busybeaver-policy

name	busybeaver-policy
description	Use BusyBeaver-50M as a local strict-JSON tool-policy helper underneath Hermes/Spark when debugging, recovering from tool errors, or choosing the next compact tool action.
version	0.1.0
author	GestaltLabs
license	MIT
metadata	{"hermes":{"tags":["busybeaver","tool-policy","debugging","harness","strict-json"],"category":"toolforge","related_skills":["systematic-debugging","test-driven-development","toolforge"],"config":[{"key":"busybeaver.server_url","description":"BusyBeaver OpenAI-compatible chat completions endpoint","default":"http://127.0.0.1:8767/v1/chat/completions","prompt":"BusyBeaver endpoint"},{"key":"busybeaver.model","description":"BusyBeaver exposed model name","default":"BusyBeaver-50M","prompt":"BusyBeaver model label"},{"key":"busybeaver.model_path","description":"Local checkpoint path or Hugging Face model id","default":"GestaltLabs/BusyBeaver-50M","prompt":"BusyBeaver model path"}]}}

BusyBeaver Policy

BusyBeaver is a tiny tool-policy model for Hermes harness use. It is not a main chat model and should not replace Hermes/Spark planning.

Use BusyBeaver when the next step should be a compact, analyzable tool action:

• inspect source after a failing test
• choose between read, search, test, shell, memory, clarify, or escalate
• recover after a command error
• gate destructive actions behind approval or clarification
• produce state/action traces for debugging the harness

Do not use BusyBeaver for open-ended prose, broad planning, browser reasoning, full report writing, or unsandboxed shell execution.

Runtime

The Windows install script creates a local command:

C:\Users\basbe\AppData\Local\hermes\bin\busybeaver-policy.ps1 -Goal "Inspect the failing test before editing." -Pretty

It calls the configured BusyBeaver endpoint:

http://127.0.0.1:8767/v1/chat/completions

If the endpoint is not running, start it from the adapter repo:

cd C:\Users\basbe\Desktop\AI_Research\BusyBeaver-Hermes-Adapter
.\scripts\start_busybeaver_server.ps1

Input Contract

Pass a compact state object with:

• repo_summary
• current_step
• recent_observations
• open_files
• last_tool
• last_error
• optional policy_features

The user message sent to the endpoint should be a JSON object:

{
  "goal": "Fix the failing calculator test.",
  "state": {
    "repo_summary": "Files: calculator.py, test_calculator.py",
    "current_step": 0,
    "recent_observations": ["pytest failed in test_calculator.py"],
    "open_files": [],
    "last_tool": "run_tests",
    "last_error": "AssertionError"
  },
  "available_tools": [
    {
      "name": "read_file",
      "description": "Read a workspace file.",
      "schema": { "path": "string" }
    }
  ]
}

Output Contract

BusyBeaver returns exactly one JSON object:

{"tool":"read_file","args":{"path":"calculator.py"},"confidence":0.82,"state_update":"Inspect source before patching."}

Hermes should treat the output as advisory policy. Validate the selected tool against the active Hermes tool registry before any side effect. Keep shell actions dry-run or sandboxed unless the user explicitly approves.

The production endpoint uses BusyBeaver's harness argument resolver after model generation. The model selects the tool; the resolver grounds exact file paths, commands, message fields, cron fields, and memory values from state.recent_observations and state.policy_features before returning the action.

Intended Boundary

BusyBeaver helps Hermes choose low-level tool policy. Hermes/Spark still owns user intent, decomposition, safety decisions, artifact quality, and final response synthesis.