Skip to content

rbudnar/harness-engineering-bootstrap

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Harness Engineering Bootstrap

A practical bootstrap template for creating a self-maintaining, token-efficient agent harness in a software repository.

The goal is not to add more documentation. The goal is to help coding agents load the right context at the right time, keep always-on instructions small, enforce drift mechanically where possible, and grow the harness only when repeated misses or real dependencies justify it.

Contents

  • Template - the current bootstrap template.
  • Dogfooding guide - how this repo keeps the template from becoming a fat harness.
  • Template fitness check - local and CI bloat guard for template changes.
  • Changelog - version history and major design changes.
  • References - source material and related work used while developing the template.

What This Template Emphasizes

  • Thin cross-agent entry points such as AGENTS.md.
  • Thin Claude, Gemini, and Copilot adapters that route back to the same source of truth.
  • Task-routed docs instead of broad context loading.
  • Decision memory, data contracts, and repo contracts.
  • Deterministic quality gates and harness validation.
  • Minimal local metrics first; PR metrics and scheduled reporting only when triggered.
  • Guide/sensor and computational/inferential control taxonomy.
  • Programmatic state surfaces before raw context dumps.
  • Optional URL-fetchable context maps for remote agents.
  • A harnessify path for turning repeated agent friction into the smallest durable control.
  • Replaceable, thin scaffolding with model/tool-upgrade review triggers.
  • Triggered agent-runtime safety docs for tools, credentials, approvals, audits, and autonomous jobs.
  • Triggered behavioral-drift sensors and evidence packs for long-running or source-heavy agent work.
  • Explicit self-correction rules so repeated agent mistakes become repo-owned harness updates or visible markers.
  • Semantic-scope checks so broad migrations and user-visible control changes verify their payoff before scaling.
  • Layer mapping, progressive disclosure, optional OTel/MCP guidance, and decision-observability checks for mature harnesses.

Dogfooding This Template

This repo uses a deliberately small harness to test the template against its own principles. AGENTS.md stays thin, Claude/Gemini/Copilot adapters point back to it, docs/dogfooding.md defines the admission test and automated PR loop for daily suggestions, and the fitness check rejects avoidable bloat in always-on guidance and template growth.

Run the local gate after template or harness edits:

node scripts/template-fitness.mjs

To check an automation proposal file before accepting it:

node scripts/template-fitness.mjs --suggestion path/to/suggestion.md

How To Use

Copy the template into a target repository and adapt it to that repository's actual stack, workflows, risks, and existing documentation. Do not copy every optional module by default. The template is intentionally more detailed than the files it asks you to create.

Start with the required core, then add optional modules only when the repository has a real trigger for them and a smaller existing control would not be enough.

License

This project is dedicated to the public domain under CC0 1.0 Universal.

You may copy, modify, distribute, and use the template without permission or attribution. Credit is appreciated but not required.

About

Bootstrap template for self-maintaining coding-agent harnesses

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors