted from Hermes's goal system and re-implemented as a native Pi extension using Pi's extension API (custom entries, message renderers, slash commands, and the agent lifecycle hooks).

Install

pi install npm:@ricoyudog/pi-goal-hermes

Or project-local:

pi install -l npm:@ricoyudog/pi-goal-hermes

How It Works

1. You set a goal via /goal <description>
2. The agent works toward the goal autonomously
3. After each turn, a secondary LLM (judge) evaluates whether the goal is achieved
4. If not done, the agent automatically continues with a follow-up prompt
5. The loop stops when: goal achieved, turn budget exhausted, or manually paused

Commands

/goal

Command	Description
/goal <text>	Set a new goal (starts autonomous loop)
/goal status	Show current goal status and progress
/goal pause	Pause the goal loop
/goal stop	Stop the goal loop (alias for pause)
/goal resume	Resume a paused goal (resets turn counter)
/goal done	Manually mark goal as achieved
/goal clear	Clear the goal entirely

/subgoal

Add additional acceptance criteria that the judge evaluates alongside the main goal.

Command	Description
/subgoal <text>	Add a subgoal criterion
/subgoal list	List all subgoals
/subgoal remove <n>	Remove subgoal by index
/subgoal clear	Clear all subgoals

Status Bar

The extension shows goal status in the Pi footer:

• ● Pursuing goal (active)
• ⏸ Goal paused
• ✓ Goal achieved

Configuration

• Default max turns: 20 (the agent gets 20 turns before the goal auto-pauses with "budget exhausted")
• Turn counter resets on /goal resume

Pause Conditions

The goal automatically pauses when:

• Turn budget exhausted (20 turns by default)
• Judge output unparseable 3 times in a row
• Agent response errors or aborts
• Session reloads
• User interrupts (Ctrl+C)

Architecture

• Judge: Uses a fast model (e.g., Haiku) as a secondary evaluator
• State persistence: Goal state is stored in the session via custom entries (survives compaction)
• Continuation: Sends follow-up prompts to keep the agent working
• Event rendering: Custom message renderers for goal events in the conversation

License

MIT