AI Agent Harness

AI Agent Harness is a public orchestration and skill framework for governed AI-assisted coding workflows.

It publishes the control surface, role contracts, and review gates for a constrained multi-agent coding pipeline. It is meant to be studied, adapted, or reused by readers who want stronger execution discipline than an unconstrained agent loop normally provides.

What This Repository Contains

• Orchestration/ Canonical constraints, schemas, guidance documents, templates, and versioned rule surfaces.
• Skill/legacy-memory-governor/ Compacts oversized role memory and archives older governance history into SQLite.
• Skill/orchestration-router/ Routes work into the correct role and orchestration entry path.
• Skill/orchestration-l1/ Defines the L1 governance and review role.
• Skill/orchestration-l2/ Defines the L2 planning and task-decomposition role.
• Skill/orchestration-l3/ Defines the L3 execution role.
• Skill/orchestration-artifact-validator/ Performs mechanical structural validation for governed artifacts.
• Skill/orchestration-critic/ Performs one bounded adversarial critique pass over a target artifact.
• README.md and LICENSE Public package metadata and licensing.

Workflow Model At A Glance

• The router decides which role should act next and which artifact should be produced.
• L1 owns governance intake, review authority, planning release, and final integration review.
• L2 owns planning artifacts such as compatibility checks, split strategy, task trees, and task packets.
• L3 owns bounded recon and bounded execution within an already approved scope.
• The artifact validator acts as the shared structural gate.
• The critic acts as a bounded adversarial pressure layer before ordinary approval or handoff.

How To Read This Repository

If you are new to the repository, read in this order:

1. Orchestration/Version-Manifest.md Use this as the authoritative pointer to the active rule set.
2. Orchestration/0. Pipeline-AI Agentic Coding.md Read the end-to-end pipeline model first.
3. Orchestration/Constraint-System.md Read the system-wide invariants and shared orchestration rules.
4. Orchestration/Constraint-L1.md, Orchestration/Constraint-L2.md, and Orchestration/Constraint-L3.md Read the role-local constraints next.
5. The relevant Skill/<name>/SKILL.md Use the matching skill contract once you know which role or helper surface you care about.

Public Package Boundary

This public package contains:

• Orchestration/
• selected public skills under Skill/
• root-level public documentation such as README.md

Local working directories that may exist beside this package are not part of the public repository contract.

Scope And Non-Goals

This repository is a governance-facing and orchestration-facing package. It is not presented as:

• a turnkey hosted service,
• a complete local runtime bundle,
• a secret-management layer,
• a model-provider abstraction product,
• or a packaged end-user application.

Unless explicitly documented here, callers still own:

• model selection,
• secret management,
• execution wiring,
• storage backends,
• local automation,
• deployment,
• and environment-specific wrappers.

Consumption Rules

When consuming this repository:

• treat Orchestration/ as the canonical rule surface,
• treat each public Skill/<name>/SKILL.md file as the contract for that skill,
• do not assume undocumented local wrappers are part of the supported surface,
• and do not treat absent runtime machinery as an accidental omission.

This repository is intentionally published as a controlled, auditable subset rather than as a full dump of every surrounding local workflow.

Portability Rules

• Repository-facing documentation is written in English.
• Committed files must not contain machine-local absolute paths.
• Environment assumptions must remain portable and caller-configurable.