Meta Harness

Meta Harness is a portable, standards-first meta-skill for designing domain-specific workflows, reusable specialist skills, and deterministic handoff artifacts.

Current project version: 0.3. See Changelog for checkpoint-based version history.

Adapted from the original Harness project and distributed here under the same Apache 2.0 license.

What This Adds

Compared to the original Claude Code-based Harness, this project adds:

• a standards-first repository layout built around AGENTS.md, .agents/skills/, and docs/harness/
• concise, human-written AGENTS.md guidance built around repo-wide WHAT / WHY / HOW plus progressive disclosure
• runtime-neutral artifact contracts based on skills, team specs, and deterministic _workspace/ handoffs
• rippable harness design guidance that keeps temporary model-specific recovery logic easy to remove
• a repo-local bootstrap installer for project-level and user-level skill installs
• a tighter maintenance loop through repo-local validation and simpler, platform-independent conventions

What It Includes

• a 6-phase workflow for analysis, architecture, generation, integration, and validation
• 6 architecture patterns: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation
• an autonomous experimentation workflow profile for user-controlled compute
• repo-local skills under .agents/skills/
• durable output specs under docs/harness/
• deterministic _workspace/ handoff conventions
• a small bootstrap installer plus validation and smoke-test scripts

Docs

• Installation
• Compatibility Guides
• Sample Prompts
• Changelog
• Harness Output Specs
• Starter Research Example
• AGENTS Authoring Guide
• Orchestrator Template

Repository Layout

meta-harness/
├── AGENTS.md
├── .agents/skills/harness/
│   ├── SKILL.md
│   └── references/
├── docs/harness/README.md
├── docs/harness/starter-research/
├── scripts/install_harness.py
├── scripts/test_install_harness.py
├── scripts/validate_codex_port.py
└── LICENSE

Install

Install into a project:

python3 scripts/install_harness.py --scope project --target /path/to/repo --layout standard

Install as a user-level shared skill:

python3 scripts/install_harness.py --scope user --layout standard

Use --layout forgecode, --layout droid, --layout openhands, or --layout aider when you want a client-specific mirror or follow-up guidance. Use --layout codex when you want Codex to see both the shared Harness tree and the native .codex/skills/harness/ mirror. Harness installs the skill tree only; the target repository keeps ownership of its own AGENTS.md, README.md, and docs. See Compatibility Guides for path mappings and agent-specific follow-up.

Use

1. Read AGENTS.md.
2. Read the main skill at .agents/skills/harness/SKILL.md.
3. When a target repository needs durable repo-wide guidance, start from the AGENTS Authoring Guide and keep AGENTS.md limited to repo-wide WHAT / WHY / HOW.
4. Generate the smallest durable artifact set that fits the domain:
   • .agents/skills/<domain>-orchestrator/SKILL.md
   • .agents/skills/<specialist>/SKILL.md
   • docs/harness/<domain>/team-spec.md
   • _workspace/{phase}_{role}_{artifact}.md

Generated SKILL.md files should begin with YAML frontmatter containing at least name and description so native skill discovery can reliably select repo-specific generated skills.

Good requests for Harness:

Build a reusable research harness for this repository.
Design a review workflow with explicit QA handoffs.
Define specialist skills and a team spec for this domain.
Design an autonomous experiment harness for this repository with a fixed metric and a deterministic results ledger.

Workflow and Patterns

The main skill preserves the 6-phase workflow:

1. Domain Analysis
2. Team Architecture Design
3. Role and Artifact Definition Generation
4. Skill Generation
5. Integration and Orchestration
6. Validation and Testing

Pattern guidance lives in .agents/skills/harness/references/agent-design-patterns.md. Output-spec conventions live in docs/harness/README.md. Use the AGENTS guide when you need a short always-loaded repo contract, and keep evolving retry or recovery logic in rippable harness docs instead of the root file. Start from the orchestrator template when you need a durable workflow spec, or adapt the starter research example when you want a concrete minimal package.

Validation

python3 scripts/test_install_harness.py
python3 scripts/validate_codex_port.py

The smoke test checks the installer across project and user scopes. The validator checks required files, README links, the short AGENTS.md contract, main-skill headings, pattern coverage, and the absence of removed runtime-specific paths in the canonical docs.

License

Apache 2.0