adrielp AI Engineering Harness — Summary

AI Engineering Harness by adrielp is a multi-tool context engineering framework that distributes a shared set of agents, skills, and commands across Claude Code, OpenCode, Gemini CLI, and Pi via a Deno-based CLI installer (ai-harness). The framework's primary innovation is format translation: the same logical commands and agents are rendered in the correct format for each target tool (Claude Code .claude/skills/, OpenCode .config/opencode/, Gemini CLI TOML, Pi prompts/). It ships 25 skills/commands including a notable OpenTelemetry-focused cluster (otel_instrument orchestrator + 5 specialized OTel sub-skills) and TDD/worktree/codebase-research workflows. Six specialized agents (codebase_analyzer, codebase_locator, codebase_pattern_finder, thoughts_analyzer, thoughts_locator, web_search_researcher) are shared across all tools. The thoughts/ directory pattern is its distinctive innovation: structured context for tickets, plans, and research, separated by person (thoughts/<username>/) and topic. Differs from seeds: closest to openspec (Archetype 2: mirror commands + skills per tool) but adds multi-tool format translation and a dedicated Deno CLI installer, making it the only framework in this batch whose install-time tooling is itself a named CLI binary.

adrielp AI Engineering Harness — Overview

Origin

GitHub: https://github.com/adrielp/ai-engineering-harness
Stars: 16
License: Apache-2.0
Language: TypeScript (Deno)
Last commit: 2026-05-15

Philosophy

From README:

"A harness for AI coding agents that provides context engineering patterns, commands, and configurations. Install once and instantly configure Claude Code, OpenCode, and Gemini CLI with battle-tested prompts, agents, and workflows."

The framework philosophy is "context engineering" — the belief that the quality of AI coding output is determined by the quality of the context provided before coding begins. The thoughts/ directory system represents this: a structured knowledge base of tickets, plans, and research that agents can read.

Key Design Decisions

1. Format translation: the same logical skill/command is rendered differently for each tool.
2. GNU Stow support: repo-mode install uses symlinks for easy git pull updates.
3. Safe re-run: unchanged files are skipped, modified files show diff + confirm.
4. Deno over Node: the CLI uses Deno for zero-install potential (npx deno fallback).
5. thoughts/ as first-class artifact: personal and shared context directories, template-driven.
6. OpenTelemetry cluster: the otel_instrument orchestrator is an unusual specialization for observability-driven development.

adrielp AI Engineering Harness — Architecture

Distribution

Deno-based CLI installer with multiple install paths:

# Register CLI (one-time)
deno install -Agf -n ai-harness \
  https://raw.githubusercontent.com/adrielp/ai-engineering-harness/main/install.ts

# Install for specific tool
ai-harness --tool=claude          # → ~/.claude/
ai-harness --tool=opencode        # → ~/.config/opencode/
ai-harness --tool=gemini          # → ~/.gemini/
ai-harness --tool=pi              # → ~/.pi/agent/
ai-harness --tool=all             # All four

# Flags
ai-harness --tool=claude --dry-run        # Preview changes
ai-harness --tool=claude --interactive    # Pick components
ai-harness --tool=claude --skill=agents   # Specific component

Directory Tree

ai-engineering-harness/
├── claude/
│   ├── agents/            # 6 shared agent definitions (.md)
│   ├── skills/            # 25 skills/commands
│   ├── settings.json      # Claude Code settings
│   └── .mcp.json          # MCP server config
├── gemini/                # Gemini CLI format (TOML)
├── opencode/              # OpenCode format
├── pi/                    # Pi agent format
├── install.ts             # Deno installer CLI
├── manifest.json          # Component manifest
├── setup.sh               # GNU Stow setup script
└── CLAUDE.md / AGENTS.md  # Project memory

Required Runtime

• Deno (or npx deno) for CLI installer
• Target tools: Claude Code, OpenCode, Gemini CLI, Pi

Install Complexity

Multi-step (one-time CLI registration + install command per tool).

Re-run Safety

Unchanged files skipped. Modified files show diff + confirmation prompt.

adrielp AI Engineering Harness — Components

Agents (6, shared across all tools)

Skills / Commands (25)

Config Files

• claude/settings.json — Claude Code permissions and hooks config
• claude/.mcp.json — MCP server configuration
• manifest.json — component manifest for installer

Format Translation

adrielp AI Engineering Harness — Uniqueness

differs_from_seeds

Closest to openspec (Archetype 2: mirror commands + skills, multi-tool portability) but openspec is npm-distributed while this uses a Deno installer CLI with format translation at install time rather than at runtime. Also resembles spec-kit (explicit command/skill pairs with TDD discipline) but spec-kit is Python-based and Claude Code-only. Unlike any seed, this framework targets 4 different AI tools (Claude Code, OpenCode, Gemini CLI, Pi) and uses the installer to transform a single component definition into each tool's native format. The thoughts/ directory pattern is not present in any seed — it provides structured multi-person knowledge management as a first-class artifact.

Most Distinctive Features

1. Multi-tool format translation at install time — no seed does this.
2. OpenTelemetry specialization — otel_instrument orchestrator with 4 specialized OTel sub-skills is a unique domain specialization.
3. thoughts/ personal/shared separation — knowledge management pattern not present in any seed.

Observable Failure Modes

1. Deno requirement: Deno is less common than Node.js — teams may need to install it.
2. Low stars: 16 stars suggests minimal community validation.
3. No hooks: no lifecycle event hooks means no automatic enforcement.
4. Format drift: maintaining 4 format translations manually requires discipline; the installer is one-directional.