Portable agent workflow design
Turn a project goal into reusable skills and inspectable handoffs.
Documentation · Source · Apache 2.0
Meta Harness is a runtime-neutral meta-skill for designing domain-specific workflows, reusable specialist skills, and deterministic handoff artifacts. It is adapted from the original Harness project and distributed under the Apache 2.0 license.
Current project version: 0.8.4. Meta Harness actively supports Pi, Codex,
Antigravity, and Cursor CLI/Agent through one portable workflow model and
runtime adapters. Generic Agent Skills use is best effort; legacy client
layouts are unverified and deprecated. See the changelog for
the checkpoint-based project history.
The Meta Harness documentation portal is the recommended entry point. It covers installation, the Phase 0 inventory and drift audit, the six execution phases, architecture patterns, prompt design, compatibility, and durable output specs.
For a quick project install, use the installed meta-harness CLI:
meta-harness install \
--scope project \
--target /path/to/repo \
--agent generic \
--non-interactiveWhen running directly from a source checkout without the CLI on PATH, use
./meta-harness (macOS/Linux), .\meta-harness (Windows), or
python scripts/install_harness.py with the same subcommand and options. This
is the current planner entry point, not the deprecated --layout interface.
Then ask for a goal-shaped workflow, for example:
Design a reusable research harness for this repository.
Keep the handoffs deterministic and validate one normal and one failure flow.
- a Phase 0 inventory and drift audit followed by six execution phases from domain analysis through validation;
- six coordination patterns: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation;
- portable skills under
.agents/skills/; - durable team specs and role contracts under
docs/harness/; - deterministic
_workspace/handoffs when inspection or resumption matters; - a shared CLI/TUI installer planner with portable layout and optional native profiles;
- audit, doctor, compile, and validation commands over inspectable plans;
- removable runtime adapters for Pi, Codex, Antigravity, and Cursor CLI/Agent without forking the canonical skill;
- explicit capability degradation and rippability rules for runtime profiles.
| Phase | Question it answers |
|---|---|
| Inventory and drift audit | What already exists, what is stale, and what operation is safe? |
| Domain analysis | What is this project, task, and quality bar? |
| Team architecture | What coordination shape earns its complexity? |
| Role and artifact definition | Who owns each output and handoff? |
| Skill generation | What reusable behavior belongs in a skill? |
| Integration and orchestration | How does information move between phases? |
| Validation and testing | Does the workflow work, fail clearly, and stay maintainable? |
Read the workflow guide and pattern guide for the operational details.
Install the meta-harness executable onto your system:
macOS / Linux:
curl -fsSL https://raw.githubusercontent.com/SaehwanPark/meta-harness/main/install.sh | bashWindows (PowerShell):
irm https://raw.githubusercontent.com/SaehwanPark/meta-harness/main/install.ps1 | iexThe installer verifies Python 3.8+, installs the executable launcher, and ensures it is available on your PATH.
From a cloned repository:
git clone https://github.com/SaehwanPark/meta-harness.git
cd meta-harness
# Install to user PATH (~/.local/bin on Unix, user PATH on Windows):
./install.sh # macOS / Linux
.\install.ps1 # Windows PowerShell
# Or run directly from the repo root:
./meta-harness # macOS / Linux
.\meta-harness # WindowsInstall into a project:
meta-harness install \
--scope project \
--target /path/to/repo \
--agent pi \
--agent cursor \
--non-interactiveInstall as a user-level shared skill:
meta-harness install \
--scope user \
--agent generic \
--non-interactive--agent is repeatable. Runtime-specific installation and capability guidance
for the actively supported Pi, Codex, Antigravity, and Cursor CLI/Agent targets
is in the compatibility matrix. Generic clients
are best-effort only; ForgeCode, Droid, OpenHands, and Aider are retained as
unverified, deprecated migration notes. Use audit, doctor, compile, and
validate for inspectable lifecycle operations.
The installer owns only explicitly planned skill/profile destinations. The
target repository keeps ownership of its AGENTS.md, README.md, and
documentation. Legacy --layout flags remain compatibility aliases and are
unverified/deprecated; use --agent for new automation.
The portable contract is authoritative: .agents/skills/ contains reusable
behavior, docs/harness/ contains role and team contracts, and _workspace/
contains durable handoffs. Runtime-native profiles are removable adapters, not
sources of truth. Read the architecture guide
and compatibility matrix before adding a
runtime-specific integration.
The canonical source is .agents/skills/harness/SKILL.md. Generated skills
must begin with YAML frontmatter containing at least name and description.
Use docs/harness/ for durable team specs and role briefs, and _workspace/
for deterministic intermediate artifacts that need inspection, resumption, or
cross-agent synthesis.
Keep AGENTS.md short and repo-wide. Put conditional detail in skills,
references, or project documentation. Prefer direct work for small tasks and
add workers only when boundaries, ownership, synthesis, and partial-failure
behavior are explicit.
Read the AGENTS Authoring Guide
(.agents/skills/harness/references/agents-md-guide.md) when a target repository
needs durable repo-wide rules. Keep temporary model-specific recovery logic in
a rippable harness layer. Every generated skill starts with YAML frontmatter
and declares its name and description.
Run the repository checks from the project root:
python3 scripts/validate_pages.py
python3 scripts/validate_skills.py
python3 scripts/validate_adapters.py
python3 scripts/test_install_harness.py
python3 scripts/test_install_planner.py
python3 scripts/test_installer_tui.py
python3 scripts/test_profile_compilation.py
python3 scripts/test_audit_harness.py
python3 scripts/validate_codex_port.pyThe Pages check protects rendered source and navigation. Portable-skill and adapter validators protect frontmatter, links, capability guidance, and fixtures. The installer tests cover project/user scopes, multi-runtime plans, legacy aliases, dry runs, idempotent updates, conflicts, profiles, and symlink mode. The Codex-port validator protects canonical paths, synchronized docs, and legacy-path exclusions.
Apache 2.0. See LICENSE.
